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

spacing changes.

Browse source
This commit is contained in:
Julia Longtin 2024-05-11 13:35:50 +00:00
parent fc23c22fd2
commit a273a9ebf2
1 changed files with 33 additions and 33 deletions
Download patch file Download diff file Expand all files Collapse all files

66
ggml-phi-knc-dot_q5_K_q8_K.c
View file

@ -201,57 +201,57 @@ void GGML_5bit_Unpack_Unaligned (const uint8x16_t * q4, const uint8_t * q1, uint
uint8_t bit5 = 0x10; uint8_t bit5 = 0x10;
__asm__ __volatile__ ( __asm__ __volatile__ (
"vprefetch0\t(%[SRC1])\n\t" // Issue our memory requests first thing. "vprefetch0\t(%[SRC1])\n\t" // Issue our memory requests first thing.
"vprefetch0\t(%[SRC4])\n\t" "vprefetch0\t(%[SRC4])\n\t"
"vprefetchenta\t(%[DST])\n\t" "vprefetchenta\t(%[DST])\n\t"
"mov\t%[SRC4],\t%%r9\n\t" // Load the address of the head of our 4-bit list. "mov\t%[SRC4],\t%%r9\n\t" // Load the address of the head of our 4-bit list.
"mov\t%[DST],\t%%r8\n\t" // Load the address of the head of our destination list. "mov\t%[DST],\t%%r8\n\t" // Load the address of the head of our destination list.
"mov\t$0,%%ecx\n\t" // Initialize our counter. "mov\t$0,%%ecx\n\t" // Initialize our counter.
"vpbroadcastd\t%[MASK]%{uint8%},\t%%zmm0\n\t" // Load our mask. "vpbroadcastd\t%[MASK]%{uint8%},\t%%zmm0\n\t" // Load our mask.
"vpbroadcastd\t%[BIT5]%{uint8},\t%%zmm1\n\t" // Load the bit we want to add (conditionally). "vpbroadcastd\t%[BIT5]%{uint8},\t%%zmm1\n\t" // Load the bit we want to add (conditionally).
"vpbroadcastd\t%[M]%{uint8%},\t%%zmm2\n\t" // Select which bit we want to test for. Start with bit 1. "vpbroadcastd\t%[M]%{uint8%},\t%%zmm2\n\t" // Select which bit we want to test for. Start with bit 1.
"vmovdqa32\t(%[SRC1])%{uint8%},\t%%zmm3\n\t" // Load 16 sets of 8 packed single bits. "vmovdqa32\t(%[SRC1])%{uint8%},\t%%zmm3\n\t" // Load 16 sets of 8 packed single bits.
"vmovdqa32\t16(%[SRC1])%{uint8%},\t%%zmm4\n\t" // Load the next 16 sets of 8 packed single bits. "vmovdqa32\t16(%[SRC1])%{uint8%},\t%%zmm4\n\t" // Load the next 16 sets of 8 packed single bits.
"1:\n\t" "1:\n\t"
"inc\t%%ecx\n\t" // We are in the loop. increment the counter. "inc\t%%ecx\n\t" // We are in the loop. increment the counter.
"vptestmd\t%%zmm3,\t%%zmm2,\t%%k1\n\t" // Test to see if our selected bit is set. "vptestmd\t%%zmm3,\t%%zmm2,\t%%k1\n\t" // Test to see if our selected bit is set.
"vptestmd\t%%zmm4,\t%%zmm2,\t%%k2\n\t" // Test to see if our selected bit is set. "vptestmd\t%%zmm4,\t%%zmm2,\t%%k2\n\t" // Test to see if our selected bit is set.
"vloadunpackld\t\t(%%r9)%{uint8%},\t%%zmm5\n\t" // Load our odd 4 bit sequences. note that it loads two 4 bit sequences into each zmm value. "vloadunpackld\t\t(%%r9)%{uint8%},\t%%zmm5\n\t" // Load our odd 4 bit sequences. note that it loads two 4 bit sequences into each zmm value.
