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

perform better prefetches, and invert the test of our clear flag for clarity.

Browse source
This commit is contained in:
Julia Longtin 2024-05-10 16:14:28 +00:00
parent a14fe02cf8
commit d8d574c56f
1 changed files with 14 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

19
ggml-phi-knc.c
View file

@ -38,11 +38,20 @@ inline static void GGML_F32x16_VEC_FMA(const float32x16_t *mvec1, const float32x
uint8_t zero = 0; uint8_t zero = 0;
__asm__ __volatile__ ( __asm__ __volatile__ (
"mov\t%[ITER],%%r8\n\t" // how many register sized chunks are we responsible for "vprefetchenta\t(%[RES])\n\t"
"mov\t%[VEC1],%%r10\n\t" // where do we start work in mvec1? "vprefetch0\t(%[VEC1])\n\t"
"mov\t%[VEC2],%%r12\n\t" // where do we start work in mvec2? "vprefetch1\t64(%[VEC1])\n\t"
"cmp\t$1,%[CLR]\n\t" // should we clear the sum before we start? "vprefetch0\t128(%[VEC1])\n\t"
"jne\t4f\n\t" "vprefetch1\t192(%[VEC1])\n\t"
"vprefetch0\t(%[VEC2])\n\t"
"vprefetch1\t64(%[VEC2])\n\t"
"vprefetch0\t128(%[VEC2])\n\t"
"vprefetch1\t192(%[VEC2])\n\t"
"mov\t%[ITER],%%r8\n\t" // How many vector sized chunks we are responsible for.
"mov\t%[VEC1],%%r10\n\t" // Where do we start work in mvec1?
"mov\t%[VEC2],%%r12\n\t" // Where do we start work in mvec2?
"cmp\t$0,%[CLR]\n\t" // Should we clear the sum before we start?
"jz\t4f\n\t"
"vbroadcastss\t%[Z]%{uint8%},\t%%zmm0\n\t" // if so, use an upscaling operator to do it. "vbroadcastss\t%[Z]%{uint8%},\t%%zmm0\n\t" // if so, use an upscaling operator to do it.
"vprefetchnta\t(%%r10)\n\t" "vprefetchnta\t(%%r10)\n\t"
"vprefetchnta\t(%%r12)\n\t" "vprefetchnta\t(%%r12)\n\t"