diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp | 273 |
1 files changed, 136 insertions, 137 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp index 4bfb36082e..7b445ef3d4 100644 --- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block8_s8_s8.hpp @@ -79,18 +79,18 @@ void interleave_block<8, 8, VLType::None, false>( "prfm pldl1keep, [x21, #0x40]\n" "blt 3f\n" "2:" // Main loop head - "ldr q26, [x28], #0x10\n" - "ldr q21, [x27], #0x10\n" + "ldr q20, [x28], #0x10\n" + "ldr q19, [x27], #0x10\n" "subs %x[width], %x[width], #0x10\n" "cmp %x[width], #0x10\n" "ldr q25, [x26], #0x10\n" "ldr q24, [x25], #0x10\n" - "zip1 v16.2d, v26.2d, v21.2d\n" + "zip1 v16.2d, v20.2d, v19.2d\n" "zip1 v18.2d, v25.2d, v24.2d\n" "ldr q23, [x24], #0x10\n" "ldr q22, [x23], #0x10\n" "zip1 v17.2d, v23.2d, v22.2d\n" - "zip2 v21.2d, v26.2d, v21.2d\n" + "zip2 v21.2d, v20.2d, v19.2d\n" "ldr q20, [x22], #0x10\n" "ldr q19, [x21], #0x10\n" "str q16, [%x[out_ptr], #0x0]\n" @@ -118,188 +118,187 @@ void interleave_block<8, 8, VLType::None, false>( "3:" // Main loop skip "cbz %x[width], 12f\n" "tbz %x[width], #3, 7f\n" - "ldr d26, [x28], #0x8\n" - "ldr d21, [x27], #0x8\n" - "ldr d25, [x26], #0x8\n" - "ldr d24, [x25], #0x8\n" - "ldr d23, [x24], #0x8\n" - "ldr d22, [x23], #0x8\n" - "ldr d20, [x22], #0x8\n" - "ldr d19, [x21], #0x8\n" + "ldr d25, [x28], #0x8\n" + "ldr d24, [x27], #0x8\n" + "ldr d23, [x26], #0x8\n" + "ldr d22, [x25], #0x8\n" + "ldr d21, [x24], #0x8\n" + "ldr d20, [x23], #0x8\n" + "ldr d19, [x22], #0x8\n" + "ldr d18, [x21], #0x8\n" "tbz %x[width], #2, 5f\n" - "ld1 { v26.s }[2], [x28], #0x4\n" - "ld1 { v21.s }[2], [x27], #0x4\n" - "ld1 { v25.s }[2], [x26], #0x4\n" - "ld1 { v24.s }[2], [x25], #0x4\n" - "ld1 { v23.s }[2], [x24], #0x4\n" - "ld1 { v22.s }[2], [x23], #0x4\n" - "ld1 { v20.s }[2], [x22], #0x4\n" - "ld1 { v19.s }[2], [x21], #0x4\n" + "ld1 { v25.s }[2], [x28], #0x4\n" + "ld1 { v24.s }[2], [x27], #0x4\n" + "ld1 { v23.s }[2], [x26], #0x4\n" + "ld1 { v22.s }[2], [x25], #0x4\n" + "ld1 { v21.s }[2], [x24], #0x4\n" + "ld1 { v20.s }[2], [x23], #0x4\n" + "ld1 { v19.s }[2], [x22], #0x4\n" + "ld1 { v18.s }[2], [x21], #0x4\n" "tbz %x[width], #1, 4f\n" - "ld1 { v26.h }[6], [x28], #0x2\n" - "ld1 { v21.h }[6], [x27], #0x2\n" + "ld1 { v25.h }[6], [x28], #0x2\n" + "ld1 { v24.h }[6], [x27], #0x2\n" "mov x20, #0x2\n" - "ld1 { v25.h }[6], [x26], #0x2\n" - "ld1 { v24.h }[6], [x25], #0x2\n" - "ld1 { v23.h }[6], [x24], #0x2\n" - "ld1 { v22.h }[6], [x23], #0x2\n" - "ld1 { v20.h }[6], [x22], #0x2\n" - "ld1 { v19.h }[6], [x21], #0x2\n" + "ld1 { v23.h }[6], [x26], #0x2\n" + "ld1 { v22.h }[6], [x25], #0x2\n" + "ld1 { v21.h }[6], [x24], #0x2\n" + "ld1 { v20.h }[6], [x23], #0x2\n" + "ld1 { v19.h }[6], [x22], #0x2\n" + "ld1 { v18.h }[6], [x21], #0x2\n" "tbz %x[width], #0, 11f\n" - "ld1 { v26.b }[14], [x28]\n" - "ld1 { v21.b }[14], [x27]\n" - "ld1 { v25.b }[14], [x26]\n" - "ld1 { v24.b }[14], [x25]\n" - "ld1 { v23.b }[14], [x24]\n" - "ld1 { v22.b }[14], [x23]\n" - "ld1 { v20.b }[14], [x22]\n" - "ld1 { v19.b }[14], [x21]\n" + "ld1 { v25.b }[14], [x28]\n" + "ld1 { v24.b }[14], [x27]\n" + "ld1 { v23.b }[14], [x26]\n" + "ld1 { v22.b }[14], [x25]\n" + "ld1 { v21.b }[14], [x24]\n" + "ld1 { v20.b }[14], [x23]\n" + "ld1 { v19.b }[14], [x22]\n" + "ld1 { v18.b }[14], [x21]\n" "b 11f\n" "4:" // odd_loads_1_12 "mov x20, #0x2\n" "tbz %x[width], #0, 11f\n" - "ld1 { v26.b }[12], [x28]\n" - "ld1 { v21.b }[12], [x27]\n" - "ld1 { v25.b }[12], [x26]\n" - "ld1 { v24.b }[12], [x25]\n" - "ld1 { v23.b }[12], [x24]\n" - "ld1 { v22.b }[12], [x23]\n" - "ld1 { v20.b }[12], [x22]\n" - "ld1 { v19.b }[12], [x21]\n" + "ld1 { v25.b }[12], [x28]\n" + "ld1 { v24.b }[12], [x27]\n" + "ld1 { v23.b }[12], [x26]\n" + "ld1 { v22.b }[12], [x25]\n" + "ld1 { v21.b }[12], [x24]\n" + "ld1 { v20.b }[12], [x23]\n" + "ld1 { v19.b }[12], [x22]\n" + "ld1 { v18.b }[12], [x21]\n" "b 11f\n" "5:" // odd_loads_2_8 "tbz %x[width], #1, 6f\n" - "ld1 { v26.h }[4], [x28], #0x2\n" - "ld1 { v21.h }[4], [x27], #0x2\n" + "ld1 { v25.h }[4], [x28], #0x2\n" + "ld1 { v24.h }[4], [x27], #0x2\n" "mov x20, #0x2\n" - "ld1 { v25.h }[4], [x26], #0x2\n" - "ld1 { v24.h }[4], [x25], #0x2\n" - "ld1 { v23.h }[4], [x24], #0x2\n" - "ld1 { v22.h }[4], [x23], #0x2\n" - "ld1 { v20.h }[4], [x22], #0x2\n" - "ld1 { v19.h }[4], [x21], #0x2\n" + "ld1 { v23.h }[4], [x26], #0x2\n" + "ld1 { v22.h }[4], [x25], #0x2\n" + "ld1 { v21.h }[4], [x24], #0x2\n" + "ld1 { v20.h }[4], [x23], #0x2\n" + "ld1 { v19.h }[4], [x22], #0x2\n" + "ld1 { v18.h }[4], [x21], #0x2\n" "tbz %x[width], #0, 11f\n" - "ld1 { v26.b }[10], [x28]\n" - "ld1 { v21.b }[10], [x27]\n" - "ld1 { v25.b }[10], [x26]\n" - "ld1 { v24.b }[10], [x25]\n" - "ld1 { v23.b }[10], [x24]\n" - "ld1 { v22.b }[10], [x23]\n" - "ld1 { v20.b }[10], [x22]\n" - "ld1 { v19.b }[10], [x21]\n" + "ld1 { v25.b }[10], [x28]\n" + "ld1 { v24.b }[10], [x27]\n" + "ld1 { v23.b }[10], [x26]\n" + "ld1 { v22.b }[10], [x25]\n" + "ld1 { v21.b }[10], [x24]\n" + "ld1 { v20.b }[10], [x23]\n" + "ld1 { v19.b }[10], [x22]\n" + "ld1 { v18.b }[10], [x21]\n" "b 11f\n" "6:" // odd_loads_1_8 "mov x20, #0x1\n" "tbz %x[width], #0, 11f\n" - "ld1 { v26.b }[8], [x28]\n" - "ld1 { v21.b }[8], [x27]\n" + "ld1 { v25.b }[8], [x28]\n" + "ld1 { v24.b }[8], [x27]\n" "mov