diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp | 80 |
1 files changed, 40 insertions, 40 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp index b3bde74635..5689f89781 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_dot_8x12/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021, 2023 Arm Limited. + * Copyright (c) 2019-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -33,108 +33,108 @@ void a64_interleaved_bf16fp32_dot_8x12( float *Cpanel, int ablocks, int bblocks, int K) { struct KernelArgs { + size_t bblocks = {}; size_t K = {}; const bfloat16 *Bpanel = {}; - size_t bblocks = {}; } ka; + ka.bblocks = bblocks; ka.K = (K/2) - 1; ka.Bpanel = Bpanel; - ka.bblocks = bblocks; __asm__ __volatile__( "1:" // Height loop - "ldr x23, [%x[args_ptr], %[offsetof_bblocks]]\n" - "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n" + "ldr x22, [%x[args_ptr], %[offsetof_bblocks]]\n" "mov x21, %x[Apanel]\n" + "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n" "2:" // Width loop - "ldr q4, [x22, #0x0]\n" - "ldr q5, [x22, #0x10]\n" + "ldr x19, [%x[args_ptr], %[offsetof_K]]\n" "mov %x[Apanel], x21\n" - "ldr q0, [%x[Apanel], #0x0]\n" - "ldr q1, [%x[Apanel], #0x10]\n" + "cmp x19, #0x2\n" "movi v8.16b, #0x0\n" - "ldr q6, [x22, #0x20]\n" - "ldr x20, [%x[args_ptr], %[offsetof_K]]\n" - "cmp x20, #0x2\n" "movi v9.16b, #0x0\n" "prfm pldl1keep, [%x[Apanel], #0x0]\n" "movi v10.16b, #0x0\n" "movi v11.16b, #0x0\n" - "prfm pldl1keep, [x22, #0x0]\n" + "prfm pldl1keep, [x20, #0x0]\n" "movi v12.16b, #0x0\n" "movi v13.16b, #0x0\n" - "prfm pldl1keep, [x22, #0x40]\n" + "prfm pldl1keep, [x20, #0x40]\n" "movi v14.16b, #0x0\n" "movi v15.16b, #0x0\n" "prfm pldl1keep, [%x[Apanel], #0x40]\n" "movi v16.16b, #0x0\n" "movi v17.16b, #0x0\n" - "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x20, #0x80]\n" "movi v18.16b, #0x0\n" "movi v19.16b, #0x0\n" + "ldr q0, [%x[Apanel], #0x0]\n" "movi v20.16b, #0x0\n" "movi v21.16b, #0x0\n" + "ldr q1, [%x[Apanel], #0x10]\n" "movi v22.16b, #0x0\n" "movi v23.16b, #0x0\n" + "ldr q4, [x20, #0x0]\n" "movi v24.16b, #0x0\n" "movi v25.16b, #0x0\n" + "ldr q5, [x20, #0x10]\n" "movi v26.16b, #0x0\n" "movi v27.16b, #0x0\n" + "ldr q6, [x20, #0x20]\n" "movi v28.16b, #0x0\n" "movi v29.16b, #0x0\n" "movi v30.16b, #0x0\n" "movi v31.16b, #0x0\n" "blt 4f\n" "3:" // main loop head - "ldr q2, [%x[Apanel], #0x20]\n" - "ldr q3, [%x[Apanel], #0x30]\n" ".inst 0x4f40f088 // bfdot v8.4s, v4.8h, v0.h[0]\n" ".inst 0x4f60f08b // bfdot v11.4s, v4.8h, v0.h[1]\n" + "ldr q2, [%x[Apanel], #0x20]\n" ".inst 0x4f40f88e // bfdot v14.4s, v4.8h, v0.h[2]\n" - "sub x20, x20, #0x2\n" ".inst 0x4f60f891 // bfdot