vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : reuse quantum structs across backends (#5943)

Browse source 
* ggml : reuse quant blocks across backends

ggml-ci

* ggml : define helper constants only for CUDA and SYCL

ggml-ci

* ggml : define helper quantum constants for SYCL

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-03-12 14:27:20 +02:00 • committed by GitHub
parent 184215e783
commit 8030da7afe
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
7 changed files with 470 additions and 881 deletions
Download patch file Download diff file Expand all files Collapse all files

78
ggml-quants.c
View file

@ -1,3 +1,6 @@
#define GGML_COMMON_IMPL_C
#include "ggml-common.h"
#include "ggml-quants.h"
#include "ggml-impl.h"
@ -951,7 +954,7 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
int sum = 0;
@ -966,7 +969,7 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict
sum += y[i].qs[QK8_1/2 + j];
}
y[i].s = sum*d;
y[i].s = GGML_FP32_TO_FP16(sum*d);
}
}
@ -994,7 +997,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
int32x4_t accv = vdupq_n_s32(0);
@ -1010,7 +1013,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
accv = vaddq_s32(accv, vi);
}
y[i].s = d * vaddvq_s32(accv);
y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
}
#elif defined(__wasm_simd128__)
for (int i = 0; i < nb; i++) {
@ -1033,7 +1036,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
v128_t accv = wasm_i32x4_splat(0);
@ -1049,10 +1052,11 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
accv = wasm_i32x4_add(accv, vi);
}
y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) +
wasm_i32x4_extract_lane(accv, 1) +
wasm_i32x4_extract_lane(accv, 2) +
wasm_i32x4_extract_lane(accv, 3));
y[i].s = GGML_FP32_TO_FP16(
d * (wasm_i32x4_extract_lane(accv, 0) +
wasm_i32x4_extract_lane(accv, 1) +
wasm_i32x4_extract_lane(accv, 2) +
wasm_i32x4_extract_lane(accv, 3)));
}
#elif defined(__AVX2__) || defined(__AVX__)
for (int i = 0; i < nb; i++) {
@ -1077,7 +1081,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
// Quantize these floats
const float d = maxScalar / 127.f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
const __m256 mul = _mm256_set1_ps( id );
@ -1101,7 +1105,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
#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].s = GGML_FP32_TO_FP16(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
@ -1131,7 +1135,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
// 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].s = d * hsum_i32_4(_mm_add_epi32(s0, s1));
y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
// Convert int32 to int16
ni0 = _mm_packs_epi32( ni0, ni1 );
@ -1162,7 +1166,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
@ -1179,7 +1183,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
// set y[i].s
int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
y[i].s = sum*d;
y[i].s = GGML_FP32_TO_FP16(sum*d);
}
#else
GGML_UNUSED(nb);
@ -4019,10 +4023,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const block_q8_1 * restrict b_y0 = &vy0[i];
const block_q8_1 * restrict b_y1 = &vy1[i];
float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * b_y0->s,
GGML_FP16_TO_FP32(b_x1->m) * b_y0->s,
GGML_FP16_TO_FP32(b_x0->m) * b_y1->s,
GGML_FP16_TO_FP32(b_x1->m) * b_y1->s};
float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)};
summs0 += summs_t;
const uint8x16_t m4b = vdupq_n_u8(0x0F);
@ -4087,7 +4091,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const block_q8_1 * restrict y0 = &y[i + 0];
const block_q8_1 * restrict y1 = &y[i + 1];
summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s;
summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
const uint8x16_t m4b = vdupq_n_u8(0x0F);
@ -4110,8 +4114,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
@ -4124,9 +4128,9 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
// Main loop
for (int i = 0; i < nb; ++i) {
const float d0 = GGML_FP16_TO_FP32(x[i].d);
const float d1 = y[i].d;
const float d1 = GGML_FP16_TO_FP32(y[i].d);
summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
const __m256 d0v = _mm256_set1_ps( d0 );
const __m256 d1v = _mm256_set1_ps( d1 );
@ -4178,7 +4182,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
}
*s = sumf;
@ -4196,7 +4200,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
}
sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
}
*s = sumf;
@ -4532,8 +4536,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const uint8x16_t m4b = vdupq_n_u8(0x0F);
summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s;
summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s;
summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
// extract the 5th bit via lookup table ((b) << 4)
memcpy(&qh0, x0->qh, sizeof(qh0));
@ -4577,10 +4581,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d);
ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d);
ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
@ -4597,7 +4601,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const block_q5_1 * restrict x0 = &x[i];
const block_q8_1 * restrict y0 = &y[i];
summs += GGML_FP16_TO_FP32(x0->m) * y0->s;
summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
const v128_t m4b = wasm_i8x16_splat(0x0F);
@ -4644,7 +4648,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d)));
wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
}
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@ -4659,14 +4663,14 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
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].s;
summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
__m256i qx = 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));
qx = _mm256_or_si256(qx, bxhi);
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
const __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@ -4686,7 +4690,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
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].s;
summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
__m256i bx_0 = bytes_from_nibbles_32(x[i].qs);
const __m256i bxhi = bytes_from_bits_32(x[i].qh);
@ -4700,7 +4704,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
bxh = _mm_or_si128(bxh, bxhih);
bx_0 = MM256_SET_M128I(bxh, bxl);
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
@ -4767,7 +4771,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
}
*s = sumf;
@ -4791,7 +4795,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
}
sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
}
*s = sumf;