mirror of
https://github.com/ggerganov/llama.cpp.git
synced 2024-11-11 13:30:35 +00:00
ggml : move FP16 <-> FP32 code to ggml-impl.h (#3861)
* ggml : move FP16 <-> FP32 stuff to ggml-impl.h ggml-ci * tests : fix ARM build * ggml : explicitly initialize deprecated type traits * ggml : add math.h to ggml-impl.h * ggml : remove duplicate static assert macros * ggml : prefix lookup tables with ggml_ ggml-ci * ggml-impl : move extern "C" to start of file
This commit is contained in:
Georgi Gerganov
GitHub
parent
6e08281e58
commit
207b51900e
237
ggml-impl.h
Normal file
237
ggml-impl.h
Normal file
@ -0,0 +1,237 @@
|
||||
#pragma once
|
||||
|
||||
#include "ggml.h"
|
||||
|
||||
// GGML internal header
|
||||
|
||||
#include <assert.h>
|
||||
#include <stddef.h>
|
||||
#include <stdbool.h>
|
||||
#include <string.h> // memcpy
|
||||
#include <math.h> // fabsf
|
||||
|
||||
#ifdef __cplusplus
|
||||
extern "C" {
|
||||
#endif
|
||||
|
||||
// static_assert should be a #define, but if it's not,
|
||||
// fall back to the _Static_assert C11 keyword.
|
||||
// if C99 - static_assert is noop
|
||||
// ref: https://stackoverflow.com/a/53923785/4039976
|
||||
#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
|
||||
|
||||
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
|
||||
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
|
||||
#ifndef __FMA__
|
||||
#define __FMA__
|
||||
#endif
|
||||
#ifndef __F16C__
|
||||
#define __F16C__
|
||||
#endif
|
||||
#ifndef __SSE3__
|
||||
#define __SSE3__
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#undef MIN
|
||||
#undef MAX
|
||||
|
||||
#define MIN(a, b) ((a) < (b) ? (a) : (b))
|
||||
#define MAX(a, b) ((a) > (b) ? (a) : (b))
|
||||
|
||||
// 16-bit float
|
||||
// on Arm, we use __fp16
|
||||
// on x86, we use uint16_t
|
||||
#if defined(__ARM_NEON) && !defined(_MSC_VER)
|
||||
|
||||
// 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>
|
||||
|
||||
#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 <wasm_simd128.h>
|
||||
#else
|
||||
#ifdef __POWER9_VECTOR__
|
||||
#include <altivec.h>
|
||||
#undef bool
|
||||
#define bool _Bool
|
||||
#else
|
||||
#if defined(_MSC_VER) || defined(__MINGW32__)
|
||||
#include <intrin.h>
|
||||
#else
|
||||
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
|
||||
#if !defined(__riscv)
|
||||
#include <immintrin.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#ifdef __riscv_v_intrinsic
|
||||
#include <riscv_vector.h>
|
||||
#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
|
||||
|
||||
// precomputed f32 table for f16 (256 KB)
|
||||
// defined in ggml.c, initialized in ggml_init()
|
||||
extern 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) || !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 ggml_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
|
||||
|
||||
// TODO: backend v2 PR
|
||||
|
||||
#ifdef __cplusplus
|
||||
}
|
||||
#endif
|
||||
350
ggml-quants.c
350
ggml-quants.c
File diff suppressed because it is too large
Load Diff
@ -1,22 +1,12 @@
|
||||
#pragma once
|
||||
|
||||
// This is a private API for quantization and dequantization
|
||||
// Should not be used directly, use ggml.h instead
|
||||
#include "ggml-impl.h"
|
||||
|
||||
#include "ggml.h"
|
||||
// GGML internal header
|
||||
|
||||
#include <stdint.h>
|
||||
#include <assert.h>
|
||||
#include <stddef.h>
|
||||
|
||||
#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
|
||||
|
||||
#define QK4_0 32
|
||||
typedef struct {
|
||||
ggml_fp16_t d; // delta
|
||||
|
||||
282
ggml.c
282
ggml.c
@ -1,6 +1,6 @@
|
||||
#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
|
||||
|
||||
#include "ggml.h"
|
||||
#include "ggml-impl.h"
|
||||
#include "ggml-quants.h"
|
||||
|
||||
#if defined(_MSC_VER) || defined(__MINGW32__)
|
||||
@ -27,18 +27,6 @@
|
||||
#include <unistd.h>
|
||||
#endif
|
||||
|
||||
// static_assert should be a #define, but if it's not,
|
||||
// fall back to the _Static_assert C11 keyword.
