Skip to content

Add and use vectorized packing in ZVL128B and ZVL256B for RISCV #5422

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: develop
Choose a base branch
from
Open

Add and use vectorized packing in ZVL128B and ZVL256B for RISCV #5422

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c00afc8
Add and use vectorized packing to ZVL128B and ZVL256B. Up to 3x+ fas…
ChipKerchner Aug 12, 2025
64401b4
Disable vectorized packing for DGEMM - since it is slower than scalar.
ChipKerchner Aug 13, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
12 changes: 11 additions & 1 deletion kernel/riscv64/KERNEL.RISCV64_ZVL128B
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,14 +96,24 @@ CGEMVTKERNEL = zgemv_t_rvv.c
ZGEMVTKERNEL = zgemv_t_rvv.c

SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N)_zvl128b.c
ifneq ($(filter $(SGEMM_UNROLL_N),4 8 16),)
SGEMMONCOPY = gemm_ncopy_$(SGEMM_UNROLL_N)_rvv.c
SGEMMOTCOPY = gemm_tcopy_$(SGEMM_UNROLL_N)_rvv.c
else
SGEMMONCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
SGEMMOTCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
endif
SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)

ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
ifneq ($(filter $(SGEMM_UNROLL_M),4 8 16),)
SGEMMINCOPY = gemm_ncopy_$(SGEMM_UNROLL_M)_rvv.c
SGEMMITCOPY = gemm_tcopy_$(SGEMM_UNROLL_M)_rvv.c
else
SGEMMINCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
SGEMMITCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
endif
SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
endif
Expand Down Expand Up @@ -255,4 +265,4 @@ SHGEMMOTCOPYOBJ = shgemm_otcopy$(TSUFFIX).$(SUFFIX)
ifndef SHGEMM_BETA
SHGEMM_BETA = gemm_beta_rvv.c
endif
endif
endif
10 changes: 10 additions & 0 deletions kernel/riscv64/KERNEL.RISCV64_ZVL256B
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,13 +96,23 @@ CTRMMKERNEL = ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N)_zvl256b.c
ZTRMMKERNEL = ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N)_zvl256b.c

SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N)_zvl256b.c
ifneq ($(filter $(SGEMM_UNROLL_N),4 8 16),)
SGEMMONCOPY = gemm_ncopy_$(SGEMM_UNROLL_N)_rvv.c
SGEMMOTCOPY = gemm_tcopy_$(SGEMM_UNROLL_N)_rvv.c
else
SGEMMONCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
SGEMMOTCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
endif
SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
ifneq ($(filter $(SGEMM_UNROLL_M),4 8 16),)
SGEMMINCOPY = gemm_ncopy_$(SGEMM_UNROLL_M)_rvv.c
SGEMMITCOPY = gemm_tcopy_$(SGEMM_UNROLL_M)_rvv.c
else
SGEMMINCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
SGEMMITCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
endif
SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
endif
Expand Down
325 changes: 325 additions & 0 deletions kernel/riscv64/gemm_ncopy_16_rvv.c
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,325 @@
/***************************************************************************
Copyright (c) 2025, The OpenBLAS Project
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the OpenBLAS project nor the names of
its contributors may be used to endorse or promote products
derived from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
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 VLSEG_FLOAT __riscv_vlse32_v_f32m1
#define VLSSEG2_FLOAT __riscv_vlsseg2e32_v_f32m1x2
#define VLSSEG4_FLOAT __riscv_vlsseg4e32_v_f32m1x4
#define VLSSEG8_FLOAT __riscv_vlsseg8e32_v_f32m1x8
#define VGET_VX2 __riscv_vget_v_f32m1x2_f32m1
#define VGET_VX4 __riscv_vget_v_f32m1x4_f32m1
#define VGET_VX8 __riscv_vget_v_f32m1x8_f32m1
#define VSET_VX2 __riscv_vset_v_f32m1_f32m1x2
#define VSET_VX4 __riscv_vset_v_f32m1_f32m1x4
#define VSET_VX8 __riscv_vset_v_f32m1_f32m1x8
#define VLEV_FLOAT __riscv_vle32_v_f32m1
#define VSEV_FLOAT __riscv_vse32_v_f32m1
#define VSSEG2_FLOAT __riscv_vsseg2e32_v_f32m1x2
#define VSSEG4_FLOAT __riscv_vsseg4e32_v_f32m1x4
#define VSSEG8_FLOAT __riscv_vsseg8e32_v_f32m1x8
#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 VLSEG_FLOAT __riscv_vlse64_v_f64m1
#define VLSSEG2_FLOAT __riscv_vlsseg2e64_v_f64m1x2
#define VLSSEG4_FLOAT __riscv_vlsseg4e64_v_f64m1x4
#define VLSSEG8_FLOAT __riscv_vlsseg8e64_v_f64m1x8
#define VGET_VX2 __riscv_vget_v_f64m1x2_f64m1
#define VGET_VX4 __riscv_vget_v_f64m1x4_f64m1
#define VGET_VX8 __riscv_vget_v_f64m1x8_f64m1
#define VSET_VX2 __riscv_vset_v_f64m1_f64m1x2
#define VSET_VX4 __riscv_vset_v_f64m1_f64m1x4
#define VSET_VX8 __riscv_vset_v_f64m1_f64m1x8
#define VLEV_FLOAT __riscv_vle64_v_f64m1
#define VSEV_FLOAT __riscv_vse64_v_f64m1
#define VSSEG2_FLOAT __riscv_vsseg2e64_v_f64m1x2
#define VSSEG4_FLOAT __riscv_vsseg4e64_v_f64m1x4
#define VSSEG8_FLOAT __riscv_vsseg8e64_v_f64m1x8
#endif

// Optimizes the implementation in ../generic/gemm_ncopy_16.c

int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b)
{
BLASLONG i, j;

FLOAT *a_offset;
FLOAT *a_offset1, *a_offset2, *a_offset3, *a_offset4;
FLOAT *a_offset5, *a_offset6, *a_offset7, *a_offset8;
FLOAT *b_offset;

FLOAT_V_T v1, v2, v3, v4, v5, v6, v7, v8;
FLOAT_V_T v9, v10, v11, v12, v13, v14, v15, v16;
FLOAT_VX2_T vx2, vx21;
FLOAT_VX4_T vx4, vx41;
FLOAT_VX8_T vx8, vx81;

size_t vl;

//fprintf(stderr, "gemm_ncopy_16 m=%ld n=%ld lda=%ld\n", m, n, lda);

a_offset = a;
b_offset = b;

j = (n >> 4);
if (j) {
vl = VSETVL(8);

do {
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda * 8;
a_offset += 16 * lda;

i = m >> 3;
if (i) {
do {
vx8 = VLSSEG8_FLOAT(a_offset1, lda * sizeof(FLOAT), vl);
vx81 = VLSSEG8_FLOAT(a_offset2, lda * sizeof(FLOAT), vl);

v1 = VGET_VX8(vx8, 0);
v2 = VGET_VX8(vx8, 1);
v3 = VGET_VX8(vx8, 2);
v4 = VGET_VX8(vx8, 3);
v5 = VGET_VX8(vx8, 4);
v6 = VGET_VX8(vx8, 5);
v7 = VGET_VX8(vx8, 6);
v8 = VGET_VX8(vx8, 7);
v9 = VGET_VX8(vx81, 0);
v10 = VGET_VX8(vx81, 1);
v11 = VGET_VX8(vx81, 2);
v12 = VGET_VX8(vx81, 3);
v13 = VGET_VX8(vx81, 4);
v14 = VGET_VX8(vx81, 5);
v15 = VGET_VX8(vx81, 6);
v16 = VGET_VX8(vx81, 7);

VSEV_FLOAT(b_offset, v1, vl);
VSEV_FLOAT(b_offset + 8, v9, vl);
VSEV_FLOAT(b_offset + 16, v2, vl);
VSEV_FLOAT(b_offset + 24, v10, vl);
VSEV_FLOAT(b_offset + 32, v3, vl);
VSEV_FLOAT(b_offset + 40, v11, vl);
VSEV_FLOAT(b_offset + 48, v4, vl);
VSEV_FLOAT(b_offset + 56, v12, vl);
VSEV_FLOAT(b_offset + 64, v5, vl);
VSEV_FLOAT(b_offset + 72, v13, vl);
VSEV_FLOAT(b_offset + 80, v6, vl);
VSEV_FLOAT(b_offset + 88, v14, vl);
VSEV_FLOAT(b_offset + 96, v7, vl);
VSEV_FLOAT(b_offset + 104, v15, vl);
VSEV_FLOAT(b_offset + 112, v8, vl);
VSEV_FLOAT(b_offset + 120, v16, vl);

a_offset1 += 8;
a_offset2 += 8;
b_offset += 128;
} while (--i);
}

if (m & 4) {
vx4 = VLSSEG4_FLOAT(a_offset1, lda * sizeof(FLOAT), vl);
vx41 = VLSSEG4_FLOAT(a_offset2, lda * sizeof(FLOAT), vl);

v1 = VGET_VX4(vx4, 0);
v2 = VGET_VX4(vx4, 1);
v3 = VGET_VX4(vx4, 2);
v4 = VGET_VX4(vx4, 3);
v5 = VGET_VX4(vx41, 0);
v6 = VGET_VX4(vx41, 1);
v7 = VGET_VX4(vx41, 2);
v8 = VGET_VX4(vx41, 3);

