Skip to content

Commit

Permalink
Merge pull request #5074 from tingboliao/develop
Browse files Browse the repository at this point in the history 
Optimize the gemm_tcopy_8_rvv to be compatible with the vlens 128 and 256.
  • Loading branch information
@martin-frbg
martin-frbg authored Jan 16, 2025
2 parents eba7338 + ef7f54b commit a7483d1
Showing 1 changed file with 25 additions and 189 deletions.
214 changes: 25 additions & 189 deletions kernel/riscv64/gemm_tcopy_8_rvv.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -28,35 +28,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "common.h"

#if !defined(DOUBLE)
#define VSETVL(n) __riscv_vsetvl_e32m1(n)
#define FLOAT_V_T vfloat32m1_t
#define FLOAT_VX2_T vfloat32m1x2_t
#define FLOAT_VX4_T vfloat32m1x4_t
#define FLOAT_VX8_T vfloat32m1x8_t
#define VLEV_FLOAT __riscv_vle32_v_f32m1
#define VLSEV_FLOAT __riscv_vlse32_v_f32m1
#define VSEV_FLOAT __riscv_vse32_v_f32m1
#define VLSSEG2_FLOAT __riscv_vlsseg2e32_v_f32m1x2
#define VSSEG2_FLOAT __riscv_vsseg2e32_v_f32m1x2
#define VLSSEG4_FLOAT __riscv_vlsseg4e32_v_f32m1x4
#define VSSEG4_FLOAT __riscv_vsseg4e32_v_f32m1x4
#define VLSSEG8_FLOAT __riscv_vlsseg8e32_v_f32m1x8
#define VSSEG8_FLOAT __riscv_vsseg8e32_v_f32m1x8
#define FLOAT_V_T vfloat32m2_t
#define FLOAT_V_T_HALF vfloat32m1_t
#define VLEV_FLOAT __riscv_vle32_v_f32m2
#define VLEV_FLOAT_HALF __riscv_vle32_v_f32m1
#define VSEV_FLOAT __riscv_vse32_v_f32m2
#define VSEV_FLOAT_HALF __riscv_vse32_v_f32m1
#else
#define VSETVL(n) __riscv_vsetvl_e64m1(n)
#define FLOAT_V_T vfloat64m1_t
#define FLOAT_VX2_T vfloat64m1x2_t
#define FLOAT_VX4_T vfloat64m1x4_t
#define FLOAT_VX8_T vfloat64m1x8_t
#define VLEV_FLOAT __riscv_vle64_v_f64m1
#define VLSEV_FLOAT __riscv_vlse64_v_f64m1
#define VSEV_FLOAT __riscv_vse64_v_f64m1
#define VLSSEG2_FLOAT __riscv_vlsseg2e64_v_f64m1x2
#define VSSEG2_FLOAT __riscv_vsseg2e64_v_f64m1x2
#define VLSSEG4_FLOAT __riscv_vlsseg4e64_v_f64m1x4
#define VSSEG4_FLOAT __riscv_vsseg4e64_v_f64m1x4
#define VLSSEG8_FLOAT __riscv_vlsseg8e64_v_f64m1x8
#define VSSEG8_FLOAT __riscv_vsseg8e64_v_f64m1x8
#define FLOAT_V_T vfloat64m4_t
#define FLOAT_V_T_HALF vfloat64m2_t
#define VLEV_FLOAT __riscv_vle64_v_f64m4
#define VLEV_FLOAT_HALF __riscv_vle64_v_f64m2
#define VSEV_FLOAT __riscv_vse64_v_f64m4
#define VSEV_FLOAT_HALF __riscv_vse64_v_f64m2
#endif

int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
Expand All @@ -69,9 +53,7 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
IFLOAT *boffset, *boffset1, *boffset2, *boffset3, *boffset4;

FLOAT_V_T v0;
FLOAT_VX2_T vx2;
FLOAT_VX4_T vx4;
FLOAT_VX8_T vx8;
FLOAT_V_T_HALF v1;

// fprintf(stderr, "gemm_tcopy_8 m=%ld n=%ld lda=%ld\n", m, n, lda);

