floraachy/OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #2023 from martin-frbg/gcc9fixes3

Browse Source 
Fix inline assembly constraints in various x86_64 GEMVN kernels
This commit is contained in:
Martin Kroeker
2019-02-17 11:48:57 +01:00
committed by GitHub
parent aec905498f 8242b1fe3f
commit 1860c9456d
4 changed files with 437 additions and 190 deletions
Download Patch File Download Diff File Expand all files Collapse all files

188
kernel/x86_64/sgemv_n_microk_bulldozer-4.c
View File

@@ -37,14 +37,14 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
__asm__ __volatile__
(
"vbroadcastss (%2), %%xmm12 \n\t" // x0
"vbroadcastss 4(%2), %%xmm13 \n\t" // x1
"vbroadcastss 8(%2), %%xmm14 \n\t" // x2
"vbroadcastss 12(%2), %%xmm15 \n\t" // x3
"vbroadcastss 16(%2), %%xmm0 \n\t" // x4
"vbroadcastss 20(%2), %%xmm1 \n\t" // x5
"vbroadcastss 24(%2), %%xmm2 \n\t" // x6
"vbroadcastss 28(%2), %%xmm3 \n\t" // x7
"vbroadcastss (%3), %%xmm12 \n\t" // x0
"vbroadcastss 4(%3), %%xmm13 \n\t" // x1
"vbroadcastss 8(%3), %%xmm14 \n\t" // x2
"vbroadcastss 12(%3), %%xmm15 \n\t" // x3
"vbroadcastss 16(%3), %%xmm0 \n\t" // x4
"vbroadcastss 20(%3), %%xmm1 \n\t" // x5
"vbroadcastss 24(%3), %%xmm2 \n\t" // x6
"vbroadcastss 28(%3), %%xmm3 \n\t" // x7
"vbroadcastss (%9), %%xmm8 \n\t" // alpha
@@ -54,22 +54,22 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vxorps %%xmm4, %%xmm4 , %%xmm4 \n\t"
"vxorps %%xmm5, %%xmm5 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%4,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmaddps %%xmm5, (%5,%0,4), %%xmm13, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%6,%0,4), %%xmm14, %%xmm4 \n\t"
"vfmaddps %%xmm5, (%7,%0,4), %%xmm15, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%5,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmaddps %%xmm5, (%6,%0,4), %%xmm13, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%7,%0,4), %%xmm14, %%xmm4 \n\t"
"vfmaddps %%xmm5, (%8,%0,4), %%xmm15, %%xmm5 \n\t"
"addq $4 , %0 \n\t"
"vfmaddps %%xmm4, (%4,%8,4), %%xmm0 , %%xmm4 \n\t"
"vfmaddps %%xmm5, (%5,%8,4), %%xmm1 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%6,%8,4), %%xmm2 , %%xmm4 \n\t"
"vfmaddps %%xmm5, (%7,%8,4), %%xmm3 , %%xmm5 \n\t"
"addq $4 , %8 \n\t"
"vfmaddps %%xmm4, (%5,%2,4), %%xmm0 , %%xmm4 \n\t"
"vfmaddps %%xmm5, (%6,%2,4), %%xmm1 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%7,%2,4), %%xmm2 , %%xmm4 \n\t"
"vfmaddps %%xmm5, (%8,%2,4), %%xmm3 , %%xmm5 \n\t"
"addq $4 , %2 \n\t"
"vaddps %%xmm5 , %%xmm4, %%xmm4 \n\t"
"vfmaddps -16(%3,%0,4) , %%xmm4, %%xmm8,%%xmm6 \n\t"
"vfmaddps -16(%4,%0,4) , %%xmm4, %%xmm8,%%xmm6 \n\t"
"subq $4 , %1 \n\t"
"vmovups %%xmm6, -16(%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm6, -16(%4,%0,4) \n\t" // 4 * y
"2: \n\t"
@@ -79,31 +79,31 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vxorps %%xmm4, %%xmm4 , %%xmm4 \n\t"
