diff --git a/third_party/ggml/fp16.c b/third_party/ggml/fp16.c new file mode 100644 index 000000000..ccba5dda6 --- /dev/null +++ b/third_party/ggml/fp16.c @@ -0,0 +1,125 @@ +/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-│ +│vi: set net ft=c ts=4 sts=4 sw=4 fenc=utf-8 :vi│ +╚──────────────────────────────────────────────────────────────────────────────╝ +│ │ +│ GGML │ +│ Copyright (c) 2023 Georgi Gerganov │ +│ │ +│ Permission is hereby granted, free of charge, to any person obtaining │ +│ a copy of this software and associated documentation files (the │ +│ "Software"), to deal in the Software without restriction, including │ +│ without limitation the rights to use, copy, modify, merge, publish, │ +│ distribute, sublicense, and/or sell copies of the Software, and to │ +│ permit persons to whom the Software is furnished to do so, subject to │ +│ the following conditions: │ +│ │ +│ The above copyright notice and this permission notice shall be │ +│ included in all copies or substantial portions of the Software. │ +│ │ +│ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, │ +│ EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF │ +│ MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. │ +│ IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY │ +│ CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, │ +│ TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE │ +│ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. │ +│ │ +╚─────────────────────────────────────────────────────────────────────────────*/ +#include "third_party/ggml/fp16.h" +#include "libc/str/str.h" +#include "third_party/ggml/fp16.internal.h" +#include "third_party/libcxx/math.h" + +asm(".ident\t\"\\n\\n\ +GGML (MIT License)\\n\ +Copyright (c) 2023 Georgi Gerganov\""); +asm(".include \"libc/disclaimer.inc\""); +// clang-format off + +#if defined(__ARM_NEON) || defined(__wasm_simd128__) +#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s +#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s) +#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s) +#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s) +#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s) +#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s) +#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s) +#define B8(c,s ) B7(c,s, c), B7(c,s, s) +// precomputed tables for expanding 8bits to 8 bytes (shl 4) +const uint64_t table_b2b_u[1 << 8] = { B8(00, 10) }; +#endif + +// +// global data +// + +// precomputed gelu table for f16 (128 KB) +ggml_fp16_t table_gelu_f16[1 << 16]; + +// precomputed silu table for f16 (128 KB) +ggml_fp16_t table_silu_f16[1 << 16]; + +// precomputed exp table for f16 (128 KB) +ggml_fp16_t table_exp_f16[1 << 16]; + +// precomputed f32 table for f16 (256 KB) +float table_f32_f16[1 << 16]; + +// note: do not use these inside ggml.c +// these are meant to be used via the ggml.h API +float ggml_fp16_to_fp32(ggml_fp16_t x) { + return (float) GGML_FP16_TO_FP32(x); +} + +ggml_fp16_t ggml_fp32_to_fp16(float x) { + return GGML_FP32_TO_FP16(x); +} + +void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, size_t n) { + for (size_t i = 0; i < n; i++) { + y[i] = GGML_FP16_TO_FP32(x[i]); + } +} + +void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, size_t n) { + size_t i = 0; +#if defined(__F16C__) + for (; i + 7 < n; i += 8) { + __m256 x_vec = _mm256_loadu_ps(x + i); + __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); + _mm_storeu_si128((__m128i *)(y + i), y_vec); + } + for(; i + 3 < n; i += 4) { + __m128 x_vec = _mm_loadu_ps(x + i); + __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); + _mm_storel_epi64((__m128i *)(y + i), y_vec); + } +#endif + for (; i < n; i++) { + y[i] = GGML_FP32_TO_FP16(x[i]); + } +} + +static const float GELU_COEF_A = 0.044715f; +static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; + +inline static float ggml_gelu_f32(float x) { + return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); +} + +// Sigmoid Linear Unit (SiLU) function +inline static float ggml_silu_f32(float x) { + return x/(1.0f + expf(-x)); +} + +void ggml_fp16_init(void) { + ggml_fp16_t ii; + for (int i = 0; i < (1 << 16); ++i) { + uint16_t ui = i; + memcpy(&ii, &ui, sizeof(ii)); + const float f = table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(ii); + table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f)); + table_silu_f16[i] = GGML_FP32_TO_FP16(ggml_silu_f32(f)); + table_exp_f16[i] = GGML_FP32_TO_FP16(expf(f)); + } +} diff --git a/third_party/ggml/fp16.h b/third_party/ggml/fp16.h new file mode 100644 index 000000000..37e746bc2 --- /dev/null +++ b/third_party/ggml/fp16.h @@ -0,0 +1,27 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_F16_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_F16_H_ +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ + +#define GGML_GELU_FP16 +#define GGML_SILU_FP16 + +#ifdef __ARM_NEON +// we use the built-in 16-bit float type +typedef __fp16 ggml_fp16_t; +#else +typedef uint16_t ggml_fp16_t; +#endif + +void ggml_fp16_init(void); + +// convert FP16 <-> FP32 +float ggml_fp16_to_fp32(ggml_fp16_t x); +ggml_fp16_t ggml_fp32_to_fp16(float x); + +void ggml_fp16_to_fp32_row(const ggml_fp16_t* x, float* y, size_t n); +void ggml_fp32_to_fp16_row(const float* x, ggml_fp16_t* y, size_t n); + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_F16_H_ */ diff --git a/third_party/ggml/fp16.internal.h b/third_party/ggml/fp16.internal.h new file mode 100644 index 000000000..d2a9e267b --- /dev/null +++ b/third_party/ggml/fp16.internal.h @@ -0,0 +1,188 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_FP16_INTERNAL_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_FP16_INTERNAL_H_ +#include "libc/literal.h" +#include "libc/str/str.h" +#include "third_party/ggml/fp16.h" +#include "third_party/intel/immintrin.internal.h" +#include "third_party/libcxx/math.h" +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ + +extern ggml_fp16_t table_gelu_f16[1 << 16]; +extern ggml_fp16_t table_silu_f16[1 << 16]; +extern ggml_fp16_t table_exp_f16[1 << 16]; +extern float table_f32_f16[1 << 16]; +#if defined(__ARM_NEON) || defined(__wasm_simd128__) +extern const uint64_t table_b2b_u[1 << 8]; +#endif + +inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t* y, + const ggml_fp16_t* x) { + const uint16_t* i16 = (const uint16_t*)x; + for (int i = 0; i < n; ++i) { + y[i] = table_gelu_f16[i16[i]]; + } +} + +// clang-format off + +// 16-bit float +// on Arm, we use __fp16 +// on x86, we use uint16_t +#ifdef __ARM_NEON + +// if YCM cannot find , make a symbolic link to it, for example: +// +// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/ +// + +#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x)) +#define GGML_COMPUTE_FP32_TO_FP16(x) (x) + +#define GGML_FP16_TO_FP32(x) ((float) (x)) +#define GGML_FP32_TO_FP16(x) (x) + +#else + +#ifdef __F16C__ + +#ifdef _MSC_VER +#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x))) +#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0) +#else +#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x) +#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0) +#endif + +#elif defined(__POWER9_VECTOR__) + +#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) +#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) +/* the inline asm below is about 12% faster than the lookup method */ +#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) +#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) + +static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + register float f; + register double d; + __asm__( + "mtfprd %0,%2\n" + "xscvhpdp %0,%0\n" + "frsp %1,%0\n" : + /* temp */ "=d"(d), + /* out */ "=f"(f): + /* in */ "r"(h)); + return f; +} + +static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { + register double d; + register ggml_fp16_t r; + __asm__( /* xscvdphp can work on double or single precision */ + "xscvdphp %0,%2\n" + "mffprd %1,%0\n" : + /* temp */ "=d"(d), + /* out */ "=r"(r): + /* in */ "f"(f)); + return r; +} + +#else + +// FP16 <-> FP32 +// ref: https://github.com/Maratyszcza/FP16 + +static inline float fp32_from_bits(uint32_t w) { + union { + uint32_t as_bits; + float as_value; + } fp32; + fp32.as_bits = w; + return fp32.as_value; +} + +static inline uint32_t fp32_to_bits(float f) { + union { + float as_value; + uint32_t as_bits; + } fp32; + fp32.as_value = f; + return fp32.as_bits; +} + +static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { + const uint32_t w = (uint32_t) h << 16; + const uint32_t sign = w & UINT32_C(0x80000000); + const uint32_t two_w = w + w; + + const uint32_t exp_offset = UINT32_C(0xE0) << 23; +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__) + const float exp_scale = 0x1.0p-112f; +#else + const float exp_scale = fp32_from_bits(UINT32_C(0x7800000)); +#endif + const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale; + + const uint32_t magic_mask = UINT32_C(126) << 23; + const float magic_bias = 0.5f; + const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias; + + const uint32_t denormalized_cutoff = UINT32_C(1) << 27; + const uint32_t result = sign | + (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); + return fp32_from_bits(result); +} + +static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { +#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__) + const float scale_to_inf = 0x1.0p+112f; + const float scale_to_zero = 0x1.0p-110f; +#else + const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000)); + const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000)); +#endif + float base = (fabsf(f) * scale_to_inf) * scale_to_zero; + + const uint32_t w = fp32_to_bits(f); + const uint32_t shl1_w = w + w; + const uint32_t sign = w & UINT32_C(0x80000000); + uint32_t bias = shl1_w & UINT32_C(0xFF000000); + if (bias < UINT32_C(0x71000000)) { + bias = UINT32_C(0x71000000); + } + + base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base; + const uint32_t bits = fp32_to_bits(base); + const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00); + const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF); + const uint32_t nonsign = exp_bits + mantissa_bits; + return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign); +} + +#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) +#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) + +#endif // __F16C__ + +#endif // __ARM_NEON + +// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, +// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON. +// This is also true for POWER9. +#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16) + +inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { + uint16_t s; + memcpy(&s, &f, sizeof(uint16_t)); + return table_f32_f16[s]; +} + +#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x) +#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) + +#endif + + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_FP16_INTERNAL_H_ */ diff --git a/third_party/ggml/ggjt.v1.c b/third_party/ggml/ggjt.v1.c new file mode 100644 index 000000000..c1c141b6e --- /dev/null +++ b/third_party/ggml/ggjt.v1.c @@ -0,0 +1,155 @@ +/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-│ +│vi: set net ft=c ts=4 sts=4 sw=4 fenc=utf-8 :vi│ +╞══════════════════════════════════════════════════════════════════════════════╡ +│ Copyright 2023 Justine Alexandra Roberts Tunney │ +│ │ +│ Permission to use, copy, modify, and/or distribute this software for │ +│ any purpose with or without fee is hereby granted, provided that the │ +│ above copyright notice and this permission notice appear in all copies. │ +│ │ +│ THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL │ +│ WARRANTIES WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED │ +│ WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE │ +│ AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL │ +│ DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR │ +│ PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER │ +│ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR │ +│ PERFORMANCE OF THIS SOFTWARE. │ +╚─────────────────────────────────────────────────────────────────────────────*/ +#include "libc/assert.h" +#include "third_party/ggml/ggjt.v1.q4_0.h" +#include "third_party/ggml/ggjt.v1.q4_1.h" +#include "third_party/ggml/ggjt.v1.q4_2.h" +#include "third_party/ggml/ggjt.v1.q5_0.h" +#include "third_party/ggml/ggjt.v1.q5_1.h" +#include "third_party/ggml/ggjt.v1.q8_0.h" +#include "third_party/ggml/ggjt.v1.q8_1.h" +#include "third_party/ggml/ggml.h" +// clang-format off + +static const int ggjt_v1_blck_size[GGML_TYPE_COUNT] = { + [GGML_TYPE_F32] = 1, + [GGML_TYPE_F16] = 1, + [GGML_TYPE_Q4_0] = V1_QK4_0, + [GGML_TYPE_Q4_1] = V1_QK4_1, + [GGML_TYPE_Q4_2] = V1_QK4_2, + [GGML_TYPE_Q5_0] = V1_QK5_0, + [GGML_TYPE_Q5_1] = V1_QK5_1, + [GGML_TYPE_Q8_0] = V1_QK8_0, + [GGML_TYPE_Q8_1] = V1_QK8_1, + [GGML_TYPE_I8] = 1, + [GGML_TYPE_I16] = 1, + [GGML_TYPE_I32] = 1, +}; + +static const size_t ggjt_v1_type_size[GGML_TYPE_COUNT] = { + [GGML_TYPE_F32] = sizeof(float), + [GGML_TYPE_F16] = sizeof(ggml_fp16_t), + [GGML_TYPE_Q4_0] = sizeof(block_v1_q4_0), + [GGML_TYPE_Q4_1] = sizeof(block_v1_q4_1), + [GGML_TYPE_Q4_2] = sizeof(block_v1_q4_2), + [GGML_TYPE_Q5_0] = sizeof(block_v1_q5_0), + [GGML_TYPE_Q5_1] = sizeof(block_v1_q5_1), + [GGML_TYPE_Q8_0] = sizeof(block_v1_q8_0), + [GGML_TYPE_Q8_1] = sizeof(block_v1_q8_1), + [GGML_TYPE_I8] = sizeof(int8_t), + [GGML_TYPE_I16] = sizeof(int16_t), + [GGML_TYPE_I32] = sizeof(int32_t), +}; + +static const char *const ggjt_v1_type_name[GGML_TYPE_COUNT] = { + [GGML_TYPE_F32] = "f32", + [GGML_TYPE_F16] = "f16", + [GGML_TYPE_Q4_0] = "q4_0", + [GGML_TYPE_Q4_1] = "q4_1", + [GGML_TYPE_Q4_2] = "q4_2", + [GGML_TYPE_Q5_0] = "q5_0", + [GGML_TYPE_Q5_1] = "q5_1", + [GGML_TYPE_Q8_0] = "q8_0", + [GGML_TYPE_Q8_1] = "q8_1", + [GGML_TYPE_I8] = "i8", + [GGML_TYPE_I16] = "i16", + [GGML_TYPE_I32] = "i32", +}; + +static const bool ggjt_v1_is_quantized[GGML_TYPE_COUNT] = { + [GGML_TYPE_F32] = false, + [GGML_TYPE_F16] = false, + [GGML_TYPE_Q4_0] = true, + [GGML_TYPE_Q4_1] = true, + [GGML_TYPE_Q4_2] = true, + [GGML_TYPE_Q5_0] = true, + [GGML_TYPE_Q5_1] = true, + [GGML_TYPE_Q8_0] = true, + [GGML_TYPE_Q8_1] = true, + [GGML_TYPE_I8] = false, + [GGML_TYPE_I16] = false, + [GGML_TYPE_I32] = false, +}; + +static const quantize_fns_t ggjt_v1_quantize_fns[GGML_TYPE_COUNT] = { + [GGML_TYPE_Q4_0] = { + .dequantize_row_q = dequantize_row_v1_q4_0, + .quantize_row_q = quantize_row_v1_q4_0, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_v1_q4_0_reference, + .quantize_row_q_dot = quantize_row_v1_q8_0, + .vec_dot_q = ggml_vec_dot_v1_q4_0_q8_0, + .vec_dot_type = GGML_TYPE_Q8_0, + }, + [GGML_TYPE_Q4_1] = { + .dequantize_row_q = dequantize_row_v1_q4_1, + .quantize_row_q = quantize_row_v1_q4_1, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_v1_q4_1_reference, + .quantize_row_q_dot = quantize_row_v1_q8_1, + .vec_dot_q = ggml_vec_dot_v1_q4_1_q8_1, + .vec_dot_type = GGML_TYPE_Q8_1, + }, + [GGML_TYPE_Q4_2] = { + .dequantize_row_q = dequantize_row_v1_q4_2, + .quantize_row_q = quantize_row_v1_q4_2, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_v1_q4_2_reference, + .quantize_row_q_dot = quantize_row_v1_q8_0, + .vec_dot_q = ggml_vec_dot_v1_q4_2_q8_0, + .vec_dot_type = GGML_TYPE_Q8_0, + }, + [GGML_TYPE_Q5_0] = { + .dequantize_row_q = dequantize_row_v1_q5_0, + .quantize_row_q = quantize_row_v1_q5_0, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_v1_q5_0_reference, + .quantize_row_q_dot = quantize_row_v1_q8_0, + .vec_dot_q = ggml_vec_dot_v1_q5_0_q8_0, + .vec_dot_type = GGML_TYPE_Q8_0, + }, + [GGML_TYPE_Q5_1] = { + .dequantize_row_q = dequantize_row_v1_q5_1, + .quantize_row_q = quantize_row_v1_q5_1, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_v1_q5_1_reference, + .quantize_row_q_dot = quantize_row_v1_q8_1, + .vec_dot_q = ggml_vec_dot_v1_q5_1_q8_1, + .vec_dot_type = GGML_TYPE_Q8_1, + }, + [GGML_TYPE_Q8_0] = { + .dequantize_row_q = dequantize_row_v1_q8_0, + .quantize_row_q = quantize_row_v1_q8_0, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_v1_q8_0_reference, + .quantize_row_q_dot = quantize_row_v1_q8_0, + .vec_dot_q = ggml_vec_dot_v1_q8_0_q8_0, + .vec_dot_type = GGML_TYPE_Q8_0, + }, + [GGML_TYPE_Q8_1] = { + .dequantize_row_q = NULL, // TODO + .quantize_row_q = quantize_row_v1_q8_1, + .quantize_row_q_reference = (quantize_row_q_t) quantize_row_v1_q8_1_reference, + .quantize_row_q_dot = quantize_row_v1_q8_1, + .vec_dot_q = NULL, // TODO + .vec_dot_type = GGML_TYPE_Q8_1, + }, +}; + +void ggjt_v1(void) { + GGML_BLCK_SIZE = ggjt_v1_blck_size; + GGML_TYPE_SIZE = ggjt_v1_type_size; + GGML_TYPE_NAME = ggjt_v1_type_name; + GGML_IS_QUANTIZED = ggjt_v1_is_quantized; + quantize_fns = ggjt_v1_quantize_fns; +} diff --git a/third_party/ggml/ggjt.v1.internal.h b/third_party/ggml/ggjt.v1.internal.h new file mode 100644 index 000000000..5d0149eea --- /dev/null +++ b/third_party/ggml/ggjt.v1.internal.h @@ -0,0 +1,134 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_INTERNAL_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_INTERNAL_H_ +#include "libc/str/str.h" +#include "third_party/intel/immintrin.internal.h" +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ +// clang-format off + +#ifdef __AVX__ +// horizontally add 8 floats +static inline float hsum_float_8(const __m256 x) { + __m128 res = _mm256_extractf128_ps(x, 1); + res = _mm_add_ps(res, _mm256_castps256_ps128(x)); + res = _mm_add_ps(res, _mm_movehl_ps(res, res)); + res = _mm_add_ss(res, _mm_movehdup_ps(res)); + return _mm_cvtss_f32(res); +} +#endif /* AVX */ + +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) +// Unpack 16 4-bit fields into 16 bytes +// The output vector contains 16 bytes, each one in [ 0 .. 15 ] interval +static inline __m128i bytes_from_nibbles_16(const uint8_t * rsi) { + // Load 8 bytes from memory + __m128i tmp = _mm_loadl_epi64( ( const __m128i* )rsi ); + // Expand bytes into uint16_t values + __m128i bytes = _mm_cvtepu8_epi16( tmp ); + // Unpack values into individual bytes + const __m128i lowMask = _mm_set1_epi8( 0xF ); + __m128i high = _mm_andnot_si128( lowMask, bytes ); + __m128i low = _mm_and_si128( lowMask, bytes ); + high = _mm_slli_epi16( high, 4 ); + bytes = _mm_or_si128( low, high ); + return bytes; +} +#endif /* AVX || AVX2 || AVX512 */ + +#if defined(__AVX2__) || defined(__AVX512F__) +// spread 32 bits to 32 bytes { 0x00, 0xFF } +static inline __m256i bytes_from_bits_32(const uint8_t * x) { + uint32_t x32; + memcpy(&x32, x, sizeof(uint32_t)); + const __m256i shuf_mask = _mm256_set_epi64x( + 0x0303030303030303, 0x0202020202020202, + 0x0101010101010101, 0x0000000000000000); + __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask); + const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe); + bytes = _mm256_or_si256(bytes, bit_mask); + return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1)); +} +#endif /* AVX2 || AVX512 */ + +#if defined(__AVX2__) || defined(__AVX512F__) + +// add int16_t pairwise and return as float vector +static inline __m256 sum_i16_pairs_float(const __m256i x) { + const __m256i ones = _mm256_set1_epi16(1); + const __m256i summed_pairs = _mm256_madd_epi16(ones, x); + return _mm256_cvtepi32_ps(summed_pairs); +} + +// Unpack 32 4-bit fields into 32 bytes +// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval +static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) { + // Load 16 bytes from memory + __m128i tmp = _mm_loadu_si128( ( const __m128i* )rsi ); + // Expand bytes into uint16_t values + __m256i bytes = _mm256_cvtepu8_epi16( tmp ); + // Unpack values into individual bytes + const __m256i lowMask = _mm256_set1_epi8( 0xF ); + __m256i high = _mm256_andnot_si256( lowMask, bytes ); + __m256i low = _mm256_and_si256( lowMask, bytes ); + high = _mm256_slli_epi16( high, 4 ); + bytes = _mm256_or_si256( low, high ); + return bytes; +} + +// multiply int8_t, add results pairwise twice and return as float vector +static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) { + // Get absolute values of x vectors + const __m256i ax = _mm256_sign_epi8(x, x); + // Sign the values of the y vectors + const __m256i sy = _mm256_sign_epi8(y, x); +#ifdef __AVXVNNI__ + const __m256i zero = _mm256_setzero_si256(); + const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy); + return _mm256_cvtepi32_ps(summed_pairs); +#else + // Perform multiplication and create 16-bit values + const __m256i dot = _mm256_maddubs_epi16(ax, sy); + return sum_i16_pairs_float(dot); +#endif +} + +static inline __m128i packNibbles( __m256i bytes ) { + // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh +#if defined(__AVX512F__) + const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4); // 0000_0000_abcd_0000 + bytes = _mm256_or_si256(bytes, bytes_srli_4); // 0000_abcd_abcd_efgh + return _mm256_cvtepi16_epi8(bytes); // abcd_efgh +#else + const __m256i lowByte = _mm256_set1_epi16( 0xFF ); + __m256i high = _mm256_andnot_si256( lowByte, bytes ); + __m256i low = _mm256_and_si256( lowByte, bytes ); + high = _mm256_srli_epi16( high, 4 ); + bytes = _mm256_or_si256( low, high ); + // Compress uint16_t lanes into bytes + __m128i r0 = _mm256_castsi256_si128( bytes ); + __m128i r1 = _mm256_extracti128_si256( bytes, 1 ); + return _mm_packus_epi16( r0, r1 ); +#endif +} + +#elif defined(__AVX__) + +static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 ) { + // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh + const __m128i lowByte = _mm_set1_epi16( 0xFF ); + __m128i high = _mm_andnot_si128( lowByte, bytes1 ); + __m128i low = _mm_and_si128( lowByte, bytes1 ); + high = _mm_srli_epi16( high, 4 ); + bytes1 = _mm_or_si128( low, high ); + high = _mm_andnot_si128( lowByte, bytes2 ); + low = _mm_and_si128( lowByte, bytes2 ); + high = _mm_srli_epi16( high, 4 ); + bytes2 = _mm_or_si128( low, high ); + return _mm_packus_epi16( bytes1, bytes2); +} + +#endif /* __AVX__ */ + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_INTERNAL_H_ */ diff --git a/third_party/ggml/ggjt.v1.q4_0.c b/third_party/ggml/ggjt.v1.q4_0.c new file mode 100644 index 000000000..c445f67a2 --- /dev/null +++ b/third_party/ggml/ggjt.v1.q4_0.c @@ -0,0 +1,700 @@ +/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-│ +│vi: set net ft=c ts=4 sts=4 sw=4 fenc=utf-8 :vi│ +╚──────────────────────────────────────────────────────────────────────────────╝ +│ │ +│ GGML │ +│ Copyright (c) 2023 Georgi Gerganov │ +│ │ +│ Permission is hereby granted, free of charge, to any person obtaining │ +│ a copy of this software and associated documentation files (the │ +│ "Software"), to deal in the Software without restriction, including │ +│ without limitation the rights to use, copy, modify, merge, publish, │ +│ distribute, sublicense, and/or sell copies of the Software, and to │ +│ permit persons to whom the Software is furnished to do so, subject to │ +│ the following conditions: │ +│ │ +│ The above copyright notice and this permission notice shall be │ +│ included in all copies or substantial portions of the Software. │ +│ │ +│ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, │ +│ EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF │ +│ MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. │ +│ IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY │ +│ CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, │ +│ TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE │ +│ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. │ +│ │ +╚─────────────────────────────────────────────────────────────────────────────*/ +#include "third_party/ggml/ggjt.v1.q4_0.h" +#include "libc/assert.h" +#include "libc/macros.internal.h" +#include "libc/str/str.h" +#include "third_party/aarch64/arm_neon.h" +#include "third_party/ggml/ggjt.v1.internal.h" +#include "third_party/ggml/ggjt.v1.q8_0.h" +#include "third_party/intel/immintrin.internal.h" +#include "third_party/libcxx/math.h" +// clang-format off + +// quantization for the ggjt.v1.q4_0 file format + +static_assert(sizeof(block_v1_q4_0) == sizeof(float) + V1_QK4_0 / 2, + "wrong q4_0 block size/padding"); +static_assert(sizeof(block_v1_q8_0) == sizeof(float) + V1_QK8_0, + "wrong q8_0 block size/padding"); + +// reference implementation for deterministic creation of model files +void quantize_row_v1_q4_0_reference(const float * restrict x, block_v1_q4_0 * restrict y, int k) { + assert(k % V1_QK4_0 == 0); + const int nb = k / V1_QK4_0; + + uint8_t pp[V1_QK4_0/2]; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + float max = 0.0f; + + for (int l = 0; l < V1_QK4_0; l++) { + const float v = x[i*V1_QK4_0 + l]; + if (amax < fabsf(v)) { + amax = fabsf(v); + max = v; + } + } + + const float d = max / -8; + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + for (int l = 0; l < V1_QK4_0; l += 2) { + const float v0 = x[i*V1_QK4_0 + l + 0]*id; + const float v1 = x[i*V1_QK4_0 + l + 1]*id; + + const uint8_t vi0 = MIN(15, (int8_t)roundf(v0) + 8); + const uint8_t vi1 = MIN(15, (int8_t)roundf(v1) + 8); + + assert(vi0 < 16); + assert(vi1 < 16); + + pp[l/2] = vi0 | (vi1 << 4); + } + + memcpy(y[i].qs, pp, sizeof(pp)); + } +} + +void quantize_row_v1_q4_0(const float * restrict x, void * restrict vy, int k) { + assert(k % V1_QK4_0 == 0); + const int nb = k / V1_QK4_0; + + block_v1_q4_0 * restrict y = vy; + +#if defined(__POWER9_VECTOR__) + const vector float v85 = vec_splats(8.5f); + const vector signed int v15 = vec_splats(15); + for (int i = 0; i < nb; i++) { + float max = 0.0f; + float min = 0.0f; + + vector float asrcv [8]; + vector float srcv [8]; + vector float maxv[8]; + vector float minv[8]; + + for (int l = 0; l < 8; l++) srcv[l] = *(vector float *)(x + i*32 + 4*l); + //for (int l = 0; l < 8; l++) asrcv[l] = vec_abs(srcv[l]); + + for (int l = 0; l < 4; l++) maxv[2*l] = vec_max(asrcv[2*l], asrcv[2*l+1]); + //for (int l = 0; l < 2; l++) maxv[4*l] = vec_max(maxv[4*l], maxv[4*l+2]); + maxv[0] = vec_max(maxv[0], maxv[2]); + maxv[4] = vec_max(maxv[4], maxv[6]); + //for (int l = 0; l < 1; l++) maxv[8*l] = vec_max(maxv[8*l], maxv[8*l+4]); + maxv[0] = vec_max(maxv[0], maxv[4]); + + for (int l = 0; l < 4; l++) minv[2*l] = vec_min(asrcv[2*l], asrcv[2*l+1]); + //for (int l = 0; l < 2; l++) minv[4*l] = vec_min(minv[4*l], minv[4*l+2]); + minv[0] = vec_min(minv[0], minv[2]); + minv[4] = vec_min(minv[4], minv[6]); + //for (int l = 0; l < 1; l++) minv[8*l] = vec_min(minv[8*l], minv[8*l+4]); + minv[0] = vec_min(minv[0], minv[4]); + + + max = MAX( + MAX(vec_extract(maxv[0], 0), vec_extract(maxv[0], 1)), + MAX(vec_extract(maxv[0], 2), vec_extract(maxv[0], 3))); + min = MIN( + MIN(vec_extract(minv[0], 0), vec_extract(minv[0], 1)), + MIN(vec_extract(minv[0], 2), vec_extract(minv[0], 3))); + + const float magnitude = max >= fabsf(min) ? max : min; + const float d = magnitude / -8; + const float id = d ? 