"vloadunpackhd\t\t16(%%r9)%{uint8%},\t%%zmm5\n\t" // Load our odd 4 bit sequences. note that it loads two 4 bit sequences into each zmm value. "vloadunpackhd\t\t16(%%r9)%{uint8%},\t%%zmm5\n\t" // Load our odd 4 bit sequences. note that it loads two 4 bit sequences into each zmm value.
"vpandd\t%%zmm0,\t%%zmm5,\t%%zmm6\n\t" // Apply a mask, storing the first set of four bits into a vector. "vpandd\t%%zmm0,\t%%zmm5,\t%%zmm6\n\t" // Apply a mask, storing the first set of four bits into a vector.
"vpord\t%%zmm1,%%zmm6,%%zmm6%{%%k1%}\n\t" // Turn on bit 5 for all values that passed the prior test. "vpord\t%%zmm1,%%zmm6,%%zmm6%{%%k1%}\n\t" // Turn on bit 5 for all values that passed the prior test.
"vmovdqa32\t\t%%zmm6%{uint8%},\t(%%r8)\n\t" // Save our result. "vmovdqa32\t\t%%zmm6%{uint8%},\t(%%r8)\n\t" // Save our result.
"vloadunpackld\t\t16(%%r9)%{uint8%},\t%%zmm7\n\t" // Load our odd 4 bit sequences. note that it loads two 4 bit sequences into each zmm value. "vloadunpackld\t\t16(%%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. "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 set of four bits into a vector. "vpandd\t%%zmm0,\t%%zmm7,\t%%zmm8\n\t" // Apply a mask, storing the next set of four bits into a vector.
"vpord\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"
"cmp\t$4,\t%%ecx\n\t" "cmp\t$4,\t%%ecx\n\t"
"vpslld\t$1,\t%%zmm2,\t%%zmm2\n\t" // Select the next bit to test for. "vpslld\t$1,\t%%zmm2,\t%%zmm2\n\t" // Select the next bit to test for.
"vptestmd\t%%zmm3,\t%%zmm2,\t%%k1\n\t" // Perform our test. "vptestmd\t%%zmm3,\t%%zmm2,\t%%k1\n\t" // Perform our test.
"vptestmd\t%%zmm4,\t%%zmm2,\t%%k2\n\t" // Perform our test. "vptestmd\t%%zmm4,\t%%zmm2,\t%%k2\n\t" // Perform our test.
"vpsrld\t$4,\t%%zmm5,\t%%zmm6\n\t" // Load our even 4 bit sequence. "vpsrld\t$4,\t%%zmm5,\t%%zmm6\n\t" // Load our even 4 bit sequence.
"vpsrld\t$4,\t%%zmm7,\t%%zmm8\n\t" // Load our next even 4 bit sequence. "vpsrld\t$4,\t%%zmm7,\t%%zmm8\n\t" // Load our next even 4 bit sequence.
"vpord\t%%zmm1,%%zmm6,%%zmm6%{%%k1%}\n\t" // Turn on bit 5 for all values that passed the prior test. "vpord\t%%zmm1,%%zmm6,%%zmm6%{%%k1%}\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. "vpord\t%%zmm1,%%zmm8,%%zmm8%{%%k2%}\n\t" // Turn on bit 5 for all values that passed the prior test.
"vmovdqa32\t\t%%zmm6%{uint8%},\t(%%r8)\n\t" // Save our result. "vmovdqa32\t\t%%zmm6%{uint8%},\t(%%r8)\n\t" // Save our result.
"vmovdqa32\t\t%%zmm8%{uint8%},\t16(%%r8)\n\t" // Save our result. "vmovdqa32\t\t%%zmm8%{uint8%},\t16(%%r8)\n\t" // Save our result.
"vprefetchenta\t32(%%r8)\n\t" "vprefetchenta\t32(%%r8)\n\t"
"je\t2f\n\t" "je\t2f\n\t"
"vprefetch0\t32(%%r9)\n\t" "vprefetch0\t32(%%r9)\n\t"
"vprefetch1\t96(%%r9)\n\t" "vprefetch1\t96(%%r9)\n\t"
"vpslld\t$1,\t%%zmm2,\t%%zmm2\n\t" // Select the next bit to test for. "vpslld\t$1,\t%%zmm2,\t%%zmm2\n\t" // Select the next bit to test for.
"add\t$32,\t%%r9\n\t" "add\t$32,\t%%r9\n\t"
"add\t$32,\t%%r8\n\t" "add\t$32,\t%%r8\n\t"
"jmp\t1b\n\t" "jmp\t1b\n\t"