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

Restore riscv64 fixes from develop branch: dot product double precision accumulation, zscal NaN handling

This commit is contained in:
Sergei Lewis 2024-02-01 10:26:02 +00:00
parent 1093def0d1
commit a3b0ef6596
4 changed files with 33 additions and 161 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
Makefile.prebuild
View File

@ -57,6 +57,7 @@ endif
ifeq ($(TARGET), CK860FV) ifeq ($(TARGET), CK860FV)
TARGET_FLAGS = -march=ck860v -mcpu=ck860fv -mfdivdu -mhard-float TARGET_FLAGS = -march=ck860v -mcpu=ck860fv -mfdivdu -mhard-float
endif
ifeq ($(TARGET), x280) ifeq ($(TARGET), x280)
TARGET_FLAGS = -march=rv64imafdcv_zba_zbb_zfh -mabi=lp64d TARGET_FLAGS = -march=rv64imafdcv_zba_zbb_zfh -mabi=lp64d

10
kernel/riscv64/dot.c
View File

@ -44,14 +44,24 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{ {
BLASLONG i=0; BLASLONG i=0;
BLASLONG ix=0,iy=0; BLASLONG ix=0,iy=0;
#if defined(DSDOT)
double dot = 0.0 ; double dot = 0.0 ;
#else
FLOAT dot = 0.0 ;
#endif
if ( n < 1 ) return(dot); if ( n < 1 ) return(dot);
while(i < n) while(i < n)
{ {
#if defined(DSDOT)
dot += (double) y[iy] * (double) x[ix] ;
#else
dot += y[iy] * x[ix] ; dot += y[iy] * x[ix] ;
#endif
ix += inc_x ; ix += inc_x ;
iy += inc_y ; iy += inc_y ;
i++ ; i++ ;

104
kernel/riscv64/zscal_rvv.c
View File

@ -69,104 +69,42 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r,FLOAT da_i, F
size_t vlmax = VSETVL_MAX; size_t vlmax = VSETVL_MAX;
FLOAT_VX2_T vx2; FLOAT_VX2_T vx2;
if(da_r == 0.0 && da_i == 0.0) { if(inc_x == 1) {
vr = VFMVVF_FLOAT(0.0, vlmax); for (size_t vl; n > 0; n -= vl, x += vl*2) {
vi = VFMVVF_FLOAT(0.0, vlmax);
if(inc_x == 1) {
for (size_t vl; n > 0; n -= vl, x += vl*2) {
vl = VSETVL(n);
vx2 = VSET_VX2(vx2, 0, vr);
vx2 = VSET_VX2(vx2, 1, vi);
VSSEG_FLOAT(x, vx2, vl);
}
} else {
for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
vl = VSETVL(n);
vx2 = VSET_VX2(vx2, 0, vr);
vx2 = VSET_VX2(vx2, 1, vi);
VSSSEG_FLOAT(x, stride_x, vx2, vl);
}
}
} else if(da_r == 0.0) {
for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
vl = VSETVL(n); vl = VSETVL(n);
vx2 = VLSSEG_FLOAT(x, stride_x, vl); vx2 = VLSEG_FLOAT(x, vl);
vr = VGET_VX2(vx2, 0); vr = VGET_VX2(vx2, 0);
vi = VGET_VX2(vx2, 1); vi = VGET_VX2(vx2, 1);
vt = VFMULVF_FLOAT(vi, -da_i, vl); vt = VFMULVF_FLOAT(vr, da_r, vl);
vi = VFMULVF_FLOAT(vr, da_i, vl); vt = VFNMSACVF_FLOAT(vt, da_i, vi, vl);
vi = VFMULVF_FLOAT(vi, da_r, vl);
vi = VFMACCVF_FLOAT(vi, da_i, vr, vl);
vx2 = VSET_VX2(vx2, 0, vt); vx2 = VSET_VX2(vx2, 0, vt);
vx2 = VSET_VX2(vx2, 1, vi); vx2 = VSET_VX2(vx2, 1, vi);
VSSEG_FLOAT(x, vx2, vl);
VSSSEG_FLOAT(x, stride_x, vx2, vl);
}
} else if(da_i == 0.0) {
for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
vl = VSETVL(n);
vx2 = VLSSEG_FLOAT(x, stride_x, vl);
vr = VGET_VX2(vx2, 0);
vi = VGET_VX2(vx2, 1);
vr = VFMULVF_FLOAT(vr, da_r, vl);
vi = VFMULVF_FLOAT(vi, da_r, vl);
vx2 = VSET_VX2(vx2, 0, vr);
vx2 = VSET_VX2(vx2, 1, vi);
VSSSEG_FLOAT(x, stride_x, vx2, vl);
} }
} else { } else {
if(inc_x == 1) { for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
vl = VSETVL(n);
for (size_t vl; n > 0; n -= vl, x += vl*2) { vx2 = VLSSEG_FLOAT(x, stride_x, vl);
vl = VSETVL(n); vr = VGET_VX2(vx2, 0);
vi = VGET_VX2(vx2, 1);
vx2 = VLSEG_FLOAT(x, vl); vt = VFMULVF_FLOAT(vr, da_r, vl);
vr = VGET_VX2(vx2, 0); vt = VFNMSACVF_FLOAT(vt, da_i, vi, vl);
vi = VGET_VX2(vx2, 1); vi = VFMULVF_FLOAT(vi, da_r, vl);
vi = VFMACCVF_FLOAT(vi, da_i, vr, vl);
vt = VFMULVF_FLOAT(vr, da_r, vl); vx2 = VSET_VX2(vx2, 0, vt);
vt = VFNMSACVF_FLOAT(vt, da_i, vi, vl); vx2 = VSET_VX2(vx2, 1, vi);
vi = VFMULVF_FLOAT(vi, da_r, vl); VSSSEG_FLOAT(x, stride_x, vx2, vl);
vi = VFMACCVF_FLOAT(vi, da_i, vr, vl);
vx2 = VSET_VX2(vx2, 0, vt);
vx2 = VSET_VX2(vx2, 1, vi);
VSSEG_FLOAT(x, vx2, vl);
}
} else {
for (size_t vl; n > 0; n -= vl, x += vl*inc_x*2) {
vl = VSETVL(n);
vx2 = VLSSEG_FLOAT(x, stride_x, vl);
vr = VGET_VX2(vx2, 0);
vi = VGET_VX2(vx2, 1);
vt = VFMULVF_FLOAT(vr, da_r, vl);
vt = VFNMSACVF_FLOAT(vt, da_i, vi, vl);
vi = VFMULVF_FLOAT(vi, da_r, vl);
vi = VFMACCVF_FLOAT(vi, da_i, vr, vl);
vx2 = VSET_VX2(vx2, 0, vt);
vx2 = VSET_VX2(vx2, 1, vi);
VSSSEG_FLOAT(x, stride_x, vx2, vl);
}
} }
} }

