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

spacing changes, eliminate dead references to k1 or zero, and use the right type when referring to src.

Browse source
This commit is contained in:
Julia Longtin 2024-03-24 12:37:47 +00:00
parent 98c9b6972a
commit 3cdfc9c596
1 changed files with 39 additions and 39 deletions
Download patch file Download diff file Expand all files Collapse all files

14
ggml-phi-knc-dot_q5_K_q8_K.c
View file

@ -22,9 +22,10 @@
#define GGML_F32_EPR 16
typedef float float32x8_t __attribute__((vector_size (64)));
typedef float float32x16_t __attribute__((vector_size (128)));
typedef int16_t int16x8_t __attribute__((vector_size (32)));
typedef int32_t int32x8_t __attribute__((vector_size (64)));
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. */
@ -69,11 +70,10 @@ inline static void GGML_I32x16_VEC_ZERO(int32x16_t *target)
__asm__ __volatile__ (
"vbroadcastI32x4\t%[Z]%{uint8%},\t%%zmm8\n\t" // use an upscaling operator to clear our value.
"kmov\t%[M],\t%%k1\n\t"
"vmovaps\t\t%%zmm8,\t%[RES]%{%%k1%}\n\t"
"vmovaps\t\t%%zmm8,\t%[RES]\n\t"
: [RES] "+m" (*target)
: [Z] "m" (zero)
: "zmm8", "k1", "memory");
: "zmm8", "memory");
}
// perform a Fused Multiply Add of an I16x8 times scalar S into I32x8.
@ -110,9 +110,9 @@ inline static void GGML_I16x16_S_FMA_I32x16 (int16x16_t *src, int32_t scale, int
"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),
: [SRC] "m" (*src),
[SCALE] "m" (scaleVec)
: "zmm0", "zmm1", "zmm2", "k1", "memory");
: "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) {
@ -176,8 +176,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * r
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;
/* FIXME: while comparing FMA output to normal output, the original had an error. hunt it down. */
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;
}