|
||||
// if C99 - static_assert is noop
|
||||
// ref: https://stackoverflow.com/a/53923785/4039976
|
||||
#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
|
||||
|
||||
#if defined(_MSC_VER)
|
||||
// disable "possible loss of data" to avoid hundreds of casts
|
||||
// we should just be careful :)
|
||||
@ -106,23 +94,11 @@ typedef void * thread_ret_t;
|
||||
#include <unistd.h>
|
||||
|
||||
#endif
|
||||
|
||||
#ifdef GGML_USE_CPU_HBM
|
||||
#include <hbwmalloc.h>
|
||||
#endif
|
||||
|
||||
// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
|
||||
#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
|
||||
#ifndef __FMA__
|
||||
#define __FMA__
|
||||
#endif
|
||||
#ifndef __F16C__
|
||||
#define __F16C__
|
||||
#endif
|
||||
#ifndef __SSE3__
|
||||
#define __SSE3__
|
||||
#endif
|
||||
#endif
|
||||
|
||||
/*#define GGML_PERF*/
|
||||
#define GGML_DEBUG 0
|
||||
#define GGML_GELU_FP16
|
||||
@ -248,213 +224,27 @@ inline static void * ggml_aligned_malloc(size_t size) {
|
||||
#include "ggml-opencl.h"
|
||||
#endif
|
||||
|
||||
#undef MIN
|
||||
#undef MAX
|
||||
#define MIN(a, b) ((a) < (b) ? (a) : (b))
|
||||
#define MAX(a, b) ((a) > (b) ? (a) : (b))
|
||||
|
||||
// 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
|
||||
#if defined(__ARM_NEON) && !defined(_MSC_VER)
|
||||
|
||||
// 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>
|
||||
|
||||
#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 <wasm_simd128.h>
|
||||
#else
|
||||
#ifdef __POWER9_VECTOR__
|
||||
#include <altivec.h>
|
||||
#undef bool
|
||||
#define bool _Bool
|
||||
#else
|
||||
#if defined(_MSC_VER) || defined(__MINGW32__)
|
||||
#include <intrin.h>
|
||||
#else
|
||||
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
|
||||
#if !defined(__riscv)
|
||||
#include <immintrin.h>
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
#endif
|
||||
|
||||
#ifdef __riscv_v_intrinsic
|
||||
#include <riscv_vector.h>
|
||||
#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];
|
||||
static ggml_fp16_t ggml_table_gelu_f16[1 << 16];
|
||||
|
||||
// precomputed quick gelu table for f16 (128 KB)
|
||||
static ggml_fp16_t table_gelu_quick_f16[1 << 16];
|
||||
static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
|
||||
|
||||
// precomputed silu table for f16 (128 KB)
|
||||
static ggml_fp16_t table_silu_f16[1 << 16];
|
||||
static ggml_fp16_t ggml_table_silu_f16[1 << 16];
|
||||
|
||||
// precomputed exp table for f16 (128 KB)
|
||||
static ggml_fp16_t table_exp_f16[1 << 16];
|
||||
static ggml_fp16_t ggml_table_exp_f16[1 << 16];
|
||||
|
||||
// precomputed f32 table for f16 (256 KB)
|
||||
static float 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) || !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
|
||||
// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
|
||||
float ggml_table_f32_f16[1 << 16];
|
||||
|
||||
// note: do not use these inside ggml.c
|
||||
// these are meant to be used via the ggml.h API
|
||||
@ -632,6 +422,28 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
|
||||
.vec_dot = ggml_vec_dot_q4_1_q8_1,
|
||||
.vec_dot_type = GGML_TYPE_Q8_1,
|
||||
},
|
||||
[4] = { // GGML_TYPE_Q4_2
|
||||
.type_name = "DEPRECATED",
|
||||
.blck_size = 0,
|
||||
.type_size = 0,
|
||||
.is_quantized = false,
|
||||
.to_float = NULL,
|
||||
.from_float = NULL,
|
||||
