diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp | 351 |
1 files changed, 181 insertions, 170 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp index 7c7e91e666..586696fcc5 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_1x8.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -41,221 +41,232 @@ void a64_transpose_interleave_16_1x8(uint8_t *out, const uint8_t *in, size_t wid __asm__ __volatile__( "1:" // Main row loop: Head "mov x9, %x[in]\n" - "add x28, x9, %x[in_stride]\n" - "add x27, x28, %x[in_stride]\n" - "add x26, x27, %x[in_stride]\n" + "cmp %x[height], #0x7\n" + "mov x28, %x[width]\n" + "mov x27, %x[out]\n" + "add x26, x9, %x[in_stride]\n" "add x25, x26, %x[in_stride]\n" "add x24, x25, %x[in_stride]\n" "add x23, x24, %x[in_stride]\n" "add x22, x23, %x[in_stride]\n" - "cmp %x[height], #0x7\n" - "add %x[in], x22, %x[in_stride]\n" + "add x21, x22, %x[in_stride]\n" + "add x20, x21, %x[in_stride]\n" + "csel x21, x21, %x[pad_row], GE\n" + "add %x[in], x20, %x[in_stride]\n" + "csel x20, x20, %x[pad_row], GT\n" + "cmp %x[height], #0x5\n" "csel x22, x22, %x[pad_row], GT\n" "csel x23, x23, %x[pad_row], GE\n" - "cmp %x[height], #0x5\n" - "mov x21, %x[width]\n" + "cmp %x[height], #0x3\n" "csel x24, x24, %x[pad_row], GT\n" "csel x25, x25, %x[pad_row], GE\n" - "cmp %x[height], #0x3\n" - "csel x26, x26, %x[pad_row], GT\n" - "csel x27, x27, %x[pad_row], GE\n" "cmp %x[height], #0x1\n" - "csel x28, x28, %x[pad_row], GT\n" - "cmp x21, #0x20\n" - "mov x20, %x[out]\n" "sub %x[height], %x[height], #0x8\n" + "csel x26, x26, %x[pad_row], GT\n" + "cmp x28, #0x20\n" "blt 3f\n" "2:" // Main row loop: Unroll column loop - "ldr q23, [x9], #0x10\n" - "ldr q22, [x28], #0x10\n" - "sub x21, x21, #0x20\n" - "cmp x21, #0x20\n" - "ldr q20, [x27], #0x10\n" - "ldr q21, [x26], #0x10\n" - "ldr q19, [x25], #0x10\n" - "ldr q18, [x24], #0x10\n" - "zip1 v5.16b, v23.16b, v19.16b\n" - "zip1 v4.16b, v22.16b, v18.16b\n" - "ldr q17, [x23], #0x10\n" - "ldr q16, [x22], #0x10\n" - "zip1 v3.16b, v20.16b, v17.16b\n" - "zip1 v31.16b, v21.16b, v16.16b\n" - "ldr q25, [x9], #0x10\n" - "ldr q24, [x28], #0x10\n" - "zip2 v2.16b, v23.16b, v19.16b\n" - "zip2 v30.16b, v20.16b, v17.16b\n" - "ldr q23, [x27], #0x10\n" - "ldr q20, [x26], #0x10\n" - "zip2 v22.16b, v22.16b, v18.16b\n" - "zip2 v21.16b, v21.16b, v16.16b\n" - "ldr q19, [x25], #0x10\n" - "ldr q18, [x24], #0x10\n" - "zip1 v29.16b, v25.16b, v19.16b\n" - "zip1 v28.16b, v24.16b, v18.16b\n" - "ldr q17, [x23], #0x10\n" - "ldr q16, [x22], #0x10\n" - "zip1 v27.16b, v23.16b, v17.16b\n" - "zip1 