From 15edb441bf827c03ab39c6dc693f6c01c1da971e Mon Sep 17 00:00:00 2001 From: gxw Date: Mon, 14 Oct 2024 17:36:56 +0800 Subject: [PATCH 1/5] LoongArch64: Opt somatcopy_rt with LASX --- kernel/loongarch64/KERNEL.LA464 | 2 + kernel/loongarch64/somatcopy_rt_lasx.c | 200 +++++++++++++++++++++++++ 2 files changed, 202 insertions(+) create mode 100644 kernel/loongarch64/somatcopy_rt_lasx.c diff --git a/kernel/loongarch64/KERNEL.LA464 b/kernel/loongarch64/KERNEL.LA464 index eff1581d9..2eec3cd55 100644 --- a/kernel/loongarch64/KERNEL.LA464 +++ b/kernel/loongarch64/KERNEL.LA464 @@ -172,4 +172,6 @@ DGEMM_SMALL_K_TN = dgemm_small_kernel_tn_lasx.S DGEMM_SMALL_K_B0_TN = dgemm_small_kernel_tn_lasx.S DGEMM_SMALL_K_TT = dgemm_small_kernel_tt_lasx.S DGEMM_SMALL_K_B0_TT = dgemm_small_kernel_tt_lasx.S + +SOMATCOPY_RT = somatcopy_rt_lasx.c endif diff --git a/kernel/loongarch64/somatcopy_rt_lasx.c b/kernel/loongarch64/somatcopy_rt_lasx.c new file mode 100644 index 000000000..51a141149 --- /dev/null +++ b/kernel/loongarch64/somatcopy_rt_lasx.c @@ -0,0 +1,200 @@ +/*************************************************************************** +Copyright (c) 2024, 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" + +#define SAVE1x4(c1) \ + "vfmul.s $vr"#c1",$vr"#c1",$vr15;vstelm.w $vr"#c1",%4,0,0;add.d %4,%4,%3;\n\t" \ + "vstelm.w $vr"#c1",%4,0,1;add.d %4,%4,%3;vstelm.w $vr"#c1",%4,0,2;add.d %4,%4,%3; \n\t" \ + "vstelm.w $vr"#c1",%4,0,3;add.d %4,%4,%3;\n\t" +#define SAVE_2x4(c1, c2, t1, t2) \ + "vilvl.w $vr"#t1",$vr"#c2",$vr"#c1";vfmul.s $vr"#t1",$vr"#t1",$vr15; \n\t" \ + "vstelm.d $vr"#t1",%4,0,0;add.d %4,%4,%3;vstelm.d $vr"#t1",%4,0,1;add.d %4,%4,%3 \n\t" \ + "vilvh.w $vr"#t2",$vr"#c2",$vr"#c1";vfmul.s $vr"#t2",$vr"#t2",$vr15; \n\t" \ + "vstelm.d $vr"#t2",%4,0,0;add.d %4,%4,%3;vstelm.d $vr"#t2",%4,0,1;add.d %4,%4,%3 \n\t" +#define SAVE_4x1(b1) \ + "vst $vr"#b1",%4,0; add.d %4,%4,%3 \n\t" +#define SAVE_4x2(b1, b2) \ + "vst $vr"#b1",%4,0; add.d %4,%4,%3; vst $vr"#b2",%4,0; add.d %4,%4,%3 \n\t" +#define SAVE_4x4(b1, b2, b3, b4) \ + "vst $vr"#b1",%4,0; add.d %4,%4,%3; vst $vr"#b2",%4,0; add.d %4,%4,%3 \n\t" \ + "vst $vr"#b3",%4,0; add.d %4,%4,%3; vst $vr"#b4",%4,0; add.d %4,%4,%3 \n\t" +#define SAVE_4x8(b1, b2, b3, b4) \ + SAVE_4x4(b1, b2, b3, b4) \ + "xvpermi.q $xr"#b1",$xr"#b1",1; xvpermi.q $xr"#b2",$xr"#b2",1; xvpermi.q $xr"#b3",$xr"#b3",1; xvpermi.q $xr"#b4",$xr"#b4",1; \n\t" \ + SAVE_4x4(b1, b2, b3, b4) + +#define TRANS_4x4(a1, a2, a3, a4, t1, t2, t3, t4) \ + "vilvl.w $vr"#t1",$vr"#a2",$vr"#a1";vilvh.w $vr"#t2",$vr"#a2",$vr"#a1"; \n\t" \ + "vilvl.w $vr"#t3",$vr"#a4",$vr"#a3";vilvh.w $vr"#t4",$vr"#a4",$vr"#a3"; \n\t" \ + "vilvl.d $vr"#a1",$vr"#t3",$vr"#t1";vilvh.d $vr"#a2",$vr"#t3",$vr"#t1"; \n\t" \ + "vilvl.d $vr"#a3",$vr"#t4",$vr"#t2";vilvh.d $vr"#a4",$vr"#t4",$vr"#t2"; \n\t" +#define TRANS_4x8(a1, a2, a3, a4, t1, t2, t3, t4) \ + "xvilvl.w $xr"#t1",$xr"#a2",$xr"#a1"; xvilvh.w $xr"#t2",$xr"#a2",$xr"#a1"; \n\t" \ + "xvilvl.w $xr"#t3",$xr"#a4",$xr"#a3"; xvilvh.w $xr"#t4",$xr"#a4",$xr"#a3"; \n\t" \ + "xvilvl.d $xr"#a1",$xr"#t3",$xr"#t1"; xvilvh.d $xr"#a2",$xr"#t3",$xr"#t1"; \n\t" \ + "xvilvl.d $xr"#a3",$xr"#t4",$xr"#t2"; xvilvh.d $xr"#a4",$xr"#t4",$xr"#t2"; \n\t" +#define COPY_4x16 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "xvld $xr0,%0,0; xvld $xr4,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; xvld $xr5,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr2,%0,0; xvld $xr6,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr3,%0,0; xvld $xr7,%0,32; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + "xvfmul.s $xr4,$xr4,$xr15;xvfmul.s $xr5,$xr5,$xr15;xvfmul.s $xr6,$xr6,$xr15;xvfmul.s $xr7,$xr7,$xr15 \n\t" \ + TRANS_4x8(0,1,2,3,8,9,10,11) SAVE_4x8(0,1,2,3) \ + TRANS_4x8(4,5,6,7,8,9,10,11) SAVE_4x8(4,5,6,7) +#define COPY_4x8 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr2,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr3,%0,0; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + TRANS_4x8(0,1,2,3,8,9,10,11) SAVE_4x8(0,1,2,3) +#define COPY_4x4 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \ + TRANS_4x4(0,1,2,3,8,9,10,11) SAVE_4x4(0,1,2,3) +#define COPY_4x2 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \ + TRANS_4x4(0,1,2,3,8,9,10,11) SAVE_4x2(0,1) +#define COPY_4x1 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "fld.s $f0,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f1,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f2,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f3,%0,0; add.d %0,%0,%2; \n\t" \ + "xvinsve0.w $xr0,$xr1,1;xvinsve0.w $xr0,$xr2,2;xvinsve0.w $xr0,$xr3,3; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15; \n\t" \ + SAVE_4x1(0) + +#define COPY_2x16 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "xvld $xr0,%0,0; xvld $xr2,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; xvld $xr3,%0,32; add.d %0,%0,%2; \n\t" \ + "xvpermi.q $xr4,$xr0,1;xvpermi.q $xr6,$xr2,1;xvpermi.q $xr5,$xr1,1;xvpermi.q $xr7,$xr3,1; \n\t" \ + SAVE_2x4(0,1,8,9) SAVE_2x4(4,5,8,9) SAVE_2x4(2,3,8,9) SAVE_2x4(6,7,8,9) +#define COPY_2x8 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \ + "xvpermi.q $xr2,$xr0,1;xvpermi.q $xr3,$xr1,1; \n\t" \ + SAVE_2x4(0,1,4,5) SAVE_2x4(2,3,4,5) +#define COPY_2x4 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + SAVE_2x4(0,1,4,5) +#define COPY_2x2 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "fld.d $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fld.d $f1,%0,0;add.d %0,%0,%2; \n\t" \ + "xvinsve0.d $xr0,$xr1,1;vfmul.s $vr0,$vr0,$vr15;vshuf4i.w $vr0,$vr0,0xd8 \n\t" \ + "vstelm.d $vr0,%4,0,0;add.d %4,%4,%3;vstelm.d $vr0,%4,0,1 \n\t" +#define COPY_2x1 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "fld.s $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fld.s $f1,%0,0;add.d %0,%0,%2; \n\t" \ + "xvinsve0.w $xr0,$xr1,1;vfmul.s $vr0,$vr0,$vr15; \n\t" \ + "vstelm.d $vr0,%4,0,0; \n\t" + +#define COPY_1x16 \ + "move %4,%1; addi.d %1,%1,4 \n\t" \ + "vld $vr1,%0,0;" SAVE1x4(1) "vld $vr2,%0,16;" SAVE1x4(2) \ + "vld $vr1,%0,32;" SAVE1x4(1) "vld $vr2,%0,48;" SAVE1x4(2) \ + "add.d %0,%0,%2 \n\t" +#define COPY_1x8 \ + "move %4,%1; addi.d %1,%1,4 \n\t" \ + "vld $vr1,%0,0;" SAVE1x4(1) "vld $vr2,%0,16;" SAVE1x4(2) \ + "add.d %0,%0,%2 \n\t" +#define COPY_1x4 \ + "move %4,%1; addi.d %1,%1,4 \n\t" \ + "vld $vr1,%0,0;" SAVE1x4(1) \ + "add.d %0,%0,%2 \n\t" +#define COPY_1x2 \ + "move %4,%1;fld.d $f1,%0,0;add.d %0,%0,%2;vfmul.s $vr1,$vr1,$vr15;vstelm.w $vr1,%4,0,0;add.d %4,%4,%3;vstelm.w $vr1,%4,0,1;\n\t" \ + "addi.d %1,%1,4;\n\t" +#define COPY_1x1 \ + "fld.s $f1,%0,0;fmul.s $f1,$f1,$f15;fst.s $f1,%1,0;add.d %0,%0,%2;addi.d %1,%1,4;\n\t" + +#define ROWS_OF_BLOCK 128 + +#define COMPUTE(ndim) \ + src = src_base; dst = dst_base; \ + __asm__ __volatile__( \ + "xvldrepl.w $xr15, %6, 0 \n\t" \ + "srli.d $r6, %5, 2 \n\t" \ + "beqz $r6, "#ndim"3f \n\t" \ + #ndim"4: \n\t" \ + COPY_4x##ndim \ + "addi.d $r6, $r6, -1 \n\t" \ + "bnez $r6, "#ndim"4b \n\t" \ + #ndim"3: \n\t" \ + "andi $r6, %5, 2 \n\t" \ + "beqz $r6, "#ndim"1f \n\t" \ + #ndim"2: \n\t" \ + COPY_2x##ndim \ + #ndim"1: \n\t" \ + "andi $r6, %5, 1 \n\t" \ + "beqz $r6, "#ndim"0f \n\t" \ + COPY_1x##ndim \ + #ndim"0: \n\t" \ + :"+r"(src),"+r"(dst),"+r"(src_ld_bytes),"+r"(dst_ld_bytes),"+r"(dst_tmp) \ + :"r"(num_rows),"r"(&ALPHA) \ + :"memory", "$r6", "$f0", "$f1", "$f2", "$f3", "$f4", "$f5", "$f6", "$f7", "$f8", "$f9", "$f10", "$f11", "$f15" \ + ); + +int CNAME(BLASLONG rows, BLASLONG cols, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG ldb){ + float *src, *dst, *dst_tmp=0, *src_base, *dst_base; + uint64_t src_ld_bytes = (uint64_t)lda * sizeof(float), dst_ld_bytes = (uint64_t)ldb * sizeof(float), num_rows = 0; + BLASLONG cols_left, rows_done; float ALPHA = alpha; + if (ALPHA == 0.0) { + dst_base = b; + for (cols_left = cols; cols_left > 0; cols_left--) {memset(dst_base, 0, rows * sizeof(float)); dst_base += ldb;} + return 0; + } + for (rows_done = 0; rows_done < rows; rows_done += num_rows) { + num_rows = rows - rows_done; + if (num_rows > ROWS_OF_BLOCK) num_rows = ROWS_OF_BLOCK; + cols_left = cols; src_base = a + (int64_t)lda * (int64_t)rows_done; dst_base = b + rows_done; + for (;cols_left > 15; cols_left -= 16) {COMPUTE(16) src_base += 16; dst_base += 16 * ldb;} + for (;cols_left > 7; cols_left -= 8) {COMPUTE(8) src_base += 8; dst_base += 8 * ldb;} + for (;cols_left > 3; cols_left -= 4) {COMPUTE(4) src_base += 4; dst_base += 4 * ldb;} + for (;cols_left > 1; cols_left -= 2) {COMPUTE(2) src_base += 2; dst_base += 2 * ldb;} + if (cols_left > 0) {COMPUTE(1) src_base ++; dst_base += ldb;} + } +} From acf6cab30478787922d03895ba27960ee293569f Mon Sep 17 00:00:00 2001 From: gxw Date: Thu, 17 Oct 2024 09:50:02 +0000 Subject: [PATCH 2/5] LoongArch64: Opt somatcopy_rn with LASX --- kernel/loongarch64/KERNEL.LA464 | 1 + kernel/loongarch64/somatcopy_rn_lasx.c | 193 +++++++++++++++++++++++++ 2 files changed, 194 insertions(+) create mode 100644 kernel/loongarch64/somatcopy_rn_lasx.c diff --git a/kernel/loongarch64/KERNEL.LA464 b/kernel/loongarch64/KERNEL.LA464 index 2eec3cd55..1a6b6ee6c 100644 --- a/kernel/loongarch64/KERNEL.LA464 +++ b/kernel/loongarch64/KERNEL.LA464 @@ -174,4 +174,5 @@ DGEMM_SMALL_K_TT = dgemm_small_kernel_tt_lasx.S DGEMM_SMALL_K_B0_TT = dgemm_small_kernel_tt_lasx.S SOMATCOPY_RT = somatcopy_rt_lasx.c +SOMATCOPY_RN = somatcopy_rn_lasx.c endif diff --git a/kernel/loongarch64/somatcopy_rn_lasx.c b/kernel/loongarch64/somatcopy_rn_lasx.c new file mode 100644 index 000000000..c9bfb4c5e --- /dev/null +++ b/kernel/loongarch64/somatcopy_rn_lasx.c @@ -0,0 +1,193 @@ +/*************************************************************************** +Copyright (c) 2024, 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" +#define SAVE_4x1(b1, b2, b3, b4) \ + "vstelm.w $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.w $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" \ + "vstelm.w $vr"#b3",%1,0,0;add.d %1,%1,%3;vstelm.w $vr"#b4",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_4x2(b1, b2, b3, b4) \ + "vstelm.d $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.d $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" \ + "vstelm.d $vr"#b3",%1,0,0;add.d %1,%1,%3;vstelm.d $vr"#b4",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_4x4(b1, b2, b3, b4) \ + "vst $vr"#b1",%1,0;add.d %1,%1,%3;vst $vr"#b2",%1,0;add.d %1,%1,%3;\n\t" \ + "vst $vr"#b3",%1,0;add.d %1,%1,%3;vst $vr"#b4",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_4x8(b1, b2, b3, b4) \ + "xvst $xr"#b1",%1,0;add.d %1,%1,%3;xvst $xr"#b2",%1,0;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b3",%1,0;add.d %1,%1,%3;xvst $xr"#b4",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_4x16(b1, b2, b3, b4, b5, b6, b7, b8) \ + "xvst $xr"#b1",%1,0;xvst $xr"#b2",%1,32;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b3",%1,0;xvst $xr"#b4",%1,32;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b5",%1,0;xvst $xr"#b6",%1,32;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b7",%1,0;xvst $xr"#b8",%1,32;add.d %1,%1,%3;\n\t" +#define SAVE_2x16(b1, b2, b3, b4) \ + "xvst $xr"#b1",%1,0;xvst $xr"#b2",%1,32;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b3",%1,0;xvst $xr"#b4",%1,32;add.d %1,%1,%3;\n\t" +#define SAVE_2x8(b1, b2) \ + "xvst $xr"#b1",%1,0;add.d %1,%1,%3;xvst $xr"#b2",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_2x4(b1, b2) \ + "vst $vr"#b1",%1,0;add.d %1,%1,%3;vst $vr"#b2",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_2x2(b1, b2) \ + "vstelm.d $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.d $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_2x1(b1, b2) \ + "vstelm.w $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.w $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_1x16(b1, b2) \ + "xvst $xr"#b1",%1,0;xvst $xr"#b2",%1,32;add.d %1,%1,%3;\n\t" +#define SAVE_1x8(b1) \ + "xvst $xr"#b1",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_1x4(b1) \ + "vst $vr"#b1",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_1x2(b1) \ + "vstelm.d $vr"#b1",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_1x1(b1) \ + "vstelm.w $vr"#b1",%1,0,0;add.d %1,%1,%3;\n\t" +#define COPY_4x16 \ + "xvld $xr0,%0,0; xvld $xr4,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; xvld $xr5,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr2,%0,0; xvld $xr6,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr3,%0,0; xvld $xr7,%0,32; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + "xvfmul.s $xr4,$xr4,$xr15;xvfmul.s $xr5,$xr5,$xr15;xvfmul.s $xr6,$xr6,$xr15;xvfmul.s $xr7,$xr7,$xr15 \n\t" \ + SAVE_4x16(0,4,1,5,2,6,3,7) +#define COPY_4x8 \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr2,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr3,%0,0; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + SAVE_4x8(0,1,2,3) +#define COPY_4x4 \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \ + SAVE_4x4(0,1,2,3) +#define COPY_4x2 \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \ + SAVE_4x2(0,1,2,3) +#define COPY_4x1 \ + "fld.s $f0,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f1,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f2,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f3,%0,0; add.d %0,%0,%2; \n\t" \ + "fmul.s $f0,$f0,$f15;fmul.s $f1,$f1,$f15;fmul.s $f2,$f2,$f15;fmul.s $f3,$f3,$f15 \n\t" \ + SAVE_4x1(0,1,2,3) +#define COPY_2x16 \ + "xvld $xr0,%0,0; xvld $xr2,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; xvld $xr3,%0,32; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + SAVE_2x16(0,2,1,3) +#define COPY_2x8 \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15; \n\t" \ + SAVE_2x8(0,1) +#define COPY_2x4 \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15; \n\t" \ + SAVE_2x4(0,1) +#define COPY_2x2 \ + "fld.d $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fld.d $f1,%0,0;add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15; \n\t" \ + SAVE_2x2(0,1) +#define COPY_2x1 \ + "fld.s $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fld.s $f1,%0,0;add.d %0,%0,%2; \n\t" \ + "fmul.s $f0,$f0,$f15;fmul.s $f1,$f1,$f15; \n\t" \ + SAVE_2x1(0,1) +#define COPY_1x16 \ + "xvld $xr0,%0,0; xvld $xr1,%0,32; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15; \n\t" \ + SAVE_1x16(0,1) +#define COPY_1x8 \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15; \n\t" \ + SAVE_1x8(0) +#define COPY_1x4 \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15; \n\t" \ + SAVE_1x4(0) +#define COPY_1x2 \ + "fld.d $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15; \n\t" \ + SAVE_1x2(0) +#define COPY_1x1 \ + "fld.s $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fmul.s $f0,$f0,$f15; \n\t" \ + SAVE_1x1(0) +#define