.from_float_reference = NULL,
|
||||
.vec_dot = NULL,
|
||||
.vec_dot_type = GGML_TYPE_COUNT,
|
||||
},
|
||||
[5] = { // GGML_TYPE_Q4_3
|
||||
.type_name = "DEPRECATED",
|
||||
.blck_size = 0,
|
||||
.type_size = 0,
|
||||
.is_quantized = false,
|
||||
.to_float = NULL,
|
||||
.from_float = NULL,
|
||||
.from_float_reference = NULL,
|
||||
.vec_dot = NULL,
|
||||
.vec_dot_type = GGML_TYPE_COUNT,
|
||||
},
|
||||
[GGML_TYPE_Q5_0] = {
|
||||
.type_name = "q5_0",
|
||||
.blck_size = QK5_0,
|
||||
@ -1551,7 +1363,7 @@ inline static float ggml_gelu_f32(float 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]];
|
||||
y[i] = ggml_table_gelu_f16[i16[i]];
|
||||
}
|
||||
}
|
||||
|
||||
@ -1561,7 +1373,7 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
|
||||
for (int i = 0; i < n; ++i) {
|
||||
ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
|
||||
memcpy(&t, &fp16, sizeof(uint16_t));
|
||||
y[i] = GGML_FP16_TO_FP32(table_gelu_f16[t]);
|
||||
y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]);
|
||||
}
|
||||
}
|
||||
#else
|
||||
@ -1579,7 +1391,7 @@ inline static float ggml_gelu_quick_f32(float x) {
|
||||
//inline static void ggml_vec_gelu_quick_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_quick_f16[i16[i]];
|
||||
// y[i] = ggml_table_gelu_quick_f16[i16[i]];
|
||||
// }
|
||||
//}
|
||||
|
||||
@ -1589,7 +1401,7 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float *
|
||||
for (int i = 0; i < n; ++i) {
|
||||
ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
|
||||
memcpy(&t, &fp16, sizeof(uint16_t));
|
||||
y[i] = GGML_FP16_TO_FP32(table_gelu_quick_f16[t]);
|
||||
y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
|
||||
}
|
||||
}
|
||||
#else
|
||||
@ -1608,7 +1420,7 @@ inline static float ggml_silu_f32(float 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) {
|
||||
// y[i] = table_silu_f16[i16[i]];
|
||||
// y[i] = ggml_table_silu_f16[i16[i]];
|
||||
// }
|
||||
//}
|
||||
|
||||
@ -1618,7 +1430,7 @@ inline static void ggml_vec_silu_f32(const int n, float * y, const float * x) {
|
||||
for (int i = 0; i < n; ++i) {
|
||||
ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
|
||||
memcpy(&t, &fp16, sizeof(uint16_t));
|
||||
y[i] = GGML_FP16_TO_FP32(table_silu_f16[t]);
|
||||
y[i] = GGML_FP16_TO_FP32(ggml_table_silu_f16[t]);
|
||||
}
|
||||
}
|
||||
#else
|
||||
@ -2334,11 +2146,11 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
|
||||
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_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_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 float f = ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(ii);
|
||||
ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
|
||||
ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f));
|
||||
ggml_table_silu_f16[i] = GGML_FP32_TO_FP16(ggml_silu_f32(f));
|
||||
ggml_table_exp_f16[i] = GGML_FP32_TO_FP16(expf(f));
|
||||
}
|
||||
|
||||
const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
|
||||
@ -10701,7 +10513,7 @@ static void ggml_compute_forward_soft_max_f32(
|
||||
// const float val = (sp[i] == -INFINITY) ? 0.0 : exp(sp[i] - max);
|
||||
ggml_fp16_t s = GGML_FP32_TO_FP16(sp[i] - max);
|
||||
memcpy(&scvt, &s, sizeof(scvt));
|
||||
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]);
|
||||
const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]);
