floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Remove unnecessary instructions from Advanced SIMD dot 

The existing kernel was issuing extra instructions to organise the arguments into the same registers they would usually be in and similarly to put the result into the appropriate register.

This has an impact on smaller sized dots and seemed like a quick fix
This commit is contained in:
Chris Sidebottom 2022-11-25 15:24:32 +00:00
parent e9a911fb9f
commit 4f7b77e08a
1 changed files with 117 additions and 128 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

245
kernel/arm64/dot_thunderx2t99.c
View File

@ -1,5 +1,6 @@
/*************************************************************************** /***************************************************************************
Copyright (c) 2017, The OpenBLAS Project Copyright (c) 2017, The OpenBLAS Project
Copyright (c) 2022, Arm Ltd
All rights reserved. All rights reserved.
Redistribution and use in source and binary forms, with or without Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are modification, are permitted provided that the following conditions are
@ -36,25 +37,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define RETURN_TYPE double #define RETURN_TYPE double
#endif #endif
#define N "x0" /* vector length */
#define X "x1" /* "X" vector address */
#define INC_X "x2" /* "X" stride */
#define Y "x3" /* "Y" vector address */
#define INC_Y "x4" /* "Y" stride */
#define J "x5" /* loop variable */
#if !defined(DOUBLE) #if !defined(DOUBLE)
#if !defined(DSDOT) #if !defined(DSDOT)
#define DOT_MOD "s"
#define REG0 "wzr" #define REG0 "wzr"
#define DOTF "s0"
#define TMPX "s16" #define TMPX "s16"
#define TMPY "s24" #define TMPY "s24"
#define INC_SHIFT "2" #define INC_SHIFT "2"
#define N_DIV_SHIFT "6" #define N_DIV_SHIFT "6"
#define N_REM_MASK "63" #define N_REM_MASK "63"
#else #else
#define DOT_MOD "d"
#define REG0 "xzr" #define REG0 "xzr"
#define DOTF "d0"
#define TMPX "s16" #define TMPX "s16"
#define TMPX1 "d2" #define TMPX1 "d2"
#define TMPY "s24" #define TMPY "s24"
@ -64,8 +58,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define N_REM_MASK "15" #define N_REM_MASK "15"
#endif #endif
#else #else
#define DOT_MOD "d"
#define REG0 "xzr" #define REG0 "xzr"
#define DOTF "d0"
#define TMPX "d16" #define TMPX "d16"
#define TMPY "d24" #define TMPY "d24"
#define INC_SHIFT "3" #define INC_SHIFT "3"
@ -73,59 +67,61 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#define N_REM_MASK "31" #define N_REM_MASK "31"
#endif #endif
#define OUT "%"DOT_MOD"[DOT_]"
#if !defined(DOUBLE) #if !defined(DOUBLE)
#if !defined(DSDOT) #if !defined(DSDOT)
#define KERNEL_F1 \ #define KERNEL_F1 \
" ldr "TMPX", ["X"] \n" \ " ldr "TMPX", [%[X_]] \n" \
" ldr "TMPY", ["Y"] \n" \ " ldr "TMPY", [%[Y_]] \n" \
