From: Ashwin Sekhar T K Date: Fri, 27 Jan 2017 09:11:58 +0000 (-0800) Subject: THUNDERX2T99: Remove Duplicate Code X-Git-Tag: upstream/0.2.20^2~51^2~18 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=e0dc5f58c5688df306291c657c1d9115688248d7;p=platform%2Fupstream%2Fopenblas.git THUNDERX2T99: Remove Duplicate Code --- diff --git a/kernel/arm64/KERNEL.THUNDERX2T99 b/kernel/arm64/KERNEL.THUNDERX2T99 index 94132bf6..c8668b67 100644 --- a/kernel/arm64/KERNEL.THUNDERX2T99 +++ b/kernel/arm64/KERNEL.THUNDERX2T99 @@ -1,19 +1,11 @@ include $(KERNELDIR)/KERNEL.CORTEXA57 -ifndef SMP -SNRM2KERNEL = snrm2_thunderx2t99.S -else -SNRM2KERNEL = nrm2_thunderx2t99_threaded.c -endif +SNRM2KERNEL = snrm2_thunderx2t99.c CNRM2KERNEL = cnrm2_thunderx2t99.S DAXPYKERNEL = daxpy_thunderx2t99.S -ifndef SMP -DDOTKERNEL = ddot_thunderx2t99.S -else -DDOTKERNEL = ddot_thunderx2t99_threaded.c -endif +DDOTKERNEL = ddot_thunderx2t99.c ifeq ($(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N), 8x4) DGEMMKERNEL = dgemm_kernel_8x4_thunderx2t99.S diff --git a/kernel/arm64/ddot_thunderx2t99.S b/kernel/arm64/ddot_thunderx2t99.S deleted file mode 100644 index 5fa39adf..00000000 --- a/kernel/arm64/ddot_thunderx2t99.S +++ /dev/null @@ -1,207 +0,0 @@ -/******************************************************************************* -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. -*******************************************************************************/ - -#define ASSEMBLER -#include "common.h" - -#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 I x5 /* loop variable */ - -/******************************************************************************* -* Macro definitions -*******************************************************************************/ - -#define REG0 xzr -#define DOTF d0 -#define TMPX d16 -#define LD1VX {v16.d}[0] -#define TMPY d24 -#define LD1VY {v24.d}[0] -#define SZ 8 - -/******************************************************************************/ - -.macro KERNEL_F1 - ldr TMPX, [X] - ldr TMPY, [Y] - add X, X, #SZ - add Y, Y, #SZ - fmadd DOTF, TMPX, TMPY, DOTF -.endm - -.macro KERNEL_F16 - ldp q16, q17, [X] - ldp q24, q25, [Y] - - ldp q18, q19, [X, #32] - ldp q26, q27, [Y, #32] - - fmla v0.2d, v16.2d, v24.2d - fmla v1.2d, v17.2d, v25.2d - - ldp q20, q21, [X, #64] - ldp q28, q29, [Y, #64] - - fmla v2.2d, v18.2d, v26.2d - fmla v3.2d, v19.2d, v27.2d - - ldp q22, q23, [X, #96] - ldp q30, q31, [Y, #96] - - add Y, Y, #128 - add X, X, #128 - - fmla v4.2d, v20.2d, v28.2d - fmla v5.2d, v21.2d, v29.2d - - PRFM PLDL1KEEP, [X, #896] - PRFM PLDL1KEEP, [Y, #896] - PRFM PLDL1KEEP, [X, #896+64] - PRFM PLDL1KEEP, [Y, #896+64] - - fmla v6.2d, v22.2d, v30.2d - fmla v7.2d, v23.2d, v31.2d -.endm - -.macro KERNEL_F32 - KERNEL_F16 - KERNEL_F16 -.endm - -.macro KERNEL_F32_FINALIZE - fadd v0.2d, v0.2d, v1.2d - fadd v2.2d, v2.2d, v3.2d - fadd v4.2d, v4.2d, v5.2d - fadd v6.2d, v6.2d, v7.2d - fadd v0.2d, v0.2d, v2.2d - fadd v4.2d, v4.2d, v6.2d - fadd v0.2d, v0.2d, v4.2d - faddp DOTF, v0.2d -.endm - -.macro INIT_S - lsl INC_X, INC_X, #3 - lsl INC_Y, INC_Y, #3 -.endm - -.macro KERNEL_S1 - ld1 LD1VX, [X], INC_X - ld1 LD1VY, [Y], INC_Y - fmadd DOTF, TMPX, TMPY, DOTF -.endm - -/******************************************************************************* -* End of macro definitions -*******************************************************************************/ - - PROLOGUE - - fmov DOTF, REG0 - fmov d1, REG0 - fmov d2, REG0 - fmov d3, REG0 - fmov d4, REG0 - fmov d5, REG0 - fmov d6, REG0 - fmov d7, REG0 - - cmp N, xzr - ble dot_kernel_L999 - - cmp INC_X, #1 - bne dot_kernel_S_BEGIN - cmp INC_Y, #1 - bne dot_kernel_S_BEGIN - -dot_kernel_F_BEGIN: - - asr I, N, #5 - cmp I, xzr - beq dot_kernel_F1 - -dot_kernel_F32: - - KERNEL_F32 - - subs I, I, #1 - bne dot_kernel_F32 - - KERNEL_F32_FINALIZE - -dot_kernel_F1: - - ands I, N, #31 - ble dot_kernel_L999 - -dot_kernel_F10: - - KERNEL_F1 - - subs I, I, #1 - bne dot_kernel_F10 - - ret - -dot_kernel_S_BEGIN: - - INIT_S - - asr I, N, #2 - cmp I, xzr - ble dot_kernel_S1 - -dot_kernel_S4: - - KERNEL_S1 - KERNEL_S1 - KERNEL_S1 - KERNEL_S1 - - subs I, I, #1 - bne dot_kernel_S4 - -dot_kernel_S1: - - ands I, N, #3 - ble dot_kernel_L999 - -dot_kernel_S10: - - KERNEL_S1 - - subs I, I, #1 - bne dot_kernel_S10 - -dot_kernel_L999: - - ret - - EPILOGUE diff --git a/kernel/arm64/ddot_thunderx2t99.c b/kernel/arm64/ddot_thunderx2t99.c new file mode 100644 index 00000000..86026733 --- /dev/null +++ b/kernel/arm64/ddot_thunderx2t99.c @@ -0,0 +1,268 @@ +/*************************************************************************** +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 */ + +#define REG0 "xzr" +#define DOTF "d0" +#define TMPX "d16" +#define LD1VX "{v16.d}[0]" +#define TMPY "d24" +#define LD1VY "{v24.d}[0]" +#define SZ "8" + +#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 FLOAT ddot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) +{ + FLOAT dot = 0.0 ; + + if ( n < 0 ) return(dot); + + __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 d1, "REG0" \n" + " fmov d2, "REG0" \n" + " fmov d3, "REG0" \n" + " fmov d4, "REG0" \n" + " fmov d5, "REG0" \n" + " fmov d6, "REG0" \n" + " fmov d7, "REG0" \n" + + " cmp "N", xzr \n" + " ble 9f //dot_kernel_L999 \n" + + " cmp "INC_X", #1 \n" + " bne 5f //dot_kernel_S_BEGIN \n" + " cmp "INC_Y", #1 \n" + " bne 5f //dot_kernel_S_BEGIN \n" + + "1: //dot_kernel_F_BEGIN \n" + " asr "J", "N", #5 \n" + " cmp "J", xzr \n" + " beq 3f //dot_kernel_F1 \n" + + " .align 5 \n" + "2: //dot_kernel_F32 \n" + " ldp q16, q17, ["X"] \n" + " ldp q24, q25, ["Y"] \n" + " ldp q18, q19, ["X", #32] \n" + " ldp q26, q27, ["Y", #32] \n" + " fmla v0.2d, v16.2d, v24.2d \n" + " fmla v1.2d, v17.2d, v25.2d \n" + " ldp q20, q21, ["X", #64] \n" + " ldp q28, q29, ["Y", #64] \n" + " fmla v2.2d, v18.2d, v26.2d \n" + " fmla v3.2d, v19.2d, v27.2d \n" + " ldp q22, q23, ["X", #96] \n" + " ldp q30, q31, ["Y", #96] \n" + " add "Y", "Y", #128 \n" + " add "X", "X", #128 \n" + " fmla v4.2d, v20.2d, v28.2d \n" + " fmla v5.2d, v21.2d, v29.2d \n" + " PRFM PLDL1KEEP, ["X", #896] \n" + " PRFM PLDL1KEEP, ["Y", #896] \n" + " PRFM PLDL1KEEP, ["X", #896+64] \n" + " PRFM PLDL1KEEP, ["Y", #896+64] \n" + " fmla v6.2d, v22.2d, v30.2d \n" + " fmla v7.2d, v23.2d, v31.2d \n" + + " ldp q16, q17, ["X"] \n" + " ldp q24, q25, ["Y"] \n" + " ldp q18, q19, ["X", #32] \n" + " ldp q26, q27, ["Y", #32] \n" + " fmla v0.2d, v16.2d, v24.2d \n" + " fmla v1.2d, v17.2d, v25.2d \n" + " ldp q20, q21, ["X", #64] \n" + " ldp q28, q29, ["Y", #64] \n" + " fmla v2.2d, v18.2d, v26.2d \n" + " fmla v3.2d, v19.2d, v27.2d \n" + " ldp q22, q23, ["X", #96] \n" + " ldp q30, q31, ["Y", #96] \n" + " add "Y", "Y", #128 \n" + " add "X", "X", #128 \n" + " fmla v4.2d, v20.2d, v28.2d \n" + " fmla v5.2d, v21.2d, v29.2d \n" + " PRFM PLDL1KEEP, ["X", #896] \n" + " PRFM PLDL1KEEP, ["Y", #896] \n" + " PRFM PLDL1KEEP, ["X", #896+64] \n" + " PRFM PLDL1KEEP, ["Y", #896+64] \n" + " fmla v6.2d, v22.2d, v30.2d \n" + " fmla v7.2d, v23.2d, v31.2d \n" + + " subs "J", "J", #1 \n" + " bne 2b //dot_kernel_F32 \n" + + " fadd v0.2d, v0.2d, v1.2d \n" + " fadd v2.2d, v2.2d, v3.2d \n" + " fadd v4.2d, v4.2d, v5.2d \n" + " fadd v6.2d, v6.2d, v7.2d \n" + " fadd v0.2d, v0.2d, v2.2d \n" + " fadd v4.2d, v4.2d, v6.2d \n" + " fadd v0.2d, v0.2d, v4.2d \n" + " faddp "DOTF", v0.2d \n" + + "3: //dot_kernel_F1 \n" + " ands "J", "N", #31 \n" + " ble 9f //dot_kernel_L999 \n" + + "4: //dot_kernel_F10 \n" + " ldr "TMPX", ["X"] \n" + " ldr "TMPY", ["Y"] \n" + " add "X", "X", #"SZ" \n" + " add "Y", "Y", #"SZ" \n" + " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" + + " subs "J", "J", #1 \n" + " bne 4b //dot_kernel_F10 \n" + + " b 9f //dot_kernel_L999 \n" + + "5: //dot_kernel_S_BEGIN \n" + " lsl "INC_X", "INC_X", #3 \n" + " lsl "INC_Y", "INC_Y", #3 \n" + " asr "J", "N", #2 \n" + " cmp "J", xzr \n" + " ble 7f //dot_kernel_S1 \n" + + "6: //dot_kernel_S4: \n" + " ld1 "LD1VX", ["X"], "INC_X" \n" + " ld1 "LD1VY", ["Y"], "INC_Y" \n" + " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" + " ld1 "LD1VX", ["X"], "INC_X" \n" + " ld1 "LD1VY", ["Y"], "INC_Y" \n" + " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" + " ld1 "LD1VX", ["X"], "INC_X" \n" + " ld1 "LD1VY", ["Y"], "INC_Y" \n" + " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" + " ld1 "LD1VX", ["X"], "INC_X" \n" + " ld1 "LD1VY", ["Y"], "INC_Y" \n" + " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" + " subs "J", "J", #1 \n" + " bne 6b //dot_kernel_S4 \n" + + "7: //dot_kernel_S1: \n" + " ands "J", "N", #3 \n" + " ble 9f //dot_kernel_L999 \n" + + "8: //dot_kernel_S10 \n" + " ld1 "LD1VX", ["X"], "INC_X" \n" + " ld1 "LD1VY", ["Y"], "INC_Y" \n" + " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" + " subs "J", "J", #1 \n" + " bne 8b //dot_kernel_S10 \n" + + "9: //dot_kernel_L999 \n" + " fmov %[DOT_], "DOTF" \n" + : [DOT_] "=r" (dot) //%0 + : [N_] "r" (n), //%1 + [X_] "r" (x), //%2 + [INCX_] "r" (inc_x), //%3 + [Y_] "r" (y), //%4 + [INCY_] "r" (inc_y) //%5 + : "cc", + "memory", + "x0", "x1", "x2", "x3", "x4", "x5", + "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7" + ); + + return(dot); +} + +#if defined(SMP) +static int ddot_thread_function(BLASLONG n, BLASLONG dummy0, + BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y, + BLASLONG inc_y, FLOAT *result, BLASLONG dummy3) +{ + *result = ddot_compute(n, x, inc_x, y, inc_y); + + return 0; +} +#endif + +FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) +{ +#if defined(SMP) + int nthreads; + FLOAT dummy_alpha; +#endif + FLOAT dot = 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) { + dot = ddot_compute(n, x, inc_x, y, inc_y); + } else { + int mode, i; + char result[MAX_CPU_NUMBER * sizeof(double) * 2]; + FLOAT *ptr; + + mode = BLAS_DOUBLE | BLAS_REAL; + + blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, + x, inc_x, y, inc_y, result, 0, + ( void *)ddot_thread_function, nthreads); + + ptr = (FLOAT *)result; + for (i = 0; i < nthreads; i++) { + dot = dot + (*ptr); + ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2); + } + } +#else + dot = ddot_compute(n, x, inc_x, y, inc_y); +#endif + + return dot; +} diff --git a/kernel/arm64/ddot_thunderx2t99_threaded.c b/kernel/arm64/ddot_thunderx2t99_threaded.c deleted file mode 100644 index 36a8d6c2..00000000 --- a/kernel/arm64/ddot_thunderx2t99_threaded.c +++ /dev/null @@ -1,258 +0,0 @@ -/*************************************************************************** -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 */ - -#define REG0 "xzr" -#define DOTF "d0" -#define TMPX "d16" -#define LD1VX "{v16.d}[0]" -#define TMPY "d24" -#define LD1VY "{v24.d}[0]" -#define SZ "8" - -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); - - -static FLOAT ddot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) -{ - FLOAT dot = 0.0 ; - - if ( n < 0 ) return(dot); - - __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 d1, "REG0" \n" - " fmov d2, "REG0" \n" - " fmov d3, "REG0" \n" - " fmov d4, "REG0" \n" - " fmov d5, "REG0" \n" - " fmov d6, "REG0" \n" - " fmov d7, "REG0" \n" - - " cmp "N", xzr \n" - " ble 9f //dot_kernel_L999 \n" - - " cmp "INC_X", #1 \n" - " bne 5f //dot_kernel_S_BEGIN \n" - " cmp "INC_Y", #1 \n" - " bne 5f //dot_kernel_S_BEGIN \n" - - "1: //dot_kernel_F_BEGIN \n" - " asr "J", "N", #5 \n" - " cmp "J", xzr \n" - " beq 3f //dot_kernel_F1 \n" - - " .align 5 \n" - "2: //dot_kernel_F32 \n" - " ldp q16, q17, ["X"] \n" - " ldp q24, q25, ["Y"] \n" - " ldp q18, q19, ["X", #32] \n" - " ldp q26, q27, ["Y", #32] \n" - " fmla v0.2d, v16.2d, v24.2d \n" - " fmla v1.2d, v17.2d, v25.2d \n" - " ldp q20, q21, ["X", #64] \n" - " ldp q28, q29, ["Y", #64] \n" - " fmla v2.2d, v18.2d, v26.2d \n" - " fmla v3.2d, v19.2d, v27.2d \n" - " ldp q22, q23, ["X", #96] \n" - " ldp q30, q31, ["Y", #96] \n" - " add "Y", "Y", #128 \n" - " add "X", "X", #128 \n" - " fmla v4.2d, v20.2d, v28.2d \n" - " fmla v5.2d, v21.2d, v29.2d \n" - " PRFM PLDL1KEEP, ["X", #896] \n" - " PRFM PLDL1KEEP, ["Y", #896] \n" - " PRFM PLDL1KEEP, ["X", #896+64] \n" - " PRFM PLDL1KEEP, ["Y", #896+64] \n" - " fmla v6.2d, v22.2d, v30.2d \n" - " fmla v7.2d, v23.2d, v31.2d \n" - - " ldp q16, q17, ["X"] \n" - " ldp q24, q25, ["Y"] \n" - " ldp q18, q19, ["X", #32] \n" - " ldp q26, q27, ["Y", #32] \n" - " fmla v0.2d, v16.2d, v24.2d \n" - " fmla v1.2d, v17.2d, v25.2d \n" - " ldp q20, q21, ["X", #64] \n" - " ldp q28, q29, ["Y", #64] \n" - " fmla v2.2d, v18.2d, v26.2d \n" - " fmla v3.2d, v19.2d, v27.2d \n" - " ldp q22, q23, ["X", #96] \n" - " ldp q30, q31, ["Y", #96] \n" - " add "Y", "Y", #128 \n" - " add "X", "X", #128 \n" - " fmla v4.2d, v20.2d, v28.2d \n" - " fmla v5.2d, v21.2d, v29.2d \n" - " PRFM PLDL1KEEP, ["X", #896] \n" - " PRFM PLDL1KEEP, ["Y", #896] \n" - " PRFM PLDL1KEEP, ["X", #896+64] \n" - " PRFM PLDL1KEEP, ["Y", #896+64] \n" - " fmla v6.2d, v22.2d, v30.2d \n" - " fmla v7.2d, v23.2d, v31.2d \n" - - " subs "J", "J", #1 \n" - " bne 2b //dot_kernel_F32 \n" - - " fadd v0.2d, v0.2d, v1.2d \n" - " fadd v2.2d, v2.2d, v3.2d \n" - " fadd v4.2d, v4.2d, v5.2d \n" - " fadd v6.2d, v6.2d, v7.2d \n" - " fadd v0.2d, v0.2d, v2.2d \n" - " fadd v4.2d, v4.2d, v6.2d \n" - " fadd v0.2d, v0.2d, v4.2d \n" - " faddp "DOTF", v0.2d \n" - - "3: //dot_kernel_F1 \n" - " ands "J", "N", #31 \n" - " ble 9f //dot_kernel_L999 \n" - - "4: //dot_kernel_F10 \n" - " ldr "TMPX", ["X"] \n" - " ldr "TMPY", ["Y"] \n" - " add "X", "X", #"SZ" \n" - " add "Y", "Y", #"SZ" \n" - " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" - - " subs "J", "J", #1 \n" - " bne 4b //dot_kernel_F10 \n" - - " b 9f //dot_kernel_L999 \n" - - "5: //dot_kernel_S_BEGIN \n" - " lsl "INC_X", "INC_X", #3 \n" - " lsl "INC_Y", "INC_Y", #3 \n" - " asr "J", "N", #2 \n" - " cmp "J", xzr \n" - " ble 7f //dot_kernel_S1 \n" - - "6: //dot_kernel_S4: \n" - " ld1 "LD1VX", ["X"], "INC_X" \n" - " ld1 "LD1VY", ["Y"], "INC_Y" \n" - " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" - " ld1 "LD1VX", ["X"], "INC_X" \n" - " ld1 "LD1VY", ["Y"], "INC_Y" \n" - " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" - " ld1 "LD1VX", ["X"], "INC_X" \n" - " ld1 "LD1VY", ["Y"], "INC_Y" \n" - " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" - " ld1 "LD1VX", ["X"], "INC_X" \n" - " ld1 "LD1VY", ["Y"], "INC_Y" \n" - " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" - " subs "J", "J", #1 \n" - " bne 6b //dot_kernel_S4 \n" - - "7: //dot_kernel_S1: \n" - " ands "J", "N", #3 \n" - " ble 9f //dot_kernel_L999 \n" - - "8: //dot_kernel_S10 \n" - " ld1 "LD1VX", ["X"], "INC_X" \n" - " ld1 "LD1VY", ["Y"], "INC_Y" \n" - " fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" - " subs "J", "J", #1 \n" - " bne 8b //dot_kernel_S10 \n" - - "9: //dot_kernel_L999 \n" - " fmov %[DOT_], "DOTF" \n" - : [DOT_] "=r" (dot) //%0 - : [N_] "r" (n), //%1 - [X_] "r" (x), //%2 - [INCX_] "r" (inc_x), //%3 - [Y_] "r" (y), //%4 - [INCY_] "r" (inc_y) //%5 - : "cc", - "memory", - "x0", "x1", "x2", "x3", "x4", "x5", - "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7" - ); - - return(dot); -} - -static int ddot_thread_function(BLASLONG n, BLASLONG dummy0, - BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y, - BLASLONG inc_y, FLOAT *result, BLASLONG dummy3) -{ - *result = ddot_compute(n, x, inc_x, y, inc_y); - - return 0; -} - -FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) -{ - int nthreads; - FLOAT dot = 0.0; - FLOAT dummy_alpha; - - nthreads = num_cpu_avail(1); - - if (inc_x == 0 || inc_y == 0) - nthreads = 1; - - if (n <= 10000) - nthreads = 1; - - if (nthreads == 1) { - dot = ddot_compute(n, x, inc_x, y, inc_y); - } else { - int mode, i; - char