OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Update RISC-V Intrinsic API.

tags/v0.3.21
Xianyi Zhang 3 years ago
parent
45786b05da
commit
968e1f51d8
32 changed files with 316 additions and 249 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +1
    -1
      Makefile.prebuild
  2. +2
    -2
      Makefile.riscv64
  3. +1
    -1
      common_riscv64.h
  4. +31
    -30
      kernel/riscv64/amax_vector.c
  5. +30
    -30
      kernel/riscv64/amin_vector.c
  6. +6
    -6
      kernel/riscv64/asum_vector.c
  7. +1
    -1
      kernel/riscv64/copy_vector.c
  8. +1
    -1
      kernel/riscv64/dgemm_kernel_8x4_c910v.c
  9. +22
    -14
      kernel/riscv64/dot_vector.c
  10. +10
    -7
      kernel/riscv64/gemv_t_vector.c
  11. +8
    -8
      kernel/riscv64/iamax_vector.c
  12. +8
    -8
      kernel/riscv64/iamin_vector.c
  13. +8
    -8
      kernel/riscv64/imax_vector.c
  14. +8
    -8
      kernel/riscv64/imin_vector.c
  15. +15
    -4
      kernel/riscv64/izamax_vector.c
  16. +14
    -4
      kernel/riscv64/izamin_vector.c
  17. +6
    -6
      kernel/riscv64/max_vector.c
  18. +6
    -6
      kernel/riscv64/min_vector.c
  19. +20
    -16
      kernel/riscv64/nrm2_vector.c
  20. +8
    -6
      kernel/riscv64/nrm2_vector_dot.c
  21. +8
    -3
      kernel/riscv64/scal_vector.c
  22. +1
    -1
      kernel/riscv64/sgemm_kernel_16x4_c910v.c
  23. +12
    -10
      kernel/riscv64/symv_L_vector.c
  24. +12
    -10
      kernel/riscv64/symv_U_vector.c
  25. +8
    -3
      kernel/riscv64/zamax_vector.c
  26. +5
    -3
      kernel/riscv64/zamin_vector.c
  27. +8
    -6
      kernel/riscv64/zasum_vector.c
  28. +8
    -6
      kernel/riscv64/zdot_vector.c
  29. +8
    -6
      kernel/riscv64/zgemv_t_vector.c
  30. +8
    -6
      kernel/riscv64/zhemv_LM_vector.c
  31. +8
    -6
      kernel/riscv64/zhemv_UV_vector.c
  32. +24
    -22
      kernel/riscv64/znrm2_vector.c

+ 1
- 1
Makefile.prebuild View File

@@ -46,7 +46,7 @@ TARGET_FLAGS = -mips64r6
endif

ifeq ($(TARGET), C910V)
TARGET_FLAGS = -march=rv64gcvxthead -mabi=lp64v
TARGET_FLAGS = -march=rv64gcv0p7_zfh_xtheadc -mabi=lp64d
endif

all: getarch_2nd


+ 2
- 2
Makefile.riscv64 View File

@@ -1,4 +1,4 @@
ifeq ($(CORE), C910V)
CCOMMON_OPT += -march=rv64gcvxthead -mabi=lp64v
FCOMMON_OPT += -march=rv64gcvxthead -mabi=lp64v -static
CCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920
FCOMMON_OPT += -march=rv64imafdcv0p7_zfh_xtheadc -mabi=lp64d -mtune=c920 -static
endif

+ 1
- 1
common_riscv64.h View File

@@ -92,7 +92,7 @@ static inline int blas_quickdivide(blasint x, blasint y){
#define SEEK_ADDRESS

#if defined(C910V)
#include <riscv-vector.h>
#include <riscv_vector.h>
#endif

#endif

+ 31
- 30
kernel/riscv64/amax_vector.c View File

@@ -88,8 +88,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -97,8 +97,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif

@@ -113,8 +113,8 @@ 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"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"+vd"(v1)
:"vd"(mask1), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -122,8 +122,8 @@ 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"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"+vd"(v1)
:"vd"(mask1), "f"(zero), "r"(gvl)
:"v0");
#endif

@@ -131,7 +131,8 @@ asm volatile(
j += gvl*2;
}
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_zero, gvl);
maxf = v_res[0];
maxf = *((FLOAT*)&v_res);
//maxf = v_res[0];
}
for(;j<n;){
gvl = VSETVL(n-j);
@@ -144,8 +145,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -153,14 +154,14 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif

v_res = VFREDMAXVS_FLOAT(v_res, v0, v_zero, gvl);
if(v_res[0] > maxf)
maxf = v_res[0];
if(*((FLOAT*)&v_res) > maxf)
maxf = *((FLOAT*)&v_res);
j += gvl;
}
}else{
@@ -179,8 +180,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -188,8 +189,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif

@@ -204,8 +205,8 @@ 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"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"+vd"(v1)
:"vd"(mask1), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -213,8 +214,8 @@ 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"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"+vd"(v1)
:"vd"(mask1), "f"(zero), "r"(gvl)
:"v0");
#endif

@@ -223,7 +224,7 @@ asm volatile(
ix += inc_xv*2;
}
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_zero, gvl);
maxf = v_res[0];
maxf = *((FLOAT*)&v_res);
}
for(;j<n;){
gvl = VSETVL(n-j);
@@ -236,8 +237,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -245,14 +246,14 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif

v_res = VFREDMAXVS_FLOAT(v_res, v0, v_zero, gvl);
if(v_res[0] > maxf)
maxf = v_res[0];
if(*((FLOAT*)&v_res) > maxf)
maxf = *((FLOAT*)&v_res);
j += gvl;
}
}


+ 30
- 30
kernel/riscv64/amin_vector.c View File

@@ -87,8 +87,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -96,8 +96,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif
v_min = VFMINVV_FLOAT(v_min, v0, gvl);
@@ -111,8 +111,8 @@ 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"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"+vd"(v1)
:"vd"(mask1), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -120,8 +120,8 @@ 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"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"+vd"(v1)
:"vd"(mask1), "f"(zero), "r"(gvl)
:"v0");
#endif

@@ -129,7 +129,7 @@ asm volatile(
j += gvl*2;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = *((FLOAT*)&v_res);
}
for(;j<n;){
gvl = VSETVL(n-j);
@@ -142,8 +142,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -151,13 +151,13 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif
v_res = VFREDMINVS_FLOAT(v_res, v0, v_max, gvl);
if(v_res[0] < minf)
minf = v_res[0];
if(*((FLOAT*)&v_res) < minf)
minf = *((FLOAT*)&v_res);
j += gvl;
}
}else{
@@ -176,8 +176,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -185,8 +185,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif
v_min = VFMINVV_FLOAT(v_min, v0, gvl);
@@ -200,8 +200,8 @@ 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"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"+vd"(v1)
:"vd"(mask1), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -209,8 +209,8 @@ 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"(v1)
:"v"(mask1), "f"(zero), "r"(gvl)
:"+vd"(v1)
:"vd"(mask1), "f"(zero), "r"(gvl)
:"v0");
#endif