"vxorps %%xmm5, %%xmm5 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%4,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%4,%0,4), %%xmm12, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%5,%0,4), %%xmm13, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm13, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%6,%0,4), %%xmm14, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%6,%0,4), %%xmm14, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%7,%0,4), %%xmm15, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%7,%0,4), %%xmm15, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%5,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm12, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%6,%0,4), %%xmm13, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%6,%0,4), %%xmm13, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%7,%0,4), %%xmm14, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%7,%0,4), %%xmm14, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%8,%0,4), %%xmm15, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%8,%0,4), %%xmm15, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%4,%8,4), %%xmm0 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%4,%8,4), %%xmm0 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%5,%8,4), %%xmm1 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%8,4), %%xmm1 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%6,%8,4), %%xmm2 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%6,%8,4), %%xmm2 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%7,%8,4), %%xmm3 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%7,%8,4), %%xmm3 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%5,%2,4), %%xmm0 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%2,4), %%xmm0 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%6,%2,4), %%xmm1 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%6,%2,4), %%xmm1 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%7,%2,4), %%xmm2 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%7,%2,4), %%xmm2 , %%xmm5 \n\t"
"vfmaddps %%xmm4, (%8,%2,4), %%xmm3 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%8,%2,4), %%xmm3 , %%xmm5 \n\t"
"vfmaddps (%3,%0,4) , %%xmm4,%%xmm8,%%xmm4 \n\t"
"vfmaddps 16(%3,%0,4) , %%xmm5,%%xmm8,%%xmm5 \n\t"
"vmovups %%xmm4, (%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm5, 16(%3,%0,4) \n\t" // 4 * y
"vfmaddps (%4,%0,4) , %%xmm4,%%xmm8,%%xmm4 \n\t"
"vfmaddps 16(%4,%0,4) , %%xmm5,%%xmm8,%%xmm5 \n\t"
"vmovups %%xmm4, (%4,%0,4) \n\t" // 4 * y
"vmovups %%xmm5, 16(%4,%0,4) \n\t" // 4 * y
"addq $8 , %0 \n\t"
"addq $8 , %8 \n\t"
"addq $8 , %2 \n\t"
"subq $8 , %1 \n\t"
@@ -120,62 +120,62 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vxorps %%xmm6, %%xmm6 , %%xmm6 \n\t"
"vxorps %%xmm7, %%xmm7 , %%xmm7 \n\t"
"prefetcht0 192(%4,%0,4) \n\t"
"vfmaddps %%xmm4, (%4,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%4,%0,4), %%xmm12, %%xmm5 \n\t"
"prefetcht0 192(%5,%0,4) \n\t"
"vfmaddps %%xmm4, (%5,%0,4), %%xmm13, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm13, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%5,%0,4), %%xmm12, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm12, %%xmm5 \n\t"