result[MAX_CPU_NUMBER * sizeof(double) * 2]; - FLOAT *ptr; - - mode = BLAS_DOUBLE | BLAS_REAL; - - blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, - x, inc_x, y, inc_y, result, 0, - ( void *)ddot_thread_function, nthreads); - - ptr = (FLOAT *)result; - for (i = 0; i < nthreads; i++) { - dot = dot + (*ptr); - ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2); - } - } - - return dot; -} diff --git a/kernel/arm64/nrm2_thunderx2t99_threaded.c b/kernel/arm64/nrm2_thunderx2t99_threaded.c deleted file mode 100644 index d810c671..00000000 --- a/kernel/arm64/nrm2_thunderx2t99_threaded.c +++ /dev/null @@ -1,249 +0,0 @@ -/*************************************************************************** -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 - -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); - -#if !defined(DOUBLE) -#define N "x0" /* vector length */ -#define X "x1" /* X vector address */ -#define INC_X "x2" /* X stride */ -#define I "x5" /* loop variable */ - -#define TMPF "s16" -#define TMPFD "d17" -#define SSQD "d0" - -#define KERNEL_F1 \ - "ldr "TMPF", ["X"], #4 \n" \ - "fcvt "TMPFD", "TMPF" \n" \ - "fmadd "SSQD", "TMPFD", "TMPFD", "SSQD"\n" - -#define KERNEL_F32 \ - "ldur q16, ["X"] \n" \ - "ldur q18, ["X", #16] \n" \ - "ldur q20, ["X", #32] \n" \ - "ldur q22, ["X", #48] \n" \ - "ldur q24, ["X", #64] \n" \ - "ldur q26, ["X", #80] \n" \ - "ldur q28, ["X", #96] \n" \ - "ldur q30, ["X", #112] \n" \ - "add "X", "X", #128 \n" \ - "fcvtl2 v17.2d, v16.4s \n" \ - "fcvtl v16.2d, v16.2s \n" \ - "fcvtl2 v19.2d, v18.4s \n" \ - "fcvtl v18.2d, v18.2s \n" \ - "fcvtl2 v21.2d, v20.4s \n" \ - "fcvtl v20.2d, v20.2s \n" \ - "fcvtl2 v23.2d, v22.4s \n" \ - "fcvtl v22.2d, v22.2s \n" \ - "fcvtl2 v25.2d, v24.4s \n" \ - "fcvtl v24.2d, v24.2s \n" \ - "fcvtl2 v27.2d, v26.4s \n" \ - "fcvtl v26.2d, v26.2s \n" \ - "fcvtl2 v29.2d, v28.4s \n" \ - "fcvtl v28.2d, v28.2s \n" \ - "fcvtl2 v31.2d, v30.4s \n" \ - "fcvtl v30.2d, v30.2s \n" \ - "fmla v0.2d, v16.2d, v16.2d \n" \ - "fmla v1.2d, v17.2d, v17.2d \n" \ - "fmla v2.2d, v18.2d, v18.2d \n" \ - "fmla v3.2d, v19.2d, v19.2d \n" \ - "fmla v4.2d, v20.2d, v20.2d \n" \ - "fmla v5.2d, v21.2d, v21.2d \n" \ - "fmla v6.2d, v22.2d, v22.2d \n" \ - "fmla v7.2d, v23.2d, v23.2d \n" \ - "fmla v0.2d, v24.2d, v24.2d \n" \ - "fmla v1.2d, v25.2d, v25.2d \n" \ - "fmla v2.2d, v26.2d, v26.2d \n" \ - "fmla v3.2d, v27.2d, v27.2d \n" \ - "fmla v4.2d, v28.2d, v28.2d \n" \ - "fmla v5.2d, v29.2d, v29.2d \n" \ - "fmla v6.2d, v30.2d, v30.2d \n" \ - "fmla v7.2d, v31.2d, v31.2d \n" \ - "prfm PLDL1KEEP, ["X", #1024] \n" \ - "prfm PLDL1KEEP, ["X", #1024+64] \n" - -#define KERNEL_F32_FINALIZE \ - "fadd v0.2d, v0.2d, v1.2d \n" \ - "fadd v2.2d, v2.2d, v3.2d \n" \ - "fadd v4.2d, v4.2d, v5.2d \n" \ - "fadd v6.2d, v6.2d, v7.2d \n" \ - "fadd v0.2d, v0.2d, v2.2d \n" \ - "fadd v4.2d, v4.2d, v6.2d \n" \ - "fadd v0.2d, v0.2d, v4.2d \n" \ - "faddp "SSQD", v0.2d \n" - -#define KERNEL_S1 \ - "ldr "TMPF", ["X"] \n" \ - "add "X", "X", "INC_X" \n" \ - "fcvt "TMPFD", "TMPF" \n" \ - "fmadd "SSQD", "TMPFD", "TMPFD", "SSQD"\n" - - -static double nrm2_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x) -{ - double ret = 0.0 ; - - if (n <= 0) return ret; - - __asm__ __volatile__ ( - " mov "N", %[N_] \n" - " mov "X", %[X_] \n" - " mov "INC_X", %[INCX_] \n" - " fmov "SSQD", xzr \n" - " fmov d1, xzr \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 8f //nrm2_kernel_L999 \n" - " cmp "INC_X", xzr \n" - " ble 8f //nrm2_kernel_L999 \n" - " cmp "INC_X", #1 \n" - " bne 5f // nrm2_kernel_S_BEGIN \n" - - "1: //nrm2_kernel_F_BEGIN: \n" - " asr "I", "N", #6 \n" - " cmp "I", xzr \n" - " beq 5f // nrm2_kernel_S_BEGIN \n" - - " .align 5 \n" - "2: //nrm2_kernel_F64: \n" - " "KERNEL_F32" \n" - " "KERNEL_F32" \n" - " subs "I", "I", #1 \n" - " bne 2b //nrm2_kernel_F64 \n" - " "KERNEL_F32_FINALIZE" \n" - - "3: // nrm2_kernel_F1: \n" - " ands "I", "N", #63 \n" - " ble 