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

formatting changes.

Browse source
This commit is contained in:
Julia Longtin 2024-04-03 20:24:00 +00:00
parent 96fdd214c8
commit 6f67ea886f
1 changed files with 36 additions and 39 deletions
Download patch file Download diff file Expand all files Collapse all files

13
ggml-phi-knc.c
View file

@ -6,14 +6,11 @@
// For memcpy. // For memcpy.
#include <string.h> #include <string.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. // We can fit 16 of these float32s in a single vector register.
#define GGML_F32_EPR 16 #define GGML_F32_EPR 16
// a single vector. 128*32=512 // A vector of 16 floats.
typedef float float32x16_t __attribute__((vector_size (128))); typedef float float32x16_t __attribute__((vector_size (64), aligned (64)));
#define GGML_F32x16 float32x16_t
// A forward declaration, to keep GCC happy... // A forward declaration, to keep GCC happy...
void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc); void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc);
@ -109,7 +106,7 @@ inline static void GGML_F32x16_VEC_FMA(const float32x16_t *mvec1, const float32x
void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc) void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc)
{ {
// our sum. // our sum.
float32x16_t sum __attribute__((aligned(64))); float32x16_t sum;
// the number of vector-sized steps we will need to do. // the number of vector-sized steps we will need to do.
const uint32_t np = (n & ~(GGML_F32_EPR - 1)); const uint32_t np = (n & ~(GGML_F32_EPR - 1));
@ -121,10 +118,10 @@ void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restri
{ {
// add the leftovers, that could not be handled by the vector loop. // add the leftovers, that could not be handled by the vector loop.
// our extended last part of x. // our extended last part of x.
float32x16_t v1 __attribute__((aligned(64))); float32x16_t v1;
GGML_F32x16_VEC_ZERO(&v1); GGML_F32x16_VEC_ZERO(&v1);
// our extended last part of y. // our extended last part of y.
float32x16_t v2 __attribute__((aligned(64))); float32x16_t v2;
GGML_F32x16_VEC_ZERO(&v2); GGML_F32x16_VEC_ZERO(&v2);
memcpy(&v1, &x[np], (n - np)*sizeof(float)); memcpy(&v1, &x[np], (n - np)*sizeof(float));