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

fix axpy implementations where y has a stride of 0

This commit is contained in:
Sergei Lewis 2024-02-16 15:58:02 +00:00
parent 5266998b9f
commit ba17758c02
2 changed files with 47 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
kernel/riscv64/axpy_rvv.c
View File

@ -30,19 +30,29 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#if !defined(DOUBLE) #if !defined(DOUBLE)
#define VSETVL(n) __riscv_vsetvl_e32m8(n) #define VSETVL(n) __riscv_vsetvl_e32m8(n)
#define FLOAT_V_T vfloat32m8_t #define FLOAT_V_T vfloat32m8_t
#define FLOAT_V_M1_T vfloat32m1_t
#define VLEV_FLOAT __riscv_vle32_v_f32m8 #define VLEV_FLOAT __riscv_vle32_v_f32m8
#define VLSEV_FLOAT __riscv_vlse32_v_f32m8 #define VLSEV_FLOAT __riscv_vlse32_v_f32m8
#define VSEV_FLOAT __riscv_vse32_v_f32m8 #define VSEV_FLOAT __riscv_vse32_v_f32m8
#define VSEV_FLOAT_M1 __riscv_vse32_v_f32m1
#define VSSEV_FLOAT __riscv_vsse32_v_f32m8 #define VSSEV_FLOAT __riscv_vsse32_v_f32m8
#define VFMACCVF_FLOAT __riscv_vfmacc_vf_f32m8 #define VFMACCVF_FLOAT __riscv_vfmacc_vf_f32m8
#define VFMVVF_FLOAT __riscv_vfmv_v_f_f32m8
#define VFREDSUMVS_FLOAT __riscv_vfredusum_vs_f32m8_f32m1
#define VFMVVF_FLOAT_M1 __riscv_vfmv_v_f_f32m1
#else #else
#define VSETVL(n) __riscv_vsetvl_e64m8(n) #define VSETVL(n) __riscv_vsetvl_e64m8(n)
#define FLOAT_V_T vfloat64m8_t #define FLOAT_V_T vfloat64m8_t
#define FLOAT_V_M1_T vfloat64m1_t
#define VLEV_FLOAT __riscv_vle64_v_f64m8 #define VLEV_FLOAT __riscv_vle64_v_f64m8
#define VLSEV_FLOAT __riscv_vlse64_v_f64m8 #define VLSEV_FLOAT __riscv_vlse64_v_f64m8
#define VSEV_FLOAT __riscv_vse64_v_f64m8 #define VSEV_FLOAT __riscv_vse64_v_f64m8
#define VSEV_FLOAT_M1 __riscv_vse64_v_f64m1
#define VSSEV_FLOAT __riscv_vsse64_v_f64m8 #define VSSEV_FLOAT __riscv_vsse64_v_f64m8
#define VFMACCVF_FLOAT __riscv_vfmacc_vf_f64m8 #define VFMACCVF_FLOAT __riscv_vfmacc_vf_f64m8
#define VFMVVF_FLOAT __riscv_vfmv_v_f_f64m8
#define VFREDSUMVS_FLOAT __riscv_vfredusum_vs_f64m8_f64m1
#define VFMVVF_FLOAT_M1 __riscv_vfmv_v_f_f64m1
#endif #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) int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
@ -76,7 +86,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
VSEV_FLOAT(y, vy, vl); VSEV_FLOAT(y, vy, vl);
} }
} else if (1 == inc_x) { } else if (1 == inc_x && 0 != inc_y) {
BLASLONG stride_y = inc_y * sizeof(FLOAT); BLASLONG stride_y = inc_y * sizeof(FLOAT);
@ -89,8 +99,20 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
VSSEV_FLOAT(y, stride_y, vy, vl); VSSEV_FLOAT(y, stride_y, vy, vl);
} }
} else { } else if( 0 == inc_y ) {
BLASLONG stride_x = inc_x * sizeof(FLOAT);
size_t in_vl = VSETVL(n);
vy = VFMVVF_FLOAT( y[0], in_vl );
for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {
vl = VSETVL(n);
vx = VLSEV_FLOAT(x, stride_x, vl);
vy = VFMACCVF_FLOAT(vy, da, vx, vl);
}
FLOAT_V_M1_T vres = VFMVVF_FLOAT_M1( 0.0f, 1 );
vres = VFREDSUMVS_FLOAT( vy, vres, in_vl );
VSEV_FLOAT_M1(y, vres, 1);
} else {
BLASLONG stride_x = inc_x * sizeof(FLOAT); BLASLONG stride_x = inc_x * sizeof(FLOAT);
BLASLONG stride_y = inc_y * sizeof(FLOAT); BLASLONG stride_y = inc_y * sizeof(FLOAT);

