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

subtract the correct amount.

Browse source
This commit is contained in:
Julia Longtin 2024-05-11 11:11:15 +00:00
parent 511ad8043f
commit 47ca67a062
1 changed files with 1 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml-phi-knc.c
View file

@ -62,7 +62,7 @@ inline static void GGML_F32x16_VEC_FMA(const float32x16_t *mvec1, const float32x
"cmp\t$4,\t%%r8\n\t" // Compare iterations to four. "cmp\t$4,\t%%r8\n\t" // Compare iterations to four.
"jnae\t6f\n\t" // If there are not four iterations left, jump to label 6. "jnae\t6f\n\t" // If there are not four iterations left, jump to label 6.
"1:\n\t" "1:\n\t"
"sub\t$3,\t%%r8\n\t" // Decrement iterations "sub\t$4,\t%%r8\n\t" // Decrement iterations
"vmovaps\t\t(%%r10),\t%%zmm1\n\t" // Load two vectors. "vmovaps\t\t(%%r10),\t%%zmm1\n\t" // Load two vectors.
"vmovaps\t\t(%%r12),\t%%zmm2\n\t" "vmovaps\t\t(%%r12),\t%%zmm2\n\t"
"vprefetchnta\t192(%%r10)\n\t" // prefetch the next float32x16_t block (192 bytes ahead) "vprefetchnta\t192(%%r10)\n\t" // prefetch the next float32x16_t block (192 bytes ahead)