"prefetcht0 192(%6,%0,4) \n\t"
"vfmaddps %%xmm4, (%6,%0,4), %%xmm14, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%6,%0,4), %%xmm14, %%xmm5 \n\t"
"vfmaddps %%xmm4, (%6,%0,4), %%xmm13, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%6,%0,4), %%xmm13, %%xmm5 \n\t"
"prefetcht0 192(%7,%0,4) \n\t"
"vfmaddps %%xmm4, (%7,%0,4), %%xmm15, %%xmm4 \n\t"
"vfmaddps %%xmm4, (%7,%0,4), %%xmm14, %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%7,%0,4), %%xmm14, %%xmm5 \n\t"
"prefetcht0 192(%8,%0,4) \n\t"
"vfmaddps %%xmm4, (%8,%0,4), %%xmm15, %%xmm4 \n\t"
".align 2 \n\t"
"vfmaddps %%xmm5, 16(%7,%0,4), %%xmm15, %%xmm5 \n\t"
"vfmaddps %%xmm5, 16(%8,%0,4), %%xmm15, %%xmm5 \n\t"
"vfmaddps %%xmm6, 32(%4,%0,4), %%xmm12, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%4,%0,4), %%xmm12, %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%5,%0,4), %%xmm13, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%5,%0,4), %%xmm13, %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%6,%0,4), %%xmm14, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%6,%0,4), %%xmm14, %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%7,%0,4), %%xmm15, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%7,%0,4), %%xmm15, %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%5,%0,4), %%xmm12, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%5,%0,4), %%xmm12, %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%6,%0,4), %%xmm13, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%6,%0,4), %%xmm13, %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%7,%0,4), %%xmm14, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%7,%0,4), %%xmm14, %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%8,%0,4), %%xmm15, %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%8,%0,4), %%xmm15, %%xmm7 \n\t"
"prefetcht0 192(%4,%8,4) \n\t"
"vfmaddps %%xmm4, (%4,%8,4), %%xmm0 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%4,%8,4), %%xmm0 , %%xmm5 \n\t"
"prefetcht0 192(%5,%8,4) \n\t"
"vfmaddps %%xmm4, (%5,%8,4), %%xmm1 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%8,4), %%xmm1 , %%xmm5 \n\t"
"prefetcht0 192(%6,%8,4) \n\t"
"vfmaddps %%xmm4, (%6,%8,4), %%xmm2 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%6,%8,4), %%xmm2 , %%xmm5 \n\t"
"prefetcht0 192(%7,%8,4) \n\t"
"vfmaddps %%xmm4, (%7,%8,4), %%xmm3 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%7,%8,4), %%xmm3 , %%xmm5 \n\t"
"prefetcht0 192(%5,%2,4) \n\t"
"vfmaddps %%xmm4, (%5,%2,4), %%xmm0 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%5,%2,4), %%xmm0 , %%xmm5 \n\t"
"prefetcht0 192(%6,%2,4) \n\t"
"vfmaddps %%xmm4, (%6,%2,4), %%xmm1 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%6,%2,4), %%xmm1 , %%xmm5 \n\t"
"prefetcht0 192(%7,%2,4) \n\t"
"vfmaddps %%xmm4, (%7,%2,4), %%xmm2 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%7,%2,4), %%xmm2 , %%xmm5 \n\t"
"prefetcht0 192(%8,%2,4) \n\t"
"vfmaddps %%xmm4, (%8,%2,4), %%xmm3 , %%xmm4 \n\t"
"vfmaddps %%xmm5, 16(%8,%2,4), %%xmm3 , %%xmm5 \n\t"
