OSchip
/
OpenBLAS
Not watched
1
0
Fork 0
Code
Releases 66 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
Browse Source

Protect align directives that are currently problematic with LLVM on WoA

pull/5076/head
Martin Kroeker GitHub 1 year ago
parent
eba7338484
commit
dda211b20e
No known key found for this signature in database GPG Key ID: B5690EEEBB952194
5 changed files with 1351 additions and 1345 deletions
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. +217
    -216
      kernel/arm64/copy_thunderx2t99.c
  2. +262
    -260
      kernel/arm64/dasum_thunderx2t99.c
  3. +346
    -345
      kernel/arm64/dot_kernel_asimd.c
  4. +263
    -262
      kernel/arm64/sasum_thunderx2t99.c
  5. +263
    -262
      kernel/arm64/zasum_thunderx2t99.c

+ 217
- 216
kernel/arm64/copy_thunderx2t99.c View File

@@ -1,216 +1,217 @@
/***************************************************************************
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 <arm_neon.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 J "x5" /* loop variable */

/*******************************************************************************
* Macro definitions
*******************************************************************************/
#if !defined(COMPLEX)
#if !defined(DOUBLE)
#define TMPF "s0"
#define INC_SHIFT "2"
#define N_DIV_SHIFT "2"
#define N_REM_MASK "3"
#else
#define TMPF "d0"
#define INC_SHIFT "3"
#define N_DIV_SHIFT "1"
#define N_REM_MASK "1"
#endif
#else
#if !defined(DOUBLE)
#define TMPF "d0"
#define INC_SHIFT "3"
#define N_DIV_SHIFT "1"
#define N_REM_MASK "1"
#else
#define TMPF "q0"
#define INC_SHIFT "4"
#define N_DIV_SHIFT "0"
#define N_REM_MASK "0"
#endif
#endif

#define KERNEL_F1 \
"ldr "TMPF", ["X"] \n" \
"add "X", "X", "INC_X" \n" \
"str "TMPF", ["Y"] \n" \
"add "Y", "Y", "INC_Y" \n"

#define KERNEL_F \
"ldr q0, ["X"], #16 \n" \
"str q0, ["Y"], #16 \n"

#define INIT \
"lsl "INC_X", "INC_X", #"INC_SHIFT" \n" \
"lsl "INC_Y", "INC_Y", #"INC_SHIFT" \n"


static int do_copy(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{
if ( n < 0 ) return 0;

__asm__ __volatile__ (
" mov "N", %[N_] \n"
" mov "X", %[X_] \n"
" mov "INC_X", %[INCX_] \n"
" mov "Y", %[Y_] \n"
" mov "INC_Y", %[INCY_] \n"
" cmp "N", xzr \n"
" ble 8f //copy_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne 4f //copy_kernel_S_BEGIN \n"
" cmp "INC_Y", #1 \n"
" bne 4f //copy_kernel_S_BEGIN \n"

"// .Lcopy_kernel_F_BEGIN: \n"
" "INIT" \n"
" asr "J", "N", #"N_DIV_SHIFT" \n"
" cmp "J", xzr \n"
" beq 2f //copy_kernel_F1 \n"
" .align 5 \n"

"1: //copy_kernel_F: \n"
" "KERNEL_F" \n"
" subs "J", "J", #1 \n"
" bne 1b //copy_kernel_F \n"

"2: //copy_kernel_F1: \n"
#if defined(COMPLEX) && defined(DOUBLE)
" b 8f //copy_kernel_L999 \n"
#else
" ands "J", "N", #"N_REM_MASK" \n"
" ble 8f //copy_kernel_L999 \n"
#endif

"3: //copy_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 3b //copy_kernel_F10 \n"
" b 8f //copy_kernel_L999 \n"

"4: //copy_kernel_S_BEGIN: \n"
" "INIT" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble 6f //copy_kernel_S1 \n"

"5: //copy_kernel_S4: \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 5b //copy_kernel_S4 \n"

"6: //copy_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble 8f //copy_kernel_L999 \n"

"7: //copy_kernel_S10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 7b //copy_kernel_S10 \n"

"8: //copy_kernel_L999: \n"

:
: [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"
);

return 0;
}

#if defined(SMP)
static int copy_thread_function(BLASLONG n, BLASLONG dummy0,
BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y,
BLASLONG inc_y, FLOAT *dummy3, BLASLONG dummy4)
{
do_copy(n, x, inc_x, y, inc_y);

return 0;
}
#endif

int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{
#if defined(SMP)
int nthreads;
FLOAT dummy_alpha;
#endif

if (n <= 0) return 0;

#if defined(SMP)
if (inc_x == 0 || n <= 10000)
nthreads = 1;
else
nthreads = num_cpu_avail(1);

if (nthreads == 1) {
do_copy(n, x, inc_x, y, inc_y);
} else {
int mode = 0;

#if !defined(COMPLEX)
mode = BLAS_REAL;
#else
mode = BLAS_COMPLEX;
#endif
#if !defined(DOUBLE)
mode |= BLAS_SINGLE;
#else
mode |= BLAS_DOUBLE;
#endif

blas_level1_thread(mode, n, 0, 0, &dummy_alpha,
x, inc_x, y, inc_y, NULL, 0,
( void *)copy_thread_function, nthreads);
}
#else
do_copy(n, x, inc_x, y, inc_y);
#endif

return 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 <arm_neon.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 J "x5" /* loop variable */
/*******************************************************************************
* Macro definitions
*******************************************************************************/
#if !defined(COMPLEX)
#if !defined(DOUBLE)
#define TMPF "s0"
#define INC_SHIFT "2"
#define N_DIV_SHIFT "2"
#define N_REM_MASK "3"
#else
#define TMPF "d0"
#define INC_SHIFT "3"
#define N_DIV_SHIFT "1"
#define N_REM_MASK "1"
#endif
#else
#if !defined(DOUBLE)
#define TMPF "d0"
#define INC_SHIFT "3"
#define N_DIV_SHIFT "1"
#define N_REM_MASK "1"
#else
#define TMPF "q0"
#define INC_SHIFT "4"
#define N_DIV_SHIFT "0"
#define N_REM_MASK "0"
#endif
#endif
#define KERNEL_F1 \
"ldr "TMPF", ["X"] \n" \
"add "X", "X", "INC_X" \n" \
"str "TMPF", ["Y"] \n" \
"add "Y", "Y", "INC_Y" \n"
#define KERNEL_F \
"ldr q0, ["X"], #16 \n" \
"str q0, ["Y"], #16 \n"
#define INIT \
"lsl "INC_X", "INC_X", #"INC_SHIFT" \n" \
"lsl "INC_Y", "INC_Y", #"INC_SHIFT" \n"
static int do_copy(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{
if ( n < 0 ) return 0;
__asm__ __volatile__ (
" mov "N", %[N_] \n"
" mov "X", %[X_] \n"
" mov "INC_X", %[INCX_] \n"
" mov "Y", %[Y_] \n"
" mov "INC_Y", %[INCY_] \n"
" cmp "N", xzr \n"
" ble 8f //copy_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne 4f //copy_kernel_S_BEGIN \n"
" cmp "INC_Y", #1 \n"
" bne 4f //copy_kernel_S_BEGIN \n"
"// .Lcopy_kernel_F_BEGIN: \n"
" "INIT" \n"
" asr "J", "N", #"N_DIV_SHIFT" \n"
" cmp "J", xzr \n"
" beq 2f //copy_kernel_F1 \n"
#if !(defined(__clang__) && defined(OS_WINDOWS))
" .align 5 \n"
#endif
"1: //copy_kernel_F: \n"
" "KERNEL_F" \n"
" subs "J", "J", #1 \n"
" bne 1b //copy_kernel_F \n"
"2: //copy_kernel_F1: \n"
#if defined(COMPLEX) && defined(DOUBLE)
" b 8f //copy_kernel_L999 \n"
#else
" ands "J", "N", #"N_REM_MASK" \n"
" ble 8f //copy_kernel_L999 \n"
#endif
"3: //copy_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 3b //copy_kernel_F10 \n"
" b 8f //copy_kernel_L999 \n"
"4: //copy_kernel_S_BEGIN: \n"
" "INIT" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble 6f //copy_kernel_S1 \n"
"5: //copy_kernel_S4: \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 5b //copy_kernel_S4 \n"
"6: //copy_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble 8f //copy_kernel_L999 \n"
"7: //copy_kernel_S10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 7b //copy_kernel_S10 \n"
"8: //copy_kernel_L999: \n"
:
: [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"
);
return 0;
}
#if defined(SMP)
static int copy_thread_function(BLASLONG n, BLASLONG dummy0,
BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y,
BLASLONG inc_y, FLOAT *dummy3, BLASLONG dummy4)
{
do_copy(n, x, inc_x, y, inc_y);
return 0;
}
#endif
int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{
#if defined(SMP)
int nthreads;
FLOAT dummy_alpha;
#endif
if (n <= 0) return 0;
#if defined(SMP)
if (inc_x == 0 || n <= 10000)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
if (nthreads == 1) {
do_copy(n, x, inc_x, y, inc_y);
} else {
int mode = 0;
#if !defined(COMPLEX)
mode = BLAS_REAL;
#else
mode = BLAS_COMPLEX;
#endif
#if !defined(DOUBLE)
mode |= BLAS_SINGLE;
#else
mode |= BLAS_DOUBLE;
#endif
blas_level1_thread(mode, n, 0, 0, &dummy_alpha,
x, inc_x, y, inc_y, NULL, 0,
( void *)copy_thread_function, nthreads);
}
#else
do_copy(n, x, inc_x, y, inc_y);
#endif
return 0;
}

+ 262
- 260
kernel/arm64/dasum_thunderx2t99.c View File

@@ -1,260 +1,262 @@
/***************************************************************************
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 <arm_neon.h>

#define N "x0" /* vector length */
#define X "x1" /* "X" vector address */
#define INC_X "x2" /* "X" stride */
#define J "x5" /* loop variable */

#define REG0 "xzr"
#define SUMF "d0"
#define TMPF "d1"

/******************************************************************************/

#define KERNEL_F1 \
"ldr "TMPF", ["X"] \n" \
"add "X", "X", #8 \n" \
"fabs "TMPF", "TMPF" \n" \
"fadd "SUMF", "SUMF", "TMPF" \n"

#define KERNEL_F32 \
"ldr q16, ["X"] \n" \
"ldr q17, ["X", #16] \n" \
"ldr q18, ["X", #32] \n" \
"ldr q19, ["X", #48] \n" \
"ldp q20, q21, ["X", #64] \n" \
"ldp q22, q23, ["X", #96] \n" \
"fabs v16.2d, v16.2d \n" \
"fabs v17.2d, v17.2d \n" \
"fabs v18.2d, v18.2d \n" \
"fabs v19.2d, v19.2d \n" \
"ldp q24, q25, ["X", #128] \n" \
"ldp q26, q27, ["X", #160] \n" \
"fabs v20.2d, v20.2d \n" \
"fabs v21.2d, v21.2d \n" \
"fabs v22.2d, v22.2d \n" \
"fabs v23.2d, v23.2d \n" \
"fadd v16.2d, v16.2d, v17.2d \n" \
"fadd v18.2d, v18.2d, v19.2d \n" \
"ldp q28, q29, ["X", #192] \n" \
"ldp q30, q31, ["X", #224] \n" \
"fabs v24.2d, v24.2d \n" \
"fabs v25.2d, v25.2d \n" \
"fabs v26.2d, v26.2d \n" \
"fabs v27.2d, v27.2d \n" \
"add "X", "X", #256 \n" \
"fadd v20.2d, v20.2d, v21.2d \n" \
"fadd v22.2d, v22.2d, v23.2d \n" \
"fabs v28.2d, v28.2d \n" \
"fabs v29.2d, v29.2d \n" \
"fabs v30.2d, v30.2d \n" \
"fabs v31.2d, v31.2d \n" \
"PRFM PLDL1KEEP, ["X", #1024] \n" \
"PRFM PLDL1KEEP, ["X", #1024+64] \n" \
"fadd v24.2d, v24.2d, v25.2d \n" \
"fadd v26.2d, v26.2d, v27.2d \n" \
"fadd v28.2d, v28.2d, v29.2d \n" \
"fadd v30.2d, v30.2d, v31.2d \n" \
"fadd v0.2d, v0.2d, v16.2d \n" \
"fadd v1.2d, v1.2d, v18.2d \n" \
"fadd v2.2d, v2.2d, v20.2d \n" \
"fadd v3.2d, v3.2d, v22.2d \n" \
"PRFM PLDL1KEEP, ["X", #1024+128] \n" \
"PRFM PLDL1KEEP, ["X", #1024+192] \n" \
"fadd v4.2d, v4.2d, v24.2d \n" \
"fadd v5.2d, v5.2d, v26.2d \n" \
"fadd v6.2d, v6.2d, v28.2d \n" \
"fadd v7.2d, v7.2d, v30.2d \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 "SUMF", v0.2d \n"

#define INIT_S \
"lsl "INC_X", "INC_X", #3 \n"

#define KERNEL_S1 \
"ldr "TMPF", ["X"] \n" \
"add "X", "X", "INC_X" \n" \
"fabs "TMPF", "TMPF" \n" \
"fadd "SUMF", "SUMF", "TMPF" \n"


#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 dasum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
{
FLOAT asum = 0.0 ;

if ( n < 0 ) return(asum);

__asm__ __volatile__ (
" mov "N", %[N_] \n"
" mov "X", %[X_] \n"
" mov "INC_X", %[INCX_] \n"
" fmov "SUMF", "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 //asum_kernel_L999 \n"
" cmp "INC_X", xzr \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne 5f //asum_kernel_S_BEGIN \n"

"1: //asum_kernel_F_BEGIN: \n"
" asr "J", "N", #5 \n"
" cmp "J", xzr \n"
" beq 3f //asum_kernel_F1 \n"

".align 5 \n"
"2: //asum_kernel_F32: \n"
" "KERNEL_F32" \n"
" subs "J", "J", #1 \n"
" bne 2b //asum_kernel_F32 \n"
" "KERNEL_F32_FINALIZE" \n"

"3: //asum_kernel_F1: \n"
" ands "J", "N", #31 \n"
" ble 9f //asum_kernel_L999 \n"

"4: //asum_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 4b //asum_kernel_F10 \n"
" b 9f //asum_kernel_L999 \n"

"5: //asum_kernel_S_BEGIN: \n"
" "INIT_S" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble 7f //asum_kernel_S1 \n"

"6: //asum_kernel_S4: \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 6b //asum_kernel_S4 \n"

"7: //asum_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble 9f //asum_kernel_L999 \n"

"8: //asum_kernel_S10: \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 8b //asum_kernel_S10 \n"

"9: //asum_kernel_L999: \n"
" fmov %[ASUM_], "SUMF" \n"

: [ASUM_] "=r" (asum) //%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 asum;
}

#if defined(SMP)
static int dasum_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 = dasum_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 asum = 0.0;

#if defined(SMP)
if (inc_x == 0 || n <= 10000)
nthreads = 1;
else
nthreads = num_cpu_avail(1);

if (nthreads == 1) {
asum = dasum_compute(n, x, inc_x);
} else {
int mode, i;
char result[MAX_CPU_NUMBER * sizeof(double) * 2];
FLOAT *ptr;

mode = BLAS_DOUBLE;

blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
x, inc_x, NULL, 0, result, 0,
( void *)dasum_thread_function, nthreads);

ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) {
asum = asum + (*ptr);
ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2);
}
}
#else
asum = dasum_compute(n, x, inc_x);
#endif

return asum;
}
/***************************************************************************
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 <arm_neon.h>
#define N "x0" /* vector length */
#define X "x1" /* "X" vector address */
#define INC_X "x2" /* "X" stride */
#define J "x5" /* loop variable */
#define REG0 "xzr"
#define SUMF "d0"
#define TMPF "d1"
/******************************************************************************/
#define KERNEL_F1 \
"ldr "TMPF", ["X"] \n" \
"add "X", "X", #8 \n" \
"fabs "TMPF", "TMPF" \n" \
"fadd "SUMF", "SUMF", "TMPF" \n"
#define KERNEL_F32 \
"ldr q16, ["X"] \n" \
"ldr q17, ["X", #16] \n" \
"ldr q18, ["X", #32] \n" \
"ldr q19, ["X", #48] \n" \
"ldp q20, q21, ["X", #64] \n" \
"ldp q22, q23, ["X", #96] \n" \
"fabs v16.2d, v16.2d \n" \
"fabs v17.2d, v17.2d \n" \
"fabs v18.2d, v18.2d \n" \
"fabs v19.2d, v19.2d \n" \
"ldp q24, q25, ["X", #128] \n" \
"ldp q26, q27, ["X", #160] \n" \
"fabs v20.2d, v20.2d \n" \
"fabs v21.2d, v21.2d \n" \
"fabs v22.2d, v22.2d \n" \
"fabs v23.2d, v23.2d \n" \
"fadd v16.2d, v16.2d, v17.2d \n" \
"fadd v18.2d, v18.2d, v19.2d \n" \
"ldp q28, q29, ["X", #192] \n" \
"ldp q30, q31, ["X", #224] \n" \
"fabs v24.2d, v24.2d \n" \
"fabs v25.2d, v25.2d \n" \
"fabs v26.2d, v26.2d \n" \
"fabs v27.2d, v27.2d \n" \
"add "X", "X", #256 \n" \
"fadd v20.2d, v20.2d, v21.2d \n" \
"fadd v22.2d, v22.2d, v23.2d \n" \
"fabs v28.2d, v28.2d \n" \
"fabs v29.2d, v29.2d \n" \
"fabs v30.2d, v30.2d \n" \
"fabs v31.2d, v31.2d \n" \
"PRFM PLDL1KEEP, ["X", #1024] \n" \
"PRFM PLDL1KEEP, ["X", #1024+64] \n" \
"fadd v24.2d, v24.2d, v25.2d \n" \
"fadd v26.2d, v26.2d, v27.2d \n" \
"fadd v28.2d, v28.2d, v29.2d \n" \
"fadd v30.2d, v30.2d, v31.2d \n" \
"fadd v0.2d, v0.2d, v16.2d \n" \
"fadd v1.2d, v1.2d, v18.2d \n" \
"fadd v2.2d, v2.2d, v20.2d \n" \
"fadd v3.2d, v3.2d, v22.2d \n" \
"PRFM PLDL1KEEP, ["X", #1024+128] \n" \
"PRFM PLDL1KEEP, ["X", #1024+192] \n" \
"fadd v4.2d, v4.2d, v24.2d \n" \
"fadd v5.2d, v5.2d, v26.2d \n" \
"fadd v6.2d, v6.2d, v28.2d \n" \
"fadd v7.2d, v7.2d, v30.2d \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 "SUMF", v0.2d \n"
#define INIT_S \
"lsl "INC_X", "INC_X", #3 \n"
#define KERNEL_S1 \
"ldr "TMPF", ["X"] \n" \
"add "X", "X", "INC_X" \n" \
"fabs "TMPF", "TMPF" \n" \
"fadd "SUMF", "SUMF", "TMPF" \n"
#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 dasum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
{
FLOAT asum = 0.0 ;
if ( n < 0 ) return(asum);
__asm__ __volatile__ (
" mov "N", %[N_] \n"
" mov "X", %[X_] \n"
" mov "INC_X", %[INCX_] \n"
" fmov "SUMF", "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 //asum_kernel_L999 \n"
" cmp "INC_X", xzr \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne 5f //asum_kernel_S_BEGIN \n"
"1: //asum_kernel_F_BEGIN: \n"
" asr "J", "N", #5 \n"
" cmp "J", xzr \n"
" beq 3f //asum_kernel_F1 \n"
#if !(defined(__clang__) && defined(OS_WINDOWS))
".align 5 \n"
#endif
"2: //asum_kernel_F32: \n"
" "KERNEL_F32" \n"
" subs "J", "J", #1 \n"
" bne 2b //asum_kernel_F32 \n"
" "KERNEL_F32_FINALIZE" \n"
"3: //asum_kernel_F1: \n"
" ands "J", "N", #31 \n"
" ble 9f //asum_kernel_L999 \n"
"4: //asum_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 4b //asum_kernel_F10 \n"
" b 9f //asum_kernel_L999 \n"
"5: //asum_kernel_S_BEGIN: \n"
" "INIT_S" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble 7f //asum_kernel_S1 \n"
"6: //asum_kernel_S4: \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 6b //asum_kernel_S4 \n"
"7: //asum_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble 9f //asum_kernel_L999 \n"
"8: //asum_kernel_S10: \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 8b //asum_kernel_S10 \n"
"9: //asum_kernel_L999: \n"
" fmov %[ASUM_], "SUMF" \n"
: [ASUM_] "=r" (asum) //%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 asum;
}
#if defined(SMP)
static int dasum_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 = dasum_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 asum = 0.0;
#if defined(SMP)
if (inc_x == 0 || n <= 10000)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
if (nthreads == 1) {
asum = dasum_compute(n, x, inc_x);
} else {
int mode, i;
char result[MAX_CPU_NUMBER * sizeof(double) * 2];
FLOAT *ptr;
mode = BLAS_DOUBLE;
blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
x, inc_x, NULL, 0, result, 0,
( void *)dasum_thread_function, nthreads);
ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) {
asum = asum + (*ptr);
ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2);
}
}
#else
asum = dasum_compute(n, x, inc_x);
#endif
return asum;
}

+ 346
- 345
kernel/arm64/dot_kernel_asimd.c View File

@@ -1,345 +1,346 @@
/***************************************************************************
Copyright (c) 2017, The OpenBLAS Project
Copyright (c) 2022, Arm Ltd
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 <arm_neon.h>

#if !defined(DSDOT)
#define RETURN_TYPE FLOAT
#else
#define RETURN_TYPE double
#endif

#if !defined(DOUBLE)
#if !defined(DSDOT)
#define DOT_MOD "s"
#define REG0 "wzr"
#define TMPX "s16"
#define TMPY "s24"
#define INC_SHIFT "2"
#define N_DIV_SHIFT "6"
#define N_REM_MASK "63"
#else
#define DOT_MOD "d"
#define REG0 "xzr"
#define TMPX "s16"
#define TMPX1 "d2"
#define TMPY "s24"
#define TMPY1 "d3"
#define INC_SHIFT "2"
#define N_DIV_SHIFT "4"
#define N_REM_MASK "15"
#endif
#else
#define DOT_MOD "d"
#define REG0 "xzr"
#define TMPX "d16"
#define TMPY "d24"
#define INC_SHIFT "3"
#define N_DIV_SHIFT "5"
#define N_REM_MASK "31"
#endif

#define OUT "%"DOT_MOD"[DOT_]"

#if !defined(DOUBLE)

#if !defined(DSDOT)
#define KERNEL_F1 \
" ldr "TMPX", [%[X_]] \n" \
" ldr "TMPY", [%[Y_]] \n" \
" add %[X_], %[X_], %[INCX_] \n" \
" add %[Y_], %[Y_], %[INCY_] \n" \
" fmadd "OUT", "TMPX", "TMPY", "OUT" \n"

#define KERNEL_F \
" ldp q16, q17, [%[X_]] \n" \
" ldp q24, q25, [%[Y_]] \n" \
" ldp q18, q19, [%[X_], #32] \n" \
" ldp q26, q27, [%[Y_], #32] \n" \
" fmla v0.4s, v16.4s, v24.4s \n" \
" fmla v1.4s, v17.4s, v25.4s \n" \
" ldp q20, q21, [%[X_], #64] \n" \
" ldp q28, q29, [%[Y_], #64] \n" \
" fmla v2.4s, v18.4s, v26.4s \n" \
" fmla v3.4s, v19.4s, v27.4s \n" \
" ldp q22, q23, [%[X_], #96] \n" \
" ldp q30, q31, [%[Y_], #96] \n" \
" add %[Y_], %[Y_], #128 \n" \
" add %[X_], %[X_], #128 \n" \
" fmla v4.4s, v20.4s, v28.4s \n" \
" fmla v5.4s, v21.4s, v29.4s \n" \
" PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v6.4s, v22.4s, v30.4s \n" \
" fmla v7.4s, v23.4s, v31.4s \n" \
" ldp q16, q17, [%[X_]] \n" \
" ldp q24, q25, [%[Y_]] \n" \
" ldp q18, q19, [%[X_], #32] \n" \
" ldp q26, q27, [%[Y_], #32] \n" \
" fmla v0.4s, v16.4s, v24.4s \n" \
" fmla v1.4s, v17.4s, v25.4s \n" \
" ldp q20, q21, [%[X_], #64] \n" \
" ldp q28, q29, [%[Y_], #64] \n" \
" fmla v2.4s, v18.4s, v26.4s \n" \
" fmla v3.4s, v19.4s, v27.4s \n" \
" ldp q22, q23, [%[X_], #96] \n" \
" ldp q30, q31, [%[Y_], #96] \n" \
" add %[Y_], %[Y_], #128 \n" \
" add %[X_], %[X_], #128 \n" \
" fmla v4.4s, v20.4s, v28.4s \n" \
" fmla v5.4s, v21.4s, v29.4s \n" \
" PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v6.4s, v22.4s, v30.4s \n" \
" fmla v7.4s, v23.4s, v31.4s \n"

#define KERNEL_F_FINALIZE \
" fadd v0.4s, v0.4s, v1.4s \n" \
" fadd v2.4s, v2.4s, v3.4s \n" \
" fadd v4.4s, v4.4s, v5.4s \n" \
" fadd v6.4s, v6.4s, v7.4s \n" \
" fadd v0.4s, v0.4s, v2.4s \n" \
" fadd v4.4s, v4.4s, v6.4s \n" \
" fadd v0.4s, v0.4s, v4.4s \n" \
" faddp v0.4s, v0.4s, v0.4s \n" \
" faddp v0.4s, v0.4s, v0.4s \n"

#else /* !defined(DSDOT) */
#define KERNEL_F1 \
" ldr "TMPX", [%[X_]] \n" \
" ldr "TMPY", [%[Y_]] \n" \
" add %[X_], %[X_], %[INCX_] \n" \
" add %[Y_], %[Y_], %[INCY_] \n" \
" fcvt "TMPX1", "TMPX" \n" \
" fcvt "TMPY1", "TMPY" \n" \
" fmul "TMPX1", "TMPX1", "TMPY1" \n" \
" fadd "OUT", "OUT", "TMPX1" \n"


#define KERNEL_F \
" ldp q18, q19, [%[X_]] \n" \
" ldp q26, q27, [%[Y_]] \n" \
" fcvtl v16.2d, v18.2s \n" \
" fcvtl2 v17.2d, v18.4s \n" \
" fcvtl v18.2d, v19.2s \n" \
" fcvtl2 v19.2d, v19.4s \n" \
" fcvtl v24.2d, v26.2s \n" \
" fcvtl2 v25.2d, v26.4s \n" \
" fcvtl v26.2d, v27.2s \n" \
" fcvtl2 v27.2d, v27.4s \n" \
" ldp q22, q23, [%[X_], #32] \n" \
" ldp q30, q31, [%[Y_], #32] \n" \
" fcvtl v20.2d, v22.2s \n" \
" fcvtl2 v21.2d, v22.4s \n" \
" fcvtl v22.2d, v23.2s \n" \
" fcvtl2 v23.2d, v23.4s \n" \
" fcvtl v28.2d, v30.2s \n" \
" fcvtl2 v29.2d, v30.4s \n" \
" fcvtl v30.2d, v31.2s \n" \
" fcvtl2 v31.2d, v31.4s \n" \
" PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v0.2d, v16.2d, v24.2d \n" \
" fmla v1.2d, v17.2d, v25.2d \n" \
" fmla v2.2d, v18.2d, v26.2d \n" \
" fmla v3.2d, v19.2d, v27.2d \n" \
" add %[Y_], %[Y_], #64 \n" \
" add %[X_], %[X_], #64 \n" \
" fmla v4.2d, v20.2d, v28.2d \n" \
" fmla v5.2d, v21.2d, v29.2d \n" \
" fmla v6.2d, v22.2d, v30.2d \n" \
" fmla v7.2d, v23.2d, v31.2d \n"

#define KERNEL_F_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 "OUT", v0.2d \n"
#endif /* !defined(DSDOT) */

#else /* !defined(DOUBLE) */
#define KERNEL_F1 \
" ldr "TMPX", [%[X_]] \n" \
" ldr "TMPY", [%[Y_]] \n" \
" add %[X_], %[X_], %[INCX_] \n" \
" add %[Y_], %[Y_], %[INCY_] \n" \
" fmadd "OUT", "TMPX", "TMPY", "OUT" \n"

#define KERNEL_F \
" 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"

#define KERNEL_F_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 "OUT", v0.2d \n"
#endif /* !defined(DOUBLE) */

static RETURN_TYPE dot_kernel_asimd(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{
RETURN_TYPE dot = 0.0;
BLASLONG j = 0;

__asm__ __volatile__ (
" fmov "OUT", "REG0" \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 %[INCX_], #1 \n"
" bne 5f //dot_kernel_S_BEGIN \n"
" cmp %[INCY_], #1 \n"
" bne 5f //dot_kernel_S_BEGIN \n"

"1: //dot_kernel_F_BEGIN: \n"
" lsl %[INCX_], %[INCX_], "INC_SHIFT" \n"
" lsl %[INCY_], %[INCY_], "INC_SHIFT" \n"
" asr %[J_], %[N_], #"N_DIV_SHIFT" \n"
" cmp %[J_], xzr \n"
" beq 3f //dot_kernel_F1 \n"

" .align 5 \n"
"2: //dot_kernel_F: \n"
" "KERNEL_F" \n"
" subs %[J_], %[J_], #1 \n"
" bne 2b //dot_kernel_F \n"
" "KERNEL_F_FINALIZE" \n"

"3: //dot_kernel_F1: \n"
" ands %[J_], %[N_], #"N_REM_MASK" \n"
" ble 9f //dot_kernel_L999 \n"

"4: //dot_kernel_F10: \n"
" "KERNEL_F1" \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 %[INCX_], %[INCX_], "INC_SHIFT" \n"
" lsl %[INCY_], %[INCY_], "INC_SHIFT" \n"
" asr %[J_], %[N_], #2 \n"
" cmp %[J_], xzr \n"
" ble 7f //dot_kernel_S1 \n"

"6: //dot_kernel_S4: \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \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"
" "KERNEL_F1" \n"
" subs %[J_], %[J_], #1 \n"
" bne 8b //dot_kernel_S10 \n"

"9: //dot_kernel_L999: \n"

: [DOT_] "=&w" (dot)
: [N_] "r" (n),
[X_] "r" (x),
[INCX_] "r" (inc_x),
[Y_] "r" (y),
[INCY_] "r" (inc_y),
[J_] "r" (j)
: "cc",
"memory",
"d1", "d2", "d3", "d4", "d5", "d6", "d7"
);

