llama.cpp/ggml-phi-knc.c

117 lines
4.0 KiB
C

#include <immintrin.h>
#include <stdint.h>
#include <stdio.h>
static inline _Bool is_aligned(const void *restrict pointer, size_t byte_count)
{ return (uintptr_t)pointer % byte_count == 0; }
// No, we have an SIMD unit.
// #define GGML_SIMD
// 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
// because we are not defining GGML_SIMD, we have to do this ourself.
#define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR)
// a single vector. 128*32=512
typedef float float32x16_t __attribute__((vector_size (128)));
#define GGML_F32x16 float32x16_t
// from chatGPT. nuke this later.
#include <string.h>
inline static void GGML_F32x16_VEC_ZERO(float32x16_t *target)
{
// we only need a mask16, but register sizes...
__mmask32 mask=0xFFFFFFFF;
// FIXME: how do we tell GNU AS to perform upconverts?
float zero[4] __attribute__((aligned(64))) = {0.0f,0.0f,0.0f,0.0f};
__asm__ __volatile__ ("movl\t%[M],\t%%eax\n\t"
"kmov %%eax,\t%%k1\n\t"
"vbroadcastf32x4\t%[Z],\t%%zmm0%{%%k1%}\n\t"
"vmovaps\t\t%%zmm0,\t%[RES]%{%%k1%}\n\t"
: [RES] "+m" (*target)
: [M] "m" (mask),
[Z] "m" (zero)
: "eax", "k1", "zmm0");
}
// multiply each item in mvec1 with the corresponding item in mvec2, adding the result to the corresponding item in sum.
inline static void GGML_F32x16_VEC_FMA(const float32x16_t *mvec1, const float32x16_t *mvec2, float32x16_t *sumvec, size_t iterations)
{
// we only need a mask16, but register sizes...
__mmask32 mask=0xFFFFFFFF;
__asm__ __volatile__ (
"vmovaps\t\t(%[RES]),\t%%zmm0\n\t" // load our initial state..
"1:\n\t"
"cmp $0,\t%[ITER]\n\t" // Compare iterations to 0
"je\t2f\n\t" // Jump to label 2 if zero (end of loop)
"vmovaps\t\t(%[VEC1]),\t%%zmm1\n\t" // Load two vectors.
"vmovaps\t\t(%[VEC2]),\t%%zmm2\n\t"
"vfmadd231ps\t%%zmm1,\t%%zmm2,\t%%zmm0\n\t" // Perform a fused multiply add.
"add $64,\t%[VEC1]\n\t" // Move to the next float32x16_t (64 bytes ahead)
"add $64,\t%[VEC2]\n\t"
"sub $1,\t%[ITER]\n\t" // Decrement iterations
"jmp 1b\n\t" // Jump back to the start of the loop
"2: \n\t" // Label for loop end
"vmovaps\t\t%%zmm0,\t(%[RES])\n\t" // save our results.
: [RES] "+r" (sumvec),
[ITER] "+r" (iterations)
: [M] "r" (mask),
[VEC1] "r" (mvec1),
[VEC2] "r" (mvec2)
: "zmm0", "zmm1", "zmm2", "cc", "memory");
}
// NOTE: all inputs must be __attribute__((aligned(64)));
float DotProduct_F32(const float * restrict inVec1, const float * restrict inVec2, uint32_t count)
{
// our single result, in the end.
float sumf = 0.0f;
// our sum.
float32x16_t sum __attribute__((aligned(64)));
// the number of vector-sized steps we will need to do.
const uint32_t np = (count & ~(GGML_F32_EPR - 1));
GGML_F32x16_VEC_ZERO(&sum);
// 0 indexed cycle count
// for (uint32_t cycle = 0; cycle < (np/GGML_F32_EPR); ++cycle)
GGML_F32x16_VEC_FMA((float32x16_t *)inVec1, (float32x16_t *)inVec2, &sum, np/GGML_F32_EPR);
if (count != np)
{
printf("handling remainder %u\n",count-np);
// add the leftovers, that could not be handled by the vector loop.
// our extended last part of inVec1.
float32x16_t v1 __attribute__((aligned(64)));
GGML_F32x16_VEC_ZERO(&v1);
// our extended last part of inVec2.
float32x16_t v2 __attribute__((aligned(64)));
GGML_F32x16_VEC_ZERO(&v2);
memcpy(&v1, &inVec1[np], (count - np)*sizeof(float));
memcpy(&v2, &inVec2[np], (count - np)*sizeof(float));
GGML_F32x16_VEC_FMA(&v1,
&v2,
&sum, 1);
}
// reduce sum0..sumX to sumf
for (uint32_t i=0; i <GGML_F32_EPR; ++i)
sumf+=((float *)&sum)[i];
return sumf;
}