diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp | 20 |
1 files changed, 10 insertions, 10 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp index 5f4fcac690..349ad1c985 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffinterleaved_bf16fp32_dot_8x12/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -54,17 +54,17 @@ void a64_ffinterleaved_bf16fp32_dot_8x12( "1:" // Height loop "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n" "ldr x25, [%x[args_ptr], %[offsetof_N]]\n" - "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "mov x24, %x[Apanel]\n" + "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "2:" // Width loop "ldr x23, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" "ldr x20, [%x[args_ptr], %[offsetof_B_stride]]\n" + "cmp x25, #0x8\n" + "mov %x[Apanel], x24\n" "add x22, x23, x20, LSL #1\n" "add x21, x22, x20, LSL #1\n" "add x20, x21, x20, LSL #1\n" "str x20, [%x[args_ptr], %[offsetof_cur_B_ptr]]\n" - "cmp x25, #0x8\n" - "mov %x[Apanel], x24\n" "bgt 3f\n" "cmp x25, #0x4\n" "mov x21, x23\n" @@ -79,12 +79,12 @@ void a64_ffinterleaved_bf16fp32_dot_8x12( "movi v9.16b, #0x0\n" "ldr q6, [x21, #0x0]\n" "ldr x20, [%x[args_ptr], %[offsetof_K]]\n" - "cmp x20, #0x2\n" "movi v10.16b, #0x0\n" "movi v11.16b, #0x0\n" "movi v12.16b, #0x0\n" "movi v13.16b, #0x0\n" "movi v14.16b, #0x0\n" + "cmp x20, #0x2\n" "movi v15.16b, #0x0\n" "movi v16.16b, #0x0\n" "movi v17.16b, #0x0\n" @@ -170,18 +170,18 @@ void a64_ffinterleaved_bf16fp32_dot_8x12( "ldr q6, [x21, #0x0]\n" "bge 4b\n" "5:" // main loop skip + "add %x[Apanel], %x[Apanel], #0x20\n" + "add x23, x23, #0x10\n" ".inst 0x4f40f088 // bfdot v8.4s, v4.8h, v0.h[0]\n" + "add x22, x22, #0x10\n" + "add x21, x21, #0x10\n" ".inst 0x4f60f08b // bfdot v11.4s, v4.8h, v0.h[1]\n" - "add %x[Apanel], %x[Apanel], #0x20\n" ".inst 0x4f40f88e // bfdot v14.4s, v4.8h, v0.h[2]\n" ".inst 0x4f60f891 // bfdot v17.4s, v4.8h, v0.h[3]\n" - "add x23, x23, #0x10\n" ".inst 0x4f41f094 // bfdot v20.4s, v4.8h, v1.h[0]\n" ".inst 0x4f61f097 // bfdot v23.4s, v4.8h, v1.h[1]\n" - "add x22, x22, #0x10\n" ".inst 0x4f41f89a // bfdot v26.4s, v4.8h, v1.h[2]\n" ".inst 0x4f61f89d // bfdot v29.4s, v4.8h, v1.h[3]\n" - "add x21, x21, #0x10\n" ".inst 0x4f40f0a9 // bfdot v9.4s, v5.8h, v0.h[0]\n" ".inst 0x4f60f0ac // bfdot v12.4s, v5.8h, v0.h[1]\n" ".inst 0x4f40f8af // bfdot v15.4s, v5.8h, v0.h[2]\n" @@ -204,8 +204,8 @@ void a64_ffinterleaved_bf16fp32_dot_8x12( "add %x[Apanel], %x[Apanel], #0x20\n" "ldr q2, [x23, #0x0]\n" "ldr q1, [x22, #0x0]\n" - ".inst 0x4f44f048 // bfdot v8.4s, v2.8h, v4.h[0]\n" "ldr q0, [x21, #0x0]\n" + ".inst 0x4f44f048 // bfdot v8.4s, v2.8h, v4.h[0]\n" ".inst 0x4f64f04b // bfdot v11.4s, v2.8h, v4.h[1]\n" ".inst 0x4f44f84e // bfdot v14.4s, v2.8h, v4.h[2]\n" ".inst 0x4f64f851 // bfdot v17.4s, v2.8h, v4.h[3]\n" |