v26.16b, v20.16b, v16.16b\n" - "zip2 v1.16b, v25.16b, v19.16b\n" - "zip2 v25.16b, v23.16b, v17.16b\n" - "zip2 v24.16b, v24.16b, v18.16b\n" - "zip2 v16.16b, v20.16b, v16.16b\n" - "zip1 v0.16b, v5.16b, v3.16b\n" - "zip1 v17.16b, v4.16b, v31.16b\n" - "zip2 v20.16b, v5.16b, v3.16b\n" - "zip2 v19.16b, v4.16b, v31.16b\n" - "zip1 v31.16b, v2.16b, v30.16b\n" - "zip1 v18.16b, v22.16b, v21.16b\n" - "zip2 v30.16b, v2.16b, v30.16b\n" + "ldr q24, [x9], #0x10\n" + "ldr q0, [x26], #0x10\n" + "sub x28, x28, #0x20\n" + "ldr q31, [x25], #0x10\n" + "ldr q30, [x24], #0x10\n" + "cmp x28, #0x20\n" + "ldr q23, [x23], #0x10\n" + "ldr q29, [x22], #0x10\n" + "ldr q22, [x21], #0x10\n" + "ldr q21, [x20], #0x10\n" + "ldr q28, [x9], #0x10\n" + "ldr q4, [x26], #0x10\n" + "ldr q27, [x25], #0x10\n" + "ldr q20, [x24], #0x10\n" + "zip1 v3.16b, v24.16b, v23.16b\n" + "zip1 v2.16b, v0.16b, v29.16b\n" + "ldr q19, [x23], #0x10\n" + "ldr q18, [x22], #0x10\n" + "zip1 v26.16b, v31.16b, v22.16b\n" + "zip1 v25.16b, v30.16b, v21.16b\n" + "ldr q17, [x21], #0x10\n" + "ldr q16, [x20], #0x10\n" + "zip2 v24.16b, v24.16b, v23.16b\n" + "zip2 v23.16b, v31.16b, v22.16b\n" + "zip2 v22.16b, v0.16b, v29.16b\n" + "zip2 v21.16b, v30.16b, v21.16b\n" + "zip1 v0.16b, v28.16b, v19.16b\n" + "zip1 v31.16b, v4.16b, v18.16b\n" + "zip1 v30.16b, v27.16b, v17.16b\n" + "zip1 v29.16b, v20.16b, v16.16b\n" + "zip2 v1.16b, v28.16b, v19.16b\n" + "zip2 v28.16b, v27.16b, v17.16b\n" + "zip2 v27.16b, v4.16b, v18.16b\n" + "zip2 v20.16b, v20.16b, v16.16b\n" + "zip1 v19.16b, v3.16b, v26.16b\n" + "zip1 v18.16b, v2.16b, v25.16b\n" + "zip2 v17.16b, v3.16b, v26.16b\n" + "zip2 v16.16b, v2.16b, v25.16b\n" + "zip1 v26.16b, v24.16b, v23.16b\n" + "zip1 v25.16b, v22.16b, v21.16b\n" + "zip2 v24.16b, v24.16b, v23.16b\n" "zip2 v23.16b, v22.16b, v21.16b\n" - "zip1 v22.16b, v29.16b, v27.16b\n" - "zip1 v21.16b, v28.16b, v26.16b\n" - "zip2 v29.16b, v29.16b, v27.16b\n" - "zip2 v28.16b, v28.16b, v26.16b\n" - "zip1 v27.16b, v1.16b, v25.16b\n" - "zip1 v26.16b, v24.16b, v16.16b\n" - "zip2 v25.16b, v1.16b, v25.16b\n" - "zip2 v24.16b, v24.16b, v16.16b\n" - "zip1 v16.16b, v0.16b, v17.16b\n" - "zip2 v17.16b, v0.16b, v17.16b\n" - "str q16, [x20, #0x0]\n" - "zip1 v16.16b, v20.16b, v19.16b\n" - "zip2 v20.16b, v20.16b, v19.16b\n" - "str q17, [x20, #0x10]\n" - "zip1 v19.16b, v31.16b, v18.16b\n" - "zip2 v18.16b, v31.16b, v18.16b\n" - "str q16, [x20, #0x20]\n" - "zip1 v17.16b, v30.16b, v23.16b\n" - "zip2 v16.16b, v30.16b, v23.16b\n" - "str q20, [x20, #0x30]\n" - "str q19, [x20, #0x40]\n" + "zip1 v22.16b, v0.16b, v30.16b\n" + "zip1 