diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp | 10 |
1 files changed, 5 insertions, 5 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp index afb06dedea..27e08135b6 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp @@ -1988,8 +1988,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ld1 { v22.4s }, [x22], #0x10\n" "ld1 { v26.4s }, [x21], #0x10\n" "tbz x11, #1, 144f\n" - "mov x19, #0x38\n" "ldr d11, [x28], #0x8\n" + "mov x19, #0x38\n" "ldr d15, [x24], #0x8\n" "ldr d19, [x23], #0x8\n" "ldr d23, [x22], #0x8\n" @@ -2042,8 +2042,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ld1 { v20.4s }, [x22], #0x10\n" "ld1 { v24.4s }, [x21], #0x10\n" "tbz x11, #1, 148f\n" - "mov x19, #0x18\n" "ldr d9, [x28], #0x8\n" + "mov x19, #0x18\n" "ldr d13, [x24], #0x8\n" "ldr d17, [x23], #0x8\n" "ldr d21, [x22], #0x8\n" @@ -2717,12 +2717,12 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ld1 { v16.4s }, [x23], #0x10\n" "ld1 { v20.4s }, [x22], #0x10\n" "ld1 { v24.4s }, [x21], #0x10\n" - "ld1 { v28.4s }, [x20], #0x10\n" "ld1 { v9.4s }, [x28], #0x10\n" "ld1 { v13.4s }, [x24], #0x10\n" "ld1 { v17.4s }, [x23], #0x10\n" "ld1 { v21.4s }, [x22], #0x10\n" "ld1 { v25.4s }, [x21], #0x10\n" + "ld1 { v28.4s }, [x20], #0x10\n" "ld1 { v29.4s }, [x20], #0x10\n" "tbz x11, #2, 180f\n" "ld1 { v10.4s }, [x28], #0x10\n" @@ -2732,8 +2732,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ld1 { v26.4s }, [x21], #0x10\n" "ld1 { v30.4s }, [x20], #0x10\n" "tbz x11, #1, 179f\n" - "mov x19, #0x38\n" "ldr d11, [x28], #0x8\n" + "mov x19, #0x38\n" "ldr d15, [x24], #0x8\n" "ldr d19, [x23], #0x8\n" "ldr d23, [x22], #0x8\n" @@ -2793,8 +2793,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ld1 { v24.4s }, [x21], #0x10\n" "ld1 { v28.4s }, [x20], #0x10\n" "tbz x11, #1, 183f\n" - "mov x19, #0x18\n" "ldr d9, [x28], #0x8\n" + "mov x19, #0x18\n" "ldr d13, [x24], #0x8\n" "ldr d17, [x23], #0x8\n" "ldr d21, [x22], #0x8\n" |