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

[ZARCH] Z14 support, BLAS 1/2 single precision implementations, Some missing double precision implementations, Gemv optimization

Browse Source
This commit is contained in:
maamountki
2018-08-06 18:20:40 +03:00
committed by GitHub
parent ee955757f9
commit 23229011db
79 changed files with 17314 additions and 2897 deletions
Download Patch File Download Diff File Expand all files Collapse all files

175
kernel/zarch/daxpy.c
View File

@@ -25,98 +25,99 @@ OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
#include "common.h"
#define PREFETCH_INS 1
#if defined(Z13_A)
#include <vecintrin.h>
static void daxpy_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT alpha)
{
BLASLONG i = 0;
__vector double v_a = {alpha,alpha};
__vector double * v_y=(__vector double *)y;
__vector double * v_x=(__vector double *)x;
for(; i<n/2; i+=16){
v_y[i] += v_a * v_x[i];
v_y[i+1] += v_a * v_x[i+1];
v_y[i+2] += v_a * v_x[i+2];
v_y[i+3] += v_a * v_x[i+3];
v_y[i+4] += v_a * v_x[i+4];
v_y[i+5] += v_a * v_x[i+5];
v_y[i+6] += v_a * v_x[i+6];
v_y[i+7] += v_a * v_x[i+7];
v_y[i+8] += v_a * v_x[i+8];
v_y[i+9] += v_a * v_x[i+9];
v_y[i+10] += v_a * v_x[i+10];
v_y[i+11] += v_a * v_x[i+11];
v_y[i+12] += v_a * v_x[i+12];
v_y[i+13] += v_a * v_x[i+13];
v_y[i+14] += v_a * v_x[i+14];
v_y[i+15] += v_a * v_x[i+15];
}
}
#else
static void daxpy_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT alpha)
static void daxpy_kernel_32(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
{
__asm__ volatile(
"vlrepg %%v0,%3 \n\t"
"srlg %%r0,%0,5 \n\t"
"xgr %%r1,%%r1 \n\t"
"0: \n\t"
"pfd 1, 1024(%%r1,%1) \n\t"
"pfd 2, 1024(%%r1,%2) \n\t"
"vl %%v16,0(%%r1,%1) \n\t"
"vl %%v17,16(%%r1,%1) \n\t"
"vl %%v18,32(%%r1,%1) \n\t"
"vl %%v19,48(%%r1,%1) \n\t"
"vl %%v20,0(%%r1,%2) \n\t"
"vl %%v21,16(%%r1,%2) \n\t"
"vl %%v22,32(%%r1,%2) \n\t"
"vl %%v23,48(%%r1,%2) \n\t"
__asm__ volatile(
#if defined(PREFETCH_INS)
"pfd 1, 0(%[x_tmp]) \n\t"
"pfd 2, 0(%[y_tmp]) \n\t"
#endif
"lgdr %%r0,%[alpha] \n\t"
"vlvgp %%v0,%%r0,%%r0 \n\t"
"srlg %%r0,%[n],5 \n\t"
"vlr %%v1,%%v0 \n\t"
".align 16 \n\t"
"1: \n\t"
#if defined(PREFETCH_INS)
"pfd 1, 256(%[x_tmp]) \n\t"
"pfd 2, 256(%[y_tmp]) \n\t"
#endif
"vlm %%v16,%%v23, 0(%[x_tmp]) \n\t"
"vlm %%v24, %%v31, 0(%[y_tmp]) \n\t"
"vfmadb %%v16,%%v0,%%v16,%%v24 \n\t"
"vfmadb %%v17,%%v1,%%v17,%%v25 \n\t"
"vfmadb %%v18,%%v0,%%v18,%%v26 \n\t"
"vfmadb %%v19,%%v1,%%v19,%%v27 \n\t"
"vfmadb %%v20,%%v0,%%v20,%%v28 \n\t"
"vfmadb %%v21,%%v1,%%v21,%%v29 \n\t"
"vfmadb %%v22,%%v0,%%v22,%%v30 \n\t"
"vfmadb %%v23,%%v1,%%v23,%%v31 \n\t"
"vstm %%v16,%%v23, 0(%[y_tmp]) \n\t"
"vlm %%v24,%%v31, 128(%[x_tmp]) \n\t"
"vlm %%v16,%%v23, 128(%[y_tmp]) \n\t"
"vfmadb %%v24,%%v0,%%v24,%%v16 \n\t"
"vfmadb %%v25,%%v1,%%v25,%%v17 \n\t"
"vfmadb %%v26,%%v0,%%v26,%%v18 \n\t"