|
||||
sum += (ggml_float)val;
|
||||
dp[i] = val;
|
||||
}
|
||||
@ -12990,7 +12802,7 @@ static void ggml_compute_forward_flash_attn_f32(
|
||||
#else
|
||||
ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max);
|
||||
memcpy(&scvt[j], &s, sizeof(uint16_t));
|
||||
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
|
||||
const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt[j]]);
|
||||
#endif
|
||||
sump[j] += (ggml_float)val;
|
||||
SS[j] = val;
|
||||
@ -13192,7 +13004,7 @@ static void ggml_compute_forward_flash_attn_f16(
|
||||
} else {
|
||||
ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max);
|
||||
memcpy(&scvt[j], &s, sizeof(uint16_t));
|
||||
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
|
||||
const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt[j]]);
|
||||
sump[j] += (ggml_float)val;
|
||||
SS[j] = val;
|
||||
}
|
||||
@ -13643,7 +13455,7 @@ static void ggml_compute_forward_flash_attn_back_f32(
|
||||
#else
|
||||
ggml_fp16_t s = GGML_FP32_TO_FP16(SR[j] - max);
|
||||
memcpy(&scvt[j], &s, sizeof(uint16_t));
|
||||
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt[j]]);
|
||||
const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt[j]]);
|
||||
#endif
|
||||
sump[j] += (ggml_float)val;
|
||||
SW[j] = val;
|
||||
@ -14393,7 +14205,7 @@ static void ggml_compute_forward_cross_entropy_loss_f32(
|
||||
#else
|
||||
ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max);
|
||||
memcpy(&scvt, &s, sizeof(scvt));
|
||||
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]);
|
||||
const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]);
|
||||
#endif
|
||||
sum += (ggml_float)val;
|
||||
st[i] = val;
|
||||
@ -14507,7 +14319,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
|
||||
#else
|
||||
ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max);
|
||||
memcpy(&scvt, &s, sizeof(scvt));
|
||||
const float val = GGML_FP16_TO_FP32(table_exp_f16[scvt]);
|
||||
const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]);
|
||||
#endif
|
||||
sum += (ggml_float)val;
|
||||
ds0[i] = val;
|
||||
|
||||
@ -1467,7 +1467,7 @@ static int32_t llama_kv_cache_cell_max(const struct llama_kv_cache & cache) {
|
||||
}
|
||||
|
||||
static void llama_kv_cache_clear(struct llama_kv_cache & cache) {
|
||||
for (int32_t i = 0; i < cache.size; ++i) {
|
||||
for (int32_t i = 0; i < (int32_t) cache.size; ++i) {
|
||||
cache.cells[i].pos = -1;
|
||||
cache.cells[i].seq_id.clear();
|
||||
}
|
||||
|
||||
@ -4,7 +4,7 @@
|
||||
|
||||
#undef NDEBUG
|
||||
#include <cassert>
|
||||
#if !defined(__riscv) && !defined(__s390__)
|
||||
#if !defined(__riscv) && !defined(__s390__) && !defined(__ARM_NEON)
|
||||
#include <immintrin.h>
|
||||
#endif
|
||||
#include <cmath>
|
||||
|
||||
@ -129,6 +129,13 @@ int main(int argc, char * argv[]) {
|
||||
ggml_type type = (ggml_type) i;
|
||||
ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
|
||||
|
||||
// deprecated - skip
|
||||
if (qfns.blck_size == 0) {
|
||||
continue;
|
||||
}
|
||||
|
||||
printf("Testing %s\n", ggml_type_name((ggml_type) i));
|
||||
|
||||
if (qfns.from_float && qfns.to_float) {
|
||||
const float total_error = total_quantization_error(qfns, test_size, test_data.data());
|
||||
const float max_quantization_error =
|
||||
|
||||
Loading…
Reference in New Issue
Block a user