diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp | 52 |
1 files changed, 26 insertions, 26 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp index 8485820c7c..f4493c6855 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -55,21 +55,21 @@ void a64_interleaved_bf16fp32_mmla_8x12( "ldr q4, [x22, #0x0]\n" "ldr q5, [x22, #0x10]\n" "mov %x[Apanel], x21\n" - "ldr q0, [%x[Apanel], #0x0]\n" - "ldr q1, [%x[Apanel], #0x10]\n" - "movi v8.16b, #0x0\n" - "ldr q2, [%x[Apanel], #0x20]\n" "ldr x20, [%x[args_ptr], %[offsetof_K]]\n" - "cmp x20, #0x2\n" + "movi v8.16b, #0x0\n" "movi v9.16b, #0x0\n" "movi v10.16b, #0x0\n" - "add x22, x22, #0x20\n" "movi v11.16b, #0x0\n" + "add x22, x22, #0x20\n" + "ldr q0, [%x[Apanel], #0x0]\n" + "ldr q1, [%x[Apanel], #0x10]\n" "movi v12.16b, #0x0\n" - "add %x[Apanel], %x[Apanel], #0x30\n" + "ldr q2, [%x[Apanel], #0x20]\n" + "cmp x20, #0x2\n" "movi v13.16b, #0x0\n" "movi v14.16b, #0x0\n" "movi v15.16b, #0x0\n" + "add %x[Apanel], %x[Apanel], #0x30\n" "movi v16.16b, #0x0\n" "movi v17.16b, #0x0\n" "movi v18.16b, #0x0\n" @@ -196,19 +196,19 @@ void a64_interleaved_bf16fp32_mmla_8x12( "cbz x20, 5f\n" "ldr q1, [x22, #0x0]\n" "ldr q7, [%x[Apanel], #0x0]\n" - ".inst 0x6e41ece8 // bfmmla v8.4s, v7.8h, v1.8h\n" "ldr q6, [%x[Apanel], #0x10]\n" "ldr q0, [x22, #0x10]\n" - ".inst 0x6e40eceb // bfmmla v11.4s, v7.8h, v0.8h\n" "ldr q5, [%x[Apanel], #0x20]\n" "ldr q4, [%x[Apanel], #0x30]\n" - ".inst 0x6e41ecce // bfmmla v14.4s, v6.8h, v1.8h\n" + "add %x[Apanel], %x[Apanel], #0x40\n" "ldr q3, [x22, #0x20]\n" "ldr q2, [x22, #0x30]\n" + ".inst 0x6e41ece8 // bfmmla v8.4s, v7.8h, v1.8h\n" + ".inst 0x6e40eceb // bfmmla v11.4s, v7.8h, v0.8h\n" + ".inst 0x6e41ecce // bfmmla v14.4s, v6.8h, v1.8h\n" ".inst 0x6e40ecd1 // bfmmla v17.4s, v6.8h, v0.8h\n" ".inst 0x6e41ecb4 // bfmmla v20.4s, v5.8h, v1.8h\n" ".inst 0x6e40ecb7 // bfmmla v23.4s, v5.8h, v0.8h\n" - "add %x[Apanel], %x[Apanel], #0x40\n" ".inst 0x6e41ec9a // bfmmla v26.4s, v4.8h, v1.8h\n" "ldr q1, [x22, #0x40]\n" ".inst 0x6e40ec9d // bfmmla v29.4s, v4.8h, v0.8h\n" @@ -216,8 +216,8 @@ void a64_interleaved_bf16fp32_mmla_8x12( ".inst 0x6e43ece9 // bfmmla v9.4s, v7.8h, v3.8h\n" ".inst 0x6e42ecec // bfmmla v12.4s, v7.8h, v2.8h\n" ".inst 0x6e43eccf // bfmmla v15.4s, v6.8h, v3.8h\n" - ".inst 0x6e42ecd2 // bfmmla v18.4s, v6.8h, v2.8h\n" "add x22, x22, #0x60\n" + ".inst 0x6e42ecd2 // bfmmla v18.4s, v6.8h, v2.8h\n" ".inst 0x6e43ecb5 // bfmmla v21.4s, v5.8h, v3.8h\n" ".inst 0x6e42ecb8 // bfmmla v24.4s, v5.8h, v2.8h\n" ".inst 0x6e43ec9b // bfmmla v27.4s, v4.8h, v3.8h\n" @@ -232,41 +232,41 @@ void a64_interleaved_bf16fp32_mmla_8x12( ".inst 0x6e40ec9f // bfmmla v31.4s, v4.8h, v0.8h\n" "5:" // multiply loop done "subs x23, x23, #0x1\n" - "uzp1 v0.2d, v8.2d, v11.2d\n" + "uzp1 v2.2d, v8.2d, v11.2d\n" "uzp2 v8.2d, v8.2d, v11.2d\n" "uzp1 v1.2d, v9.2d, v12.2d\n" "uzp2 v9.2d, v9.2d, v12.2d\n" - "str q0, [%x[Cpanel], #0x0]\n" "uzp1 v0.2d, v10.2d, v13.2d\n" "uzp2 v10.2d, v10.2d, v13.2d\n" - "str q1, [%x[Cpanel], #0x10]\n" - "str q0, [%x[Cpanel], #0x20]\n" - "uzp1 v0.2d, v14.2d, v17.2d\n" + "str q2, [%x[Cpanel], #0x0]\n" + "uzp1 v3.2d, v14.2d, v17.2d\n" "uzp2 v14.2d, v14.2d, v17.2d\n" - "str q8, [%x[Cpanel], #0x30]\n" + "str q1, [%x[Cpanel], #0x10]\n" "uzp1 v2.2d, v15.2d, v18.2d\n" "uzp2 v15.2d, v15.2d, v18.2d\n" - "str q9, [%x[Cpanel], #0x40]\n" + "str q0, [%x[Cpanel], #0x20]\n" "uzp1 v17.2d, v16.2d, v19.2d\n" "uzp2 v16.2d, v16.2d, v19.2d\n" - "str q10, [%x[Cpanel], #0x50]\n" + "str q8, [%x[Cpanel], #0x30]\n" "uzp1 v1.2d, v20.2d, v23.2d\n" "uzp2 v20.2d, v20.2d, v23.2d\n" - "str q0, [%x[Cpanel], #0x60]\n" + "str q9, [%x[Cpanel], #0x40]\n" "uzp1 v0.2d, v21.2d, v24.2d\n" "uzp2 v21.2d, v21.2d, v24.2d\n" - "str q2, [%x[Cpanel], #0x70]\n" + "str q10, [%x[Cpanel], #0x50]\n" "uzp1 v23.2d, v22.2d, v25.2d\n" "uzp2 v22.2d, v22.2d, v25.2d\n" - "str q17, [%x[Cpanel], #0x80]\n" + "str q3, [%x[Cpanel], #0x60]\n" "uzp1 v19.2d, v26.2d, v29.2d\n" "uzp2 v26.2d, v26.2d, v29.2d\n" - "str q14, [%x[Cpanel], #0x90]\n" + "str q2, [%x[Cpanel], #0x70]\n" "uzp1 v18.2d, v27.2d, v30.2d\n" "uzp2 v27.2d, v27.2d, v30.2d\n" - "str q15, [%x[Cpanel], #0xa0]\n" + "str q17, [%x[Cpanel], #0x80]\n" "uzp1 v17.2d, v28.2d, v31.2d\n" "uzp2 v28.2d, v28.2d, v31.2d\n" + "str q14, [%x[Cpanel], #0x90]\n" + "str q15, [%x[Cpanel], #0xa0]\n" "str q16, [%x[Cpanel], #0xb0]\n" "str q1, [%x[Cpanel], #0xc0]\n" "str q0, [%x[Cpanel], #0xd0]\n" |