24
kernel/riscv64/axpy_vector.c
View File

@ -51,11 +51,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define VSETVL JOIN(RISCV_RVV(vsetvl), _e, ELEN, LMUL, _) #define VSETVL JOIN(RISCV_RVV(vsetvl), _e, ELEN, LMUL, _)
#define FLOAT_V_T JOIN(vfloat, ELEN, LMUL, _t, _) #define FLOAT_V_T JOIN(vfloat, ELEN, LMUL, _t, _)
#define FLOAT_V_M1_T JOIN(vfloat, ELEN, m1, _t, _)
#define VLEV_FLOAT JOIN(RISCV_RVV(vle), ELEN, _v_f, ELEN, LMUL) #define VLEV_FLOAT JOIN(RISCV_RVV(vle), ELEN, _v_f, ELEN, LMUL)
#define VLSEV_FLOAT JOIN(RISCV_RVV(vlse), ELEN, _v_f, ELEN, LMUL) #define VLSEV_FLOAT JOIN(RISCV_RVV(vlse), ELEN, _v_f, ELEN, LMUL)
#define VSEV_FLOAT JOIN(RISCV_RVV(vse), ELEN, _v_f, ELEN, LMUL) #define VSEV_FLOAT JOIN(RISCV_RVV(vse), ELEN, _v_f, ELEN, LMUL)
#define VSSEV_FLOAT JOIN(RISCV_RVV(vsse), ELEN, _v_f, ELEN, LMUL) #define VSSEV_FLOAT JOIN(RISCV_RVV(vsse), ELEN, _v_f, ELEN, LMUL)
#define VFMACCVF_FLOAT JOIN(RISCV_RVV(vfmacc), _vf_f, ELEN, LMUL, _) #define VFMACCVF_FLOAT JOIN(RISCV_RVV(vfmacc), _vf_f, ELEN, LMUL, _)
#define VFMVVF_FLOAT JOIN(RISCV_RVV(vfmv), _v_f_f, ELEN, LMUL, _)
#define VFMVVF_FLOAT_M1 JOIN(RISCV_RVV(vfmv), _v_f_f, ELEN, m1, _)
#ifdef RISCV_0p10_INTRINSICS
#define VFREDSUMVS_FLOAT(va, vb, gvl) JOIN(RISCV_RVV(vfredusum_vs_f), ELEN, LMUL, _f, JOIN2( ELEN, m1))(v_res, va, vb, gvl)
#else
#define VFREDSUMVS_FLOAT JOIN(RISCV_RVV(vfredusum_vs_f), ELEN, LMUL, _f, JOIN2( ELEN, m1))
#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) int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
{ {
@ -123,7 +132,7 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
VSEV_FLOAT(&y[j], vy0, gvl); VSEV_FLOAT(&y[j], vy0, gvl);
j += gvl; j += gvl;
} }
}else if(inc_x == 1){ } else if (1 == inc_x && 0 != inc_y) {
stride_y = inc_y * sizeof(FLOAT); stride_y = inc_y * sizeof(FLOAT);
gvl = VSETVL(n); gvl = VSETVL(n);
if(gvl <= n/2){ if(gvl <= n/2){
@ -151,6 +160,19 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
VSSEV_FLOAT(&y[j*inc_y], stride_y, vy0, gvl); VSSEV_FLOAT(&y[j*inc_y], stride_y, vy0, gvl);
j += gvl; j += gvl;
} }
} else if( 0 == inc_y ) {
BLASLONG stride_x = inc_x * sizeof(FLOAT);
size_t in_vl = VSETVL(n);
vy0 = VFMVVF_FLOAT( y[0], in_vl );
for (size_t vl; n > 0; n -= vl, x += vl*inc_x) {
vl = VSETVL(n);
vx0 = VLSEV_FLOAT(x, stride_x, vl);
vy0 = VFMACCVF_FLOAT(vy0, da, vx0, vl);
}
FLOAT_V_M1_T v_res = VFMVVF_FLOAT_M1( 0.0f, 1 );
v_res = VFREDSUMVS_FLOAT( vy0, v_res, in_vl );
y[0] = EXTRACT_FLOAT(v_res);
}else{ }else{
stride_x = inc_x * sizeof(FLOAT); stride_x = inc_x * sizeof(FLOAT);
stride_y = inc_y * sizeof(FLOAT); stride_y = inc_y * sizeof(FLOAT);