v17.4s, v4.8h, v0.h[3]\n" + "ldr q3, [%x[Apanel], #0x30]\n" ".inst 0x4f41f094 // bfdot v20.4s, v4.8h, v1.h[0]\n" - "cmp x20, #0x2\n" ".inst 0x4f61f097 // bfdot v23.4s, v4.8h, v1.h[1]\n" + "sub x19, x19, #0x2\n" ".inst 0x4f41f89a // bfdot v26.4s, v4.8h, v1.h[2]\n" - "prfm pldl1keep, [%x[Apanel], #0x80]\n" ".inst 0x4f61f89d // bfdot v29.4s, v4.8h, v1.h[3]\n" - "ldr q4, [x22, #0x30]\n" + "ldr q4, [x20, #0x30]\n" ".inst 0x4f40f0a9 // bfdot v9.4s, v5.8h, v0.h[0]\n" ".inst 0x4f60f0ac // bfdot v12.4s, v5.8h, v0.h[1]\n" + "cmp x19, #0x2\n" ".inst 0x4f40f8af // bfdot v15.4s, v5.8h, v0.h[2]\n" - "add %x[Apanel], %x[Apanel], #0x40\n" ".inst 0x4f60f8b2 // bfdot v18.4s, v5.8h, v0.h[3]\n" + "prfm pldl1keep, [%x[Apanel], #0x80]\n" ".inst 0x4f41f0b5 // bfdot v21.4s, v5.8h, v1.h[0]\n" - "prfm pldl1keep, [x22, #0x100]\n" ".inst 0x4f61f0b8 // bfdot v24.4s, v5.8h, v1.h[1]\n" + "add %x[Apanel], %x[Apanel], #0x40\n" ".inst 0x4f41f8bb // bfdot v27.4s, v5.8h, v1.h[2]\n" - "prfm pldl1keep, [x22, #0x140]\n" ".inst 0x4f61f8be // bfdot v30.4s, v5.8h, v1.h[3]\n" - "ldr q5, [x22, #0x40]\n" + "ldr q5, [x20, #0x40]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" ".inst 0x4f60f0cd // bfdot v13.4s, v6.8h, v0.h[1]\n" + "prfm pldl1keep, [x20, #0x100]\n" ".inst 0x4f40f8d0 // bfdot v16.4s, v6.8h, v0.h[2]\n" ".inst 0x4f60f8d3 // bfdot v19.4s, v6.8h, v0.h[3]\n" - "ldr q0, [%x[Apanel], #0x0]\n" + "prfm pldl1keep, [x20, #0x140]\n" ".inst 0x4f41f0d6 // bfdot v22.4s, v6.8h, v1.h[0]\n" ".inst 0x4f61f0d9 // bfdot v25.4s, v6.8h, v1.h[1]\n" + "ldr q0, [%x[Apanel], #0x0]\n" ".inst 0x4f41f8dc // bfdot v28.4s, v6.8h, v1.h[2]\n" ".inst 0x4f61f8df // bfdot v31.4s, v6.8h, v1.h[3]\n" - "ldr q6, [x22, #0x50]\n" - "ldr q1, [%x[Apanel], #0x10]\n" - "add x22, x22, #0x60\n" + "ldr q6, [x20, #0x50]\n" + "add x20, x20, #0x60\n" ".inst 0x4f42f088 // bfdot v8.4s, v4.8h, v2.h[0]\n" ".inst 0x4f62f08b // bfdot v11.4s, v4.8h, v2.h[1]\n" + "ldr q1, [%x[Apanel], #0x10]\n" ".inst 0x4f42f88e // bfdot v14.4s, v4.8h, v2.h[2]\n" ".inst 0x4f62f891 // bfdot v17.4s, v4.8h, v2.h[3]\n" ".inst 0x4f43f094 // bfdot v20.4s, v4.8h, v3.h[0]\n" ".inst 0x4f63f097 // bfdot v23.4s, v4.8h, v3.h[1]\n" ".inst 0x4f43f89a // bfdot v26.4s, v4.8h, v3.h[2]\n" ".inst 0x4f63f89d // bfdot v29.4s, v4.8h, v3.h[3]\n" - "ldr q4, [x22, #0x0]\n" + "ldr q4, [x20, #0x0]\n" ".inst 0x4f42f0a9 // bfdot v9.4s, v5.8h, v2.h[0]\n" ".inst 0x4f62f0ac // bfdot v12.4s, v5.8h, v2.h[1]\n" ".inst 0x4f42f8af // bfdot v15.4s, v5.8h, v2.h[2]\n" @@ -143,7 +143,7 @@ void a64_interleaved_bf16fp32_dot_8x12( ".inst 0x4f63f0b8 // bfdot v24.4s, v5.8h, v3.h[1]\n" ".inst 0x4f43f8bb // bfdot v27.4s, v5.8h, v3.h[2]\n" ".inst 0x4f63f8be // bfdot v30.4s, v5.8h, v3.h[3]\n" - "ldr q5, [x22, #0x10]\n" + "ldr q5, [x20, #0x10]\n" ".inst 0x4f42f0ca // bfdot v10.4s, v6.8h, v2.h[0]\n" ".inst 0x4f62f0cd // bfdot v13.4s, v6.8h, v2.h[1]\n" ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" @@ -152,13 +152,13 @@ void a64_interleaved_bf16fp32_dot_8x12( ".inst 0x4f63f0d9 // bfdot v25.4s, v6.8h, v3.h[1]\n" ".inst 0x4f43f8dc // bfdot v28.4s, v6.8h, v3.h[2]\n" ".inst 0x4f63f8df // bfdot v31.4s, v6.8h, v3.h[3]\n" - "ldr q6, [x22, #0x20]\n" + "ldr q6, [x20, #0x20]\n" "bge 3b\n" "4:" // main loop skip "add %x[Apanel], %x[Apanel], #0x20\n" ".inst 0x4f40f088 // bfdot v8.4s, v4.8h, v0.h[0]\n" ".inst 0x4f60f08b // bfdot v11.4s, v4.8h, v0.h[1]\n" - "add x22, x22, #0x30\n" + "add x20, x20, #0x30\n" ".inst 0x4f40f88e // bfdot v14.4s, v4.8h, v0.h[2]\n" ".inst 0x4f60f891 // bfdot v17.4s, v4.8h, v0.h[3]\n" ".inst 0x4f41f094 // bfdot v20.4s, v4.8h, v1.h[0]\n" @@ -181,19 +181,19 @@ void a64_interleaved_bf16fp32_dot_8x12( ".inst 0x4f61f0d9 // bfdot v25.4s, v6.8h, v1.h[1]\n" ".inst 0x4f41f8dc // bfdot v28.4s, v6.8h, v1.h[2]\n" ".inst 0x4f61f8df // bfdot v31.4s, v6.8h, v1.h[3]\n" - "cbz x20, 5f\n" + "cbz x19, 5f\n" "ldr q0, [%x[Apanel], #0x0]\n" "ldr q1, [%x[Apanel], #0x10]\n" "add %x[Apanel], %x[Apanel], #0x20\n" - "ldr q7, [x22, #0x0]\n" - "ldr q4, [x22, #0x10]\n" + "ldr q7, [x20, #0x0]\n" + "ldr q4, [x20, #0x10]\n" ".inst 0x4f40f0e8 // bfdot v8.4s, v7.8h, v0.h[0]\n" - "ldr q5, [x22, #0x20]\n" + "ldr q5, [x20, #0x20]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f40f8ee // bfdot v14.4s, v7.8h, v0.h[2]\n" + "add x20, x20, #0x30\n" ".inst 0x4f60f8f1 // bfdot v17.4s, v7.8h, v0.h[3]\n" ".inst 0x4f41f0f4 // bfdot v20.4s, v7.8h, v1.h[0]\n" - "add x22, x22, #0x30\n" ".inst 0x4f61f0f7 // bfdot v23.4s, v7.8h, v1.h[1]\n" ".inst 0x4f41f8fa // bfdot v26.4s, v7.8h, v1.h[2]\n" ".inst 0x4f61f8fd // bfdot v29.4s, v7.8h, v1.h[3]\n" @@ -214,7 +214,7 @@ void a64_interleaved_bf16fp32_dot_8x12( ".inst 0x4f41f8bc // bfdot v28.4s, v5.8h, v1.h[2]\n" ".inst 0x4f61f8bf // bfdot v31.4s, v5.8h, v1.h[3]\n" "5:" // multiply loop done - "subs x23, x23, #0x1\n" + "subs x22, x22, #0x1\n" "str q8, [%x[Cpanel], #0x0]\n" "str q9, [%x[Cpanel], #0x10]\n" "str q10, [%x[Cpanel], #0x20]\n" @@ -245,7 +245,7 @@ void a64_interleaved_bf16fp32_dot_8x12( "bne 1b\n" : [Apanel] "+&r" (Apanel), [Cpanel] "+&r" (Cpanel), [ablocks] "+&r" (ablocks) : [args_ptr] "r" (&ka), [offsetof_Bpanel] "I" (offsetof(KernelArgs, Bpanel)), [offsetof_K] "I" (offsetof(KernelArgs, K)), [offsetof_bblocks] "I" (offsetof(KernelArgs, bblocks)) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x19", "x20", "x21", "x22" ); } |