mirror of
https://github.com/amd/blis.git
synced 2026-04-19 23:28:52 +00:00
Remove alignment restrictions on C in haswell kernel.
This commit is contained in:
@@ -935,23 +935,6 @@ void bli_dgemm_asm_6x8
|
|||||||
//"leaq (%%r13,%%rsi,4), %%r10 \n\t" // r10 = 7*cs_c;
|
//"leaq (%%r13,%%rsi,4), %%r10 \n\t" // r10 = 7*cs_c;
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
|
||||||
" \n\t" // determine if
|
|
||||||
" \n\t" // c % 32 == 0, AND
|
|
||||||
" \n\t" // 8*rs_c % 32 == 0, AND
|
|
||||||
" \n\t" // cs_c == 1
|
|
||||||
" \n\t" // ie: aligned, ldim aligned, and
|
|
||||||
" \n\t" // row-stored
|
|
||||||
" \n\t"
|
|
||||||
"cmpq $8, %%rsi \n\t" // set ZF if (8*cs_c) == 8.
|
|
||||||
"sete %%bl \n\t" // bl = ( ZF == 1 ? 1 : 0 );
|
|
||||||
"testq $31, %%rcx \n\t" // set ZF if c & 32 is zero.
|
|
||||||
"setz %%bh \n\t" // bh = ( ZF == 0 ? 1 : 0 );
|
|
||||||
"testq $31, %%rdi \n\t" // set ZF if (8*rs_c) & 32 is zero.
|
|
||||||
"setz %%al \n\t" // al = ( ZF == 0 ? 1 : 0 );
|
|
||||||
" \n\t" // and(bl,bh) followed by
|
|
||||||
" \n\t" // and(bh,al) will reveal result
|
|
||||||
" \n\t"
|
|
||||||
" \n\t" // now avoid loading C if beta == 0
|
" \n\t" // now avoid loading C if beta == 0
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vxorpd %%ymm0, %%ymm0, %%ymm0 \n\t" // set ymm0 to zero.
|
"vxorpd %%ymm0, %%ymm0, %%ymm0 \n\t" // set ymm0 to zero.
|
||||||
@@ -959,10 +942,8 @@ void bli_dgemm_asm_6x8
|
|||||||
"je .DBETAZERO \n\t" // if ZF = 1, jump to beta == 0 case
|
"je .DBETAZERO \n\t" // if ZF = 1, jump to beta == 0 case
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t" // check if aligned/row-stored
|
"cmpq $8, %%rsi \n\t" // set ZF if (8*cs_c) == 8.
|
||||||
"andb %%bl, %%bh \n\t" // set ZF if bl & bh == 1.
|
"jz .DROWSTORED \n\t" // jump to row storage case
|
||||||
"andb %%bh, %%al \n\t" // set ZF if bh & al == 1.
|
|
||||||
"jne .DROWSTORED \n\t" // jump to row storage case
|
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
@@ -1050,63 +1031,51 @@ void bli_dgemm_asm_6x8
|
|||||||
".DROWSTORED: \n\t"
|
".DROWSTORED: \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps (%%rcx), %%ymm0 \n\t"
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm4 \n\t"
|
||||||
"vfmadd213pd %%ymm4, %%ymm3, %%ymm0 \n\t"
|
"vmovups %%ymm4, (%%rcx) \n\t"
|
||||||
"vmovaps %%ymm0, (%%rcx) \n\t"
|
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps (%%rdx), %%ymm1 \n\t"
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm5 \n\t"
|
||||||
"vfmadd213pd %%ymm5, %%ymm3, %%ymm1 \n\t"
|
"vmovups %%ymm5, (%%rdx) \n\t"
|
||||||
"vmovaps %%ymm1, (%%rdx) \n\t"
|
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps (%%rcx), %%ymm0 \n\t"
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm6 \n\t"
|
||||||
"vfmadd213pd %%ymm6, %%ymm3, %%ymm0 \n\t"
|
"vmovups %%ymm6, (%%rcx) \n\t"
|
||||||
"vmovaps %%ymm0, (%%rcx) \n\t"
|
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps (%%rdx), %%ymm1 \n\t"
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm7 \n\t"
|
||||||
"vfmadd213pd %%ymm7, %%ymm3, %%ymm1 \n\t"
|
"vmovups %%ymm7, (%%rdx) \n\t"
|
||||||
"vmovaps %%ymm1, (%%rdx) \n\t"
|
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps (%%rcx), %%ymm0 \n\t"
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm8 \n\t"
|
||||||
"vfmadd213pd %%ymm8, %%ymm3, %%ymm0 \n\t"
|
"vmovups %%ymm8, (%%rcx) \n\t"
|
||||||
"vmovaps %%ymm0, (%%rcx) \n\t"
|
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps (%%rdx), %%ymm1 \n\t"
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm9 \n\t"
|
||||||
"vfmadd213pd %%ymm9, %%ymm3, %%ymm1 \n\t"
|
"vmovups %%ymm9, (%%rdx) \n\t"
|
||||||
"vmovaps %%ymm1, (%%rdx) \n\t"
|
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps (%%rcx), %%ymm0 \n\t"
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm10 \n\t"
|
||||||
