From cfa2bba98169cb5ab1945462514be1b6badf7d98 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Thu, 27 Jun 2019 17:00:52 +0100 Subject: COMPMID-2178: Update GEMM assembly code. Perform offset reduction and requantization within the assembly wrapper. Change-Id: I5d5b3e1f6f9ef4c71805362c57f88ff199c027a3 Signed-off-by: Georgios Pinitas Reviewed-on: https://review.mlplatform.org/c/1541 Comments-Addressed: Pablo Marquez Reviewed-by: Gian Marco Iodice Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins --- src/core/NEON/kernels/arm_gemm/merges/list.hpp | 5 +- .../arm_gemm/merges/sve_merge_fp32_3VLx8.hpp | 38 +- .../arm_gemm/merges/sve_merge_s32_3VLx8.hpp | 1348 ++++++++++++++++++++ .../arm_gemm/merges/sve_merge_u32_3VLx8.hpp | 1348 ++++++++++++++++++++ 4 files changed, 2718 insertions(+), 21 deletions(-) create mode 100644 src/core/NEON/kernels/arm_gemm/merges/sve_merge_s32_3VLx8.hpp create mode 100644 src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp (limited to 'src/core/NEON/kernels/arm_gemm/merges') diff --git a/src/core/NEON/kernels/arm_gemm/merges/list.hpp b/src/core/NEON/kernels/arm_gemm/merges/list.hpp index 181d1a4e3f..788a957d5e 100644 --- a/src/core/NEON/kernels/arm_gemm/merges/list.hpp +++ b/src/core/NEON/kernels/arm_gemm/merges/list.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -26,5 +26,6 @@ #include "a64_merge_float_to_half_12x8.hpp" #include "a64_merge_half_24x8.hpp" #include "a64_merge_int32_12x8.hpp" -#include "sve_merge_fp32_2VLx8.hpp" #include "sve_merge_fp32_3VLx8.hpp" +#include "sve_merge_s32_3VLx8.hpp" +#include "sve_merge_u32_3VLx8.hpp" diff --git a/src/core/NEON/kernels/arm_gemm/merges/sve_merge_fp32_3VLx8.hpp b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_fp32_3VLx8.hpp index 27084c3598..4b51066ebe 100644 --- a/src/core/NEON/kernels/arm_gemm/merges/sve_merge_fp32_3VLx8.hpp +++ b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_fp32_3VLx8.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 Arm Limited. + * Copyright (c) 2018-2019 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -84,7 +84,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 2: { long w = xmax - i; @@ -135,7 +135,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 3: { long w = xmax - i; @@ -198,7 +198,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 4: { long w = xmax - i; @@ -272,7 +272,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 5: { long w = xmax - i; @@ -358,7 +358,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 6: { long w = xmax - i; @@ -456,7 +456,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 7: { long w = xmax - i; @@ -566,7 +566,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + default: case 8: { @@ -688,8 +688,8 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - - + + } } else @@ -739,7 +739,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 2: { long w = xmax - i; @@ -802,7 +802,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 3: { long w = xmax - i; @@ -883,7 +883,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 4: { long w = xmax - i; @@ -981,7 +981,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 5: { long w = xmax - i; @@ -1097,7 +1097,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 6: { long w = xmax - i; @@ -1231,7 +1231,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + case 7: { long w = xmax - i; @@ -1383,7 +1383,7 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - + default: case 8: { @@ -1553,8 +1553,8 @@ inline void MergeResults<3, 8, true>(float *out, const float *in, const int ldou ); } break; - - + + } } } diff --git a/src/core/NEON/kernels/arm_gemm/merges/sve_merge_s32_3VLx8.hpp b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_s32_3VLx8.hpp new file mode 100644 index 0000000000..2ea38a78df --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_s32_3VLx8.hpp @@ -0,0 +1,1348 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#pragma once + +#ifdef __ARM_FEATURE_SVE + +template<> +inline void MergeResults<3, 8, true>(int32_t *out, const int32_t *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const int32_t alpha, const int32_t beta) +{ + const int32_t *inptr = in; + + for (int y=y0; y())) { + if (beta==0) + { + switch(height) { + case 1: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 2: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" + "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 3: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" + "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 4: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 5: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-6, MUL VL]\n" + "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-3, MUL VL]\n" + "st1w z10.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr4], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" + "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 6: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #7, MUL VL]\n" + "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-3, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [x8]\n" + "st1w z8.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr5], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 7: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "st1w z10.s, p0, [%[outptr6]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" + "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-3, MUL VL]\n" + "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [x8]\n" + "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #3, MUL VL]\n" + "st1w z10.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr5], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr6], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z11.s, p0/z, [x8, #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "ld1w z8.s, p0/z, [x8, #4, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" + "st1w z11.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "st1w z8.s, p0, [%[outptr6], #2, MUL VL]\n" + "addvl %[outptr6], %[outptr6], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + default: + case 8: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "ld1w z11.s, p0/z, [x8, #5, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "st1w z10.s, p0, [%[outptr6]]\n" + "st1w z11.s, p0, [%[outptr7]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-3, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8]\n" + "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #3, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr5], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr6], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr7], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "ld1w z10.s, p0/z, [x8, #4, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" + "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "ld1w z11.s, p0/z, [x8, #7, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr7], #0x60]\n" + "st1w z10.s, p0, [%[outptr6], #2, MUL VL]\n" + "addvl %[outptr6], %[outptr6], #3\n" + "st1w z11.s, p0, [%[outptr7], #2, MUL VL]\n" + "addvl %[outptr7], %[outptr7], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + + } + } + else + { + switch(height) { + case 1: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "ld1w z5.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 2: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z6.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 3: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z7.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z7.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z4.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 4: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z4.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 5: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z5.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z8.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr4], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-3, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z10.s, p0, [%[outptr1], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr4], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z7.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z6.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 6: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z5.s, p0/z, [%[outptr5]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z6.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z8.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z9.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr4], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-3, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z7.s, p0/z, [%[outptr5], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [x8]\n" + "st1w z8.s, p0, [%[outptr2], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr5], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z4.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr5], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 7: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z5.s, p0/z, [%[outptr5]]\n" + "ld1w z6.s, p0/z, [%[outptr6]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "st1w z10.s, p0, [%[outptr6]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z7.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr4], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-3, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z4.s, p0/z, [%[outptr5], #1, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr6], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z8.s, p0/z, [x8]\n" + "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8, #3, MUL VL]\n" + "st1w z10.s, p0, [%[outptr3], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr5], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr6], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z7.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z6.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z7.s, p0/z, [%[outptr5], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #1, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z4.s, p0/z, [%[outptr6], #2, MUL VL]\n" + "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z8.s, p0/z, [x8, #4, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "st1w z8.s, p0, [%[outptr6], #2, MUL VL]\n" + "addvl %[outptr6], %[outptr6], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + default: + case 8: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z5.s, p0/z, [%[outptr5]]\n" + "ld1w z6.s, p0/z, [%[outptr6]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr7]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [x8, #5, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "st1w z10.s, p0, [%[outptr6]]\n" + "st1w z11.s, p0, [%[outptr7]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z4.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-3, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z5.s, p0/z, [%[outptr5], #1, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr6], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8]\n" + "ld1w z7.s, p0/z, [%[outptr7], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z10.s, p0/z, [x8, #3, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [x8, #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr5], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr6], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr7], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z4.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z5.s, p0/z, [%[outptr5], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z6.s, p0/z, [%[outptr6], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr7], #0x60]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #4, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "ld1w z7.s, p0/z, [%[outptr7], #2, MUL VL]\n" + "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [x8, #7, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr6], #2, MUL VL]\n" + "addvl %[outptr6], %[outptr6], #3\n" + "st1w z11.s, p0, [%[outptr7], #2, MUL VL]\n" + "addvl %[outptr7], %[outptr7], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + + } + } + } + } +} + +#endif // __ARM_FEATURE_SVE diff --git a/src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp new file mode 100644 index 0000000000..eb684e2118 --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/merges/sve_merge_u32_3VLx8.hpp @@ -0,0 +1,1348 @@ +/* + * Copyright (c) 2019 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#pragma once + +#ifdef __ARM_FEATURE_SVE + +template<> +inline void MergeResults<3, 8, true>(uint32_t *out, const uint32_t *in, const int ldout, const int y0, const int ymax, const int x0, const int xmax, const uint32_t alpha, const uint32_t beta) +{ + const uint32_t *inptr = in; + + for (int y=y0; y())) { + if (beta==0u) + { + switch(height) { + case 1: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 2: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" + "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 3: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" + "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 4: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 5: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-6, MUL VL]\n" + "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-3, MUL VL]\n" + "st1w z10.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr4], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" + "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 6: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #7, MUL VL]\n" + "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-3, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [x8]\n" + "st1w z8.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr5], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 7: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "st1w z10.s, p0, [%[outptr6]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" + "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-3, MUL VL]\n" + "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [x8]\n" + "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #3, MUL VL]\n" + "st1w z10.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr5], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr6], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z11.s, p0/z, [x8, #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "ld1w z8.s, p0/z, [x8, #4, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" + "st1w z11.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "st1w z8.s, p0, [%[outptr6], #2, MUL VL]\n" + "addvl %[outptr6], %[outptr6], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + default: + case 8: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "ld1w z11.s, p0/z, [x8, #5, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "st1w z10.s, p0, [%[outptr6]]\n" + "st1w z11.s, p0, [%[outptr7]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" + "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-3, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8]\n" + "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #3, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr5], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr6], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr7], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PSTL1KEEP, [%[outptr0], #0x60]\n" + "prfm PSTL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr2], #0x60]\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr3], #0x60]\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr4], #0x60]\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr5], #0x60]\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "ld1w z10.s, p0/z, [x8, #4, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr6], #0x60]\n" + "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "ld1w z11.s, p0/z, [x8, #7, MUL VL]\n" + "prfm PSTL1KEEP, [%[outptr7], #0x60]\n" + "st1w z10.s, p0, [%[outptr6], #2, MUL VL]\n" + "addvl %[outptr6], %[outptr6], #3\n" + "st1w z11.s, p0, [%[outptr7], #2, MUL VL]\n" + "addvl %[outptr7], %[outptr7], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + + } + } + else + { + switch(height) { + case 1: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "ld1w z5.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 2: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z6.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 3: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z7.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z7.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z4.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 4: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z4.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 5: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z5.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z8.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr4], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-3, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z10.s, p0, [%[outptr1], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr2], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr4], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z7.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z6.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 6: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z5.s, p0/z, [%[outptr5]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z6.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z10.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z8.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z9.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr4], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z10.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-3, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z7.s, p0/z, [%[outptr5], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr1], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [x8]\n" + "st1w z8.s, p0, [%[outptr2], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr3], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr5], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z4.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr5], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + case 7: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z5.s, p0/z, [%[outptr5]]\n" + "ld1w z6.s, p0/z, [%[outptr6]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "st1w z10.s, p0, [%[outptr6]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z7.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z11.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z9.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr4], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-3, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z4.s, p0/z, [%[outptr5], #1, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr6], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr1], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z8.s, p0/z, [x8]\n" + "st1w z9.s, p0, [%[outptr2], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8, #3, MUL VL]\n" + "st1w z10.s, p0, [%[outptr3], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z11.s, p0, [%[outptr4], #1, MUL VL]\n" + "st1w z8.s, p0, [%[outptr5], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr6], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z6.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z10.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z7.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "ld1w z8.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr1], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z6.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr2], #2, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #-2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z7.s, p0/z, [%[outptr5], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr3], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #1, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z4.s, p0/z, [%[outptr6], #2, MUL VL]\n" + "st1w z10.s, p0, [%[outptr4], #2, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z8.s, p0/z, [x8, #4, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "st1w z8.s, p0, [%[outptr6], #2, MUL VL]\n" + "addvl %[outptr6], %[outptr6], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + default: + case 8: + { + long w = xmax - i; + long p = 0; + /* Optimized routine to copy an entire block */ + __asm __volatile ( + "mov z2.s, %s[alpha]\n" + "addvl x8, %[inptr], #16\n" + "mov z3.s, %s[beta]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x180]\n" + "ld1w z4.s, p0/z, [%[outptr0]]\n" + "prfm PLDL1KEEP, [%[inptr], #0x240]\n" + "ld1w z8.s, p0/z, [%[inptr]]\n" + "ld1w z5.s, p0/z, [%[outptr1]]\n" + "ld1w z9.s, p0/z, [%[inptr], #3, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #6, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-7, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4]]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0]]\n" + "ld1w z8.s, p0/z, [x8, #-4, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z5.s, p0/z, [%[outptr5]]\n" + "ld1w z6.s, p0/z, [%[outptr6]]\n" + "st1w z9.s, p0, [%[outptr1]]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8, #-1, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr7]]\n" + "st1w z10.s, p0, [%[outptr2]]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z10.s, p0/z, [x8, #2, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3]]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [x8, #5, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4]]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr5]]\n" + "st1w z10.s, p0, [%[outptr6]]\n" + "st1w z11.s, p0, [%[outptr7]]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "incw %[p], all, mul #1\n" + "prfm PLDL1KEEP, [%[inptr], #0x1c0]\n" + "ld1w z4.s, p0/z, [%[outptr0], #1, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x280]\n" + "ld1w z8.s, p0/z, [%[inptr], #1, MUL VL]\n" + "ld1w z5.s, p0/z, [%[outptr1], #1, MUL VL]\n" + "ld1w z9.s, p0/z, [%[inptr], #4, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z10.s, p0/z, [%[inptr], #7, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z11.s, p0/z, [x8, #-6, MUL VL]\n" + "ld1w z4.s, p0/z, [%[outptr4], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "st1w z8.s, p0, [%[outptr0], #1, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-3, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "ld1w z5.s, p0/z, [%[outptr5], #1, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr6], #1, MUL VL]\n" + "st1w z9.s, p0, [%[outptr1], #1, MUL VL]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [x8]\n" + "ld1w z7.s, p0/z, [%[outptr7], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr2], #1, MUL VL]\n" + "add z9.s, z9.s, z5.s\n" + "ld1w z10.s, p0/z, [x8, #3, MUL VL]\n" + "st1w z11.s, p0, [%[outptr3], #1, MUL VL]\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [x8, #6, MUL VL]\n" + "st1w z8.s, p0, [%[outptr4], #1, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z9.s, p0, [%[outptr5], #1, MUL VL]\n" + "st1w z10.s, p0, [%[outptr6], #1, MUL VL]\n" + "st1w z11.s, p0, [%[outptr7], #1, MUL VL]\n" + "whilelt p0.s, %[p], %[w]\n" + "b.none 1f\n" + "prfm PLDL1KEEP, [%[outptr0], #0x60]\n" + "prfm PLDL1KEEP, [%[outptr1], #0x60]\n" + "ld1w z4.s, p0/z, [%[outptr0], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x200]\n" + "ld1w z8.s, p0/z, [%[inptr], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr2], #0x60]\n" + "ld1w z5.s, p0/z, [%[outptr1], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr3], #0x60]\n" + "add z8.s, z8.s, z4.s\n" + "ld1w z9.s, p0/z, [%[inptr], #5, MUL VL]\n" + "ld1w z6.s, p0/z, [%[outptr2], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr4], #0x60]\n" + "ld1w z10.s, p0/z, [x8, #-8, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr5], #0x60]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr0], #2, MUL VL]\n" + "ld1w z7.s, p0/z, [%[outptr3], #2, MUL VL]\n" + "addvl %[outptr0], %[outptr0], #3\n" + "add z10.s, z10.s, z6.s\n" + "st1w z9.s, p0, [%[outptr1], #2, MUL VL]\n" + "ld1w z11.s, p0/z, [x8, #-5, MUL VL]\n" + "addvl %[outptr1], %[outptr1], #3\n" + "ld1w z4.s, p0/z, [%[outptr4], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[inptr], #0x2c0]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr2], #2, MUL VL]\n" + "ld1w z8.s, p0/z, [x8, #-2, MUL VL]\n" + "addvl %[outptr2], %[outptr2], #3\n" + "ld1w z5.s, p0/z, [%[outptr5], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr6], #0x60]\n" + "add z8.s, z8.s, z4.s\n" + "st1w z11.s, p0, [%[outptr3], #2, MUL VL]\n" + "ld1w z9.s, p0/z, [x8, #1, MUL VL]\n" + "addvl %[outptr3], %[outptr3], #3\n" + "ld1w z6.s, p0/z, [%[outptr6], #2, MUL VL]\n" + "prfm PLDL1KEEP, [%[outptr7], #0x60]\n" + "add z9.s, z9.s, z5.s\n" + "st1w z8.s, p0, [%[outptr4], #2, MUL VL]\n" + "ld1w z10.s, p0/z, [x8, #4, MUL VL]\n" + "addvl %[outptr4], %[outptr4], #3\n" + "ld1w z7.s, p0/z, [%[outptr7], #2, MUL VL]\n" + "st1w z9.s, p0, [%[outptr5], #2, MUL VL]\n" + "addvl %[outptr5], %[outptr5], #3\n" + "add z10.s, z10.s, z6.s\n" + "ld1w z11.s, p0/z, [x8, #7, MUL VL]\n" + "add z11.s, z11.s, z7.s\n" + "st1w z10.s, p0, [%[outptr6], #2, MUL VL]\n" + "addvl %[outptr6], %[outptr6], #3\n" + "st1w z11.s, p0, [%[outptr7], #2, MUL VL]\n" + "addvl %[outptr7], %[outptr7], #3\n" + "1:\n" + "addvl %[inptr], %[inptr], #24\n" + : [outptr0] "+r" (outptr0), [outptr1] "+r" (outptr1), [outptr2] "+r" (outptr2), [outptr3] "+r" (outptr3), [outptr4] "+r" (outptr4), [outptr5] "+r" (outptr5), [outptr6] "+r" (outptr6), [outptr7] "+r" (outptr7), + [inptr] "+r" (inptr), [p] "+r" (p) + : [alpha] "w" (alpha), [beta] "w" (beta), [w] "r" (w) + : "x8", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "memory", "cc" + ); + } + break; + + + } + } + } + } +} + +#endif // __ARM_FEATURE_SVE -- cgit v1.2.1