8f //nrm2_kernel_L999 \n" - - "4: // nrm2_kernel_F10: \n" - " "KERNEL_F1" \n" - " subs "I", "I", #1 \n" - " bne 4b //nrm2_kernel_F10 \n" - " b 8f //nrm2_kernel_L999 \n" - - "5: // nrm2_kernel_S_BEGIN: \n" - " lsl "INC_X", "INC_X", #2 \n" - " asr "I", "N", #2 \n" - " cmp "I", xzr \n" - " ble 6f //nrm2_kernel_S1 \n" - - "4: //nrm2_kernel_S4: \n" - " "KERNEL_S1" \n" - " "KERNEL_S1" \n" - " "KERNEL_S1" \n" - " "KERNEL_S1" \n" - " subs "I", "I", #1 \n" - " bne 4b //nrm2_kernel_S4 \n" - - "6: //nrm2_kernel_S1: \n" - " ands "I", "N", #3 \n" - " ble 8f //nrm2_kernel_L999 \n" - - "7: //nrm2_kernel_S10: \n" - " "KERNEL_S1" \n" - " subs "I", "I", #1 \n" - " bne 7b //nrm2_kernel_S10 \n" - - "8: //nrm2_kernel_L999: \n" - " fmov %[RET_], "SSQD" \n" - - : [RET_] "=r" (ret) //%0 - : [N_] "r" (n), //%1 - [X_] "r" (x), //%2 - [INCX_] "r" (inc_x) //%3 - : "cc", - "memory", - "x0", "x1", "x2", "x3", "x4", "x5", - "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7" - ); - - return ret; -} -#else //!defined(DOUBLE) -#endif - -static int nrm2_thread_function(BLASLONG n, BLASLONG dummy0, - BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *dummy3, - BLASLONG dummy4, FLOAT *result, BLASLONG dummy5) -{ - *(double *)result = nrm2_compute(n, x, inc_x); - - return 0; -} - -FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) -{ - int nthreads; - FLOAT nrm2 = 0.0; - double nrm2_double = 0.0; - FLOAT dummy_alpha; - - if (n <= 0 || inc_x <= 0) return 0.0; - if (n == 1) return fabs(x[0]); - - nthreads = num_cpu_avail(1); - - if (n <= 10000) - nthreads = 1; - - if (nthreads == 1) { - nrm2_double = nrm2_compute(n, x, inc_x); - } else { - int mode, i; - char result[MAX_CPU_NUMBER * sizeof(double) * 2]; - double *ptr; - - mode = BLAS_SINGLE | BLAS_REAL; - - blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, - x, inc_x, NULL, 0, result, 0, - ( void *)nrm2_thread_function, nthreads); - - ptr = (double *)result; - for (i = 0; i < nthreads; i++) { - nrm2_double = nrm2_double + (*ptr) * (*ptr); - ptr = (double *)(((char *)ptr) + sizeof(double) * 2); - } - } - nrm2 = sqrt(nrm2_double); - - return nrm2; -} diff --git a/kernel/arm64/snrm2_thunderx2t99.S b/kernel/arm64/snrm2_thunderx2t99.S deleted file mode 100644 index d69441d0..00000000 --- a/kernel/arm64/snrm2_thunderx2t99.S +++ /dev/null @@ -1,228 +0,0 @@ -/******************************************************************************* -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. -*******************************************************************************/ - -#define ASSEMBLER -#include "common.h" - -#define N x0 /* vector length */ -#define X x1 /* X vector address */ -#define INC_X x2 /* X stride */ -#define I x5 /* loop variable */ - -/******************************************************************************* -* Macro definitions -*******************************************************************************/ - -#define TMPF s16 -#define TMPFD d17 -#define SSQ s0 -#define SSQD d0 -#define TMPVF {v16.s}[0] -#define TMPVFD {v17.s}[0] -#define SZ 4 - -/******************************************************************************/ - -.macro INIT - fmov SSQD, xzr - fmov d1, xzr - fmov d2, xzr - fmov d3, xzr - fmov d4, xzr - fmov d5, xzr - fmov d6, xzr - fmov d7, xzr -.endm - -.macro KERNEL_F1 - ldr TMPF, [X], #SZ - fcvt TMPFD, TMPF - fmadd SSQD, TMPFD, TMPFD, SSQD -.endm - -.macro KERNEL_F32 - ldur q16, [X] - ldur q18, [X, #16] - ldur q20, [X, #32] - ldur q22, [X, #48] - ldur q24, [X, #64] - ldur q26, [X, #80] - ldur q28, [X, #96] - ldur q30, [X, #112] - - add X, X, #128 - - fcvtl2 v17.2d, v16.4s - fcvtl v16.2d, v16.2s - fcvtl2 v19.2d, v18.4s - fcvtl v18.2d, v18.2s - fcvtl2 v21.2d, v20.4s - fcvtl v20.2d, v20.2s - fcvtl2 v23.2d, v22.4s - fcvtl v22.2d, v22.2s - fcvtl2 v25.2d, v24.4s - fcvtl v24.2d, v24.2s - fcvtl2 v27.2d, v26.4s - fcvtl v26.2d, v26.2s - fcvtl2 v29.2d, v28.4s - fcvtl v28.2d, v28.2s - fcvtl2 v31.2d, v30.4s - fcvtl v30.2d, v30.2s - - fmla v0.2d, v16.2d, v16.2d - fmla v1.2d, v17.2d, v17.2d - fmla v2.2d, v18.2d, v18.2d - fmla v3.2d, v19.2d, v19.2d - fmla v4.2d, v20.2d, v20.2d - fmla v5.2d, v21.2d, v21.2d - fmla v6.2d, v22.2d, v22.2d - fmla v7.2d, v23.2d, v23.2d - - fmla v0.2d, v24.2d, v24.2d - fmla v1.2d, v25.2d, v25.2d - fmla v2.2d, v26.2d, v26.2d - fmla v3.2d, v27.2d, v27.2d - fmla v4.2d, v28.2d, v28.2d - fmla v5.2d, v29.2d, v29.2d - fmla v6.2d, v30.2d, v30.2d - fmla v7.2d, v31.2d, v31.2d - - prfm PLDL1KEEP, [X, #1024] - prfm PLDL1KEEP, [X, #1024+64] -.endm - -.macro KERNEL_F32_FINALIZE - fadd v0.2d, v0.2d, v1.2d - fadd v2.2d, v2.2d, v3.2d - fadd v4.2d, v4.2d, v5.2d - fadd v6.2d, v6.2d, v7.2d - - fadd v0.2d, v0.2d, v2.2d - fadd v4.2d, v4.2d, v6.2d - - fadd v0.2d, v0.2d, v4.2d - faddp SSQD, v0.2d -.endm - -.macro INIT_S - lsl INC_X, INC_X, #2 -.endm - -.macro KERNEL_S1 - ldr TMPF, [X] - add X, X, INC_X - fcvt TMPFD, TMPF - fmadd SSQD, TMPFD, TMPFD, SSQD -.endm - -/******************************************************************************* -* End of macro definitions -*******************************************************************************/ - - PROLOGUE - - INIT - - cmp N, xzr - ble nrm2_kernel_zero - cmp INC_X, xzr - ble nrm2_kernel_zero - cmp INC_X, #1 - bne nrm2_kernel_S_BEGIN - -nrm2_kernel_F_BEGIN: - - asr I, N, #6 - cmp I, xzr - beq nrm2_kernel_S_BEGIN - - .align 5 -nrm2_kernel_F64: - - KERNEL_F32 - KERNEL_F32 - - subs I, I, #1 - bne nrm2_kernel_F64 - - KERNEL_F32_FINALIZE - -nrm2_kernel_F1: - - ands I, N, #63 - ble nrm2_kernel_L999 - -nrm2_kernel_F10: - - KERNEL_F1 - - subs I, I, #1 - bne nrm2_kernel_F10 - - b nrm2_kernel_L999 - -nrm2_kernel_S_BEGIN: - - INIT_S - - asr I, N, #2 - cmp I, xzr - ble nrm2_kernel_S1 - -nrm2_kernel_S4: - - KERNEL_S1 - KERNEL_S1 - KERNEL_S1 - KERNEL_S1 - - subs I, I, #1 - bne nrm2_kernel_S4 - -nrm2_kernel_S1: - - ands I, N, #3 - ble nrm2_kernel_L999 - -nrm2_kernel_S10: - - KERNEL_S1 - - subs I, I, #1 - bne nrm2_kernel_S10 - -nrm2_kernel_L999: - fsqrt SSQD, SSQD - fcvt SSQ, SSQD - ret - -nrm2_kernel_zero: - fmov SSQ, wzr - - ret - - EPILOGUE diff --git a/kernel/arm64/snrm2_thunderx2t99.c b/kernel/arm64/snrm2_thunderx2t99.c new file mode 100644 index 00000000..90595582 --- /dev/null +++ b/kernel/arm64/snrm2_thunderx2t99.c @@ -0,0 +1,256 @@ +/*************************************************************************** +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 + +#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 + +#define N "x0" /* vector length */ +#define X "x1" /* X vector address */ +#define INC_X "x2" /* X stride */ +#define I "x5" /* loop variable */ + +#define TMPF "s16" +#define TMPFD "d17" +#define SSQD "d0" + +#define KERNEL_F1 \ + "ldr "TMPF", ["X"], #4 \n" \ + "fcvt "TMPFD", "TMPF" \n" \ + "fmadd "SSQD", "TMPFD", "TMPFD", "SSQD"\n" + +#define KERNEL_F32 \ + "ldur q16, ["X"] \n" \ + "ldur q18, ["X", #16] \n" \ + "ldur q20, ["X", #32] \n" \ + "ldur q22, ["X", #48] \n" \ + "ldur q24, ["X", #64] \n" \ + "ldur q26, ["X", #80] \n" \ + "ldur q28, ["X", #96] \n" \ + "ldur q30, ["X", #112] \n" \ + "add "X", "X", #128 \n" \ + "fcvtl2 v17.2d, v16.4s \n" \ + "fcvtl v16.2d, v16.2s \n" \ + "fcvtl2 v19.2d, v18.4s \n" \ + "fcvtl v18.2d, v18.2s \n" \ + "fcvtl2 v21.2d, v20.4s \n" \ + "fcvtl v20.2d, v20.2s \n" \ + "fcvtl2 v23.2d, v22.4s \n" \ + "fcvtl v22.2d, v22.2s \n" \ + "fcvtl2 v25.2d, v24.4s \n" \ + "fcvtl v24.2d, v24.2s \n" \ + "fcvtl2 v27.2d, v26.4s \n" \ + "fcvtl v26.2d, v26.2s \n" \ + "fcvtl2 v29.2d, v28.4s \n" \ + "fcvtl v28.2d, v28.2s \n" \ + "fcvtl2 v31.2d, v30.4s \n" \ + "fcvtl v30.2d, v30.2s \n" \ + "fmla v0.2d, v16.2d, v16.2d \n" \ + "fmla v1.2d, v17.2d, v17.2d \n" \ + "fmla v2.2d, v18.2d, v18.2d \n" \ + "fmla v3.2d, v19.2d, v19.2d \n" \ + "fmla v4.2d, v20.2d, v20.2d \n" \ + "fmla v5.2d, v21.2d, v21.2d \n" \ + "fmla v6.2d, v22.2d, v22.2d \n" \ + "fmla v7.2d, v23.2d, v23.2d \n" \ + "fmla v0.2d, v24.2d, v24.2d \n" \ + "fmla v1.2d, v25.2d, v25.2d \n" \ + "fmla v2.2d, v26.2d, v26.2d \n" \ + "fmla v3.2d, v27.2d, v27.2d \n" \ + "fmla v4.2d, v28.2d, v28.2d \n" \ + "fmla v5.2d, v29.2d, v29.2d \n" \ + "fmla v6.2d, v30.2d, v30.2d \n" \ + "fmla v7.2d, v31.2d, v31.2d \n" \ + "prfm PLDL1KEEP, ["X", #1024] \n" \ + "prfm PLDL1KEEP, ["X", #1024+64] \n" + +#define KERNEL_F32_FINALIZE \ + "fadd v0.2d, v0.2d, v1.2d \n" \ + "fadd v2.2d, v2.2d, v3.2d \n" \ + "fadd v4.2d, v4.2d, v5.2d \n" \ + "fadd v6.2d, v6.2d, v7.2d \n" \ + "fadd v0.2d, v0.2d, v2.2d \n" \ + "fadd v4.2d, v4.2d, v6.2d \n" \ + "fadd v0.2d, v0.2d, v4.2d \n" \ + "faddp "SSQD", v0.2d \n" + +#define KERNEL_S1 \ + "ldr "TMPF", ["X"] \n" \ + "add "X", "X", "INC_X" \n" \ + "fcvt "TMPFD", "TMPF" \n" \ + "fmadd "SSQD", "TMPFD", "TMPFD", "SSQD"\n" + + +static double nrm2_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x) +{ + double ret = 0.0 ; + + if (n <= 0) return ret; + + __asm__ __volatile__ ( + " mov "N", %[N_] \n" + " mov "X", %[X_] \n" + " mov "INC_X", %[INCX_] \n" + " fmov "SSQD", xzr \n" + " fmov d1, xzr \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 8f //nrm2_kernel_L999 \n" + " cmp "INC_X", xzr \n" + " ble 8f //nrm2_kernel_L999 \n" + " cmp "INC_X", #1 \n" + " bne 5f // nrm2_kernel_S_BEGIN \n" + + "1: //nrm2_kernel_F_BEGIN: \n" + " asr "I", "N", #6 \n" + " cmp "I", xzr \n" + " beq 5f // nrm2_kernel_S_BEGIN \n" + + " .align 5 \n" + "2: //nrm2_kernel_F64: \n" + " "KERNEL_F32" \n" + " "KERNEL_F32" \n" + " subs "I", "I", #1 \n" + " bne 2b //nrm2_kernel_F64 \n" + " "KERNEL_F32_FINALIZE" \n" + + "3: // nrm2_kernel_F1: \n" + " ands "I", "N", #63 \n" + " ble 8f //nrm2_kernel_L999 \n" + + "4: // nrm2_kernel_F10: \n" + " "KERNEL_F1" \n" + " subs "I", "I", #1 \n" + " bne 4b //nrm2_kernel_F10 \n" + " b 8f //nrm2_kernel_L999 \n" + + "5: // nrm2_kernel_S_BEGIN: \n" + " lsl "INC_X", "INC_X", #2 \n" + " asr "I", "N", #2 \n" + " cmp "I", xzr \n" + " ble 6f //nrm2_kernel_S1 \n" + + "4: //nrm2_kernel_S4: \n" + " "KERNEL_S1" \n" + " "KERNEL_S1" \n" + " "KERNEL_S1" \n" + " "KERNEL_S1" \n" + " subs "I", "I", #1 \n" + " bne 4b //nrm2_kernel_S4 \n" + + "6: //nrm2_kernel_S1: \n" + " ands "I", "N", #3 \n" + " ble 8f //nrm2_kernel_L999 \n" + + "7: //nrm2_kernel_S10: \n" + " "KERNEL_S1" \n" + " subs "I", "I", #1 \n" + " bne 7b //nrm2_kernel_S10 \n" + + "8: //nrm2_kernel_L999: \n" + " fmov %[RET_], "SSQD" \n" + + : [RET_] "=r" (ret) //%0 + : [N_] "r" (n), //%1 + [X_] "r" (x), //%2 + [INCX_] "r" (inc_x) //%3 + : "cc", + "memory", + "x0", "x1", "x2", "x3", "x4", "x5", + "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7" + ); + + return ret; +} + +#if defined(SMP) +static int nrm2_thread_function(BLASLONG n, BLASLONG dummy0, + BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *dummy3, + BLASLONG dummy4, FLOAT *result, BLASLONG dummy5) +{ + *(double *)result = nrm2_compute(n, x, inc_x); + + return 0; +} +#endif + +FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) +{ +#if defined(SMP) + int nthreads; + FLOAT dummy_alpha; +#endif + FLOAT nrm2 = 0.0; + double nrm2_double = 0.0; + + if (n <= 0 || inc_x <= 0) return 0.0; + if (n == 1) return fabs(x[0]); + +#if defined(SMP) + nthreads = num_cpu_avail(1); + + if (n <= 10000) + nthreads = 1; + + if (nthreads == 1) { + nrm2_double = nrm2_compute(n, x, inc_x); + } else { + int mode, i; + char result[MAX_CPU_NUMBER * sizeof(double) * 2]; + double *ptr; + + mode = BLAS_SINGLE | BLAS_REAL; + + blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha, + x, inc_x, NULL, 0, result, 0, + ( void *)nrm2_thread_function, nthreads); + + ptr = (double *)result; + for (i = 0; i < nthreads; i++) { + nrm2_double = nrm2_double + (*ptr) * (*ptr); + ptr = (double *)(((char *)ptr) + sizeof(double) * 2); + } + } +#else + nrm2_double = nrm2_compute(n, x, inc_x); +#endif + nrm2 = sqrt(nrm2_double); + + return nrm2; +}