1.0/d : 0.0; + + y[i].d = d; + + const vector float vid = vec_splats(id); + uint8_t * restrict pb = y[i].qs; + for (int l = 0; l < 8; l++) { + const vector float vf = vec_madd(srcv[l], vid, v85); + const vector signed int vi = vec_signed(vf); + const vector signed int vc = vec_min(vi, v15); + + pb[2*l + 0] = vec_extract(vc, 0) | (vec_extract(vc, 1) << 4); + pb[2*l + 1] = vec_extract(vc, 2) | (vec_extract(vc, 3) << 4); + } + } +#elif __ARM_NEON + for (int i = 0; i < nb; i++) { + float32x4_t srcv [8]; + float32x4_t maxv[8]; + float32x4_t minv[8]; + + for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); + + for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l+1]); + for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l+2]); + for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l+4]); + + for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l+1]); + for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l+2]); + for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l+4]); + + const float max = vmaxvq_f32(maxv[0]); + const float min = vminvq_f32(minv[0]); + + const float magnitude = max >= fabsf(min) ? max : min; + const float d = magnitude / -8; + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + for (int l = 0; l < 8; l++) { + const float32x4_t v = vmulq_n_f32(srcv[l], id); + const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(8.5f)); + const int32x4_t vi = vcvtq_s32_f32(vf); + const int32x4_t vc = vminq_s32(vi, vdupq_n_s32(15)); + + y[i].qs[2*l + 0] = vgetq_lane_s32(vc, 0) | (vgetq_lane_s32(vc, 1) << 4); + y[i].qs[2*l + 1] = vgetq_lane_s32(vc, 2) | (vgetq_lane_s32(vc, 3) << 4); + } + } +#elif defined(__AVX2__) + for (int i = 0; i < nb; i++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x ); + __m256 v1 = _mm256_loadu_ps( x + 8 ); + __m256 v2 = _mm256_loadu_ps( x + 16 ); + __m256 v3 = _mm256_loadu_ps( x + 24 ); + x += 32; + + // Compute max for the block + __m256 max = _mm256_max_ps( v0, v1 ); + __m256 maxTmp = _mm256_max_ps( v2, v3 ); + max = _mm256_max_ps( max, maxTmp ); + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + // Compute min for the block + __m256 min = _mm256_min_ps( v0, v1 ); + __m256 minTmp = _mm256_min_ps( v2, v3 ); + min = _mm256_min_ps( min, minTmp ); + + __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) ); + min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); + min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); + const float minScalar = _mm_cvtss_f32( min4 ); + + // Quantize these floats + const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar; + const float d = magnitude / -8.0f; + y[i].d = d; + const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f; + const __m256 mul = _mm256_set1_ps( id ); + + // Apply the multiplier + v0 = _mm256_mul_ps( v0, mul ); + v1 = _mm256_mul_ps( v1, mul ); + v2 = _mm256_mul_ps( v2, mul ); + v3 = _mm256_mul_ps( v3, mul ); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); + + // Convert int32 to int16 + i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 + i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 + // Convert int16 to int8 + i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 + + // We got our precious signed bytes, but the order is now wrong + // These AVX2 pack instructions process 16-byte pieces independently + // The following instruction is fixing the order + const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); + i0 = _mm256_permutevar8x32_epi32( i0, perm ); + + // Apply offset and clamp to translate the range from [ -8 .. +8 ] into [ +0 .. +15 ] + const __m256i off = _mm256_set1_epi8( 8 ); + i0 = _mm256_add_epi8( i0, off ); + const __m256i maxNibble = _mm256_set1_epi8( 15 ); + i0 = _mm256_min_epi8( i0, maxNibble ); + + // Compress the vector into 4 bit/value, and store + __m128i res = packNibbles( i0 ); + _mm_storeu_si128( ( __m128i* )y[i].qs, res ); + } +#elif defined(__AVX__) + for (int i = 0; i < nb; i++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x ); + __m256 v1 = _mm256_loadu_ps( x + 8 ); + __m256 v2 = _mm256_loadu_ps( x + 16 ); + __m256 v3 = _mm256_loadu_ps( x + 24 ); + x += 32; + + // Compute max for the block + __m256 max = _mm256_max_ps( v0, v1 ); + __m256 maxTmp = _mm256_max_ps( v2, v3 ); + max = _mm256_max_ps( max, maxTmp ); + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + // Compute min for the block + __m256 min = _mm256_min_ps( v0, v1 ); + __m256 minTmp = _mm256_min_ps( v2, v3 ); + min = _mm256_min_ps( min, minTmp ); + + __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) ); + min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); + min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); + const float minScalar = _mm_cvtss_f32( min4 ); + + // Quantize these floats + const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar; + const float d = magnitude / -8.0f; + y[i].d = d; + const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f; + const __m256 mul = _mm256_set1_ps( id ); + + // Apply the multiplier + v0 = _mm256_mul_ps( v0, mul ); + v1 = _mm256_mul_ps( v1, mul ); + v2 = _mm256_mul_ps( v2, mul ); + v3 = _mm256_mul_ps( v3, mul ); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); + + // Since we don't have in AVX some necessary functions, + // we split the registers in half and call AVX2 analogs from SSE + __m128i ni0 = _mm256_castsi256_si128( i0 ); + __m128i ni1 = _mm256_extractf128_si256( i0, 1); + __m128i ni2 = _mm256_castsi256_si128( i1 ); + __m128i ni3 = _mm256_extractf128_si256( i1, 1); + __m128i ni4 = _mm256_castsi256_si128( i2 ); + __m128i ni5 = _mm256_extractf128_si256( i2, 1); + __m128i ni6 = _mm256_castsi256_si128( i3 ); + __m128i ni7 = _mm256_extractf128_si256( i3, 1); + + // Convert int32 to int16 + ni0 = _mm_packs_epi32( ni0, ni1 ); + ni2 = _mm_packs_epi32( ni2, ni3 ); + ni4 = _mm_packs_epi32( ni4, ni5 ); + ni6 = _mm_packs_epi32( ni6, ni7 ); + // Convert int16 to int8 + ni0 = _mm_packs_epi16( ni0, ni2 ); + ni4 = _mm_packs_epi16( ni4, ni6 ); + + // Apply offset and clamp to translate the range from [ -8 .. +8 ] into [ +0 .. +15 ] + const __m128i off = _mm_set1_epi8( 8 ); + ni0 = _mm_add_epi8( ni0, off ); + ni4 = _mm_add_epi8( ni4, off ); + const __m128i maxNibble = _mm_set1_epi8( 15 ); + ni0 = _mm_min_epi8( ni0, maxNibble ); + ni4 = _mm_min_epi8( ni4, maxNibble ); + + // Compress the vector into 4 bit/value, and store + __m128i res = packNibbles( ni0, ni4 ); + _mm_storeu_si128( ( __m128i* )y[i].qs, res ); + } +#elif defined(__wasm_simd128__) + for (int i = 0; i < nb; i++) { + float max = 0.0f; + float min = 0.0f; + + v128_t srcv [8]; + v128_t maxv[8]; + v128_t minv[8]; + + for (int l = 0; l < 8; l++) srcv[l] = wasm_v128_load(x + i*32 + 4*l); + + for (int l = 0; l < 4; l++) maxv[2*l] = wasm_f32x4_max(srcv[2*l], srcv[2*l+1]); + for (int l = 0; l < 2; l++) maxv[4*l] = wasm_f32x4_max(maxv[4*l], maxv[4*l+2]); + for (int l = 0; l < 1; l++) maxv[8*l] = wasm_f32x4_max(maxv[8*l], maxv[8*l+4]); + + for (int l = 0; l < 4; l++) minv[2*l] = wasm_f32x4_min(srcv[2*l], srcv[2*l+1]); + for (int l = 0; l < 2; l++) minv[4*l] = wasm_f32x4_min(minv[4*l], minv[4*l+2]); + for (int l = 0; l < 1; l++) minv[8*l] = wasm_f32x4_min(minv[8*l], minv[8*l+4]); + + max = MAX( + MAX(wasm_f32x4_extract_lane(maxv[0], 0), wasm_f32x4_extract_lane(maxv[0], 1)), + MAX(wasm_f32x4_extract_lane(maxv[0], 2), wasm_f32x4_extract_lane(maxv[0], 3))); + min = MIN( + MIN(wasm_f32x4_extract_lane(minv[0], 0), wasm_f32x4_extract_lane(minv[0], 1)), + MIN(wasm_f32x4_extract_lane(minv[0], 2), wasm_f32x4_extract_lane(minv[0], 3))); + + const float magnitude = max >= fabsf(min) ? max : min; + const float d = magnitude / -8; + const float id = d ? 1.0/d : 0.0; + + y[i].d = d; + + for (int l = 0; l < 8; l++) { + const v128_t v = wasm_f32x4_mul(srcv[l], wasm_f32x4_splat(id)); + const v128_t vf = wasm_f32x4_add(v, wasm_f32x4_splat(8.5f)); + const v128_t vi = wasm_i32x4_trunc_sat_f32x4(vf); + const v128_t vc = wasm_i32x4_min(vi, wasm_i32x4_splat(15)); + + y[i].qs[2*l + 0] = wasm_i32x4_extract_lane(vc, 0) | (wasm_i32x4_extract_lane(vc, 1) << 4); + y[i].qs[2*l + 1] = wasm_i32x4_extract_lane(vc, 2) | (wasm_i32x4_extract_lane(vc, 3) << 4); + } + } +#else + // scalar + quantize_row_v1_q4_0_reference(x, y, k); +#endif +} + +size_t ggml_quantize_v1_q4_0(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % V1_QK4_0 == 0); + const int nb = k / V1_QK4_0; + + for (int j = 0; j < n; j += k) { + block_v1_q4_0 * restrict y = (block_v1_q4_0 *)dst + j/V1_QK4_0; + + quantize_row_v1_q4_0_reference(src + j, y, k); + + for (int i = 0; i < nb; i++) { + for (int l = 0; l < V1_QK4_0; l += 2) { + const uint8_t vi0 = y[i].qs[l/2] & 0x0F; + const uint8_t vi1 = y[i].qs[l/2] >> 4; + + hist[vi0]++; + hist[vi1]++; + } + } + } + + return (n/V1_QK4_0*sizeof(block_v1_q4_0)); +} + +void dequantize_row_v1_q4_0(const void * restrict vx, float * restrict y, int k) { + assert(k % V1_QK4_0 == 0); + const int nb = k / V1_QK4_0; + + const block_v1_q4_0 * restrict x = vx; + +#if defined(__AVX2__) + for (int i = 0; i < nb; i++) { + // scale factor + const __m256 d_v = _mm256_broadcast_ss(&x[i].d); + + const uint8_t * restrict pp = x[i].qs; + + for (int l = 0; l < V1_QK4_0; l += 32) { + // Load 32x4-bit integers into 32x8-bit integers + __m256i vx8 = bytes_from_nibbles_32(pp+l/2); + + // Subtract 8 from the integers + vx8 = _mm256_sub_epi8(vx8, _mm256_set1_epi8(8)); + + // Convert to 16-bit int + const __m256i vx16_lo = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 0)); + const __m256i vx16_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 1)); + + // Convert to 32-bit int -> float 32 + const __m256 vf[4] = { + _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 0))), + _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 1))), + _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 0))), + _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 1))) + }; + + // Scale and store + for (int j = 0; j < 4; j++) { + const __m256 result = _mm256_mul_ps(vf[j], d_v); + _mm256_storeu_ps(y + i * V1_QK4_0 + l + j*8, result); + } + } + } +#elif defined(__ARM_NEON) + for (int i = 0; i < nb; i++) { + const float32x4_t vd = vdupq_n_f32(x[i].d); + + const uint8_t * restrict pp = x[i].qs; + + for (int l = 0; l < V1_QK4_0; l += 16) { + // Load 16x4-bit integers into 8x8-bit integers + const uint8x8_t v8 = vld1_u8(pp + l/2); + + // Expand 4-bit qs to 8-bit bytes + const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0F)); + const uint8x8_t v1 = vshr_n_u8(v8, 4); + + // Convert to signed 8-bit integers + const int8x8_t vs_0 = vreinterpret_s8_u8(v0); + const int8x8_t vs_1 = vreinterpret_s8_u8(v1); + + // Subtract 8 from each byte + const int8x8_t vb_0 = vsub_s8(vs_0, vdup_n_s8(8)); + const int8x8_t vb_1 = vsub_s8(vs_1, vdup_n_s8(8)); + + // Interleave and combine + const int8x8_t vx_0 = vzip1_s8(vb_0, vb_1); + const int8x8_t vx_1 = vzip2_s8(vb_0, vb_1); + + const int8x16_t vq = vcombine_s8(vx_0, vx_1); + + // convert to 2x int16x8_t + const int16x8_t vi_0 = vmovl_s8(vget_low_s8 (vq)); + const int16x8_t vi_1 = vmovl_s8(vget_high_s8(vq)); + + // convert to 4x float32x4_t + const float32x4_t vf_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16 (vi_0))); + const float32x4_t vf_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vi_0))); + const float32x4_t vf_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16 (vi_1))); + const float32x4_t vf_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vi_1))); + + // Multiply by d + const float32x4_t r0 = vmulq_f32(vf_0, vd); + const float32x4_t r1 = vmulq_f32(vf_1, vd); + const float32x4_t r2 = vmulq_f32(vf_2, vd); + const float32x4_t r3 = vmulq_f32(vf_3, vd); + + // Store + vst1q_f32(y + i*V1_QK4_0 + l + 0, r0); + vst1q_f32(y + i*V1_QK4_0 + l + 4, r1); + vst1q_f32(y + i*V1_QK4_0 + l + 8, r2); + vst1q_f32(y + i*V1_QK4_0 + l + 12, r3); + } + } +#else + // scalar + for (int i = 0; i < nb; i++) { + const float d = x[i].d; + + const uint8_t * restrict pp = x[i].qs; + + for (int l = 0; l < V1_QK4_0; l += 2) { + const uint8_t vi = pp[l/2]; + + const int8_t vi0 = vi & 0x0F; + const int8_t vi1 = vi >> 4; + + const float v0 = (vi0 - 8)*d; + const float v1 = (vi1 - 8)*d; + + //printf("d = %f, vi = %d, vi0 = %d, vi1 = %d, v0 = %f, v1 = %f\n", d, vi, vi0, vi1, v0, v1); + + y[i*V1_QK4_0 + l + 0] = v0; + y[i*V1_QK4_0 + l + 1] = v1; + + assert(!isnan(y[i*V1_QK4_0 + l + 0])); + assert(!isnan(y[i*V1_QK4_0 + l + 1])); + } + } +#endif +} + +void ggml_vec_dot_v1_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / V1_QK8_0; + + assert(n % V1_QK8_0 == 0); + assert(nb % 2 == 0); + + const block_v1_q4_0 * restrict x = vx; + const block_v1_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i += 2) { + const block_v1_q4_0 * restrict x0 = &x[i + 0]; + const block_v1_q4_0 * restrict x1 = &x[i + 1]; + const block_v1_q8_0 * restrict y0 = &y[i + 0]; + const block_v1_q8_0 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + const int8x16_t s8b = vdupq_n_s8(0x8); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // sub 8 + const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); + const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); + const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); + const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); + + // interleave + const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs); + const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs); + const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs); + const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + // dot product into int32x4_t + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; ++i) { + /* Compute combined scale for the block */ + const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + + // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. + const __m256i off = _mm256_set1_epi8( 8 ); + bx = _mm256_sub_epi8( bx, off ); + + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + /* Multiply q with scale and accumulate */ + acc = _mm256_fmadd_ps( d, q, acc ); + } + + *s = hsum_float_8(acc); +#elif defined(__AVX__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; ++i) { + // Compute combined scale for the block + const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); + + __m128i i32[2]; + for (int j = 0; j < 2; ++j) { + // Load 8 bytes, and unpack 4 bit fields into bytes, making 16 bytes + __m128i bx = bytes_from_nibbles_16(x[i].qs + 8*j); + __m128i by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16*j)); + + // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. + const __m128i off = _mm_set1_epi8( 8 ); + bx = _mm_sub_epi8( bx, off ); + + // Get absolute values of x vectors + const __m128i ax = _mm_sign_epi8(bx, bx); + + // Sign the values of the y vectors + const __m128i sy = _mm_sign_epi8(by, bx); + + // Perform multiplication and create 16-bit values + const __m128i dot = _mm_maddubs_epi16(ax, sy); + + const __m128i ones = _mm_set1_epi16(1); + i32[j] = _mm_madd_epi16(ones, dot); + } + + // Convert int32_t to float + __m256 p = _mm256_cvtepi32_ps( _mm256_set_m128i( i32[0], i32[1] )); + // Apply the scale, and accumulate + acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc); + } + + *s = hsum_float_8(acc); +#else + // scalar + float sumf = 0.0; + for (int i = 0; i < nb; i++) { + const float d0 = x[i].d; + const float d1 = y[i].d; + + const uint8_t * restrict p0 = x[i].qs; + const int8_t * restrict p1 = y[i].qs; + + int sumi = 0; + for (int j = 0; j < V1_QK8_0/2; j++) { + const uint8_t v0 = p0[j]; + + const int i0 = (int8_t) (v0 & 0x0F) - 8; + const int i1 = (int8_t) (v0 >> 4) - 8; + + const int i2 = p1[2*j + 0]; + const int i3 = p1[2*j + 1]; + + sumi += i0*i2 + i1*i3; + } + sumf += d0*d1*sumi; + } + *s = sumf; +#endif +} diff --git a/third_party/ggml/ggjt.v1.q4_0.h b/third_party/ggml/ggjt.v1.q4_0.h new file mode 100644 index 000000000..d67f5c51f --- /dev/null +++ b/third_party/ggml/ggjt.v1.q4_0.h @@ -0,0 +1,22 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_V1_Q4_0_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_V1_Q4_0_H_ +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ + +#define V1_QK4_0 32 +typedef struct { + float d; // delta + uint8_t qs[V1_QK4_0 / 2]; // nibbles / quants +} block_v1_q4_0; + +void dequantize_row_v1_q4_0(const void* restrict, float* restrict, int); +size_t ggml_quantize_v1_q4_0(const float*, void*, int, int, int64_t*); +void quantize_row_v1_q4_0(const float* restrict, void* restrict, int); +void quantize_row_v1_q4_0_reference(const float* restrict, + block_v1_q4_0* restrict, int); +void ggml_vec_dot_v1_q4_0_q8_0(const int, float* restrict, const void* restrict, + const void* restrict); + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_V1_Q4_0_H_ */ diff --git a/third_party/ggml/ggjt.v1.q4_1.c b/third_party/ggml/ggjt.v1.q4_1.c new file mode 100644 index 000000000..bdcffecf8 --- /dev/null +++ b/third_party/ggml/ggjt.v1.q4_1.c @@ -0,0 +1,472 @@ +/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-│ +│vi: set net ft=c ts=4 sts=4 sw=4 fenc=utf-8 :vi│ +╚──────────────────────────────────────────────────────────────────────────────╝ +│ │ +│ GGML │ +│ Copyright (c) 2023 Georgi Gerganov │ +│ │ +│ Permission is hereby granted, free of charge, to any person obtaining │ +│ a copy of this software and associated documentation files (the │ +│ "Software"), to deal in the Software without restriction, including │ +│ without limitation the rights to use, copy, modify, merge, publish, │ +│ distribute, sublicense, and/or sell copies of the Software, and to │ +│ permit persons to whom the Software is furnished to do so, subject to │ +│ the following conditions: │ +│ │ +│ The above copyright notice and this permission notice shall be │ +│ included in all copies or substantial portions of the Software. │ +│ │ +│ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, │ +│ EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF │ +│ MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. │ +│ IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY │ +│ CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, │ +│ TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE │ +│ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. │ +│ │ +╚─────────────────────────────────────────────────────────────────────────────*/ +#include "third_party/ggml/ggjt.v1.q4_1.h" +#include "libc/assert.h" +#include "libc/macros.internal.h" +#include "libc/str/str.h" +#include "third_party/aarch64/arm_neon.h" +#include "third_party/ggml/ggjt.v1.internal.h" +#include "third_party/ggml/ggjt.v1.q4_1.h" +#include "third_party/ggml/ggjt.v1.q8_1.h" +#include "third_party/intel/immintrin.internal.h" +#include "third_party/libcxx/math.h" +// clang-format off + +// quantization for the ggjt.v1.q4_1 file format + +static_assert(sizeof(block_v1_q4_1) == 2 * sizeof(float) + V1_QK4_1 / 2, + "wrong q4_1 block size/padding"); +static_assert(sizeof(block_v1_q8_1) == 3*sizeof(float) + V1_QK8_1, + "wrong q8_1 block size/padding"); + +void quantize_row_v1_q4_1_reference(const float * restrict x, void * restrict vy, int k) { + assert(k % V1_QK4_1 == 0); + const int nb = k / V1_QK4_1; + + block_v1_q4_1 * restrict y = vy; + + uint8_t pp[V1_QK4_1/2]; + + for (int i = 0; i < nb; i++) { + float min = FLT_MAX; + float max = -FLT_MAX; + + for (int l = 0; l < V1_QK4_1; l++) { + const float v = x[i*V1_QK4_1 + l]; + if (v < min) min = v; + if (v > max) max = v; + } + + const float d = (max - min) / ((1 << 4) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + y[i].m = min; + + for (int l = 0; l < V1_QK4_1; l += 2) { + const float v0 = (x[i*V1_QK4_1 + l + 0] - min)*id; + const float v1 = (x[i*V1_QK4_1 + l + 1] - min)*id; + + const uint8_t vi0 = roundf(v0); + const uint8_t vi1 = roundf(v1); + + assert(vi0 < 16); + assert(vi1 < 16); + + pp[l/2] = vi0 | (vi1 << 4); + } + + memcpy(y[i].qs, pp, sizeof(pp)); + } +} + +void quantize_row_v1_q4_1(const float * restrict x, void * restrict vy, int k) { + assert(k % V1_QK4_1 == 0); + + const int nb = k / V1_QK4_1; + + block_v1_q4_1 * restrict y = vy; + +#if defined(__AVX2__) + for (int i = 0; i < nb; i++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x ); + __m256 v1 = _mm256_loadu_ps( x + 8 ); + __m256 v2 = _mm256_loadu_ps( x + 16 ); + __m256 v3 = _mm256_loadu_ps( x + 24 ); + x += 32; + + // Compute max for the block + __m256 vmax; + vmax = _mm256_max_ps( v0, v1 ); + vmax = _mm256_max_ps( vmax, v2 ); + vmax = _mm256_max_ps( vmax, v3 ); + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( vmax, 1 ), _mm256_castps256_ps128( vmax ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + // Compute min for the block + __m256 vmin; + vmin = _mm256_min_ps( v0, v1 ); + vmin = _mm256_min_ps( vmin, v2 ); + vmin = _mm256_min_ps( vmin, v3 ); + + __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( vmin, 1 ), _mm256_castps256_ps128( vmin ) ); + min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); + min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); + const float minScalar = _mm_cvtss_f32( min4 ); + + // Quantize these floats + const float d = (maxScalar - minScalar) / ((1 << 4) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].m = minScalar; + y[i].d = d; + + // x = (x-min)*id + const __m256 mul = _mm256_set1_ps( id ); + const __m256 off = _mm256_set1_ps( minScalar ); + v0 = _mm256_mul_ps( _mm256_sub_ps( v0, off ), mul ); + v1 = _mm256_mul_ps( _mm256_sub_ps( v1, off ), mul ); + v2 = _mm256_mul_ps( _mm256_sub_ps( v2, off ), mul ); + v3 = _mm256_mul_ps( _mm256_sub_ps( v3, off ), mul ); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); + + // Convert int32 to int16 + i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 + i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 + // Convert int16 to int8 + i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 + + // We got our precious signed bytes, but the order is now wrong + // These AVX2 pack instructions process 16-byte pieces independently + // The following instruction is fixing the order + const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); + i0 = _mm256_permutevar8x32_epi32( i0, perm ); + + // Compress the vector into 4 bit/value, and store + __m128i res = packNibbles( i0 ); + _mm_storeu_si128( ( __m128i* )y[i].qs, res ); + } +#elif __ARM_NEON + for (int i = 0; i < nb; i++) { + float32x4_t srcv[8]; + float32x4_t minv[8]; + float32x4_t maxv[8]; + + for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*V1_QK4_1 + 4*l); + + for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l + 1]); + for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l + 2]); + for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l + 4]); + + for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l + 1]); + for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l + 2]); + for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l + 4]); + + const float min = vminvq_f32(minv[0]); + const float max = vmaxvq_f32(maxv[0]); + + const float d = (max - min) / ((1 << 4) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + y[i].m = min; + + const float32x4_t minv0 = vdupq_n_f32(min); + + for (int l = 0; l < 8; l++) { + const float32x4_t v = vmulq_n_f32(vsubq_f32(srcv[l], minv0), id); + const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(0.5f)); // needed to round to nearest + const int32x4_t vi = vcvtq_s32_f32(vf); + + y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4); + y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4); + } + } +#else + // scalar + quantize_row_v1_q4_1_reference(x, vy, k); +#endif +} + +size_t ggml_quantize_v1_q4_1(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % V1_QK4_1 == 0); + const int nb = k / V1_QK4_1; + + for (int j = 0; j < n; j += k) { + block_v1_q4_1 * restrict y = (block_v1_q4_1 *)dst + j/V1_QK4_1; + + quantize_row_v1_q4_1_reference(src + j, y, k); + + for (int i = 0; i < nb; i++) { + for (int l = 0; l < V1_QK4_1; l += 2) { + const uint8_t vi0 = y[i].qs[l/2] & 0x0F; + const uint8_t vi1 = y[i].qs[l/2] >> 4; + + hist[vi0]++; + hist[vi1]++; + } + } + } + + return (n/V1_QK4_1*sizeof(block_v1_q4_1)); +} + +void dequantize_row_v1_q4_1(const void * restrict vx, float * restrict y, int k) { + assert(k % V1_QK4_1 == 0); + const int nb = k / V1_QK4_1; + + const block_v1_q4_1 * restrict x = vx; + +#if defined(__AVX2__) + for (int i = 0; i < nb; i++) { + const __m256 d_v = _mm256_broadcast_ss(&x[i].d); + const __m256 d_m = _mm256_broadcast_ss(&x[i].m); + + const uint8_t * restrict pp = x[i].qs; + + for (int l = 0; l < V1_QK4_1; l += 32) { + // Load 32x4-bit integers into 32x8-bit integers + __m256i vx8 = bytes_from_nibbles_32(pp+l/2); + + // Convert to 16-bit int + const __m256i vx16_lo = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 