From ab3ffab96a4addbbed420e9c9304e909b778ce03 Mon Sep 17 00:00:00 2001 From: Ashwin Sekhar T K Date: Tue, 21 Feb 2017 03:25:00 -0800 Subject: [PATCH] THUNDERX2T99: Add Optimized C/Z DOT Implementation --- kernel/arm64/KERNEL.THUNDERX2T99 | 3 +- kernel/arm64/zdot_thunderx2t99.c | 357 +++++++++++++++++++++++++++++++ 2 files changed, 359 insertions(+), 1 deletion(-) create mode 100644 kernel/arm64/zdot_thunderx2t99.c diff --git a/kernel/arm64/KERNEL.THUNDERX2T99 b/kernel/arm64/KERNEL.THUNDERX2T99 index 1c31cea51..15537ccda 100644 --- a/kernel/arm64/KERNEL.THUNDERX2T99 +++ b/kernel/arm64/KERNEL.THUNDERX2T99 @@ -27,7 +27,8 @@ DAXPYKERNEL = daxpy_thunderx2t99.S DDOTKERNEL = dot_thunderx2t99.c SDOTKERNEL = dot_thunderx2t99.c - +CDOTKERNEL = zdot_thunderx2t99.c +ZDOTKERNEL = zdot_thunderx2t99.c ifeq ($(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N), 8x4) DGEMMKERNEL = dgemm_kernel_8x4_thunderx2t99.S diff --git a/kernel/arm64/zdot_thunderx2t99.c b/kernel/arm64/zdot_thunderx2t99.c new file mode 100644 index 000000000..64823871f --- /dev/null +++ b/kernel/arm64/zdot_thunderx2t99.c @@ -0,0 +1,357 @@ +/*************************************************************************** +Copyright (c) 2017, 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" + +#include + +#define N "x0" /* vector length */ +#define X "x1" /* "X" vector address */ +#define INC_X "x2" /* "X" stride */ +#define Y "x3" /* "Y" vector address */ +#define INC_Y "x4" /* "Y" stride */ +#define J "x5" /* loop variable */ + +#if !defined(DOUBLE) +#define REG0 "wzr" +#define DOTF "s0" +#define DOTI "s1" +#define INC_SHIFT "3" +#define N_DIV_SHIFT "4" +#define N_REM_MASK "15" +#else +#define REG0 "xzr" +#define DOTF "d0" +#define DOTI "d1" +#define INC_SHIFT "4" +#define N_DIV_SHIFT "3" +#define N_REM_MASK "7" +#endif + +#if !defined(CONJ) +#define f_ii "fmls" +#define f_ir "fmla" +#define a_ii "fsub" +#define a_ir "fadd" +#else +#define f_ii "fmla" +#define f_ir "fmls" +#define a_ii "fadd" +#define a_ir "fsub" +#endif + +#if !defined(DOUBLE) +#define KERNEL_F1 \ + " ldr d16, ["X"] \n" \ + " ldr d24, ["Y"] \n" \ + " add "X", "X", "INC_X" \n" \ + " add "Y", "Y", "INC_Y" \n" \ + " ins v17.s[0], v16.s[1] \n" \ + " fmla "DOTF", s16, v24.s[0] \n" \ + " "f_ii" "DOTF", s17, v24.s[1] \n" \ + " "f_ir" "DOTI", s17, v24.s[0] \n" \ + " fmla "DOTI", s16, v24.s[1] \n" + +#define KERNEL_F \ + " ld2 {v16.4s, v17.4s}, ["X"] \n" \ + " ld2 {v24.4s, v25.4s}, ["Y"] \n" \ + " add "X", "X", #32 \n" \ + " add "Y", "Y", #32 \n" \ + " ld2 {v18.4s, v19.4s}, ["X"] \n" \ + " ld2 {v26.4s, v27.4s}, ["Y"] \n" \ + " add "X", "X", #32 \n" \ + " add "Y", "Y", #32 \n" \ + " fmla v0.4s, v16.4s, v24.4s \n" \ + " fmla v1.4s, v17.4s, v25.4s \n" \ + " fmla v2.4s, v16.4s, v25.4s \n" \ + " fmla v3.4s, v17.4s, v24.4s \n" \ + " ld2 {v20.4s, v21.4s}, ["X"] \n" \ + " ld2 {v28.4s, v29.4s}, ["Y"] \n" \ + " add "X", "X", #32 \n" \ + " add "Y", "Y", #32 \n" \ + " fmla v4.4s, v18.4s, v26.4s \n" \ + " fmla v5.4s, v19.4s, v27.4s \n" \ + " fmla v6.4s, v18.4s, v27.4s \n" \ + " fmla v7.4s, v19.4s, v26.4s \n" \ + " ld2 {v22.4s, v23.4s}, ["X"] \n" \ + " ld2 {v30.4s, v31.4s}, ["Y"] \n" \ + " fmla v0.4s, v20.4s, v28.4s \n" \ + " fmla v1.4s, v21.4s, v29.4s \n" \ + " fmla v2.4s, v20.4s, v29.4s \n" \ + " fmla v3.4s, v21.4s, v28.4s \n" \ + " add "X", "X", #32 \n" \ + " add "Y", "Y", #32 \n" \ + " PRFM PLDL1KEEP, ["X", #1024] \n" \ + " PRFM PLDL1KEEP, ["Y", #1024] \n" \ + " PRFM PLDL1KEEP, ["X", #1024+64] \n" \ + " PRFM PLDL1KEEP, ["Y", #1024+64] \n" \ + " fmla v4.4s, v22.4s, v30.4s \n" \ + " fmla v5.4s, v23.4s, v31.4s \n" \ + " fmla v6.4s, v22.4s, v31.4s \n" \ + " fmla v7.4s, v23.4s, v30.4s \n" + +#define KERNEL_F_FINALIZE \ + " fadd v0.4s, v0.4s, v4.4s \n" \ + " fadd v1.4s, v1.4s, v5.4s \n" \ + " fadd v2.4s, v2.4s, v6.4s \n" \ + " fadd v3.4s, v3.4s, v7.4s \n" \ + " "a_ii" v0.4s, v0.4s, v1.4s \n" \ + " "a_ir" v1.4s, v2.4s, v3.4s \n" \ + " faddp v0.4s, v0.4s, v0.4s \n" \ + " faddp v0.4s, v0.4s, v0.4s \n" \ + " faddp v1.4s, v1.4s, v1.4s \n" \ + " faddp v1.4s, v1.4s, v1.4s \n" + +#else + +#define KERNEL_F1 \ + " ldr q16, ["X"] \n" \ + " ldr q24, ["Y"] \n" \ + " add "X", "X", "INC_X" \n" \ + " add "Y", "Y", "INC_Y" \n" \ + " ins v17.d[0], v16.d[1] \n" \ + " fmla "DOTF", d16, v24.d[0] \n" \ + " "f_ii" "DOTF", d17, v24.d[1] \n" \ + " "f_ir" "DOTI", d17, v24.d[0] \n" \ + " fmla "DOTI", d16, v24.d[1] \n" + +#define KERNEL_F \ + " ld2 {v16.2d, v17.2d}, ["X"] \n" \ + " ld2 {v24.2d, v25.2d}, ["Y"] \n" \ + " add "X", "X", #32 \n" \ + " add "Y", "Y", #32 \n" \ + " ld2 {v18.2d, v19.2d}, ["X"] \n" \ + " ld2 {v26.2d, v27.2d}, ["Y"] \n" \ + " add "X", "X", #32 \n" \ + " add "Y", "Y", #32 \n" \ + " fmla v0.2d, v16.2d, v24.2d \n" \ + " fmla v1.2d, v17.2d, v25.2d \n" \ + " fmla v2.2d, v16.2d, v25.2d \n" \ + " fmla v3.2d, v17.2d, v24.2d \n" \ + " ld2 {v20.2d, v21.2d}, ["X"] \n" \ + " ld2 {v28.2d, v29.2d}, ["Y"] \n" \ + " add "X", "X", #32 \n" \ + " add "Y", "Y", #32 \n" \ + " fmla v4.2d, v18.2d, v26.2d \n" \ + " fmla v5.2d, v19.2d, v27.2d \n" \ + " fmla v6.2d, v18.2d, v27.2d \n" \ + " fmla v7.2d, v19.2d, v26.2d \n" \ + " ld2 {v22.2d, v23.2d}, ["X"] \n" \ + " ld2 {v30.2d, v31.2d}, ["Y"] \n" \ + " fmla v0.2d, v20.2d, v28.2d \n" \ + " fmla v1.2d, v21.2d, v29.2d \n" \ + " fmla v2.2d, v20.2d, v29.2d \n" \ + " fmla v3.2d, v21.2d, v28.2d \n" \ + " add "X", "X", #32 \n" \ + " add "Y", "Y", #32 \n" \ + " PRFM PLDL1KEEP, ["X", #1024] \n" \ + " PRFM PLDL1KEEP, ["Y", #1024] \n" \ + " PRFM PLDL1KEEP, ["X", #1024+64] \n" \ + " PRFM PLDL1KEEP, ["Y", #1024+64] \n" \ + " fmla v4.2d, v22.2d, v30.2d \n" \ + " fmla v5.2d, v23.2d, v31.2d \n" \ + " fmla v6.2d, v22.2d, v31.2d \n" \ + " fmla v7.2d, v23.2d, v30.2d \n" + +#define KERNEL_F_FINALIZE \ + " fadd v0.2d, v0.2d, v4.2d \n" \ + " fadd v1.2d, v1.2d, v5.2d \n" \ + " fadd v2.2d, v2.2d, v6.2d \n" \ + " fadd v3.2d, v3.2d, v7.2d \n" \ + " "a_ii" v0.2d, v0.2d, v1.2d \n" \ + " "a_ir" v1.2d, v2.2d, v3.2d \n" \ + " faddp "DOTF", v0.2d \n" \ + " faddp "DOTI", v1.2d \n" +#endif + +#if defined(SMP) +extern int blas_level1_thread_with_return_value(int mode, BLASLONG m, BLASLONG n, + BLASLONG k, void *alpha, void *a, BLASLONG lda, void *b, BLASLONG ldb, + void *c, BLASLONG ldc, int (*function)(), int nthreads); +#endif + +static void zdot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, OPENBLAS_COMPLEX_FLOAT *result) +{ + FLOAT dotr = 0.0, doti = 0.0; + CREAL(*result) = 0.0; + CIMAG(*result) = 0.0; + + if ( n < 0 ) return; + + __asm__ __volatile__ ( + " mov "N", %[N_] \n" + " mov "X", %[X_] \n" + " mov "INC_X", %[INCX_] \n" + " mov "Y", %[Y_] \n" + " mov "INC_Y", %[INCY_] \n" + " fmov "DOTF", "REG0" \n" + " fmov "DOTI", "REG0" \n" + " fmov d2, xzr \n" + " fmov d3, xzr \n" + " fmov d4, xzr \n" + " fmov d5, xzr \n" + " fmov d6, xzr \n" + " fmov d7, xzr \n" + " cmp "N", xzr \n" + " ble .Ldot_kernel_L999 \n" + " cmp "INC_X", #1 \n" + " bne .Ldot_kernel_S_BEGIN \n" + " cmp "INC_Y", #1 \n" + " bne .Ldot_kernel_S_BEGIN \n" + + ".Ldot_kernel_F_BEGIN: \n" + " lsl "INC_X", "INC_X", "INC_SHIFT" \n" + " lsl "INC_Y", "INC_Y", "INC_SHIFT" \n" + " asr "J", "N", #"N_DIV_SHIFT" \n" + " cmp "J", xzr \n" + " beq .Ldot_kernel_F1 \n" + + " .align 5 \n" + ".Ldot_kernel_F: \n" + " "KERNEL_F" \n" + " subs "J", "J", #1 \n" + " bne .Ldot_kernel_F \n" + " "KERNEL_F_FINALIZE" \n" + + ".Ldot_kernel_F1: \n" + " ands "J", "N", #"N_REM_MASK" \n" + " ble .Ldot_kernel_L999 \n" + + ".Ldot_kernel_F10: \n" + " "KERNEL_F1" \n" + " subs "J", "J", #1 \n" + " bne .Ldot_kernel_F10 \n" + " b .Ldot_kernel_L999 \n" + + ".Ldot_kernel_S_BEGIN: \n" + " lsl "INC_X", "INC_X", "INC_SHIFT" \n" + " lsl "INC_Y", "INC_Y", "INC_SHIFT" \n" + " asr "J", "N", #2 \n" + " cmp "J", xzr \n" + " ble .Ldot_kernel_S1 \n" + + ".Ldot_kernel_S4: \n" + " "KERNEL_F1" \n" + " "KERNEL_F1" \n" + " "KERNEL_F1" \n" + " "KERNEL_F1" \n" + " subs "J", "J", #1 \n" + " bne .Ldot_kernel_S4 \n" + + ".Ldot_kernel_S1: \n" + " ands "J", "N", #3 \n" + " ble .Ldot_kernel_L999 \n" + + ".Ldot_kernel_S10: \n" + " "KERNEL_F1" \n" + " subs "J", "J", #1 \n" + " bne .Ldot_kernel_S10 \n" + + ".Ldot_kernel_L999: \n" + " str "DOTF", [%[DOTR_]] \n" + " str "DOTI", [%[DOTI_]] \n" + + : + : [DOTR_] "r" (&dotr), //%0 + [DOTI_] "r" (&doti), //%1 + [N_] "r" (n), //%2 + [X_] "r" (x), //%3 + [INCX_] "r" (inc_x), //%4 + [Y_] "r" (y), //%5 + [INCY_] "r" (inc_y) //%6 + : "cc", + "memory", + "x0", "x1", "x2", "x3", "x4", "x5", + "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7" + ); + + CREAL(*result) = dotr; + CIMAG(*result) = doti; + return; +} + +#if defined(SMP) +static int zdot_thread_function(BLASLONG n, BLASLONG dummy0, + BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y, + BLASLONG inc_y, FLOAT *result, BLASLONG dummy3) +{ + zdot_compute(n, x, inc_x, y, inc_y, (void *)result); + + return 0; +} +#endif + +OPENBLAS_COMPLEX_FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) +{ +#if defined(SMP) + int nthreads; + FLOAT dummy_alpha; +#endif + OPENBLAS_COMPLEX_FLOAT zdot; + CREAL(zdot) = 0.0; + CIMAG(zdot) = 0.0; + +#if defined(SMP) + nthreads = num_cpu_avail(1); + + if (inc_x == 0 || inc_y == 0) + nthreads = 1; + + if (n <= 10000) + nthreads = 1; + + if (nthreads == 1) { + zdot_compute(n, x, inc_x, y, inc_y, &zdot); + } else { + int mode, i; + char result[MAX_CPU_NUMBER * sizeof(double) * 2]; + OPENBLAS_COMPLEX_FLOAT *ptr; + +#if !defined(DOUBLE) + mode = BLAS_SINGLE | BLAS_COMPLEX; +#else + mode = BLAS_DOUBLE | BLAS_COMPLEX; +#endif + + blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, + x, inc_x, y, inc_y, result, 0, + ( void *)zdot_thread_function, nthreads); + + ptr = (OPENBLAS_COMPLEX_FLOAT *)result; + for (i = 0; i < nthreads; i++) { + CREAL(zdot) = CREAL(zdot) + CREAL(*ptr); + CIMAG(zdot) = CIMAG(zdot) + CIMAG(*ptr); + ptr = (void *)(((char *)ptr) + sizeof(double) * 2); + } + } +#else + zdot_compute(n, x, inc_x, y, inc_y, &zdot); +#endif + + return zdot; +}