return dot;
}
/***************************************************************************
Copyright (c) 2017, The OpenBLAS Project
Copyright (c) 2022, Arm Ltd
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 <arm_neon.h>
#if !defined(DSDOT)
#define RETURN_TYPE FLOAT
#else
#define RETURN_TYPE double
#endif
#if !defined(DOUBLE)
#if !defined(DSDOT)
#define DOT_MOD "s"
#define REG0 "wzr"
#define TMPX "s16"
#define TMPY "s24"
#define INC_SHIFT "2"
#define N_DIV_SHIFT "6"
#define N_REM_MASK "63"
#else
#define DOT_MOD "d"
#define REG0 "xzr"
#define TMPX "s16"
#define TMPX1 "d2"
#define TMPY "s24"
#define TMPY1 "d3"
#define INC_SHIFT "2"
#define N_DIV_SHIFT "4"
#define N_REM_MASK "15"
#endif
#else
#define DOT_MOD "d"
#define REG0 "xzr"
#define TMPX "d16"
#define TMPY "d24"
#define INC_SHIFT "3"
#define N_DIV_SHIFT "5"
#define N_REM_MASK "31"
#endif
#define OUT "%"DOT_MOD"[DOT_]"
#if !defined(DOUBLE)
#if !defined(DSDOT)
#define KERNEL_F1 \
" ldr "TMPX", [%[X_]] \n" \
" ldr "TMPY", [%[Y_]] \n" \
" add %[X_], %[X_], %[INCX_] \n" \
" add %[Y_], %[Y_], %[INCY_] \n" \
" fmadd "OUT", "TMPX", "TMPY", "OUT" \n"
#define KERNEL_F \
" ldp q16, q17, [%[X_]] \n" \
" ldp q24, q25, [%[Y_]] \n" \
" ldp q18, q19, [%[X_], #32] \n" \
" ldp q26, q27, [%[Y_], #32] \n" \
" fmla v0.4s, v16.4s, v24.4s \n" \
" fmla v1.4s, v17.4s, v25.4s \n" \
" ldp q20, q21, [%[X_], #64] \n" \
" ldp q28, q29, [%[Y_], #64] \n" \
" fmla v2.4s, v18.4s, v26.4s \n" \
" fmla v3.4s, v19.4s, v27.4s \n" \
" ldp q22, q23, [%[X_], #96] \n" \
" ldp q30, q31, [%[Y_], #96] \n" \
" add %[Y_], %[Y_], #128 \n" \
" add %[X_], %[X_], #128 \n" \
" fmla v4.4s, v20.4s, v28.4s \n" \
" fmla v5.4s, v21.4s, v29.4s \n" \
" PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v6.4s, v22.4s, v30.4s \n" \
" fmla v7.4s, v23.4s, v31.4s \n" \
" ldp q16, q17, [%[X_]] \n" \
" ldp q24, q25, [%[Y_]] \n" \
" ldp q18, q19, [%[X_], #32] \n" \
" ldp q26, q27, [%[Y_], #32] \n" \
" fmla v0.4s, v16.4s, v24.4s \n" \
" fmla v1.4s, v17.4s, v25.4s \n" \
" ldp q20, q21, [%[X_], #64] \n" \
" ldp q28, q29, [%[Y_], #64] \n" \
" fmla v2.4s, v18.4s, v26.4s \n" \
" fmla v3.4s, v19.4s, v27.4s \n" \
" ldp q22, q23, [%[X_], #96] \n" \
" ldp q30, q31, [%[Y_], #96] \n" \
" add %[Y_], %[Y_], #128 \n" \
" add %[X_], %[X_], #128 \n" \
" fmla v4.4s, v20.4s, v28.4s \n" \
" fmla v5.4s, v21.4s, v29.4s \n" \
" PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v6.4s, v22.4s, v30.4s \n" \
" fmla v7.4s, v23.4s, v31.4s \n"
#define KERNEL_F_FINALIZE \
" fadd v0.4s, v0.4s, v1.4s \n" \
" fadd v2.4s, v2.4s, v3.4s \n" \
" fadd v4.4s, v4.4s, v5.4s \n" \
" fadd v6.4s, v6.4s, v7.4s \n" \
" fadd v0.4s, v0.4s, v2.4s \n" \
" fadd v4.4s, v4.4s, v6.4s \n" \
" fadd v0.4s, v0.4s, v4.4s \n" \
" faddp v0.4s, v0.4s, v0.4s \n" \
" faddp v0.4s, v0.4s, v0.4s \n"
#else /* !defined(DSDOT) */
#define KERNEL_F1 \
" ldr "TMPX", [%[X_]] \n" \
" ldr "TMPY", [%[Y_]] \n" \
" add %[X_], %[X_], %[INCX_] \n" \
" add %[Y_], %[Y_], %[INCY_] \n" \
" fcvt "TMPX1", "TMPX" \n" \
" fcvt "TMPY1", "TMPY" \n" \
" fmul "TMPX1", "TMPX1", "TMPY1" \n" \
" fadd "OUT", "OUT", "TMPX1" \n"
#define KERNEL_F \
" ldp q18, q19, [%[X_]] \n" \
" ldp q26, q27, [%[Y_]] \n" \
" fcvtl v16.2d, v18.2s \n" \
" fcvtl2 v17.2d, v18.4s \n" \
" fcvtl v18.2d, v19.2s \n" \
" fcvtl2 v19.2d, v19.4s \n" \
" fcvtl v24.2d, v26.2s \n" \
" fcvtl2 v25.2d, v26.4s \n" \
" fcvtl v26.2d, v27.2s \n" \
" fcvtl2 v27.2d, v27.4s \n" \
" ldp q22, q23, [%[X_], #32] \n" \
" ldp q30, q31, [%[Y_], #32] \n" \
" fcvtl v20.2d, v22.2s \n" \
" fcvtl2 v21.2d, v22.4s \n" \
" fcvtl v22.2d, v23.2s \n" \
" fcvtl2 v23.2d, v23.4s \n" \
" fcvtl v28.2d, v30.2s \n" \
" fcvtl2 v29.2d, v30.4s \n" \
" fcvtl v30.2d, v31.2s \n" \
" fcvtl2 v31.2d, v31.4s \n" \
" PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v0.2d, v16.2d, v24.2d \n" \
" fmla v1.2d, v17.2d, v25.2d \n" \
" fmla v2.2d, v18.2d, v26.2d \n" \
" fmla v3.2d, v19.2d, v27.2d \n" \
" add %[Y_], %[Y_], #64 \n" \
" add %[X_], %[X_], #64 \n" \
" fmla v4.2d, v20.2d, v28.2d \n" \
" fmla v5.2d, v21.2d, v29.2d \n" \
" fmla v6.2d, v22.2d, v30.2d \n" \
" fmla v7.2d, v23.2d, v31.2d \n"
#define KERNEL_F_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 "OUT", v0.2d \n"
#endif /* !defined(DSDOT) */
#else /* !defined(DOUBLE) */
#define KERNEL_F1 \
" ldr "TMPX", [%[X_]] \n" \
" ldr "TMPY", [%[Y_]] \n" \
" add %[X_], %[X_], %[INCX_] \n" \
" add %[Y_], %[Y_], %[INCY_] \n" \
" fmadd "OUT", "TMPX", "TMPY", "OUT" \n"
#define KERNEL_F \
" 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"
#define KERNEL_F_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 "OUT", v0.2d \n"
#endif /* !defined(DOUBLE) */
static RETURN_TYPE dot_kernel_asimd(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{
RETURN_TYPE dot = 0.0;
BLASLONG j = 0;
__asm__ __volatile__ (
" fmov "OUT", "REG0" \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 %[INCX_], #1 \n"
" bne 5f //dot_kernel_S_BEGIN \n"
" cmp %[INCY_], #1 \n"
" bne 5f //dot_kernel_S_BEGIN \n"
"1: //dot_kernel_F_BEGIN: \n"
" lsl %[INCX_], %[INCX_], "INC_SHIFT" \n"
" lsl %[INCY_], %[INCY_], "INC_SHIFT" \n"
" asr %[J_], %[N_], #"N_DIV_SHIFT" \n"
" cmp %[J_], xzr \n"
" beq 3f //dot_kernel_F1 \n"
#if !(defined(__clang__) && defined(OS_WINDOWS))
" .align 5 \n"
#endif
"2: //dot_kernel_F: \n"
" "KERNEL_F" \n"
" subs %[J_], %[J_], #1 \n"
" bne 2b //dot_kernel_F \n"
" "KERNEL_F_FINALIZE" \n"
"3: //dot_kernel_F1: \n"
" ands %[J_], %[N_], #"N_REM_MASK" \n"
" ble 9f //dot_kernel_L999 \n"
"4: //dot_kernel_F10: \n"
" "KERNEL_F1" \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 %[INCX_], %[INCX_], "INC_SHIFT" \n"
" lsl %[INCY_], %[INCY_], "INC_SHIFT" \n"
" asr %[J_], %[N_], #2 \n"
" cmp %[J_], xzr \n"
" ble 7f //dot_kernel_S1 \n"
"6: //dot_kernel_S4: \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \n"
" "KERNEL_F1" \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"
" "KERNEL_F1" \n"
" subs %[J_], %[J_], #1 \n"
" bne 8b //dot_kernel_S10 \n"
"9: //dot_kernel_L999: \n"
: [DOT_] "=&w" (dot)
: [N_] "r" (n),
[X_] "r" (x),
[INCX_] "r" (inc_x),
[Y_] "r" (y),
[INCY_] "r" (inc_y),
[J_] "r" (j)
: "cc",
"memory",
"d1", "d2", "d3", "d4", "d5", "d6", "d7"
);
return dot;
}

+ 263
- 262
kernel/arm64/sasum_thunderx2t99.c View File

@@ -1,262 +1,263 @@
/***************************************************************************
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 <arm_neon.h>

#define N "x0" /* vector length */
#define X "x1" /* "X" vector address */
#define INC_X "x2" /* "X" stride */
#define J "x5" /* loop variable */

#define REG0 "wzr"
#define SUMF "s0"
#define SUMFD "d0"

/******************************************************************************/

#define KERNEL_F1 \
"ldr s1, ["X"] \n" \
"add "X", "X", #4 \n" \
"fabs s1, s1 \n" \
"fadd "SUMF", "SUMF", s1 \n"

#define KERNEL_F64 \
"ldr q16, ["X"] \n" \
"ldr q17, ["X", #16] \n" \
"ldr q18, ["X", #32] \n" \
"ldr q19, ["X", #48] \n" \
"ldp q20, q21, ["X", #64] \n" \
"ldp q22, q23, ["X", #96] \n" \
"fabs v16.4s, v16.4s \n" \
"fabs v17.4s, v17.4s \n" \
"fabs v18.4s, v18.4s \n" \
"fabs v19.4s, v19.4s \n" \
"ldp q24, q25, ["X", #128] \n" \
"ldp q26, q27, ["X", #160] \n" \
"fabs v20.4s, v20.4s \n" \
"fabs v21.4s, v21.4s \n" \
"fabs v22.4s, v22.4s \n" \
"fabs v23.4s, v23.4s \n" \
"fadd v16.4s, v16.4s, v17.4s \n" \
"fadd v18.4s, v18.4s, v19.4s \n" \
"ldp q28, q29, ["X", #192] \n" \
"ldp q30, q31, ["X", #224] \n" \
"fabs v24.4s, v24.4s \n" \
"fabs v25.4s, v25.4s \n" \
"fabs v26.4s, v26.4s \n" \
"fabs v27.4s, v27.4s \n" \
"add "X", "X", #256 \n" \
"fadd v20.4s, v20.4s, v21.4s \n" \
"fadd v22.4s, v22.4s, v23.4s \n" \
"fabs v28.4s, v28.4s \n" \
"fabs v29.4s, v29.4s \n" \
"fabs v30.4s, v30.4s \n" \
"fabs v31.4s, v31.4s \n" \
"PRFM PLDL1KEEP, ["X", #1024] \n" \
"PRFM PLDL1KEEP, ["X", #1024+64] \n" \
"fadd v24.4s, v24.4s, v25.4s \n" \
"fadd v26.4s, v26.4s, v27.4s \n" \
"fadd v0.4s, v0.4s, v16.4s \n" \
"fadd v1.4s, v1.4s, v18.4s \n" \
"fadd v2.4s, v2.4s, v20.4s \n" \
"fadd v3.4s, v3.4s, v22.4s \n" \
"PRFM PLDL1KEEP, ["X", #1024+128] \n" \
"PRFM PLDL1KEEP, ["X", #1024+192] \n" \
"fadd v28.4s, v28.4s, v29.4s \n" \
"fadd v30.4s, v30.4s, v31.4s \n" \
"fadd v4.4s, v4.4s, v24.4s \n" \
"fadd v5.4s, v5.4s, v26.4s \n" \
"fadd v6.4s, v6.4s, v28.4s \n" \
"fadd v7.4s, v7.4s, v30.4s \n"

#define KERNEL_F64_FINALIZE \
"fadd v0.4s, v0.4s, v1.4s \n" \
"fadd v2.4s, v2.4s, v3.4s \n" \
"fadd v4.4s, v4.4s, v5.4s \n" \
"fadd v6.4s, v6.4s, v7.4s \n" \
"fadd v0.4s, v0.4s, v2.4s \n" \
"fadd v4.4s, v4.4s, v6.4s \n" \
"fadd v0.4s, v0.4s, v4.4s \n" \
"ext v1.16b, v0.16b, v0.16b, #8 \n" \
"fadd v0.2s, v0.2s, v1.2s \n" \
"faddp "SUMF", v0.2s \n"

#define INIT_S \
"lsl "INC_X", "INC_X", #2 \n"

#define KERNEL_S1 \
"ldr s1, ["X"] \n" \
"add "X", "X", "INC_X" \n" \
"fabs s1, s1 \n" \
"fadd "SUMF", "SUMF", s1 \n"


#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 sasum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
{
FLOAT asum = 0.0 ;

if ( n < 0 ) return(asum);

__asm__ __volatile__ (
" mov "N", %[N_] \n"
" mov "X", %[X_] \n"
" mov "INC_X", %[INCX_] \n"
" fmov "SUMF", "REG0" \n"
" fmov s1, "REG0" \n"
" fmov s2, "REG0" \n"
" fmov s3, "REG0" \n"
" fmov s4, "REG0" \n"
" fmov s5, "REG0" \n"
" fmov s6, "REG0" \n"
" fmov s7, "REG0" \n"
" cmp "N", xzr \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", xzr \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne 5f //asum_kernel_S_BEGIN \n"

"1: //asum_kernel_F_BEGIN: \n"
" asr "J", "N", #6 \n"
" cmp "J", xzr \n"
" beq 3f //asum_kernel_F1 \n"

".align 5 \n"
"2: //asum_kernel_F64: \n"
" "KERNEL_F64" \n"
" subs "J", "J", #1 \n"
" bne 2b //asum_kernel_F64 \n"
" "KERNEL_F64_FINALIZE" \n"

"3: //asum_kernel_F1: \n"
" ands "J", "N", #63 \n"
" ble 9f //asum_kernel_L999 \n"

"4: //asum_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 4b //asum_kernel_F10 \n"
" b 9f //asum_kernel_L999 \n"

"5: //asum_kernel_S_BEGIN: \n"
" "INIT_S" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble 7f //asum_kernel_S1 \n"

"6: //asum_kernel_S4: \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 6b //asum_kernel_S4 \n"

"7: //asum_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble 9f //asum_kernel_L999 \n"

"8: //asum_kernel_S10: \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 8b //asum_kernel_S10 \n"

"9: //asum_kernel_L999: \n"
" fmov %[ASUM_], "SUMFD" \n"

: [ASUM_] "=r" (asum) //%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 asum;
}

#if defined(SMP)
static int sasum_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 = sasum_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 asum = 0.0;

#if defined(SMP)
if (inc_x == 0 || n <= 10000)
nthreads = 1;
else
nthreads = num_cpu_avail(1);

if (nthreads == 1) {
asum = sasum_compute(n, x, inc_x);
} else {
int mode, i;
char result[MAX_CPU_NUMBER * sizeof(double) * 2];
FLOAT *ptr;

mode = BLAS_SINGLE;

blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
x, inc_x, NULL, 0, result, 0,
( void *)sasum_thread_function, nthreads);

ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) {
asum = asum + (*ptr);
ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2);
}
}
#else
asum = sasum_compute(n, x, inc_x);
#endif