x20, #0x2\n" - "ld1 { v25.b }[8], [x26]\n" - "ld1 { v24.b }[8], [x25]\n" - "ld1 { v23.b }[8], [x24]\n" - "ld1 { v22.b }[8], [x23]\n" - "ld1 { v20.b }[8], [x22]\n" - "ld1 { v19.b }[8], [x21]\n" + "ld1 { v23.b }[8], [x26]\n" + "ld1 { v22.b }[8], [x25]\n" + "ld1 { v21.b }[8], [x24]\n" + "ld1 { v20.b }[8], [x23]\n" + "ld1 { v19.b }[8], [x22]\n" + "ld1 { v18.b }[8], [x21]\n" "b 11f\n" "7:" // odd_loads_4_0 "tbz %x[width], #2, 9f\n" - "ldr s26, [x28], #0x4\n" - "ldr s21, [x27], #0x4\n" - "ldr s25, [x26], #0x4\n" - "ldr s24, [x25], #0x4\n" - "ldr s23, [x24], #0x4\n" - "ldr s22, [x23], #0x4\n" - "ldr s20, [x22], #0x4\n" - "ldr s19, [x21], #0x4\n" + "ldr s25, [x28], #0x4\n" + "ldr s24, [x27], #0x4\n" + "ldr s23, [x26], #0x4\n" + "ldr s22, [x25], #0x4\n" + "ldr s21, [x24], #0x4\n" + "ldr s20, [x23], #0x4\n" + "ldr s19, [x22], #0x4\n" + "ldr s18, [x21], #0x4\n" "tbz %x[width], #1, 8f\n" - "ld1 { v26.h }[2], [x28], #0x2\n" - "ld1 { v21.h }[2], [x27], #0x2\n" + "ld1 { v25.h }[2], [x28], #0x2\n" + "ld1 { v24.h }[2], [x27], #0x2\n" "mov x20, #0x1\n" - "ld1 { v25.h }[2], [x26], #0x2\n" - "ld1 { v24.h }[2], [x25], #0x2\n" - "ld1 { v23.h }[2], [x24], #0x2\n" - "ld1 { v22.h }[2], [x23], #0x2\n" - "ld1 { v20.h }[2], [x22], #0x2\n" - "ld1 { v19.h }[2], [x21], #0x2\n" + "ld1 { v23.h }[2], [x26], #0x2\n" + "ld1 { v22.h }[2], [x25], #0x2\n" + "ld1 { v21.h }[2], [x24], #0x2\n" + "ld1 { v20.h }[2], [x23], #0x2\n" + "ld1 { v19.h }[2], [x22], #0x2\n" + "ld1 { v18.h }[2], [x21], #0x2\n" "tbz %x[width], #0, 11f\n" - "ld1 { v26.b }[6], [x28]\n" - "ld1 { v21.b }[6], [x27]\n" - "ld1 { v25.b }[6], [x26]\n" - "ld1 { v24.b }[6], [x25]\n" - "ld1 { v23.b }[6], [x24]\n" - "ld1 { v22.b }[6], [x23]\n" - "ld1 { v20.b }[6], [x22]\n" - "ld1 { v19.b }[6], [x21]\n" + "ld1 { v25.b }[6], [x28]\n" + "ld1 { v24.b }[6], [x27]\n" + "ld1 { v23.b }[6], [x26]\n" + "ld1 { v22.b }[6], [x25]\n" + "ld1 { v21.b }[6], [x24]\n" + "ld1 { v20.b }[6], [x23]\n" + "ld1 { v19.b }[6], [x22]\n" + "ld1 { v18.b }[6], [x21]\n" "b 11f\n" "8:" // odd_loads_1_4 "mov x20, #0x1\n" "tbz %x[width], #0, 11f\n" - "ld1 { v26.b }[4], [x28]\n" - "ld1 { v21.b }[4], [x27]\n" - "ld1 { v25.b }[4], [x26]\n" - "ld1 { v24.b }[4], [x25]\n" - "ld1 { v23.b }[4], [x24]\n" - "ld1 { v22.b }[4], [x23]\n" - "ld1 { v20.b }[4], [x22]\n" - "ld1 { v19.b }[4], [x21]\n" + "ld1 { v25.b }[4], [x28]\n" + "ld1 { v24.b }[4], [x27]\n" + "ld1 { v23.b }[4], [x26]\n" + "ld1 { v22.b }[4], [x25]\n" + "ld1 { v21.b }[4], [x24]\n" + "ld1 { v20.b }[4], [x23]\n" + "ld1 { v19.b }[4], [x22]\n" + "ld1 { v18.b }[4], [x21]\n" "b 11f\n" "9:" // odd_loads_2_0 "tbz %x[width], #1, 10f\n" - "ldr