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

update the amin function for riscv platform 

improve the amin function with less vector instruction.
only verify it on QEMU.
This commit is contained in:
Wu Zhigang 2021-03-10 11:07:57 -08:00
parent ef0238ba2b
commit 8c6e532358
1 changed files with 64 additions and 132 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

184
kernel/riscv64/amin_vector.c
View File

@ -58,90 +58,56 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
{ {
BLASLONG i=0, j=0; BLASLONG i=0, j=0;
if (n <= 0 || inc_x <= 0) return(0.0); if (n <= 0 || inc_x <= 0)
return(0.0);
FLOAT minf=FLT_MAX; FLOAT minf=FLT_MAX;
unsigned int gvl = 0; unsigned int gvl = 0;
FLOAT_V_T v0, v1, v_min; volatile FLOAT_V_T v0;
FLOAT_V_T v_min, v1;
MASK_T mask0, mask1;
FLOAT zero = 0.0; FLOAT zero = 0.0;
if(inc_x == 1){ if(inc_x == 1){
gvl = vsetvli(n, RVV_EFLOAT, RVV_M); gvl = vsetvli(n, RVV_EFLOAT, RVV_M);
if(gvl <= n/2){ if(gvl <= n/2){
v_min = VFMVVF_FLOAT(FLT_MAX, gvl); v_min = VFMVVF_FLOAT(FLT_MAX, gvl);
for(i=0,j=0; i<n/(gvl*2); i++){ for(i=0,j=0; i<n/(gvl*2); i++){
v0 = VLEV_FLOAT(&x[j], gvl);
mask0 = VMFLTVF_FLOAT(v0, 0, gvl);
//v0 = VFRSUBVF_MASK_FLOAT(v0, 0, mask0, gvl);
#if defined(DOUBLE)
asm volatile(
"vor.vv v0, %1, %1\n\t"
"vsetvli x0, %3, e64,m8 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v0)
:"v"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
"vor.vv v0, %1, %1\n\t"
"vsetvli x0, %3, e32,m8 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v0)
:"v"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif
v_min = VFMINVV_FLOAT(v_min, v0, gvl);
v1 = VLEV_FLOAT(&x[j+gvl], gvl); asm volatile(
mask1 = VMFLTVF_FLOAT(v1, 0, gvl); "vle.v %1, (%4) \n\t"
//v1 = VFRSUBVF_MASK_FLOAT(v1, 0, mask1, gvl); "vle.v %2, (%5) \n\t"
#if defined(DOUBLE)
asm volatile( "vmflt.vf v0, %1, %3 \n\t"
"vor.vv v0, %1, %1\n\t" "vfrsub.vf %1, %1, %3, v0.t \n\t"
"vsetvli x0, %3, e64,m8 \n\t" "vmflt.vf v0, %2, %3 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t" "vfrsub.vf %2, %2, %3, v0.t \n\t"
:"+v"(v1)
:"v"(mask1), "f"(zero), "r"(gvl) "vfmin.vv %0, %0, %1 \n\t"
:"v0"); "vfmin.vv %0, %0, %2 \n\t"
#else :"+v"(v_min)
asm volatile( :"v"(v0), "v"(v1), "f"(zero), "r"(&x[j]), "r"(&x[j+gvl])
"vor.vv v0, %1, %1\n\t" :"v0"
"vsetvli x0, %3, e32,m8 \n\t" );
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"v0");
#endif
v_min = VFMINVV_FLOAT(v_min, v1, gvl);
j += gvl*2; j += gvl*2;
} }
v1 = VFMVVF_FLOAT(FLT_MAX, gvl); v1 = VFMVVF_FLOAT(FLT_MAX, gvl);
v0 = VFREDMINVS_FLOAT(v_min, v1, gvl); v0 = VFREDMINVS_FLOAT(v_min, v1, gvl);
minf = v0[0]; minf = v0[0];
} }
for(;j<n;){ for(;j<n;){
gvl = vsetvli(n-j, RVV_EFLOAT, RVV_M); gvl = vsetvli(n-j, RVV_EFLOAT, RVV_M);
v0 = VLEV_FLOAT(&x[j], gvl); asm volatile(
mask0 = VMFLTVF_FLOAT(v0, 0, gvl); "vle.v %0, (%2) \n\t"
//v0 = VFRSUBVF_MASK_FLOAT(v0, 0, mask0, gvl); "vmflt.vf v0, %0, %1 \n\t"
#if defined(DOUBLE) "vfrsub.vf %0, %0, %1, v0.t \n\t"
asm volatile(
"vor.vv v0, %1, %1\n\t"
"vsetvli x0, %3, e64,m8 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v0) :"+v"(v0)
:"v"(mask0), "f"(zero), "r"(gvl) :"f"(zero), "r"(&x[j])
:"v0"); :"v0"
#else );
asm volatile(
"vor.vv v0, %1, %1\n\t"
"vsetvli x0, %3, e32,m8 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v0)
:"v"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif
v1 = VFMVVF_FLOAT(FLT_MAX, gvl); v1 = VFMVVF_FLOAT(FLT_MAX, gvl);
v0 = VFREDMINVS_FLOAT(v0, v1, gvl); v0 = VFREDMINVS_FLOAT(v0, v1, gvl);
if(v0[0] < minf) if(v0[0] < minf)
@ -151,54 +117,29 @@ asm volatile(
}else{ }else{
gvl = vsetvli(n, RVV_EFLOAT, RVV_M); gvl = vsetvli(n, RVV_EFLOAT, RVV_M);
BLASLONG stride_x = inc_x * sizeof(FLOAT); BLASLONG stride_x = inc_x * sizeof(FLOAT);
if(gvl <= n/2){ if(gvl <= n/2){
BLASLONG idx = 0, inc_xv = inc_x * gvl; BLASLONG idx = 0, inc_xv = inc_x * gvl;
v_min = VFMVVF_FLOAT(FLT_MAX, gvl); v_min = VFMVVF_FLOAT(FLT_MAX, gvl);
for(i=0,j=0; i<n/(gvl*2); i++){ for(i=0,j=0; i<n/(gvl*2); i++){
v0 = VLSEV_FLOAT(&x[idx], stride_x, gvl);
mask0 = VMFLTVF_FLOAT(v0, 0, gvl);
//v0 = VFRSUBVF_MASK_FLOAT(v0, 0, mask0, gvl);
#if defined(DOUBLE)
asm volatile(
"vor.vv v0, %1, %1\n\t"
"vsetvli x0, %3, e64,m8 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v0)
:"v"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
"vor.vv v0, %1, %1\n\t"
"vsetvli x0, %3, e32,m8 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v0)
:"v"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif
v_min = VFMINVV_FLOAT(v_min, v0, gvl);
v1 = VLSEV_FLOAT(&x[idx+inc_xv], stride_x, gvl); asm volatile(
mask1 = VMFLTVF_FLOAT(v1, 0, gvl); "vlse.v %1, (%4), %6 \n\t"
//v1 = VFRSUBVF_MASK_FLOAT(v1, 0, mask1, gvl); "vlse.v %2, (%5), %6 \n\t"
#if defined(DOUBLE)
asm volatile( "vmflt.vf v0, %1, %3 \n\t"
"vor.vv v0, %1, %1\n\t" "vfrsub.vf %1, %1, %3, v0.t \n\t"
"vsetvli x0, %3, e64,m8 \n\t" "vmflt.vf v0, %2, %3 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t" "vfrsub.vf %2, %2, %3, v0.t \n\t"
:"+v"(v1)
:"v"(mask1), "f"(zero), "r"(gvl) "vfmin.vv %0, %0, %1 \n\t"
:"v0"); "vfmin.vv %0, %0, %2 \n\t"
#else :"+v"(v_min)
asm volatile( :"v"(v0), "v"(v1), "f"(zero), "r"(&x[idx]), "r"(&x[idx+inc_xv]), "r"(stride_x)
"vor.vv v0, %1, %1\n\t" :"v0"
"vsetvli x0, %3, e32,m8 \n\t" );
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"v0");
#endif
v_min = VFMINVV_FLOAT(v_min, v1, gvl);
j += gvl*2; j += gvl*2;
idx += inc_xv*2; idx += inc_xv*2;
} }
@ -206,28 +147,18 @@ asm volatile(
v0 = VFREDMINVS_FLOAT(v_min, v1, gvl); v0 = VFREDMINVS_FLOAT(v_min, v1, gvl);
minf = v0[0]; minf = v0[0];
} }
for(;j<n;){ for(;j<n;){
gvl = vsetvli(n-j, RVV_EFLOAT, RVV_M); gvl = vsetvli(n-j, RVV_EFLOAT, RVV_M);
v0 = VLSEV_FLOAT(&x[j*inc_x], stride_x, gvl); asm volatile(
mask0 = VMFLTVF_FLOAT(v0, 0, gvl); "vlse.v %0, (%2), %3 \n\t"
//v0 = VFRSUBVF_MASK_FLOAT(v0, 0, mask0, gvl); "vmflt.vf v0, %0, %1 \n\t"
#if defined(DOUBLE) "vfrsub.vf %0, %0, %1, v0.t \n\t"
asm volatile(
"vor.vv v0, %1, %1\n\t"
"vsetvli x0, %3, e64,m8 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v0) :"+v"(v0)
:"v"(mask0), "f"(zero), "r"(gvl) :"f"(zero), "r"(&x[j*inc_x]), "r"(stride_x)
:"v0"); :"v0"
#else );
asm volatile(
"vor.vv v0, %1, %1\n\t"
"vsetvli x0, %3, e32,m8 \n\t"
"vfrsub.vf %0, %0, %2, v0.t \n\t"
:"+v"(v0)
:"v"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif
v1 = VFMVVF_FLOAT(FLT_MAX, gvl); v1 = VFMVVF_FLOAT(FLT_MAX, gvl);
v0 = VFREDMINVS_FLOAT(v0, v1, gvl); v0 = VFREDMINVS_FLOAT(v0, v1, gvl);
if(v0[0] < minf) if(v0[0] < minf)
@ -235,6 +166,7 @@ asm volatile(
j += gvl; j += gvl;
} }
} }
return(minf); return(minf);
} }