v21.16b, v31.16b, v29.16b\n" + "zip2 v0.16b, v0.16b, v30.16b\n" + "zip2 v31.16b, v31.16b, v29.16b\n" + "zip1 v30.16b, v1.16b, v28.16b\n" + "zip1 v29.16b, v27.16b, v20.16b\n" + "zip2 v28.16b, v1.16b, v28.16b\n" + "zip2 v27.16b, v27.16b, v20.16b\n" + "zip1 v20.16b, v19.16b, v18.16b\n" + "zip2 v19.16b, v19.16b, v18.16b\n" + "zip1 v18.16b, v17.16b, v16.16b\n" + "zip2 v17.16b, v17.16b, v16.16b\n" + "zip1 v16.16b, v26.16b, v25.16b\n" + "zip2 v26.16b, v26.16b, v25.16b\n" + "zip1 v25.16b, v24.16b, v23.16b\n" + "zip2 v24.16b, v24.16b, v23.16b\n" + "str q20, [x27, #0x0]\n" + "str q19, [x27, #0x10]\n" "zip1 v23.16b, v22.16b, v21.16b\n" "zip2 v22.16b, v22.16b, v21.16b\n" - "str q18, [x20, #0x50]\n" - "zip1 v21.16b, v29.16b, v28.16b\n" - "zip2 v20.16b, v29.16b, v28.16b\n" - "str q17, [x20, #0x60]\n" - "zip1 v19.16b, v27.16b, v26.16b\n" - "zip2 v18.16b, v27.16b, v26.16b\n" - "str q16, [x20, #0x70]\n" - "add x20, x20, %x[out_stride]\n" - "zip1 v17.16b, v25.16b, v24.16b\n" - "zip2 v16.16b, v25.16b, v24.16b\n" - "str q23, [x20, #0x0]\n" - "str q22, [x20, #0x10]\n" - "str q21, [x20, #0x20]\n" - "str q20, [x20, #0x30]\n" - "str q19, [x20, #0x40]\n" - "str q18, [x20, #0x50]\n" - "str q17, [x20, #0x60]\n" - "str q16, [x20, #0x70]\n" - "add x20, x20, %x[out_stride]\n" + "str q18, [x27, #0x20]\n" + "zip1 v21.16b, v0.16b, v31.16b\n" + "zip2 v20.16b, v0.16b, v31.16b\n" + "str q17, [x27, #0x30]\n" + "zip1 v19.16b, v30.16b, v29.16b\n" + "zip2 v18.16b, v30.16b, v29.16b\n" + "str q16, [x27, #0x40]\n" + "zip1 v17.16b, v28.16b, v27.16b\n" + "zip2 v16.16b, v28.16b, v27.16b\n" + "str q26, [x27, #0x50]\n" + "str q25, [x27, #0x60]\n" + "str q24, [x27, #0x70]\n" + "add x27, x27, %x[out_stride]\n" + "str q23, [x27, #0x0]\n" + "str q22, [x27, #0x10]\n" + "str q21, [x27, #0x20]\n" + "str q20, [x27, #0x30]\n" + "str q19, [x27, #0x40]\n" + "str q18, [x27, #0x50]\n" + "str q17, [x27, #0x60]\n" + "str q16, [x27, #0x70]\n" + "add x27, x27, %x[out_stride]\n" "bge 2b\n" "3:" // Main row loop: Unroll column loop skip - "cmp x21, #0x10\n" + "cmp x28, #0x10\n" "blt 5f\n" "4:" // Main row loop: Column loop - "ldr q25, [x9], #0x10\n" - "ldr q27, [x28], #0x10\n" - "sub x21, x21, #0x10\n" - "cmp x21, #0x10\n" - "ldr q26, [x27], #0x10\n" - "ldr q24, [x26], #0x10\n" - "ldr q22, [x25], #0x10\n" - "ldr q21, [x24], #0x10\n" - "zip1 v20.16b, v25.16b, v22.16b\n" - "zip1 v23.16b, v27.16b, v21.16b\n" - "ldr q17, [x23], #0x10\n" - "ldr q16, [x22], #0x10\n" - "zip1 v19.16b, v26.16b, v17.16b\n" - "zip1 v18.16b, v24.16b, v16.16b\n" - "zip2 v25.16b, v25.16b, v22.16b\n" - "zip2 v22.16b, v26.16b, v17.16b\n" - "zip2 