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

separate filling aux16 from consuming aux16 by making it an array of vectors.

Browse source
This commit is contained in:
Julia Longtin 2024-03-24 14:18:08 +00:00
parent e579af1e95
commit 2a47e5f05f
1 changed files with 11 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

18
ggml-phi-knc-dot_q5_K_q8_K.c
View file

@ -147,7 +147,7 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
float32x16_t sums __attribute__((aligned(128))); float32x16_t sums __attribute__((aligned(128)));
int8x16_t aux8[QK_K/16] __attribute__((aligned(32))); int8x16_t aux8[QK_K/16] __attribute__((aligned(32)));
int16x16_t aux16 __attribute__((aligned(64))); int16x16_t aux16[QK_K/16] __attribute__((aligned(64)));
int32x16_t aux32 __attribute__((aligned(128))); int32x16_t aux32 __attribute__((aligned(128)));
GGML_F32x16_VEC_ZERO(&sums); GGML_F32x16_VEC_ZERO(&sums);
@ -188,15 +188,19 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
int is = 0; int is = 0;
for (int j = 0; j < QK_K/32; ++j) { for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++]; for (int l = 0; l < 16; ++l) ((int16_t *)&aux16[j*2])[l] = q8[l] * a[l];
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; q8 += 16; a += 16;
for (int l = 0; l < 16; ++l) ((int16_t *)&aux16)[l] = q8[l] * a[l]; for (int l = 0; l < 16; ++l) ((int16_t *)&aux16[(j*2)+1])[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; q8 += 16; a += 16;
} }
// FIXME: while comparing FMA output to the original output, the original had an error. hunt it down.
for (int j = 0; j < QK_K/32; ++j) {
int32_t scale = scales[is++];
GGML_I16x16_S_FMA_I32x16 (&aux16[j*2], scale, &aux32);
GGML_I16x16_S_FMA_I32x16 (&aux16[(j*2)+1], scale, &aux32);
}
const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d; 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]; 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; const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;