return asum;
}
/***************************************************************************
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 <arm_neon.h>
#define N "x0" /* vector length */
#define X "x1" /* "X" vector address */
#define INC_X "x2" /* "X" stride */
#define J "x5" /* loop variable */
#define REG0 "wzr"
#define SUMF "s0"
#define SUMFD "d0"
/******************************************************************************/
#define KERNEL_F1 \
"ldr s1, ["X"] \n" \
"add "X", "X", #4 \n" \
"fabs s1, s1 \n" \
"fadd "SUMF", "SUMF", s1 \n"
#define KERNEL_F64 \
"ldr q16, ["X"] \n" \
"ldr q17, ["X", #16] \n" \
"ldr q18, ["X", #32] \n" \
"ldr q19, ["X", #48] \n" \
"ldp q20, q21, ["X", #64] \n" \
"ldp q22, q23, ["X", #96] \n" \
"fabs v16.4s, v16.4s \n" \
"fabs v17.4s, v17.4s \n" \
"fabs v18.4s, v18.4s \n" \
"fabs v19.4s, v19.4s \n" \
"ldp q24, q25, ["X", #128] \n" \
"ldp q26, q27, ["X", #160] \n" \
"fabs v20.4s, v20.4s \n" \
"fabs v21.4s, v21.4s \n" \
"fabs v22.4s, v22.4s \n" \
"fabs v23.4s, v23.4s \n" \
"fadd v16.4s, v16.4s, v17.4s \n" \
"fadd v18.4s, v18.4s, v19.4s \n" \
"ldp q28, q29, ["X", #192] \n" \
"ldp q30, q31, ["X", #224] \n" \
"fabs v24.4s, v24.4s \n" \
"fabs v25.4s, v25.4s \n" \
"fabs v26.4s, v26.4s \n" \
"fabs v27.4s, v27.4s \n" \
"add "X", "X", #256 \n" \
"fadd v20.4s, v20.4s, v21.4s \n" \
"fadd v22.4s, v22.4s, v23.4s \n" \
"fabs v28.4s, v28.4s \n" \
"fabs v29.4s, v29.4s \n" \
"fabs v30.4s, v30.4s \n" \
"fabs v31.4s, v31.4s \n" \
"PRFM PLDL1KEEP, ["X", #1024] \n" \
"PRFM PLDL1KEEP, ["X", #1024+64] \n" \
"fadd v24.4s, v24.4s, v25.4s \n" \
"fadd v26.4s, v26.4s, v27.4s \n" \
"fadd v0.4s, v0.4s, v16.4s \n" \
"fadd v1.4s, v1.4s, v18.4s \n" \
"fadd v2.4s, v2.4s, v20.4s \n" \
"fadd v3.4s, v3.4s, v22.4s \n" \
"PRFM PLDL1KEEP, ["X", #1024+128] \n" \
"PRFM PLDL1KEEP, ["X", #1024+192] \n" \
"fadd v28.4s, v28.4s, v29.4s \n" \
"fadd v30.4s, v30.4s, v31.4s \n" \
"fadd v4.4s, v4.4s, v24.4s \n" \
"fadd v5.4s, v5.4s, v26.4s \n" \
"fadd v6.4s, v6.4s, v28.4s \n" \
"fadd v7.4s, v7.4s, v30.4s \n"
#define KERNEL_F64_FINALIZE \
"fadd v0.4s, v0.4s, v1.4s \n" \
"fadd v2.4s, v2.4s, v3.4s \n" \
"fadd v4.4s, v4.4s, v5.4s \n" \
"fadd v6.4s, v6.4s, v7.4s \n" \
"fadd v0.4s, v0.4s, v2.4s \n" \
"fadd v4.4s, v4.4s, v6.4s \n" \
"fadd v0.4s, v0.4s, v4.4s \n" \
"ext v1.16b, v0.16b, v0.16b, #8 \n" \
"fadd v0.2s, v0.2s, v1.2s \n" \
"faddp "SUMF", v0.2s \n"
#define INIT_S \
"lsl "INC_X", "INC_X", #2 \n"
#define KERNEL_S1 \
"ldr s1, ["X"] \n" \
"add "X", "X", "INC_X" \n" \
"fabs s1, s1 \n" \
"fadd "SUMF", "SUMF", s1 \n"
#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 sasum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
{
FLOAT asum = 0.0 ;
if ( n < 0 ) return(asum);
__asm__ __volatile__ (
" mov "N", %[N_] \n"
" mov "X", %[X_] \n"
" mov "INC_X", %[INCX_] \n"
" fmov "SUMF", "REG0" \n"
" fmov s1, "REG0" \n"
" fmov s2, "REG0" \n"
" fmov s3, "REG0" \n"
" fmov s4, "REG0" \n"
" fmov s5, "REG0" \n"
" fmov s6, "REG0" \n"
" fmov s7, "REG0" \n"
" cmp "N", xzr \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", xzr \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne 5f //asum_kernel_S_BEGIN \n"
"1: //asum_kernel_F_BEGIN: \n"
" asr "J", "N", #6 \n"
" cmp "J", xzr \n"
" beq 3f //asum_kernel_F1 \n"
#if !(defined(__clang__) && defined(OS_WINDOWS))
".align 5 \n"
#endif
"2: //asum_kernel_F64: \n"
" "KERNEL_F64" \n"
" subs "J", "J", #1 \n"
" bne 2b //asum_kernel_F64 \n"
" "KERNEL_F64_FINALIZE" \n"
"3: //asum_kernel_F1: \n"
" ands "J", "N", #63 \n"
" ble 9f //asum_kernel_L999 \n"
"4: //asum_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 4b //asum_kernel_F10 \n"
" b 9f //asum_kernel_L999 \n"
"5: //asum_kernel_S_BEGIN: \n"
" "INIT_S" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble 7f //asum_kernel_S1 \n"
"6: //asum_kernel_S4: \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 6b //asum_kernel_S4 \n"
"7: //asum_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble 9f //asum_kernel_L999 \n"
"8: //asum_kernel_S10: \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 8b //asum_kernel_S10 \n"
"9: //asum_kernel_L999: \n"
" fmov %[ASUM_], "SUMFD" \n"
: [ASUM_] "=r" (asum) //%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 asum;
}
#if defined(SMP)
static int sasum_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 = sasum_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 asum = 0.0;
#if defined(SMP)
if (inc_x == 0 || n <= 10000)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
if (nthreads == 1) {
asum = sasum_compute(n, x, inc_x);
} else {
int mode, i;
char result[MAX_CPU_NUMBER * sizeof(double) * 2];
FLOAT *ptr;
mode = BLAS_SINGLE;
blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
x, inc_x, NULL, 0, result, 0,
( void *)sasum_thread_function, nthreads);
ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) {
asum = asum + (*ptr);
ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2);
}
}
#else
asum = sasum_compute(n, x, inc_x);
#endif
return asum;
}

+ 263
- 262
kernel/arm64/zasum_thunderx2t99.c View File

@@ -1,262 +1,263 @@
/***************************************************************************
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 <arm_neon.h>

#define N "x0" /* vector length */
#define X "x1" /* "X" vector address */
#define INC_X "x2" /* "X" stride */
#define J "x5" /* loop variable */

#define REG0 "xzr"
#define SUMF "d0"
#define TMPF "d1"

/******************************************************************************/

#define KERNEL_F1 \
"ldr q1, ["X"] \n" \
"add "X", "X", #16 \n" \
"fabs v1.2d, v1.2d \n" \
"faddp d1, v1.2d \n" \
"fadd "SUMF", "SUMF", d1 \n"

#define KERNEL_F16 \
"ldr q16, ["X"] \n" \
"ldr q17, ["X", #16] \n" \
"ldr q18, ["X", #32] \n" \
"ldr q19, ["X", #48] \n" \
"ldp q20, q21, ["X", #64] \n" \
"ldp q22, q23, ["X", #96] \n" \
"fabs v16.2d, v16.2d \n" \
"fabs v17.2d, v17.2d \n" \
"fabs v18.2d, v18.2d \n" \
"fabs v19.2d, v19.2d \n" \
"ldp q24, q25, ["X", #128] \n" \
"ldp q26, q27, ["X", #160] \n" \
"fabs v20.2d, v20.2d \n" \
"fabs v21.2d, v21.2d \n" \
"fabs v22.2d, v22.2d \n" \
"fabs v23.2d, v23.2d \n" \
"fadd v16.2d, v16.2d, v17.2d \n" \
"fadd v18.2d, v18.2d, v19.2d \n" \
"ldp q28, q29, ["X", #192] \n" \
"ldp q30, q31, ["X", #224] \n" \
"fabs v24.2d, v24.2d \n" \
"fabs v25.2d, v25.2d \n" \
"fabs v26.2d, v26.2d \n" \
"fabs v27.2d, v27.2d \n" \
"add "X", "X", #256 \n" \
"fadd v20.2d, v20.2d, v21.2d \n" \
"fadd v22.2d, v22.2d, v23.2d \n" \
"fabs v28.2d, v28.2d \n" \
"fabs v29.2d, v29.2d \n" \
"fabs v30.2d, v30.2d \n" \
"fabs v31.2d, v31.2d \n" \
"PRFM PLDL1KEEP, ["X", #1024] \n" \
"PRFM PLDL1KEEP, ["X", #1024+64] \n" \
"fadd v24.2d, v24.2d, v25.2d \n" \
"fadd v26.2d, v26.2d, v27.2d \n" \
"fadd v28.2d, v28.2d, v29.2d \n" \
"fadd v30.2d, v30.2d, v31.2d \n" \
"fadd v0.2d, v0.2d, v16.2d \n" \
"fadd v1.2d, v1.2d, v18.2d \n" \
"fadd v2.2d, v2.2d, v20.2d \n" \
"fadd v3.2d, v3.2d, v22.2d \n" \
"PRFM PLDL1KEEP, ["X", #1024+128] \n" \
"PRFM PLDL1KEEP, ["X", #1024+192] \n" \
"fadd v4.2d, v4.2d, v24.2d \n" \
"fadd v5.2d, v5.2d, v26.2d \n" \
"fadd v6.2d, v6.2d, v28.2d \n" \
"fadd v7.2d, v7.2d, v30.2d \n"

#define KERNEL_F16_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 "SUMF", v0.2d \n"

#define INIT_S \
"lsl "INC_X", "INC_X", #4 \n"

#define KERNEL_S1 \
"ldr q1, ["X"] \n" \
"add "X", "X", "INC_X" \n" \
"fabs v1.2d, v1.2d \n" \
"faddp d1, v1.2d \n" \
"fadd "SUMF", "SUMF", d1 \n"


#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 zasum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
{
FLOAT asum = 0.0 ;

if ( n < 0 ) return(asum);

__asm__ __volatile__ (
" mov "N", %[N_] \n"
" mov "X", %[X_] \n"
" mov "INC_X", %[INCX_] \n"
" fmov "SUMF", "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 //asum_kernel_L999 \n"
" cmp "INC_X", xzr \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne 5f //asum_kernel_S_BEGIN \n"

"1: //asum_kernel_F_BEGIN: \n"
" asr "J", "N", #4 \n"
" cmp "J", xzr \n"
" beq 3f //asum_kernel_F1 \n"

".align 5 \n"
"2: //asum_kernel_F16: \n"
" "KERNEL_F16" \n"
" subs "J", "J", #1 \n"
" bne 2b //asum_kernel_F16 \n"
" "KERNEL_F16_FINALIZE" \n"

"3: //asum_kernel_F1: \n"
" ands "J", "N", #15 \n"
" ble 9f //asum_kernel_L999 \n"

"4: //asum_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 4b //asum_kernel_F10 \n"
" b 9f //asum_kernel_L999 \n"

"5: //asum_kernel_S_BEGIN: \n"
" "INIT_S" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble 7f //asum_kernel_S1 \n"

"6: //asum_kernel_S4: \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 6b //asum_kernel_S4 \n"

"7: //asum_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble 9f //asum_kernel_L999 \n"

"8: //asum_kernel_S10: \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 8b //asum_kernel_S10 \n"

"9: //asum_kernel_L999: \n"
" fmov %[ASUM_], "SUMF" \n"

: [ASUM_] "=r" (asum) //%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 asum;
}

#if defined(SMP)
static int zasum_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 = zasum_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 asum = 0.0;

#if defined(SMP)
if (inc_x == 0 || n <= 10000)
nthreads = 1;
else
nthreads = num_cpu_avail(1);

if (nthreads == 1) {
asum = zasum_compute(n, x, inc_x);
} else {
int mode, i;
char result[MAX_CPU_NUMBER * sizeof(double) * 2];
FLOAT *ptr;

mode = BLAS_DOUBLE | BLAS_COMPLEX;

blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
x, inc_x, NULL, 0, result, 0,
( void *)zasum_thread_function, nthreads);

ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) {
asum = asum + (*ptr);
ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2);
}
}
#else
asum = zasum_compute(n, x, inc_x);
#endif

return asum;
}
/***************************************************************************
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 <arm_neon.h>
#define N "x0" /* vector length */
#define X "x1" /* "X" vector address */
#define INC_X "x2" /* "X" stride */
#define J "x5" /* loop variable */
#define REG0 "xzr"
#define SUMF "d0"
#define TMPF "d1"
/******************************************************************************/
#define KERNEL_F1 \
"ldr q1, ["X"] \n" \
"add "X", "X", #16 \n" \
"fabs v1.2d, v1.2d \n" \
"faddp d1, v1.2d \n" \
"fadd "SUMF", "SUMF", d1 \n"
#define KERNEL_F16 \
"ldr q16, ["X"] \n" \
"ldr q17, ["X", #16] \n" \
"ldr q18, ["X", #32] \n" \
"ldr q19, ["X", #48] \n" \
"ldp q20, q21, ["X", #64] \n" \
"ldp q22, q23, ["X", #96] \n" \
"fabs v16.2d, v16.2d \n" \
"fabs v17.2d, v17.2d \n" \
"fabs v18.2d, v18.2d \n" \
"fabs v19.2d, v19.2d \n" \
"ldp q24, q25, ["X", #128] \n" \
"ldp q26, q27, ["X", #160] \n" \
"fabs v20.2d, v20.2d \n" \
"fabs v21.2d, v21.2d \n" \
"fabs v22.2d, v22.2d \n" \
"fabs v23.2d, v23.2d \n" \
"fadd v16.2d, v16.2d, v17.2d \n" \
"fadd v18.2d, v18.2d, v19.2d \n" \
"ldp q28, q29, ["X", #192] \n" \
"ldp q30, q31, ["X", #224] \n" \
"fabs v24.2d, v24.2d \n" \
"fabs v25.2d, v25.2d \n" \
"fabs v26.2d, v26.2d \n" \
"fabs v27.2d, v27.2d \n" \
"add "X", "X", #256 \n" \
"fadd v20.2d, v20.2d, v21.2d \n" \
"fadd v22.2d, v22.2d, v23.2d \n" \
"fabs v28.2d, v28.2d \n" \
"fabs v29.2d, v29.2d \n" \
"fabs v30.2d, v30.2d \n" \
"fabs v31.2d, v31.2d \n" \
"PRFM PLDL1KEEP, ["X", #1024] \n" \
"PRFM PLDL1KEEP, ["X", #1024+64] \n" \
"fadd v24.2d, v24.2d, v25.2d \n" \
"fadd v26.2d, v26.2d, v27.2d \n" \
"fadd v28.2d, v28.2d, v29.2d \n" \
"fadd v30.2d, v30.2d, v31.2d \n" \
"fadd v0.2d, v0.2d, v16.2d \n" \
"fadd v1.2d, v1.2d, v18.2d \n" \
"fadd v2.2d, v2.2d, v20.2d \n" \
"fadd v3.2d, v3.2d, v22.2d \n" \
"PRFM PLDL1KEEP, ["X", #1024+128] \n" \
"PRFM PLDL1KEEP, ["X", #1024+192] \n" \
"fadd v4.2d, v4.2d, v24.2d \n" \
"fadd v5.2d, v5.2d, v26.2d \n" \
"fadd v6.2d, v6.2d, v28.2d \n" \
"fadd v7.2d, v7.2d, v30.2d \n"
#define KERNEL_F16_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 "SUMF", v0.2d \n"
#define INIT_S \
"lsl "INC_X", "INC_X", #4 \n"
#define KERNEL_S1 \
"ldr q1, ["X"] \n" \
"add "X", "X", "INC_X" \n" \
"fabs v1.2d, v1.2d \n" \
"faddp d1, v1.2d \n" \
"fadd "SUMF", "SUMF", d1 \n"
#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 zasum_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x)
{
FLOAT asum = 0.0 ;
if ( n < 0 ) return(asum);
__asm__ __volatile__ (
" mov "N", %[N_] \n"
" mov "X", %[X_] \n"
" mov "INC_X", %[INCX_] \n"
" fmov "SUMF", "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 //asum_kernel_L999 \n"
" cmp "INC_X", xzr \n"
" ble 9f //asum_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne 5f //asum_kernel_S_BEGIN \n"
"1: //asum_kernel_F_BEGIN: \n"
" asr "J", "N", #4 \n"
" cmp "J", xzr \n"
" beq 3f //asum_kernel_F1 \n"
#if !(defined(__clang__) && defined(OS_WINDOWS))
".align 5 \n"
#endif
"2: //asum_kernel_F16: \n"
" "KERNEL_F16" \n"
" subs "J", "J", #1 \n"
" bne 2b //asum_kernel_F16 \n"
" "KERNEL_F16_FINALIZE" \n"
"3: //asum_kernel_F1: \n"
" ands "J", "N", #15 \n"
" ble 9f //asum_kernel_L999 \n"
"4: //asum_kernel_F10: \n"
" "KERNEL_F1" \n"
" subs "J", "J", #1 \n"
" bne 4b //asum_kernel_F10 \n"
" b 9f //asum_kernel_L999 \n"
"5: //asum_kernel_S_BEGIN: \n"
" "INIT_S" \n"
" asr "J", "N", #2 \n"
" cmp "J", xzr \n"
" ble 7f //asum_kernel_S1 \n"
"6: //asum_kernel_S4: \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 6b //asum_kernel_S4 \n"
"7: //asum_kernel_S1: \n"
" ands "J", "N", #3 \n"
" ble 9f //asum_kernel_L999 \n"
"8: //asum_kernel_S10: \n"
" "KERNEL_S1" \n"
" subs "J", "J", #1 \n"
" bne 8b //asum_kernel_S10 \n"
"9: //asum_kernel_L999: \n"
" fmov %[ASUM_], "SUMF" \n"
: [ASUM_] "=r" (asum) //%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 asum;
}
#if defined(SMP)
static int zasum_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 = zasum_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 asum = 0.0;
#if defined(SMP)
if (inc_x == 0 || n <= 10000)
nthreads = 1;
else
nthreads = num_cpu_avail(1);
if (nthreads == 1) {
asum = zasum_compute(n, x, inc_x);
} else {
int mode, i;
char result[MAX_CPU_NUMBER * sizeof(double) * 2];
FLOAT *ptr;
mode = BLAS_DOUBLE | BLAS_COMPLEX;
blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
x, inc_x, NULL, 0, result, 0,
( void *)zasum_thread_function, nthreads);
ptr = (FLOAT *)result;
for (i = 0; i < nthreads; i++) {
asum = asum + (*ptr);
ptr = (FLOAT *)(((char *)ptr) + sizeof(double) * 2);
}
}
#else
asum = zasum_compute(n, x, inc_x);
#endif
return asum;
}

Write Preview
Loading…
Cancel
Save