@@ -219,7 +219,7 @@ asm volatile(
idx += inc_xv*2;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = *((FLOAT*)&v_res);
}
for(;j<n;){
gvl = VSETVL(n-j);
@@ -232,8 +232,8 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#else
asm volatile(
@@ -241,13 +241,13 @@ 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)
:"+vd"(v0)
:"vd"(mask0), "f"(zero), "r"(gvl)
:"v0");
#endif
v_res = VFREDMINVS_FLOAT(v_res, v0, v_max, gvl);
if(v_res[0] < minf)
minf = v_res[0];
if(*((FLOAT*)&v_res) < minf)
minf = *((FLOAT*)&v_res);
j += gvl;
}
}


+ 6
- 6
kernel/riscv64/asum_vector.c View File

@@ -35,7 +35,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define FLOAT_V_T_M1 vfloat32m1_t
#define VLEV_FLOAT vle_v_f32m8
#define VLSEV_FLOAT vlse_v_f32m8
#define VFREDSUMVS_FLOAT vfredsum_vs_f32m8_f32m1
#define VFREDSUMVS_FLOAT vfredosum_vs_f32m8_f32m1
#define MASK_T vbool4_t
#define VMFLTVF_FLOAT vmflt_vf_f32m8_b4
#define VFMVVF_FLOAT vfmv_v_f_f32m8
@@ -49,7 +49,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define FLOAT_V_T_M1 vfloat64m1_t
#define VLEV_FLOAT vle_v_f64m8
#define VLSEV_FLOAT vlse_v_f64m8
#define VFREDSUMVS_FLOAT vfredsum_vs_f64m8_f64m1
#define VFREDSUMVS_FLOAT vfredusum_vs_f64m8_f64m1
#define MASK_T vbool8_t
#define VMFLTVF_FLOAT vmflt_vf_f64m8_b8
#define VFMVVF_FLOAT vfmv_v_f_f64m8
@@ -89,7 +89,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
j += gvl * 2;
}
v_res = VFREDSUMVS_FLOAT(v_res, v_sum, v_z0, gvl);
asumf += v_res[0];
asumf += *((FLOAT*)&v_res);
}
for(;j<n;){
gvl = VSETVL(n-j);
@@ -97,7 +97,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
mask0 = VMFLTVF_FLOAT(v0, 0, gvl);
v0 = VFRSUBVF_MASK_FLOAT(mask0, v0, v0, 0, gvl);
v_res = VFREDSUMVS_FLOAT(v_res, v0, v_z0, gvl);
asumf += v_res[0];
asumf += *((FLOAT*)&v_res);
j += gvl;
}
}else{
@@ -121,7 +121,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
inc_xv += inc_xv * 2;
}
v_res = VFREDSUMVS_FLOAT(v_res, v_sum, v_z0, gvl);
asumf += v_res[0];
asumf += *((FLOAT*)&v_res);
}
for(;j<n;){
gvl = VSETVL(n-j);
@@ -129,7 +129,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
mask0 = VMFLTVF_FLOAT(v0, 0, gvl);
v0 = VFRSUBVF_MASK_FLOAT(mask0, v0, v0, 0, gvl);
v_res = VFREDSUMVS_FLOAT(v_res, v0, v_z0, gvl);
asumf += v_res[0];
asumf += *((FLOAT*)&v_res);
j += gvl;
}
}


+ 1
- 1
kernel/riscv64/copy_vector.c View File

@@ -33,7 +33,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSEV_FLOAT vse_v_f32m8
#define VSSEV_FLOAT vsse_v_f32m8
#else
#define VSETVL(n) vsetvl_e32m8(n)
#define VSETVL(n) vsetvl_e64m8(n)
#define FLOAT_V_T vfloat64m8_t
#define VLEV_FLOAT vle_v_f64m8
#define VLSEV_FLOAT vlse_v_f64m8


+ 1
- 1
kernel/riscv64/dgemm_kernel_8x4_c910v.c View File

@@ -1,5 +1,5 @@
#include "common.h"
#include <riscv-vector.h>
#include <riscv_vector.h>

#define KERNEL8x4_I \
"addi t1, %[PB], 1*8 \n\t"\


+ 22
- 14
kernel/riscv64/dot_vector.c View File

@@ -31,9 +31,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLEV_FLOAT vle_v_f32m4
#define VLSEV_FLOAT vlse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredosum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
@@ -43,9 +44,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLEV_FLOAT vle_v_f64m4
#define VLSEV_FLOAT vlse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
@@ -81,7 +83,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
}
if(j > 0){
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
dot += v_res[0];
dot += (double)VFMVFS_FLOAT(v_res);
}
//tail
if(j < n){
@@ -92,12 +94,12 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
//vr = VFDOTVV_FLOAT(vx, vy, gvl);
vr = VFMACCVV_FLOAT(vz, vx, vy, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
dot += v_res[0];
dot += (double)VFMVFS_FLOAT(v_res);
}
}else if(inc_y == 1){
gvl = VSETVL(n);
vr = VFMVVF_FLOAT(0, gvl);
unsigned int stride_x = inc_x * sizeof(FLOAT);
int stride_x = inc_x * sizeof(FLOAT);
for(i=0,j=0; i<n/gvl; i++){
vx = VLSEV_FLOAT(&x[j*inc_x], stride_x, gvl);
vy = VLEV_FLOAT(&y[j], gvl);
@@ -106,7 +108,8 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
}
if(j > 0){
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
dot += v_res[0];
dot += (double)VFMVFS_FLOAT(v_res);

}
//tail
if(j < n){
@@ -117,12 +120,13 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
//vr = VFDOTVV_FLOAT(vx, vy, gvl);
vr = VFMACCVV_FLOAT(vz, vx, vy, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
dot += v_res[0];
dot += (double)VFMVFS_FLOAT(v_res);

}
}else if(inc_x == 1){
gvl = VSETVL(n);
vr = VFMVVF_FLOAT(0, gvl);
unsigned int stride_y = inc_y * sizeof(FLOAT);
int stride_y = inc_y * sizeof(FLOAT);
for(i=0,j=0; i<n/gvl; i++){
vx = VLEV_FLOAT(&x[j], gvl);
vy = VLSEV_FLOAT(&y[j*inc_y], stride_y, gvl);
@@ -131,7 +135,8 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
}
if(j > 0){
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
dot += v_res[0];
dot += (double)VFMVFS_FLOAT(v_res);

}
//tail
if(j < n){
@@ -142,13 +147,14 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
//vr = VFDOTVV_FLOAT(vx, vy, gvl);
vr = VFMACCVV_FLOAT(vz, vx, vy, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
dot += v_res[0];
dot += (double)VFMVFS_FLOAT(v_res);

}
}else{
gvl = VSETVL(n);
vr = VFMVVF_FLOAT(0, gvl);
unsigned int stride_x = inc_x * sizeof(FLOAT);
unsigned int stride_y = inc_y * sizeof(FLOAT);
int stride_x = inc_x * sizeof(FLOAT);
int stride_y = inc_y * sizeof(FLOAT);
for(i=0,j=0; i<n/gvl; i++){
vx = VLSEV_FLOAT(&x[j*inc_x], stride_x, gvl);
vy = VLSEV_FLOAT(&y[j*inc_y], stride_y, gvl);
@@ -157,7 +163,8 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
}
if(j > 0){
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
dot += v_res[0];
dot += (double)VFMVFS_FLOAT(v_res);

}
//tail
if(j < n){
@@ -168,7 +175,8 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
//vr = VFDOTVV_FLOAT(vx, vy, gvl);
vr = VFMACCVV_FLOAT(vz, vx, vy, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
dot += v_res[0];
dot += (double)VFMVFS_FLOAT(v_res);

}
}
return(dot);


+ 10
- 7
kernel/riscv64/gemv_t_vector.c View File

@@ -31,9 +31,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLEV_FLOAT vle_v_f32m4
#define VLSEV_FLOAT vlse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredosum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
@@ -44,9 +45,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLEV_FLOAT vle_v_f64m4
#define VLSEV_FLOAT vlse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
@@ -80,7 +82,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
j += gvl;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp = v_res[0];
temp = (FLOAT)VFMVFS_FLOAT(v_res);
if(j < m){
gvl = VSETVL(m-j);
va = VLEV_FLOAT(&a_ptr[j], gvl);
@@ -88,7 +90,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
vr = VFMULVV_FLOAT(va, vx, gvl);

v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp += v_res[0];
temp += (FLOAT)VFMVFS_FLOAT(v_res);
}
y[iy] += alpha * temp;
iy += inc_y;
@@ -96,9 +98,10 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
}
}else{
BLASLONG stride_x = inc_x * sizeof(FLOAT);
BLASLONG inc_xv = inc_x * gvl;
for(i = 0; i < n; i++){
gvl = VSETVL(m);
BLASLONG inc_xv = inc_x * gvl;
j = 0;
ix = 0;
vr = VFMVVF_FLOAT(0, gvl);
@@ -110,7 +113,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
ix += inc_xv;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp = v_res[0];
temp = (FLOAT)VFMVFS_FLOAT(v_res);
if(j < m){
gvl = VSETVL(m-j);
va = VLEV_FLOAT(&a_ptr[j], gvl);
@@ -118,7 +121,7 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLO
vr = VFMULVV_FLOAT(va, vx, gvl);

v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp += v_res[0];
temp += (FLOAT)VFMVFS_FLOAT(v_res);
}
y[iy] += alpha * temp;
iy += inc_y;


+ 8
- 8
kernel/riscv64/iamax_vector.c View File

@@ -117,10 +117,10 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
j += gvl;
}
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_z0, gvl);
maxf = v_res[0];
maxf = *((FLOAT*)&v_res);
mask = VMFGEVF_FLOAT(v_max, maxf, gvl);
max_index = VMFIRSTM(mask,gvl);
max_index = v_max_index[max_index];
max_index = *((unsigned int*)&v_max_index+max_index);

if(j < n){
gvl = VSETVL(n-j);
@@ -130,7 +130,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v_max = VFRSUBVF_MASK_FLOAT(mask, vx, vx, 0, gvl);

v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_z0, gvl);
FLOAT cur_maxf = v_res[0];
FLOAT cur_maxf = *((FLOAT*)&v_res);
if(cur_maxf > maxf){
//tail index
v_max_index = VIDV_UINT(gvl);
@@ -138,7 +138,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)

mask = VMFGEVF_FLOAT(v_max, cur_maxf, gvl);
max_index = VMFIRSTM(mask,gvl);
max_index = v_max_index[max_index];
max_index = *((unsigned int*)&v_max_index+max_index);
}
}
}else{
@@ -165,10 +165,10 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
idx += inc_v;
}
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_z0, gvl);
maxf = v_res[0];
maxf = *((FLOAT*)&v_res);
mask = VMFGEVF_FLOAT(v_max, maxf, gvl);
max_index = VMFIRSTM(mask,gvl);
max_index = v_max_index[max_index];
max_index = *((unsigned int*)&v_max_index+max_index);

if(j < n){
gvl = VSETVL(n-j);
@@ -178,7 +178,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v_max = VFRSUBVF_MASK_FLOAT(mask, vx, vx, 0, gvl);

v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_z0, gvl);
FLOAT cur_maxf = v_res[0];
FLOAT cur_maxf = *((FLOAT*)&v_res);
if(cur_maxf > maxf){
//tail index
v_max_index = VIDV_UINT(gvl);
@@ -186,7 +186,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)

mask = VMFGEVF_FLOAT(v_max, cur_maxf, gvl);
max_index = VMFIRSTM(mask,gvl);
max_index = v_max_index[max_index];
max_index = *((unsigned int*)&v_max_index+max_index);
}
}
}


+ 8
- 8
kernel/riscv64/iamin_vector.c View File

@@ -118,10 +118,10 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
j += gvl;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = *((FLOAT*)&v_res);
mask = VMFLEVF_FLOAT(v_min, minf, gvl);
min_index = VMFIRSTM(mask,gvl);
min_index = v_min_index[min_index];
min_index = *((unsigned int*)&v_min_index+min_index);

if(j < n){
gvl = VSETVL(n-j);
@@ -131,7 +131,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v_min = VFRSUBVF_MASK_FLOAT(mask, vx, vx, 0, gvl);

v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
FLOAT cur_minf = v_res[0];
FLOAT cur_minf = *((FLOAT*)&v_res);
if(cur_minf < minf){
//tail index
v_min_index = VIDV_UINT(gvl);
@@ -139,7 +139,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)

mask = VMFLEVF_FLOAT(v_min, cur_minf, gvl);
min_index = VMFIRSTM(mask,gvl);
min_index = v_min_index[min_index];
min_index = *((unsigned int*)&v_min_index+min_index);
}
}
}else{
@@ -166,10 +166,10 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
idx += inc_v;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = *((FLOAT*)&v_res);
mask = VMFLEVF_FLOAT(v_min, minf, gvl);
min_index = VMFIRSTM(mask,gvl);
min_index = v_min_index[min_index];
min_index = *((unsigned int*)&v_min_index+min_index);

if(j < n){
gvl = VSETVL(n-j);
@@ -179,7 +179,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v_min = VFRSUBVF_MASK_FLOAT(mask, vx, vx, 0, gvl);

v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
FLOAT cur_minf = v_res[0];
FLOAT cur_minf = *((FLOAT*)&v_res);
if(cur_minf < minf){
//tail index
v_min_index = VIDV_UINT(gvl);
@@ -187,7 +187,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)

mask = VMFLEVF_FLOAT(v_min, cur_minf, gvl);
min_index = VMFIRSTM(mask,gvl);
min_index = v_min_index[min_index];
min_index = *((unsigned int*)&v_min_index+min_index);
}
}
}


+ 8
- 8
kernel/riscv64/imax_vector.c View File

@@ -111,17 +111,17 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
j += gvl;
}
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_min, gvl);
maxf = v_res[0];
maxf = *((FLOAT*)&v_res);
mask = VMFGEVF_FLOAT(v_max, maxf, gvl);
max_index = VMFIRSTM(mask,gvl);
max_index = v_max_index[max_index];
max_index = *((unsigned int*)&v_max_index+max_index);

if(j < n){
gvl = VSETVL(n-j);
v_max = VLEV_FLOAT(&x[j], gvl);

v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_min, gvl);
FLOAT cur_maxf = v_res[0];
FLOAT cur_maxf = *((FLOAT*)&v_res);
if(cur_maxf > maxf){
//tail index
v_max_index = VIDV_UINT(gvl);
@@ -129,7 +129,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)

mask = VMFGEVF_FLOAT(v_max, cur_maxf, gvl);
max_index = VMFIRSTM(mask,gvl);
max_index = v_max_index[max_index];
max_index = *((unsigned int*)&v_max_index+max_index);
}
}
}else{
@@ -153,17 +153,17 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
idx += inc_v;
}
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_min, gvl);
maxf = v_res[0];
maxf = *((FLOAT*)&v_res);
mask = VMFGEVF_FLOAT(v_max, maxf, gvl);
max_index = VMFIRSTM(mask,gvl);
max_index = v_max_index[max_index];
max_index = *((unsigned int*)&v_max_index+max_index);

if(j < n){
gvl = VSETVL(n-j);
v_max = VLSEV_FLOAT(&x[idx], stride_x, gvl);

v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_min, gvl);
FLOAT cur_maxf = v_res[0];
FLOAT cur_maxf = *((FLOAT*)&v_res);
if(cur_maxf > maxf){
//tail index
v_max_index = VIDV_UINT(gvl);
@@ -171,7 +171,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)

mask = VMFGEVF_FLOAT(v_max, cur_maxf, gvl);
max_index = VMFIRSTM(mask,gvl);
max_index = v_max_index[max_index];
max_index = *((unsigned int*)&v_max_index+max_index);
}
}
}


+ 8
- 8
kernel/riscv64/imin_vector.c View File

@@ -129,24 +129,24 @@ asm volatile(
j += gvl;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = *((FLOAT*)&v_res);
mask = VMFLEVF_FLOAT(v_min, minf, gvl);
min_index = VMFIRSTM(mask,gvl);
min_index = v_min_index[min_index];
min_index = *((unsigned int*)&v_min_index+min_index);

if(j < n){
gvl = VSETVL(n-j);
v_min = VLEV_FLOAT(&x[j], gvl);

v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
FLOAT cur_minf = v_res[0];
FLOAT cur_minf = *((FLOAT*)&v_res);
if(cur_minf < minf){
//tail index
v_min_index = VIDV_UINT(gvl);
v_min_index = VADDVX_UINT(v_min_index, j, gvl);
mask = VMFLEVF_FLOAT(v_min, cur_minf, gvl);
min_index = VMFIRSTM(mask,gvl);
min_index = v_min_index[min_index];
min_index = *((unsigned int*)&v_min_index+min_index);
}
}
}else{
@@ -190,24 +190,24 @@ asm volatile(
idx += inc_v;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = *((FLOAT*)&v_res);
mask = VMFLEVF_FLOAT(v_min, minf, gvl);
min_index = VMFIRSTM(mask,gvl);
min_index = v_min_index[min_index];
min_index = *((unsigned int*)&v_min_index+min_index);

if(j < n){
gvl = VSETVL(n-j);
v_min = VLSEV_FLOAT(&x[idx], stride_x, gvl);

v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
FLOAT cur_minf = v_res[0];
FLOAT cur_minf = *((FLOAT*)&v_res);
if(cur_minf < minf){
//tail index
v_min_index = VIDV_UINT(gvl);
v_min_index = VADDVX_UINT(v_min_index, j, gvl);
mask = VMFLEVF_FLOAT(v_min, cur_minf, gvl);
min_index = VMFIRSTM(mask,gvl);
min_index = v_min_index[min_index];
min_index = *((unsigned int*)&v_min_index+min_index);
}
}
}


+ 15
- 4
kernel/riscv64/izamax_vector.c View File

@@ -34,6 +34,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m8_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLSEV_FLOAT vlse_v_f64m8
#define VFREDMAXVS_FLOAT vfredmax_vs_f64m8_f64m1
#define MASK_T vbool8_t
@@ -46,6 +47,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VMFGEVF_FLOAT vmfge_vf_f64m8_b8
#define VMFIRSTM vmfirst_m_b8
#define UINT_V_T vuint64m8_t
#define VSEVU_UINT vse64_v_u64m8
#define UINT_T long unsigned int
#define VIDV_MASK_UINT vid_v_u64m8_m
#define VIDV_UINT vid_v_u64m8
#define VADDVX_MASK_UINT vadd_vx_u64m8_m
@@ -59,6 +62,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m8_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLSEV_FLOAT vlse_v_f32m8
#define VFREDMAXVS_FLOAT vfredmax_vs_f32m8_f32m1
#define MASK_T vbool4_t
@@ -71,6 +75,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VMFGEVF_FLOAT vmfge_vf_f32m8_b4
#define VMFIRSTM vmfirst_m_b4
#define UINT_V_T vuint32m8_t
#define UINT_T unsigned int
#define VSEVU_UINT vse32_v_u32m8
#define VIDV_MASK_UINT vid_v_u32m8_m
#define VIDV_UINT vid_v_u32m8
#define VADDVX_MASK_UINT vadd_vx_u32m8_m
@@ -98,6 +104,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v_z0 = VFMVVF_FLOAT_M1(0, gvl);

gvl = VSETVL(n);
UINT_T temp_uint[gvl];
v_max_index = VMVVX_UINT(0, gvl);
v_max = VFMVVF_FLOAT(-1, gvl);
BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
@@ -183,10 +190,12 @@ asm volatile(
}
vx0 = VFMVVF_FLOAT(0, gvl);
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_z0, gvl);
maxf = v_res[0];
maxf = VFMVFS_FLOAT(v_res);
mask0 = VMFGEVF_FLOAT(v_max, maxf, gvl);
max_index = VMFIRSTM(mask0,gvl);
max_index = v_max_index[max_index];
VSEVU_UINT(temp_uint,v_max_index,gvl);
max_index = temp_uint[max_index];


if(j < n){
gvl = VSETVL(n-j);
@@ -239,7 +248,7 @@ asm volatile(
*/
v_max = VFADDVV_FLOAT(vx0, vx1, gvl);
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_z0, gvl);
FLOAT cur_maxf = v_res[0];
FLOAT cur_maxf = VFMVFS_FLOAT(v_res);
if(cur_maxf > maxf){
//tail index
v_max_index = VIDV_UINT(gvl);
@@ -247,7 +256,9 @@ asm volatile(

mask0 = VMFGEVF_FLOAT(v_max, cur_maxf, gvl);
max_index = VMFIRSTM(mask0,gvl);
max_index = v_max_index[max_index];
VSEVU_UINT(temp_uint,v_max_index,gvl);
max_index = temp_uint[max_index];

}
}
return(max_index+1);


+ 14
- 4
kernel/riscv64/izamin_vector.c View File

@@ -35,6 +35,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m8_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLSEV_FLOAT vlse_v_f64m8
#define VFREDMINVS_FLOAT vfredmin_vs_f64m8_f64m1
#define MASK_T vbool8_t
@@ -47,6 +48,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VMFLEVF_FLOAT vmfle_vf_f64m8_b8
#define VMFIRSTM vmfirst_m_b8
#define UINT_V_T vuint64m8_t
#define VSEVU_UINT vse64_v_u64m8
#define UINT_T long unsigned int
#define VIDV_MASK_UINT vid_v_u64m8_m
#define VIDV_UINT vid_v_u64m8
#define VADDVX_MASK_UINT vadd_vx_u64m8_m
@@ -60,6 +63,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m8_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLSEV_FLOAT vlse_v_f32m8
#define VFREDMINVS_FLOAT vfredmin_vs_f32m8_f32m1
#define MASK_T vbool4_t
@@ -72,6 +76,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VMFLEVF_FLOAT vmfle_vf_f32m8_b4
#define VMFIRSTM vmfirst_m_b4
#define UINT_V_T vuint32m8_t
#define UINT_T unsigned int
#define VSEVU_UINT vse32_v_u32m8
#define VIDV_MASK_UINT vid_v_u32m8_m
#define VIDV_UINT vid_v_u32m8
#define VADDVX_MASK_UINT vadd_vx_u32m8_m
@@ -98,6 +104,7 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v_max = VFMVVF_FLOAT_M1(FLT_MAX, gvl);

gvl = VSETVL(n);
UINT_T temp_uint[gvl];
v_min_index = VMVVX_UINT(0, gvl);
v_min = VFMVVF_FLOAT(FLT_MAX, gvl);
BLASLONG stride_x = inc_x * 2 * sizeof(FLOAT);
@@ -182,10 +189,11 @@ asm volatile(
ix += inc_xv;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = VFMVFS_FLOAT(v_res);
mask0 = VMFLEVF_FLOAT(v_min, minf, gvl);
min_index = VMFIRSTM(mask0,gvl);
min_index = v_min_index[min_index];
VSEVU_UINT(temp_uint,v_min_index,gvl);
min_index = temp_uint[min_index];

if(j < n){
gvl = VSETVL(n-j);
@@ -238,7 +246,7 @@ asm volatile(
*/
v_min = VFADDVV_FLOAT(vx0, vx1, gvl);
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
FLOAT cur_minf = v_res[0];
FLOAT cur_minf = VFMVFS_FLOAT(v_res);
if(cur_minf < minf){
//tail index
v_min_index = VIDV_UINT(gvl);
@@ -246,7 +254,9 @@ asm volatile(

mask0 = VMFLEVF_FLOAT(v_min, cur_minf, gvl);
min_index = VMFIRSTM(mask0,gvl);
min_index = v_min_index[min_index];
VSEVU_UINT(temp_uint,v_min_index,gvl);
min_index = temp_uint[min_index];

}
}
return(min_index+1);


+ 6
- 6
kernel/riscv64/max_vector.c View File

@@ -77,14 +77,14 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
j += gvl * 2;
}
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_min, gvl);
maxf = v_res[0];
maxf = *((FLOAT*)&v_res);
}
for(;j<n;){
gvl = VSETVL(n-j);
v0 = VLEV_FLOAT(&x[j], gvl);
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_min, gvl);
if(v_res[0] > maxf)
maxf = v_res[0];
if(*((FLOAT*)&v_res) > maxf)
maxf = *((FLOAT*)&v_res);
j += gvl;
}
}else{
@@ -103,14 +103,14 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
idx += inc_xv * 2;
}
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_min, gvl);
maxf = v_res[0];
maxf = *((FLOAT*)&v_res);
}
for(;j<n;){
gvl = VSETVL(n-j);
v0 = VLSEV_FLOAT(&x[j*inc_x], stride_x, gvl);
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_min, gvl);
if(v_res[0] > maxf)
maxf = v_res[0];
if(*((FLOAT*)&v_res) > maxf)
maxf = *((FLOAT*)&v_res);
j += gvl;
}
}


+ 6
- 6
kernel/riscv64/min_vector.c View File

@@ -77,14 +77,14 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
j += gvl * 2;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = *((FLOAT*)&v_res);
}
for(;j<n;){
gvl = VSETVL(n-j);
v0 = VLEV_FLOAT(&x[j], gvl);
v_res = VFREDMINVS_FLOAT(v_res, v0, v_max, gvl);
if(v_res[0] < minf)
minf = v_res[0];
if(*((FLOAT*)&v_res) < minf)
minf = *((FLOAT*)&v_res);
j += gvl;
}
}else{
@@ -103,14 +103,14 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
idx += inc_xv * 2;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = *((FLOAT*)&v_res);
}
for(;j<n;){
gvl = VSETVL(n-j);
v0 = VLSEV_FLOAT(&x[j*inc_x], stride_x, gvl);
v_res = VFREDMINVS_FLOAT(v_res, v0, v_max, gvl);
if(v_res[0] < minf)
minf = v_res[0];
if(*((FLOAT*)&v_res) < minf)
minf = *((FLOAT*)&v_res);
j += gvl;
}
}


+ 20
- 16
kernel/riscv64/nrm2_vector.c View File

@@ -30,10 +30,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL(n) vsetvl_e32m4(n)
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define VFMVFS_FLOATM4 vfmv_f_s_f32m4_f32
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLEV_FLOAT vle_v_f32m4
#define VLSEV_FLOAT vlse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
@@ -50,10 +52,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL(n) vsetvl_e64m4(n)
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define VFMVFS_FLOATM4 vfmv_f_s_f64m4_f64
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLEV_FLOAT vle_v_f64m4
#define VLSEV_FLOAT vlse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
@@ -106,13 +110,13 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq before current vector
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = v_res[0];
scale = VFMVFS_FLOAT(v_res);
//ssq in vector vr
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
@@ -122,7 +126,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq now
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);

//tail
if(j < n){
@@ -141,16 +145,16 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = v_res[0];
scale = VFMVFS_FLOAT(v_res);
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
}
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq now
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
}
}else{
gvl = VSETVL(n);
@@ -175,13 +179,13 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq before current vector
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = v_res[0];
scale = VFMVFS_FLOAT(v_res);
//ssq in vector vr
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
@@ -192,7 +196,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq now
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);

//tail
if(j < n){
@@ -211,16 +215,16 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = vr[0];
scale = VFMVFS_FLOATM4(vr);
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
}
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq now
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
}
}
return(scale * sqrt(ssq));


+ 8
- 6
kernel/riscv64/nrm2_vector_dot.c View File

@@ -31,9 +31,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m8_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLEV_FLOAT vle_v_f32m8
#define VLSEV_FLOAT vlse_v_f32m8
#define VFREDSUM_FLOAT vfredsum_vs_f32m8_f32m1
#define VFREDSUM_FLOAT vfredusum_vs_f32m8_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m8
#define VFMVVF_FLOAT vfmv_v_f_f32m8
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
@@ -44,9 +45,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m8_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLEV_FLOAT vle_v_f64m8
#define VLSEV_FLOAT vlse_v_f64m8
#define VFREDSUM_FLOAT vfredsum_vs_f64m8_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m8_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m8
#define VFMVVF_FLOAT vfmv_v_f_f64m8
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
@@ -83,7 +85,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
j += gvl;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
len += v_res[0];
len += VFMVFS_FLOAT(v_res);
}
//tail
for(;j < n;){
@@ -94,7 +96,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//vr = VFDOTVV_FLOAT(v0, v0, gvl);
vr = VFMACCVV_FLOAT(v1, v0, v0, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
len += v_res[0];
len += VFMVFS_FLOAT(v_res);

j += gvl;
}
@@ -113,7 +115,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
j += gvl;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
len += v_res[0];
len += VFMVFS_FLOAT(v_res);
}
//tail
for(;j < n;){
@@ -124,7 +126,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//vr = VFDOTVV_FLOAT(v0, v0, gvl);
vr = VFMACCVV_FLOAT(v1, v0, v0, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
len += v_res[0];
len += VFMVFS_FLOAT(v_res);

j += gvl;
}


+ 8
- 3
kernel/riscv64/scal_vector.c View File

@@ -85,19 +85,24 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
}else{
if(da == 0.0){
gvl = VSETVL(n);
BLASLONG stride_x = inc_x * sizeof(FLOAT);
BLASLONG ix = 0;
if(gvl <= n / 2){
long int inc_xv = gvl * inc_x;
v0 = VFMVVF_FLOAT(0, gvl);
for(i = 0, j = 0; i < n/(2*gvl); i++, j+=2*gvl){
VSEV_FLOAT(&x[j], v0, gvl);
VSEV_FLOAT(&x[j+gvl], v0, gvl);
VSSEV_FLOAT(&x[ix], stride_x, v0, gvl);
VSSEV_FLOAT(&x[ix + inc_xv], stride_x, v0, gvl);
ix += inc_xv * 2;
}
}
//tail
for(; j <n; ){
gvl = VSETVL(n-j);
v0 = VFMVVF_FLOAT(0, gvl);
VSEV_FLOAT(&x[j], v0, gvl);
VSSEV_FLOAT(&x[ix], stride_x, v0, gvl);
j += gvl;
ix += inc_x * gvl;
}
}else{
gvl = VSETVL(n);


+ 1
- 1
kernel/riscv64/sgemm_kernel_16x4_c910v.c View File

@@ -1,5 +1,5 @@
#include "common.h"
#include <riscv-vector.h>
#include <riscv_vector.h>

#define KERNEL16x4_I \
"addi t1, %[PB], 1*4 \n\t"\


+ 12
- 10
kernel/riscv64/symv_L_vector.c View File

@@ -31,11 +31,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLEV_FLOAT vle_v_f32m4
#define VLSEV_FLOAT vlse_v_f32m4
#define VSEV_FLOAT vse_v_f32m4
#define VSSEV_FLOAT vsse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFMACCVF_FLOAT vfmacc_vf_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
@@ -46,11 +47,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLEV_FLOAT vle_v_f64m4
#define VLSEV_FLOAT vlse_v_f64m4
#define VSEV_FLOAT vse_v_f64m4
#define VSSEV_FLOAT vsse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFMACCVF_FLOAT vfmacc_vf_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
@@ -98,7 +100,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
i += gvl;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 = v_res[0];
temp2 = VFMVFS_FLOAT(v_res);
if(i < m){
gvl = VSETVL(m-i);
vy = VLEV_FLOAT(&y[i], gvl);
@@ -109,7 +111,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
vx = VLEV_FLOAT(&x[i], gvl);
vr = VFMULVV_FLOAT(vx, va, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 += v_res[0];
temp2 += VFMVFS_FLOAT(v_res);
}
}
y[j] += alpha * temp2;
@@ -143,7 +145,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
iy += inc_yv;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 = v_res[0];
temp2 = VFMVFS_FLOAT(v_res);
if(i < m){
gvl = VSETVL(m-i);
vy = VLSEV_FLOAT(&y[iy], stride_y, gvl);
@@ -154,7 +156,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
vx = VLEV_FLOAT(&x[i], gvl);
vr = VFMULVV_FLOAT(vx, va, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 += v_res[0];
temp2 += VFMVFS_FLOAT(v_res);
}
}
y[jy] += alpha * temp2;
@@ -189,7 +191,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
ix += inc_xv;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 = v_res[0];
temp2 = VFMVFS_FLOAT(v_res);
if(i < m){
gvl = VSETVL(m-i);
vy = VLEV_FLOAT(&y[i], gvl);
@@ -200,7 +202,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
vx = VLSEV_FLOAT(&x[ix], stride_x, gvl);
vr = VFMULVV_FLOAT(vx, va, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 += v_res[0];
temp2 += VFMVFS_FLOAT(v_res);
}
}
y[j] += alpha * temp2;
@@ -240,7 +242,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
iy += inc_yv;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 = v_res[0];
temp2 = VFMVFS_FLOAT(v_res);
if(i < m){
gvl = VSETVL(m-i);
vy = VLSEV_FLOAT(&y[iy], stride_y, gvl);
@@ -251,7 +253,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
vx = VLSEV_FLOAT(&x[ix], stride_x, gvl);
vr = VFMULVV_FLOAT(vx, va, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 += v_res[0];
temp2 += VFMVFS_FLOAT(v_res);
}
}
y[jy] += alpha * temp2;


+ 12
- 10
kernel/riscv64/symv_U_vector.c View File

@@ -31,11 +31,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLEV_FLOAT vle_v_f32m4
#define VLSEV_FLOAT vlse_v_f32m4
#define VSEV_FLOAT vse_v_f32m4
#define VSSEV_FLOAT vsse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFMACCVF_FLOAT vfmacc_vf_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
@@ -47,11 +48,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLEV_FLOAT vle_v_f64m4
#define VLSEV_FLOAT vlse_v_f64m4
#define VSEV_FLOAT vse_v_f64m4
#define VSSEV_FLOAT vsse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFMACCVF_FLOAT vfmacc_vf_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
@@ -100,7 +102,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
i += gvl;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 = v_res[0];
temp2 = VFMVFS_FLOAT(v_res);
if(i < j){
gvl = VSETVL(j-i);
vy = VLEV_FLOAT(&y[i], gvl);
@@ -111,7 +113,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
vx = VLEV_FLOAT(&x[i], gvl);
vr = VFMULVV_FLOAT(vx, va, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 += v_res[0];
temp2 += VFMVFS_FLOAT(v_res);
}
}
y[j] += temp1 * a_ptr[j] + alpha * temp2;
@@ -144,7 +146,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
iy += inc_yv;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 = v_res[0];
temp2 = VFMVFS_FLOAT(v_res);
if(i < j){
gvl = VSETVL(j-i);
vy = VLSEV_FLOAT(&y[iy], stride_y, gvl);
@@ -155,7 +157,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
vx = VLEV_FLOAT(&x[i], gvl);
vr = VFMULVV_FLOAT(vx, va, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 += v_res[0];
temp2 += VFMVFS_FLOAT(v_res);
}
}
y[jy] += temp1 * a_ptr[j] + alpha * temp2;
@@ -189,7 +191,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
ix += inc_xv;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 = v_res[0];
temp2 = VFMVFS_FLOAT(v_res);
if(i < j){
gvl = VSETVL(j-i);
vy = VLEV_FLOAT(&y[i], gvl);
@@ -200,7 +202,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
vx = VLSEV_FLOAT(&x[ix], stride_x, gvl);
vr = VFMULVV_FLOAT(vx, va, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 += v_res[0];
temp2 += VFMVFS_FLOAT(v_res);
}
}
y[j] += temp1 * a_ptr[j] + alpha * temp2;
@@ -239,7 +241,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
iy += inc_yv;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 = v_res[0];
temp2 = VFMVFS_FLOAT(v_res);
if(i < j){
gvl = VSETVL(j-i);
vy = VLSEV_FLOAT(&y[iy], stride_y, gvl);
@@ -250,7 +252,7 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOA
vx = VLSEV_FLOAT(&x[ix], stride_x, gvl);
vr = VFMULVV_FLOAT(vx, va, gvl);
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp2 += v_res[0];
temp2 += VFMVFS_FLOAT(v_res);
}
}
y[jy] += temp1 * a_ptr[j] + alpha * temp2;


+ 8
- 3
kernel/riscv64/zamax_vector.c View File

@@ -33,6 +33,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m8_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLSEV_FLOAT vlse_v_f32m8
#define VFREDMAXVS_FLOAT vfredmax_vs_f32m8_f32m1
#define MASK_T vbool4_t
@@ -42,11 +43,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VFRSUBVF_MASK_FLOAT vfrsub_vf_f32m8_m
#define VFMAXVV_FLOAT vfmax_vv_f32m8
#define VFADDVV_FLOAT vfadd_vv_f32m8

#else
#define VSETVL(n) vsetvl_e64m8(n)
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m8_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLSEV_FLOAT vlse_v_f64m8
#define VFREDMAXVS_FLOAT vfredmax_vs_f64m8_f64m1
#define MASK_T vbool8_t
@@ -56,6 +59,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VFRSUBVF_MASK_FLOAT vfrsub_vf_f64m8_m
#define VFMAXVV_FLOAT vfmax_vv_f64m8
#define VFADDVV_FLOAT vfadd_vv_f64m8

#endif

FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
@@ -91,7 +95,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
ix += inc_xv;
}
v_res = VFREDMAXVS_FLOAT(v_res, v_max, v_z0, gvl);
maxf = v_res[0];
maxf = VFMVFS_FLOAT(v_res);

if(j<n){
gvl = VSETVL(n-j);
@@ -103,8 +107,9 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v1 = VFRSUBVF_MASK_FLOAT(mask1, v1, v1, 0, gvl);
v1 = VFADDVV_FLOAT(v0, v1, gvl);
v_res = VFREDMAXVS_FLOAT(v_res, v1, v_z0, gvl);
if(v_res[0] > maxf)
maxf = v_res[0];
if(VFMVFS_FLOAT(v_res)> maxf)
maxf = VFMVFS_FLOAT(v_res);
}
return(maxf);
}

+ 5
- 3
kernel/riscv64/zamin_vector.c View File

@@ -34,6 +34,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m8_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLSEV_FLOAT vlse_v_f32m8
#define VFREDMINVS_FLOAT vfredmin_vs_f32m8_f32m1
#define MASK_T vbool4_t
@@ -48,6 +49,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat64m8_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLSEV_FLOAT vlse_v_f64m8
#define VFREDMINVS_FLOAT vfredmin_vs_f64m8_f64m1
#define MASK_T vbool8_t
@@ -92,7 +94,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
ix += inc_xv;
}
v_res = VFREDMINVS_FLOAT(v_res, v_min, v_max, gvl);
minf = v_res[0];
minf = VFMVFS_FLOAT(v_res);

if(j<n){
gvl = VSETVL(n-j);
@@ -104,8 +106,8 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v1 = VFRSUBVF_MASK_FLOAT(mask1, v1, v1, 0, gvl);
v1 = VFADDVV_FLOAT(v0, v1, gvl);
v_res = VFREDMINVS_FLOAT(v_res, v1, v_max, gvl);
if(v_res[0] < minf)
minf = v_res[0];
if(VFMVFS_FLOAT(v_res) < minf)
minf = VFMVFS_FLOAT(v_res);
}
return(minf);
}

+ 8
- 6
kernel/riscv64/zasum_vector.c View File

@@ -33,9 +33,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m8_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLEV_FLOAT vle_v_f32m8
#define VLSEV_FLOAT vlse_v_f32m8
#define VFREDSUMVS_FLOAT vfredsum_vs_f32m8_f32m1
#define VFREDSUMVS_FLOAT vfredusum_vs_f32m8_f32m1
#define MASK_T vbool4_t
#define VMFLTVF_FLOAT vmflt_vf_f32m8_b4
#define VFMVVF_FLOAT vfmv_v_f_f32m8
@@ -47,9 +48,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m8_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLEV_FLOAT vle_v_f64m8
#define VLSEV_FLOAT vlse_v_f64m8
#define VFREDSUMVS_FLOAT vfredsum_vs_f64m8_f64m1
#define VFREDSUMVS_FLOAT vfredusum_vs_f64m8_f64m1
#define MASK_T vbool8_t
#define VMFLTVF_FLOAT vmflt_vf_f64m8_b8
#define VFMVVF_FLOAT vfmv_v_f_f64m8
@@ -90,7 +92,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
j += gvl * 2;
}
v_res = VFREDSUMVS_FLOAT(v_res, v_sum, v_z0, gvl);
asumf += v_res[0];
asumf += VFFMVFS_FLOAT(v_res);
}
for(;j<n2;){
gvl = VSETVL(n2-j);
@@ -98,7 +100,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
mask0 = VMFLTVF_FLOAT(v0, 0, gvl);
v0 = VFRSUBVF_MASK_FLOAT(mask0, v0, v0, 0, gvl);
v_res = VFREDSUMVS_FLOAT(v_res, v0, v_z0, gvl);
asumf += v_res[0];
asumf += VFFMVFS_FLOAT(v_res);
j += gvl;
}
}else{
@@ -123,7 +125,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
ix += inc_xv;
}
v_res = VFREDSUMVS_FLOAT(v_res, v_sum, v_z0, gvl);
asumf += v_res[0];
asumf += VFFMVFS_FLOAT(v_res);
if(j<n){
gvl = VSETVL(n-j);
v0 = VLSEV_FLOAT(&x[ix], stride_x, gvl);
@@ -135,7 +137,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
v1 = VFRSUBVF_MASK_FLOAT(mask1, v1, v1, 0, gvl);
v_sum = VFADDVV_FLOAT(v0, v1, gvl);
v_res = VFREDSUMVS_FLOAT(v_res, v_sum, v_z0, gvl);
asumf += v_res[0];
asumf += VFFMVFS_FLOAT(v_res);
}
}
return(asumf);


+ 8
- 6
kernel/riscv64/zdot_vector.c View File

@@ -31,9 +31,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLEV_FLOAT vle_v_f32m4
#define VLSEV_FLOAT vlse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
@@ -46,9 +47,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLEV_FLOAT vle_v_f64m4
#define VLSEV_FLOAT vlse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
@@ -108,9 +110,9 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
iy += inc_yv;
}
v_res = VFREDSUM_FLOAT(v_res, vr0, v_z0, gvl);
dot[0] += v_res[0];
dot[0] += VFMVFS_FLOAT(v_res);
v_res = VFREDSUM_FLOAT(v_res, vr1, v_z0, gvl);
dot[1] += v_res[0];
dot[1] += VFMVFS_FLOAT(v_res);
//tail
if(j < n){
gvl = VSETVL(n-j);
@@ -131,9 +133,9 @@ OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLA
vr1 = VFMSACVV_FLOAT(vr1, vx0, vy1, gvl);
#endif
v_res = VFREDSUM_FLOAT(v_res, vr0, v_z0, gvl);
dot[0] += v_res[0];
dot[0] += VFMVFS_FLOAT(v_res);
v_res = VFREDSUM_FLOAT(v_res, vr1, v_z0, gvl);
dot[1] += v_res[0];
dot[1] += VFMVFS_FLOAT(v_res);
}
CREAL(result) = dot[0];
CIMAG(result) = dot[1];