0)); + const __m256i vx16_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 1)); + + // Convert to 32-bit int -> float 32 + const __m256 vf[4] = { + _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 0))), + _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 1))), + _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 0))), + _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 1))) + }; + + // Scale, add m and store + for (int j = 0; j < 4; j++) { + const __m256 result = _mm256_add_ps(_mm256_mul_ps(vf[j], d_v), d_m); + _mm256_storeu_ps(y + i * V1_QK4_1 + l + j*8, result); + } + } + } +#elif defined(__ARM_NEON) + for (int i = 0; i < nb; i++) { + const float32x4_t vd = vdupq_n_f32(x[i].d); + const float32x4_t vm = vdupq_n_f32(x[i].m); + + const uint8_t * restrict pp = x[i].qs; + + for (int l = 0; l < V1_QK4_1; l += 16) { + // Load 16x4-bit integers into 8x8-bit integers + const uint8x8_t v8 = vld1_u8(pp + l/2); + + // Expand 4-bit qs to 8-bit bytes + const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0F)); + const uint8x8_t v1 = vshr_n_u8(v8, 4); + + // Interleave and combine + const uint8x8_t vx_0 = vzip1_u8(v0, v1); + const uint8x8_t vx_1 = vzip2_u8(v0, v1); + + const uint8x16_t vq = vcombine_u8(vx_0, vx_1); + + // convert to 2x uint16x8_t + const uint16x8_t vi_0 = vmovl_u8(vget_low_u8 (vq)); + const uint16x8_t vi_1 = vmovl_u8(vget_high_u8(vq)); + + // convert to 4x float32x4_t + const float32x4_t vf_0 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_0))); + const float32x4_t vf_1 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_0))); + const float32x4_t vf_2 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_1))); + const float32x4_t vf_3 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_1))); + + // multiply by d and add m + const float32x4_t r0 = vmlaq_f32(vm, vf_0, vd); + const float32x4_t r1 = vmlaq_f32(vm, vf_1, vd); + const float32x4_t r2 = vmlaq_f32(vm, vf_2, vd); + const float32x4_t r3 = vmlaq_f32(vm, vf_3, vd); + + // Store + vst1q_f32(y + i*V1_QK4_1 + l + 0, r0); + vst1q_f32(y + i*V1_QK4_1 + l + 4, r1); + vst1q_f32(y + i*V1_QK4_1 + l + 8, r2); + vst1q_f32(y + i*V1_QK4_1 + l + 12, r3); + } + } +#else + for (int i = 0; i < nb; i++) { + const float d = x[i].d; + const float m = x[i].m; + + const uint8_t * restrict pp = x[i].qs; + + for (int l = 0; l < V1_QK4_1; l += 2) { + const uint8_t vi = pp[l/2]; + + const int8_t vi0 = vi & 0x0F; + const int8_t vi1 = vi >> 4; + + const float v0 = vi0*d + m; + const float v1 = vi1*d + m; + + y[i*V1_QK4_1 + l + 0] = v0; + y[i*V1_QK4_1 + l + 1] = v1; + + assert(!isnan(y[i*V1_QK4_1 + l + 0])); + assert(!isnan(y[i*V1_QK4_1 + l + 1])); + } + } +#endif +} + +void ggml_vec_dot_v1_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / V1_QK8_1; + + assert(n % V1_QK8_1 == 0); + assert(nb % 2 == 0); + + const block_v1_q4_1 * restrict x = vx; + const block_v1_q8_1 * restrict y = vy; + + // TODO: add AVX / WASM SIMD / etc +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + float summs = 0; + + for (int i = 0; i < nb; i += 2) { + const block_v1_q4_1 * restrict x0 = &x[i + 0]; + const block_v1_q4_1 * restrict x1 = &x[i + 1]; + const block_v1_q8_1 * restrict y0 = &y[i + 0]; + const block_v1_q8_1 * restrict y1 = &y[i + 1]; + + summs += x0->m * (y0->s0 + y0->s1) + x1->m * (y1->s0 + y1->s1); + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // interleave + const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h); + const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h); + const int8x16_t v0_1lz = vzip1q_s8(v0_1l, v0_1h); + const int8x16_t v0_1hz = vzip2q_s8(v0_1l, v0_1h); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + // dot product into int32x4_t + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs; +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + float summs = 0; + + // Main loop + for (int i = 0; i < nb; ++i) { + const float * d0 = &x[i].d; + const float * d1 = &y[i].d; + + summs += x[i].m * (y[i].s0 + y[i].s1); + + const __m256 d0v = _mm256_broadcast_ss( d0 ); + const __m256 d1v = _mm256_broadcast_ss( d1 ); + + // Compute combined scales + const __m256 d0d1 = _mm256_mul_ps( d0v, d1v ); + + // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes + const __m256i bx = bytes_from_nibbles_32(x[i].qs); + const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs ); + + const __m256 xy = mul_sum_i8_pairs_float(bx, by); + + // Accumulate d0*d1*x*y + acc = _mm256_fmadd_ps( d0d1, xy, acc ); + } + + *s = hsum_float_8(acc) + summs; +#else + // scalar + float sumf = 0.0; + for (int i = 0; i < nb; i++) { + const float d0 = x[i].d; + const float m0 = x[i].m; + const float d1 = y[i].d; + + const uint8_t * restrict p0 = x[i].qs; + const int8_t * restrict p1 = y[i].qs; + + // TODO: this is very slow .. + for (int j = 0; j < V1_QK8_1/2; j++) { + const uint8_t v0 = p0[j]; + + const float f0 = d0*(v0 & 0x0F) + m0; + const float f1 = d0*(v0 >> 4) + m0; + + const float f2 = d1*p1[2*j + 0]; + const float f3 = d1*p1[2*j + 1]; + + sumf += f0*f2 + f1*f3; + } + } + *s = sumf; +#endif +} diff --git a/third_party/ggml/ggjt.v1.q4_1.h b/third_party/ggml/ggjt.v1.q4_1.h new file mode 100644 index 000000000..ef74d95d2 --- /dev/null +++ b/third_party/ggml/ggjt.v1.q4_1.h @@ -0,0 +1,22 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_V1_Q4_1_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_V1_Q4_1_H_ +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ + +#define V1_QK4_1 32 +typedef struct { + float d; // delta + float m; // min + uint8_t qs[V1_QK4_1 / 2]; // nibbles / quants +} block_v1_q4_1; + +void dequantize_row_v1_q4_1(const void* restrict, float* restrict, int); +size_t ggml_quantize_v1_q4_1(const float*, void*, int, int, int64_t*); +void quantize_row_v1_q4_1(const float* restrict, void* restrict, int); +void ggml_vec_dot_v1_q4_1_q8_1(const int, float* restrict, const void* restrict, + const void* restrict); +void quantize_row_v1_q4_1_reference(const float* restrict, void* restrict, int); + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_V1_Q4_1_H_ */ diff --git a/third_party/ggml/ggjt.v1.q4_2.c b/third_party/ggml/ggjt.v1.q4_2.c new file mode 100644 index 000000000..abe164e78 --- /dev/null +++ b/third_party/ggml/ggjt.v1.q4_2.c @@ -0,0 +1,297 @@ +/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-│ +│vi: set net ft=c ts=4 sts=4 sw=4 fenc=utf-8 :vi│ +╚──────────────────────────────────────────────────────────────────────────────╝ +│ │ +│ GGML │ +│ Copyright (c) 2023 Georgi Gerganov │ +│ │ +│ Permission is hereby granted, free of charge, to any person obtaining │ +│ a copy of this software and associated documentation files (the │ +│ "Software"), to deal in the Software without restriction, including │ +│ without limitation the rights to use, copy, modify, merge, publish, │ +│ distribute, sublicense, and/or sell copies of the Software, and to │ +│ permit persons to whom the Software is furnished to do so, subject to │ +│ the following conditions: │ +│ │ +│ The above copyright notice and this permission notice shall be │ +│ included in all copies or substantial portions of the Software. │ +│ │ +│ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, │ +│ EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF │ +│ MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. │ +│ IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY │ +│ CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, │ +│ TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE │ +│ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. │ +│ │ +╚─────────────────────────────────────────────────────────────────────────────*/ +#include "third_party/ggml/ggjt.v1.q4_2.h" +#include "libc/assert.h" +#include "libc/macros.internal.h" +#include "libc/str/str.h" +#include "third_party/aarch64/arm_neon.h" +#include "third_party/ggml/fp16.internal.h" +#include "third_party/ggml/ggjt.v1.internal.h" +#include "third_party/ggml/ggjt.v1.q8_0.h" +#include "third_party/intel/immintrin.internal.h" +#include "third_party/libcxx/math.h" +// clang-format off + +static_assert(sizeof(block_v1_q4_2) == sizeof(ggml_fp16_t) + V1_QK4_2 / 2, + "wrong q4_2 block size/padding"); + +// reference implementation for deterministic creation of model files +void quantize_row_v1_q4_2_reference(const float * restrict x, block_v1_q4_2 * restrict y, int k) { + assert(k % V1_QK4_2 == 0); + + const int nb = k / V1_QK4_2; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + float max = 0.0f; + + for (int l = 0; l < V1_QK4_2; l++) { + const float v = x[i*V1_QK4_2 + l]; + if (amax < fabsf(v)) { + amax = fabsf(v); + max = v; + } + } + + const float d = max / -8; + + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + for (int l = 0; l < V1_QK4_2; l += 2) { + const float v0 = x[i*V1_QK4_2 + l + 0]*id; + const float v1 = x[i*V1_QK4_2 + l + 1]*id; + + const uint8_t vi0 = MIN(15, (uint8_t)(v0 + 8.5f)); + const uint8_t vi1 = MIN(15, (uint8_t)(v1 + 8.5f)); + + assert(vi0 < 16); + assert(vi1 < 16); + + y[i].qs[l/2] = vi0 | (vi1 << 4); + } + } +} + +void quantize_row_v1_q4_2(const float * restrict x, void * restrict vy, int k) { + assert(k % V1_QK4_2 == 0); + + block_v1_q4_2 * restrict y = vy; + + quantize_row_v1_q4_2_reference(x, y, k); +} + +size_t ggml_quantize_v1_q4_2(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % V1_QK4_2 == 0); + const int nb = k / V1_QK4_2; + + for (int j = 0; j < n; j += k) { + block_v1_q4_2 * restrict y = (block_v1_q4_2 *)dst + j/V1_QK4_2; + + quantize_row_v1_q4_2_reference(src + j, y, k); + + for (int i = 0; i < nb; i++) { + for (int l = 0; l < V1_QK4_2; l += 2) { + const uint8_t vi0 = y[i].qs[l/2] & 0x0F; + const uint8_t vi1 = y[i].qs[l/2] >> 4; + + hist[vi0]++; + hist[vi1]++; + } + } + } + + return (n/V1_QK4_2*sizeof(block_v1_q4_2)); +} + +void dequantize_row_v1_q4_2(const void * restrict vx, float * restrict y, int k) { + assert(k % V1_QK4_2 == 0); + const int nb = k / V1_QK4_2; + + const block_v1_q4_2 * restrict x = vx; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + + const uint8_t * restrict pp = x[i].qs; + + for (int l = 0; l < V1_QK4_2; l += 2) { + const uint8_t vi = pp[l/2]; + + const int8_t vi0 = vi & 0x0F; + const int8_t vi1 = vi >> 4; + + const float v0 = (vi0 - 8)*d; + const float v1 = (vi1 - 8)*d; + + y[i*V1_QK4_2 + l + 0] = v0; + y[i*V1_QK4_2 + l + 1] = v1; + + assert(!isnan(y[i*V1_QK4_2 + l + 0])); + assert(!isnan(y[i*V1_QK4_2 + l + 1])); + } + } +} + +void ggml_vec_dot_v1_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / V1_QK8_0; + + assert(n % V1_QK8_0 == 0); + assert(nb % 2 == 0); + assert(V1_QK8_0 == 2*V1_QK4_2); + + const block_v1_q4_2 * restrict x = vx; + const block_v1_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i += 2) { + const block_v1_q4_2 * restrict x0_0 = &x[2*(i + 0) + 0]; + const block_v1_q4_2 * restrict x0_1 = &x[2*(i + 0) + 1]; + const block_v1_q4_2 * restrict x1_0 = &x[2*(i + 1) + 0]; + const block_v1_q4_2 * restrict x1_1 = &x[2*(i + 1) + 1]; + + const block_v1_q8_0 * restrict y0 = &y[i + 0]; + const block_v1_q8_0 * restrict y1 = &y[i + 1]; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + const int8x16_t s8b = vdupq_n_s8(0x8); + + const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs)); + const uint8x16_t v0_1 = vcombine_u8(vld1_u8(x1_0->qs), vld1_u8(x1_1->qs)); + + // 4-bit -> 8-bit + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + + // sub 8 + const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); + const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); + const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); + const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); + + // interleave + const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs); + const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs); + const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs); + const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs); + + // load y + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( + vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), GGML_FP16_TO_FP32(x0_0->d)), + vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), GGML_FP16_TO_FP32(x0_1->d))), y0->d); + + sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( + vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l)), GGML_FP16_TO_FP32(x1_0->d)), + vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1hz, v1_1h)), GGML_FP16_TO_FP32(x1_1->d))), y1->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); + + sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( + vmulq_n_f32(vcvtq_f32_s32(pl0), GGML_FP16_TO_FP32(x0_0->d)), + vmulq_n_f32(vcvtq_f32_s32(ph0), GGML_FP16_TO_FP32(x0_1->d))), y0->d); + + sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( + vmulq_n_f32(vcvtq_f32_s32(pl1), GGML_FP16_TO_FP32(x1_0->d)), + vmulq_n_f32(vcvtq_f32_s32(ph1), GGML_FP16_TO_FP32(x1_1->d))), y1->d); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; i++) { + /* Compute combined scale for the block */ + const __m128 d0 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 0].d)); + const __m128 d1 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 1].d)); + const __m256 d = _mm256_mul_ps(_mm256_set_m128(d1, d0), _mm256_broadcast_ss(&y[i].d)); + + __m128i bx0 = bytes_from_nibbles_16(x[2*i + 0].qs); + __m128i bx1 = bytes_from_nibbles_16(x[2*i + 1].qs); + __m256i bx = _mm256_set_m128i(bx1, bx0); + + // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. + const __m256i off = _mm256_set1_epi8(8); + bx = _mm256_sub_epi8(bx, off); + + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + /* Multiply q with scale and accumulate */ + acc = _mm256_fmadd_ps(d, q, acc); + } + + *s = hsum_float_8(acc); +#else + // scalar + float sumf = 0.0; + for (int i = 0; i < nb; i++) { + const uint8_t * restrict x0 = x[2*i + 0].qs; + const uint8_t * restrict x1 = x[2*i + 1].qs; + const int8_t * restrict y0 = y[i].qs; + + const float d0 = GGML_FP16_TO_FP32(x[2*i + 0].d); + const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d); + + int sumi_0 = 0; + int sumi_1 = 0; + + for (int j = 0; j < V1_QK8_0/4; j++) { + const uint8_t v0 = x0[j]; + const uint8_t v1 = x1[j]; + + const int i0_0 = (int8_t) (v0 & 0x0F) - 8; + const int i1_0 = (int8_t) (v0 >> 4) - 8; + + const int i0_1 = (int8_t) (v1 & 0x0F) - 8; + const int i1_1 = (int8_t) (v1 >> 4) - 8; + + const int i2_0 = y0[2*j + 0]; + const int i3_0 = y0[2*j + 1]; + + const int i2_1 = y0[2*(j + V1_QK8_0/4) + 0]; + const int i3_1 = y0[2*(j + V1_QK8_0/4) + 1]; + + sumi_0 += i0_0*i2_0 + i1_0*i3_0; + sumi_1 += i0_1*i2_1 + i1_1*i3_1; + } + + sumf += (d0 * y[i].d) * sumi_0; + sumf += (d1 * y[i].d) * sumi_1; + } + *s = sumf; +#endif +} diff --git a/third_party/ggml/ggjt.v1.q4_2.h b/third_party/ggml/ggjt.v1.q4_2.h new file mode 100644 index 000000000..3f5d0c5b8 --- /dev/null +++ b/third_party/ggml/ggjt.v1.q4_2.h @@ -0,0 +1,23 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q4_2_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q4_2_H_ +#include "third_party/ggml/fp16.h" +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ + +#define V1_QK4_2 16 +typedef struct { + ggml_fp16_t d; // delta + uint8_t qs[V1_QK4_2 / 2]; // nibbles / quants +} block_v1_q4_2; + +void dequantize_row_v1_q4_2(const void* restrict, float* restrict, int); +void quantize_row_v1_q4_2(const float* restrict, void* restrict, int); +size_t ggml_quantize_v1_q4_2(const float*, void*, int, int, int64_t*); +void ggml_vec_dot_v1_q4_2_q8_0(const int, float* restrict, const void* restrict, + const void* restrict); +void quantize_row_v1_q4_2_reference(const float* restrict, + block_v1_q4_2* restrict, int); + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q4_2_H_ */ diff --git a/third_party/ggml/ggjt.v1.q5_0.c b/third_party/ggml/ggjt.v1.q5_0.c new file mode 100644 index 000000000..cc959e6cf --- /dev/null +++ b/third_party/ggml/ggjt.v1.q5_0.c @@ -0,0 +1,350 @@ +/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-│ +│vi: set net ft=c ts=4 sts=4 sw=4 fenc=utf-8 :vi│ +╚──────────────────────────────────────────────────────────────────────────────╝ +│ │ +│ GGML │ +│ Copyright (c) 2023 Georgi Gerganov │ +│ │ +│ Permission is hereby granted, free of charge, to any person obtaining │ +│ a copy of this software and associated documentation files (the │ +│ "Software"), to deal in the Software without restriction, including │ +│ without limitation the rights to use, copy, modify, merge, publish, │ +│ distribute, sublicense, and/or sell copies of the Software, and to │ +│ permit persons to whom the Software is furnished to do so, subject to │ +│ the following conditions: │ +│ │ +│ The above copyright notice and this permission notice shall be │ +│ included in all copies or substantial portions of the Software. │ +│ │ +│ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, │ +│ EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF │ +│ MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. │ +│ IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY │ +│ CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, │ +│ TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE │ +│ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. │ +│ │ +╚─────────────────────────────────────────────────────────────────────────────*/ +#include "third_party/ggml/ggjt.v1.q5_0.h" +#include "libc/assert.h" +#include "libc/macros.internal.h" +#include "third_party/aarch64/arm_neon.h" +#include "third_party/ggml/fp16.internal.h" +#include "third_party/ggml/ggjt.v1.internal.h" +#include "third_party/ggml/ggjt.v1.q8_0.h" +// clang-format off + +static_assert(sizeof(block_v1_q5_0) == + sizeof(ggml_fp16_t) + sizeof(uint32_t) + V1_QK5_0 / 2, + "wrong q5_0 block size/padding"); + +void quantize_row_v1_q5_0_reference(const float * restrict x, block_v1_q5_0 * restrict y, int k) { + assert(k % V1_QK5_0 == 0); + const int nb = k / V1_QK5_0; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + float max = 0.0f; + + for (int l = 0; l < V1_QK5_0; l++) { + const float v = x[i*V1_QK5_0 + l]; + if (amax < fabsf(v)) { + amax = fabsf(v); + max = v; + } + } + + const float d = max / -16; + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + + uint32_t qh = 0; + + for (int l = 0; l < V1_QK5_0; l += 2) { + const float v0 = x[i*V1_QK5_0 + l + 0]*id; + const float v1 = x[i*V1_QK5_0 + l + 1]*id; + + const uint32_t vi0 = MIN(31, (int) (v0 + 16.5f)); + const uint32_t vi1 = MIN(31, (int) (v1 + 16.5f)); + + y[i].qs[l/2] = (vi0 & 0x0F) | ((vi1 & 0x0F) << 4); + + // get the 5-th bit and store it in qh at the right position + qh |= ((vi0 & 0x10) >> 4) << (l + 0); + qh |= ((vi1 & 0x10) >> 4) << (l + 1); + } + + memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); + } +} + +void quantize_row_v1_q5_0(const float * restrict x, void * restrict vy, int k) { + assert(k % V1_QK5_0 == 0); + + block_v1_q5_0 * restrict y = vy; + + quantize_row_v1_q5_0_reference(x, y, k); +} + +size_t ggml_quantize_v1_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % V1_QK5_0 == 0); + const int nb = k / V1_QK5_0; + + for (int j = 0; j < n; j += k) { + block_v1_q5_0 * restrict y = (block_v1_q5_0 *)dst + j/V1_QK5_0; + + quantize_row_v1_q5_0_reference(src + j, y, k); + + for (int i = 0; i < nb; i++) { + uint32_t qh; + memcpy(&qh, &y[i].qh, sizeof(qh)); + + for (int l = 0; l < V1_QK5_0; l += 2) { + const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; + const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + + // cast to 16 bins + const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2; + const uint8_t vi1 = ((y[i].qs[l/2] >> 4) | vh1) / 2; + + hist[vi0]++; + hist[vi1]++; + } + } + } + + return (n/V1_QK5_0*sizeof(block_v1_q5_0)); +} + +void dequantize_row_v1_q5_0(const void * restrict vx, float * restrict y, int k) { + assert(k % V1_QK5_0 == 0); + const int nb = k / V1_QK5_0; + + const block_v1_q5_0 * restrict x = vx; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + + const uint8_t * restrict pp = x[i].qs; + + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + for (int l = 0; l < V1_QK5_0; l += 2) { + const uint8_t vi = pp[l/2]; + + // extract the 5-th bit from qh + const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; + const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + + const int8_t vi0 = (vi & 0x0F) | vh0; + const int8_t vi1 = (vi >> 4) | vh1; + + const float v0 = (vi0 - 16)*d; + const float v1 = (vi1 - 16)*d; + + y[i*V1_QK5_0 + l + 0] = v0; + y[i*V1_QK5_0 + l + 1] = v1; + + assert(!isnan(y[i*V1_QK5_0 + l + 0])); + assert(!isnan(y[i*V1_QK5_0 + l + 1])); + } + } +} + +void ggml_vec_dot_v1_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / V1_QK8_0; + + assert(n % V1_QK8_0 == 0); + assert(nb % 2 == 0); + assert(V1_QK8_0 == V1_QK5_0); + + const block_v1_q5_0 * restrict x = vx; + const block_v1_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv = vdupq_n_f32(0.0f); + + uint64_t tmp[4]; + + for (int i = 0; i < nb; ++i) { + const block_v1_q5_0 * restrict x0 = &x[i]; + const block_v1_q8_0 * restrict y0 = &y[i]; + + const uint8x16_t m4b = vdupq_n_u8(0x0F); + const int8x16_t s16b = vdupq_n_s8(0x10); + + // extract the 5th bit + uint32_t qh; + memcpy(&qh, x0->qh, sizeof(qh)); + + tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_u[(qh >> 24) ]; + + const int8x16_t qhl = vld1q_s8((const int8_t *)(tmp + 0)); + const int8x16_t qhh = vld1q_s8((const int8_t *)(tmp + 2)); + + const uint8x16_t v0 = vld1q_u8(x0->qs); + + // 4-bit -> 8-bit + const int8x16_t v0l = vreinterpretq_s8_u8(vandq_u8 (v0, m4b)); + const int8x16_t v0h = vreinterpretq_s8_u8(vshrq_n_u8(v0, 4)); + + // interleave + const int8x16_t v0lz = vzip1q_s8(v0l, v0h); + const int8x16_t v0hz = vzip2q_s8(v0l, v0h); + + // add high bit and sub 16 + const int8x16_t v0lf = vsubq_s8(vorrq_s8(v0lz, qhl), s16b); + const int8x16_t v0hf = vsubq_s8(vorrq_s8(v0hz, qhh), s16b); + + // load y + const int8x16_t v1l = vld1q_s8(y0->qs); + const int8x16_t v1h = vld1q_s8(y0->qs + 16); + + const float x0d = GGML_FP16_TO_FP32(x0->d); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0lf, v1l), + vdotq_s32(vdupq_n_s32(0), v0hf, v1h))), x0d*y0->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0lf), vget_low_s8 (v1l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0lf), vget_high_s8(v1l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0hf), vget_low_s8 (v1h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0hf), vget_high_s8(v1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + + sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); +#endif + } + + *s = vaddvq_f32(sumv); +#elif defined(__wasm_simd128__) + v128_t sumv = wasm_f32x4_splat(0.0f); + + uint64_t tmp[4]; + + for (int i = 0; i < nb; ++i) { + const block_v1_q5_0 * restrict x0 = &x[i]; + const block_v1_q8_0 * restrict y0 = &y[i]; + + const v128_t m4b = wasm_i8x16_splat(0x0F); + const v128_t s16b = wasm_i8x16_splat(0x10); + + // extract the 5th bit + uint32_t qh; + memcpy(&qh, x0->qh, sizeof(qh)); + + tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_u[(qh >> 24) ]; + + const v128_t qhl = wasm_v128_load(tmp + 0); + const v128_t qhh = wasm_v128_load(tmp + 2); + + const v128_t v0 = wasm_v128_load(x0->qs); + + // 4-bit -> 8-bit + const v128_t v0l = wasm_v128_and (v0, m4b); + const v128_t v0h = wasm_u8x16_shr(v0, 4); + + // interleave + const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23); + const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31); + + // add high bit and sub 16 + const v128_t v0lf = wasm_i8x16_sub(wasm_v128_or(v0lz, qhl), s16b); + const v128_t v0hf = wasm_i8x16_sub(wasm_v128_or(v0hz, qhh), s16b); + + // load y + const v128_t v1l = wasm_v128_load(y0->qs); + const v128_t v1h = wasm_v128_load(y0->qs + 16); + + // int8x16 -> int16x8 + const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf); + const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf); + const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf); + const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf); + + const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l); + const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l); + const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h); + const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h); + + const float x0d = GGML_FP16_TO_FP32(x0->d); + + // dot product + sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4( + wasm_i32x4_add( + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), + wasm_i32x4_dot_i16x8(v0lfh, v1lh)), + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), + wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d))); + } + + *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3); +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; i++) { + /* Compute combined scale for the block */ + const __m256 d = _mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)), _mm256_broadcast_ss(&y[i].d)); + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i bxhi = bytes_from_bits_32(x[i].qh); + bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0)); + bx = _mm256_or_si256(bx, bxhi); + + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + /* Multiply q with scale and accumulate */ + acc = _mm256_fmadd_ps(d, q, acc); + } + + *s = hsum_float_8(acc); +#else + // scalar + float sumf = 0.0; + for (int i = 0; i < nb; i++) { + const uint8_t * restrict x0 = x[i].qs; + const int8_t * restrict y0 = y[i].qs; + + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + const float d = GGML_FP16_TO_FP32(x[i].d); + + int sxy = 0; + + for (int j = 0; j < V1_QK8_0/2; j++) { + const uint8_t v0 = x0[j]; + + const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4; + const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4; + + const int x0_0 = ((v0 & 0x0F) | x0_0h) - 16; + const int x1_0 = ((v0 >> 4) | x1_0h) - 16; + + const int y0_0 = y0[2*j + 0]; + const int y1_0 = y0[2*j + 1]; + + sxy += x0_0*y0_0 + x1_0*y1_0; + } + + sumf += (d*sxy)*y[i].d; + } + *s = sumf; +#endif +} diff --git a/third_party/ggml/ggjt.v1.q5_0.h b/third_party/ggml/ggjt.v1.q5_0.h new file mode 100644 index 000000000..13c06e31d --- /dev/null +++ b/third_party/ggml/ggjt.v1.q5_0.h @@ -0,0 +1,24 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q5_0_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q5_0_H_ +#include "third_party/ggml/fp16.h" +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ + +#define V1_QK5_0 32 +typedef struct { + ggml_fp16_t d; // delta + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[V1_QK5_0 / 2]; // nibbles / quants +} block_v1_q5_0; + +void dequantize_row_v1_q5_0(const void* restrict, float* restrict, int); +void quantize_row_v1_q5_0(const float* restrict, void* restrict, int); +size_t ggml_quantize_v1_q5_0(const float*, void*, int, int, int64_t*); +void ggml_vec_dot_v1_q5_0_q8_0(const int, float* restrict, const void* restrict, + const void* restrict); +void quantize_row_v1_q5_0_reference(const float* restrict, + block_v1_q5_0* restrict, int); + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q5_0_H_ */ diff --git a/third_party/ggml/ggjt.v1.q5_1.c b/third_party/ggml/ggjt.v1.q5_1.c new file mode 100644 index 000000000..fb7a34cd7 --- /dev/null +++ b/third_party/ggml/ggjt.v1.q5_1.c @@ -0,0 +1,352 @@ +/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:2;tab-width:8;coding:utf-8 -*-│ +│vi: set net ft=c ts=2 sts=2 sw=2 fenc=utf-8 :vi│ +╞══════════════════════════════════════════════════════════════════════════════╡ +│ Copyright 2023 Justine Alexandra Roberts Tunney │ +│ │ +│ Permission to use, copy, modify, and/or distribute this software for │ +│ any purpose with or without fee is hereby granted, provided that the │ +│ above copyright notice and this permission notice appear in all copies. │ +│ │ +│ THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL │ +│ WARRANTIES WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED │ +│ WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE │ +│ AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL │ +│ DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR │ +│ PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER │ +│ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR │ +│ PERFORMANCE OF THIS SOFTWARE. │ +╚─────────────────────────────────────────────────────────────────────────────*/ +#include "third_party/ggml/ggjt.v1.q5_1.h" +#include "libc/assert.h" +#include "libc/math.h" +#include "libc/str/str.h" +#include "third_party/aarch64/arm_neon.h" +#include "third_party/ggml/fp16.internal.h" +#include "third_party/ggml/ggjt.v1.internal.h" +#include "third_party/ggml/ggjt.v1.q8_1.h" +// clang-format off + +static_assert(sizeof(block_v1_q5_1) == + 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + V1_QK5_1 / 2, + "wrong q5_1 block size/padding"); + +void quantize_row_v1_q5_1_reference(const float * restrict x, block_v1_q5_1 * restrict y, int k) { + assert(k % V1_QK5_1 == 0); + const int nb = k / V1_QK5_1; + + for (int i = 0; i < nb; i++) { + float min = FLT_MAX; + float max = -FLT_MAX; + + for (int l = 0; l < V1_QK5_1; l++) { + const float v = x[i*V1_QK5_1 + l]; + if (v < min) min = v; + if (v > max) max = v; + } + + const float d = (max - min) / ((1 << 5) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = GGML_FP32_TO_FP16(d); + y[i].m = GGML_FP32_TO_FP16(min); + + uint32_t qh = 0; + + for (int l = 0; l < V1_QK5_1; l += 2) { + const float v0 = (x[i*V1_QK5_1 + l + 0] - min)*id; + const float v1 = (x[i*V1_QK5_1 + l + 1] - min)*id; + + const uint32_t vi0 = (int) (v0 + 0.5f); + const uint32_t vi1 = (int) (v1 + 0.5f); + + y[i].qs[l/2] = (vi0 & 0x0F) | ((vi1 & 0x0F) << 4); + + // get the 5-th bit and store it in qh at the right position + qh |= ((vi0 & 0x10) >> 4) << (l + 0); + qh |= ((vi1 & 0x10) >> 4) << (l + 1); + } + + memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); + } +} + +void quantize_row_v1_q5_1(const float * restrict x, void * restrict vy, int k) { + assert(k % V1_QK5_1 == 0); + + block_v1_q5_1 * restrict y = vy; + + quantize_row_v1_q5_1_reference(x, y, k); +} + +size_t ggml_quantize_v1_q5_1(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % V1_QK5_1 == 0); + const int nb = k / V1_QK5_1; + + for (int j = 0; j < n; j += k) { + block_v1_q5_1 * restrict y = (block_v1_q5_1 *)dst + j/V1_QK5_1; + + quantize_row_v1_q5_1_reference(src + j, y, k); + + for (int i = 0; i < nb; i++) { + uint32_t qh; + memcpy(&qh, &y[i].qh, sizeof(qh)); + + for (int l = 0; l < V1_QK5_1; l += 2) { + const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; + const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + + // cast to 16 bins + const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2; + const uint8_t vi1 = ((y[i].qs[l/2] >> 4) | vh1) / 2; + + hist[vi0]++; + hist[vi1]++; + } + } + } + + return (n/V1_QK5_1*sizeof(block_v1_q5_1)); +} + +void dequantize_row_v1_q5_1(const void * restrict vx, float * restrict y, int k) { + assert(k % V1_QK5_1 == 0); + const int nb = k / V1_QK5_1; + + const block_v1_q5_1 * restrict x = vx; + + for (int i = 0; i < nb; i++) { + const float d = GGML_FP16_TO_FP32(x[i].d); + const float m = GGML_FP16_TO_FP32(x[i].m); + + const uint8_t * restrict pp = x[i].qs; + + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + for (int l = 0; l < V1_QK5_1; l += 2) { + const uint8_t vi = pp[l/2]; + + // extract the 5-th bit from qh + const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; + const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + + const uint8_t vi0 = (vi & 0x0F) | vh0; + const uint8_t vi1 = (vi >> 4) | vh1; + + const float v0 = vi0*d + m; + const float v1 = vi1*d + m; + + y[i*V1_QK5_1 + l + 0] = v0; + y[i*V1_QK5_1 + l + 1] = v1; + + assert(!isnan(y[i*V1_QK5_1 + l + 0])); + assert(!isnan(y[i*V1_QK5_1 + l + 1])); + } + } +} + +void ggml_vec_dot_v1_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / V1_QK8_1; + + assert(n % V1_QK8_1 == 0); + assert(nb % 2 == 0); + assert(V1_QK8_1 == V1_QK5_1); + + const block_v1_q5_1 * restrict x = vx; + const block_v1_q8_1 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv = vdupq_n_f32(0.0f); + + float summs = 0.0f; + + uint64_t tmp[4]; + + for (int i = 0; i < nb; ++i) { + const block_v1_q5_1 * restrict x0 = &x[i]; + const block_v1_q8_1 * restrict y0 = &y[i]; + + summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1); + + // extract the 5th bit + uint32_t qh; + memcpy(&qh, x0->qh, sizeof(qh)); + + tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_u[(qh >> 24) ]; + + const int8x16_t qhl = vld1q_s8((const int8_t *)(tmp + 0)); + const int8x16_t qhh = vld1q_s8((const int8_t *)(tmp + 2)); + + const uint8x16_t v0 = vld1q_u8(x0->qs); + + // 4-bit -> 8-bit + const int8x16_t v0l = vreinterpretq_s8_u8(vandq_u8 (v0, vdupq_n_u8(0x0F))); + const int8x16_t v0h = vreinterpretq_s8_u8(vshrq_n_u8(v0, 4)); + + // interleave + const int8x16_t v0lz = vzip1q_s8(v0l, v0h); + const int8x16_t v0hz = vzip2q_s8(v0l, v0h); + + // add + const int8x16_t v0lf = vorrq_s8(v0lz, qhl); + const int8x16_t v0hf = vorrq_s8(v0hz, qhh); + + // load y + const int8x16_t v1l = vld1q_s8(y0->qs); + const int8x16_t v1h = vld1q_s8(y0->qs + 16); + + const float x0d = GGML_FP16_TO_FP32(x0->d); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0lf, v1l), + vdotq_s32(vdupq_n_s32(0), v0hf, v1h))), x0d*y0->d); +#else + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0lf), vget_low_s8 (v1l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0lf), vget_high_s8(v1l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0hf), vget_low_s8 (v1h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0hf), vget_high_s8(v1h)); + + const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); + const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + + sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); +#endif + } + + *s = vaddvq_f32(sumv) + summs; +#elif defined(__wasm_simd128__) + v128_t sumv = wasm_f32x4_splat(0.0f); + + float summs = 0.0f; + + uint64_t tmp[4]; + + for (int i = 0; i < nb; ++i) { + const block_v1_q5_1 * restrict x0 = &x[i]; + const block_v1_q8_1 * restrict y0 = &y[i]; + + summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1); + + const v128_t m4b = wasm_i8x16_splat(0x0F); + + // extract the 5th bit + uint32_t qh; + memcpy(&qh, x0->qh, sizeof(qh)); + + tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_u[(qh >> 24) ]; + + const v128_t qhl = wasm_v128_load(tmp + 0); + const v128_t qhh = wasm_v128_load(tmp + 2); + + const v128_t v0 = wasm_v128_load(x0->qs); + + // 4-bit -> 8-bit + const v128_t v0l = wasm_v128_and (v0, m4b); + const v128_t v0h = wasm_u8x16_shr(v0, 4); + + static bool x = true; + + // interleave + const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23); + const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31); + + // add high bit + const v128_t v0lf = wasm_v128_or(v0lz, qhl); + const v128_t v0hf = wasm_v128_or(v0hz, qhh); + + // load y + const v128_t v1l = wasm_v128_load(y0->qs); + const v128_t v1h = wasm_v128_load(y0->qs + 16); + + // int8x16 -> int16x8 + const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf); + const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf); + const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf); + const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf); + + const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l); + const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l); + const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h); + const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h); + + const float x0d = GGML_FP16_TO_FP32(x0->d); + + // dot product + sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4( + wasm_i32x4_add( + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll), + wasm_i32x4_dot_i16x8(v0lfh, v1lh)), + wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl), + wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d))); + } + + *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) + + wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs; +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + float summs = 0.0f; + + // Main loop + for (int i = 0; i < nb; i++) { + const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); + + summs += GGML_FP16_TO_FP32(x[i].m) * (y[i].s0 + y[i].s1); + + __m256i bx = bytes_from_nibbles_32(x[i].qs); + __m256i bxhi = bytes_from_bits_32(x[i].qh); + bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10)); + bx = _mm256_or_si256(bx, bxhi); + + const __m256 dy = _mm256_broadcast_ss(&y[i].d); + const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc); + } + + *s = hsum_float_8(acc) + summs; +#else + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + const uint8_t * restrict x0 = x[i].qs; + const int8_t * restrict y0 = y[i].qs; + + uint32_t qh; + memcpy(&qh, x[i].qh, sizeof(qh)); + + const float d = GGML_FP16_TO_FP32(x[i].d); + const float m = GGML_FP16_TO_FP32(x[i].m); + + int sxy = 0; + + for (int j = 0; j < V1_QK8_1/2; j++) { + const uint8_t v0 = x0[j]; + + const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4; + const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4; + + const int x0_0 = (v0 & 0x0F) | x0_0h; + const int x1_0 = (v0 >> 4) | x1_0h; + + const int y0_0 = y0[2*j + 0]; + const int y1_0 = y0[2*j + 1]; + + sxy += x0_0*y0_0 + x1_0*y1_0; + } + + sumf += (d*sxy)*y[i].d + m*(y[i].s0 + y[i].s1); + } + + *s = sumf; +#endif +} diff --git a/third_party/ggml/ggjt.v1.q5_1.h b/third_party/ggml/ggjt.v1.q5_1.h new file mode 100644 index 000000000..c335f060d --- /dev/null +++ b/third_party/ggml/ggjt.v1.q5_1.h @@ -0,0 +1,25 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q5_1_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q5_1_H_ +#include "third_party/ggml/fp16.h" +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ + +#define V1_QK5_1 32 +typedef struct { + ggml_fp16_t d; // delta + ggml_fp16_t m; // min + uint8_t qh[4]; // 5-th bit of quants + uint8_t qs[V1_QK5_1 / 2]; // nibbles / quants +} block_v1_q5_1; + +void dequantize_row_v1_q5_1(const void* restrict, float* restrict, int); +void quantize_row_v1_q5_1(const float* restrict, void* restrict, int); +size_t ggml_quantize_v1_q5_1(const float*, void*, int, int, int64_t*); +void ggml_vec_dot_v1_q5_1_q8_1(const int, float* restrict, const void* restrict, + const void* restrict); +void quantize_row_v1_q5_1_reference(const float* restrict, + block_v1_q5_1* restrict, int); + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q5_1_H_ */ diff --git a/third_party/ggml/ggjt.v1.q8_0.c b/third_party/ggml/ggjt.v1.q8_0.c new file mode 100644 index 000000000..952eaf9de --- /dev/null +++ b/third_party/ggml/ggjt.v1.q8_0.c @@ -0,0 +1,325 @@ +/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:2;tab-width:8;coding:utf-8 -*-│ +│vi: set net ft=c ts=2 sts=2 sw=2 fenc=utf-8 :vi│ +╞══════════════════════════════════════════════════════════════════════════════╡ +│ Copyright 2023 Justine Alexandra Roberts Tunney │ +│ │ +│ Permission to use, copy, modify, and/or distribute this software for │ +│ any purpose with or without fee is hereby granted, provided that the │ +│ above copyright notice and this permission notice appear in all copies. │ +│ │ +│ THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL │ +│ WARRANTIES WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED │ +│ WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE │ +│ AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL │ +│ DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR │ +│ PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER │ +│ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR │ +│ PERFORMANCE OF THIS SOFTWARE. │ +╚─────────────────────────────────────────────────────────────────────────────*/ +#include "third_party/ggml/ggjt.v1.q8_0.h" +#include "libc/assert.h" +#include "libc/macros.internal.h" +#include "third_party/aarch64/arm_neon.h" +#include "third_party/ggml/ggjt.v1.internal.h" +#include "third_party/ggml/ggjt.v1.q8_0.h" +#include "third_party/intel/immintrin.internal.h" +#include "third_party/libcxx/math.h" +// clang-format off + +static_assert(sizeof(block_v1_q8_0) == sizeof(float) + V1_QK8_0, + "wrong q8_0 block size/padding"); + +// reference implementation for deterministic creation of model files +void quantize_row_v1_q8_0_reference(const float * restrict x, block_v1_q8_0 * restrict y, int k) { + assert(k % V1_QK8_0 == 0); + const int nb = k / V1_QK8_0; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + + for (int l = 0; l < V1_QK8_0; l++) { + const float v = x[i*V1_QK8_0 + l]; + amax = MAX(amax, fabsf(v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + for (int l = 0; l < V1_QK8_0; ++l) { + const float v0 = x[i*V1_QK8_0 + l]*id; + + y[i].qs[l] = roundf(v0); + } + } +} + +void quantize_row_v1_q8_0(const float * restrict x, void * restrict vy, int k) { + assert(V1_QK8_0 == 32); + assert(k % V1_QK8_0 == 0); + const int nb = k / V1_QK8_0; + + block_v1_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + for (int i = 0; i < nb; i++) { + float32x4_t srcv [8]; + float32x4_t asrcv[8]; + float32x4_t amaxv[8]; + + for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); + for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]); + + for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]); + for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]); + for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]); + + const float amax = vmaxvq_f32(amaxv[0]); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + for (int l = 0; l < 8; l++) { + const float32x4_t v = vmulq_n_f32(srcv[l], id); + const int32x4_t vi = vcvtnq_s32_f32(v); + + y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); + } + } +#elif defined(__AVX2__) || defined(__AVX__) + for (int i = 0; i < nb; i++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x ); + __m256 v1 = _mm256_loadu_ps( x + 8 ); + __m256 v2 = _mm256_loadu_ps( x + 16 ); + __m256 v3 = _mm256_loadu_ps( x + 24 ); + x += 32; + + // Compute max(abs(e)) for the block + const __m256 signBit = _mm256_set1_ps( -0.0f ); + __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + // Quantize these floats + const float d = maxScalar / 127.f; + y[i].d = d; + const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; + const __m256 mul = _mm256_set1_ps( id ); + + // Apply the multiplier + v0 = _mm256_mul_ps( v0, mul ); + v1 = _mm256_mul_ps( v1, mul ); + v2 = _mm256_mul_ps( v2, mul ); + v3 = _mm256_mul_ps( v3, mul ); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); + +#if defined(__AVX2__) + // Convert int32 to int16 + i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 + i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 + // Convert int16 to int8 + i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 + + // We got our precious signed bytes, but the order is now wrong + // These AVX2 pack instructions process 16-byte pieces independently + // The following instruction is fixing the order + const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); + i0 = _mm256_permutevar8x32_epi32( i0, perm ); + + _mm256_storeu_si256((__m256i *)y[i].qs, i0); +#else + // Since we don't have in AVX some necessary functions, + // we split the registers in half and call AVX2 analogs from SSE + __m128i ni0 = _mm256_castsi256_si128( i0 ); + __m128i ni1 = _mm256_extractf128_si256( i0, 1); + __m128i ni2 = _mm256_castsi256_si128( i1 ); + __m128i ni3 = _mm256_extractf128_si256( i1, 1); + __m128i ni4 = _mm256_castsi256_si128( i2 ); + __m128i ni5 = _mm256_extractf128_si256( i2, 1); + __m128i ni6 = _mm256_castsi256_si128( i3 ); + __m128i ni7 = _mm256_extractf128_si256( i3, 1); + + // Convert int32 to int16 + ni0 = _mm_packs_epi32( ni0, ni1 ); + ni2 = _mm_packs_epi32( ni2, ni3 ); + ni4 = _mm_packs_epi32( ni4, ni5 ); + ni6 = _mm_packs_epi32( ni6, ni7 ); + // Convert int16 to int8 + ni0 = _mm_packs_epi16( ni0, ni2 ); + ni4 = _mm_packs_epi16( ni4, ni6 ); + + _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0); + _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4); +#endif + } +#else + // scalar + quantize_row_v1_q8_0_reference(x, y, k); +#endif +} + +size_t ggml_quantize_v1_q8_0(const float * src, void * dst, int n, int k, int64_t * hist) { + assert(k % V1_QK8_0 == 0); + const int nb = k / V1_QK8_0; + + for (int j = 0; j < n; j += k) { + block_v1_q8_0 * restrict y = (block_v1_q8_0 *)dst + j/V1_QK8_0; + + quantize_row_v1_q8_0_reference(src + j, y, k); + + for (int i = 0; i < nb; i++) { + for (int l = 0; l < V1_QK8_0; ++l) { + const int8_t vi = y[i].qs[l]; + + hist[vi/16 + 8]++; + } + } + } + + return (n/V1_QK8_0*sizeof(block_v1_q8_0)); +} + +void dequantize_row_v1_q8_0(const void * restrict vx, float * restrict y, int k) { + assert(k % V1_QK8_0 == 0); + const int nb = k / V1_QK8_0; + + const block_v1_q8_0 * restrict x = vx; + + for (int i = 0; i < nb; i++) { + const float d = x[i].d; + + const int8_t * restrict pp = x[i].qs; + + for (int l = 0; l < V1_QK8_0; ++l) { + y[i*V1_QK8_0 + l] = pp[l]*d; + } + } +} + +void ggml_vec_dot_v1_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int nb = n / V1_QK8_0; + + assert(n % V1_QK8_0 == 0); + assert(nb % 2 == 0); + assert(V1_QK8_0 == V1_QK8_0); + + const block_v1_q8_0 * restrict x = vx; + const block_v1_q8_0 * restrict y = vy; + +#if defined(__ARM_NEON) + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); + + for (int i = 0; i < nb; i += 2) { + const block_v1_q8_0 * restrict x0 = &x[i + 0]; + const block_v1_q8_0 * restrict x1 = &x[i + 1]; + const block_v1_q8_0 * restrict y0 = &y[i + 0]; + const block_v1_q8_0 * restrict y1 = &y[i + 1]; + + const int8x16_t x0_0 = vld1q_s8(x0->qs); + const int8x16_t x0_1 = vld1q_s8(x0->qs + 16); + const int8x16_t x1_0 = vld1q_s8(x1->qs); + const int8x16_t x1_1 = vld1q_s8(x1->qs + 16); + + // load y + const int8x16_t y0_0 = vld1q_s8(y0->qs); + const int8x16_t y0_1 = vld1q_s8(y0->qs + 16); + const int8x16_t y1_0 = vld1q_s8(y1->qs); + const int8x16_t y1_1 = vld1q_s8(y1->qs + 16); + +#if defined(__ARM_FEATURE_DOTPROD) + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), x0_0, y0_0), + vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), x0->d*y0->d); + + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), x1_0, y1_0), + vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), x1->d*y1->d); + +#else + const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0)); + const int16x8_t p0_1 = vmull_s8(vget_high_s8(x0_0), vget_high_s8(y0_0)); + const int16x8_t p0_2 = vmull_s8(vget_low_s8 (x0_1), vget_low_s8 (y0_1)); + const int16x8_t p0_3 = vmull_s8(vget_high_s8(x0_1), vget_high_s8(y0_1)); + + const int16x8_t p1_0 = vmull_s8(vget_low_s8 (x1_0), vget_low_s8 (y1_0)); + const int16x8_t p1_1 = vmull_s8(vget_high_s8(x1_0), vget_high_s8(y1_0)); + const int16x8_t p1_2 = vmull_s8(vget_low_s8 (x1_1), vget_low_s8 (y1_1)); + const int16x8_t p1_3 = vmull_s8(vget_high_s8(x1_1), vget_high_s8(y1_1)); + + const int32x4_t p0 = vaddq_s32(vpaddlq_s16(p0_0), vpaddlq_s16(p0_1)); + const int32x4_t p1 = vaddq_s32(vpaddlq_s16(p0_2), vpaddlq_s16(p0_3)); + const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1)); + const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3)); + + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), x0->d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), x1->d*y1->d); +#endif + } + + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); +#elif defined(__AVX2__) + // Initialize accumulator with zeros + __m256 acc = _mm256_setzero_ps(); + + // Main loop + for (int i = 0; i < nb; ++i) { + // Compute combined scale for the block + const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); + __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs); + __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); + + const __m256 q = mul_sum_i8_pairs_float(bx, by); + + // Multiply q with scale and accumulate + acc = _mm256_fmadd_ps( d, q, acc ); + } + + *s = hsum_float_8(acc); +#else + // scalar + float sumf = 0.0; + + for (int i = 0; i < nb; i++) { + const int8_t * restrict x0 = x[i].qs; + const int8_t * restrict y0 = y[i].qs; + + int sumi = 0; + + for (int j = 0; j < V1_QK8_0; j++) { + const int v0 = x0[j]; + const int v1 = y0[j]; + + sumi += v0*v1; + } + + sumf += (x[i].d*y[i].d)*sumi; + } + + *s = sumf; +#endif +} diff --git a/third_party/ggml/ggjt.v1.q8_0.h b/third_party/ggml/ggjt.v1.q8_0.h new file mode 100644 index 000000000..54765da77 --- /dev/null +++ b/third_party/ggml/ggjt.v1.q8_0.h @@ -0,0 +1,22 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q8_0_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q8_0_H_ +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ + +#define V1_QK8_0 32 +typedef struct { + float d; // delta + int8_t qs[V1_QK8_0]; // quants +} block_v1_q8_0; + +void dequantize_row_v1_q8_0(const void* restrict, float* restrict, int); +void quantize_row_v1_q8_0(const float* restrict, void* restrict, int); +size_t ggml_quantize_v1_q8_0(const float*, void*, int, int, int64_t*); +void ggml_vec_dot_v1_q8_0_q8_0(const int, float* restrict, const void* restrict, + const void* restrict); +void quantize_row_v1_q8_0_reference(const float* restrict, + block_v1_q8_0* restrict, int); + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q8_0_H_ */ diff --git a/third_party/ggml/ggjt.v1.q8_1.c b/third_party/ggml/ggjt.v1.q8_1.c new file mode 100644 index 000000000..2aabb28c8 --- /dev/null +++ b/third_party/ggml/ggjt.v1.q8_1.c @@ -0,0 +1,240 @@ +/*-*- mode:c;indent-tabs-mode:nil;c-basic-offset:2;tab-width:8;coding:utf-8 -*-│ +│vi: set net ft=c ts=2 sts=2 sw=2 fenc=utf-8 :vi│ +╞══════════════════════════════════════════════════════════════════════════════╡ +│ Copyright 2023 Justine Alexandra Roberts Tunney │ +│ │ +│ Permission to use, copy, modify, and/or distribute this software for │ +│ any purpose with or without fee is hereby granted, provided that the │ +│ above copyright notice and this permission notice appear in all copies. │ +│ │ +│ THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL │ +│ WARRANTIES WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED │ +│ WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE │ +│ AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL │ +│ DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR │ +│ PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER │ +│ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR │ +│ PERFORMANCE OF THIS SOFTWARE. │ +╚─────────────────────────────────────────────────────────────────────────────*/ +#include "third_party/ggml/ggjt.v1.q8_1.h" +#include "libc/assert.h" +#include "libc/macros.internal.h" +#include "third_party/aarch64/arm_neon.h" +#include "third_party/ggml/ggml.h" +#include "third_party/intel/immintrin.internal.h" +#include "third_party/libcxx/math.h" +// clang-format off + +static_assert(sizeof(block_v1_q8_1) == 3 * sizeof(float) + V1_QK8_1, + "wrong q8_1 block size/padding"); + +#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) +// horizontally add 8 int32_t +static inline int hsum_i32_8(const __m256i a) { + const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1)); + const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128); + const __m128i sum64 = _mm_add_epi32(hi64, sum128); + const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1)); + return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32)); +} +#endif /* AVX || AVX2 || AVX512F */ + +// reference implementation for deterministic creation of model files +void quantize_row_v1_q8_1_reference(const float * restrict x, block_v1_q8_1 * restrict y, int k) { + assert(V1_QK8_1 == 32); + assert(k % V1_QK8_1 == 0); + const int nb = k / V1_QK8_1; + + for (int i = 0; i < nb; i++) { + float amax = 0.0f; // absolute max + + for (int l = 0; l < V1_QK8_1; l++) { + const float v = x[i*V1_QK8_1 + l]; + amax = MAX(amax, fabsf(v)); + } + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + int sum0 = 0; + int sum1 = 0; + + for (int l = 0; l < V1_QK8_1/2; ++l) { + const float v0 = x[i*V1_QK8_1 + l]*id; + const float v1 = x[i*V1_QK8_1 + V1_QK8_1/2 + l]*id; + + y[i].qs[ l] = roundf(v0); + y[i].qs[V1_QK8_1/2 + l] = roundf(v1); + + sum0 += y[i].qs[ l]; + sum1 += y[i].qs[V1_QK8_1/2 + l]; + } + + y[i].s0 = d * sum0; + y[i].s1 = d * sum1; + } +} + +void quantize_row_v1_q8_1(const float * restrict x, void * restrict vy, int k) { + assert(k % V1_QK8_1 == 0); + const int nb = k / V1_QK8_1; + + block_v1_q8_1 * restrict y = vy; + +#if defined(__ARM_NEON) + for (int i = 0; i < nb; i++) { + float32x4_t srcv [8]; + float32x4_t asrcv[8]; + float32x4_t amaxv[8]; + + for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); + for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]); + + for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]); + for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]); + for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]); + + const float amax = vmaxvq_f32(amaxv[0]); + + const float d = amax / ((1 << 7) - 1); + const float id = d ? 