v21.16b, v27.16b, v21.16b\n" - "zip2 v16.16b, v24.16b, v16.16b\n" - "zip1 v24.16b, v20.16b, v19.16b\n" - "zip1 v17.16b, v23.16b, v18.16b\n" - "zip2 v20.16b, v20.16b, v19.16b\n" - "zip2 v19.16b, v23.16b, v18.16b\n" - "zip1 v23.16b, v25.16b, v22.16b\n" - "zip1 v18.16b, v21.16b, v16.16b\n" - "zip2 v22.16b, v25.16b, v22.16b\n" - "zip2 v21.16b, v21.16b, v16.16b\n" - "zip1 v16.16b, v24.16b, v17.16b\n" - "zip2 v17.16b, v24.16b, v17.16b\n" - "str q16, [x20, #0x0]\n" - "zip1 v16.16b, v20.16b, v19.16b\n" + "ldr q24, [x9], #0x10\n" + "ldr q25, [x26], #0x10\n" + "sub x28, x28, #0x10\n" + "ldr q23, [x25], #0x10\n" + "ldr q22, [x24], #0x10\n" + "cmp x28, #0x10\n" + "ldr q18, [x23], #0x10\n" + "ldr q17, [x22], #0x10\n" + "ldr q21, [x21], #0x10\n" + "ldr q16, [x20], #0x10\n" + "zip1 v20.16b, v24.16b, v18.16b\n" + "zip1 v19.16b, v25.16b, v17.16b\n" + "zip2 v24.16b, v24.16b, v18.16b\n" + "zip2 v25.16b, v25.16b, v17.16b\n" + "zip1 v18.16b, v23.16b, v21.16b\n" + "zip1 v17.16b, v22.16b, v16.16b\n" + "zip2 v23.16b, v23.16b, v21.16b\n" + "zip2 v16.16b, v22.16b, v16.16b\n" + "zip1 v22.16b, v20.16b, v18.16b\n" + "zip1 v21.16b, v19.16b, v17.16b\n" + "zip2 v20.16b, v20.16b, v18.16b\n" + "zip2 v19.16b, v19.16b, v17.16b\n" + "zip1 v18.16b, v24.16b, v23.16b\n" + "zip1 v17.16b, v25.16b, v16.16b\n" + "zip2 v24.16b, v24.16b, v23.16b\n" + "zip2 v16.16b, v25.16b, v16.16b\n" + "zip1 v23.16b, v22.16b, v21.16b\n" + "zip2 v22.16b, v22.16b, v21.16b\n" + "zip1 v21.16b, v20.16b, v19.16b\n" "zip2 v20.16b, v20.16b, v19.16b\n" - "str q17, [x20, #0x10]\n" - "zip1 v19.16b, v23.16b, v18.16b\n" - "zip2 v18.16b, v23.16b, v18.16b\n" - "str q16, [x20, #0x20]\n" - "zip1 v17.16b, v22.16b, v21.16b\n" - "zip2 v16.16b, v22.16b, v21.16b\n" - "str q20, [x20, #0x30]\n" - "str q19, [x20, #0x40]\n" - "str q18, [x20, #0x50]\n" - "str q17, [x20, #0x60]\n" - "str q16, [x20, #0x70]\n" - "add x20, x20, %x[out_stride]\n" + "zip1 v19.16b, v18.16b, v17.16b\n" + "zip2 v18.16b, v18.16b, v17.16b\n" + "zip1 v17.16b, v24.16b, v16.16b\n" + "zip2 v16.16b, v24.16b, v16.16b\n" + "str q23, [x27, #0x0]\n" + "str q22, [x27, #0x10]\n" + "str q21, [x27, #0x20]\n" + "str q20, [x27, #0x30]\n" + "str q19, [x27, #0x40]\n" + "str q18, [x27, #0x50]\n" + "str q17, [x27, #0x60]\n" + "str q16, [x27, #0x70]\n" + "add x27, x27, %x[out_stride]\n" "bge 4b\n" "5:" // Main row loop: Column loop skip - "cmp x21, #0x4\n" + "cbz x28, 10f\n" + "cmp x28, #0x4\n" + "movi v16.16b, #0x0\n" + "str q16, [x27, #0x0]\n" + "str q16, [x27, #0x10]\n" + "str q16, [x27, #0x20]\n" + "str q16, [x27, #0x30]\n" + "str