"vfmaddps %%xmm6, 32(%4,%8,4), %%xmm0 , %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%4,%8,4), %%xmm0 , %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%5,%8,4), %%xmm1 , %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%5,%8,4), %%xmm1 , %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%6,%8,4), %%xmm2 , %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%6,%8,4), %%xmm2 , %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%7,%8,4), %%xmm3 , %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%7,%8,4), %%xmm3 , %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%5,%2,4), %%xmm0 , %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%5,%2,4), %%xmm0 , %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%6,%2,4), %%xmm1 , %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%6,%2,4), %%xmm1 , %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%7,%2,4), %%xmm2 , %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%7,%2,4), %%xmm2 , %%xmm7 \n\t"
"vfmaddps %%xmm6, 32(%8,%2,4), %%xmm3 , %%xmm6 \n\t"
"vfmaddps %%xmm7, 48(%8,%2,4), %%xmm3 , %%xmm7 \n\t"
"vfmaddps (%3,%0,4) , %%xmm4,%%xmm8,%%xmm4 \n\t"
"vfmaddps 16(%3,%0,4) , %%xmm5,%%xmm8,%%xmm5 \n\t"
"vfmaddps 32(%3,%0,4) , %%xmm6,%%xmm8,%%xmm6 \n\t"
"vfmaddps 48(%3,%0,4) , %%xmm7,%%xmm8,%%xmm7 \n\t"
"vfmaddps (%4,%0,4) , %%xmm4,%%xmm8,%%xmm4 \n\t"
"vfmaddps 16(%4,%0,4) , %%xmm5,%%xmm8,%%xmm5 \n\t"
"vfmaddps 32(%4,%0,4) , %%xmm6,%%xmm8,%%xmm6 \n\t"
"vfmaddps 48(%4,%0,4) , %%xmm7,%%xmm8,%%xmm7 \n\t"
"addq $16, %0 \n\t"
"vmovups %%xmm4,-64(%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm5,-48(%3,%0,4) \n\t" // 4 * y
"addq $16, %8 \n\t"
"vmovups %%xmm6,-32(%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm7,-16(%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm4,-64(%4,%0,4) \n\t" // 4 * y
"vmovups %%xmm5,-48(%4,%0,4) \n\t" // 4 * y
"addq $16, %2 \n\t"
"vmovups %%xmm6,-32(%4,%0,4) \n\t" // 4 * y
"vmovups %%xmm7,-16(%4,%0,4) \n\t" // 4 * y
"subq $16, %1 \n\t"
"jnz 1b \n\t"
@@ -184,15 +184,15 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
:
"+r" (i), // 0
"+r" (n) // 1
"+r" (n), // 1
"+r" (lda4) // 2
:
"r" (x), // 2
"r" (y), // 3
"r" (ap[0]), // 4
"r" (ap[1]), // 5
"r" (ap[2]), // 6
"r" (ap[3]), // 7
"r" (lda4), // 8
"r" (x), // 3
"r" (y), // 4
"r" (ap[0]), // 5
"r" (ap[1]), // 6
"r" (ap[2]), // 7
"r" (ap[3]), // 8
"r" (alpha) // 9
: "cc",
"%xmm0", "%xmm1",

54
kernel/x86_64/sgemv_n_microk_nehalem-4.c
View File

@@ -37,19 +37,19 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
__asm__ __volatile__
(
"movss (%2), %%xmm12 \n\t" // x0
"movss 4(%2), %%xmm13 \n\t" // x1
"movss 8(%2), %%xmm14 \n\t" // x2
"movss 12(%2), %%xmm15 \n\t" // x3
"movss (%3), %%xmm12 \n\t" // x0
"movss 4(%3), %%xmm13 \n\t" // x1
"movss 8(%3), %%xmm14 \n\t" // x2
"movss 12(%3), %%xmm15 \n\t" // x3
"shufps $0, %%xmm12, %%xmm12\n\t"
"shufps $0, %%xmm13, %%xmm13\n\t"
"shufps $0, %%xmm14, %%xmm14\n\t"
"shufps $0, %%xmm15, %%xmm15\n\t"
"movss 16(%2), %%xmm0 \n\t" // x4
"movss 20(%2), %%xmm1 \n\t" // x5
