happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
llama.cpp/ggml-phi-knc-dot_q5_K_q8_K.c

208 lines
8.4 KiB
C
Raw Blame History

// For uint32_t
#include <stdint.h>
// For size_t
#include <stdio.h>
// Yes, we have to tell this header to actually export stuff.
#define GGML_COMMON_IMPL_C
#include "ggml-common.h"
#include "ggml-quants.h"
#include "ggml-impl.h"
// FIXME: why do we have to import this twice?
#define GGML_COMMON_IMPL_C
// For block_q5_K and block_q8_K. only given the second time.
#include "ggml-common.h"
// This SIMD unit can work with 32 float32s at once.
#define GGML_F32_STEP 32
// We can fit 16 of these float32s in a single vector register.
#define GGML_F32_EPR 16
typedef float float32x8_t __attribute__((vector_size (64)));
typedef float float32x16_t __attribute__((vector_size (128)));
typedef int8_t int8x16_t __attribute__((vector_size (32)));
typedef int16_t int16x8_t __attribute__((vector_size (32)));
typedef int16_t int16x16_t __attribute__((vector_size (64)));
typedef int32_t int32x8_t __attribute__((vector_size (64)));
typedef int32_t int32x16_t __attribute__((vector_size (128)));
/* A forward declaration, to keep GCC happy. */
void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc);
/* clear a vector of 8 floats. */
inline static void GGML_F32x8_VEC_ZERO(float32x8_t *target)
{
uint8_t zero[4] __attribute__((aligned(32))) = {0,0,0,0};
uint32_t mask=0x000000FF;
__asm__ __volatile__ (
"vbroadcastf32x4\t%[Z]%{uint8%},\t%%zmm8\n\t" // use an upscaling operator to clear our register.
"kmov\t%[M],\t%%k1\n\t"
"vmovaps\t\t%%zmm8,\t%[RES]%{%%k1%}\n\t"
: [RES] "+m" (*target)
: [Z] "m" (zero),
[M] "r" (mask)
: "zmm8", "k1", "memory");
}
/* clear a vector of 16 floats. */
inline static void GGML_F32x16_VEC_ZERO(float32x16_t *target)
{
uint8_t zero[4] __attribute__((aligned(32))) = {0,0,0,0};
__asm__ __volatile__ (
"vbroadcastf32x4\t%[Z]%{uint8%},\t%%zmm8\n\t" // use an upscaling operator to clear our register.
"vmovaps\t\t%%zmm8,\t%[RES]\n\t"
: [RES] "+m" (*target)
: [Z] "m" (zero)
: "zmm8", "memory");
}
/* clear a vector of 8 int32_ts. */
inline static void GGML_I32x8_VEC_ZERO(int32x8_t *target)
{
uint8_t zero[4] __attribute__((aligned(32))) = {0,0,0,0};
uint32_t mask=0x000000FF;
__asm__ __volatile__ (
"vbroadcastI32x4\t%[Z]%{uint8%},\t%%zmm8\n\t" // use an upscaling operator to clear our register.
"kmov\t%[M],\t%%k1\n\t"
"vmovaps\t\t%%zmm8,\t%[RES]%{%%k1%}\n\t"
: [RES] "+m" (*target)
: [Z] "m" (zero),
[M] "r" (mask)
: "zmm8", "k1", "memory");
}
/* clear a vector of 16 int32_ts. */
inline static void GGML_I32x16_VEC_ZERO(int32x16_t *target)
{
uint8_t zero[4] __attribute__((aligned(32))) = {0,0,0,0};
__asm__ __volatile__ (
"vbroadcastI32x4\t%[Z]%{uint8%},\t%%zmm8\n\t" // use an upscaling operator to clear our register.
"vmovaps\t\t%%zmm8,\t%[RES]\n\t"
: [RES] "+m" (*target)
: [Z] "m" (zero)
: "zmm8", "memory");
}
// perform a Fused Multiply Add of an I16x8 times scalar S into I32x8.
inline static void GGML_I16x8_S_FMA_I32x8 (int16x8_t *src, int32_t scale, int32x8_t *dest)
{
uint32_t mask=0x000000FF;
int32_t scaleVec[4] = {scale, scale, scale, scale};
__asm__ __volatile__ (
"kmov\t%[M],\t%%k1\n\t" // we will only be working with 8 values at a time. le sigh.
"vmovdqa32\t\t%[SRC]%{sint16%},\t%%zmm0%{%%k1%}\n\t" // load the item we will be summing from. upscale it from int16.