q16, [x27, #0x40]\n" + "str q16, [x27, #0x50]\n" + "str q16, [x27, #0x60]\n" + "str q16, [x27, #0x70]\n" "blt 7f\n" "6:" // Main row loop: width 4 loop: loop - "ldr s18, [x9], #0x4\n" - "ldr s19, [x28], #0x4\n" - "sub x21, x21, #0x4\n" - "cmp x21, #0x4\n" - "ldr s21, [x27], #0x4\n" - "ldr s20, [x26], #0x4\n" - "ldr s17, [x25], #0x4\n" - "ldr s16, [x24], #0x4\n" - "zip1 v18.16b, v18.16b, v17.16b\n" - "zip1 v19.16b, v19.16b, v16.16b\n" - "ldr s17, [x23], #0x4\n" - "ldr s16, [x22], #0x4\n" + "ldr s23, [x9], #0x4\n" + "ldr s22, [x26], #0x4\n" + "sub x28, x28, #0x4\n" + "ldr s21, [x25], #0x4\n" + "ldr s20, [x24], #0x4\n" + "cmp x28, #0x4\n" + "ldr s18, [x23], #0x4\n" + "ldr s19, [x22], #0x4\n" + "ldr s17, [x21], #0x4\n" + "ldr s16, [x20], #0x4\n" + "zip1 v18.16b, v23.16b, v18.16b\n" + "zip1 v19.16b, v22.16b, v19.16b\n" "zip1 v17.16b, v21.16b, v17.16b\n" "zip1 v16.16b, v20.16b, v16.16b\n" "zip1 v18.16b, v18.16b, v17.16b\n" "zip1 v16.16b, v19.16b, v16.16b\n" "zip1 v17.16b, v18.16b, v16.16b\n" "zip2 v16.16b, v18.16b, v16.16b\n" - "str q17, [x20, #0x0]\n" - "str q16, [x20, #0x10]\n" - "add x20, x20, #0x20\n" + "str q17, [x27, #0x0]\n" + "str q16, [x27, #0x10]\n" + "add x27, x27, #0x20\n" "bge 6b\n" "7:" // Main row loop: width 4 loop: skip - "cmp x21, #0x1\n" + "cmp x28, #0x1\n" "blt 9f\n" "8:" // Main row loop: width 1 loop: loop - "ldr b19, [x9], #0x1\n" - "ldr b18, [x28], #0x1\n" - "sub x21, x21, #0x1\n" - "cmp x21, #0x1\n" - "ldr b21, [x27], #0x1\n" - "ldr b20, [x26], #0x1\n" - "ldr b17, [x25], #0x1\n" - "ldr b16, [x24], #0x1\n" - "zip1 v19.16b, v19.16b, v17.16b\n" - "zip1 v18.16b, v18.16b, v16.16b\n" - "ldr b17, [x23], #0x1\n" - "ldr b16, [x22], #0x1\n" + "ldr b23, [x9], #0x1\n" + "ldr b22, [x26], #0x1\n" + "sub x28, x28, #0x1\n" + "ldr b21, [x25], #0x1\n" + "ldr b20, [x24], #0x1\n" + "cmp x28, #0x1\n" + "ldr b19, [x23], #0x1\n" + "ldr b18, [x22], #0x1\n" + "ldr b17, [x21], #0x1\n" + "ldr b16, [x20], #0x1\n" + "zip1 v19.16b, v23.16b, v19.16b\n" + "zip1 v18.16b, v22.16b, v18.16b\n" "zip1 v17.16b, v21.16b, v17.16b\n" "zip1 v16.16b, v20.16b, v16.16b\n" "zip1 v17.16b, v19.16b, v17.16b\n" "zip1 v16.16b, v18.16b, v16.16b\n" "zip1 v16.16b, v17.16b, v16.16b\n" - "str d16, [x20, #0x0]\n" - "add x20, x20, #0x8\n" + "str d16, [x27, #0x0]\n" + "add x27, x27, #0x8\n" "bge 8b\n" "9:" // Main row loop: width 1 loop: skip + "10:" // Main row loop: odd col skip "cmp %x[height], #0x1\n" "add %x[out], %x[out], #0x80\n" "bge 1b\n" : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out) : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |