amd/blis
1
0
Fork 0
mirror of https://github.com/amd/blis.git synced 2026-05-11 09:39:59 +00:00
Code Issues Packages Projects Releases Wiki Activity

More minor tweaks to sandybridge/avx micro-kernel.

Browse Source 
Details:
- Re-enabled use of b_next for dgemm and cgemm micro-kernels.
This commit is contained in:
Field G. Van Zee
2014-10-12 13:43:47 -05:00
parent 7b6fe4cae5
commit d1e86e1876
1 changed files with 16 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
kernels/x86_64/avx/3/bli_gemm_asm_d8x4.c
View File

@@ -1043,7 +1043,7 @@ void bli_dgemm_asm_8x4(
)
{
//void* a_next = bli_auxinfo_next_a( data );
//void* b_next = bli_auxinfo_next_b( data );
void* b_next = bli_auxinfo_next_b( data );
dim_t k_iter = k / 4;
dim_t k_left = k % 4;
@@ -1054,8 +1054,9 @@ void bli_dgemm_asm_8x4(
" \n\t"
"movq %2, %%rax \n\t" // load address of a.
"movq %3, %%rbx \n\t" // load address of b.
//"movq %9, %%r15 \n\t" // load address of b_next.
"movq %9, %%r15 \n\t" // load address of b_next.
//"movq %10, %%r14 \n\t" // load address of a_next.
"addq $-4 * 64, %%r15 \n\t"
" \n\t"
"vmovapd 0 * 32(%%rax), %%ymm0 \n\t" // initialize loop by pre-loading
"vmovapd 0 * 32(%%rbx), %%ymm2 \n\t" // elements of a and b.
@@ -1090,6 +1091,7 @@ void bli_dgemm_asm_8x4(
" \n\t"
".DLOOPKITER: \n\t" // MAIN LOOP
" \n\t"
"addq $4 * 4 * 8, %%r15 \n\t" // b_next += 4*4 (unroll x nr)
" \n\t"
" \n\t" // iteration 0
"vmovapd 1 * 32(%%rax), %%ymm1 \n\t"
@@ -1113,6 +1115,7 @@ void bli_dgemm_asm_8x4(
"vmovapd 2 * 32(%%rax), %%ymm0 \n\t"
"vaddpd %%ymm11, %%ymm6, %%ymm11 \n\t"
"vaddpd %%ymm9, %%ymm7, %%ymm9 \n\t"
"prefetcht0 0 * 32(%%r15) \n\t" // prefetch b_next[0*4]
" \n\t"
"vmulpd %%ymm1, %%ymm4, %%ymm6 \n\t"
"vmulpd %%ymm1, %%ymm5, %%ymm7 \n\t"
@@ -1172,6 +1175,7 @@ void bli_dgemm_asm_8x4(
"vmovapd 6 * 32(%%rax), %%ymm0 \n\t"
"vaddpd %%ymm11, %%ymm6, %%ymm11 \n\t"
"vaddpd %%ymm9, %%ymm7, %%ymm9 \n\t"
"prefetcht0 2 * 32(%%r15) \n\t" // prefetch b_next[2*4]
" \n\t"
"vmulpd %%ymm1, %%ymm4, %%ymm6 \n\t"
"vmulpd %%ymm1, %%ymm5, %%ymm7 \n\t"
@@ -1699,8 +1703,8 @@ void bli_dgemm_asm_8x4(
"m" (beta), // 5
"m" (c), // 6
"m" (rs_c), // 7
"m" (cs_c)/*, // 8
"m" (b_next), // 9
"m" (cs_c), // 8
"m" (b_next)/*, // 9
"m" (a_next)*/ // 10
: // register clobber list
"rax", "rbx", "rcx", "rdx", "rsi", "rdi",
@@ -1724,7 +1728,7 @@ void bli_cgemm_asm_8x4(
)
{
//void* a_next = bli_auxinfo_next_a( data );
//void* b_next = bli_auxinfo_next_b( data );
void* b_next = bli_auxinfo_next_b( data );
dim_t k_iter = k / 4;
dim_t k_left = k % 4;
@@ -1735,8 +1739,9 @@ void bli_cgemm_asm_8x4(
" \n\t"
"movq %2, %%rax \n\t" // load address of a.
"movq %3, %%rbx \n\t" // load address of b.
//"movq %9, %%r15 \n\t" // load address of b_next.
"movq %9, %%r15 \n\t" // load address of b_next.
//"movq %10, %%r14 \n\t" // load address of a_next.
"addq $-4 * 64, %%r15 \n\t"
" \n\t"
"vmovaps 0 * 32(%%rax), %%ymm0 \n\t" // initialize loop by pre-loading
"vmovsldup 0 * 32(%%rbx), %%ymm2 \n\t"
@@ -1771,6 +1776,7 @@ void bli_cgemm_asm_8x4(
" \n\t"
".CLOOPKITER: \n\t" // MAIN LOOP
" \n\t"
"addq $4 * 4 * 8, %%r15 \n\t" // b_next += 4*4 (unroll x nr)
" \n\t"
" \n\t" // iteration 0
"prefetcht0 8 * 32(%%rax) \n\t"
@@ -1801,6 +1807,7 @@ void bli_cgemm_asm_8x4(
"vperm2f128 $0x3, %%ymm3, %%ymm3, %%ymm5 \n\t"
"vaddps %%ymm6, %%ymm10, %%ymm10 \n\t"
"vaddps %%ymm7, %%ymm8, %%ymm8 \n\t"
"prefetcht0 0 * 32(%%r15) \n\t" // prefetch b_next[0*4]
" \n\t"
"vpermilps $0xb1, %%ymm1, %%ymm1 \n\t"
"vmulps %%ymm0, %%ymm2, %%ymm6 \n\t"
@@ -1911,6 +1918,7 @@ void bli_cgemm_asm_8x4(
"vperm2f128 $0x3, %%ymm3, %%ymm3, %%ymm5 \n\t"
"vaddps %%ymm6, %%ymm10, %%ymm10 \n\t"
"vaddps %%ymm7, %%ymm8, %%ymm8 \n\t"
"prefetcht0 2 * 32(%%r15) \n\t" // prefetch b_next[2*4]
" \n\t"
"vpermilps $0xb1, %%ymm1, %%ymm1 \n\t"
"vmulps %%ymm0, %%ymm2, %%ymm6 \n\t"
@@ -2667,8 +2675,8 @@ void bli_cgemm_asm_8x4(
"m" (beta), // 5
"m" (c), // 6
"m" (rs_c), // 7
"m" (cs_c)/*, // 8
"m" (b_next), // 9
"m" (cs_c), // 8
"m" (b_next)/*, // 9
"m" (a_next)*/ // 10
: // register clobber list
"rax", "rbx", "rcx", "rdx", "rsi", "rdi",