"movss 24(%2), %%xmm2 \n\t" // x6
"movss 28(%2), %%xmm3 \n\t" // x7
"movss 16(%3), %%xmm0 \n\t" // x4
"movss 20(%3), %%xmm1 \n\t" // x5
"movss 24(%3), %%xmm2 \n\t" // x6
"movss 28(%3), %%xmm3 \n\t" // x7
"shufps $0, %%xmm0 , %%xmm0 \n\t"
"shufps $0, %%xmm1 , %%xmm1 \n\t"
"shufps $0, %%xmm2 , %%xmm2 \n\t"
@@ -63,13 +63,13 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"1: \n\t"
"xorps %%xmm4 , %%xmm4 \n\t"
"xorps %%xmm5 , %%xmm5 \n\t"
"movups (%3,%0,4), %%xmm7 \n\t" // 4 * y
"movups (%4,%0,4), %%xmm7 \n\t" // 4 * y
".p2align 1 \n\t"
"movups (%4,%0,4), %%xmm8 \n\t"
"movups (%5,%0,4), %%xmm9 \n\t"
"movups (%6,%0,4), %%xmm10 \n\t"
"movups (%7,%0,4), %%xmm11 \n\t"
"movups (%5,%0,4), %%xmm8 \n\t"
"movups (%6,%0,4), %%xmm9 \n\t"
"movups (%7,%0,4), %%xmm10 \n\t"
"movups (%8,%0,4), %%xmm11 \n\t"
".p2align 1 \n\t"
"mulps %%xmm12, %%xmm8 \n\t"
"mulps %%xmm13, %%xmm9 \n\t"
@@ -80,10 +80,10 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"addps %%xmm10, %%xmm4 \n\t"
"addps %%xmm11, %%xmm5 \n\t"
"movups (%4,%8,4), %%xmm8 \n\t"
"movups (%5,%8,4), %%xmm9 \n\t"
"movups (%6,%8,4), %%xmm10 \n\t"
"movups (%7,%8,4), %%xmm11 \n\t"
"movups (%5,%2,4), %%xmm8 \n\t"
"movups (%6,%2,4), %%xmm9 \n\t"
"movups (%7,%2,4), %%xmm10 \n\t"
"movups (%8,%2,4), %%xmm11 \n\t"
".p2align 1 \n\t"
"mulps %%xmm0 , %%xmm8 \n\t"
"mulps %%xmm1 , %%xmm9 \n\t"
@@ -94,28 +94,28 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"addps %%xmm10, %%xmm4 \n\t"
"addps %%xmm11, %%xmm5 \n\t"
"addq $4 , %8 \n\t"
"addq $4 , %2 \n\t"
"addps %%xmm5 , %%xmm4 \n\t"
"addq $4 , %0 \n\t"
"mulps %%xmm6 , %%xmm4 \n\t"
"subq $4 , %1 \n\t"
"addps %%xmm4 , %%xmm7 \n\t"
"movups %%xmm7 , -16(%3,%0,4) \n\t" // 4 * y
"movups %%xmm7 , -16(%4,%0,4) \n\t" // 4 * y
"jnz 1b \n\t"
:
"+r" (i), // 0
"+r" (n) // 1
"+r" (n), // 1
"+r" (lda4) // 2
:
"r" (x), // 2
"r" (y), // 3
"r" (ap[0]), // 4
"r" (ap[1]), // 5
"r" (ap[2]), // 6
"r" (ap[3]), // 7
"r" (lda4), // 8
"r" (x), // 3
"r" (y), // 4
"r" (ap[0]), // 5
"r" (ap[1]), // 6
"r" (ap[2]), // 7
"r" (ap[3]), // 8
"r" (alpha) // 9
: "cc",
"%xmm0", "%xmm1",

138
kernel/x86_64/sgemv_n_microk_sandy-4.c
View File

@@ -39,14 +39,14 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
__asm__ __volatile__
(
"vzeroupper \n\t"
"vbroadcastss (%2), %%ymm12 \n\t" // x0
"vbroadcastss 4(%2), %%ymm13 \n\t" // x1
"vbroadcastss 8(%2), %%ymm14 \n\t" // x2
"vbroadcastss 12(%2), %%ymm15 \n\t" // x3
"vbroadcastss 16(%2), %%ymm0 \n\t" // x4
"vbroadcastss 20(%2), %%ymm1 \n\t" // x5
"vbroadcastss 24(%2), %%ymm2 \n\t" // x6
"vbroadcastss 28(%2), %%ymm3 \n\t" // x7
"vbroadcastss (%3), %%ymm12 \n\t" // x0
"vbroadcastss 4(%3), %%ymm13 \n\t" // x1
"vbroadcastss 8(%3), %%ymm14 \n\t" // x2
"vbroadcastss 12(%3), %%ymm15 \n\t" // x3
"vbroadcastss 16(%3), %%ymm0 \n\t" // x4
"vbroadcastss 20(%3), %%ymm1 \n\t" // x5
"vbroadcastss 24(%3), %%ymm2 \n\t" // x6
"vbroadcastss 28(%3), %%ymm3 \n\t" // x7
"vbroadcastss (%9), %%ymm6 \n\t" // alpha
@@ -55,21 +55,21 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vxorps %%xmm4 , %%xmm4 , %%xmm4 \n\t"
