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 | 134 |
1 files changed, 67 insertions, 67 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 94c72a31c9..2938639048 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 Arm Limited. + * Copyright (c) 2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -33,46 +33,46 @@ void a64_interleaved_bf16fp32_mmla_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/4) - 1; ka.Bpanel = Bpanel; + ka.bblocks = bblocks; __asm__ __volatile__( "1:" // Height loop - "ldr x22, [%x[args_ptr], %[offsetof_bblocks]]\n" + "ldr x23, [%x[args_ptr], %[offsetof_bblocks]]\n" + "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n" "mov x21, %x[Apanel]\n" - "ldr x20, [%x[args_ptr], %[offsetof_Bpanel]]\n" "2:" // Width loop - "ldr x19, [%x[args_ptr], %[offsetof_K]]\n" + "ldr q4, [x22, #0x0]\n" + "ldr q5, [x22, #0x10]\n" "mov %x[Apanel], x21\n" - "cmp x19, #0x2\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 v9.16b, #0x0\n" - "ldr q4, [x20, #0x0]\n" "movi v10.16b, #0x0\n" + "add x22, x22, #0x20\n" "movi v11.16b, #0x0\n" - "ldr q0, [%x[Apanel], #0x0]\n" "movi v12.16b, #0x0\n" + "add %x[Apanel], %x[Apanel], #0x30\n" "movi v13.16b, #0x0\n" - "ldr q1, [%x[Apanel], #0x10]\n" "movi v14.16b, #0x0\n" "movi v15.16b, #0x0\n" - "ldr q5, [x20, #0x10]\n" "movi v16.16b, #0x0\n" "movi v17.16b, #0x0\n" - "ldr q2, [%x[Apanel], #0x20]\n" "movi v18.16b, #0x0\n" "movi v19.16b, #0x0\n" - "add x20, x20, #0x20\n" "movi v20.16b, #0x0\n" "movi v21.16b, #0x0\n" - "add %x[Apanel], %x[Apanel], #0x30\n" "movi v22.16b, #0x0\n" "movi v23.16b, #0x0\n" "movi v24.16b, #0x0\n" @@ -86,149 +86,149 @@ void a64_interleaved_bf16fp32_mmla_8x12( "blt 4f\n" "3:" // main loop head "ldr q3, [%x[Apanel], #0x0]\n" + "ldr q6, [x22, #0x0]\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec2e // bfmmla v14.4s, v1.8h, v4.8h\n" + "ldr q7, [x22, #0x10]\n" ".inst 0x6e45ec0b // bfmmla v11.4s, v0.8h, v5.8h\n" + ".inst 0x6e44ec2e // bfmmla v14.4s, v1.8h, v4.8h\n" ".inst 0x6e45ec31 // bfmmla v17.4s, v1.8h, v5.8h\n" - "ldr q6, [x20, #0x0]\n" ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" + "sub x20, x20, #0x2\n" ".inst 0x6e45ec57 // bfmmla v23.4s, v2.8h, v5.8h\n" - "ldr q7, [x20, #0x10]\n" ".inst 0x6e44ec7a // bfmmla v26.4s, v3.8h, v4.8h\n" + "ldr q4, [x22, #0x20]\n" ".inst 0x6e45ec7d // bfmmla v29.4s, v3.8h, v5.8h\n" - "ldr q4, [x20, #0x20]\n" - "ldr q5, [x20, #0x30]\n" + "ldr q5, [x22, #0x30]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" + ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + "cmp x20, #0x2\n" + ".inst 0x6e47ec32 // bfmmla v18.4s, v1.8h, v7.8h\n" ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" + ".inst 0x6e47ec58 // bfmmla v24.4s, v2.8h, v7.8h\n" ".inst 0x6e46ec7b // bfmmla v27.4s, v3.8h, v6.8h\n" - "ldr q6, [x20, #0x40]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" + "ldr q6, [x22, #0x40]\n" + ".inst 0x6e47ec7e // bfmmla v30.4s, v3.8h, v7.8h\n" + "ldr q7, [x22, #0x50]\n" ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "sub x19, x19, #0x2\n" ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n" - ".inst 0x6e47ec32 // bfmmla v18.4s, v1.8h, v7.8h\n" "ldr q0, [%x[Apanel], #0x10]\n" ".inst 0x6e44ec30 // bfmmla v16.4s, v1.8h, v4.8h\n" ".inst 0x6e45ec33 // bfmmla v19.4s, v1.8h, v5.8h\n" "ldr q1, [%x[Apanel], #0x20]\n" - ".inst 0x6e47ec58 // bfmmla v24.