" add "X", "X", "INC_X" \n" \ " add %[X_], %[X_], %[INCX_] \n" \
" add "Y", "Y", "INC_Y" \n" \ " add %[Y_], %[Y_], %[INCY_] \n" \
" fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" " fmadd "OUT", "TMPX", "TMPY", "OUT" \n"
#define KERNEL_F \ #define KERNEL_F \
" ldp q16, q17, ["X"] \n" \ " ldp q16, q17, [%[X_]] \n" \
" ldp q24, q25, ["Y"] \n" \ " ldp q24, q25, [%[Y_]] \n" \
" ldp q18, q19, ["X", #32] \n" \ " ldp q18, q19, [%[X_], #32] \n" \
" ldp q26, q27, ["Y", #32] \n" \ " ldp q26, q27, [%[Y_], #32] \n" \
" fmla v0.4s, v16.4s, v24.4s \n" \ " fmla v0.4s, v16.4s, v24.4s \n" \
" fmla v1.4s, v17.4s, v25.4s \n" \ " fmla v1.4s, v17.4s, v25.4s \n" \
" ldp q20, q21, ["X", #64] \n" \ " ldp q20, q21, [%[X_], #64] \n" \
" ldp q28, q29, ["Y", #64] \n" \ " ldp q28, q29, [%[Y_], #64] \n" \
" fmla v2.4s, v18.4s, v26.4s \n" \ " fmla v2.4s, v18.4s, v26.4s \n" \
" fmla v3.4s, v19.4s, v27.4s \n" \ " fmla v3.4s, v19.4s, v27.4s \n" \
" ldp q22, q23, ["X", #96] \n" \ " ldp q22, q23, [%[X_], #96] \n" \
" ldp q30, q31, ["Y", #96] \n" \ " ldp q30, q31, [%[Y_], #96] \n" \
" add "Y", "Y", #128 \n" \ " add %[Y_], %[Y_], #128 \n" \
" add "X", "X", #128 \n" \ " add %[X_], %[X_], #128 \n" \
" fmla v4.4s, v20.4s, v28.4s \n" \ " fmla v4.4s, v20.4s, v28.4s \n" \
" fmla v5.4s, v21.4s, v29.4s \n" \ " fmla v5.4s, v21.4s, v29.4s \n" \
" PRFM PLDL1KEEP, ["X", #896] \n" \ " PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, ["Y", #896] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, ["X", #896+64] \n" \ " PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, ["Y", #896+64] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v6.4s, v22.4s, v30.4s \n" \ " fmla v6.4s, v22.4s, v30.4s \n" \
" fmla v7.4s, v23.4s, v31.4s \n" \ " fmla v7.4s, v23.4s, v31.4s \n" \
" ldp q16, q17, ["X"] \n" \ " ldp q16, q17, [%[X_]] \n" \
" ldp q24, q25, ["Y"] \n" \ " ldp q24, q25, [%[Y_]] \n" \
" ldp q18, q19, ["X", #32] \n" \ " ldp q18, q19, [%[X_], #32] \n" \
" ldp q26, q27, ["Y", #32] \n" \ " ldp q26, q27, [%[Y_], #32] \n" \
" fmla v0.4s, v16.4s, v24.4s \n" \ " fmla v0.4s, v16.4s, v24.4s \n" \
" fmla v1.4s, v17.4s, v25.4s \n" \ " fmla v1.4s, v17.4s, v25.4s \n" \
" ldp q20, q21, ["X", #64] \n" \ " ldp q20, q21, [%[X_], #64] \n" \
" ldp q28, q29, ["Y", #64] \n" \ " ldp q28, q29, [%[Y_], #64] \n" \
" fmla v2.4s, v18.4s, v26.4s \n" \ " fmla v2.4s, v18.4s, v26.4s \n" \
" fmla v3.4s, v19.4s, v27.4s \n" \ " fmla v3.4s, v19.4s, v27.4s \n" \
" ldp q22, q23, ["X", #96] \n" \ " ldp q22, q23, [%[X_], #96] \n" \
" ldp q30, q31, ["Y", #96] \n" \ " ldp q30, q31, [%[Y_], #96] \n" \
" add "Y", "Y", #128 \n" \ " add %[Y_], %[Y_], #128 \n" \
" add "X", "X", #128 \n" \ " add %[X_], %[X_], #128 \n" \
" fmla v4.4s, v20.4s, v28.4s \n" \ " fmla v4.4s, v20.4s, v28.4s \n" \
" fmla v5.4s, v21.4s, v29.4s \n" \ " fmla v5.4s, v21.4s, v29.4s \n" \
" PRFM PLDL1KEEP, ["X", #896] \n" \ " PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, ["Y", #896] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, ["X", #896+64] \n" \ " PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, ["Y", #896+64] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v6.4s, v22.4s, v30.4s \n" \ " fmla v6.4s, v22.4s, v30.4s \n" \
" fmla v7.4s, v23.4s, v31.4s \n" " fmla v7.4s, v23.4s, v31.4s \n"
@ -142,19 +138,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#else /* !defined(DSDOT) */ #else /* !defined(DSDOT) */
#define KERNEL_F1 \ #define KERNEL_F1 \
" ldr "TMPX", ["X"] \n" \ " ldr "TMPX", [%[X_]] \n" \
" ldr "TMPY", ["Y"] \n" \ " ldr "TMPY", [%[Y_]] \n" \
" add "X", "X", "INC_X" \n" \ " add %[X_], %[X_], %[INCX_] \n" \
" add "Y", "Y", "INC_Y" \n" \ " add %[Y_], %[Y_], %[INCY_] \n" \
" fcvt "TMPX1", "TMPX" \n" \ " fcvt "TMPX1", "TMPX" \n" \
" fcvt "TMPY1", "TMPY" \n" \ " fcvt "TMPY1", "TMPY" \n" \
" fmul "TMPX1", "TMPX1", "TMPY1" \n" \ " fmul "TMPX1", "TMPX1", "TMPY1" \n" \
" fadd "DOTF", "DOTF", "TMPX1" \n" " fadd "OUT", "OUT", "TMPX1" \n"
#define KERNEL_F \ #define KERNEL_F \
" ldp q18, q19, ["X"] \n" \ " ldp q18, q19, [%[X_]] \n" \
" ldp q26, q27, ["Y"] \n" \ " ldp q26, q27, [%[Y_]] \n" \
" fcvtl v16.2d, v18.2s \n" \ " fcvtl v16.2d, v18.2s \n" \
" fcvtl2 v17.2d, v18.4s \n" \ " fcvtl2 v17.2d, v18.4s \n" \
" fcvtl v18.2d, v19.2s \n" \ " fcvtl v18.2d, v19.2s \n" \
@ -163,8 +159,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
" fcvtl2 v25.2d, v26.4s \n" \ " fcvtl2 v25.2d, v26.4s \n" \
" fcvtl v26.2d, v27.2s \n" \ " fcvtl v26.2d, v27.2s \n" \
" fcvtl2 v27.2d, v27.4s \n" \ " fcvtl2 v27.2d, v27.4s \n" \
" ldp q22, q23, ["X", #32] \n" \ " ldp q22, q23, [%[X_], #32] \n" \
" ldp q30, q31, ["Y", #32] \n" \ " ldp q30, q31, [%[Y_], #32] \n" \
" fcvtl v20.2d, v22.2s \n" \ " fcvtl v20.2d, v22.2s \n" \
" fcvtl2 v21.2d, v22.4s \n" \ " fcvtl2 v21.2d, v22.4s \n" \
" fcvtl v22.2d, v23.2s \n" \ " fcvtl v22.2d, v23.2s \n" \
@ -173,16 +169,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
" fcvtl2 v29.2d, v30.4s \n" \ " fcvtl2 v29.2d, v30.4s \n" \
" fcvtl v30.2d, v31.2s \n" \ " fcvtl v30.2d, v31.2s \n" \
" fcvtl2 v31.2d, v31.4s \n" \ " fcvtl2 v31.2d, v31.4s \n" \
" PRFM PLDL1KEEP, ["X", #896] \n" \ " PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, ["Y", #896] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, ["X", #896+64] \n" \ " PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, ["Y", #896+64] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v0.2d, v16.2d, v24.2d \n" \ " fmla v0.2d, v16.2d, v24.2d \n" \
" fmla v1.2d, v17.2d, v25.2d \n" \ " fmla v1.2d, v17.2d, v25.2d \n" \
" fmla v2.2d, v18.2d, v26.2d \n" \ " fmla v2.2d, v18.2d, v26.2d \n" \
" fmla v3.2d, v19.2d, v27.2d \n" \ " fmla v3.2d, v19.2d, v27.2d \n" \
" add "Y", "Y", #64 \n" \ " add %[Y_], %[Y_], #64 \n" \
" add "X", "X", #64 \n" \ " add %[X_], %[X_], #64 \n" \
" fmla v4.2d, v20.2d, v28.2d \n" \ " fmla v4.2d, v20.2d, v28.2d \n" \
" fmla v5.2d, v21.2d, v29.2d \n" \ " fmla v5.2d, v21.2d, v29.2d \n" \
" fmla v6.2d, v22.2d, v30.2d \n" \ " fmla v6.2d, v22.2d, v30.2d \n" \
@ -196,60 +192,60 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
" fadd v0.2d, v0.2d, v2.2d \n" \ " fadd v0.2d, v0.2d, v2.2d \n" \
" fadd v4.2d, v4.2d, v6.2d \n" \ " fadd v4.2d, v4.2d, v6.2d \n" \
" fadd v0.2d, v0.2d, v4.2d \n" \ " fadd v0.2d, v0.2d, v4.2d \n" \
" faddp "DOTF", v0.2d \n" " faddp "OUT", v0.2d \n"
#endif /* !defined(DSDOT) */ #endif /* !defined(DSDOT) */
#else /* !defined(DOUBLE) */ #else /* !defined(DOUBLE) */
#define KERNEL_F1 \ #define KERNEL_F1 \
" ldr "TMPX", ["X"] \n" \ " ldr "TMPX", [%[X_]] \n" \
" ldr "TMPY", ["Y"] \n" \ " ldr "TMPY", [%[Y_]] \n" \
" add "X", "X", "INC_X" \n" \ " add %[X_], %[X_], %[INCX_] \n" \
" add "Y", "Y", "INC_Y" \n" \ " add %[Y_], %[Y_], %[INCY_] \n" \
" fmadd "DOTF", "TMPX", "TMPY", "DOTF" \n" " fmadd "OUT", "TMPX", "TMPY", "OUT" \n"
#define KERNEL_F \ #define KERNEL_F \
" ldp q16, q17, ["X"] \n" \ " ldp q16, q17, [%[X_]] \n" \
" ldp q24, q25, ["Y"] \n" \ " ldp q24, q25, [%[Y_]] \n" \
" ldp q18, q19, ["X", #32] \n" \ " ldp q18, q19, [%[X_], #32] \n" \
" ldp q26, q27, ["Y", #32] \n" \ " ldp q26, q27, [%[Y_], #32] \n" \
" fmla v0.2d, v16.2d, v24.2d \n" \ " fmla v0.2d, v16.2d, v24.2d \n" \
" fmla v1.2d, v17.2d, v25.2d \n" \ " fmla v1.2d, v17.2d, v25.2d \n" \
" ldp q20, q21, ["X", #64] \n" \ " ldp q20, q21, [%[X_], #64] \n" \
" ldp q28, q29, ["Y", #64] \n" \ " ldp q28, q29, [%[Y_], #64] \n" \
" fmla v2.2d, v18.2d, v26.2d \n" \ " fmla v2.2d, v18.2d, v26.2d \n" \
" fmla v3.2d, v19.2d, v27.2d \n" \ " fmla v3.2d, v19.2d, v27.2d \n" \
" ldp q22, q23, ["X", #96] \n" \ " ldp q22, q23, [%[X_], #96] \n" \
" ldp q30, q31, ["Y", #96] \n" \ " ldp q30, q31, [%[Y_], #96] \n" \
" add "Y", "Y", #128 \n" \ " add %[Y_], %[Y_], #128 \n" \
" add "X", "X", #128 \n" \ " add %[X_], %[X_], #128 \n" \
" fmla v4.2d, v20.2d, v28.2d \n" \ " fmla v4.2d, v20.2d, v28.2d \n" \
" fmla v5.2d, v21.2d, v29.2d \n" \ " fmla v5.2d, v21.2d, v29.2d \n" \
" PRFM PLDL1KEEP, ["X", #896] \n" \ " PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, ["Y", #896] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, ["X", #896+64] \n" \ " PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, ["Y", #896+64] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v6.2d, v22.2d, v30.2d \n" \ " fmla v6.2d, v22.2d, v30.2d \n" \
" fmla v7.2d, v23.2d, v31.2d \n" \ " fmla v7.2d, v23.2d, v31.2d \n" \
" ldp q16, q17, ["X"] \n" \ " ldp q16, q17, [%[X_]] \n" \
" ldp q24, q25, ["Y"] \n" \ " ldp q24, q25, [%[Y_]] \n" \
" ldp q18, q19, ["X", #32] \n" \ " ldp q18, q19, [%[X_], #32] \n" \
" ldp q26, q27, ["Y", #32] \n" \ " ldp q26, q27, [%[Y_], #32] \n" \
" fmla v0.2d, v16.2d, v24.2d \n" \ " fmla v0.2d, v16.2d, v24.2d \n" \
" fmla v1.2d, v17.2d, v25.2d \n" \ " fmla v1.2d, v17.2d, v25.2d \n" \
" ldp q20, q21, ["X", #64] \n" \ " ldp q20, q21, [%[X_], #64] \n" \
" ldp q28, q29, ["Y", #64] \n" \ " ldp q28, q29, [%[Y_], #64] \n" \
" fmla v2.2d, v18.2d, v26.2d \n" \ " fmla v2.2d, v18.2d, v26.2d \n" \
" fmla v3.2d, v19.2d, v27.2d \n" \ " fmla v3.2d, v19.2d, v27.2d \n" \
" ldp q22, q23, ["X", #96] \n" \ " ldp q22, q23, [%[X_], #96] \n" \
" ldp q30, q31, ["Y", #96] \n" \ " ldp q30, q31, [%[Y_], #96] \n" \
" add "Y", "Y", #128 \n" \ " add %[Y_], %[Y_], #128 \n" \
" add "X", "X", #128 \n" \ " add %[X_], %[X_], #128 \n" \
" fmla v4.2d, v20.2d, v28.2d \n" \ " fmla v4.2d, v20.2d, v28.2d \n" \
" fmla v5.2d, v21.2d, v29.2d \n" \ " fmla v5.2d, v21.2d, v29.2d \n" \
" PRFM PLDL1KEEP, ["X", #896] \n" \ " PRFM PLDL1KEEP, [%[X_], #896] \n" \
" PRFM PLDL1KEEP, ["Y", #896] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896] \n" \
" PRFM PLDL1KEEP, ["X", #896+64] \n" \ " PRFM PLDL1KEEP, [%[X_], #896+64] \n" \
" PRFM PLDL1KEEP, ["Y", #896+64] \n" \ " PRFM PLDL1KEEP, [%[Y_], #896+64] \n" \
" fmla v6.2d, v22.2d, v30.2d \n" \ " fmla v6.2d, v22.2d, v30.2d \n" \
" fmla v7.2d, v23.2d, v31.2d \n" " fmla v7.2d, v23.2d, v31.2d \n"
@ -261,7 +257,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
" fadd v0.2d, v0.2d, v2.2d \n" \ " fadd v0.2d, v0.2d, v2.2d \n" \
" fadd v4.2d, v4.2d, v6.2d \n" \ " fadd v4.2d, v4.2d, v6.2d \n" \
" fadd v0.2d, v0.2d, v4.2d \n" \ " fadd v0.2d, v0.2d, v4.2d \n" \
" faddp "DOTF", v0.2d \n" " faddp "OUT", v0.2d \n"
#endif /* !defined(DOUBLE) */ #endif /* !defined(DOUBLE) */
#if defined(SMP) #if defined(SMP)
@ -272,17 +268,14 @@ extern int blas_level1_thread_with_return_value(int mode, BLASLONG m, BLASLONG n
static RETURN_TYPE dot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y) static RETURN_TYPE dot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
{ {
RETURN_TYPE dot = 0.0 ; RETURN_TYPE dot = 0.0;
if ( n < 0 ) return dot; if ( n < 0 ) return dot;
BLASLONG j = 0;
__asm__ __volatile__ ( __asm__ __volatile__ (
" mov "N", %[N_] \n" " fmov "OUT", "REG0" \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, xzr \n" " fmov d1, xzr \n"
" fmov d2, xzr \n" " fmov d2, xzr \n"
" fmov d3, xzr \n" " fmov d3, xzr \n"
@ -290,42 +283,40 @@ static RETURN_TYPE dot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, B
" fmov d5, xzr \n" " fmov d5, xzr \n"
" fmov d6, xzr \n" " fmov d6, xzr \n"
" fmov d7, xzr \n" " fmov d7, xzr \n"
" cmp "N", xzr \n" " cmp %[INCX_], #1 \n"
" ble 9f //dot_kernel_L999 \n"
" cmp "INC_X", #1 \n"
" bne 5f //dot_kernel_S_BEGIN \n" " bne 5f //dot_kernel_S_BEGIN \n"
" cmp "INC_Y", #1 \n" " cmp %[INCY_], #1 \n"
" bne 5f //dot_kernel_S_BEGIN \n" " bne 5f //dot_kernel_S_BEGIN \n"
"1: //dot_kernel_F_BEGIN: \n" "1: //dot_kernel_F_BEGIN: \n"
" lsl "INC_X", "INC_X", "INC_SHIFT" \n" " lsl %[INCX_], %[INCX_], "INC_SHIFT" \n"
" lsl "INC_Y", "INC_Y", "INC_SHIFT" \n" " lsl %[INCY_], %[INCY_], "INC_SHIFT" \n"
" asr "J", "N", #"N_DIV_SHIFT" \n" " asr %[J_], %[N_], #"N_DIV_SHIFT" \n"
" cmp "J", xzr \n" " cmp %[J_], xzr \n"
" beq 3f //dot_kernel_F1 \n" " beq 3f //dot_kernel_F1 \n"
" .align 5 \n" " .align 5 \n"
"2: //dot_kernel_F: \n" "2: //dot_kernel_F: \n"
" "KERNEL_F" \n" " "KERNEL_F" \n"
" subs "J", "J", #1 \n" " subs %[J_], %[J_], #1 \n"
" bne 2b //dot_kernel_F \n" " bne 2b //dot_kernel_F \n"
" "KERNEL_F_FINALIZE" \n" " "KERNEL_F_FINALIZE" \n"
"3: //dot_kernel_F1: \n" "3: //dot_kernel_F1: \n"
" ands "J", "N", #"N_REM_MASK" \n" " ands %[J_], %[N_], #"N_REM_MASK" \n"
" ble 9f //dot_kernel_L999 \n" " ble 9f //dot_kernel_L999 \n"
"4: //dot_kernel_F10: \n" "4: //dot_kernel_F10: \n"
" "KERNEL_F1" \n" " "KERNEL_F1" \n"
" subs "J", "J", #1 \n" " subs %[J_], %[J_], #1 \n"
" bne 4b //dot_kernel_F10 \n" " bne 4b //dot_kernel_F10 \n"
" b 9f //dot_kernel_L999 \n" " b 9f //dot_kernel_L999 \n"
"5: //dot_kernel_S_BEGIN: \n" "5: //dot_kernel_S_BEGIN: \n"
" lsl "INC_X", "INC_X", "INC_SHIFT" \n" " lsl %[INCX_], %[INCX_], "INC_SHIFT" \n"
" lsl "INC_Y", "INC_Y", "INC_SHIFT" \n" " lsl %[INCY_], %[INCY_], "INC_SHIFT" \n"
" asr "J", "N", #2 \n" " asr %[J_], %[N_], #2 \n"
" cmp "J", xzr \n" " cmp %[J_], xzr \n"
" ble 7f //dot_kernel_S1 \n" " ble 7f //dot_kernel_S1 \n"
"6: //dot_kernel_S4: \n" "6: //dot_kernel_S4: \n"
@ -333,32 +324,30 @@ static RETURN_TYPE dot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, B
" "KERNEL_F1" \n" " "KERNEL_F1" \n"
" "KERNEL_F1" \n" " "KERNEL_F1" \n"
" "KERNEL_F1" \n" " "KERNEL_F1" \n"
" subs "J", "J", #1 \n" " subs %[J_], %[J_], #1 \n"
" bne 6b //dot_kernel_S4 \n" " bne 6b //dot_kernel_S4 \n"
"7: //dot_kernel_S1: \n" "7: //dot_kernel_S1: \n"
" ands "J", "N", #3 \n" " ands %[J_], %[N_], #3 \n"
" ble 9f //dot_kernel_L999 \n" " ble 9f //dot_kernel_L999 \n"
"8: //dot_kernel_S10: \n" "8: //dot_kernel_S10: \n"
" "KERNEL_F1" \n" " "KERNEL_F1" \n"
" subs "J", "J", #1 \n" " subs %[J_], %[J_], #1 \n"
" bne 8b //dot_kernel_S10 \n" " bne 8b //dot_kernel_S10 \n"
"9: //dot_kernel_L999: \n" "9: //dot_kernel_L999: \n"
" str "DOTF", [%[DOT_]] \n"
: : [DOT_] "=&w" (dot)
: [DOT_] "r" (&dot), //%0 : [N_] "r" (n),
[N_] "r" (n), //%1 [X_] "r" (x),
[X_] "r" (x), //%2 [INCX_] "r" (inc_x),
[INCX_] "r" (inc_x), //%3 [Y_] "r" (y),
[Y_] "r" (y), //%4 [INCY_] "r" (inc_y),
[INCY_] "r" (inc_y) //%5 [J_] "r" (j)
: "cc", : "cc",
"memory", "memory",
"x0", "x1", "x2", "x3", "x4", "x5", "d1", "d2", "d3", "d4", "d5", "d6", "d7"
"d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7"
); );
return dot; return dot;