"vxorps %%xmm5 , %%xmm5 , %%xmm5 \n\t"
"vmovups (%3,%0,4), %%xmm7 \n\t" // 4 * y
"vmovups (%4,%0,4), %%xmm7 \n\t" // 4 * y
"vmulps (%4,%0,4), %%xmm12, %%xmm8 \n\t"
"vmulps (%5,%0,4), %%xmm13, %%xmm10 \n\t"
"vmulps (%6,%0,4), %%xmm14, %%xmm9 \n\t"
"vmulps (%7,%0,4), %%xmm15, %%xmm11 \n\t"
"vmulps (%5,%0,4), %%xmm12, %%xmm8 \n\t"
"vmulps (%6,%0,4), %%xmm13, %%xmm10 \n\t"
"vmulps (%7,%0,4), %%xmm14, %%xmm9 \n\t"
"vmulps (%8,%0,4), %%xmm15, %%xmm11 \n\t"
"vaddps %%xmm4, %%xmm8 , %%xmm4 \n\t"
"vaddps %%xmm5, %%xmm10, %%xmm5 \n\t"
"vaddps %%xmm4, %%xmm9 , %%xmm4 \n\t"
"vaddps %%xmm5, %%xmm11, %%xmm5 \n\t"
"vmulps (%4,%8,4), %%xmm0 , %%xmm8 \n\t"
"vmulps (%5,%8,4), %%xmm1 , %%xmm10 \n\t"
"vmulps (%6,%8,4), %%xmm2 , %%xmm9 \n\t"
"vmulps (%7,%8,4), %%xmm3 , %%xmm11 \n\t"
"vmulps (%5,%2,4), %%xmm0 , %%xmm8 \n\t"
"vmulps (%6,%2,4), %%xmm1 , %%xmm10 \n\t"
"vmulps (%7,%2,4), %%xmm2 , %%xmm9 \n\t"
"vmulps (%8,%2,4), %%xmm3 , %%xmm11 \n\t"
"vaddps %%xmm4, %%xmm8 , %%xmm4 \n\t"
"vaddps %%xmm5, %%xmm10, %%xmm5 \n\t"
"vaddps %%xmm4, %%xmm9 , %%xmm4 \n\t"
@@ -79,9 +79,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vmulps %%xmm6, %%xmm4 , %%xmm5 \n\t"
"vaddps %%xmm5, %%xmm7 , %%xmm5 \n\t"
"vmovups %%xmm5, (%3,%0,4) \n\t" // 4 * y
"vmovups %%xmm5, (%4,%0,4) \n\t" // 4 * y
"addq $4, %8 \n\t"
"addq $4, %2 \n\t"
"addq $4, %0 \n\t"
"subq $4, %1 \n\t"
@@ -92,21 +92,21 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vxorps %%ymm4 , %%ymm4 , %%ymm4 \n\t"
"vxorps %%ymm5 , %%ymm5 , %%ymm5 \n\t"
"vmovups (%3,%0,4), %%ymm7 \n\t" // 8 * y
"vmovups (%4,%0,4), %%ymm7 \n\t" // 8 * y
"vmulps (%4,%0,4), %%ymm12, %%ymm8 \n\t"
"vmulps (%5,%0,4), %%ymm13, %%ymm10 \n\t"
"vmulps (%6,%0,4), %%ymm14, %%ymm9 \n\t"
"vmulps (%7,%0,4), %%ymm15, %%ymm11 \n\t"
"vmulps (%5,%0,4), %%ymm12, %%ymm8 \n\t"
"vmulps (%6,%0,4), %%ymm13, %%ymm10 \n\t"
"vmulps (%7,%0,4), %%ymm14, %%ymm9 \n\t"
"vmulps (%8,%0,4), %%ymm15, %%ymm11 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm10, %%ymm5 \n\t"
"vaddps %%ymm4, %%ymm9 , %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
"vmulps (%4,%8,4), %%ymm0 , %%ymm8 \n\t"
"vmulps (%5,%8,4), %%ymm1 , %%ymm10 \n\t"
"vmulps (%6,%8,4), %%ymm2 , %%ymm9 \n\t"
"vmulps (%7,%8,4), %%ymm3 , %%ymm11 \n\t"
"vmulps (%5,%2,4), %%ymm0 , %%ymm8 \n\t"
"vmulps (%6,%2,4), %%ymm1 , %%ymm10 \n\t"
"vmulps (%7,%2,4), %%ymm2 , %%ymm9 \n\t"
"vmulps (%8,%2,4), %%ymm3 , %%ymm11 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm10, %%ymm5 \n\t"
"vaddps %%ymm4, %%ymm9 , %%ymm4 \n\t"
@@ -116,9 +116,9 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vmulps %%ymm6, %%ymm4 , %%ymm5 \n\t"
"vaddps %%ymm5, %%ymm7 , %%ymm5 \n\t"
"vmovups %%ymm5, (%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm5, (%4,%0,4) \n\t" // 8 * y
"addq $8, %8 \n\t"
"addq $8, %2 \n\t"
"addq $8, %0 \n\t"
"subq $8, %1 \n\t"
@@ -134,45 +134,45 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vxorps %%ymm4 , %%ymm4 , %%ymm4 \n\t"
"vxorps %%ymm5 , %%ymm5 , %%ymm5 \n\t"
"prefetcht0 192(%4,%0,4) \n\t"
"vmulps (%4,%0,4), %%ymm12, %%ymm8 \n\t"
