From f1e628b88974ba984adf0ea08c817aef60548720 Mon Sep 17 00:00:00 2001 From: "Iha, Taisei" Date: Fri, 11 Apr 2025 20:00:33 +0900 Subject: [PATCH] Further performance improvements to [SD]GEMV. --- kernel/arm64/KERNEL.A64FX | 4 +- kernel/arm64/KERNEL.NEOVERSEV1 | 2 + kernel/arm64/gemv_n_sve_v1x3.c | 138 ++++++++++++++++++++++ kernel/arm64/gemv_n_sve_v4x3.c | 207 +++++++++++++++++++++++++++++++++ 4 files changed, 349 insertions(+), 2 deletions(-) create mode 100644 kernel/arm64/gemv_n_sve_v1x3.c create mode 100644 kernel/arm64/gemv_n_sve_v4x3.c diff --git a/kernel/arm64/KERNEL.A64FX b/kernel/arm64/KERNEL.A64FX index 75f0f39a7..3d68271da 100644 --- a/kernel/arm64/KERNEL.A64FX +++ b/kernel/arm64/KERNEL.A64FX @@ -1,6 +1,6 @@ include $(KERNELDIR)/KERNEL.ARMV8SVE -SGEMVNKERNEL = gemv_n_sve.c -DGEMVNKERNEL = gemv_n_sve.c +SGEMVNKERNEL = gemv_n_sve_v4x3.c +DGEMVNKERNEL = gemv_n_sve_v4x3.c SGEMVTKERNEL = gemv_t_sve_v4x3.c DGEMVTKERNEL = gemv_t_sve_v4x3.c diff --git a/kernel/arm64/KERNEL.NEOVERSEV1 b/kernel/arm64/KERNEL.NEOVERSEV1 index bacedf8cf..3e622bcbf 100644 --- a/kernel/arm64/KERNEL.NEOVERSEV1 +++ b/kernel/arm64/KERNEL.NEOVERSEV1 @@ -1,5 +1,7 @@ include $(KERNELDIR)/KERNEL.ARMV8SVE +SGEMVNKERNEL = gemv_n_sve_v1x3.c +DGEMVNKERNEL = gemv_n_sve_v1x3.c SGEMVTKERNEL = gemv_t_sve_v1x3.c DGEMVTKERNEL = gemv_t_sve_v1x3.c ifeq ($(BUILD_BFLOAT16), 1) diff --git a/kernel/arm64/gemv_n_sve_v1x3.c b/kernel/arm64/gemv_n_sve_v1x3.c new file mode 100644 index 000000000..44c9a89b9 --- /dev/null +++ b/kernel/arm64/gemv_n_sve_v1x3.c @@ -0,0 +1,138 @@ +/*************************************************************************** +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 COPYRIGHT OWNER 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 + +#include "common.h" + +#ifdef DOUBLE +#define SV_COUNT svcntd +#define SV_TYPE svfloat64_t +#define SV_TRUE svptrue_b64 +#define SV_WHILE svwhilelt_b64_s64 +#define SV_DUP svdup_f64 +#else +#define SV_COUNT svcntw +#define SV_TYPE svfloat32_t +#define SV_TRUE svptrue_b32 +#define SV_WHILE svwhilelt_b32_s64 +#define SV_DUP svdup_f32 +#endif + +int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, + BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, + FLOAT *buffer) +{ + BLASLONG i; + BLASLONG ix,iy; + BLASLONG j; + FLOAT *a_ptr; + FLOAT temp; + + ix = 0; + a_ptr = a; + + if (inc_y == 1) { + BLASLONG width = (n + 3 - 1) / 3; + + FLOAT *a0_ptr = a_ptr + lda * width * 0; + FLOAT *a1_ptr = a_ptr + lda * width * 1; + FLOAT *a2_ptr = a_ptr + lda * width * 2; + + FLOAT *x0_ptr = x + inc_x * width * 0; + FLOAT *x1_ptr = x + inc_x * width * 1; + FLOAT *x2_ptr = x + inc_x * width * 2; + + for (j = 0; j < width; j++) { + svbool_t pg00 = ((j + width * 0) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg01 = ((j + width * 1) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg02 = ((j + width * 2) < n) ? SV_TRUE() : svpfalse(); + + SV_TYPE temp0_vec = SV_DUP(alpha * x0_ptr[ix]); + SV_TYPE temp1_vec = SV_DUP(alpha * x1_ptr[ix]); + SV_TYPE temp2_vec = SV_DUP(alpha * x2_ptr[ix]); + i = 0; + BLASLONG sve_size = SV_COUNT(); + while ((i + sve_size * 1 - 1) < m) { + SV_TYPE y0_vec = svld1_vnum(SV_TRUE(), y + i, 0); + + SV_TYPE a00_vec = svld1_vnum(pg00, a0_ptr + i, 0); + SV_TYPE a01_vec = svld1_vnum(pg01, a1_ptr + i, 0); + SV_TYPE a02_vec = svld1_vnum(pg02, a2_ptr + i, 0); + + y0_vec = svmla_m(pg00, y0_vec, temp0_vec, a00_vec); + y0_vec = svmla_m(pg01, y0_vec, temp1_vec, a01_vec); + y0_vec = svmla_m(pg02, y0_vec, temp2_vec, a02_vec); + + svst1_vnum(SV_TRUE(), y + i, 0, y0_vec); + i += sve_size * 1; + } + + if (i < m) { + svbool_t pg0 = SV_WHILE(i + sve_size * 0, m); + + pg00 = svand_z(SV_TRUE(), pg0, pg00); + pg01 = svand_z(SV_TRUE(), pg0, pg01); + pg02 = svand_z(SV_TRUE(), pg0, pg02); + + SV_TYPE y0_vec = svld1_vnum(pg0, y + i, 0); + + SV_TYPE a00_vec = svld1_vnum(pg00, a0_ptr + i, 0); + SV_TYPE a01_vec = svld1_vnum(pg01, a1_ptr + i, 0); + SV_TYPE a02_vec = svld1_vnum(pg02, a2_ptr + i, 0); + + y0_vec = svmla_m(pg00, y0_vec, temp0_vec, a00_vec); + y0_vec = svmla_m(pg01, y0_vec, temp1_vec, a01_vec); + y0_vec = svmla_m(pg02, y0_vec, temp2_vec, a02_vec); + + svst1_vnum(pg0, y + i, 0, y0_vec); + } + a0_ptr += lda; + a1_ptr += lda; + a2_ptr += lda; + ix += inc_x; + } + return(0); + } + + for (j = 0; j < n; j++) { + temp = alpha * x[ix]; + iy = 0; + for (i = 0; i < m; i++) { + y[iy] += temp * a_ptr[i]; + iy += inc_y; + } + a_ptr += lda; + ix += inc_x; + } + return (0); +} diff --git a/kernel/arm64/gemv_n_sve_v4x3.c b/kernel/arm64/gemv_n_sve_v4x3.c new file mode 100644 index 000000000..92e4f75b6 --- /dev/null +++ b/kernel/arm64/gemv_n_sve_v4x3.c @@ -0,0 +1,207 @@ +/*************************************************************************** +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 COPYRIGHT OWNER 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 + +#include "common.h" + +#ifdef DOUBLE +#define SV_COUNT svcntd +#define SV_TYPE svfloat64_t +#define SV_TRUE svptrue_b64 +#define SV_WHILE svwhilelt_b64_s64 +#define SV_DUP svdup_f64 +#else +#define SV_COUNT svcntw +#define SV_TYPE svfloat32_t +#define SV_TRUE svptrue_b32 +#define SV_WHILE svwhilelt_b32_s64 +#define SV_DUP svdup_f32 +#endif + +int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, + BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, + FLOAT *buffer) +{ + BLASLONG i; + BLASLONG ix,iy; + BLASLONG j; + FLOAT *a_ptr; + FLOAT temp; + + ix = 0; + a_ptr = a; + + if (inc_y == 1) { + BLASLONG width = (n + 3 - 1) / 3; + + FLOAT *a0_ptr = a_ptr + lda * width * 0; + FLOAT *a1_ptr = a_ptr + lda * width * 1; + FLOAT *a2_ptr = a_ptr + lda * width * 2; + + FLOAT *x0_ptr = x + inc_x * width * 0; + FLOAT *x1_ptr = x + inc_x * width * 1; + FLOAT *x2_ptr = x + inc_x * width * 2; + + for (j = 0; j < width; j++) { + svbool_t pg00 = ((j + width * 0) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg10 = ((j + width * 0) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg20 = ((j + width * 0) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg30 = ((j + width * 0) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg01 = ((j + width * 1) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg11 = ((j + width * 1) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg21 = ((j + width * 1) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg31 = ((j + width * 1) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg02 = ((j + width * 2) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg12 = ((j + width * 2) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg22 = ((j + width * 2) < n) ? SV_TRUE() : svpfalse(); + svbool_t pg32 = ((j + width * 2) < n) ? SV_TRUE() : svpfalse(); + + SV_TYPE temp0_vec = SV_DUP(alpha * x0_ptr[ix]); + SV_TYPE temp1_vec = SV_DUP(alpha * x1_ptr[ix]); + SV_TYPE temp2_vec = SV_DUP(alpha * x2_ptr[ix]); + i = 0; + BLASLONG sve_size = SV_COUNT(); + while ((i + sve_size * 4 - 1) < m) { + SV_TYPE y0_vec = svld1_vnum(SV_TRUE(), y + i, 0); + SV_TYPE y1_vec = svld1_vnum(SV_TRUE(), y + i, 1); + SV_TYPE y2_vec = svld1_vnum(SV_TRUE(), y + i, 2); + SV_TYPE y3_vec = svld1_vnum(SV_TRUE(), y + i, 3); + + SV_TYPE a00_vec = svld1_vnum(pg00, a0_ptr + i, 0); + SV_TYPE a10_vec = svld1_vnum(pg10, a0_ptr + i, 1); + SV_TYPE a20_vec = svld1_vnum(pg20, a0_ptr + i, 2); + SV_TYPE a30_vec = svld1_vnum(pg30, a0_ptr + i, 3); + SV_TYPE a01_vec = svld1_vnum(pg01, a1_ptr + i, 0); + SV_TYPE a11_vec = svld1_vnum(pg11, a1_ptr + i, 1); + SV_TYPE a21_vec = svld1_vnum(pg21, a1_ptr + i, 2); + SV_TYPE a31_vec = svld1_vnum(pg31, a1_ptr + i, 3); + SV_TYPE a02_vec = svld1_vnum(pg02, a2_ptr + i, 0); + SV_TYPE a12_vec = svld1_vnum(pg12, a2_ptr + i, 1); + SV_TYPE a22_vec = svld1_vnum(pg22, a2_ptr + i, 2); + SV_TYPE a32_vec = svld1_vnum(pg32, a2_ptr + i, 3); + + y0_vec = svmla_m(pg00, y0_vec, temp0_vec, a00_vec); + y1_vec = svmla_m(pg10, y1_vec, temp0_vec, a10_vec); + y2_vec = svmla_m(pg20, y2_vec, temp0_vec, a20_vec); + y3_vec = svmla_m(pg30, y3_vec, temp0_vec, a30_vec); + y0_vec = svmla_m(pg01, y0_vec, temp1_vec, a01_vec); + y1_vec = svmla_m(pg11, y1_vec, temp1_vec, a11_vec); + y2_vec = svmla_m(pg21, y2_vec, temp1_vec, a21_vec); + y3_vec = svmla_m(pg31, y3_vec, temp1_vec, a31_vec); + y0_vec = svmla_m(pg02, y0_vec, temp2_vec, a02_vec); + y1_vec = svmla_m(pg12, y1_vec, temp2_vec, a12_vec); + y2_vec = svmla_m(pg22, y2_vec, temp2_vec, a22_vec); + y3_vec = svmla_m(pg32, y3_vec, temp2_vec, a32_vec); + + svst1_vnum(SV_TRUE(), y + i, 0, y0_vec); + svst1_vnum(SV_TRUE(), y + i, 1, y1_vec); + svst1_vnum(SV_TRUE(), y + i, 2, y2_vec); + svst1_vnum(SV_TRUE(), y + i, 3, y3_vec); + i += sve_size * 4; + } + + if (i < m) { + svbool_t pg0 = SV_WHILE(i + sve_size * 0, m); + svbool_t pg1 = SV_WHILE(i + sve_size * 1, m); + svbool_t pg2 = SV_WHILE(i + sve_size * 2, m); + svbool_t pg3 = SV_WHILE(i + sve_size * 3, m); + + pg00 = svand_z(SV_TRUE(), pg0, pg00); + pg10 = svand_z(SV_TRUE(), pg1, pg10); + pg20 = svand_z(SV_TRUE(), pg2, pg20); + pg30 = svand_z(SV_TRUE(), pg3, pg30); + pg01 = svand_z(SV_TRUE(), pg0, pg01); + pg11 = svand_z(SV_TRUE(), pg1, pg11); + pg21 = svand_z(SV_TRUE(), pg2, pg21); + pg31 = svand_z(SV_TRUE(), pg3, pg31); + pg02 = svand_z(SV_TRUE(), pg0, pg02); + pg12 = svand_z(SV_TRUE(), pg1, pg12); + pg22 = svand_z(SV_TRUE(), pg2, pg22); + pg32 = svand_z(SV_TRUE(), pg3, pg32); + + SV_TYPE y0_vec = svld1_vnum(pg0, y + i, 0); + SV_TYPE y1_vec = svld1_vnum(pg1, y + i, 1); + SV_TYPE y2_vec = svld1_vnum(pg2, y + i, 2); + SV_TYPE y3_vec = svld1_vnum(pg3, y + i, 3); + + SV_TYPE a00_vec = svld1_vnum(pg00, a0_ptr + i, 0); + SV_TYPE a10_vec = svld1_vnum(pg10, a0_ptr + i, 1); + SV_TYPE a20_vec = svld1_vnum(pg20, a0_ptr + i, 2); + SV_TYPE a30_vec = svld1_vnum(pg30, a0_ptr + i, 3); + SV_TYPE a01_vec = svld1_vnum(pg01, a1_ptr + i, 0); + SV_TYPE a11_vec = svld1_vnum(pg11, a1_ptr + i, 1); + SV_TYPE a21_vec = svld1_vnum(pg21, a1_ptr + i, 2); + SV_TYPE a31_vec = svld1_vnum(pg31, a1_ptr + i, 3); + SV_TYPE a02_vec = svld1_vnum(pg02, a2_ptr + i, 0); + SV_TYPE a12_vec = svld1_vnum(pg12, a2_ptr + i, 1); + SV_TYPE a22_vec = svld1_vnum(pg22, a2_ptr + i, 2); + SV_TYPE a32_vec = svld1_vnum(pg32, a2_ptr + i, 3); + + y0_vec = svmla_m(pg00, y0_vec, temp0_vec, a00_vec); + y1_vec = svmla_m(pg10, y1_vec, temp0_vec, a10_vec); + y2_vec = svmla_m(pg20, y2_vec, temp0_vec, a20_vec); + y3_vec = svmla_m(pg30, y3_vec, temp0_vec, a30_vec); + y0_vec = svmla_m(pg01, y0_vec, temp1_vec, a01_vec); + y1_vec = svmla_m(pg11, y1_vec, temp1_vec, a11_vec); + y2_vec = svmla_m(pg21, y2_vec, temp1_vec, a21_vec); + y3_vec = svmla_m(pg31, y3_vec, temp1_vec, a31_vec); + y0_vec = svmla_m(pg02, y0_vec, temp2_vec, a02_vec); + y1_vec = svmla_m(pg12, y1_vec, temp2_vec, a12_vec); + y2_vec = svmla_m(pg22, y2_vec, temp2_vec, a22_vec); + y3_vec = svmla_m(pg32, y3_vec, temp2_vec, a32_vec); + + svst1_vnum(pg0, y + i, 0, y0_vec); + svst1_vnum(pg1, y + i, 1, y1_vec); + svst1_vnum(pg2, y + i, 2, y2_vec); + svst1_vnum(pg3, y + i, 3, y3_vec); + } + a0_ptr += lda; + a1_ptr += lda; + a2_ptr += lda; + ix += inc_x; + } + return(0); + } + + for (j = 0; j < n; j++) { + temp = alpha * x[ix]; + iy = 0; + for (i = 0; i < m; i++) { + y[iy] += temp * a_ptr[i]; + iy += inc_y; + } + a_ptr += lda; + ix += inc_x; + } + return (0); +}