"vbroadcastI32x4\t%[SCALE],\t%%zmm1\n\t" // load the item we will be multiplying by.
"vmovdqa32\t\t%[RES],\t%%zmm2%{%%k1%}\n\t" // load the item we will be summing onto.
"vpmadd231d\t%%zmm0,\t%%zmm1,\t%%zmm2%{%%k1%}\n\t" // perform our multiply-add.
"vmovdqa32\t\t%%zmm2,\t%[RES]%{%%k1}\n\t" // save the result.
: [RES] "+m" (*dest)
: [M] "r" (mask),
[SRC] "m" (*src),
[SCALE] "m" (scaleVec)
: "zmm0", "zmm1", "zmm2", "k1", "memory");
}
// perform a Fused Multiply Add of an I16x16 times scalar S into I32x16.
inline static void GGML_I16x16_S_FMA_I32x16 (int16x16_t *src, int32_t scale, int32x16_t *dest)
{
int32_t scaleVec[4] = {scale, scale, scale, scale};
__asm__ __volatile__ (
"vmovdqa32\t\t%[SRC]%{sint16%},\t%%zmm0\n\t" // load the item we will be summing from. upscale it from int16.
"vbroadcastI32x4\t%[SCALE],\t%%zmm1\n\t" // load the item we will be multiplying by.
"vmovdqa32\t\t%[RES],\t%%zmm2\n\t" // load the item we will be summing onto.
"vpmadd231d\t%%zmm0,\t%%zmm1,\t%%zmm2\n\t" // perform our multiply-add.
"vmovdqa32\t\t%%zmm2,\t%[RES]\n\t" // save the result.
: [RES] "+m" (*dest)
: [SRC] "m" (*src),
[SCALE] "m" (scaleVec)
: "zmm0", "zmm1", "zmm2", "memory");
}
void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
/* interpret X and Y as vectors. */
const block_q5_K * restrict x = vx;
const block_q8_K * restrict y = vy;
/* the number of blocks we will process this in. */
const int nb = n / QK_K;
static const uint32_t kmask1 = 0x3f3f3f3f;
static const uint32_t kmask2 = 0x0f0f0f0f;
static const uint32_t kmask3 = 0x03030303;
uint32_t utmp[4];
const uint8_t * scales = (const uint8_t*)&utmp[0];
const uint8_t * mins = (const uint8_t*)&utmp[2];
float32x16_t sums __attribute__((aligned(128)));
int8x16_t aux8[QK_K/16] __attribute__((aligned(32)));
int16x16_t aux16 __attribute__((aligned(64)));
int32x16_t aux32 __attribute__((aligned(128)));
GGML_F32x16_VEC_ZERO(&sums);
float sumf = 0;
for (int i = 0; i < nb; ++i) {
const uint8_t * restrict q4 = x[i].qs;
const uint8_t * restrict hm = x[i].qh;
const int8_t * restrict q8 = y[i].qs;
int8_t * restrict a = (int8_t * restrict)aux8;
uint8_t m = 1;
// Fill the 8 bit vector a with our 5 bit quantization data, 64 blocks at a time.
for (int j = 0; j < QK_K/64; ++j) {
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
a += 32; m <<= 1;
q4 += 32;
}
memcpy(utmp, x[i].scales, 12);
utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
const uint32_t uaux = utmp[1] & kmask1;
utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
utmp[2] = uaux;
utmp[0] &= kmask1;
a = (int8_t * restrict)aux8;
int sumi = 0;
GGML_I32x16_VEC_ZERO(&aux32);
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
int is = 0;
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
for (int l = 0; l < 16; ++l) ((int16_t *)&aux16)[l] = q8[l] * a[l];
GGML_I16x16_S_FMA_I32x16 (&aux16, scale, &aux32);
q8 += 16; a += 16;
for (int l = 0; l < 16; ++l) ((int16_t *)&aux16)[l] = q8[l] * a[l];
// FIXME: while comparing FMA output to the original output, the original had an error. hunt it down.
GGML_I16x16_S_FMA_I32x16 (&aux16, scale, &aux32);
q8 += 16; a += 16;
}
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
for (int l = 0; l < 16; ++l) ((float *)&sums)[l] += d * ((int32_t *)&aux32)[l];
const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
sumf -= dmin * sumi;
}
for (int l = 0; l < 16; ++l) sumf += ((float *)&sums)[l];
*s = sumf;
}
Reference in New Issue View Git Blame Copy Permalink