Expand All @@ -81,156 +63,12 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
boffset3 = b + m * (n & ~3);
boffset4 = b + m * (n & ~1);

for(j = (m >> 3); j > 0; j--) {

aoffset1 = aoffset;
aoffset += 8 * lda;

boffset1 = boffset;
boffset += 64;

for(i = (n >> 3); i > 0; i--) {
size_t vl = 8;

vx8 = VLSSEG8_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSSEG8_FLOAT(boffset1, vx8, vl);

aoffset1 += 8;
boffset1 += m * 8;
}

if (n & 4) {
size_t vl = 8;

vx4 = VLSSEG4_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSSEG4_FLOAT(boffset2, vx4, vl);

aoffset1 += 4;
boffset2 += 32;
}

if (n & 2) {
size_t vl = 8;

vx2 = VLSSEG2_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSSEG2_FLOAT(boffset3, vx2, vl);

aoffset1 += 2;
boffset3 += 16;
}

if (n & 1) {
size_t vl = 8;

v0 = VLSEV_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSEV_FLOAT(boffset4, v0, vl);

aoffset1 += 1;
boffset4 += 8;
}

}

if (m & 4) {

aoffset1 = aoffset;
aoffset += 4 * lda;

boffset1 = boffset;
boffset += 32;

for(i = (n >> 3); i > 0; i--) {
size_t vl = 4;

vx8 = VLSSEG8_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSSEG8_FLOAT(boffset1, vx8, vl);

aoffset1 += 8;
boffset1 += m * 8;
}

if (n & 4) {
size_t vl = 4;

vx4 = VLSSEG4_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSSEG4_FLOAT(boffset2, vx4, vl);

aoffset1 += 4;
boffset2 += 16;
}

if (n & 2) {
size_t vl = 4;

vx2 = VLSSEG2_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSSEG2_FLOAT(boffset3, vx2, vl);

aoffset1 += 2;
boffset3 += 8;
}

if (n & 1) {
size_t vl = 4;

v0 = VLSEV_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSEV_FLOAT(boffset4, v0, vl);

aoffset1 += 1;
boffset4 += 4;
}
}

if (m & 2) {
for(j = m; j > 0; j--) {
aoffset1 = aoffset;
aoffset += 2 * lda;

boffset1 = boffset;
boffset += 16;

for(i = (n >> 3); i > 0; i--) {
size_t vl = 2;

vx8 = VLSSEG8_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSSEG8_FLOAT(boffset1, vx8, vl);

aoffset1 += 8;
boffset1 += m * 8;
}

if (n & 4) {
size_t vl = 2;

vx4 = VLSSEG4_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSSEG4_FLOAT(boffset2, vx4, vl);

aoffset1 += 4;
boffset2 += 8;
}

if (n & 2) {
size_t vl = 2;

vx2 = VLSSEG2_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSSEG2_FLOAT(boffset3, vx2, vl);

aoffset1 += 2;
boffset3 += 4;
}

if (n & 1) {
size_t vl = 2;

v0 = VLSEV_FLOAT(aoffset1, lda * sizeof(FLOAT), vl);
VSEV_FLOAT(boffset4, v0, vl);

aoffset1 += 1;
boffset4 += 2;
}
}

if (m & 1) {
aoffset1 = aoffset;
boffset1 = boffset;
aoffset += lda;
boffset += 8;

for(i = (n >> 3); i > 0; i--) {
size_t vl = 8;
Expand All @@ -245,27 +83,25 @@ int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b)
if (n & 4) {
size_t vl = 4;

v0 = VLEV_FLOAT(aoffset1, vl);
VSEV_FLOAT(boffset2, v0, vl);
v1 = VLEV_FLOAT_HALF(aoffset1, vl);
VSEV_FLOAT_HALF(boffset2, v1, vl);

aoffset1 += 4;
//boffset2 += 4;
boffset2 += 4;
}

if (n & 2) {
size_t vl = 2;

v0 = VLEV_FLOAT(aoffset1, vl);
VSEV_FLOAT(boffset3, v0, vl);
*(boffset3) = *(aoffset1);
*(boffset3 + 1) = *(aoffset1 + 1);

aoffset1 += 2;
// boffset3 += 2;
boffset3 += 2;
}

if (n & 1) {
*(boffset4) = *(aoffset1);
// aoffset1 ++;
// boffset4 ++;
*(boffset4) = *(aoffset1);
aoffset1 ++;
boffset4 ++;
}
}

Expand Down

0 comments on commit a7483d1

Please sign in to comment.