ROWS_OF_BLOCK 128 +#define COMPUTE(ndim) \ + src = src_base; dst = dst_base; \ + __asm__ __volatile__( \ + "xvldrepl.w $xr15, %6, 0 \n\t" \ + "srli.d $r6, %5, 2 \n\t" \ + "beqz $r6, "#ndim"3f \n\t" \ + #ndim"4: \n\t" \ + COPY_4x##ndim \ + "addi.d $r6, $r6, -1 \n\t" \ + "bnez $r6, "#ndim"4b \n\t" \ + #ndim"3: \n\t" \ + "andi $r6, %5, 2 \n\t" \ + "beqz $r6, "#ndim"1f \n\t" \ + #ndim"2: \n\t" \ + COPY_2x##ndim \ + #ndim"1: \n\t" \ + "andi $r6, %5, 1 \n\t" \ + "beqz $r6, "#ndim"0f \n\t" \ + COPY_1x##ndim \ + #ndim"0: \n\t" \ + :"+r"(src),"+r"(dst),"+r"(src_ld_bytes),"+r"(dst_ld_bytes),"+r"(dst_tmp) \ + :"r"(num_rows),"r"(&ALPHA) \ + :"memory", "$r6", "$f0", "$f1", "$f2", "$f3", "$f4", "$f5", "$f6", "$f7", "$f8", "$f9", "$f10", "$f11", "$f15" \ + ); + +int CNAME(BLASLONG rows, BLASLONG cols, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG ldb){ + float *src, *dst, *dst_tmp=0, *src_base, *dst_base; + uint64_t src_ld_bytes = (uint64_t)lda * sizeof(float), dst_ld_bytes = (uint64_t)ldb * sizeof(float), num_rows = 0; + BLASLONG cols_left, rows_done; float ALPHA = alpha; + if (ALPHA == 0.0) { + dst_base = b; + for (cols_left = cols; cols_left > 0; cols_left--) {memset(dst_base, 0, rows * sizeof(float)); dst_base += ldb;} + return 0; + } + for (rows_done = 0; rows_done < rows; rows_done += num_rows) { + num_rows = rows - rows_done; + if (num_rows > ROWS_OF_BLOCK) num_rows = ROWS_OF_BLOCK; + cols_left = cols; src_base = a + (int64_t)lda * (int64_t)rows_done; dst_base = b + rows_done; + for (;cols_left > 15; cols_left -= 16) {COMPUTE(16) src_base += 16; dst_base += 16;} + for (;cols_left > 7; cols_left -= 8) {COMPUTE(8) src_base += 8; dst_base += 8;} + for (;cols_left > 3; cols_left -= 4) {COMPUTE(4) src_base += 4; dst_base += 4;} + for (;cols_left > 1; cols_left -= 2) {COMPUTE(2) src_base += 2; dst_base += 2;} + if (cols_left > 0) {COMPUTE(1) src_base ++; dst_base ++;} + } +} From b37129341ba4372caf2fdc680edbb09f0628d06b Mon Sep 17 00:00:00 2001 From: gxw Date: Thu, 17 Oct 2024 11:27:55 +0000 Subject: [PATCH 3/5] LoongArch64: Opt somatcopy_cn with LASX --- kernel/loongarch64/KERNEL.LA464 | 1 + kernel/loongarch64/somatcopy_cn_lasx.c | 193 +++++++++++++++++++++++++ 2 files changed, 194 insertions(+) create mode 100644 kernel/loongarch64/somatcopy_cn_lasx.c diff --git a/kernel/loongarch64/KERNEL.LA464 b/kernel/loongarch64/KERNEL.LA464 index 1a6b6ee6c..f664f551b 100644 --- a/kernel/loongarch64/KERNEL.LA464 +++ b/kernel/loongarch64/KERNEL.LA464 @@ -175,4 +175,5 @@ DGEMM_SMALL_K_B0_TT = dgemm_small_kernel_tt_lasx.S SOMATCOPY_RT = somatcopy_rt_lasx.c SOMATCOPY_RN = somatcopy_rn_lasx.c +SOMATCOPY_CN = somatcopy_cn_lasx.c endif diff --git a/kernel/loongarch64/somatcopy_cn_lasx.c b/kernel/loongarch64/somatcopy_cn_lasx.c new file mode 100644 index 000000000..e4d1f9f90 --- /dev/null +++ b/kernel/loongarch64/somatcopy_cn_lasx.c @@ -0,0 +1,193 @@ +/*************************************************************************** +Copyright (c) 2024, 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" +#define SAVE_4x1(b1, b2, b3, b4) \ + "vstelm.w $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.w $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" \ + "vstelm.w $vr"#b3",%1,0,0;add.d %1,%1,%3;vstelm.w $vr"#b4",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_4x2(b1, b2, b3, b4) \ + "vstelm.d $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.d $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" \ + "vstelm.d $vr"#b3",%1,0,0;add.d %1,%1,%3;vstelm.d $vr"#b4",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_4x4(b1, b2, b3, b4) \ + "vst $vr"#b1",%1,0;add.d %1,%1,%3;vst $vr"#b2",%1,0;add.d %1,%1,%3;\n\t" \ + "vst $vr"#b3",%1,0;add.d %1,%1,%3;vst $vr"#b4",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_4x8(b1, b2, b3, b4) \ + "xvst $xr"#b1",%1,0;add.d %1,%1,%3;xvst $xr"#b2",%1,0;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b3",%1,0;add.d %1,%1,%3;xvst $xr"#b4",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_4x16(b1, b2, b3, b4, b5, b6, b7, b8) \ + "xvst $xr"#b1",%1,0;xvst $xr"#b2",%1,32;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b3",%1,0;xvst $xr"#b4",%1,32;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b5",%1,0;xvst $xr"#b6",%1,32;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b7",%1,0;xvst $xr"#b8",%1,32;add.d %1,%1,%3;\n\t" +#define SAVE_2x16(b1, b2, b3, b4) \ + "xvst $xr"#b1",%1,0;xvst $xr"#b2",%1,32;add.d %1,%1,%3;\n\t" \ + "xvst $xr"#b3",%1,0;xvst $xr"#b4",%1,32;add.d %1,%1,%3;\n\t" +#define SAVE_2x8(b1, b2) \ + "xvst $xr"#b1",%1,0;add.d %1,%1,%3;xvst $xr"#b2",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_2x4(b1, b2) \ + "vst $vr"#b1",%1,0;add.d %1,%1,%3;vst $vr"#b2",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_2x2(b1, b2) \ + "vstelm.d $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.d $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_2x1(b1, b2) \ + "vstelm.w $vr"#b1",%1,0,0;add.d %1,%1,%3;vstelm.w $vr"#b2",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_1x16(b1, b2) \ + "xvst $xr"#b1",%1,0;xvst $xr"#b2",%1,32;add.d %1,%1,%3;\n\t" +#define SAVE_1x8(b1) \ + "xvst $xr"#b1",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_1x4(b1) \ + "vst $vr"#b1",%1,0;add.d %1,%1,%3;\n\t" +#define SAVE_1x2(b1) \ + "vstelm.d $vr"#b1",%1,0,0;add.d %1,%1,%3;\n\t" +#define SAVE_1x1(b1) \ + "vstelm.w $vr"#b1",%1,0,0;add.d %1,%1,%3;\n\t" +#define COPY_4x16 \ + "xvld $xr0,%0,0; xvld $xr4,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; xvld $xr5,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr2,%0,0; xvld $xr6,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr3,%0,0; xvld $xr7,%0,32; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + "xvfmul.s $xr4,$xr4,$xr15;xvfmul.s $xr5,$xr5,$xr15;xvfmul.s $xr6,$xr6,$xr15;xvfmul.s $xr7,$xr7,$xr15 \n\t" \ + SAVE_4x16(0,4,1,5,2,6,3,7) +#define COPY_4x8 \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr2,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr3,%0,0; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + SAVE_4x8(0,1,2,3) +#define COPY_4x4 \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \ + SAVE_4x4(0,1,2,3) +#define COPY_4x2 \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \ + SAVE_4x2(0,1,2,3) +#define COPY_4x1 \ + "fld.s $f0,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f1,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f2,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f3,%0,0; add.d %0,%0,%2; \n\t" \ + "fmul.s $f0,$f0,$f15;fmul.s $f1,$f1,$f15;fmul.s $f2,$f2,$f15;fmul.s $f3,$f3,$f15 \n\t" \ + SAVE_4x1(0,1,2,3) +#define COPY_2x16 \ + "xvld $xr0,%0,0; xvld $xr2,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; xvld $xr3,%0,32; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + SAVE_2x16(0,2,1,3) +#define COPY_2x8 \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15; \n\t" \ + SAVE_2x8(0,1) +#define COPY_2x4 \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15; \n\t" \ + SAVE_2x4(0,1) +#define COPY_2x2 \ + "fld.d $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fld.d $f1,%0,0;add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15; \n\t" \ + SAVE_2x2(0,1) +#define COPY_2x1 \ + "fld.s $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fld.s $f1,%0,0;add.d %0,%0,%2; \n\t" \ + "fmul.s $f0,$f0,$f15;fmul.s $f1,$f1,$f15; \n\t" \ + SAVE_2x1(0,1) +#define COPY_1x16 \ + "xvld $xr0,%0,0; xvld $xr1,%0,32; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15; \n\t" \ + SAVE_1x16(0,1) +#define COPY_1x8 \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15; \n\t" \ + SAVE_1x8(0) +#define COPY_1x4 \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15; \n\t" \ + SAVE_1x4(0) +#define COPY_1x2 \ + "fld.d $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15; \n\t" \ + SAVE_1x2(0) +#define COPY_1x1 \ + "fld.s $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fmul.s $f0,$f0,$f15; \n\t" \ + SAVE_1x1(0) +#define ROWS_OF_BLOCK 128 +#define COMPUTE(ndim) \ + src = src_base; dst = dst_base; \ + __asm__ __volatile__( \ + "xvldrepl.w $xr15, %6, 0 \n\t" \ + "srli.d $r6, %5, 2 \n\t" \ + "beqz $r6, "#ndim"3f \n\t" \ + #ndim"4: \n\t" \ + COPY_4x##ndim \ + "addi.d $r6, $r6, -1 \n\t" \ + "bnez $r6, "#ndim"4b \n\t" \ + #ndim"3: \n\t" \ + "andi $r6, %5, 2 \n\t" \ + "beqz $r6, "#ndim"1f \n\t" \ + #ndim"2: \n\t" \ + COPY_2x##ndim \ + #ndim"1: \n\t" \ + "andi $r6, %5, 1 \n\t" \ + "beqz $r6, "#ndim"0f \n\t" \ + COPY_1x##ndim \ + #ndim"0: \n\t" \ + :"+r"(src),"+r"(dst),"+r"(src_ld_bytes),"+r"(dst_ld_bytes),"+r"(dst_tmp) \ + :"r"(num_cols),"r"(&ALPHA) \ + :"memory", "$r6", "$f0", "$f1", "$f2", "$f3", "$f4", "$f5", "$f6", "$f7", "$f8", "$f9", "$f10", "$f11", "$f15" \ + ); + +int CNAME(BLASLONG rows, BLASLONG cols, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG ldb){ + float *src, *dst, *dst_tmp=0, *src_base, *dst_base; + uint64_t src_ld_bytes = (uint64_t)lda * sizeof(float), dst_ld_bytes = (uint64_t)ldb * sizeof(float), num_cols = 0; + BLASLONG rows_left, cols_done; float ALPHA = alpha; + if (ALPHA == 0.0) { + dst_base = b; + for (rows_left = rows; rows_left > 0; rows_left--) {memset(dst_base, 0, cols * sizeof(float)); dst_base += ldb;} + return 0; + } + for (cols_done = 0; cols_done < cols; cols_done += num_cols) { + num_cols = cols - cols_done; + if (num_cols > ROWS_OF_BLOCK) num_cols = ROWS_OF_BLOCK; + rows_left = rows; src_base = a + (int64_t)lda * (int64_t)cols_done; dst_base = b + cols_done; + for (;rows_left > 15; rows_left -= 16) {COMPUTE(16) src_base += 16; dst_base += 16;} + for (;rows_left > 7; rows_left -= 8) {COMPUTE(8) src_base += 8; dst_base += 8;} + for (;rows_left > 3; rows_left -= 4) {COMPUTE(4) src_base += 4; dst_base += 4;} + for (;rows_left > 1; rows_left -= 2) {COMPUTE(2) src_base += 2; dst_base += 2;} + if (rows_left > 0) {COMPUTE(1) src_base ++; dst_base ++;} + } +} From bb31bbef522b8b105ae2cb1bfce484ded5839b22 Mon Sep 17 00:00:00 2001 From: gxw Date: Thu, 17 Oct 2024 11:45:13 +0000 Subject: [PATCH 4/5] LoongArch64: Opt somatcopy_ct with LASX --- kernel/loongarch64/KERNEL.LA464 | 1 + kernel/loongarch64/somatcopy_ct_lasx.c | 200 +++++++++++++++++++++++++ 2 files changed, 201 insertions(+) create mode 100644 kernel/loongarch64/somatcopy_ct_lasx.c diff --git a/kernel/loongarch64/KERNEL.LA464 b/kernel/loongarch64/KERNEL.LA464 index f664f551b..ca8c4d388 100644 --- a/kernel/loongarch64/KERNEL.LA464 +++ b/kernel/loongarch64/KERNEL.LA464 @@ -175,5 +175,6 @@ DGEMM_SMALL_K_B0_TT = dgemm_small_kernel_tt_lasx.S SOMATCOPY_RT = somatcopy_rt_lasx.c SOMATCOPY_RN = somatcopy_rn_lasx.c +SOMATCOPY_CT = somatcopy_ct_lasx.c SOMATCOPY_CN = somatcopy_cn_lasx.c endif diff --git a/kernel/loongarch64/somatcopy_ct_lasx.c b/kernel/loongarch64/somatcopy_ct_lasx.c new file mode 100644 index 000000000..c400d186e --- /dev/null +++ b/kernel/loongarch64/somatcopy_ct_lasx.c @@ -0,0 +1,200 @@ +/*************************************************************************** +Copyright (c) 2024, 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" + +#define SAVE1x4(c1) \ + "vfmul.s $vr"#c1",$vr"#c1",$vr15;vstelm.w $vr"#c1",%4,0,0;add.d %4,%4,%3;\n\t" \ + "vstelm.w $vr"#c1",%4,0,1;add.d %4,%4,%3;vstelm.w $vr"#c1",%4,0,2;add.d %4,%4,%3; \n\t" \ + "vstelm.w $vr"#c1",%4,0,3;add.d %4,%4,%3;\n\t" +#define SAVE_2x4(c1, c2, t1, t2) \ + "vilvl.w $vr"#t1",$vr"#c2",$vr"#c1";vfmul.s $vr"#t1",$vr"#t1",$vr15; \n\t" \ + "vstelm.d $vr"#t1",%4,0,0;add.d %4,%4,%3;vstelm.d $vr"#t1",%4,0,1;add.d %4,%4,%3 \n\t" \ + "vilvh.w $vr"#t2",$vr"#c2",$vr"#c1";vfmul.s $vr"#t2",$vr"#t2",$vr15; \n\t" \ + "vstelm.d $vr"#t2",%4,0,0;add.d %4,%4,%3;vstelm.d $vr"#t2",%4,0,1;add.d %4,%4,%3 \n\t" +#define SAVE_4x1(b1) \ + "vst $vr"#b1",%4,0; add.d %4,%4,%3 \n\t" +#define SAVE_4x2(b1, b2) \ + "vst $vr"#b1",%4,0; add.d %4,%4,%3; vst $vr"#b2",%4,0; add.d %4,%4,%3 \n\t" +#define SAVE_4x4(b1, b2, b3, b4) \ + "vst $vr"#b1",%4,0; add.d %4,%4,%3; vst $vr"#b2",%4,0; add.d %4,%4,%3 \n\t" \ + "vst $vr"#b3",%4,0; add.d %4,%4,%3; vst $vr"#b4",%4,0; add.d %4,%4,%3 \n\t" +#define SAVE_4x8(b1, b2, b3, b4) \ + SAVE_4x4(b1, b2, b3, b4) \ + "xvpermi.q $xr"#b1",$xr"#b1",1; xvpermi.q $xr"#b2",$xr"#b2",1; xvpermi.q $xr"#b3",$xr"#b3",1; xvpermi.q $xr"#b4",$xr"#b4",1; \n\t" \ + SAVE_4x4(b1, b2, b3, b4) + +#define TRANS_4x4(a1, a2, a3, a4, t1, t2, t3, t4) \ + "vilvl.w $vr"#t1",$vr"#a2",$vr"#a1";vilvh.w $vr"#t2",$vr"#a2",$vr"#a1"; \n\t" \ + "vilvl.w $vr"#t3",$vr"#a4",$vr"#a3";vilvh.w $vr"#t4",$vr"#a4",$vr"#a3"; \n\t" \ + "vilvl.d $vr"#a1",$vr"#t3",$vr"#t1";vilvh.d $vr"#a2",$vr"#t3",$vr"#t1"; \n\t" \ + "vilvl.d $vr"#a3",$vr"#t4",$vr"#t2";vilvh.d $vr"#a4",$vr"#t4",$vr"#t2"; \n\t" +#define TRANS_4x8(a1, a2, a3, a4, t1, t2, t3, t4) \ + "xvilvl.w $xr"#t1",$xr"#a2",$xr"#a1"; xvilvh.w $xr"#t2",$xr"#a2",$xr"#a1"; \n\t" \ + "xvilvl.w $xr"#t3",$xr"#a4",$xr"#a3"; xvilvh.w $xr"#t4",$xr"#a4",$xr"#a3"; \n\t" \ + "xvilvl.d $xr"#a1",$xr"#t3",$xr"#t1"; xvilvh.d $xr"#a2",$xr"#t3",$xr"#t1"; \n\t" \ + "xvilvl.d $xr"#a3",$xr"#t4",$xr"#t2"; xvilvh.d $xr"#a4",$xr"#t4",$xr"#t2"; \n\t" +#define COPY_4x16 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "xvld $xr0,%0,0; xvld $xr4,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; xvld $xr5,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr2,%0,0; xvld $xr6,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr3,%0,0; xvld $xr7,%0,32; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + "xvfmul.s $xr4,$xr4,$xr15;xvfmul.s $xr5,$xr5,$xr15;xvfmul.s $xr6,$xr6,$xr15;xvfmul.s $xr7,$xr7,$xr15 \n\t" \ + TRANS_4x8(0,1,2,3,8,9,10,11) SAVE_4x8(0,1,2,3) \ + TRANS_4x8(4,5,6,7,8,9,10,11) SAVE_4x8(4,5,6,7) +#define COPY_4x8 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr2,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr3,%0,0; add.d %0,%0,%2; \n\t" \ + "xvfmul.s $xr0,$xr0,$xr15;xvfmul.s $xr1,$xr1,$xr15;xvfmul.s $xr2,$xr2,$xr15;xvfmul.s $xr3,$xr3,$xr15 \n\t" \ + TRANS_4x8(0,1,2,3,8,9,10,11) SAVE_4x8(0,1,2,3) +#define COPY_4x4 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \ + TRANS_4x4(0,1,2,3,8,9,10,11) SAVE_4x4(0,1,2,3) +#define COPY_4x2 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr2,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr3,%0,0; add.d %0,%0,%2; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15;vfmul.s $vr1,$vr1,$vr15;vfmul.s $vr2,$vr2,$vr15;vfmul.s $vr3,$vr3,$vr15 \n\t" \ + TRANS_4x4(0,1,2,3,8,9,10,11) SAVE_4x2(0,1) +#define COPY_4x1 \ + "move %4,%1; addi.d %1,%1,16 \n\t" \ + "fld.s $f0,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f1,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f2,%0,0; add.d %0,%0,%2; \n\t" \ + "fld.s $f3,%0,0; add.d %0,%0,%2; \n\t" \ + "xvinsve0.w $xr0,$xr1,1;xvinsve0.w $xr0,$xr2,2;xvinsve0.w $xr0,$xr3,3; \n\t" \ + "vfmul.s $vr0,$vr0,$vr15; \n\t" \ + SAVE_4x1(0) + +#define COPY_2x16 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "xvld $xr0,%0,0; xvld $xr2,%0,32; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; xvld $xr3,%0,32; add.d %0,%0,%2; \n\t" \ + "xvpermi.q $xr4,$xr0,1;xvpermi.q $xr6,$xr2,1;xvpermi.q $xr5,$xr1,1;xvpermi.q $xr7,$xr3,1; \n\t" \ + SAVE_2x4(0,1,8,9) SAVE_2x4(4,5,8,9) SAVE_2x4(2,3,8,9) SAVE_2x4(6,7,8,9) +#define COPY_2x8 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "xvld $xr0,%0,0; add.d %0,%0,%2; \n\t" \ + "xvld $xr1,%0,0; add.d %0,%0,%2; \n\t" \ + "xvpermi.q $xr2,$xr0,1;xvpermi.q $xr3,$xr1,1; \n\t" \ + SAVE_2x4(0,1,4,5) SAVE_2x4(2,3,4,5) +#define COPY_2x4 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "vld $vr0,%0,0; add.d %0,%0,%2; \n\t" \ + "vld $vr1,%0,0; add.d %0,%0,%2; \n\t" \ + SAVE_2x4(0,1,4,5) +#define COPY_2x2 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "fld.d $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fld.d $f1,%0,0;add.d %0,%0,%2; \n\t" \ + "xvinsve0.d $xr0,$xr1,1;vfmul.s $vr0,$vr0,$vr15;vshuf4i.w $vr0,$vr0,0xd8 \n\t" \ + "vstelm.d $vr0,%4,0,0;add.d %4,%4,%3;vstelm.d $vr0,%4,0,1 \n\t" +#define COPY_2x1 \ + "move %4,%1; addi.d %1,%1,8 \n\t" \ + "fld.s $f0,%0,0;add.d %0,%0,%2; \n\t" \ + "fld.s $f1,%0,0;add.d %0,%0,%2; \n\t" \ + "xvinsve0.w $xr0,$xr1,1;vfmul.s $vr0,$vr0,$vr15; \n\t" \ + "vstelm.d $vr0,%4,0,0; \n\t" + +#define COPY_1x16 \ + "move %4,%1; addi.d %1,%1,4 \n\t" \ + "vld $vr1,%0,0;" SAVE1x4(1) "vld $vr2,%0,16;" SAVE1x4(2) \ + "vld $vr1,%0,32;" SAVE1x4(1) "vld $vr2,%0,48;" SAVE1x4(2) \ + "add.d %0,%0,%2 \n\t" +#define COPY_1x8 \ + "move %4,%1; addi.d %1,%1,4 \n\t" \ + "vld $vr1,%0,0;" SAVE1x4(1) "vld $vr2,%0,16;" SAVE1x4(2) \ + "add.d %0,%0,%2 \n\t" +#define COPY_1x4 \ + "move %4,%1; addi.d %1,%1,4 \n\t" \ + "vld $vr1,%0,0;" SAVE1x4(1) \ + "add.d %0,%0,%2 \n\t" +#define COPY_1x2 \ + "move %4,%1;fld.d $f1,%0,0;add.d %0,%0,%2;vfmul.s $vr1,$vr1,$vr15;vstelm.w $vr1,%4,0,0;add.d %4,%4,%3;vstelm.w $vr1,%4,0,1;\n\t" \ + "addi.d %1,%1,4;\n\t" +#define COPY_1x1 \ + "fld.s $f1,%0,0;fmul.s $f1,$f1,$f15;fst.s $f1,%1,0;add.d %0,%0,%2;addi.d %1,%1,4;\n\t" + +#define ROWS_OF_BLOCK 128 + +#define COMPUTE(ndim) \ + src = src_base; dst = dst_base; \ + __asm__ __volatile__( \ + "xvldrepl.w $xr15, %6, 0 \n\t" \ + "srli.d $r6, %5, 2 \n\t" \ + "beqz $r6, "#ndim"3f \n\t" \ + #ndim"4: \n\t" \ + COPY_4x##ndim \ + "addi.d $r6, $r6, -1 \n\t" \ + "bnez $r6, "#ndim"4b \n\t" \ + #ndim"3: \n\t" \ + "andi $r6, %5, 2 \n\t" \ + "beqz $r6, "#ndim"1f \n\t" \ + #ndim"2: \n\t" \ + COPY_2x##ndim \ + #ndim"1: \n\t" \ + "andi $r6, %5, 1 \n\t" \ + "beqz $r6, "#ndim"0f \n\t" \ + COPY_1x##ndim \ + #ndim"0: \n\t" \ + :"+r"(src),"+r"(dst),"+r"(src_ld_bytes),"+r"(dst_ld_bytes),"+r"(dst_tmp) \ + :"r"(num_cols),"r"(&ALPHA) \ + :"memory", "$r6", "$f0", "$f1", "$f2", "$f3", "$f4", "$f5", "$f6", "$f7", "$f8", "$f9", "$f10", "$f11", "$f15" \ + ); + +int CNAME(BLASLONG rows, BLASLONG cols, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG ldb){ + float *src, *dst, *dst_tmp=0, *src_base, *dst_base; + uint64_t src_ld_bytes = (uint64_t)lda * sizeof(float), dst_ld_bytes = (uint64_t)ldb * sizeof(float), num_cols = 0; + BLASLONG rows_left, cols_done; float ALPHA = alpha; + if (ALPHA == 0.0) { + dst_base = b; + for (rows_left = rows; rows_left > 0; rows_left--) {memset(dst_base, 0, cols * sizeof(float)); dst_base += ldb;} + return 0; + } + for (cols_done = 0; cols_done < cols; cols_done += num_cols) { + num_cols = cols - cols_done; + if (num_cols > ROWS_OF_BLOCK) num_cols = ROWS_OF_BLOCK; + rows_left = rows; src_base = a + (int64_t)lda * (int64_t)cols_done; dst_base = b + cols_done; + for (;rows_left > 15; rows_left -= 16) {COMPUTE(16) src_base += 16; dst_base += 16 * ldb;} + for (;rows_left > 7; rows_left -= 8) {COMPUTE(8) src_base += 8; dst_base += 8 * ldb;} + for (;rows_left > 3; rows_left -= 4) {COMPUTE(4) src_base += 4; dst_base += 4 * ldb;} + for (;rows_left > 1; rows_left -= 2) {COMPUTE(2) src_base += 2; dst_base += 2 * ldb;} + if (rows_left > 0) {COMPUTE(1) src_base ++; dst_base += ldb;} + } +} From ffaa5765a41e283916049820a8fc9d805c72f04f Mon Sep 17 00:00:00 2001 From: gxw Date: Thu, 17 Oct 2024 12:32:54 +0000 Subject: [PATCH 5/5] Bench: Add omatcopy --- benchmark/Makefile | 39 +++++++++++++- benchmark/omatcopy.c | 122 +++++++++++++++++++++++++++++++++++++++++++ 2 files changed, 160 insertions(+), 1 deletion(-) create mode 100644 benchmark/omatcopy.c diff --git a/benchmark/Makefile b/benchmark/Makefile index b7493950a..c295b1458 100644 --- a/benchmark/Makefile +++ b/benchmark/Makefile @@ -103,6 +103,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \ sgetri.goto dgetri.goto cgetri.goto zgetri.goto \ spotrf.goto dpotrf.goto cpotrf.goto zpotrf.goto \ ssymm.goto dsymm.goto csymm.goto zsymm.goto \ + somatcopy.goto domatcopy.goto comatcopy.goto zomatcopy.goto \ saxpby.goto daxpby.goto caxpby.goto zaxpby.goto $(GOTO_HALF_TARGETS) acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \ @@ -276,6 +277,7 @@ goto :: sgemm.goto dgemm.goto cgemm.goto zgemm.goto \ samin.goto damin.goto camin.goto zamin.goto \ smin.goto dmin.goto \ saxpby.goto daxpby.goto caxpby.goto zaxpby.goto \ + somatcopy.goto domatcopy.goto comatcopy.goto zomatcopy.goto \ snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto $(GOTO_LAPACK_TARGETS) $(GOTO_HALF_TARGETS) acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \ @@ -2906,6 +2908,29 @@ dznrm2.goto : dznrm2.$(SUFFIX) ../$(LIBNAME) dznrm2.atlas : dznrm2.$(SUFFIX) -$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) +################################################################################################### + +############################################ SOMATCOPY ############################################ +somatcopy.goto : somatcopy.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +################################################################################################### + +############################################ DOMATCOPY ############################################ +domatcopy.goto : domatcopy.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +################################################################################################### + +############################################ COMATCOPY ############################################ +comatcopy.goto : comatcopy.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm + +################################################################################################### + +############################################ ZOMATCOPY ############################################ +zomatcopy.goto : zomatcopy.$(SUFFIX) ../$(LIBNAME) + $(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm ################################################################################################### @@ -3435,6 +3460,18 @@ scnrm2.$(SUFFIX) : nrm2.c dznrm2.$(SUFFIX) : nrm2.c $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ +somatcopy.$(SUFFIX) : omatcopy.c + $(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^ + +domatcopy.$(SUFFIX) : omatcopy.c + $(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^ + +comatcopy.$(SUFFIX) : omatcopy.c + $(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^ + +zomatcopy.$(SUFFIX) : omatcopy.c + $(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^ + smallscaling: smallscaling.c ../$(LIBNAME) $(CC) $(CFLAGS) -o $(@F) $^ $(EXTRALIB) -fopenmp -lm -lpthread @@ -3442,4 +3479,4 @@ smallscaling: smallscaling.c ../$(LIBNAME) clean :: @rm -f *.goto *.mkl *.acml *.atlas *.veclib *.essl smallscaling -include $(TOPDIR)/Makefile.tail \ No newline at end of file +include $(TOPDIR)/Makefile.tail diff --git a/benchmark/omatcopy.c b/benchmark/omatcopy.c new file mode 100644 index 000000000..c3348c9be --- /dev/null +++ b/benchmark/omatcopy.c @@ -0,0 +1,122 @@ +/*************************************************************************** +Copyright (c) 2024, 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 "bench.h" + +#undef OMATCOPY + +#ifndef COMPLEX +#ifdef DOUBLE +#define OMATCOPY BLASFUNC(domatcopy) +#else +#define OMATCOPY BLASFUNC(somatcopy) +#endif +#else +#ifdef DOUBLE +#define OMATCOPY BLASFUNC(zomatcopy) +#else +#define OMATCOPY BLASFUNC(comatcopy) +#endif +#endif +int main(int argc, char *argv[]){ + FLOAT *a, *b; + FLOAT alpha[] = {1.0, 0.0}; + char trans = 'N'; + char order = 'C'; + blasint crows, ccols, clda, cldb; + int loops = 1; + char *p; + + int from = 1; + int to = 200; + int step = 1; + int i, j; + + double time1, timeg; + + argc--;argv++; + + if (argc > 0) { from = atol(*argv); argc--; argv++; } + if (argc > 0) { to = MAX(atol(*argv), from); argc--; argv++; } + if (argc > 0) { step = atol(*argv); argc--; argv++; } + + if ((p = getenv("OPENBLAS_TRANS"))) { + trans=*p; + } + if ((p = getenv("OPENBLAS_ORDER"))) { + order=*p; + } + TOUPPER(trans); + TOUPPER(order); + fprintf(stderr, "From : %3d To : %3d Step=%d : Trans=%c : Order=%c\n", from, to, step, trans, order); + p = getenv("OPENBLAS_LOOPS"); + if ( p != NULL ) { + loops = atoi(p); + } + + if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL) { + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + if (( b = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL) { + fprintf(stderr,"Out of Memory!!\n");exit(1); + } + +#ifdef __linux + srandom(getpid()); +#endif + + for (i = 0; i < to * to * COMPSIZE; i++) { + a[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + for (i = 0; i < to * to * COMPSIZE; i++) { + b[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5; + } + + fprintf(stderr, " SIZE Flops Time\n"); + for (i = from; i <= to; i += step) { + cldb = clda = crows = ccols = i; + fprintf(stderr, " ROWS=%4d, COLS=%4d : ", (int)crows, (int)ccols); + begin(); + + for (j=0; j