h26, [x28], #0x2\n" - "ldr h21, [x27], #0x2\n" + "ldr h25, [x28], #0x2\n" + "ldr h24, [x27], #0x2\n" "mov x20, #0x1\n" - "ldr h25, [x26], #0x2\n" - "ldr h24, [x25], #0x2\n" - "ldr h23, [x24], #0x2\n" - "ldr h22, [x23], #0x2\n" - "ldr h20, [x22], #0x2\n" - "ldr h19, [x21], #0x2\n" + "ldr h23, [x26], #0x2\n" + "ldr h22, [x25], #0x2\n" + "ldr h21, [x24], #0x2\n" + "ldr h20, [x23], #0x2\n" + "ldr h19, [x22], #0x2\n" + "ldr h18, [x21], #0x2\n" "tbz %x[width], #0, 11f\n" - "ld1 { v26.b }[2], [x28]\n" - "ld1 { v21.b }[2], [x27]\n" - "ld1 { v25.b }[2], [x26]\n" - "ld1 { v24.b }[2], [x25]\n" - "ld1 { v23.b }[2], [x24]\n" - "ld1 { v22.b }[2], [x23]\n" - "ld1 { v20.b }[2], [x22]\n" - "ld1 { v19.b }[2], [x21]\n" + "ld1 { v25.b }[2], [x28]\n" + "ld1 { v24.b }[2], [x27]\n" + "ld1 { v23.b }[2], [x26]\n" + "ld1 { v22.b }[2], [x25]\n" + "ld1 { v21.b }[2], [x24]\n" + "ld1 { v20.b }[2], [x23]\n" + "ld1 { v19.b }[2], [x22]\n" + "ld1 { v18.b }[2], [x21]\n" "b 11f\n" "10:" // odd_loads_1_0 - "ldr b26, [x28, #0x0]\n" - "ldr b21, [x27, #0x0]\n" + "ldr b25, [x28, #0x0]\n" + "ldr b24, [x27, #0x0]\n" "mov x20, #0x1\n" - "ldr b25, [x26, #0x0]\n" - "ldr b24, [x25, #0x0]\n" - "ldr b23, [x24, #0x0]\n" - "ldr b22, [x23, #0x0]\n" - "ldr b20, [x22, #0x0]\n" - "ldr b19, [x21, #0x0]\n" + "ldr b23, [x26, #0x0]\n" + "ldr b22, [x25, #0x0]\n" + "ldr b21, [x24, #0x0]\n" + "ldr b20, [x23, #0x0]\n" + "ldr b19, [x22, #0x0]\n" + "ldr b18, [x21, #0x0]\n" "11:" // Odd load end "subs x20, x20, #0x1\n" - "zip1 v16.2d, v26.2d, v21.2d\n" + "zip1 v16.2d, v25.2d, v24.2d\n" "str q16, [%x[out_ptr], #0x0]\n" - "zip1 v18.2d, v25.2d, v24.2d\n" - "str q18, [%x[out_ptr], #0x10]\n" - "zip1 v17.2d, v23.2d, v22.2d\n" - "zip1 v16.2d, v20.2d, v19.2d\n" + "zip1 v16.2d, v23.2d, v22.2d\n" + "str q16, [%x[out_ptr], #0x10]\n" + "zip1 v17.2d, v21.2d, v20.2d\n" + "zip1 v16.2d, v19.2d, v18.2d\n" "str q17, [%x[out_ptr], #0x20]\n" "str q16, [%x[out_ptr], #0x30]\n" "add %x[out_ptr], %x[out_ptr], #0x40\n" "beq 12f\n" - "zip2 v21.2d, v26.2d, v21.2d\n" - "str q21, [%x[out_ptr], #0x0]\n" - "zip2 v18.2d, v25.2d, v24.2d\n" - "str q18, [%x[out_ptr], #0x10]\n" - "zip2 v17.2d, v23.2d, v22.2d\n" - "zip2 v16.2d, v20.2d, v19.2d\n" + "zip2 v16.2d, v25.2d, v24.2d\n" + "str q16, [%x[out_ptr], #0x0]\n" + "zip2 v16.2d, v23.2d, v22.2d\n" + "str q16, [%x[out_ptr], #0x10]\n" + "zip2 v17.2d, v21.2d, v20.2d\n" + "zip2 v16.2d, v19.2d, v18.2d\n" "str q17, [%x[out_ptr], #0x20]\n" "str q16, [%x[out_ptr], #0x30]\n" "add %x[out_ptr], %x[out_ptr], #0x40\n" "12:" // 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", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |