From 7d9a626aaba9837cb82d189a9c4f0bcef58825bb Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 1 Feb 2023 16:37:07 +0000 Subject: Update CPU kernels to remove x19 and w19 Resolves: COMPMID-5805 Change-Id: Idf720bbb136474810086f5089c5ed23b3f79835a Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9081 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Reviewed-by: Gunes Bayir Reviewed-by: Viet-Hoa Do --- .../a64_interleave8_block1_fp32_fp32.hpp | 212 ++++++++++----------- 1 file changed, 106 insertions(+), 106 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp32_fp32.hpp') diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp32_fp32.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp32_fp32.hpp index eefb8549ea..6c009b34b8 100644 --- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp32_fp32.hpp +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block1_fp32_fp32.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -10,16 +10,16 @@ * 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 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. + * 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. */ #ifdef __aarch64__ @@ -31,44 +31,45 @@ void interleave_block<8, 1, VLType::None, false>( ) { __asm__ __volatile__( - "ldr x27, [%x[in], #0x0]\n" + "ldr x28, [%x[in], #0x0]\n" + "ldr x27, [%x[in], #0x8]\n" "cmp %x[height], #0x8\n" - "ldr x26, [%x[in], #0x8]\n" + "add x28, x28, %x[row_offset], LSL #2\n" + "ldr x26, [%x[in], #0x10]\n" + "ldr x25, [%x[in], #0x18]\n" "add x27, x27, %x[row_offset], LSL #2\n" - "ldr x25, [%x[in], #0x10]\n" - "ldr x24, [%x[in], #0x18]\n" "add x26, x26, %x[row_offset], LSL #2\n" - "ldr x23, [%x[in], #0x20]\n" + "ldr x24, [%x[in], #0x20]\n" + "ldr x23, [%x[in], #0x28]\n" "add x25, x25, %x[row_offset], LSL #2\n" - "ldr x22, [%x[in], #0x28]\n" - "ldr x21, [%x[in], #0x30]\n" "add x24, x24, %x[row_offset], LSL #2\n" - "ldr x20, [%x[in], #0x38]\n" + "ldr x22, [%x[in], #0x30]\n" + "ldr x21, [%x[in], #0x38]\n" "add x23, x23, %x[row_offset], LSL #2\n" "add x22, x22, %x[row_offset], LSL #2\n" "add x21, x21, %x[row_offset], LSL #2\n" - "add x20, x20, %x[row_offset], LSL #2\n" "beq 1f\n" - "mov x20, x27\n" "cmp %x[height], #0x2\n" - "csel x26, x26, x27, GE\n" - "csel x25, x25, x27, GT\n" + "csel x27, x27, x28, GE\n" + "csel x26, x26, x28, GT\n" "cmp %x[height], #0x4\n" - "csel x24, x24, x27, GE\n" - "csel x23, x23, x27, GT\n" + "csel x25, x25, x28, GE\n" + "csel x24, x24, x28, GT\n" "cmp %x[height], #0x6\n" - "csel x22, x22, x27, GE\n" - "csel x21, x21, x27, GT\n" + "mov x21, x28\n" + "csel x23, x23, x28, GE\n" + "csel x22, x22, x28, GT\n" "1:" // no_pointer_adj - "prfm pldl1keep, [x27, #0x0]\n" "cmp %x[width], #0x4\n" + "prfm pldl1keep, [x28, #0x0]\n" + "prfm pldl1keep, [x27, #0x0]\n" "prfm pldl1keep, [x26, #0x0]\n" "prfm pldl1keep, [x25, #0x0]\n" "prfm pldl1keep, [x24, #0x0]\n" "prfm pldl1keep, [x23, #0x0]\n" "prfm pldl1keep, [x22, #0x0]\n" "prfm pldl1keep, [x21, #0x0]\n" - "prfm pldl1keep, [x20, #0x0]\n" + "prfm pldl1keep, [x28, #0x40]\n" "prfm pldl1keep, [x27, #0x40]\n" "prfm pldl1keep, [x26, #0x40]\n" "prfm