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

Fixed ldim alignment bug in core2 gemm ukernel.

Browse Source 
Details:
- Fixed a bug in the dunnington/core2 gemm micro-kernels that resulted in
  a segmentation fault if a column-stored matrix's starting address was
  aligned, but its leading dimension was such that its second column was
  unaligned. Basically, the micro-kernel was assuming that aligned load
  instructions were safe when they actually were not. An extra condition
  that checks the alignment of cs_c (ie: the leading dimension in the
  column storage case) has now been added. Thanks to Michael Lehn for
  reporting this bug.
This commit is contained in:
Field G. Van Zee
2014-05-21 11:34:42 -05:00
parent 77a2d8dac8
commit 3fc60e4914
1 changed files with 26 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
kernels/x86_64/core2-sse3/3/bli_gemm_opt_d4x4.c
View File

@@ -68,6 +68,7 @@ void bli_sgemm_opt_8x4(
"movq %6, %%rcx \n\t" // load address of c
"movq %8, %%rdi \n\t" // load cs_c
"leaq (,%%rdi,4), %%rdi \n\t" // cs_c *= sizeof(float)
"movq %%rdi, %%r12 \n\t" // make a copy of cs_c (in bytes)
"leaq (%%rcx,%%rdi,2), %%r10 \n\t" // load address of c + 2*cs_c;
" \n\t"
"prefetcht2 0 * 4(%%r9) \n\t" // prefetch b_next
@@ -364,15 +365,20 @@ void bli_sgemm_opt_8x4(
" \n\t"
" \n\t"
" \n\t" // determine if
" \n\t" // c % 16 == 0, AND
" \n\t" // rs_c == 1
" \n\t" // ie: aligned and column-stored
" \n\t" // c % 16 == 0, AND
" \n\t" // 8*cs_c % 16 == 0, AND
" \n\t" // rs_c == 1
" \n\t" // ie: aligned, ldim aligned, and
" \n\t" // column-stored
" \n\t"
"cmpq $1, %%r8 \n\t" // set ZF if rs_c == 1.
"sete %%bl \n\t" // bl = ( ZF == 1 ? 1 : 0 );
"testq $15, %%rcx \n\t" // set ZF if c & 16 is zero.
"setz %%bh \n\t" // bh = ( ZF == 1 ? 1 : 0 );
" \n\t" // and(bl,bh) will reveal result
"testq $15, %%r12 \n\t" // set ZF if (4*cs_c) & 16 is zero.
"setz %%al \n\t" // al = ( ZF == 1 ? 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"
@@ -383,6 +389,7 @@ void bli_sgemm_opt_8x4(
" \n\t"
" \n\t" // check if aligned/column-stored
"andb %%bl, %%bh \n\t" // set ZF if bl & bh == 1.
"andb %%bh, %%al \n\t" // set ZF if bh & al == 1.
"jne .SCOLSTORED \n\t" // jump to column storage case
" \n\t"
" \n\t"
@@ -632,6 +639,7 @@ void bli_sgemm_opt_8x4(
".SBETAZERO: \n\t"
" \n\t" // check if aligned/column-stored
"andb %%bl, %%bh \n\t" // set ZF if bl & bh == 1.
"andb %%bh, %%al \n\t" // set ZF if bh & al == 1.
"jne .SCOLSTORBZ \n\t" // jump to column storage case
" \n\t"
" \n\t"
@@ -706,7 +714,7 @@ void bli_sgemm_opt_8x4(
"m" (cs_c),
"m" (b_next)
: // register clobber list
"rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
"rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11", "r12",
"xmm0", "xmm1", "xmm2", "xmm3",
"xmm4", "xmm5", "xmm6", "xmm7",
"xmm8", "xmm9", "xmm10", "xmm11",
@@ -750,6 +758,7 @@ void bli_dgemm_opt_4x4(
"movq %6, %%rcx \n\t" // load address of c
"movq %8, %%rdi \n\t" // load cs_c
"leaq (,%%rdi,8), %%rdi \n\t" // cs_c *= sizeof(double)
"movq %%rdi, %%r12 \n\t" // make a copy of cs_c (in bytes)
"leaq (%%rcx,%%rdi,2), %%r10 \n\t" // load address of c + 2*cs_c;
" \n\t"
"prefetcht2 0 * 8(%%r9) \n\t" // prefetch b_next
@@ -1043,15 +1052,20 @@ void bli_dgemm_opt_4x4(
" \n\t"
" \n\t"
" \n\t" // determine if
" \n\t" // c % 16 == 0, AND
" \n\t" // rs_c == 1
" \n\t" // ie: aligned and column-stored
" \n\t" // c % 16 == 0, AND
" \n\t" // 8*cs_c % 16 == 0, AND
" \n\t" // rs_c == 1
" \n\t" // ie: aligned, ldim aligned, and
" \n\t" // column-stored
" \n\t"
"cmpq $1, %%r8 \n\t" // set ZF if rs_c == 1.
"sete %%bl \n\t" // bl = ( ZF == 1 ? 1 : 0 );
"testq $15, %%rcx \n\t" // set ZF if c & 16 is zero.
"setz %%bh \n\t" // bh = ( ZF == 1 ? 1 : 0 );
" \n\t" // and(bl,bh) will reveal result
"testq $15, %%r12 \n\t" // set ZF if (8*cs_c) & 16 is zero.
"setz %%al \n\t" // al = ( ZF == 1 ? 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"
@@ -1062,6 +1076,7 @@ void bli_dgemm_opt_4x4(
" \n\t"
" \n\t" // check if aligned/column-stored
"andb %%bl, %%bh \n\t" // set ZF if bl & bh == 1.
"andb %%bh, %%al \n\t" // set ZF if bh & al == 1.
"jne .DCOLSTORED \n\t" // jump to column storage case
" \n\t"
" \n\t"
@@ -1220,6 +1235,7 @@ void bli_dgemm_opt_4x4(
".DBETAZERO: \n\t"
" \n\t" // check if aligned/column-stored
"andb %%bl, %%bh \n\t" // set ZF if bl & bh == 1.
"andb %%bh, %%al \n\t" // set ZF if bh & al == 1.
"jne .DCOLSTORBZ \n\t" // jump to column storage case
" \n\t"
" \n\t"
@@ -1339,7 +1355,7 @@ void bli_dgemm_opt_4x4(
"m" (b_next), // 9
"m" (a_next) // 10
: // register clobber list
"rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11",
"rax", "rbx", "rcx", "rdx", "rsi", "rdi", "r8", "r9", "r10", "r11", "r12",
"xmm0", "xmm1", "xmm2", "xmm3",
"xmm4", "xmm5", "xmm6", "xmm7",
"xmm8", "xmm9", "xmm10", "xmm11",