VSEV_FLOAT(b_offset, v1, vl);
VSEV_FLOAT(b_offset + 8, v5, vl);
VSEV_FLOAT(b_offset + 16, v2, vl);
VSEV_FLOAT(b_offset + 24, v6, vl);
VSEV_FLOAT(b_offset + 32, v3, vl);
VSEV_FLOAT(b_offset + 40, v7, vl);
VSEV_FLOAT(b_offset + 48, v4, vl);
VSEV_FLOAT(b_offset + 56, v8, vl);

a_offset1 += 4;
a_offset2 += 4;
b_offset += 64;
}

if (m & 2) {
vx2 = VLSSEG2_FLOAT(a_offset1, lda * sizeof(FLOAT), vl);
vx21 = VLSSEG2_FLOAT(a_offset2, lda * sizeof(FLOAT), vl);

v1 = VGET_VX2(vx2, 0);
v2 = VGET_VX2(vx2, 1);
v3 = VGET_VX2(vx21, 0);
v4 = VGET_VX2(vx21, 1);

VSEV_FLOAT(b_offset, v1, vl);
VSEV_FLOAT(b_offset + 8, v3, vl);
VSEV_FLOAT(b_offset + 16, v2, vl);
VSEV_FLOAT(b_offset + 24, v4, vl);

a_offset1 += 2;
a_offset2 += 2;
b_offset += 32;
}

if (m & 1) {
v1 = VLSEG_FLOAT(a_offset1, lda * sizeof(FLOAT), vl);
v2 = VLSEG_FLOAT(a_offset2, lda * sizeof(FLOAT), vl);

VSEV_FLOAT(b_offset, v1, vl);
VSEV_FLOAT(b_offset + 8, v2, vl);

b_offset += 16;
}
} while (--j);
}

if (n & 8) {
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda;
a_offset3 = a_offset2 + lda;
a_offset4 = a_offset3 + lda;
a_offset5 = a_offset4 + lda;
a_offset6 = a_offset5 + lda;
a_offset7 = a_offset6 + lda;
a_offset8 = a_offset7 + lda;
a_offset += 8 * lda;

for(i = m; i > 0; i -= vl) {
vl = VSETVL(i);

v1 = VLEV_FLOAT(a_offset1, vl);
v2 = VLEV_FLOAT(a_offset2, vl);
v3 = VLEV_FLOAT(a_offset3, vl);
v4 = VLEV_FLOAT(a_offset4, vl);
v5 = VLEV_FLOAT(a_offset5, vl);
v6 = VLEV_FLOAT(a_offset6, vl);
v7 = VLEV_FLOAT(a_offset7, vl);
v8 = VLEV_FLOAT(a_offset8, vl);

vx8 = VSET_VX8(vx8, 0, v1);
vx8 = VSET_VX8(vx8, 1, v2);
vx8 = VSET_VX8(vx8, 2, v3);
vx8 = VSET_VX8(vx8, 3, v4);
vx8 = VSET_VX8(vx8, 4, v5);
vx8 = VSET_VX8(vx8, 5, v6);
vx8 = VSET_VX8(vx8, 6, v7);
vx8 = VSET_VX8(vx8, 7, v8);

VSSEG8_FLOAT(b_offset, vx8, vl);

a_offset1 += vl;
a_offset2 += vl;
a_offset3 += vl;
a_offset4 += vl;
a_offset5 += vl;
a_offset6 += vl;
a_offset7 += vl;
a_offset8 += vl;
b_offset += vl*8;
}
}

if (n & 4) {
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda;
a_offset3 = a_offset2 + lda;
a_offset4 = a_offset3 + lda;
a_offset += 4 * lda;

for(i = m; i > 0; i -= vl) {
vl = VSETVL(i);

v1 = VLEV_FLOAT(a_offset1, vl);
v2 = VLEV_FLOAT(a_offset2, vl);
v3 = VLEV_FLOAT(a_offset3, vl);
v4 = VLEV_FLOAT(a_offset4, vl);

vx4 = VSET_VX4(vx4, 0, v1);
vx4 = VSET_VX4(vx4, 1, v2);
vx4 = VSET_VX4(vx4, 2, v3);
vx4 = VSET_VX4(vx4, 3, v4);

VSSEG4_FLOAT(b_offset, vx4, vl);

a_offset1 += vl;
a_offset2 += vl;
a_offset3 += vl;
a_offset4 += vl;
b_offset += vl*4;
}
}

if (n & 2) {
a_offset1 = a_offset;
a_offset2 = a_offset1 + lda;
a_offset += 2 * lda;

for(i = m; i > 0; i -= vl) {
vl = VSETVL(i);

v1 = VLEV_FLOAT(a_offset1, vl);
v2 = VLEV_FLOAT(a_offset2, vl);

vx2 = VSET_VX2(vx2, 0, v1);
vx2 = VSET_VX2(vx2, 1, v2);

VSSEG2_FLOAT(b_offset, vx2, vl);

a_offset1 += vl;
a_offset2 += vl;
b_offset += vl*2;
}
}

if (n & 1) {
a_offset1 = a_offset;

for(i = m; i > 0; i -= vl) {
vl = VSETVL(i);

v1 = VLEV_FLOAT(a_offset1, vl);

VSEV_FLOAT(b_offset, v1, vl);

a_offset1 += vl;
b_offset += vl;
}
}

return 0;
}
Loading
Loading