+ 8
- 6
kernel/riscv64/zgemv_t_vector.c View File

@@ -31,8 +31,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLSEV_FLOAT vlse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFNMSACVV_FLOAT vfnmsac_vv_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
@@ -43,8 +44,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLSEV_FLOAT vlse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFNMSACVV_FLOAT vfnmsac_vv_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
@@ -100,9 +102,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,
ix += inc_xv;
}
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp_r = v_res[0];
temp_r = VFMVFS_FLOAT(v_res);
v_res = VFREDSUM_FLOAT(v_res, vi, v_z0, gvl);
temp_i = v_res[0];
temp_i = VFMVFS_FLOAT(v_res);
if(j/2 < m){
gvl = VSETVL(m-j/2);
va0 = VLSEV_FLOAT(&a_ptr[j], stride_a, gvl);
@@ -122,9 +124,9 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i,

#endif
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
temp_r += v_res[0];
temp_r += VFMVFS_FLOAT(v_res);
v_res = VFREDSUM_FLOAT(v_res, vi, v_z0, gvl);
temp_i += v_res[0];
temp_i += VFMVFS_FLOAT(v_res);
}
#if !defined(XCONJ)
y[iy] += alpha_r * temp_r - alpha_i * temp_i;


+ 8
- 6
kernel/riscv64/zhemv_LM_vector.c View File

@@ -31,9 +31,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLSEV_FLOAT vlse_v_f32m4
#define VSSEV_FLOAT vsse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFMACCVF_FLOAT vfmacc_vf_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
@@ -46,9 +47,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLSEV_FLOAT vlse_v_f64m4
#define VSSEV_FLOAT vsse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFMACCVF_FLOAT vfmacc_vf_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
@@ -142,9 +144,9 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, B
ia += inc_av;
}
v_res = VFREDSUM_FLOAT(v_res, vr0, v_z0, gvl);
temp_r2 = v_res[0];
temp_r2 = VFMVFS_FLOAT(v_res);
v_res = VFREDSUM_FLOAT(v_res, vr1, v_z0, gvl);
temp_i2 = v_res[0];
temp_i2 = VFMVFS_FLOAT(v_res);
if(i < m){
gvl = VSETVL(m-i);
va0 = VLSEV_FLOAT(&a_ptr[ia], stride_a, gvl);
@@ -180,9 +182,9 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, B
#endif

v_res = VFREDSUM_FLOAT(v_res, vr0, v_z0, gvl);
temp_r2 += v_res[0];
temp_r2 += VFMVFS_FLOAT(v_res);
v_res = VFREDSUM_FLOAT(v_res, vr1, v_z0, gvl);
temp_i2 += v_res[0];
temp_i2 += VFMVFS_FLOAT(v_res);
}
}
y[jy] += alpha_r * temp_r2 - alpha_i * temp_i2;


+ 8
- 6
kernel/riscv64/zhemv_UV_vector.c View File

@@ -31,9 +31,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLSEV_FLOAT vlse_v_f32m4
#define VSSEV_FLOAT vsse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFMACCVF_FLOAT vfmacc_vf_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
@@ -46,9 +47,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLSEV_FLOAT vlse_v_f64m4
#define VSSEV_FLOAT vsse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFMACCVF_FLOAT vfmacc_vf_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
@@ -141,9 +143,9 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, B
ia += inc_av;
}
v_res = VFREDSUM_FLOAT(v_res, vr0, v_z0, gvl);
temp_r2 = v_res[0];
temp_r2 = VFMVFS_FLOAT(v_res);
v_res = VFREDSUM_FLOAT(v_res, vr1, v_z0, gvl);
temp_i2 = v_res[0];
temp_i2 = VFMVFS_FLOAT(v_res);
if(i < j){
gvl = VSETVL(j-i);
va0 = VLSEV_FLOAT(&a_ptr[ia], stride_a, gvl);
@@ -179,9 +181,9 @@ int CNAME(BLASLONG m, BLASLONG offset, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, B
#endif

v_res = VFREDSUM_FLOAT(v_res, vr0, v_z0, gvl);
temp_r2 += v_res[0];
temp_r2 += VFMVFS_FLOAT(v_res);
v_res = VFREDSUM_FLOAT(v_res, vr1, v_z0, gvl);
temp_i2 += v_res[0];
temp_i2 += VFMVFS_FLOAT(v_res);
}
}
y[jy] += temp_r1 * a_ptr[ja];


+ 24
- 22
kernel/riscv64/znrm2_vector.c View File

@@ -31,9 +31,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e32m1()
#define FLOAT_V_T vfloat32m4_t
#define FLOAT_V_T_M1 vfloat32m1_t
#define VFMVFS_FLOAT vfmv_f_s_f32m1_f32
#define VLEV_FLOAT vle_v_f32m4
#define VLSEV_FLOAT vlse_v_f32m4
#define VFREDSUM_FLOAT vfredsum_vs_f32m4_f32m1
#define VFREDSUM_FLOAT vfredusum_vs_f32m4_f32m1
#define VFMACCVV_FLOAT vfmacc_vv_f32m4
#define VFMVVF_FLOAT vfmv_v_f_f32m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f32m1
@@ -51,9 +52,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL_MAX vsetvlmax_e64m1()
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_M1 vfloat64m1_t
#define VFMVFS_FLOAT vfmv_f_s_f64m1_f64
#define VLEV_FLOAT vle_v_f64m4
#define VLSEV_FLOAT vlse_v_f64m4
#define VFREDSUM_FLOAT vfredsum_vs_f64m4_f64m1
#define VFREDSUM_FLOAT vfredusum_vs_f64m4_f64m1
#define VFMACCVV_FLOAT vfmacc_vv_f64m4
#define VFMVVF_FLOAT vfmv_v_f_f64m4
#define VFMVVF_FLOAT_M1 vfmv_v_f_f64m1
@@ -107,13 +109,13 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq before current vector
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = v_res[0];
scale = VFMVFS_FLOAT(v_res);
//ssq in vector vr
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
@@ -123,7 +125,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq now
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);

//tail
if(j < n2){
@@ -142,16 +144,16 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = v_res[0];
scale = VFMVFS_FLOAT(v_res);
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
}
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq now
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
}
}else{
gvl = VSETVL(n);
@@ -176,13 +178,13 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq before current vector
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = v_res[0];
scale = VFMVFS_FLOAT(v_res);
//ssq in vector vr
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
@@ -204,13 +206,13 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq before current vector
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = v_res[0];
scale = VFMVFS_FLOAT(v_res);
//ssq in vector vr
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
@@ -221,7 +223,7 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq now
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);

//tail
if(j < n){
@@ -242,9 +244,9 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = v_res[0];
scale = VFMVFS_FLOAT(v_res);
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
}
@@ -265,20 +267,20 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq before current vector
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
//find max
v_res = VFREDMAXVS_FLOAT(v_res, v0, v_z0, gvl);
//update ssq before max_index
ssq = ssq * (scale/v_res[0])*(scale/v_res[0]);
ssq = ssq * (scale/VFMVFS_FLOAT(v_res))*(scale/VFMVFS_FLOAT(v_res));
//update scale
scale = v_res[0];
scale = VFMVFS_FLOAT(v_res);
v0 = VFDIVVF_FLOAT(v0, scale, gvl);
vr = VFMACCVV_FLOAT(v_zero, v0, v0, gvl);
}
//ssq in vector vr: vr[0]
v_res = VFREDSUM_FLOAT(v_res, vr, v_z0, gvl);
//total ssq now
ssq += v_res[0];
ssq += VFMVFS_FLOAT(v_res);
}
}
return(scale * sqrt(ssq));


Write Preview
Loading…
Cancel
Save