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

use or, instead of and. bug fix?

Browse source
This commit is contained in:
Julia Longtin 2024-04-24 17:50:12 +00:00
parent dc1f639bf0
commit 0124f7acd8
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-dot_q5_K_q8_K.c
View file

@ -215,7 +215,7 @@ void GGML_5bit_Unpack_Unaligned (const uint8x16_t * q4, const uint8_t * q1, uint
"vloadunpackhd\t\t32(%%r9)%{uint8%},\t%%zmm7\n\t" // load our odd 4 bit sequences. note that it loads two 4 bit sequences into each zmm value. "vloadunpackhd\t\t32(%%r9)%{uint8%},\t%%zmm7\n\t" // load our odd 4 bit sequences. note that it loads two 4 bit sequences into each zmm value.
"vprefetch1\t32(%%r9)\n\t" // pull the next set of 4 bit sequences into the L2 cache. "vprefetch1\t32(%%r9)\n\t" // pull the next set of 4 bit sequences into the L2 cache.
"vpandd\t%%zmm0,\t%%zmm7,\t%%zmm8\n\t" // apply a mask, storing the next low four bits of vector zmm1 into zmm5. "vpandd\t%%zmm0,\t%%zmm7,\t%%zmm8\n\t" // apply a mask, storing the next low four bits of vector zmm1 into zmm5.
"vpaddd\t%%zmm1,%%zmm8,%%zmm8%{%%k2%}\n\t" // turn on bit 5 for all values that passed the prior test. "vpord\t%%zmm1,%%zmm8,%%zmm8%{%%k2%}\n\t" // turn on bit 5 for all values that passed the prior test.
"vmovdqa32\t\t%%zmm8%{uint8%},\t16(%%r8)\n\t" // save our result. "vmovdqa32\t\t%%zmm8%{uint8%},\t16(%%r8)\n\t" // save our result.
"add\t$32,\t%%r8\n\t" "add\t$32,\t%%r8\n\t"