pldl1keep, [x25, #0x40]\n" @@ -76,49 +77,48 @@ void interleave_block<8, 1, VLType::None, false>( "prfm pldl1keep, [x23, #0x40]\n" "prfm pldl1keep, [x22, #0x40]\n" "prfm pldl1keep, [x21, #0x40]\n" - "prfm pldl1keep, [x20, #0x40]\n" "blt 3f\n" "2:" // Main loop head - "ldr q28, [x27], #0x10\n" + "ldr q28, [x28], #0x10\n" + "ldr q27, [x27], #0x10\n" "subs %x[width], %x[width], #0x4\n" - "ldr q29, [x26], #0x10\n" "cmp %x[width], #0x4\n" - "ldr q25, [x25], #0x10\n" - "zip1 v22.4s, v28.4s, v25.4s\n" - "ldr q21, [x24], #0x10\n" - "zip2 v28.4s, v28.4s, v25.4s\n" - "ldr q27, [x23], #0x10\n" - "ldr q26, [x22], #0x10\n" - "zip1 v20.4s, v29.4s, v21.4s\n" - "ldr q19, [x21], #0x10\n" - "zip2 v25.4s, v29.4s, v21.4s\n" - "ldr q24, [x20], #0x10\n" - "zip1 v23.4s, v22.4s, v20.4s\n" + "ldr q22, [x26], #0x10\n" + "ldr q21, [x25], #0x10\n" + "zip1 v26.4s, v28.4s, v22.4s\n" + "zip1 v25.4s, v27.4s, v21.4s\n" + "ldr q24, [x24], #0x10\n" + "ldr q23, [x23], #0x10\n" + "zip2 v22.4s, v28.4s, v22.4s\n" + "zip2 v21.4s, v27.4s, v21.4s\n" + "ldr q19, [x22], #0x10\n" + "ldr q18, [x21], #0x10\n" + "zip1 v20.4s, v24.4s, v19.4s\n" + "zip1 v17.4s, v23.4s, v18.4s\n" + "zip2 v19.4s, v24.4s, v19.4s\n" + "zip2 v18.4s, v23.4s, v18.4s\n" + "prfm pldl1keep, [x28, #0x70]\n" "prfm pldl1keep, [x27, #0x70]\n" - "zip2 v22.4s, v22.4s, v20.4s\n" "prfm pldl1keep, [x26, #0x70]\n" - "zip1 v21.4s, v28.4s, v25.4s\n" "prfm pldl1keep, [x25, #0x70]\n" - "zip1 v18.4s, v27.4s, v19.4s\n" + "zip1 v16.4s, v26.4s, v25.4s\n" + "str q16, [%x[out_ptr], #0x0]\n" "prfm pldl1keep, [x24, #0x70]\n" - "zip1 v16.4s, v26.4s, v24.4s\n" "prfm pldl1keep, [x23, #0x70]\n" - "zip1 v17.4s, v18.4s, v16.4s\n" + "zip1 v16.4s, v20.4s, v17.4s\n" + "str q16, [%x[out_ptr], #0x10]\n" "prfm pldl1keep, [x22, #0x70]\n" - "zip2 v20.4s, v18.4s, v16.4s\n" "prfm pldl1keep, [x21, #0x70]\n" - "zip2 v19.4s, v27.4s, v19.4s\n" - "prfm pldl1keep, [x20, #0x70]\n" - "zip2 v16.4s, v26.4s, v24.4s\n" - "str q23, [%x[out_ptr], #0x0]\n" - "zip1 v18.4s, v19.4s, v16.4s\n" - "str q17, [%x[out_ptr], #0x10]\n" - "zip2 v17.4s, v28.4s, v25.4s\n" - "str q22, [%x[out_ptr], #0x20]\n" - "zip2 v16.4s, v19.4s, v16.4s\n" - "str q20, [%x[out_ptr], #0x30]\n" - "str q21, [%x[out_ptr], #0x40]\n" - "str q18, [%x[out_ptr], #0x50]\n" + "zip2 v16.4s, v26.4s, v25.4s\n" + "str q16, [%x[out_ptr], #0x20]\n" + "zip2 v16.4s, v20.4s, v17.4s\n" + "str q16, [%x[out_ptr], #0x30]\n" + "zip1 v16.4s, v22.4s, v21.4s\n" + "str q16, [%x[out_ptr], #0x40]\n" + "zip1 v16.4s, v19.4s, v18.4s\n" + "zip2 v17.4s, v22.4s, v21.4s\n" + "str q16, [%x[out_ptr], #0x50]\n" + "zip2 v16.4s, v19.4s, v18.4s\n" "str q17, [%x[out_ptr], #0x60]\n" "str q16, [%x[out_ptr], #0x70]\n" "add %x[out_ptr], %x[out_ptr], #0x80\n" @@ -126,69 +126,69 @@ void interleave_block<8, 1, VLType::None, false>( "3:" // Main loop skip "cbz %x[width], 6f\n" "tbz %x[width], #1, 4f\n" - "ldr d28, [x27], #0x8\n" - "ldr d29, [x26], #0x8\n" - "mov x19, #0x2\n" - "ldr d25, [x25], #0x8\n" - "ldr d21, [x24], #0x8\n" - "ldr d27, [x23], #0x8\n" - "ldr d26, [x22], #0x8\n" - "ldr d19, [x21], #0x8\n" - "ldr d24, [x20], #0x8\n" + "ldr d28, [x28], #0x8\n" + "ldr