79
kernel/riscv64/zscal_vector.c
View File

@ -59,84 +59,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r,FLOAT da_i, F
unsigned int gvl = 0; unsigned int gvl = 0;
FLOAT_V_T vt, v0, v1; FLOAT_V_T vt, v0, v1;
if(da_r == 0.0 && da_i == 0.0){ {
gvl = VSETVL(n);
BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
BLASLONG inc_xv = inc_x * 2 * gvl;
vt = VFMVVF_FLOAT(0.0, gvl);
for(i=0,j=0; i < n/(gvl*2); i++){
VSSEV_FLOAT(&x[ix], stride_x, vt, gvl);
VSSEV_FLOAT(&x[ix+1], stride_x, vt, gvl);
VSSEV_FLOAT(&x[ix+inc_xv], stride_x, vt, gvl);
VSSEV_FLOAT(&x[ix+inc_xv+1], stride_x, vt, gvl);
j += gvl*2;
ix += inc_xv*2;
}
for(; j < n; ){
gvl = VSETVL(n-j);
VSSEV_FLOAT(&x[ix], stride_x, vt, gvl);
VSSEV_FLOAT(&x[ix+1], stride_x, vt, gvl);
j += gvl;
ix += inc_x * 2 * gvl;
}
}else if(da_r == 0.0){
gvl = VSETVL(n);
BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
BLASLONG inc_xv = inc_x * 2 * gvl;
for(i=0,j=0; i < n/gvl; i++){
v0 = VLSEV_FLOAT(&x[ix], stride_x, gvl);
v1 = VLSEV_FLOAT(&x[ix+1], stride_x, gvl);
vt = VFMULVF_FLOAT(v1, -da_i, gvl);
v1 = VFMULVF_FLOAT(v0, da_i, gvl);
VSSEV_FLOAT(&x[ix], stride_x, vt, gvl);
VSSEV_FLOAT(&x[ix+1], stride_x, v1, gvl);
j += gvl;
ix += inc_xv;
}
if(j < n){
gvl = VSETVL(n-j);
v0 = VLSEV_FLOAT(&x[ix], stride_x, gvl);
v1 = VLSEV_FLOAT(&x[ix+1], stride_x, gvl);
vt = VFMULVF_FLOAT(v1, -da_i, gvl);
v1 = VFMULVF_FLOAT(v0, da_i, gvl);
VSSEV_FLOAT(&x[ix], stride_x, vt, gvl);
VSSEV_FLOAT(&x[ix+1], stride_x, v1, gvl);
}
}else if(da_i == 0.0){
gvl = VSETVL(n);
BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
BLASLONG inc_xv = inc_x * 2 * gvl;
for(i=0,j=0; i < n/gvl; i++){
v0 = VLSEV_FLOAT(&x[ix], stride_x, gvl);
v1 = VLSEV_FLOAT(&x[ix+1], stride_x, gvl);
vt = VFMULVF_FLOAT(v0, da_r, gvl);
v1 = VFMULVF_FLOAT(v1, da_r, gvl);
VSSEV_FLOAT(&x[ix], stride_x, vt, gvl);
VSSEV_FLOAT(&x[ix+1], stride_x, v1, gvl);
j += gvl;
ix += inc_xv;
}
if(j < n){
gvl = VSETVL(n-j);
v0 = VLSEV_FLOAT(&x[ix], stride_x, gvl);
v1 = VLSEV_FLOAT(&x[ix+1], stride_x, gvl);
vt = VFMULVF_FLOAT(v0, da_r, gvl);
v1 = VFMULVF_FLOAT(v1, da_r, gvl);
VSSEV_FLOAT(&x[ix], stride_x, vt, gvl);
VSSEV_FLOAT(&x[ix+1], stride_x, v1, gvl);
}
}else{
gvl = VSETVL(n); gvl = VSETVL(n);
BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT); BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
BLASLONG inc_xv = inc_x * 2 * gvl; BLASLONG inc_xv = inc_x * 2 * gvl;