ggml : build backends as libraries (#10256)
* ggml : build backends as libraries --------- Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com> Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> Co-authored-by: R0CKSTAR <xiaodong.ye@mthreads.com>
This commit is contained in:
Diego Devesa
GitHub
parent
4a8ccb37ad
commit
ae8de6d50a
191 changed files with 17541 additions and 16879 deletions
|
|
@ -3,13 +3,29 @@
|
|||
// GGML internal header
|
||||
|
||||
#include "ggml.h"
|
||||
|
||||
#include <assert.h>
|
||||
#include <math.h>
|
||||
#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
|
||||
#include <stdbool.h>
|
||||
#include <stdint.h>
|
||||
#include <string.h>
|
||||
|
||||
#ifdef __ARM_FEATURE_SVE
|
||||
#include <arm_sve.h>
|
||||
#endif // __ARM_FEATURE_SVE
|
||||
|
||||
#if defined(__ARM_NEON)
|
||||
// if YCM cannot find <arm_neon.h>, 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 <arm_neon.h>
|
||||
#endif
|
||||
|
||||
#if defined(__F16C__)
|
||||
#include <immintrin.h>
|
||||
#endif
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
|
@ -28,13 +44,13 @@ extern "C" {
|
|||
// if C99 - static_assert is noop
|
||||
// ref: https://stackoverflow.com/a/53923785/4039976
|
||||
#ifndef __cplusplus
|
||||
#ifndef static_assert
|
||||
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
|
||||
#define static_assert(cond, msg) _Static_assert(cond, msg)
|
||||
#else
|
||||
#define static_assert(cond, msg) struct global_scope_noop_trick
|
||||
#endif
|
||||
#endif
|
||||
#ifndef static_assert
|
||||
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
|
||||
#define static_assert(cond, msg) _Static_assert(cond, msg)
|
||||
#else
|
||||
#define static_assert(cond, msg) struct global_scope_noop_trick
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
static inline int ggml_up32(int n) {
|
||||
|
|
@ -120,14 +136,12 @@ struct ggml_map_custom1_op_params {
|
|||
void * userdata;
|
||||
};
|
||||
|
||||
|
||||
struct ggml_map_custom2_op_params {
|
||||
ggml_custom2_op_t fun;
|
||||
int n_tasks;
|
||||
void * userdata;
|
||||
};
|
||||
|
||||
|
||||
struct ggml_map_custom3_op_params {
|
||||
ggml_custom3_op_t fun;
|
||||
int n_tasks;
|
||||
|
|
@ -287,9 +301,249 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
|
|||
void * ggml_aligned_malloc(size_t size);
|
||||
void ggml_aligned_free(void * ptr, size_t size);
|
||||
|
||||
// TODO: move to threading file
|
||||
void ggml_critical_section_start(void);
|
||||
void ggml_critical_section_end(void);
|
||||
// FP16 to FP32 conversion
|
||||
|
||||
#if defined(__ARM_NEON)
|
||||
#ifdef _MSC_VER
|
||||
typedef uint16_t ggml_fp16_internal_t;
|
||||
#else
|
||||
typedef __fp16 ggml_fp16_internal_t;
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#if defined(__ARM_NEON) && !defined(_MSC_VER)
|
||||
#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)
|
||||
|
||||
#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
|
||||
|
||||
static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
|
||||
ggml_fp16_internal_t tmp;
|
||||
memcpy(&tmp, &h, sizeof(ggml_fp16_t));
|
||||
return (float)tmp;
|
||||
}
|
||||
|
||||
static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
|
||||
ggml_fp16_t res;
|
||||
ggml_fp16_internal_t tmp = f;
|
||||
memcpy(&res, &tmp, sizeof(ggml_fp16_t));
|
||||
return res;
|
||||
}
|
||||
|
||||
#elif defined(__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__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
|
||||
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__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
|
||||
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 // defined(__ARM_NEON) && (!defined(__MSC_VER)
|
||||
|
||||
// precomputed f32 table for f16 (256 KB)
|
||||
// defined in ggml.c, initialized in ggml_init()
|
||||
GGML_API float ggml_table_f32_f16[1 << 16];
|
||||
|
||||
// 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)
|
||||
inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
|
||||
uint16_t s;
|
||||
memcpy(&s, &f, sizeof(uint16_t));
|
||||
return ggml_table_f32_f16[s];
|
||||
}
|
||||
|
||||
#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
|
||||
#endif
|
||||
|
||||
#if !defined(GGML_FP32_TO_FP16)
|
||||
#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
|
||||
#endif
|
||||
|
||||
/**
|
||||
* Converts brain16 to float32.
|
||||
*
|
||||
* The bfloat16 floating point format has the following structure:
|
||||
*
|
||||
* ┌sign
|
||||
* │
|
||||
* │ ┌exponent
|
||||
* │ │
|
||||
* │ │ ┌mantissa
|
||||
* │ │ │
|
||||
* │┌──┴───┐┌─┴───┐
|
||||
* 0b0000000000000000 brain16
|
||||
*
|
||||
* Since bf16 has the same number of exponent bits as a 32bit float,
|
||||
* encoding and decoding numbers becomes relatively straightforward.
|
||||
*
|
||||
* ┌sign
|
||||
* │
|
||||
* │ ┌exponent
|
||||
* │ │
|
||||
* │ │ ┌mantissa
|
||||
* │ │ │
|
||||
* │┌──┴───┐┌─┴───────────────────┐
|
||||
* 0b00000000000000000000000000000000 IEEE binary32
|
||||
*
|
||||
* For comparison, the standard fp16 format has fewer exponent bits.
|
||||
*
|
||||
* ┌sign
|
||||
* │
|
||||
* │ ┌exponent
|
||||
* │ │
|
||||
* │ │ ┌mantissa
|
||||
* │ │ │
|
||||
* │┌─┴─┐┌─┴──────┐
|
||||
* 0b0000000000000000 IEEE binary16
|
||||
*
|
||||
* @see IEEE 754-2008
|
||||
*/
|
||||
static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
|
||||
union {
|
||||
float f;
|
||||
uint32_t i;
|
||||
} u;
|
||||
u.i = (uint32_t)h.bits << 16;
|
||||
return u.f;
|
||||
}
|
||||
|
||||
/**
|
||||
* Converts float32 to brain16.
|
||||
*
|
||||
* This is binary identical with Google Brain float conversion.
|
||||
* Floats shall round to nearest even, and NANs shall be quiet.
|
||||
* Subnormals aren't flushed to zero, except perhaps when used.
|
||||
* This code should vectorize nicely if using modern compilers.
|
||||
*/
|
||||
static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
|
||||
ggml_bf16_t h;
|
||||
union {
|
||||
float f;
|
||||
uint32_t i;
|
||||
} u;
|
||||
u.f = s;
|
||||
if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
|
||||
h.bits = (u.i >> 16) | 64; /* force to quiet */
|
||||
return h;
|
||||
}
|
||||
h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
|
||||
return h;
|
||||
}
|
||||
|
||||
#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
|
||||
#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue