diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp | 10 |
1 files changed, 5 insertions, 5 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp index ba57ad493a..ab0c88a3b2 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp @@ -1819,8 +1819,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "ld1 { v22.4s }, [x21], #0x10\n" "ld1 { v26.4s }, [x20], #0x10\n" "tbz x10, #1, 139f\n" - "mov x24, #0x38\n" "ldr d11, [x28], #0x8\n" + "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" "ldr d23, [x21], #0x8\n" @@ -1873,8 +1873,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v24.4s }, [x20], #0x10\n" "tbz x10, #1, 143f\n" - "mov x24, #0x18\n" "ldr d9, [x28], #0x8\n" + "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" "ldr d21, [x21], #0x8\n" @@ -2487,12 +2487,12 @@ void a64_hybrid_u8u32_dot_6x16 ( "ld1 { v16.4s }, [x22], #0x10\n" "ld1 { v20.4s }, [x21], #0x10\n" "ld1 { v24.4s }, [x20], #0x10\n" - "ld1 { v28.4s }, [x19], #0x10\n" "ld1 { v9.4s }, [x28], #0x10\n" "ld1 { v13.4s }, [x23], #0x10\n" "ld1 { v17.4s }, [x22], #0x10\n" "ld1 { v21.4s }, [x21], #0x10\n" "ld1 { v25.4s }, [x20], #0x10\n" + "ld1 { v28.4s }, [x19], #0x10\n" "ld1 { v29.4s }, [x19], #0x10\n" "tbz x10, #2, 174f\n" "ld1 { v10.4s }, [x28], #0x10\n" @@ -2502,8 +2502,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "ld1 { v26.4s }, [x20], #0x10\n" "ld1 { v30.4s }, [x19], #0x10\n" "tbz x10, #1, 173f\n" - "mov x24, #0x38\n" "ldr d11, [x28], #0x8\n" + "mov x24, #0x38\n" "ldr d15, [x23], #0x8\n" "ldr d19, [x22], #0x8\n" "ldr d23, [x21], #0x8\n" @@ -2563,8 +2563,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "ld1 { v24.4s }, [x20], #0x10\n" "ld1 { v28.4s }, [x19], #0x10\n" "tbz x10, #1, 177f\n" - "mov x24, #0x18\n" "ldr d9, [x28], #0x8\n" + "mov x24, #0x18\n" "ldr d13, [x23], #0x8\n" "ldr d17, [x22], #0x8\n" "ldr d21, [x21], #0x8\n" |