"vfmadd213pd %%ymm10, %%ymm3, %%ymm0 \n\t"
|
"vmovups %%ymm10, (%%rcx) \n\t"
|
||||||
"vmovaps %%ymm0, (%%rcx) \n\t"
|
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps (%%rdx), %%ymm1 \n\t"
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm11 \n\t"
|
||||||
"vfmadd213pd %%ymm11, %%ymm3, %%ymm1 \n\t"
|
"vmovups %%ymm11, (%%rdx) \n\t"
|
||||||
"vmovaps %%ymm1, (%%rdx) \n\t"
|
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps (%%rcx), %%ymm0 \n\t"
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm12 \n\t"
|
||||||
"vfmadd213pd %%ymm12, %%ymm3, %%ymm0 \n\t"
|
"vmovups %%ymm12, (%%rcx) \n\t"
|
||||||
"vmovaps %%ymm0, (%%rcx) \n\t"
|
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps (%%rdx), %%ymm1 \n\t"
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm13 \n\t"
|
||||||
"vfmadd213pd %%ymm13, %%ymm3, %%ymm1 \n\t"
|
"vmovups %%ymm13, (%%rdx) \n\t"
|
||||||
"vmovaps %%ymm1, (%%rdx) \n\t"
|
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps (%%rcx), %%ymm0 \n\t"
|
"vfmadd231pd (%%rcx), %%ymm3, %%ymm14 \n\t"
|
||||||
"vfmadd213pd %%ymm14, %%ymm3, %%ymm0 \n\t"
|
"vmovups %%ymm14, (%%rcx) \n\t"
|
||||||
"vmovaps %%ymm0, (%%rcx) \n\t"
|
|
||||||
//"addq %%rdi, %%rcx \n\t"
|
//"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps (%%rdx), %%ymm1 \n\t"
|
"vfmadd231pd (%%rdx), %%ymm3, %%ymm15 \n\t"
|
||||||
"vfmadd213pd %%ymm15, %%ymm3, %%ymm1 \n\t"
|
"vmovups %%ymm15, (%%rdx) \n\t"
|
||||||
"vmovaps %%ymm1, (%%rdx) \n\t"
|
|
||||||
//"addq %%rdi, %%rdx \n\t"
|
//"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
@@ -1116,10 +1085,9 @@ void bli_dgemm_asm_6x8
|
|||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
".DBETAZERO: \n\t"
|
".DBETAZERO: \n\t"
|
||||||
" \n\t" // check if aligned/row-stored
|
" \n\t"
|
||||||
"andb %%bl, %%bh \n\t" // set ZF if bl & bh == 1.
|
"cmpq $8, %%rsi \n\t" // set ZF if (8*cs_c) == 8.
|
||||||
"andb %%bh, %%al \n\t" // set ZF if bh & al == 1.
|
"jz .DROWSTORBZ \n\t" // jump to row storage case
|
||||||
"jne .DROWSTORBZ \n\t" // jump to row storage case
|
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
@@ -1195,38 +1163,38 @@ void bli_dgemm_asm_6x8
|
|||||||
".DROWSTORBZ: \n\t"
|
".DROWSTORBZ: \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps %%ymm4, (%%rcx) \n\t"
|
"vmovups %%ymm4, (%%rcx) \n\t"
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps %%ymm5, (%%rdx) \n\t"
|
"vmovups %%ymm5, (%%rdx) \n\t"
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps %%ymm6, (%%rcx) \n\t"
|
"vmovups %%ymm6, (%%rcx) \n\t"
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps %%ymm7, (%%rdx) \n\t"
|
"vmovups %%ymm7, (%%rdx) \n\t"
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps %%ymm8, (%%rcx) \n\t"
|
"vmovups %%ymm8, (%%rcx) \n\t"
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps %%ymm9, (%%rdx) \n\t"
|
"vmovups %%ymm9, (%%rdx) \n\t"
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps %%ymm10, (%%rcx) \n\t"
|
"vmovups %%ymm10, (%%rcx) \n\t"
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps %%ymm11, (%%rdx) \n\t"
|
"vmovups %%ymm11, (%%rdx) \n\t"
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps %%ymm12, (%%rcx) \n\t"
|
"vmovups %%ymm12, (%%rcx) \n\t"
|
||||||
"addq %%rdi, %%rcx \n\t"
|
"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps %%ymm13, (%%rdx) \n\t"
|
"vmovups %%ymm13, (%%rdx) \n\t"
|
||||||
"addq %%rdi, %%rdx \n\t"
|
"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
"vmovaps %%ymm14, (%%rcx) \n\t"
|
"vmovups %%ymm14, (%%rcx) \n\t"
|
||||||
//"addq %%rdi, %%rcx \n\t"
|
//"addq %%rdi, %%rcx \n\t"
|
||||||
"vmovaps %%ymm15, (%%rdx) \n\t"
|
"vmovups %%ymm15, (%%rdx) \n\t"
|
||||||
//"addq %%rdi, %%rdx \n\t"
|
//"addq %%rdi, %%rdx \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
" \n\t"
|
" \n\t"
|
||||||
|
|||||||
Reference in New Issue
Block a user