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

Merge pull request #1440 from quickwritereader/develop 

small corrections
This commit is contained in:
Martin Kroeker 2018-01-31 23:48:47 +01:00 committed by GitHub
parent e388459a27 f653e7a18d
commit f45776ec1f
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
4 changed files with 13 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
kernel/zarch/dcopy.c
View File

@ -44,7 +44,7 @@ static void dcopy_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y) {
"brctg %[n_tmp],1b" "brctg %[n_tmp],1b"
: [mem_y] "=m" (*(double (*)[n])y), [n_tmp] "+&r"(n) : [mem_y] "=m" (*(double (*)[n])y), [n_tmp] "+&r"(n)
: [mem_x] "m" (*(const double (*)[n])x), : [mem_x] "m" (*(const double (*)[n])x),
[ptr_x] "a"(x), [ptr_y] "a"(y) [ptr_x] "+&a"(x), [ptr_y] "+&a"(y)
: "cc" : "cc"
); );
return; return;

4
kernel/zarch/dswap.c
View File

@ -139,7 +139,7 @@ static void dswap_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y)
#else #else
static void __attribute__ ((noinline)) dswap_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y) static void dswap_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y)
{ {
__asm__ volatile( __asm__ volatile(
"pfd 2, 0(%[ptr_x]) \n\t" "pfd 2, 0(%[ptr_x]) \n\t"
@ -227,7 +227,7 @@ static void __attribute__ ((noinline)) dswap_kernel_32(BLASLONG n, FLOAT *x, F
[mem_y] "+m" (*(double (*)[n])y), [mem_y] "+m" (*(double (*)[n])y),
[n_tmp] "+&r"(n) [n_tmp] "+&r"(n)
: [ptr_x] "a"(x), [ptr_y] "a"(y) : [ptr_x] "a"(x), [ptr_y] "a"(y)
: "cc", "memory","r1", "v0","v1","v2","v3","v4","v5","v6","v7","v16", : "cc", "r1", "v0","v1","v2","v3","v4","v5","v6","v7","v16",
"v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31" "v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
); );
return; return;

17
kernel/zarch/zscal.c
View File

@ -96,7 +96,7 @@ static void zscal_kernel_8(BLASLONG n, FLOAT da_r,FLOAT da_i, FLOAT *x) {
"clgrjl %[x_ptr],%%r0,1b \n\t" "clgrjl %[x_ptr],%%r0,1b \n\t"
: [mem] "+m" (*(double (*)[2*n])x) ,[x_ptr] "+&a"(x) : [mem] "+m" (*(double (*)[2*n])x) ,[x_ptr] "+&a"(x)
: [n] "r"(n), [alpha_r] "f"(da_r),[alpha_i] "f"(da_i) : [n] "r"(n), [alpha_r] "f"(da_r),[alpha_i] "f"(da_i)
: "cc", "memory","r0","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25" : "cc", "r0","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25"
); );
@ -106,10 +106,9 @@ static void zscal_kernel_8_zero_r(BLASLONG n, FLOAT da_i, FLOAT *x) {
__asm__ ( "pfd 2, 0(%1) \n\t" __asm__ ( "pfd 2, 0(%1) \n\t"
"lgdr %%r0,%[alpha] \n\t" "lgdr %%r0,%[alpha] \n\t"
"vlvgg %%v16,%%r0,0 \n\t" "vlvgp %%v16,%%r0,%%r0 \n\t" //load both from disjoint
"lcdbr %[alpha],%[alpha] \n\t" "vflcdb %%v16,%%v16 \n\t" //complement both
"lgdr %%r0,%[alpha] \n\t" "vlvgg %%v16,%%r0,0 \n\t" //restore 1st
"vlvgg %%v16,%%r0,1 \n\t"
"vlr %%v17 ,%%v16 \n\t" "vlr %%v17 ,%%v16 \n\t"
"sllg %%r0,%[n],4 \n\t" "sllg %%r0,%[n],4 \n\t"
"agr %%r0,%[x_ptr] \n\t" "agr %%r0,%[x_ptr] \n\t"
@ -129,8 +128,8 @@ static void zscal_kernel_8_zero_r(BLASLONG n, FLOAT da_i, FLOAT *x) {
"vsteg %%v26, 40(%[x_ptr]),0 \n\t" "vsteg %%v26, 40(%[x_ptr]),0 \n\t"
"vl %%v27, 48(%[x_ptr]) \n\t" "vl %%v27, 48(%[x_ptr]) \n\t"
"vfmdb %%v27,%%v27,%%v17 \n\t" "vfmdb %%v27,%%v27,%%v17 \n\t"
"vsteg %%v27, 40(%[x_ptr]),1 \n\t" "vsteg %%v27, 48(%[x_ptr]),1 \n\t"
"vsteg %%v27, 48(%[x_ptr]),0 \n\t" "vsteg %%v27, 56(%[x_ptr]),0 \n\t"
"vl %%v28, 64(%[x_ptr]) \n\t" "vl %%v28, 64(%[x_ptr]) \n\t"
"vfmdb %%v28,%%v28,%%v16 \n\t" "vfmdb %%v28,%%v28,%%v16 \n\t"
"vsteg %%v28, 64(%[x_ptr]),1 \n\t" "vsteg %%v28, 64(%[x_ptr]),1 \n\t"
@ -141,8 +140,8 @@ static void zscal_kernel_8_zero_r(BLASLONG n, FLOAT da_i, FLOAT *x) {
"vsteg %%v29, 88(%[x_ptr]),0 \n\t" "vsteg %%v29, 88(%[x_ptr]),0 \n\t"
"vl %%v30, 96(%[x_ptr]) \n\t" "vl %%v30, 96(%[x_ptr]) \n\t"
"vfmdb %%v30,%%v30,%%v16 \n\t" "vfmdb %%v30,%%v30,%%v16 \n\t"
"vsteg %%v27, 96(%[x_ptr]),1 \n\t" "vsteg %%v30, 96(%[x_ptr]),1 \n\t"
"vsteg %%v27, 104(%[x_ptr]),0 \n\t" "vsteg %%v30, 104(%[x_ptr]),0 \n\t"
"vl %%v31, 112(%[x_ptr]) \n\t" "vl %%v31, 112(%[x_ptr]) \n\t"
"vfmdb %%v31,%%v31,%%v17 \n\t" "vfmdb %%v31,%%v31,%%v17 \n\t"
"vsteg %%v31, 112(%[x_ptr]),1 \n\t" "vsteg %%v31, 112(%[x_ptr]),1 \n\t"

4
kernel/zarch/zswap.c
View File

@ -137,7 +137,7 @@ static void zswap_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y)
#else #else
static void __attribute__ ((noinline)) zswap_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y) static void zswap_kernel_16(BLASLONG n, FLOAT *x, FLOAT *y)
{ {
__asm__ volatile( __asm__ volatile(
"pfd 2, 0(%[ptr_x]) \n\t" "pfd 2, 0(%[ptr_x]) \n\t"
@ -225,7 +225,7 @@ static void __attribute__ ((noinline)) zswap_kernel_16(BLASLONG n, FLOAT *x, F
[mem_y] "+m" (*(double (*)[2*n])y), [mem_y] "+m" (*(double (*)[2*n])y),
[n_tmp] "+&r"(n) [n_tmp] "+&r"(n)
: [ptr_x] "a"(x), [ptr_y] "a"(y) : [ptr_x] "a"(x), [ptr_y] "a"(y)
: "cc", "memory", "r1", "v0","v1","v2","v3","v4","v5","v6","v7","v16", : "cc", "r1", "v0","v1","v2","v3","v4","v5","v6","v7","v16",
"v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31" "v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
); );
return; return;