From e28cf395b57a091d0850cd28cecc81046153b843 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Sun, 31 Jan 2021 05:18:43 +0000 Subject: Regenerate kernels and update A55 versions Signed-off-by: Georgios Pinitas Change-Id: I9eae76c77db03b8806af65729da34ab2d77f95f2 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4965 Tested-by: Arm Jenkins Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- .../a64_interleave8_block2_fp32_fp32.hpp | 74 +++++++++++----------- 1 file changed, 37 insertions(+), 37 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_fp32_fp32.hpp') diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_fp32_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_fp32_fp32.hpp index 3aea6a8999..064207c0fa 100644 --- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_fp32_fp32.hpp +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block2_fp32_fp32.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -80,39 +80,39 @@ void interleave_block<8, 2, VLType::None, false>( "blt 3f\n" "2:" // Main loop head "ldr q27, [x27], #0x10\n" - "prfm pldl1keep, [x27, #0x70]\n" + "subs %x[width], %x[width], #0x4\n" "ldr q24, [x26], #0x10\n" "zip1 v26.2d, v27.2d, v24.2d\n" - "prfm pldl1keep, [x26, #0x70]\n" "ldr q25, [x25], #0x10\n" + "cmp %x[width], #0x4\n" "zip2 v24.2d, v27.2d, v24.2d\n" - "prfm pldl1keep, [x25, #0x70]\n" "ldr q21, [x24], #0x10\n" - "zip1 v23.2d, v25.2d, v21.2d\n" - "prfm pldl1keep, [x24, #0x70]\n" - "ldr q22, [x23], #0x10\n" + "ldr q23, [x23], #0x10\n" + "zip1 v22.2d, v25.2d, v21.2d\n" + "ldr q18, [x22], #0x10\n" "zip2 v21.2d, v25.2d, v21.2d\n" + "ldr q20, [x21], #0x10\n" + "ldr q16, [x20], #0x10\n" + "zip1 v19.2d, v23.2d, v18.2d\n" + "prfm pldl1keep, [x27, #0x70]\n" + "zip2 v18.2d, v23.2d, v18.2d\n" + "prfm pldl1keep, [x26, #0x70]\n" + "zip1 v17.2d, v20.2d, v16.2d\n" + "prfm pldl1keep, [x25, #0x70]\n" + "zip2 v16.2d, v20.2d, v16.2d\n" + "prfm pldl1keep, [x24, #0x70]\n" "prfm pldl1keep, [x23, #0x70]\n" - "ldr q18, [x22], #0x10\n" - "zip1 v20.2d, v22.2d, v18.2d\n" "prfm pldl1keep, [x22, #0x70]\n" - "ldr q19, [x21], #0x10\n" - "zip2 v18.2d, v22.2d, v18.2d\n" "prfm pldl1keep, [x21, #0x70]\n" - "ldr q16, [x20], #0x10\n" - "zip1 v17.2d, v19.2d, v16.2d\n" "prfm pldl1keep, [x20, #0x70]\n" "str q26, [%x[out_ptr], #0x0]\n" - "zip2 v16.2d, v19.2d, v16.2d\n" - "str q23, [%x[out_ptr], #0x10]\n" - "str q20, [%x[out_ptr], #0x20]\n" + "str q22, [%x[out_ptr], #0x10]\n" + "str q19, [%x[out_ptr], #0x20]\n" "str q17, [%x[out_ptr], #0x30]\n" "str q24, [%x[out_ptr], #0x40]\n" "str