d27, [x27], #0x8\n" + "mov x20, #0x2\n" + "ldr d22, [x26], #0x8\n" + "ldr d21, [x25], #0x8\n" + "ldr d24, [x24], #0x8\n" + "ldr d23, [x23], #0x8\n" + "ldr d19, [x22], #0x8\n" + "ldr d18, [x21], #0x8\n" "tbz %x[width], #0, 5f\n" - "ld1 { v28.s }[2], [x27]\n" - "mov x19, #0x3\n" - "ld1 { v29.s }[2], [x26]\n" - "ld1 { v25.s }[2], [x25]\n" - "ld1 { v21.s }[2], [x24]\n" - "ld1 { v27.s }[2], [x23]\n" - "ld1 { v26.s }[2], [x22]\n" - "ld1 { v19.s }[2], [x21]\n" - "ld1 { v24.s }[2], [x20]\n" + "ld1 { v28.s }[2], [x28]\n" + "ld1 { v27.s }[2], [x27]\n" + "mov x20, #0x3\n" + "ld1 { v22.s }[2], [x26]\n" + "ld1 { v21.s }[2], [x25]\n" + "ld1 { v24.s }[2], [x24]\n" + "ld1 { v23.s }[2], [x23]\n" + "ld1 { v19.s }[2], [x22]\n" + "ld1 { v18.s }[2], [x21]\n" "b 5f\n" "4:" // odd_loads_1_0 - "ldr s28, [x27, #0x0]\n" - "mov x19, #0x1\n" - "ldr s29, [x26, #0x0]\n" - "ldr s25, [x25, #0x0]\n" - "ldr s21, [x24, #0x0]\n" - "ldr s27, [x23, #0x0]\n" - "ldr s26, [x22, #0x0]\n" - "ldr s19, [x21, #0x0]\n" - "ldr s24, [x20, #0x0]\n" + "ldr s28, [x28, #0x0]\n" + "ldr s27, [x27, #0x0]\n" + "mov x20, #0x1\n" + "ldr s22, [x26, #0x0]\n" + "ldr s21, [x25, #0x0]\n" + "ldr s24, [x24, #0x0]\n" + "ldr s23, [x23, #0x0]\n" + "ldr s19, [x22, #0x0]\n" + "ldr s18, [x21, #0x0]\n" "5:" // Odd load end - "zip1 v22.4s, v28.4s, v25.4s\n" - "subs x19, x19, #0x1\n" - "zip1 v20.4s, v29.4s, v21.4s\n" - "zip1 v23.4s, v22.4s, v20.4s\n" - "str q23, [%x[out_ptr], #0x0]\n" - "zip1 v18.4s, v27.4s, v19.4s\n" - "zip1 v16.4s, v26.4s, v24.4s\n" - "zip1 v17.4s, v18.4s, v16.4s\n" - "str q17, [%x[out_ptr], #0x10]\n" + "zip1 v26.4s, v28.4s, v22.4s\n" + "zip1 v25.4s, v27.4s, v21.4s\n" + "subs x20, x20, #0x1\n" + "zip1 v20.4s, v24.4s, v19.4s\n" + "zip1 v17.4s, v23.4s, v18.4s\n" + "zip1 v16.4s, v26.4s, v25.4s\n" + "str q16, [%x[out_ptr], #0x0]\n" + "zip1 v16.4s, v20.4s, v17.4s\n" + "str q16, [%x[out_ptr], #0x10]\n" "add %x[out_ptr], %x[out_ptr], #0x20\n" "beq 6f\n" - "zip2 v22.4s, v22.4s, v20.4s\n" - "str q22, [%x[out_ptr], #0x0]\n" - "zip2 v20.4s, v18.4s, v16.4s\n" - "subs x19, x19, #0x1\n" - "str q20, [%x[out_ptr], #0x10]\n" + "subs x20, x20, #0x1\n" + "zip2 v16.4s, v26.4s, v25.4s\n" + "str q16, [%x[out_ptr], #0x0]\n" + "zip2 v16.4s, v20.4s, v17.4s\n" + "str q16, [%x[out_ptr], #0x10]\n" "add %x[out_ptr], %x[out_ptr], #0x20\n" "beq 6f\n" - "zip2 v28.4s, v28.4s, v25.4s\n" - "zip2 v25.4s, v29.4s, v21.4s\n" - "zip1 v21.4s, v28.4s, v25.4s\n" - "str q21, [%x[out_ptr], #0x0]\n" - "zip2 v19.4s, v27.4s, v19.4s\n" - "zip2 v16.4s, v26.4s, v24.4s\n" - "zip1 v18.4s, v19.4s, v16.4s\n" - "str q18, [%x[out_ptr], #0x10]\n" + "zip2 v22.4s, v28.4s, v22.4s\n" + "zip2 v21.4s, v27.4s, v21.4s\n" + "zip2 v19.4s, v24.4s, v19.4s\n" + "zip2 v18.4s, v23.4s, v18.4s\n" + "zip1 v16.4s, v22.4s, v21.4s\n" + "str q16, [%x[out_ptr], #0x0]\n" + "zip1 v16.4s, v19.4s, v18.4s\n" + "str q16, [%x[out_ptr], #0x10]\n" "add %x[out_ptr], %x[out_ptr], #0x20\n" "6:" // Odds skip : [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", "v28", "v29", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27" + : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } -- cgit v1.2.1