"vfmadb %%v27,%%v1,%%v27,%%v19 \n\t"
"vfmadb %%v28,%%v0,%%v28,%%v20 \n\t"
"vfmadb %%v29,%%v1,%%v29,%%v21 \n\t"
"vfmadb %%v30,%%v0,%%v30,%%v22 \n\t"
"vfmadb %%v31,%%v1,%%v31,%%v23 \n\t"
"la %[x_tmp],256(%[x_tmp]) \n\t"
"vstm %%v24, %%v31, 128(%[y_tmp]) \n\t"
"la %[y_tmp],256(%[y_tmp]) \n\t"
"brctg %%r0,1b"
: [mem_y] "+m" (*(double (*)[n])y), [x_tmp] "+&a"(x), [y_tmp] "+&a"(y)
: [mem_x] "m" (*(const double (*)[n])x), [n] "r"(n), [alpha] "f"(alpha)
:"cc", "r0", "v0","v1","v16","v17","v18","v19","v20","v21",
"v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
);
"vfmadb %%v16,%%v0,%%v16,%%v20 \n\t"
"vfmadb %%v17,%%v0,%%v17,%%v21 \n\t"
"vfmadb %%v18,%%v0,%%v18,%%v22 \n\t"
"vfmadb %%v19,%%v0,%%v19,%%v23 \n\t"
"vl %%v24,64(%%r1,%1) \n\t"
"vl %%v25,80(%%r1,%1) \n\t"
"vl %%v26,96(%%r1,%1) \n\t"
"vl %%v27,112(%%r1,%1) \n\t"
"vl %%v28,64(%%r1,%2) \n\t"
"vl %%v29,80(%%r1,%2) \n\t"
"vl %%v30,96(%%r1,%2) \n\t"
"vl %%v31,112(%%r1,%2) \n\t"
"vfmadb %%v20,%%v0,%%v24,%%v28 \n\t"
"vfmadb %%v21,%%v0,%%v25,%%v29 \n\t"
"vfmadb %%v22,%%v0,%%v26,%%v30 \n\t"
"vfmadb %%v23,%%v0,%%v27,%%v31 \n\t"
"vst %%v16,0(%%r1,%2) \n\t"
"vst %%v17,16(%%r1,%2) \n\t"
"vst %%v18,32(%%r1,%2) \n\t"
"vst %%v19,48(%%r1,%2) \n\t"
"vst %%v20,64(%%r1,%2) \n\t"
"vst %%v21,80(%%r1,%2) \n\t"
"vst %%v22,96(%%r1,%2) \n\t"
"vst %%v23,112(%%r1,%2) \n\t"
"vl %%v16,128(%%r1,%1) \n\t"
"vl %%v17,144(%%r1,%1) \n\t"
"vl %%v18,160(%%r1,%1) \n\t"
"vl %%v19,176(%%r1,%1) \n\t"
"vl %%v20,128(%%r1,%2) \n\t"
"vl %%v21,144(%%r1,%2) \n\t"
"vl %%v22,160(%%r1,%2) \n\t"
"vl %%v23,176(%%r1,%2) \n\t"
"vfmadb %%v16,%%v0,%%v16,%%v20 \n\t"
"vfmadb %%v17,%%v0,%%v17,%%v21 \n\t"
"vfmadb %%v18,%%v0,%%v18,%%v22 \n\t"
"vfmadb %%v19,%%v0,%%v19,%%v23 \n\t"
"vl %%v24,192(%%r1,%1) \n\t"
"vl %%v25,208(%%r1,%1) \n\t"
"vl %%v26,224(%%r1,%1) \n\t"
"vl %%v27,240(%%r1,%1) \n\t"
"vl %%v28,192(%%r1,%2) \n\t"
"vl %%v29,208(%%r1,%2) \n\t"
"vl %%v30,224(%%r1,%2) \n\t"
"vl %%v31,240(%%r1,%2) \n\t"
"vfmadb %%v20,%%v0,%%v24,%%v28 \n\t"
"vfmadb %%v21,%%v0,%%v25,%%v29 \n\t"
"vfmadb %%v22,%%v0,%%v26,%%v30 \n\t"
"vfmadb %%v23,%%v0,%%v27,%%v31 \n\t"
"vst %%v16,128(%%r1,%2) \n\t"
"vst %%v17,144(%%r1,%2) \n\t"
"vst %%v18,160(%%r1,%2) \n\t"
"vst %%v19,176(%%r1,%2) \n\t"
"vst %%v20,192(%%r1,%2) \n\t"
"vst %%v21,208(%%r1,%2) \n\t"
"vst %%v22,224(%%r1,%2) \n\t"
"vst %%v23,240(%%r1,%2) \n\t"
"agfi %%r1,256 \n\t"
"brctg %%r0,0b "
:
:"r"(n),"ZR"((const FLOAT (*)[n])x),"ZR"((FLOAT (*)[n])y),"m"(*alpha)
:"memory","cc","r0","r1","v0","v16","v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
);
}
#endif
int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
{
@@ -131,7 +132,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
BLASLONG n1 = n & -32;
if ( n1 )
daxpy_kernel_32(n1, x, y , da );
daxpy_kernel_32(n1, x, y , &da);
i = n1;
while(i < n)