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

Fix inline assembly constraints 

rework indices to allow marking argument lda as input and output.
This commit is contained in:
Martin Kroeker 2019-02-16 18:36:39 +01:00 committed by GitHub
parent 9d8be15789
commit e976557d29
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 69 additions and 69 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

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",