"vmulps 32(%4,%0,4), %%ymm12, %%ymm9 \n\t"
"prefetcht0 192(%5,%0,4) \n\t"
"vmulps (%5,%0,4), %%ymm13, %%ymm10 \n\t"
"vmulps 32(%5,%0,4), %%ymm13, %%ymm11 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
"vmulps (%5,%0,4), %%ymm12, %%ymm8 \n\t"
"vmulps 32(%5,%0,4), %%ymm12, %%ymm9 \n\t"
"prefetcht0 192(%6,%0,4) \n\t"
"vmulps (%6,%0,4), %%ymm14, %%ymm8 \n\t"
"vmulps 32(%6,%0,4), %%ymm14, %%ymm9 \n\t"
"vmulps (%6,%0,4), %%ymm13, %%ymm10 \n\t"
"vmulps 32(%6,%0,4), %%ymm13, %%ymm11 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
"prefetcht0 192(%7,%0,4) \n\t"
"vmulps (%7,%0,4), %%ymm15, %%ymm10 \n\t"
"vmulps 32(%7,%0,4), %%ymm15, %%ymm11 \n\t"
"vmulps (%7,%0,4), %%ymm14, %%ymm8 \n\t"
"vmulps 32(%7,%0,4), %%ymm14, %%ymm9 \n\t"
"prefetcht0 192(%8,%0,4) \n\t"
"vmulps (%8,%0,4), %%ymm15, %%ymm10 \n\t"
"vmulps 32(%8,%0,4), %%ymm15, %%ymm11 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
"prefetcht0 192(%4,%8,4) \n\t"
"vmulps (%4,%8,4), %%ymm0 , %%ymm8 \n\t"
"vmulps 32(%4,%8,4), %%ymm0 , %%ymm9 \n\t"
"prefetcht0 192(%5,%8,4) \n\t"
"vmulps (%5,%8,4), %%ymm1 , %%ymm10 \n\t"
"vmulps 32(%5,%8,4), %%ymm1 , %%ymm11 \n\t"
"prefetcht0 192(%5,%2,4) \n\t"
"vmulps (%5,%2,4), %%ymm0 , %%ymm8 \n\t"
"vmulps 32(%5,%2,4), %%ymm0 , %%ymm9 \n\t"
"prefetcht0 192(%6,%2,4) \n\t"
"vmulps (%6,%2,4), %%ymm1 , %%ymm10 \n\t"
"vmulps 32(%6,%2,4), %%ymm1 , %%ymm11 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm11, %%ymm5 \n\t"
"prefetcht0 192(%6,%8,4) \n\t"
"vmulps (%6,%8,4), %%ymm2 , %%ymm8 \n\t"
"vmulps 32(%6,%8,4), %%ymm2 , %%ymm9 \n\t"
"prefetcht0 192(%7,%8,4) \n\t"
"vmulps (%7,%8,4), %%ymm3 , %%ymm10 \n\t"
"vmulps 32(%7,%8,4), %%ymm3 , %%ymm11 \n\t"
"prefetcht0 192(%7,%2,4) \n\t"
"vmulps (%7,%2,4), %%ymm2 , %%ymm8 \n\t"
"vmulps 32(%7,%2,4), %%ymm2 , %%ymm9 \n\t"
"prefetcht0 192(%8,%2,4) \n\t"
"vmulps (%8,%2,4), %%ymm3 , %%ymm10 \n\t"
"vmulps 32(%8,%2,4), %%ymm3 , %%ymm11 \n\t"
"vaddps %%ymm4, %%ymm8 , %%ymm4 \n\t"
"vaddps %%ymm5, %%ymm9 , %%ymm5 \n\t"
"vaddps %%ymm4, %%ymm10, %%ymm4 \n\t"
@@ -181,13 +181,13 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
"vmulps %%ymm6, %%ymm4 , %%ymm4 \n\t"
"vmulps %%ymm6, %%ymm5 , %%ymm5 \n\t"
"vaddps (%3,%0,4), %%ymm4 , %%ymm4 \n\t" // 8 * y
"vaddps 32(%3,%0,4), %%ymm5 , %%ymm5 \n\t" // 8 * y
"vaddps (%4,%0,4), %%ymm4 , %%ymm4 \n\t" // 8 * y
"vaddps 32(%4,%0,4), %%ymm5 , %%ymm5 \n\t" // 8 * y
"vmovups %%ymm4, (%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm5, 32(%3,%0,4) \n\t" // 8 * y
"vmovups %%ymm4, (%4,%0,4) \n\t" // 8 * y
"vmovups %%ymm5, 32(%4,%0,4) \n\t" // 8 * y
"addq $16, %8 \n\t"
"addq $16, %2 \n\t"
"addq $16, %0 \n\t"
"subq $16, %1 \n\t"
"jnz 1b \n\t"
@@ -197,15 +197,15 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
:
"+r" (i), // 0
"+r" (n) // 1
"+r" (n), // 1
"+r" (lda4) // 2
:
"r" (x), // 2
"r" (y), // 3
"r" (ap[0]), // 4
"r" (ap[1]), // 5
"r" (ap[2]), // 6
"r" (ap[3]), // 7
"r" (lda4), // 8
"r" (x), // 3
"r" (y), // 4
"r" (ap[0]), // 5
"r" (ap[1]), // 6
"r" (ap[2]), // 7
"r" (ap[3]), // 8
"r" (alpha) // 9
: "cc",
"%xmm0", "%xmm1",