1.0f/d : 0.0f; + + y[i].d = d; + + int32x4_t accv0 = vdupq_n_s32(0); + int32x4_t accv1 = vdupq_n_s32(0); + + // low half + for (int l = 0; l < 4; l++) { + const float32x4_t v = vmulq_n_f32(srcv[l], id); + const int32x4_t vi = vcvtnq_s32_f32(v); + + y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); + + accv0 = vaddq_s32(accv0, vi); + } + + // high half + for (int l = 4; l < 8; l++) { + const float32x4_t v = vmulq_n_f32(srcv[l], id); + const int32x4_t vi = vcvtnq_s32_f32(v); + + y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); + + accv1 = vaddq_s32(accv1, vi); + } + + const int32_t sum0 = vaddvq_s32(accv0); + const int32_t sum1 = vaddvq_s32(accv1); + + y[i].s0 = d * sum0; + y[i].s1 = d * sum1; + } +#elif defined(__AVX2__) || defined(__AVX__) + for (int i = 0; i < nb; i++) { + // Load elements into 4 AVX vectors + __m256 v0 = _mm256_loadu_ps( x ); + __m256 v1 = _mm256_loadu_ps( x + 8 ); + __m256 v2 = _mm256_loadu_ps( x + 16 ); + __m256 v3 = _mm256_loadu_ps( x + 24 ); + x += 32; + + // Compute max(abs(e)) for the block + const __m256 signBit = _mm256_set1_ps( -0.0f ); + __m256 maxAbs = _mm256_andnot_ps( signBit, v0 ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) ); + maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) ); + + __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) ); + max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); + max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); + const float maxScalar = _mm_cvtss_f32( max4 ); + + // Quantize these floats + const float d = maxScalar / 127.f; + y[i].d = d; + const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; + const __m256 mul = _mm256_set1_ps( id ); + + // Apply the multiplier + v0 = _mm256_mul_ps( v0, mul ); + v1 = _mm256_mul_ps( v1, mul ); + v2 = _mm256_mul_ps( v2, mul ); + v3 = _mm256_mul_ps( v3, mul ); + + // Round to nearest integer + v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); + v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); + v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); + v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); + + // Convert floats to integers + __m256i i0 = _mm256_cvtps_epi32( v0 ); + __m256i i1 = _mm256_cvtps_epi32( v1 ); + __m256i i2 = _mm256_cvtps_epi32( v2 ); + __m256i i3 = _mm256_cvtps_epi32( v3 ); + +#if defined(__AVX2__) + // Compute the sum of the quants and set y[i].s + //y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))); + y[i].s0 = d * hsum_i32_8(_mm256_add_epi32(i0, i1)); + y[i].s1 = d * hsum_i32_8(_mm256_add_epi32(i2, i3)); + + // Convert int32 to int16 + i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 + i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 + // Convert int16 to int8 + i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 + + // We got our precious signed bytes, but the order is now wrong + // These AVX2 pack instructions process 16-byte pieces independently + // The following instruction is fixing the order + const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); + i0 = _mm256_permutevar8x32_epi32( i0, perm ); + + _mm256_storeu_si256((__m256i *)y[i].qs, i0); +#else + // Since we don't have in AVX some necessary functions, + // we split the registers in half and call AVX2 analogs from SSE + __m128i ni0 = _mm256_castsi256_si128( i0 ); + __m128i ni1 = _mm256_extractf128_si256( i0, 1); + __m128i ni2 = _mm256_castsi256_si128( i1 ); + __m128i ni3 = _mm256_extractf128_si256( i1, 1); + __m128i ni4 = _mm256_castsi256_si128( i2 ); + __m128i ni5 = _mm256_extractf128_si256( i2, 1); + __m128i ni6 = _mm256_castsi256_si128( i3 ); + __m128i ni7 = _mm256_extractf128_si256( i3, 1); + + // Compute the sum of the quants and set y[i].s + const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3)); + const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7)); + y[i].s0 = d * hsum_i32_4(s0); + y[i].s1 = d * hsum_i32_4(s1); + + // Convert int32 to int16 + ni0 = _mm_packs_epi32( ni0, ni1 ); + ni2 = _mm_packs_epi32( ni2, ni3 ); + ni4 = _mm_packs_epi32( ni4, ni5 ); + ni6 = _mm_packs_epi32( ni6, ni7 ); + // Convert int16 to int8 + ni0 = _mm_packs_epi16( ni0, ni2 ); + ni4 = _mm_packs_epi16( ni4, ni6 ); + + _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0); + _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4); +#endif + } +#else + // scalar + quantize_row_v1_q8_1_reference(x, y, k); +#endif +} diff --git a/third_party/ggml/ggjt.v1.q8_1.h b/third_party/ggml/ggjt.v1.q8_1.h new file mode 100644 index 000000000..47fe9ec15 --- /dev/null +++ b/third_party/ggml/ggjt.v1.q8_1.h @@ -0,0 +1,20 @@ +#ifndef COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q8_1_H_ +#define COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q8_1_H_ +#if !(__ASSEMBLER__ + __LINKER__ + 0) +COSMOPOLITAN_C_START_ + +#define V1_QK8_1 32 +typedef struct { + float d; // delta + float s0; // d * sum(qs[i]) low + float s1; // d * sum(qs[i]) high + int8_t qs[V1_QK8_1]; // quants +} block_v1_q8_1; + +void quantize_row_v1_q8_1(const float* restrict, void* restrict, int); +void quantize_row_v1_q8_1_reference(const float* restrict, + block_v1_q8_1* restrict, int); + +COSMOPOLITAN_C_END_ +#endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ +#endif /* COSMOPOLITAN_THIRD_PARTY_GGML_GGJT_V1_Q8_1_H_ */ diff --git a/third_party/ggml/ggml.c b/third_party/ggml/ggml.c index d9ebd89a4..4e75185df 100644 --- a/third_party/ggml/ggml.c +++ b/third_party/ggml/ggml.c @@ -42,10 +42,10 @@ #include "third_party/aarch64/arm_neon.h" #include "third_party/intel/immintrin.internal.h" #include "libc/macros.internal.h" -#include "libc/intrin/kprintf.h" -#include "libc/intrin/kprintf.h" -#include "libc/tinymath/magicu.h" #include "libc/tinymath/magicu.h" +#include "third_party/ggml/fp16.h" +#include "third_party/ggml/fp16.internal.h" +#include "libc/assert.h" #include "third_party/libcxx/math.h" asm(".ident\t\"\\n\\n\ @@ -58,8 +58,6 @@ typedef void* thread_ret_t; /*#define GGML_PERF*/ #define GGML_DEBUG 0 -#define GGML_GELU_FP16 -#define GGML_SILU_FP16 #define GGML_SOFT_MAX_UNROLL 4 #define GGML_VEC_DOT_UNROLL 2 @@ -99,244 +97,6 @@ inline static void* ggml_aligned_malloc(size_t size) { // floating point type used to accumulate sums typedef double ggml_float; -// 16-bit float -// on Arm, we use __fp16 -// on x86, we use uint16_t -#ifdef __ARM_NEON - -// if YCM cannot find , make a symbolic link to it, for example: -// -// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/ -// -//#include - -#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x)) -#define GGML_COMPUTE_FP32_TO_FP16(x) (x) - -#define GGML_FP16_TO_FP32(x) ((float) (x)) -#define GGML_FP32_TO_FP16(x) (x) - -#else - -#ifdef __wasm_simd128__ -//#include -#else -#ifdef __POWER9_VECTOR__ -//#include -#undef bool -#define bool _Bool -#else -#if defined(_MSC_VER) || defined(__MINGW32__) -//#include -#else -//#include -#endif -#endif -#endif - -#ifdef __F16C__ - -#ifdef _MSC_VER -#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x))) -#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0) -#else -#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0) -#endif - -#elif defined(__POWER9_VECTOR__) - -#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) -/* the inline asm below is about 12% faster than the lookup method */ -#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x) -#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) - -static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - register float f; - register double d; - __asm__( - "mtfprd %0,%2\n" - "xscvhpdp %0,%0\n" - "frsp %1,%0\n" : - /* temp */ "=d"(d), - /* out */ "=f"(f): - /* in */ "r"(h)); - return f; -} - -static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { - register double d; - register ggml_fp16_t r; - __asm__( /* xscvdphp can work on double or single precision */ - "xscvdphp %0,%2\n" - "mffprd %1,%0\n" : - /* temp */ "=d"(d), - /* out */ "=r"(r): - /* in */ "f"(f)); - return r; -} - -#else - -// FP16 <-> FP32 -// ref: https://github.com/Maratyszcza/FP16 - -static inline float fp32_from_bits(uint32_t w) { - union { - uint32_t as_bits; - float as_value; - } fp32; - fp32.as_bits = w; - return fp32.as_value; -} - -static inline uint32_t fp32_to_bits(float f) { - union { - float as_value; - uint32_t as_bits; - } fp32; - fp32.as_value = f; - return fp32.as_bits; -} - -static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) { - const uint32_t w = (uint32_t) h << 16; - const uint32_t sign = w & UINT32_C(0x80000000); - const uint32_t two_w = w + w; - - const uint32_t exp_offset = UINT32_C(0xE0) << 23; -#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__) - const float exp_scale = 0x1.0p-112f; -#else - const float exp_scale = fp32_from_bits(UINT32_C(0x7800000)); -#endif - const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale; - - const uint32_t magic_mask = UINT32_C(126) << 23; - const float magic_bias = 0.5f; - const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias; - - const uint32_t denormalized_cutoff = UINT32_C(1) << 27; - const uint32_t result = sign | - (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value)); - return fp32_from_bits(result); -} - -static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) { -#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__) - const float scale_to_inf = 0x1.0p+112f; - const float scale_to_zero = 0x1.0p-110f; -#else - const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000)); - const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000)); -#endif - float base = (fabsf(f) * scale_to_inf) * scale_to_zero; - - const uint32_t w = fp32_to_bits(f); - const uint32_t shl1_w = w + w; - const uint32_t sign = w & UINT32_C(0x80000000); - uint32_t bias = shl1_w & UINT32_C(0xFF000000); - if (bias < UINT32_C(0x71000000)) { - bias = UINT32_C(0x71000000); - } - - base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base; - const uint32_t bits = fp32_to_bits(base); - const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00); - const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF); - const uint32_t nonsign = exp_bits + mantissa_bits; - return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign); -} - -#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x) -#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x) - -#endif // __F16C__ - -#endif // __ARM_NEON - -// -// global data -// - -// precomputed gelu table for f16 (128 KB) -static ggml_fp16_t table_gelu_f16[1 << 16]; - -// precomputed silu table for f16 (128 KB) -static ggml_fp16_t table_silu_f16[1 << 16]; - -// precomputed exp table for f16 (128 KB) -static ggml_fp16_t table_exp_f16[1 << 16]; - -// precomputed f32 table for f16 (256 KB) -static float table_f32_f16[1 << 16]; - -#if defined(__ARM_NEON) || defined(__wasm_simd128__) -#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s -#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s) -#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s) -#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s) -#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s) -#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s) -#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s) -#define B8(c,s ) B7(c,s, c), B7(c,s, s) - -// precomputed tables for expanding 8bits to 8 bytes (shl 4) -static const uint64_t table_b2b_u[1 << 8] = { B8(00, 10) }; -#endif - -// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32, -// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON. -// This is also true for POWER9. -#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16) - -inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) { - uint16_t s; - memcpy(&s, &f, sizeof(uint16_t)); - return table_f32_f16[s]; -} - -#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x) -#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x) - -#endif - -// note: do not use these inside ggml.c -// these are meant to be used via the ggml.h API -float ggml_fp16_to_fp32(ggml_fp16_t x) { - return (float) GGML_FP16_TO_FP32(x); -} - -ggml_fp16_t ggml_fp32_to_fp16(float x) { - return GGML_FP32_TO_FP16(x); -} - -void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, size_t n) { - for (size_t i = 0; i < n; i++) { - y[i] = GGML_FP16_TO_FP32(x[i]); - } -} - -void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, size_t n) { - size_t i = 0; -#if defined(__F16C__) - for (; i + 7 < n; i += 8) { - __m256 x_vec = _mm256_loadu_ps(x + i); - __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); - _mm_storeu_si128((__m128i *)(y + i), y_vec); - } - for(; i + 3 < n; i += 4) { - __m128 x_vec = _mm_loadu_ps(x + i); - __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT); - _mm_storel_epi64((__m128i *)(y + i), y_vec); - } -#endif - for (; i < n; i++) { - y[i] = GGML_FP32_TO_FP16(x[i]); - } -} - // // timing @@ -415,23 +175,16 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float); // #if __AVX__ || __AVX2__ || __AVX512F__ -// Unpack 16 4-bit fields into 16 bytes -// The output vector contains 16 bytes, each one in [ 0 .. 15 ] interval -static inline __m128i bytes_from_nibbles_16(const uint8_t * rsi) -{ - // Load 8 bytes from memory - __m128i tmp = _mm_loadl_epi64( ( const __m128i* )rsi ); - - // Expand bytes into uint16_t values - __m128i bytes = _mm_cvtepu8_epi16( tmp ); - - // Unpack values into individual bytes - const __m128i lowMask = _mm_set1_epi8( 0xF ); - __m128i high = _mm_andnot_si128( lowMask, bytes ); - __m128i low = _mm_and_si128( lowMask, bytes ); - high = _mm_slli_epi16( high, 4 ); - bytes = _mm_or_si128( low, high ); - return bytes; +// multiply int8_t, add results pairwise twice +static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) { + // Get absolute values of x vectors + const __m128i ax = _mm_sign_epi8(x, x); + // Sign the values of the y vectors + const __m128i sy = _mm_sign_epi8(y, x); + // Perform multiplication and create 16-bit values + const __m128i dot = _mm_maddubs_epi16(ax, sy); + const __m128i ones = _mm_set1_epi16(1); + return _mm_madd_epi16(ones, dot); } // horizontally add 8 floats @@ -466,8 +219,8 @@ static inline __m256i bytes_from_bits_32(const uint8_t * x) { uint32_t x32; memcpy(&x32, x, sizeof(uint32_t)); const __m256i shuf_mask = _mm256_set_epi64x( - 0x0303030303030303, 0x0202020202020202, - 0x0101010101010101, 0x0000000000000000); + 0x0303030303030303, 0x0202020202020202, + 0x0101010101010101, 0x0000000000000000); __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask); const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe); bytes = _mm256_or_si256(bytes, bit_mask); @@ -478,19 +231,10 @@ static inline __m256i bytes_from_bits_32(const uint8_t * x) { // The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) { - // Load 16 bytes from memory - __m128i tmp = _mm_loadu_si128( ( const __m128i* )rsi ); - - // Expand bytes into uint16_t values - __m256i bytes = _mm256_cvtepu8_epi16( tmp ); - - // Unpack values into individual bytes + const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi); + const __m256i bytes = _mm256_set_m128i(_mm_srli_epi16(tmp, 4), tmp); const __m256i lowMask = _mm256_set1_epi8( 0xF ); - __m256i high = _mm256_andnot_si256( lowMask, bytes ); - __m256i low = _mm256_and_si256( lowMask, bytes ); - high = _mm256_slli_epi16( high, 4 ); - bytes = _mm256_or_si256( low, high ); - return bytes; + return _mm256_and_si256(lowMask, bytes); } // add int16_t pairwise and return as float vector @@ -620,94 +364,6 @@ float vmaxvq_f32(float32x4_t v) { MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3))); } -int8x8_t vzip1_s8(int8x8_t a, int8x8_t b) { - int8x8_t res; - - res[0] = a[0]; res[1] = b[0]; - res[2] = a[1]; res[3] = b[1]; - res[4] = a[2]; res[5] = b[2]; - res[6] = a[3]; res[7] = b[3]; - - return res; -} - -int8x8_t vzip2_s8(int8x8_t a, int8x8_t b) { - int8x8_t res; - - res[0] = a[4]; res[1] = b[4]; - res[2] = a[5]; res[3] = b[5]; - res[4] = a[6]; res[5] = b[6]; - res[6] = a[7]; res[7] = b[7]; - - return res; -} - -uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) { - uint8x8_t res; - - res[0] = a[0]; res[1] = b[0]; - res[2] = a[1]; res[3] = b[1]; - res[4] = a[2]; res[5] = b[2]; - res[6] = a[3]; res[7] = b[3]; - - return res; -} - -uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) { - uint8x8_t res; - - res[0] = a[4]; res[1] = b[4]; - res[2] = a[5]; res[3] = b[5]; - res[4] = a[6]; res[5] = b[6]; - res[6] = a[7]; res[7] = b[7]; - - return res; -} - -int8x16_t vzip1q_s8(int8x16_t a, int8x16_t b) { - int8x16_t res; - - res[0] = a[0]; res[1] = b[0]; res[2] = a[1]; res[3] = b[1]; - res[4] = a[2]; res[5] = b[2]; res[6] = a[3]; res[7] = b[3]; - res[8] = a[4]; res[9] = b[4]; res[10] = a[5]; res[11] = b[5]; - res[12] = a[6]; res[13] = b[6]; res[14] = a[7]; res[15] = b[7]; - - return res; -} - -int8x16_t vzip2q_s8(int8x16_t a, int8x16_t b) { - int8x16_t res; - - res[0] = a[8]; res[1] = b[8]; res[2] = a[9]; res[3] = b[9]; - res[4] = a[10]; res[5] = b[10]; res[6] = a[11]; res[7] = b[11]; - res[8] = a[12]; res[9] = b[12]; res[10] = a[13]; res[11] = b[13]; - res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15]; - - return res; -} - -uint8x16_t vzip1q_u8(uint8x16_t a, uint8x16_t b) { - uint8x16_t res; - - res[0] = a[0]; res[1] = b[0]; res[2] = a[1]; res[3] = b[1]; - res[4] = a[2]; res[5] = b[2]; res[6] = a[3]; res[7] = b[3]; - res[8] = a[4]; res[9] = b[4]; res[10] = a[5]; res[11] = b[5]; - res[12] = a[6]; res[13] = b[6]; res[14] = a[7]; res[15] = b[7]; - - return res; -} - -uint8x16_t vzip2q_u8(uint8x16_t a, uint8x16_t b) { - uint8x16_t res; - - res[0] = a[8]; res[1] = b[8]; res[2] = a[9]; res[3] = b[9]; - res[4] = a[10]; res[5] = b[10]; res[6] = a[11]; res[7] = b[11]; - res[8] = a[12]; res[9] = b[12]; res[10] = a[13]; res[11] = b[13]; - res[12] = a[14]; res[13] = b[14]; res[14] = a[15]; res[15] = b[15]; - - return res; -} - int32x4_t vcvtnq_s32_f32(float32x4_t v) { int32x4_t res; @@ -738,13 +394,6 @@ typedef struct { } block_q4_1; static_assert(sizeof(block_q4_1) == 2 * sizeof(float) + QK4_1 / 2, "wrong q4_1 block size/padding"); -#define QK4_2 16 -typedef struct { - ggml_fp16_t d; // delta - uint8_t qs[QK4_2 / 2]; // nibbles / quants -} block_q4_2; -static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding"); - #define QK5_0 32 typedef struct { ggml_fp16_t d; // delta @@ -771,634 +420,162 @@ static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block siz #define QK8_1 32 typedef struct { - float d; // delta - float s0; // d * sum(qs[i]) low - float s1; // d * sum(qs[i]) high - int8_t qs[QK8_1]; // quants + float d; // delta + float s; // d * sum(qs[i]) + int8_t qs[QK8_1]; // quants } block_q8_1; -static_assert(sizeof(block_q8_1) == 3*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); +static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding"); // reference implementation for deterministic creation of model files static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) { - assert(k % QK4_0 == 0); - const int nb = k / QK4_0; + static const int qk = QK4_0; - uint8_t pp[QK4_0/2]; + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max - float max = 0.0f; + float max = 0.0f; - for (int l = 0; l < QK4_0; l++) { - const float v = x[i*QK4_0 + l]; + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; if (amax < fabsf(v)) { amax = fabsf(v); - max = v; + max = v; } } - const float d = max / -8; + const float d = max / -8; const float id = d ? 1.0f/d : 0.0f; y[i].d = d; - for (int l = 0; l < QK4_0; l += 2) { - const float v0 = x[i*QK4_0 + l + 0]*id; - const float v1 = x[i*QK4_0 + l + 1]*id; + for (int j = 0; j < qk/2; ++j) { + const float x0 = x[i*qk + 0 + j]*id; + const float x1 = x[i*qk + qk/2 + j]*id; - const uint8_t vi0 = MIN(15, (int8_t)roundf(v0) + 8); - const uint8_t vi1 = MIN(15, (int8_t)roundf(v1) + 8); + const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f)); + const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f)); - assert(vi0 < 16); - assert(vi1 < 16); - - pp[l/2] = vi0 | (vi1 << 4); + y[i].qs[j] = xi0; + y[i].qs[j] |= xi1 << 4; } - - memcpy(y[i].qs, pp, sizeof(pp)); } } -static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_0 == 0); - const int nb = k / QK4_0; - - block_q4_0 * restrict y = vy; - -#if defined(__POWER9_VECTOR__) - const vector float v85 = vec_splats(8.5f); - const vector signed int v15 = vec_splats(15); - for (int i = 0; i < nb; i++) { - float max = 0.0f; - float min = 0.0f; - - vector float asrcv [8]; - vector float srcv [8]; - vector float maxv[8]; - vector float minv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = *(vector float *)(x + i*32 + 4*l); - //for (int l = 0; l < 8; l++) asrcv[l] = vec_abs(srcv[l]); - - for (int l = 0; l < 4; l++) maxv[2*l] = vec_max(asrcv[2*l], asrcv[2*l+1]); - //for (int l = 0; l < 2; l++) maxv[4*l] = vec_max(maxv[4*l], maxv[4*l+2]); - maxv[0] = vec_max(maxv[0], maxv[2]); - maxv[4] = vec_max(maxv[4], maxv[6]); - //for (int l = 0; l < 1; l++) maxv[8*l] = vec_max(maxv[8*l], maxv[8*l+4]); - maxv[0] = vec_max(maxv[0], maxv[4]); - - for (int l = 0; l < 4; l++) minv[2*l] = vec_min(asrcv[2*l], asrcv[2*l+1]); - //for (int l = 0; l < 2; l++) minv[4*l] = vec_min(minv[4*l], minv[4*l+2]); - minv[0] = vec_min(minv[0], minv[2]); - minv[4] = vec_min(minv[4], minv[6]); - //for (int l = 0; l < 1; l++) minv[8*l] = vec_min(minv[8*l], minv[8*l+4]); - minv[0] = vec_min(minv[0], minv[4]); - - - max = MAX( - MAX(vec_extract(maxv[0], 0), vec_extract(maxv[0], 1)), - MAX(vec_extract(maxv[0], 2), vec_extract(maxv[0], 3))); - min = MIN( - MIN(vec_extract(minv[0], 0), vec_extract(minv[0], 1)), - MIN(vec_extract(minv[0], 2), vec_extract(minv[0], 3))); - - const float magnitude = max >= fabsf(min) ? max : min; - const float d = magnitude / -8; - const float id = d ? 1.0/d : 0.0; - - y[i].d = d; - - const vector float vid = vec_splats(id); - uint8_t * restrict pb = y[i].qs; - for (int l = 0; l < 8; l++) { - const vector float vf = vec_madd(srcv[l], vid, v85); - const vector signed int vi = vec_signed(vf); - const vector signed int vc = vec_min(vi, v15); - - pb[2*l + 0] = vec_extract(vc, 0) | (vec_extract(vc, 1) << 4); - pb[2*l + 1] = vec_extract(vc, 2) | (vec_extract(vc, 3) << 4); - } - } -#elif __ARM_NEON - for (int i = 0; i < nb; i++) { - float32x4_t srcv [8]; - float32x4_t maxv[8]; - float32x4_t minv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); - - for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l+1]); - for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l+2]); - for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l+4]); - - for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l+1]); - for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l+2]); - for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l+4]); - - const float max = vmaxvq_f32(maxv[0]); - const float min = vminvq_f32(minv[0]); - - const float magnitude = max >= fabsf(min) ? max : min; - const float d = magnitude / -8; - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - - for (int l = 0; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(srcv[l], id); - const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(8.5f)); - const int32x4_t vi = vcvtq_s32_f32(vf); - const int32x4_t vc = vminq_s32(vi, vdupq_n_s32(15)); - - y[i].qs[2*l + 0] = vgetq_lane_s32(vc, 0) | (vgetq_lane_s32(vc, 1) << 4); - y[i].qs[2*l + 1] = vgetq_lane_s32(vc, 2) | (vgetq_lane_s32(vc, 3) << 4); - } - } -#elif defined(__AVX2__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max for the block - __m256 max = _mm256_max_ps( v0, v1 ); - __m256 maxTmp = _mm256_max_ps( v2, v3 ); - max = _mm256_max_ps( max, maxTmp ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Compute min for the block - __m256 min = _mm256_min_ps( v0, v1 ); - __m256 minTmp = _mm256_min_ps( v2, v3 ); - min = _mm256_min_ps( min, minTmp ); - - __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) ); - min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); - min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); - const float minScalar = _mm_cvtss_f32( min4 ); - - // Quantize these floats - const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar; - const float d = magnitude / -8.0f; - y[i].d = d; - const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f; - const __m256 mul = _mm256_set1_ps( id ); - - // Apply the multiplier - v0 = _mm256_mul_ps( v0, mul ); - v1 = _mm256_mul_ps( v1, mul ); - v2 = _mm256_mul_ps( v2, mul ); - v3 = _mm256_mul_ps( v3, mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - - // Convert int32 to int16 - i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 - i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 - // Convert int16 to int8 - i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 - - // We got our precious signed bytes, but the order is now wrong - // These AVX2 pack instructions process 16-byte pieces independently - // The following instruction is fixing the order - const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); - i0 = _mm256_permutevar8x32_epi32( i0, perm ); - - // Apply offset and clamp to translate the range from [ -8 .. +8 ] into [ +0 .. +15 ] - const __m256i off = _mm256_set1_epi8( 8 ); - i0 = _mm256_add_epi8( i0, off ); - const __m256i maxNibble = _mm256_set1_epi8( 15 ); - i0 = _mm256_min_epi8( i0, maxNibble ); - - // Compress the vector into 4 bit/value, and store - __m128i res = packNibbles( i0 ); - _mm_storeu_si128( ( __m128i* )y[i].qs, res ); - } -#elif defined(__AVX__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max for the block - __m256 max = _mm256_max_ps( v0, v1 ); - __m256 maxTmp = _mm256_max_ps( v2, v3 ); - max = _mm256_max_ps( max, maxTmp ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( max, 1 ), _mm256_castps256_ps128( max ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Compute min for the block - __m256 min = _mm256_min_ps( v0, v1 ); - __m256 minTmp = _mm256_min_ps( v2, v3 ); - min = _mm256_min_ps( min, minTmp ); - - __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( min, 1 ), _mm256_castps256_ps128( min ) ); - min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); - min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); - const float minScalar = _mm_cvtss_f32( min4 ); - - // Quantize these floats - const float magnitude = maxScalar >= fabsf(minScalar) ? maxScalar : minScalar; - const float d = magnitude / -8.0f; - y[i].d = d; - const float id = ( magnitude != 0.0f ) ? -8.0f / magnitude : 0.0f; - const __m256 mul = _mm256_set1_ps( id ); - - // Apply the multiplier - v0 = _mm256_mul_ps( v0, mul ); - v1 = _mm256_mul_ps( v1, mul ); - v2 = _mm256_mul_ps( v2, mul ); - v3 = _mm256_mul_ps( v3, mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - - // Since we don't have in AVX some necessary functions, - // we split the registers in half and call AVX2 analogs from SSE - __m128i ni0 = _mm256_castsi256_si128( i0 ); - __m128i ni1 = _mm256_extractf128_si256( i0, 1); - __m128i ni2 = _mm256_castsi256_si128( i1 ); - __m128i ni3 = _mm256_extractf128_si256( i1, 1); - __m128i ni4 = _mm256_castsi256_si128( i2 ); - __m128i ni5 = _mm256_extractf128_si256( i2, 1); - __m128i ni6 = _mm256_castsi256_si128( i3 ); - __m128i ni7 = _mm256_extractf128_si256( i3, 1); - - // Convert int32 to int16 - ni0 = _mm_packs_epi32( ni0, ni1 ); - ni2 = _mm_packs_epi32( ni2, ni3 ); - ni4 = _mm_packs_epi32( ni4, ni5 ); - ni6 = _mm_packs_epi32( ni6, ni7 ); - // Convert int16 to int8 - ni0 = _mm_packs_epi16( ni0, ni2 ); - ni4 = _mm_packs_epi16( ni4, ni6 ); - - // Apply offset and clamp to translate the range from [ -8 .. +8 ] into [ +0 .. +15 ] - const __m128i off = _mm_set1_epi8( 8 ); - ni0 = _mm_add_epi8( ni0, off ); - ni4 = _mm_add_epi8( ni4, off ); - const __m128i maxNibble = _mm_set1_epi8( 15 ); - ni0 = _mm_min_epi8( ni0, maxNibble ); - ni4 = _mm_min_epi8( ni4, maxNibble ); - - // Compress the vector into 4 bit/value, and store - __m128i res = packNibbles( ni0, ni4 ); - _mm_storeu_si128( ( __m128i* )y[i].qs, res ); - } -#elif defined(__wasm_simd128__) - for (int i = 0; i < nb; i++) { - float max = 0.0f; - float min = 0.0f; - - v128_t srcv [8]; - v128_t maxv[8]; - v128_t minv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = wasm_v128_load(x + i*32 + 4*l); - - for (int l = 0; l < 4; l++) maxv[2*l] = wasm_f32x4_max(srcv[2*l], srcv[2*l+1]); - for (int l = 0; l < 2; l++) maxv[4*l] = wasm_f32x4_max(maxv[4*l], maxv[4*l+2]); - for (int l = 0; l < 1; l++) maxv[8*l] = wasm_f32x4_max(maxv[8*l], maxv[8*l+4]); - - for (int l = 0; l < 4; l++) minv[2*l] = wasm_f32x4_min(srcv[2*l], srcv[2*l+1]); - for (int l = 0; l < 2; l++) minv[4*l] = wasm_f32x4_min(minv[4*l], minv[4*l+2]); - for (int l = 0; l < 1; l++) minv[8*l] = wasm_f32x4_min(minv[8*l], minv[8*l+4]); - - max = MAX( - MAX(wasm_f32x4_extract_lane(maxv[0], 0), wasm_f32x4_extract_lane(maxv[0], 1)), - MAX(wasm_f32x4_extract_lane(maxv[0], 2), wasm_f32x4_extract_lane(maxv[0], 3))); - min = MIN( - MIN(wasm_f32x4_extract_lane(minv[0], 0), wasm_f32x4_extract_lane(minv[0], 1)), - MIN(wasm_f32x4_extract_lane(minv[0], 2), wasm_f32x4_extract_lane(minv[0], 3))); - - const float magnitude = max >= fabsf(min) ? max : min; - const float d = magnitude / -8; - const float id = d ? 1.0/d : 0.0; - - y[i].d = d; - - for (int l = 0; l < 8; l++) { - const v128_t v = wasm_f32x4_mul(srcv[l], wasm_f32x4_splat(id)); - const v128_t vf = wasm_f32x4_add(v, wasm_f32x4_splat(8.5f)); - const v128_t vi = wasm_i32x4_trunc_sat_f32x4(vf); - const v128_t vc = wasm_i32x4_min(vi, wasm_i32x4_splat(15)); - - y[i].qs[2*l + 0] = wasm_i32x4_extract_lane(vc, 0) | (wasm_i32x4_extract_lane(vc, 1) << 4); - y[i].qs[2*l + 1] = wasm_i32x4_extract_lane(vc, 2) | (wasm_i32x4_extract_lane(vc, 3) << 4); - } - } -#else - // scalar +static void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) { quantize_row_q4_0_reference(x, y, k); -#endif } -static void quantize_row_q4_1_reference(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_1 == 0); - const int nb = k / QK4_1; +static void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k) { + const int qk = QK4_1; - block_q4_1 * restrict y = vy; + assert(k % qk == 0); - uint8_t pp[QK4_1/2]; + const int nb = k / qk; for (int i = 0; i < nb; i++) { float min = FLT_MAX; float max = -FLT_MAX; - for (int l = 0; l < QK4_1; l++) { - const float v = x[i*QK4_1 + l]; + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + if (v < min) min = v; if (v > max) max = v; } - const float d = (max - min) / ((1 << 4) - 1); + const float d = (max - min) / ((1 << 4) - 1); const float id = d ? 1.0f/d : 0.0f; y[i].d = d; y[i].m = min; - for (int l = 0; l < QK4_1; l += 2) { - const float v0 = (x[i*QK4_1 + l + 0] - min)*id; - const float v1 = (x[i*QK4_1 + l + 1] - min)*id; + for (int j = 0; j < qk/2; ++j) { + const float x0 = (x[i*qk + 0 + j] - min)*id; + const float x1 = (x[i*qk + qk/2 + j] - min)*id; - const uint8_t vi0 = roundf(v0); - const uint8_t vi1 = roundf(v1); + const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f)); + const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f)); - assert(vi0 < 16); - assert(vi1 < 16); - - pp[l/2] = vi0 | (vi1 << 4); - } - - memcpy(y[i].qs, pp, sizeof(pp)); - } -} - -static void quantize_row_q4_1(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_1 == 0); - - const int nb = k / QK4_1; - - block_q4_1 * restrict y = vy; - -#if defined(__AVX2__) - for (int i = 0; i < nb; i++) { - // Load elements into 4 AVX vectors - __m256 v0 = _mm256_loadu_ps( x ); - __m256 v1 = _mm256_loadu_ps( x + 8 ); - __m256 v2 = _mm256_loadu_ps( x + 16 ); - __m256 v3 = _mm256_loadu_ps( x + 24 ); - x += 32; - - // Compute max for the block - __m256 vmax; - vmax = _mm256_max_ps( v0, v1 ); - vmax = _mm256_max_ps( vmax, v2 ); - vmax = _mm256_max_ps( vmax, v3 ); - - __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( vmax, 1 ), _mm256_castps256_ps128( vmax ) ); - max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) ); - max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) ); - const float maxScalar = _mm_cvtss_f32( max4 ); - - // Compute min for the block - __m256 vmin; - vmin = _mm256_min_ps( v0, v1 ); - vmin = _mm256_min_ps( vmin, v2 ); - vmin = _mm256_min_ps( vmin, v3 ); - - __m128 min4 = _mm_min_ps( _mm256_extractf128_ps( vmin, 1 ), _mm256_castps256_ps128( vmin ) ); - min4 = _mm_min_ps( min4, _mm_movehl_ps( min4, min4 ) ); - min4 = _mm_min_ss( min4, _mm_movehdup_ps( min4 ) ); - const float minScalar = _mm_cvtss_f32( min4 ); - - // Quantize these floats - const float d = (maxScalar - minScalar) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].m = minScalar; - y[i].d = d; - - // x = (x-min)*id - const __m256 mul = _mm256_set1_ps( id ); - const __m256 off = _mm256_set1_ps( minScalar ); - v0 = _mm256_mul_ps( _mm256_sub_ps( v0, off ), mul ); - v1 = _mm256_mul_ps( _mm256_sub_ps( v1, off ), mul ); - v2 = _mm256_mul_ps( _mm256_sub_ps( v2, off ), mul ); - v3 = _mm256_mul_ps( _mm256_sub_ps( v3, off ), mul ); - - // Round to nearest integer - v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST ); - v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST ); - v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST ); - v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST ); - - // Convert floats to integers - __m256i i0 = _mm256_cvtps_epi32( v0 ); - __m256i i1 = _mm256_cvtps_epi32( v1 ); - __m256i i2 = _mm256_cvtps_epi32( v2 ); - __m256i i3 = _mm256_cvtps_epi32( v3 ); - - // Convert int32 to int16 - i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 - i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 - // Convert int16 to int8 - i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 - - // We got our precious signed bytes, but the order is now wrong - // These AVX2 pack instructions process 16-byte pieces independently - // The following instruction is fixing the order - const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 ); - i0 = _mm256_permutevar8x32_epi32( i0, perm ); - - // Compress the vector into 4 bit/value, and store - __m128i res = packNibbles( i0 ); - _mm_storeu_si128( ( __m128i* )y[i].qs, res ); - } -#elif __ARM_NEON - for (int i = 0; i < nb; i++) { - float32x4_t srcv[8]; - float32x4_t minv[8]; - float32x4_t maxv[8]; - - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*QK4_1 + 4*l); - - for (int l = 0; l < 4; l++) minv[2*l] = vminq_f32(srcv[2*l], srcv[2*l + 1]); - for (int l = 0; l < 2; l++) minv[4*l] = vminq_f32(minv[4*l], minv[4*l + 2]); - for (int l = 0; l < 1; l++) minv[8*l] = vminq_f32(minv[8*l], minv[8*l + 4]); - - for (int l = 0; l < 4; l++) maxv[2*l] = vmaxq_f32(srcv[2*l], srcv[2*l + 1]); - for (int l = 0; l < 2; l++) maxv[4*l] = vmaxq_f32(maxv[4*l], maxv[4*l + 2]); - for (int l = 0; l < 1; l++) maxv[8*l] = vmaxq_f32(maxv[8*l], maxv[8*l + 4]); - - const float min = vminvq_f32(minv[0]); - const float max = vmaxvq_f32(maxv[0]); - - const float d = (max - min) / ((1 << 4) - 1); - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = d; - y[i].m = min; - - const float32x4_t minv0 = vdupq_n_f32(min); - - for (int l = 0; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(vsubq_f32(srcv[l], minv0), id); - const float32x4_t vf = vaddq_f32(v, vdupq_n_f32(0.5f)); // needed to round to nearest - const int32x4_t vi = vcvtq_s32_f32(vf); - - y[i].qs[2*l + 0] = vgetq_lane_s32(vi, 0) | (vgetq_lane_s32(vi, 1) << 4); - y[i].qs[2*l + 1] = vgetq_lane_s32(vi, 2) | (vgetq_lane_s32(vi, 3) << 4); - } - } -#else - // scalar - quantize_row_q4_1_reference(x, vy, k); -#endif -} - -// reference implementation for deterministic creation of model files -static void quantize_row_q4_2_reference(const float * restrict x, block_q4_2 * restrict y, int k) { - assert(k % QK4_2 == 0); - - const int nb = k / QK4_2; - - for (int i = 0; i < nb; i++) { - float amax = 0.0f; // absolute max - float max = 0.0f; - - for (int l = 0; l < QK4_2; l++) { - const float v = x[i*QK4_2 + l]; - if (amax < fabsf(v)) { - amax = fabsf(v); - max = v; - } - } - - const float d = max / -8; - - const float id = d ? 1.0f/d : 0.0f; - - y[i].d = GGML_FP32_TO_FP16(d); - - for (int l = 0; l < QK4_2; l += 2) { - const float v0 = x[i*QK4_2 + l + 0]*id; - const float v1 = x[i*QK4_2 + l + 1]*id; - - const uint8_t vi0 = MIN(15, (uint8_t)(v0 + 8.5f)); - const uint8_t vi1 = MIN(15, (uint8_t)(v1 + 8.5f)); - - assert(vi0 < 16); - assert(vi1 < 16); - - y[i].qs[l/2] = vi0 | (vi1 << 4); + y[i].qs[j] = xi0; + y[i].qs[j] |= xi1 << 4; } } } -static void quantize_row_q4_2(const float * restrict x, void * restrict vy, int k) { - assert(k % QK4_2 == 0); - - block_q4_2 * restrict y = vy; - - quantize_row_q4_2_reference(x, y, k); +static void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) { + quantize_row_q4_1_reference(x, y, k); } static void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) { - assert(k % QK5_0 == 0); - const int nb = k / QK5_0; + static const int qk = QK5_0; + + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max - float max = 0.0f; + float max = 0.0f; - for (int l = 0; l < QK5_0; l++) { - const float v = x[i*QK5_0 + l]; + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; if (amax < fabsf(v)) { amax = fabsf(v); - max = v; + max = v; } } - const float d = max / -16; + const float d = max / -16; const float id = d ? 1.0f/d : 0.0f; y[i].d = GGML_FP32_TO_FP16(d); uint32_t qh = 0; - for (int l = 0; l < QK5_0; l += 2) { - const float v0 = x[i*QK5_0 + l + 0]*id; - const float v1 = x[i*QK5_0 + l + 1]*id; + for (int j = 0; j < qk/2; ++j) { + const float x0 = x[i*qk + 0 + j]*id; + const float x1 = x[i*qk + qk/2 + j]*id; - const uint32_t vi0 = MIN(31, (int) (v0 + 16.5f)); - const uint32_t vi1 = MIN(31, (int) (v1 + 16.5f)); + const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f)); + const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f)); - y[i].qs[l/2] = (vi0 & 0x0F) | ((vi1 & 0x0F) << 4); + y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); // get the 5-th bit and store it in qh at the right position - qh |= ((vi0 & 0x10) >> 4) << (l + 0); - qh |= ((vi1 & 0x10) >> 4) << (l + 1); + qh |= ((xi0 & 0x10) >> 4) << (j + 0); + qh |= ((xi1 & 0x10) >> 4) << (j + qk/2); } - memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); + memcpy(&y[i].qh, &qh, sizeof(qh)); } } -static void quantize_row_q5_0(const float * restrict x, void * restrict vy, int k) { - assert(k % QK5_0 == 0); - - block_q5_0 * restrict y = vy; - +static void quantize_row_q5_0(const float * restrict x, void * restrict y, int k) { quantize_row_q5_0_reference(x, y, k); } static void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k) { - assert(k % QK5_1 == 0); - const int nb = k / QK5_1; + const int qk = QK5_1; + + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { float min = FLT_MAX; float max = -FLT_MAX; - for (int l = 0; l < QK5_1; l++) { - const float v = x[i*QK5_1 + l]; + for (int j = 0; j < qk; j++) { + const float v = x[i*qk + j]; + if (v < min) min = v; if (v > max) max = v; } - const float d = (max - min) / ((1 << 5) - 1); + const float d = (max - min) / ((1 << 5) - 1); const float id = d ? 1.0f/d : 0.0f; y[i].d = GGML_FP32_TO_FP16(d); @@ -1406,29 +583,25 @@ static void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * r uint32_t qh = 0; - for (int l = 0; l < QK5_1; l += 2) { - const float v0 = (x[i*QK5_1 + l + 0] - min)*id; - const float v1 = (x[i*QK5_1 + l + 1] - min)*id; + for (int j = 0; j < qk/2; ++j) { + const float x0 = (x[i*qk + 0 + j] - min)*id; + const float x1 = (x[i*qk + qk/2 + j] - min)*id; - const uint32_t vi0 = (int) (v0 + 0.5f); - const uint32_t vi1 = (int) (v1 + 0.5f); + const uint8_t xi0 = (uint8_t)(x0 + 0.5f); + const uint8_t xi1 = (uint8_t)(x1 + 0.5f); - y[i].qs[l/2] = (vi0 & 0x0F) | ((vi1 & 0x0F) << 4); + y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4); // get the 5-th bit and store it in qh at the right position - qh |= ((vi0 & 0x10) >> 4) << (l + 0); - qh |= ((vi1 & 0x10) >> 4) << (l + 1); + qh |= ((xi0 & 0x10) >> 4) << (j + 0); + qh |= ((xi1 & 0x10) >> 4) << (j + qk/2); } memcpy(&y[i].qh, &qh, sizeof(y[i].qh)); } } -static void quantize_row_q5_1(const float * restrict x, void * restrict vy, int k) { - assert(k % QK5_1 == 0); - - block_q5_1 * restrict y = vy; - +static void quantize_row_q5_1(const float * restrict x, void * restrict y, int k) { quantize_row_q5_1_reference(x, y, k); } @@ -1440,8 +613,8 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max - for (int l = 0; l < QK8_0; l++) { - const float v = x[i*QK8_0 + l]; + for (int j = 0; j < QK8_0; j++) { + const float v = x[i*QK8_0 + j]; amax = MAX(amax, fabsf(v)); } @@ -1450,10 +623,10 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r y[i].d = d; - for (int l = 0; l < QK8_0; ++l) { - const float v0 = x[i*QK8_0 + l]*id; + for (int j = 0; j < QK8_0; ++j) { + const float x0 = x[i*QK8_0 + j]*id; - y[i].qs[l] = roundf(v0); + y[i].qs[j] = roundf(x0); } } } @@ -1471,12 +644,12 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int float32x4_t asrcv[8]; float32x4_t amaxv[8]; - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); - for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]); + for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); - for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]); - for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]); - for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]); + for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); const float amax = vmaxvq_f32(amaxv[0]); @@ -1485,14 +658,14 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int y[i].d = d; - for (int l = 0; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(srcv[l], id); + for (int j = 0; j < 8; j++) { + const float32x4_t v = vmulq_n_f32(srcv[j], id); const int32x4_t vi = vcvtnq_s32_f32(v); - y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); + y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); } } #elif defined(__AVX2__) || defined(__AVX__) @@ -1594,8 +767,8 @@ static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * r for (int i = 0; i < nb; i++) { float amax = 0.0f; // absolute max - for (int l = 0; l < QK8_1; l++) { - const float v = x[i*QK8_1 + l]; + for (int j = 0; j < QK8_1; j++) { + const float v = x[i*QK8_1 + j]; amax = MAX(amax, fabsf(v)); } @@ -1604,22 +777,20 @@ static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * r y[i].d = d; - int sum0 = 0; - int sum1 = 0; + int sum = 0; - for (int l = 0; l < QK8_1/2; ++l) { - const float v0 = x[i*QK8_1 + l]*id; - const float v1 = x[i*QK8_1 + QK8_1/2 + l]*id; + for (int j = 0; j < QK8_1/2; ++j) { + const float v0 = x[i*QK8_1 + j]*id; + const float v1 = x[i*QK8_1 + QK8_1/2 + j]*id; - y[i].qs[ l] = roundf(v0); - y[i].qs[QK8_1/2 + l] = roundf(v1); + y[i].qs[ j] = roundf(v0); + y[i].qs[QK8_1/2 + j] = roundf(v1); - sum0 += y[i].qs[ l]; - sum1 += y[i].qs[QK8_1/2 + l]; + sum += y[i].qs[ j]; + sum += y[i].qs[QK8_1/2 + j]; } - y[i].s0 = d * sum0; - y[i].s1 = d * sum1; + y[i].s = d * sum; } } @@ -1635,12 +806,12 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int float32x4_t asrcv[8]; float32x4_t amaxv[8]; - for (int l = 0; l < 8; l++) srcv[l] = vld1q_f32(x + i*32 + 4*l); - for (int l = 0; l < 8; l++) asrcv[l] = vabsq_f32(srcv[l]); + for (int j = 0; j < 8; j++) srcv[j] = vld1q_f32(x + i*32 + 4*j); + for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]); - for (int l = 0; l < 4; l++) amaxv[2*l] = vmaxq_f32(asrcv[2*l], asrcv[2*l+1]); - for (int l = 0; l < 2; l++) amaxv[4*l] = vmaxq_f32(amaxv[4*l], amaxv[4*l+2]); - for (int l = 0; l < 1; l++) amaxv[8*l] = vmaxq_f32(amaxv[8*l], amaxv[8*l+4]); + for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]); + for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]); + for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]); const float amax = vmaxvq_f32(amaxv[0]); @@ -1649,40 +820,21 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int y[i].d = d; - int32x4_t accv0 = vdupq_n_s32(0); - int32x4_t accv1 = vdupq_n_s32(0); + int32x4_t accv = vdupq_n_s32(0); - // low half - for (int l = 0; l < 4; l++) { - const float32x4_t v = vmulq_n_f32(srcv[l], id); + for (int j = 0; j < 8; j++) { + const float32x4_t v = vmulq_n_f32(srcv[j], id); const int32x4_t vi = vcvtnq_s32_f32(v); - y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); + y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0); + y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1); + y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2); + y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3); - accv0 = vaddq_s32(accv0, vi); + accv = vaddq_s32(accv, vi); } - // high half - for (int l = 4; l < 8; l++) { - const float32x4_t v = vmulq_n_f32(srcv[l], id); - const int32x4_t vi = vcvtnq_s32_f32(v); - - y[i].qs[4*l + 0] = vgetq_lane_s32(vi, 0); - y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1); - y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2); - y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3); - - accv1 = vaddq_s32(accv1, vi); - } - - const int32_t sum0 = vaddvq_s32(accv0); - const int32_t sum1 = vaddvq_s32(accv1); - - y[i].s0 = d * sum0; - y[i].s1 = d * sum1; + y[i].s = d * vaddvq_s32(accv); } #elif defined(__AVX2__) || defined(__AVX__) for (int i = 0; i < nb; i++) { @@ -1731,9 +883,7 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int #if defined(__AVX2__) // Compute the sum of the quants and set y[i].s - //y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))); - y[i].s0 = d * hsum_i32_8(_mm256_add_epi32(i0, i1)); - y[i].s1 = d * hsum_i32_8(_mm256_add_epi32(i2, i3)); + y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))); // Convert int32 to int16 i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 @@ -1763,8 +913,7 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int // Compute the sum of the quants and set y[i].s const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3)); const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7)); - y[i].s0 = d * hsum_i32_4(s0); - y[i].s1 = d * hsum_i32_4(s1); + y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1)); // Convert int32 to int16 ni0 = _mm_packs_epi32( ni0, ni1 ); @@ -1785,359 +934,127 @@ static void quantize_row_q8_1(const float * restrict x, void * restrict vy, int #endif } -static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, int k) { - assert(k % QK4_0 == 0); - const int nb = k / QK4_0; +static void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k) { + static const int qk = QK4_0; - const block_q4_0 * restrict x = vx; + assert(k % qk == 0); -#if defined(__AVX2__) - for (int i = 0; i < nb; i++) { - // scale factor - const __m256 d_v = _mm256_broadcast_ss(&x[i].d); + const int nb = k / qk; - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_0; l += 32) { - // Load 32x4-bit integers into 32x8-bit integers - __m256i vx8 = bytes_from_nibbles_32(pp+l/2); - - // Subtract 8 from the integers - vx8 = _mm256_sub_epi8(vx8, _mm256_set1_epi8(8)); - - // Convert to 16-bit int - const __m256i vx16_lo = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 0)); - const __m256i vx16_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 1)); - - // Convert to 32-bit int -> float 32 - const __m256 vf[4] = { - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 1))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 1))) - }; - - // Scale and store - for (int j = 0; j < 4; j++) { - const __m256 result = _mm256_mul_ps(vf[j], d_v); - _mm256_storeu_ps(y + i * QK4_0 + l + j*8, result); - } - } - } -#elif defined(__ARM_NEON) - for (int i = 0; i < nb; i++) { - const float32x4_t vd = vdupq_n_f32(x[i].d); - - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_0; l += 16) { - // Load 16x4-bit integers into 8x8-bit integers - const uint8x8_t v8 = vld1_u8(pp + l/2); - - // Expand 4-bit qs to 8-bit bytes - const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0F)); - const uint8x8_t v1 = vshr_n_u8(v8, 4); - - // Convert to signed 8-bit integers - const int8x8_t vs_0 = vreinterpret_s8_u8(v0); - const int8x8_t vs_1 = vreinterpret_s8_u8(v1); - - // Subtract 8 from each byte - const int8x8_t vb_0 = vsub_s8(vs_0, vdup_n_s8(8)); - const int8x8_t vb_1 = vsub_s8(vs_1, vdup_n_s8(8)); - - // Interleave and combine - const int8x8_t vx_0 = vzip1_s8(vb_0, vb_1); - const int8x8_t vx_1 = vzip2_s8(vb_0, vb_1); - - const int8x16_t vq = vcombine_s8(vx_0, vx_1); - - // convert to 2x int16x8_t - const int16x8_t vi_0 = vmovl_s8(vget_low_s8 (vq)); - const int16x8_t vi_1 = vmovl_s8(vget_high_s8(vq)); - - // convert to 4x float32x4_t - const float32x4_t vf_0 = vcvtq_f32_s32(vmovl_s16(vget_low_s16 (vi_0))); - const float32x4_t vf_1 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vi_0))); - const float32x4_t vf_2 = vcvtq_f32_s32(vmovl_s16(vget_low_s16 (vi_1))); - const float32x4_t vf_3 = vcvtq_f32_s32(vmovl_s16(vget_high_s16(vi_1))); - - // Multiply by d - const float32x4_t r0 = vmulq_f32(vf_0, vd); - const float32x4_t r1 = vmulq_f32(vf_1, vd); - const float32x4_t r2 = vmulq_f32(vf_2, vd); - const float32x4_t r3 = vmulq_f32(vf_3, vd); - - // Store - vst1q_f32(y + i*QK4_0 + l + 0, r0); - vst1q_f32(y + i*QK4_0 + l + 4, r1); - vst1q_f32(y + i*QK4_0 + l + 8, r2); - vst1q_f32(y + i*QK4_0 + l + 12, r3); - } - } -#else - // scalar for (int i = 0; i < nb; i++) { const float d = x[i].d; - const uint8_t * restrict pp = x[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0x0F) - 8; + const int x1 = (x[i].qs[j] >> 4) - 8; - for (int l = 0; l < QK4_0; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0x0F; - const int8_t vi1 = vi >> 4; - - const float v0 = (vi0 - 8)*d; - const float v1 = (vi1 - 8)*d; - - //printf("d = %f, vi = %d, vi0 = %d, vi1 = %d, v0 = %f, v1 = %f\n", d, vi, vi0, vi1, v0, v1); - - y[i*QK4_0 + l + 0] = v0; - y[i*QK4_0 + l + 1] = v1; - - assert(!isnan(y[i*QK4_0 + l + 0])); - assert(!isnan(y[i*QK4_0 + l + 1])); + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; } } -#endif } -static void dequantize_row_q4_1(const void * restrict vx, float * restrict y, int k) { - assert(k % QK4_1 == 0); - const int nb = k / QK4_1; +static void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k) { + static const int qk = QK4_1; - const block_q4_1 * restrict x = vx; + assert(k % qk == 0); -#if defined(__AVX2__) - for (int i = 0; i < nb; i++) { - const __m256 d_v = _mm256_broadcast_ss(&x[i].d); - const __m256 d_m = _mm256_broadcast_ss(&x[i].m); + const int nb = k / qk; - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_1; l += 32) { - // Load 32x4-bit integers into 32x8-bit integers - __m256i vx8 = bytes_from_nibbles_32(pp+l/2); - - // Convert to 16-bit int - const __m256i vx16_lo = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 0)); - const __m256i vx16_hi = _mm256_cvtepi8_epi16(_mm256_extracti128_si256(vx8, 1)); - - // Convert to 32-bit int -> float 32 - const __m256 vf[4] = { - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_lo, 1))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 0))), - _mm256_cvtepi32_ps(_mm256_cvtepi16_epi32(_mm256_extracti128_si256(vx16_hi, 1))) - }; - - // Scale, add m and store - for (int j = 0; j < 4; j++) { - const __m256 result = _mm256_add_ps(_mm256_mul_ps(vf[j], d_v), d_m); - _mm256_storeu_ps(y + i * QK4_1 + l + j*8, result); - } - } - } -#elif defined(__ARM_NEON) - for (int i = 0; i < nb; i++) { - const float32x4_t vd = vdupq_n_f32(x[i].d); - const float32x4_t vm = vdupq_n_f32(x[i].m); - - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_1; l += 16) { - // Load 16x4-bit integers into 8x8-bit integers - const uint8x8_t v8 = vld1_u8(pp + l/2); - - // Expand 4-bit qs to 8-bit bytes - const uint8x8_t v0 = vand_u8(v8, vdup_n_u8(0x0F)); - const uint8x8_t v1 = vshr_n_u8(v8, 4); - - // Interleave and combine - const uint8x8_t vx_0 = vzip1_u8(v0, v1); - const uint8x8_t vx_1 = vzip2_u8(v0, v1); - - const uint8x16_t vq = vcombine_u8(vx_0, vx_1); - - // convert to 2x uint16x8_t - const uint16x8_t vi_0 = vmovl_u8(vget_low_u8 (vq)); - const uint16x8_t vi_1 = vmovl_u8(vget_high_u8(vq)); - - // convert to 4x float32x4_t - const float32x4_t vf_0 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_0))); - const float32x4_t vf_1 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_0))); - const float32x4_t vf_2 = vcvtq_f32_u32(vmovl_u16(vget_low_u16 (vi_1))); - const float32x4_t vf_3 = vcvtq_f32_u32(vmovl_u16(vget_high_u16(vi_1))); - - // multiply by d and add m - const float32x4_t r0 = vmlaq_f32(vm, vf_0, vd); - const float32x4_t r1 = vmlaq_f32(vm, vf_1, vd); - const float32x4_t r2 = vmlaq_f32(vm, vf_2, vd); - const float32x4_t r3 = vmlaq_f32(vm, vf_3, vd); - - // Store - vst1q_f32(y + i*QK4_1 + l + 0, r0); - vst1q_f32(y + i*QK4_1 + l + 4, r1); - vst1q_f32(y + i*QK4_1 + l + 8, r2); - vst1q_f32(y + i*QK4_1 + l + 12, r3); - } - } -#else for (int i = 0; i < nb; i++) { const float d = x[i].d; const float m = x[i].m; - const uint8_t * restrict pp = x[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int x0 = (x[i].qs[j] & 0x0F); + const int x1 = (x[i].qs[j] >> 4); - for (int l = 0; l < QK4_1; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0x0F; - const int8_t vi1 = vi >> 4; - - const float v0 = vi0*d + m; - const float v1 = vi1*d + m; - - y[i*QK4_1 + l + 0] = v0; - y[i*QK4_1 + l + 1] = v1; - - assert(!isnan(y[i*QK4_1 + l + 0])); - assert(!isnan(y[i*QK4_1 + l + 1])); - } - } -#endif -} - -static void dequantize_row_q4_2(const void * restrict vx, float * restrict y, int k) { - assert(k % QK4_2 == 0); - const int nb = k / QK4_2; - - const block_q4_2 * restrict x = vx; - - for (int i = 0; i < nb; i++) { - const float d = GGML_FP16_TO_FP32(x[i].d); - - const uint8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK4_2; l += 2) { - const uint8_t vi = pp[l/2]; - - const int8_t vi0 = vi & 0x0F; - const int8_t vi1 = vi >> 4; - - const float v0 = (vi0 - 8)*d; - const float v1 = (vi1 - 8)*d; - - y[i*QK4_2 + l + 0] = v0; - y[i*QK4_2 + l + 1] = v1; - - assert(!isnan(y[i*QK4_2 + l + 0])); - assert(!isnan(y[i*QK4_2 + l + 1])); + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; } } } -static void dequantize_row_q5_0(const void * restrict vx, float * restrict y, int k) { - assert(k % QK5_0 == 0); - const int nb = k / QK5_0; +static void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k) { + static const int qk = QK5_0; - const block_q5_0 * restrict x = vx; + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { const float d = GGML_FP16_TO_FP32(x[i].d); - const uint8_t * restrict pp = x[i].qs; - uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_0; l += 2) { - const uint8_t vi = pp[l/2]; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - // extract the 5-th bit from qh - const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; - const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; - const int8_t vi0 = (vi & 0x0F) | vh0; - const int8_t vi1 = (vi >> 4) | vh1; - - const float v0 = (vi0 - 16)*d; - const float v1 = (vi1 - 16)*d; - - y[i*QK5_0 + l + 0] = v0; - y[i*QK5_0 + l + 1] = v1; - - assert(!isnan(y[i*QK5_0 + l + 0])); - assert(!isnan(y[i*QK5_0 + l + 1])); + y[i*qk + j + 0 ] = x0*d; + y[i*qk + j + qk/2] = x1*d; } } } -static void dequantize_row_q5_1(const void * restrict vx, float * restrict y, int k) { - assert(k % QK5_1 == 0); - const int nb = k / QK5_1; +static void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k) { + static const int qk = QK5_1; - const block_q5_1 * restrict x = vx; + assert(k % qk == 0); + + const int nb = k / qk; for (int i = 0; i < nb; i++) { const float d = GGML_FP16_TO_FP32(x[i].d); const float m = GGML_FP16_TO_FP32(x[i].m); - const uint8_t * restrict pp = x[i].qs; - uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_1; l += 2) { - const uint8_t vi = pp[l/2]; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - // extract the 5-th bit from qh - const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; - const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + const int x0 = (x[i].qs[j] & 0x0F) | xh_0; + const int x1 = (x[i].qs[j] >> 4) | xh_1; - const uint8_t vi0 = (vi & 0x0F) | vh0; - const uint8_t vi1 = (vi >> 4) | vh1; - - const float v0 = vi0*d + m; - const float v1 = vi1*d + m; - - y[i*QK5_1 + l + 0] = v0; - y[i*QK5_1 + l + 1] = v1; - - assert(!isnan(y[i*QK5_1 + l + 0])); - assert(!isnan(y[i*QK5_1 + l + 1])); + y[i*qk + j + 0 ] = x0*d + m; + y[i*qk + j + qk/2] = x1*d + m; } } } static void dequantize_row_q8_0(const void * restrict vx, float * restrict y, int k) { - assert(k % QK8_0 == 0); - const int nb = k / QK8_0; + static const int qk = QK8_0; + + assert(k % qk == 0); + + const int nb = k / qk; const block_q8_0 * restrict x = vx; for (int i = 0; i < nb; i++) { const float d = x[i].d; - const int8_t * restrict pp = x[i].qs; - - for (int l = 0; l < QK8_0; ++l) { - y[i*QK8_0 + l] = pp[l]*d; + for (int j = 0; j < qk; ++j) { + y[i*qk + j] = x[i].qs[j]*d; } } } static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); -static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy); -static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { +const quantize_fns_t ggjt_v2_quantize_fns[GGML_TYPE_COUNT] = { [GGML_TYPE_Q4_0] = { - .dequantize_row_q = dequantize_row_q4_0, + .dequantize_row_q = (dequantize_row_q_t) dequantize_row_q4_0, .quantize_row_q = quantize_row_q4_0, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_0_reference, .quantize_row_q_dot = quantize_row_q8_0, @@ -2145,23 +1062,15 @@ static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { .vec_dot_type = GGML_TYPE_Q8_0, }, [GGML_TYPE_Q4_1] = { - .dequantize_row_q = dequantize_row_q4_1, + .dequantize_row_q = (dequantize_row_q_t)dequantize_row_q4_1, .quantize_row_q = quantize_row_q4_1, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_1_reference, .quantize_row_q_dot = quantize_row_q8_1, .vec_dot_q = ggml_vec_dot_q4_1_q8_1, .vec_dot_type = GGML_TYPE_Q8_1, }, - [GGML_TYPE_Q4_2] = { - .dequantize_row_q = dequantize_row_q4_2, - .quantize_row_q = quantize_row_q4_2, - .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_2_reference, - .quantize_row_q_dot = quantize_row_q8_0, - .vec_dot_q = ggml_vec_dot_q4_2_q8_0, - .vec_dot_type = GGML_TYPE_Q8_0, - }, [GGML_TYPE_Q5_0] = { - .dequantize_row_q = dequantize_row_q5_0, + .dequantize_row_q = (dequantize_row_q_t) dequantize_row_q5_0, .quantize_row_q = quantize_row_q5_0, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q5_0_reference, .quantize_row_q_dot = quantize_row_q8_0, @@ -2169,7 +1078,7 @@ static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { .vec_dot_type = GGML_TYPE_Q8_0, }, [GGML_TYPE_Q5_1] = { - .dequantize_row_q = dequantize_row_q5_1, + .dequantize_row_q = (dequantize_row_q_t) dequantize_row_q5_1, .quantize_row_q = quantize_row_q5_1, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q5_1_reference, .quantize_row_q_dot = quantize_row_q8_1, @@ -2194,6 +1103,8 @@ static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { }, }; +const quantize_fns_t *quantize_fns; + // For internal test use quantize_fns_t ggml_internal_get_quantize_fn(size_t i) { GGML_ASSERT(i < GGML_TYPE_COUNT); @@ -2794,9 +1705,10 @@ inline static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t } static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; + const int qk = QK8_0; + const int nb = n / qk; - assert(n % QK8_0 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); const block_q4_0 * restrict x = vx; @@ -2830,12 +1742,6 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); - // interleave - const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs); - const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs); - // load y const int8x16_t v1_0l = vld1q_s8(y0->qs); const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); @@ -2844,21 +1750,21 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * #if defined(__ARM_FEATURE_DOTPROD) // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h); + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0h)); - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); @@ -2904,31 +1810,24 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * // Compute combined scale for the block const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) ); - __m128i i32[2]; - for (int j = 0; j < 2; ++j) { - // Load 8 bytes, and unpack 4 bit fields into bytes, making 16 bytes - __m128i bx = bytes_from_nibbles_16(x[i].qs + 8*j); - __m128i by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16*j)); + const __m128i lowMask = _mm_set1_epi8(0xF); + const __m128i off = _mm_set1_epi8(8); - // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. - const __m128i off = _mm_set1_epi8( 8 ); - bx = _mm_sub_epi8( bx, off ); + const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs); - // Get absolute values of x vectors - const __m128i ax = _mm_sign_epi8(bx, bx); + __m128i bx = _mm_and_si128(lowMask, tmp); + __m128i by = _mm_loadu_si128((const __m128i *)y[i].qs); + bx = _mm_sub_epi8(bx, off); + const __m128i i32_0 = mul_sum_i8_pairs(bx, by); - // Sign the values of the y vectors - const __m128i sy = _mm_sign_epi8(by, bx); - - // Perform multiplication and create 16-bit values - const __m128i dot = _mm_maddubs_epi16(ax, sy); - - const __m128i ones = _mm_set1_epi16(1); - i32[j] = _mm_madd_epi16(ones, dot); - } + bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4)); + by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16)); + bx = _mm_sub_epi8(bx, off); + const __m128i i32_1 = mul_sum_i8_pairs(bx, by); // Convert int32_t to float - __m256 p = _mm256_cvtepi32_ps( _mm256_set_m128i( i32[0], i32[1] )); + __m256 p = _mm256_cvtepi32_ps(_mm256_set_m128i(i32_0, i32_1)); + // Apply the scale, and accumulate acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc); } @@ -2937,41 +1836,35 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * #else // scalar float sumf = 0.0; + for (int i = 0; i < nb; i++) { - const float d0 = x[i].d; - const float d1 = y[i].d; - - const uint8_t * restrict p0 = x[i].qs; - const int8_t * restrict p1 = y[i].qs; - int sumi = 0; - for (int j = 0; j < QK8_0/2; j++) { - const uint8_t v0 = p0[j]; - const int i0 = (int8_t) (v0 & 0x0F) - 8; - const int i1 = (int8_t) (v0 >> 4) - 8; + for (int j = 0; j < qk/2; ++j) { + const int v0 = (x[i].qs[j] & 0x0F) - 8; + const int v1 = (x[i].qs[j] >> 4) - 8; - const int i2 = p1[2*j + 0]; - const int i3 = p1[2*j + 1]; - - sumi += i0*i2 + i1*i3; + sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); } - sumf += d0*d1*sumi; + + sumf += (x[i].d*y[i].d)*sumi; } + *s = sumf; #endif } static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_1; + const int qk = QK8_1; + const int nb = n / qk; - assert(n % QK8_1 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); const block_q4_1 * restrict x = vx; const block_q8_1 * restrict y = vy; - // TODO: add AVX / WASM SIMD / etc + // TODO: add WASM SIMD #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); @@ -2984,7 +1877,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const block_q8_1 * restrict y0 = &y[i + 0]; const block_q8_1 * restrict y1 = &y[i + 1]; - summs += x0->m * (y0->s0 + y0->s1) + x1->m * (y1->s0 + y1->s1); + summs += x0->m * y0->s + x1->m * y1->s; const uint8x16_t m4b = vdupq_n_u8(0x0F); @@ -2997,12 +1890,6 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - // interleave - const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h); - const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h); - const int8x16_t v0_1lz = vzip1q_s8(v0_1l, v0_1h); - const int8x16_t v0_1hz = vzip2q_s8(v0_1l, v0_1h); - // load y const int8x16_t v1_0l = vld1q_s8(y0->qs); const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); @@ -3011,21 +1898,21 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * #if defined(__ARM_FEATURE_DOTPROD) // dot product into int32x4_t - const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h); - const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h); + const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h); + const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h); sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d); sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0h)); - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); @@ -3049,7 +1936,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * const float * d0 = &x[i].d; const float * d1 = &y[i].d; - summs += x[i].m * (y[i].s0 + y[i].s1); + summs += x[i].m * y[i].s; const __m256 d0v = _mm256_broadcast_ss( d0 ); const __m256 d1v = _mm256_broadcast_ss( d1 ); @@ -3071,77 +1958,86 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * #else // scalar float sumf = 0.0; + for (int i = 0; i < nb; i++) { - const float d0 = x[i].d; - const float m0 = x[i].m; - const float d1 = y[i].d; + int sumi = 0; - const uint8_t * restrict p0 = x[i].qs; - const int8_t * restrict p1 = y[i].qs; + for (int j = 0; j < qk/2; ++j) { + const int v0 = (x[i].qs[j] & 0x0F); + const int v1 = (x[i].qs[j] >> 4); - // TODO: this is very slow .. - for (int j = 0; j < QK8_1/2; j++) { - const uint8_t v0 = p0[j]; - - const float f0 = d0*(v0 & 0x0F) + m0; - const float f1 = d0*(v0 >> 4) + m0; - - const float f2 = d1*p1[2*j + 0]; - const float f3 = d1*p1[2*j + 1]; - - sumf += f0*f2 + f1*f3; + sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]); } + + sumf += (x[i].d*y[i].d)*sumi + x[i].m*y[i].s; } + *s = sumf; #endif } -static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; +static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { + const int qk = QK8_0; + const int nb = n / qk; - assert(n % QK8_0 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); - assert(QK8_0 == 2*QK4_2); + assert(qk == QK5_0); - const block_q4_2 * restrict x = vx; + const block_q5_0 * restrict x = vx; const block_q8_0 * restrict y = vy; #if defined(__ARM_NEON) float32x4_t sumv0 = vdupq_n_f32(0.0f); float32x4_t sumv1 = vdupq_n_f32(0.0f); - for (int i = 0; i < nb; i += 2) { - const block_q4_2 * restrict x0_0 = &x[2*(i + 0) + 0]; - const block_q4_2 * restrict x0_1 = &x[2*(i + 0) + 1]; - const block_q4_2 * restrict x1_0 = &x[2*(i + 1) + 0]; - const block_q4_2 * restrict x1_1 = &x[2*(i + 1) + 1]; + uint32_t qh0; + uint32_t qh1; - const block_q8_0 * restrict y0 = &y[i + 0]; + uint64_t tmp0[4]; + uint64_t tmp1[4]; + + for (int i = 0; i < nb; i += 2) { + const block_q5_0 * restrict x0 = &x[i]; + const block_q5_0 * restrict x1 = &x[i + 1]; + const block_q8_0 * restrict y0 = &y[i]; const block_q8_0 * restrict y1 = &y[i + 1]; - const uint8x16_t m4b = vdupq_n_u8(0x0F); - const int8x16_t s8b = vdupq_n_s8(0x8); + const uint8x16_t m4b = vdupq_n_u8(0x0F); - const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs)); - const uint8x16_t v0_1 = vcombine_u8(vld1_u8(x1_0->qs), vld1_u8(x1_1->qs)); + // extract the 5th bit via lookup table ((!b) << 4) + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); + + tmp0[0] = table_b2b_1[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_1[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_1[(qh0 >> 24) ]; + + tmp1[0] = table_b2b_1[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_1[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_1[(qh1 >> 24) ]; + + const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); + const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); + const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); + const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); // 4-bit -> 8-bit - const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); - const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); - const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); - const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); + int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - // sub 8 - const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b); - const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b); - const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b); - const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b); - - // interleave - const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs); - const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs); - const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs); + // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) + const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0); + const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0); + const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1); + const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1); // load y const int8x16_t v1_0l = vld1q_s8(y0->qs); @@ -3149,187 +2045,45 @@ static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * const int8x16_t v1_1l = vld1q_s8(y1->qs); const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); + const float x0d = GGML_FP16_TO_FP32(x0->d); + const float x1d = GGML_FP16_TO_FP32(x1->d); + #if defined(__ARM_FEATURE_DOTPROD) - sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), GGML_FP16_TO_FP32(x0_0->d)), - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), GGML_FP16_TO_FP32(x0_1->d))), y0->d); - - sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l)), GGML_FP16_TO_FP32(x1_0->d)), - vmulq_n_f32(vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1hz, v1_1h)), GGML_FP16_TO_FP32(x1_1->d))), y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); - const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l)); - const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l)); - const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h)); - const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h)); + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - sumv0 = vmlaq_n_f32(sumv0, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(pl0), GGML_FP16_TO_FP32(x0_0->d)), - vmulq_n_f32(vcvtq_f32_s32(ph0), GGML_FP16_TO_FP32(x0_1->d))), y0->d); - - sumv1 = vmlaq_n_f32(sumv1, vaddq_f32( - vmulq_n_f32(vcvtq_f32_s32(pl1), GGML_FP16_TO_FP32(x1_0->d)), - vmulq_n_f32(vcvtq_f32_s32(ph1), GGML_FP16_TO_FP32(x1_1->d))), y1->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d); #endif } *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1); -#elif defined(__AVX2__) - // Initialize accumulator with zeros - __m256 acc = _mm256_setzero_ps(); - - // Main loop - for (int i = 0; i < nb; i++) { - /* Compute combined scale for the block */ - const __m128 d0 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 0].d)); - const __m128 d1 = _mm_set1_ps(GGML_FP16_TO_FP32(x[2*i + 1].d)); - const __m256 d = _mm256_mul_ps(_mm256_set_m128(d1, d0), _mm256_broadcast_ss(&y[i].d)); - - __m128i bx0 = bytes_from_nibbles_16(x[2*i + 0].qs); - __m128i bx1 = bytes_from_nibbles_16(x[2*i + 1].qs); - __m256i bx = _mm256_set_m128i(bx1, bx0); - - // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval. - const __m256i off = _mm256_set1_epi8(8); - bx = _mm256_sub_epi8(bx, off); - - __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs); - - const __m256 q = mul_sum_i8_pairs_float(bx, by); - - /* Multiply q with scale and accumulate */ - acc = _mm256_fmadd_ps(d, q, acc); - } - - *s = hsum_float_8(acc); -#else - // scalar - float sumf = 0.0; - for (int i = 0; i < nb; i++) { - const uint8_t * restrict x0 = x[2*i + 0].qs; - const uint8_t * restrict x1 = x[2*i + 1].qs; - const int8_t * restrict y0 = y[i].qs; - - const float d0 = GGML_FP16_TO_FP32(x[2*i + 0].d); - const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d); - - int sumi_0 = 0; - int sumi_1 = 0; - - for (int j = 0; j < QK8_0/4; j++) { - const uint8_t v0 = x0[j]; - const uint8_t v1 = x1[j]; - - const int i0_0 = (int8_t) (v0 & 0x0F) - 8; - const int i1_0 = (int8_t) (v0 >> 4) - 8; - - const int i0_1 = (int8_t) (v1 & 0x0F) - 8; - const int i1_1 = (int8_t) (v1 >> 4) - 8; - - const int i2_0 = y0[2*j + 0]; - const int i3_0 = y0[2*j + 1]; - - const int i2_1 = y0[2*(j + QK8_0/4) + 0]; - const int i3_1 = y0[2*(j + QK8_0/4) + 1]; - - sumi_0 += i0_0*i2_0 + i1_0*i3_0; - sumi_1 += i0_1*i2_1 + i1_1*i3_1; - } - - sumf += (d0 * y[i].d) * sumi_0; - sumf += (d1 * y[i].d) * sumi_1; - } - *s = sumf; -#endif -} - -static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; - - assert(n % QK8_0 == 0); - assert(nb % 2 == 0); - assert(QK8_0 == QK5_0); - - const block_q5_0 * restrict x = vx; - const block_q8_0 * restrict y = vy; - -#if defined(__ARM_NEON) - float32x4_t sumv = vdupq_n_f32(0.0f); - - uint64_t tmp[4]; - - for (int i = 0; i < nb; ++i) { - const block_q5_0 * restrict x0 = &x[i]; - const block_q8_0 * restrict y0 = &y[i]; - - const uint8x16_t m4b = vdupq_n_u8(0x0F); - const int8x16_t s16b = vdupq_n_s8(0x10); - - // extract the 5th bit - uint32_t qh; - memcpy(&qh, x0->qh, sizeof(qh)); - - tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_u[(qh >> 24) ]; - - const int8x16_t qhl = vld1q_s8((const int8_t *)(tmp + 0)); - const int8x16_t qhh = vld1q_s8((const int8_t *)(tmp + 2)); - - const uint8x16_t v0 = vld1q_u8(x0->qs); - - // 4-bit -> 8-bit - const int8x16_t v0l = vreinterpretq_s8_u8(vandq_u8 (v0, m4b)); - const int8x16_t v0h = vreinterpretq_s8_u8(vshrq_n_u8(v0, 4)); - - // interleave - const int8x16_t v0lz = vzip1q_s8(v0l, v0h); - const int8x16_t v0hz = vzip2q_s8(v0l, v0h); - - // add high bit and sub 16 - const int8x16_t v0lf = vsubq_s8(vorrq_s8(v0lz, qhl), s16b); - const int8x16_t v0hf = vsubq_s8(vorrq_s8(v0hz, qhh), s16b); - - // load y - const int8x16_t v1l = vld1q_s8(y0->qs); - const int8x16_t v1h = vld1q_s8(y0->qs + 16); - - const float x0d = GGML_FP16_TO_FP32(x0->d); - -#if defined(__ARM_FEATURE_DOTPROD) - sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0lf, v1l), - vdotq_s32(vdupq_n_s32(0), v0hf, v1h))), x0d*y0->d); -#else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0lf), vget_low_s8 (v1l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0lf), vget_high_s8(v1l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0hf), vget_low_s8 (v1h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0hf), vget_high_s8(v1h)); - - const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); - const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); - - sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); -#endif - } - - *s = vaddvq_f32(sumv); #elif defined(__wasm_simd128__) v128_t sumv = wasm_f32x4_splat(0.0f); + uint32_t qh; uint64_t tmp[4]; + // TODO: check if unrolling this is better for (int i = 0; i < nb; ++i) { const block_q5_0 * restrict x0 = &x[i]; const block_q8_0 * restrict y0 = &y[i]; @@ -3338,13 +2092,12 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * const v128_t s16b = wasm_i8x16_splat(0x10); // extract the 5th bit - uint32_t qh; memcpy(&qh, x0->qh, sizeof(qh)); - tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_u[(qh >> 24) ]; + tmp[0] = table_b2b_1[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_1[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_1[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_1[(qh >> 24) ]; const v128_t qhl = wasm_v128_load(tmp + 0); const v128_t qhh = wasm_v128_load(tmp + 2); @@ -3355,13 +2108,9 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * const v128_t v0l = wasm_v128_and (v0, m4b); const v128_t v0h = wasm_u8x16_shr(v0, 4); - // interleave - const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23); - const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31); - - // add high bit and sub 16 - const v128_t v0lf = wasm_i8x16_sub(wasm_v128_or(v0lz, qhl), s16b); - const v128_t v0hf = wasm_i8x16_sub(wasm_v128_or(v0hz, qhh), s16b); + // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero) + const v128_t v0lf = wasm_i8x16_sub(v0l, qhl); + const v128_t v0hf = wasm_i8x16_sub(v0h, qhh); // load y const v128_t v1l = wasm_v128_load(y0->qs); @@ -3417,134 +2166,161 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * #else // scalar float sumf = 0.0; - for (int i = 0; i < nb; i++) { - const uint8_t * restrict x0 = x[i].qs; - const int8_t * restrict y0 = y[i].qs; + for (int i = 0; i < nb; i++) { uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - const float d = GGML_FP16_TO_FP32(x[i].d); + int sumi = 0; - int sxy = 0; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); - for (int j = 0; j < QK8_0/2; j++) { - const uint8_t v0 = x0[j]; + const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16; + const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16; - const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4; - const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4; - - const int x0_0 = ((v0 & 0x0F) | x0_0h) - 16; - const int x1_0 = ((v0 >> 4) | x1_0h) - 16; - - const int y0_0 = y0[2*j + 0]; - const int y1_0 = y0[2*j + 1]; - - sxy += x0_0*y0_0 + x1_0*y1_0; + sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); } - sumf += (d*sxy)*y[i].d; + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi; } + *s = sumf; #endif } static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_1; + const int qk = QK8_1; + const int nb = n / qk; - assert(n % QK8_1 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); - assert(QK8_1 == QK5_1); + assert(qk == QK5_1); const block_q5_1 * restrict x = vx; const block_q8_1 * restrict y = vy; #if defined(__ARM_NEON) - float32x4_t sumv = vdupq_n_f32(0.0f); + float32x4_t sumv0 = vdupq_n_f32(0.0f); + float32x4_t sumv1 = vdupq_n_f32(0.0f); - float summs = 0.0f; + float summs0 = 0.0f; + float summs1 = 0.0f; - uint64_t tmp[4]; + uint32_t qh0; + uint32_t qh1; - for (int i = 0; i < nb; ++i) { + uint64_t tmp0[4]; + uint64_t tmp1[4]; + + for (int i = 0; i < nb; i += 2) { const block_q5_1 * restrict x0 = &x[i]; + const block_q5_1 * restrict x1 = &x[i + 1]; const block_q8_1 * restrict y0 = &y[i]; + const block_q8_1 * restrict y1 = &y[i + 1]; - summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1); + const uint8x16_t m4b = vdupq_n_u8(0x0F); - // extract the 5th bit - uint32_t qh; - memcpy(&qh, x0->qh, sizeof(qh)); + summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s; + summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s; - tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_u[(qh >> 24) ]; + // extract the 5th bit via lookup table ((b) << 4) + memcpy(&qh0, x0->qh, sizeof(qh0)); + memcpy(&qh1, x1->qh, sizeof(qh1)); - const int8x16_t qhl = vld1q_s8((const int8_t *)(tmp + 0)); - const int8x16_t qhh = vld1q_s8((const int8_t *)(tmp + 2)); + tmp0[0] = table_b2b_0[(qh0 >> 0) & 0xFF]; + tmp0[1] = table_b2b_0[(qh0 >> 8) & 0xFF]; + tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF]; + tmp0[3] = table_b2b_0[(qh0 >> 24) ]; - const uint8x16_t v0 = vld1q_u8(x0->qs); + tmp1[0] = table_b2b_0[(qh1 >> 0) & 0xFF]; + tmp1[1] = table_b2b_0[(qh1 >> 8) & 0xFF]; + tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF]; + tmp1[3] = table_b2b_0[(qh1 >> 24) ]; + + const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0)); + const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2)); + const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0)); + const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2)); + + const uint8x16_t v0_0 = vld1q_u8(x0->qs); + const uint8x16_t v0_1 = vld1q_u8(x1->qs); // 4-bit -> 8-bit - const int8x16_t v0l = vreinterpretq_s8_u8(vandq_u8 (v0, vdupq_n_u8(0x0F))); - const int8x16_t v0h = vreinterpretq_s8_u8(vshrq_n_u8(v0, 4)); + const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8 (v0_0, m4b)); + const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4)); + const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8 (v0_1, m4b)); + const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4)); - // interleave - const int8x16_t v0lz = vzip1q_s8(v0l, v0h); - const int8x16_t v0hz = vzip2q_s8(v0l, v0h); - - // add - const int8x16_t v0lf = vorrq_s8(v0lz, qhl); - const int8x16_t v0hf = vorrq_s8(v0hz, qhh); + // add high bit + const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0); + const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0); + const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1); + const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1); // load y - const int8x16_t v1l = vld1q_s8(y0->qs); - const int8x16_t v1h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_0l = vld1q_s8(y0->qs); + const int8x16_t v1_0h = vld1q_s8(y0->qs + 16); + const int8x16_t v1_1l = vld1q_s8(y1->qs); + const int8x16_t v1_1h = vld1q_s8(y1->qs + 16); const float x0d = GGML_FP16_TO_FP32(x0->d); + const float x1d = GGML_FP16_TO_FP32(x1->d); #if defined(__ARM_FEATURE_DOTPROD) - sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32( - vdotq_s32(vdupq_n_s32(0), v0lf, v1l), - vdotq_s32(vdupq_n_s32(0), v0hf, v1h))), x0d*y0->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l), + vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32( + vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l), + vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d); #else - const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0lf), vget_low_s8 (v1l)); - const int16x8_t pl0h = vmull_s8(vget_high_s8(v0lf), vget_high_s8(v1l)); - const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0hf), vget_low_s8 (v1h)); - const int16x8_t ph0h = vmull_s8(vget_high_s8(v0hf), vget_high_s8(v1h)); + const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l)); + const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l)); + const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hf), vget_low_s8 (v1_0h)); + const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hf), vget_high_s8(v1_0h)); + + const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lf), vget_low_s8 (v1_1l)); + const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lf), vget_high_s8(v1_1l)); + const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hf), vget_low_s8 (v1_1h)); + const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hf), vget_high_s8(v1_1h)); const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h)); const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h)); + const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h)); + const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h)); - sumv = vmlaq_n_f32(sumv, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); + sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d); + sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d); #endif } - *s = vaddvq_f32(sumv) + summs; + *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1; #elif defined(__wasm_simd128__) v128_t sumv = wasm_f32x4_splat(0.0f); float summs = 0.0f; + uint32_t qh; uint64_t tmp[4]; + // TODO: check if unrolling this is better for (int i = 0; i < nb; ++i) { const block_q5_1 * restrict x0 = &x[i]; const block_q8_1 * restrict y0 = &y[i]; - summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1); + summs += GGML_FP16_TO_FP32(x0->m) * y0->s; const v128_t m4b = wasm_i8x16_splat(0x0F); // extract the 5th bit - uint32_t qh; memcpy(&qh, x0->qh, sizeof(qh)); - tmp[0] = table_b2b_u[(qh >> 0) & 0xFF]; - tmp[1] = table_b2b_u[(qh >> 8) & 0xFF]; - tmp[2] = table_b2b_u[(qh >> 16) & 0xFF]; - tmp[3] = table_b2b_u[(qh >> 24) ]; + tmp[0] = table_b2b_0[(qh >> 0) & 0xFF]; + tmp[1] = table_b2b_0[(qh >> 8) & 0xFF]; + tmp[2] = table_b2b_0[(qh >> 16) & 0xFF]; + tmp[3] = table_b2b_0[(qh >> 24) ]; const v128_t qhl = wasm_v128_load(tmp + 0); const v128_t qhh = wasm_v128_load(tmp + 2); @@ -3557,13 +2333,9 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * static bool x = true; - // interleave - const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23); - const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31); - // add high bit - const v128_t v0lf = wasm_v128_or(v0lz, qhl); - const v128_t v0hf = wasm_v128_or(v0hz, qhh); + const v128_t v0lf = wasm_v128_or(v0l, qhl); + const v128_t v0hf = wasm_v128_or(v0h, qhh); // load y const v128_t v1l = wasm_v128_load(y0->qs); @@ -3596,13 +2368,14 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * #elif defined(__AVX2__) // Initialize accumulator with zeros __m256 acc = _mm256_setzero_ps(); + float summs = 0.0f; // Main loop for (int i = 0; i < nb; i++) { const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)); - summs += GGML_FP16_TO_FP32(x[i].m) * (y[i].s0 + y[i].s1); + summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s; __m256i bx = bytes_from_nibbles_32(x[i].qs); __m256i bxhi = bytes_from_bits_32(x[i].qh); @@ -3619,36 +2392,26 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * *s = hsum_float_8(acc) + summs; #else + // scalar float sumf = 0.0; for (int i = 0; i < nb; i++) { - const uint8_t * restrict x0 = x[i].qs; - const int8_t * restrict y0 = y[i].qs; - uint32_t qh; memcpy(&qh, x[i].qh, sizeof(qh)); - const float d = GGML_FP16_TO_FP32(x[i].d); - const float m = GGML_FP16_TO_FP32(x[i].m); + int sumi = 0; - int sxy = 0; + for (int j = 0; j < qk/2; ++j) { + const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; + const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; - for (int j = 0; j < QK8_1/2; j++) { - const uint8_t v0 = x0[j]; + const int32_t x0 = (x[i].qs[j] & 0xF) | xh_0; + const int32_t x1 = (x[i].qs[j] >> 4) | xh_1; - const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4; - const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4; - - const int x0_0 = (v0 & 0x0F) | x0_0h; - const int x1_0 = (v0 >> 4) | x1_0h; - - const int y0_0 = y0[2*j + 0]; - const int y1_0 = y0[2*j + 1]; - - sxy += x0_0*y0_0 + x1_0*y1_0; + sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]); } - sumf += (d*sxy)*y[i].d + m*(y[i].s0 + y[i].s1); + sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s; } *s = sumf; @@ -3656,11 +2419,11 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * } static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) { - const int nb = n / QK8_0; + const int qk = QK8_0; + const int nb = n / qk; - assert(n % QK8_0 == 0); + assert(n % qk == 0); assert(nb % 2 == 0); - assert(QK8_0 == QK8_0); const block_q8_0 * restrict x = vx; const block_q8_0 * restrict y = vy; @@ -3740,16 +2503,10 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * float sumf = 0.0; for (int i = 0; i < nb; i++) { - const int8_t * restrict x0 = x[i].qs; - const int8_t * restrict y0 = y[i].qs; - int sumi = 0; - for (int j = 0; j < QK8_0; j++) { - const int v0 = x0[j]; - const int v1 = y0[j]; - - sumi += v0*v1; + for (int j = 0; j < qk; j++) { + sumi += x[i].qs[j]*y[i].qs[j]; } sumf += (x[i].d*y[i].d)*sumi; @@ -3883,20 +2640,6 @@ inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; } inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; } -static const float GELU_COEF_A = 0.044715f; -static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f; - -inline static float ggml_gelu_f32(float x) { - return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x))); -} - -inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { - const uint16_t * i16 = (const uint16_t *) x; - for (int i = 0; i < n; ++i) { - y[i] = table_gelu_f16[i16[i]]; - } -} - #ifdef GGML_GELU_FP16 inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) { uint16_t t; @@ -3914,11 +2657,6 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) { } #endif -// Sigmoid Linear Unit (SiLU) function -inline static float ggml_silu_f32(float x) { - return x/(1.0f + expf(-x)); -} - inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) { const uint16_t * i16 = (const uint16_t *) x; for (int i = 0; i < n; ++i) { @@ -4008,12 +2746,11 @@ inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x // data types // -static const int GGML_BLCK_SIZE[GGML_TYPE_COUNT] = { +const int ggjt_v2_blck_size[GGML_TYPE_COUNT] = { [GGML_TYPE_F32] = 1, [GGML_TYPE_F16] = 1, [GGML_TYPE_Q4_0] = QK4_0, [GGML_TYPE_Q4_1] = QK4_1, - [GGML_TYPE_Q4_2] = QK4_2, [GGML_TYPE_Q5_0] = QK5_0, [GGML_TYPE_Q5_1] = QK5_1, [GGML_TYPE_Q8_0] = QK8_0, @@ -4022,14 +2759,14 @@ static const int GGML_BLCK_SIZE[GGML_TYPE_COUNT] = { [GGML_TYPE_I16] = 1, [GGML_TYPE_I32] = 1, }; +const int *GGML_BLCK_SIZE; static_assert(GGML_TYPE_COUNT == 13, "GGML_BLCK_SIZE is outdated"); -static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = { +const size_t ggjt_v2_type_size[GGML_TYPE_COUNT] = { [GGML_TYPE_F32] = sizeof(float), [GGML_TYPE_F16] = sizeof(ggml_fp16_t), [GGML_TYPE_Q4_0] = sizeof(block_q4_0), [GGML_TYPE_Q4_1] = sizeof(block_q4_1), - [GGML_TYPE_Q4_2] = sizeof(block_q4_2), [GGML_TYPE_Q5_0] = sizeof(block_q5_0), [GGML_TYPE_Q5_1] = sizeof(block_q5_1), [GGML_TYPE_Q8_0] = sizeof(block_q8_0), @@ -4038,15 +2775,14 @@ static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = { [GGML_TYPE_I16] = sizeof(int16_t), [GGML_TYPE_I32] = sizeof(int32_t), }; +const size_t *GGML_TYPE_SIZE; static_assert(GGML_TYPE_COUNT == 13, "GGML_TYPE_SIZE is outdated"); - -static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = { +const char *const ggjt_v2_type_name[GGML_TYPE_COUNT] = { [GGML_TYPE_F32] = "f32", [GGML_TYPE_F16] = "f16", [GGML_TYPE_Q4_0] = "q4_0", [GGML_TYPE_Q4_1] = "q4_1", - [GGML_TYPE_Q4_2] = "q4_2", [GGML_TYPE_Q5_0] = "q5_0", [GGML_TYPE_Q5_1] = "q5_1", [GGML_TYPE_Q8_0] = "q8_0", @@ -4055,14 +2791,14 @@ static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = { [GGML_TYPE_I16] = "i16", [GGML_TYPE_I32] = "i32", }; +const char *const *GGML_TYPE_NAME; static_assert(GGML_TYPE_COUNT == 13, "GGML_TYPE_NAME is outdated"); -static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = { +const bool ggjt_v2_is_quantized[GGML_TYPE_COUNT] = { [GGML_TYPE_F32] = false, [GGML_TYPE_F16] = false, [GGML_TYPE_Q4_0] = true, [GGML_TYPE_Q4_1] = true, - [GGML_TYPE_Q4_2] = true, [GGML_TYPE_Q5_0] = true, [GGML_TYPE_Q5_1] = true, [GGML_TYPE_Q8_0] = true, @@ -4071,8 +2807,17 @@ static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = { [GGML_TYPE_I16] = false, [GGML_TYPE_I32] = false, }; +const bool *GGML_IS_QUANTIZED; static_assert(GGML_TYPE_COUNT == 13, "GGML_IS_QUANTIZED is outdated"); +void ggjt_v2(void) { + GGML_BLCK_SIZE = ggjt_v2_blck_size; + GGML_TYPE_SIZE = ggjt_v2_type_size; + GGML_TYPE_NAME = ggjt_v2_type_name; + GGML_IS_QUANTIZED = ggjt_v2_is_quantized; + quantize_fns = ggjt_v2_quantize_fns; +} + static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { "NONE", @@ -4430,28 +3175,11 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { static bool is_first_call = true; + assert(quantize_fns); // you must call ggjt_v2() or ggjt_v1() first + if (is_first_call) { - // initialize time system (required on Windows) ggml_time_init(); - - // initialize GELU, SILU and EXP F32 tables - { - const uint64_t t_start = ggml_time_us(); UNUSED(t_start); - - ggml_fp16_t ii; - for (int i = 0; i < (1 << 16); ++i) { - uint16_t ui = i; - memcpy(&ii, &ui, sizeof(ii)); - const float f = table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(ii); - table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f)); - table_silu_f16[i] = GGML_FP32_TO_FP16(ggml_silu_f32(f)); - table_exp_f16[i] = GGML_FP32_TO_FP16(expf(f)); - } - - const uint64_t t_end = ggml_time_us(); UNUSED(t_end); - - GGML_PRINT_DEBUG("%s: GELU, SILU and EXP tables initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f); - } + ggml_fp16_init(); // initialize g_state { @@ -13068,15 +11796,15 @@ size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * assert(k % QK4_0 == 0); const int nb = k / QK4_0; - for (int j = 0; j < n; j += k) { - block_q4_0 * restrict y = (block_q4_0 *)dst + j/QK4_0; + for (int b = 0; b < n; b += k) { + block_q4_0 * restrict y = (block_q4_0 *) dst + b/QK4_0; - quantize_row_q4_0_reference(src + j, y, k); + quantize_row_q4_0_reference(src + b, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK4_0; l += 2) { - const uint8_t vi0 = y[i].qs[l/2] & 0x0F; - const uint8_t vi1 = y[i].qs[l/2] >> 4; + for (int j = 0; j < QK4_0; j += 2) { + const uint8_t vi0 = y[i].qs[j/2] & 0x0F; + const uint8_t vi1 = y[i].qs[j/2] >> 4; hist[vi0]++; hist[vi1]++; @@ -13091,15 +11819,15 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * assert(k % QK4_1 == 0); const int nb = k / QK4_1; - for (int j = 0; j < n; j += k) { - block_q4_1 * restrict y = (block_q4_1 *)dst + j/QK4_1; + for (int b = 0; b < n; b += k) { + block_q4_1 * restrict y = (block_q4_1 *) dst + b/QK4_1; - quantize_row_q4_1_reference(src + j, y, k); + quantize_row_q4_1_reference(src + b, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK4_1; l += 2) { - const uint8_t vi0 = y[i].qs[l/2] & 0x0F; - const uint8_t vi1 = y[i].qs[l/2] >> 4; + for (int j = 0; j < QK4_1; j += 2) { + const uint8_t vi0 = y[i].qs[j/2] & 0x0F; + const uint8_t vi1 = y[i].qs[j/2] >> 4; hist[vi0]++; hist[vi1]++; @@ -13110,49 +11838,26 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * return (n/QK4_1*sizeof(block_q4_1)); } -size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist) { - assert(k % QK4_2 == 0); - const int nb = k / QK4_2; - - for (int j = 0; j < n; j += k) { - block_q4_2 * restrict y = (block_q4_2 *)dst + j/QK4_2; - - quantize_row_q4_2_reference(src + j, y, k); - - for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK4_2; l += 2) { - const uint8_t vi0 = y[i].qs[l/2] & 0x0F; - const uint8_t vi1 = y[i].qs[l/2] >> 4; - - hist[vi0]++; - hist[vi1]++; - } - } - } - - return (n/QK4_2*sizeof(block_q4_2)); -} - size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) { assert(k % QK5_0 == 0); const int nb = k / QK5_0; - for (int j = 0; j < n; j += k) { - block_q5_0 * restrict y = (block_q5_0 *)dst + j/QK5_0; + for (int b = 0; b < n; b += k) { + block_q5_0 * restrict y = (block_q5_0 *)dst + b/QK5_0; - quantize_row_q5_0_reference(src + j, y, k); + quantize_row_q5_0_reference(src + b, y, k); for (int i = 0; i < nb; i++) { uint32_t qh; memcpy(&qh, &y[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_0; l += 2) { - const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; - const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + for (int j = 0; j < QK5_0; j += 2) { + const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12)); // cast to 16 bins - const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2; - const uint8_t vi1 = ((y[i].qs[l/2] >> 4) | vh1) / 2; + const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2; + const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2; hist[vi0]++; hist[vi1]++; @@ -13167,22 +11872,22 @@ size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * assert(k % QK5_1 == 0); const int nb = k / QK5_1; - for (int j = 0; j < n; j += k) { - block_q5_1 * restrict y = (block_q5_1 *)dst + j/QK5_1; + for (int b = 0; b < n; b += k) { + block_q5_1 * restrict y = (block_q5_1 *)dst + b/QK5_1; - quantize_row_q5_1_reference(src + j, y, k); + quantize_row_q5_1_reference(src + b, y, k); for (int i = 0; i < nb; i++) { uint32_t qh; memcpy(&qh, &y[i].qh, sizeof(qh)); - for (int l = 0; l < QK5_1; l += 2) { - const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4; - const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4; + for (int j = 0; j < QK5_1; j += 2) { + const uint8_t vh0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; + const uint8_t vh1 = ((qh & (1u << (j + 16))) >> (j + 12)); // cast to 16 bins - const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2; - const uint8_t vi1 = ((y[i].qs[l/2] >> 4) | vh1) / 2; + const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2; + const uint8_t vi1 = ((y[i].qs[j/2] >> 4) | vh1) / 2; hist[vi0]++; hist[vi1]++; @@ -13197,14 +11902,14 @@ size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * assert(k % QK8_0 == 0); const int nb = k / QK8_0; - for (int j = 0; j < n; j += k) { - block_q8_0 * restrict y = (block_q8_0 *)dst + j/QK8_0; + for (int b = 0; b < n; b += k) { + block_q8_0 * restrict y = (block_q8_0 *)dst + b/QK8_0; - quantize_row_q8_0_reference(src + j, y, k); + quantize_row_q8_0_reference(src + b, y, k); for (int i = 0; i < nb; i++) { - for (int l = 0; l < QK8_0; ++l) { - const int8_t vi = y[i].qs[l]; + for (int j = 0; j < QK8_0; ++j) { + const int8_t vi = y[i].qs[j]; hist[vi/16 + 8]++; } @@ -13229,12 +11934,6 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i block_q4_1 * block = (block_q4_1*)dst + start / QK4_1; result = ggml_quantize_q4_1(src + start, block, n, n, hist); } break; - case GGML_TYPE_Q4_2: - { - GGML_ASSERT(start % QK4_2 == 0); - block_q4_2 * block = (block_q4_2*)dst + start / QK4_2; - result = ggml_quantize_q4_2(src + start, block, n, n, hist); - } break; case GGML_TYPE_Q5_0: { GGML_ASSERT(start % QK5_0 == 0); diff --git a/third_party/ggml/ggml.h b/third_party/ggml/ggml.h index 20b661383..98ae577e1 100644 --- a/third_party/ggml/ggml.h +++ b/third_party/ggml/ggml.h @@ -205,20 +205,6 @@ COSMOPOLITAN_C_START_ } \ } while (0) -#ifdef __ARM_NEON - // we use the built-in 16-bit float type - typedef __fp16 ggml_fp16_t; -#else - typedef uint16_t ggml_fp16_t; -#endif - - // convert FP16 <-> FP32 - GGML_API float ggml_fp16_to_fp32(ggml_fp16_t x); - GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x); - - GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, size_t n); - GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, size_t n); - struct ggml_object; struct ggml_context; @@ -385,6 +371,13 @@ COSMOPOLITAN_C_START_ bool no_alloc; // don't allocate memory for the tensor data }; + // + // compatibilty + // + + GGML_API void ggjt_v2(void); + GGML_API void ggjt_v1(void); + // misc GGML_API void ggml_time_init(void); // call this once at the beginning of the program @@ -926,6 +919,12 @@ COSMOPOLITAN_C_START_ quantize_fns_t ggml_internal_get_quantize_fn(size_t i); + extern const int *GGML_BLCK_SIZE; + extern const size_t *GGML_TYPE_SIZE; + extern const bool *GGML_IS_QUANTIZED; + extern const char *const *GGML_TYPE_NAME; + extern const quantize_fns_t *quantize_fns; + COSMOPOLITAN_C_END_ #endif /* !(__ASSEMBLER__ + __LINKER__ + 0) */ #endif /* COSMOPOLITAN_THIRD_PARTY_LLAMA_CPP_GGML_H_ */ diff --git a/third_party/ggml/ggml.mk b/third_party/ggml/ggml.mk index ad9251aac..df6b6450f 100644 --- a/third_party/ggml/ggml.mk +++ b/third_party/ggml/ggml.mk @@ -10,12 +10,35 @@ PKGS += THIRD_PARTY_GGML THIRD_PARTY_GGML_ARTIFACTS += THIRD_PARTY_GGML_A THIRD_PARTY_GGML = $(THIRD_PARTY_GGML_A_DEPS) $(THIRD_PARTY_GGML_A) THIRD_PARTY_GGML_A = o/$(MODE)/third_party/ggml/ggml.a -THIRD_PARTY_GGML_A_HDRS = third_party/ggml/ggml.h -THIRD_PARTY_GGML_A_SRCS = third_party/ggml/ggml.c THIRD_PARTY_GGML_A_OBJS = $(THIRD_PARTY_GGML_A_SRCS:%.c=o/$(MODE)/%.o) THIRD_PARTY_GGML_A_FILES = $(THIRD_PARTY_GGML_A_SRCS) $(THIRD_PARTY_GGML_A_HDRS) THIRD_PARTY_GGML_A_CHECKS = $(THIRD_PARTY_GGML_A).pkg $(THIRD_PARTY_GGML_A_HDRS:%=o/$(MODE)/%.ok) +THIRD_PARTY_GGML_A_HDRS = \ + third_party/ggml/fp16.h \ + third_party/ggml/ggml.h \ + third_party/ggml/ggjt.v1.q4_0.h \ + third_party/ggml/ggjt.v1.q4_1.h \ + third_party/ggml/ggjt.v1.q4_2.h \ + third_party/ggml/ggjt.v1.q5_0.h \ + third_party/ggml/ggjt.v1.q5_1.h \ + third_party/ggml/ggjt.v1.q8_0.h \ + third_party/ggml/ggjt.v1.q8_1.h \ + third_party/ggml/fp16.internal.h \ + third_party/ggml/ggjt.v1.internal.h + +THIRD_PARTY_GGML_A_SRCS = \ + third_party/ggml/fp16.c \ + third_party/ggml/ggml.c \ + third_party/ggml/ggjt.v1.c \ + third_party/ggml/ggjt.v1.q4_0.c \ + third_party/ggml/ggjt.v1.q4_1.c \ + third_party/ggml/ggjt.v1.q4_2.c \ + third_party/ggml/ggjt.v1.q5_0.c \ + third_party/ggml/ggjt.v1.q5_1.c \ + third_party/ggml/ggjt.v1.q8_0.c \ + third_party/ggml/ggjt.v1.q8_1.c + THIRD_PARTY_GGML_A_DIRECTDEPS = \ LIBC_CALLS \ LIBC_INTRIN \ diff --git a/third_party/ggml/llama.cc b/third_party/ggml/llama.cc index 268deaf5f..f104bf962 100644 --- a/third_party/ggml/llama.cc +++ b/third_party/ggml/llama.cc @@ -30,6 +30,8 @@ #include "libc/assert.h" #include "libc/intrin/bits.h" #include "libc/macros.internal.h" +#include "libc/stdio/stdio.h" +#include "third_party/ggml/fp16.h" #include "third_party/ggml/ggml.h" #include "third_party/ggml/llama_util.h" #include "third_party/libcxx/algorithm" @@ -431,7 +433,8 @@ struct llama_load_tensors_map { enum llama_file_version { LLAMA_FILE_VERSION_GGML, LLAMA_FILE_VERSION_GGMF_V1, // added version field and scores in vocab - LLAMA_FILE_VERSION_GGJT_V1, // added padding + LLAMA_FILE_VERSION_GGJT_V1, // adopted unified aligned mappable layout + LLAMA_FILE_VERSION_GGJT_V2, // changed quantization format }; struct llama_file_loader { @@ -458,10 +461,16 @@ struct llama_file_loader { if (magic == READ32BE("ggml") && version == 0) { file_version = LLAMA_FILE_VERSION_GGML; + ggjt_v1(); } else if (magic == READ32BE("ggmf") && version == 1) { file_version = LLAMA_FILE_VERSION_GGMF_V1; + ggjt_v1(); } else if (magic == READ32BE("ggjt") && version == 1) { file_version = LLAMA_FILE_VERSION_GGJT_V1; + ggjt_v1(); + } else if (magic == READ32BE("ggjt") && version == 2) { + file_version = LLAMA_FILE_VERSION_GGJT_V2; + ggjt_v2(); } else { Die("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?", magic, version); @@ -870,8 +879,9 @@ bool llama_mlock_supported() { static const char *llama_file_version_name(llama_file_version version) { switch (version) { case LLAMA_FILE_VERSION_GGML: return "'ggml' (old version with low tokenizer quality and no mmap support)"; - case LLAMA_FILE_VERSION_GGMF_V1: return "ggmf v1 (old version with no mmap support)"; - case LLAMA_FILE_VERSION_GGJT_V1: return "ggjt v1 (latest)"; + case LLAMA_FILE_VERSION_GGMF_V1: return "ggmf v1 (pre #613 with sharded files and no mmap support)"; + case LLAMA_FILE_VERSION_GGJT_V1: return "ggjt v1 (pre #1405)"; + case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (latest)"; default: LLAMA_ASSERT(false); } } @@ -951,6 +961,19 @@ static void llama_model_load_internal( fprintf(stderr, "%s: model size = %s\n", __func__, llama_model_type_name(model.type)); } + // check for consistency between ftype and version. for example, + // Q4_2 was removed when GGJT v2 was introduced, so we reject it + // unless the file is using an earlier version number. + if (((unsigned)hparams.ftype >= GGML_TYPE_COUNT || + !GGML_BLCK_SIZE[hparams.ftype] || + !GGML_TYPE_SIZE[hparams.ftype] || + !GGML_TYPE_NAME[hparams.ftype])) { + fprintf(stderr, "%s: error: '%s' isn't specified by '%s'\n", + __func__, llama_ftype_name(hparams.ftype), + llama_file_version_name(file_version)); + exit(1); + } + if (vocab_only) { return; } diff --git a/tool/emacs/cosmo-cpp-constants.el b/tool/emacs/cosmo-cpp-constants.el index 73bf14899..ff32dac86 100644 --- a/tool/emacs/cosmo-cpp-constants.el +++ b/tool/emacs/cosmo-cpp-constants.el @@ -47,6 +47,8 @@ "__AES__" "__AVX__" "__AVX2__" + "__AVX512F__" + "__AVXVNNI__" "__ABM__" "__BMI__" "__BMI2__"