diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp | 89 |
1 files changed, 44 insertions, 45 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp index dfec14358b..730bfd6342 100644 --- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/a64_interleave8_block4_fp32_bf16.hpp @@ -79,14 +79,14 @@ void interleave_block<8, 4, VLType::None, false>( "prfm pldl1keep, [x21, #0x40]\n" "blt 3f\n" "2:" // Main loop head - "ldr q23, [x28], #0x10\n" - "ldr q22, [x26], #0x10\n" - ".inst 0x0ea16af7 // bfcvtn v23.4h, v23.4s\n" - ".inst 0x0ea16ad6 // bfcvtn v22.4h, v22.4s\n" - "ldr q21, [x24], #0x10\n" - "ldr q20, [x22], #0x10\n" - ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n" - ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n" + "ldr q17, [x28], #0x10\n" + "ldr q16, [x26], #0x10\n" + ".inst 0x0ea16a37 // bfcvtn v23.4h, v17.4s\n" + ".inst 0x0ea16a16 // bfcvtn v22.4h, v16.4s\n" + "ldr q17, [x24], #0x10\n" + "ldr q16, [x22], #0x10\n" + ".inst 0x0ea16a35 // bfcvtn v21.4h, v17.4s\n" + ".inst 0x0ea16a14 // bfcvtn v20.4h, v16.4s\n" "ldr q19, [x27], #0x10\n" "ldr q18, [x25], #0x10\n" "subs %x[width], %x[width], #0x4\n" @@ -114,51 +114,50 @@ void interleave_block<8, 4, VLType::None, false>( "3:" // Main loop skip "cbz %x[width], 6f\n" "tbz %x[width], #1, 4f\n" - "ldr d23, [x28], #0x8\n" - "ldr d19, [x27], #0x8\n" + "ldr d19, [x28], #0x8\n" + "ldr d23, [x27], #0x8\n" "mov x20, #0x1\n" - "ldr d22, [x26], #0x8\n" - "ldr d18, [x25], #0x8\n" - "ldr d21, [x24], #0x8\n" - "ldr d17, [x23], #0x8\n" - "ldr d20, [x22], #0x8\n" - "ldr d16, [x21], #0x8\n" + "ldr d18, [x26], #0x8\n" + "ldr d22, [x25], #0x8\n" + "ldr d17, [x24], #0x8\n" + "ldr d21, [x23], #0x8\n" + "ldr d16, [x22], #0x8\n" + "ldr d20, [x21], #0x8\n" "tbz %x[width], #0, 5f\n" - "ld1 { v23.s }[2], [x28]\n" - "ld1 { v19.s }[2], [x27]\n" - "ld1 { v22.s }[2], [x26]\n" - "ld1 { v18.s }[2], [x25]\n" - "ld1 { v21.s }[2], [x24]\n" - "ld1 { v17.s }[2], [x23]\n" - "ld1 { v20.s }[2], [x22]\n" - "ld1 { v16.s }[2], [x21]\n" + "ld1 { v19.s }[2], [x28]\n" + "ld1 { v23.s }[2], [x27]\n" + "ld1 { v18.s }[2], [x26]\n" + "ld1 { v22.s }[2], [x25]\n" + "ld1 { v17.s }[2], [x24]\n" + "ld1 { v21.s }[2], [x23]\n" + "ld1 { v16.s }[2], [x22]\n" + "ld1 { v20.s }[2], [x21]\n" "b 5f\n" "4:" // odd_loads_1_0 - "ldr s23, [x28, #0x0]\n" - "ldr s19, [x27, #0x0]\n" + "ldr s19, [x28, #0x0]\n" + "ldr s23, [x27, #0x0]\n" "mov x20, #0x1\n" - "ldr s22, [x26, #0x0]\n" - "ldr s18, [x25, #0x0]\n" - "ldr s21, [x24, #0x0]\n" - "ldr s17, [x23, #0x0]\n" - "ldr s20, [x22, #0x0]\n" - "ldr s16, [x21, #0x0]\n" + "ldr s18, [x26, #0x0]\n" + "ldr s22, [x25, #0x0]\n" + "ldr s17, [x24, #0x0]\n" + "ldr s21, [x23, #0x0]\n" + "ldr s16, [x22, #0x0]\n" + "ldr s20, [x21, #0x0]\n" "5:" // Odd load end - ".inst 0x0ea16af7 // bfcvtn v23.4h, v23.4s\n" - ".inst 0x0ea16ad6 // bfcvtn v22.4h, v22.4s\n" - ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n" - ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n" - ".inst 0x4ea16a77 // bfcvtn2 v23.8h, v19.4s\n" - ".inst 0x4ea16a56 // bfcvtn2 v22.8h, v18.4s\n" - "str q23, [%x[out_ptr], #0x0]\n" - ".inst 0x4ea16a35 // bfcvtn2 v21.8h, v17.4s\n" - ".inst 0x4ea16a14 // bfcvtn2 v20.8h, v16.4s\n" - "str q22, [%x[out_ptr], #0x10]\n" - "str q21, [%x[out_ptr], #0x20]\n" - "str q20, [%x[out_ptr], #0x30]\n" + ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n" + ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n" + ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n" + ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" + ".inst 0x4ea16af3 // bfcvtn2 v19.8h, v23.4s\n" + ".inst 0x4ea16ad2 // bfcvtn2 v18.8h, v22.4s\n" + "str q19, [%x[out_ptr], #0x0]\n" + ".inst 0x4ea16ab1 // bfcvtn2 v17.8h, v21.4s\n" + ".inst 0x4ea16a90 // bfcvtn2 v16.8h, v20.4s\n" + "str q18, [%x[out_ptr], #0x10]\n" + "str q17, [%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) : [height] "r" (height), [in] "r" (in), [row_offset] "r" (row_offset) : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" |