q21, [%x[out_ptr], #0x50]\n" "str q18, [%x[out_ptr], #0x60]\n" "str q16, [%x[out_ptr], #0x70]\n" - "subs %x[width], %x[width], #0x4\n" - "cmp %x[width], #0x4\n" "add %x[out_ptr], %x[out_ptr], #0x80\n" "bge 2b\n" "3:" // Main loop skip @@ -120,58 +120,58 @@ void interleave_block<8, 2, VLType::None, false>( "tbz %x[width], #1, 4f\n" "ldr d27, [x27], #0x8\n" "ldr d24, [x26], #0x8\n" + "mov x19, #0x1\n" "ldr d25, [x25], #0x8\n" "ldr d21, [x24], #0x8\n" - "ldr d22, [x23], #0x8\n" + "ldr d23, [x23], #0x8\n" "ldr d18, [x22], #0x8\n" - "ldr d19, [x21], #0x8\n" + "ldr d20, [x21], #0x8\n" "ldr d16, [x20], #0x8\n" - "mov x19, #0x1\n" "tbz %x[width], #0, 5f\n" "ld1 { v27.s }[2], [x27]\n" + "mov x19, #0x2\n" "ld1 { v24.s }[2], [x26]\n" "ld1 { v25.s }[2], [x25]\n" "ld1 { v21.s }[2], [x24]\n" - "ld1 { v22.s }[2], [x23]\n" + "ld1 { v23.s }[2], [x23]\n" "ld1 { v18.s }[2], [x22]\n" - "ld1 { v19.s }[2], [x21]\n" + "ld1 { v20.s }[2], [x21]\n" "ld1 { v16.s }[2], [x20]\n" - "mov x19, #0x2\n" "b 5f\n" "4:" // odd_loads_1_0 "ldr s27, [x27, #0x0]\n" + "mov x19, #0x1\n" "ldr s24, [x26, #0x0]\n" "ldr s25, [x25, #0x0]\n" "ldr s21, [x24, #0x0]\n" - "ldr s22, [x23, #0x0]\n" + "ldr s23, [x23, #0x0]\n" "ldr s18, [x22, #0x0]\n" - "ldr s19, [x21, #0x0]\n" + "ldr s20, [x21, #0x0]\n" "ldr s16, [x20, #0x0]\n" - "mov x19, #0x1\n" "5:" // Odd load end "zip1 v26.2d, v27.2d, v24.2d\n" - "subs x19, x19, #0x1\n" - "zip1 v23.2d, v25.2d, v21.2d\n" "str q26, [%x[out_ptr], #0x0]\n" - "zip1 v20.2d, v22.2d, v18.2d\n" - "str q23, [%x[out_ptr], #0x10]\n" - "zip1 v17.2d, v19.2d, v16.2d\n" - "str q20, [%x[out_ptr], #0x20]\n" + "zip1 v22.2d, v25.2d, v21.2d\n" + "subs x19, x19, #0x1\n" + "zip1 v19.2d, v23.2d, v18.2d\n" + "str q22, [%x[out_ptr], #0x10]\n" + "zip1 v17.2d, v20.2d, v16.2d\n" + "str q19, [%x[out_ptr], #0x20]\n" "str q17, [%x[out_ptr], #0x30]\n" "add %x[out_ptr], %x[out_ptr], #0x40\n" "beq 6f\n" "zip2 v24.2d, v27.2d, v24.2d\n" - "zip2 v21.2d, v25.2d, v21.2d\n" "str q24, [%x[out_ptr], #0x0]\n" - "zip2 v18.2d, v22.2d, v18.2d\n" + "zip2 v21.2d, v25.2d, v21.2d\n" + "zip2 v18.2d, v23.2d, v18.2d\n" "str q21, [%x[out_ptr], #0x10]\n" - "zip2 v16.2d, v19.2d, v16.2d\n" + "zip2 v16.2d, v20.2d, v16.2d\n" "str q18, [%x[out_ptr], #0x20]\n" "str q16, [%x[out_ptr], #0x30]\n" "add %x[out_ptr], %x[out_ptr], #0x40\n" "6:" // Odds skip - : [out_ptr] "+r" (out_ptr), [width] "+r" (width) + : [out_ptr] "+&r" (out_ptr), [width] "+&r" (width) : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset) : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27" ); -- cgit v1.2.1