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec7e // bfmmla v30.4s, v3.8h, v7.8h\n" - "ldr q7, [x20, #0x50]\n" ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" ".inst 0x6e45ec59 // bfmmla v25.4s, v2.8h, v5.8h\n" "ldr q2, [%x[Apanel], #0x30]\n" ".inst 0x6e44ec7c // bfmmla v28.4s, v3.8h, v4.8h\n" + "ldr q4, [x22, #0x60]\n" ".inst 0x6e45ec7f // bfmmla v31.4s, v3.8h, v5.8h\n" "ldr q3, [%x[Apanel], #0x40]\n" + "ldr q5, [x22, #0x70]\n" ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - "ldr q4, [x20, #0x60]\n" ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" ".inst 0x6e47ec31 // bfmmla v17.4s, v1.8h, v7.8h\n" - "ldr q5, [x20, #0x70]\n" ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" - "cmp x19, #0x2\n" ".inst 0x6e46ec7a // bfmmla v26.4s, v3.8h, v6.8h\n" + "ldr q6, [x22, #0x80]\n" ".inst 0x6e47ec7d // bfmmla v29.4s, v3.8h, v7.8h\n" - "ldr q6, [x20, #0x80]\n" - "ldr q7, [x20, #0x90]\n" + "ldr q7, [x22, #0x90]\n" ".inst 0x6e44ec09 // bfmmla v9.4s, v0.8h, v4.8h\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" ".inst 0x6e44ec2f // bfmmla v15.4s, v1.8h, v4.8h\n" + ".inst 0x6e45ec32 // bfmmla v18.4s, v1.8h, v5.8h\n" ".inst 0x6e44ec55 // bfmmla v21.4s, v2.8h, v4.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" ".inst 0x6e44ec7b // bfmmla v27.4s, v3.8h, v4.8h\n" - "ldr q4, [x20, #0xa0]\n" - ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" + "ldr q4, [x22, #0xa0]\n" + ".inst 0x6e45ec7e // bfmmla v30.4s, v3.8h, v5.8h\n" + "ldr q5, [x22, #0xb0]\n" ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - ".inst 0x6e45ec32 // bfmmla v18.4s, v1.8h, v5.8h\n" "ldr q0, [%x[Apanel], #0x50]\n" ".inst 0x6e46ec30 // bfmmla v16.4s, v1.8h, v6.8h\n" ".inst 0x6e47ec33 // bfmmla v19.4s, v1.8h, v7.8h\n" "ldr q1, [%x[Apanel], #0x60]\n" - ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" - ".inst 0x6e45ec7e // bfmmla v30.4s, v3.8h, v5.8h\n" - "ldr q5, [x20, #0xb0]\n" ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" ".inst 0x6e47ec59 // bfmmla v25.4s, v2.8h, v7.8h\n" "ldr q2, [%x[Apanel], #0x70]\n" ".inst 0x6e46ec7c // bfmmla v28.4s, v3.8h, v6.8h\n" ".inst 0x6e47ec7f // bfmmla v31.4s, v3.8h, v7.8h\n" "add %x[Apanel], %x[Apanel], #0x80\n" - "add x20, x20, #0xc0\n" + "add x22, x22, #0xc0\n" "bge 3b\n" "4:" // main loop skip "ldr q3, [%x[Apanel], #0x0]\n" + "ldr q6, [x22, #0x0]\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec2e // bfmmla v14.4s, v1.8h, v4.8h\n" + "ldr q7, [x22, #0x10]\n" ".inst 0x6e45ec0b // bfmmla v11.4s, v0.8h, v5.8h\n" + ".inst 0x6e44ec2e // bfmmla v14.4s, v1.8h, v4.8h\n" ".inst 0x6e45ec31 // bfmmla v17.4s, v1.8h, v5.8h\n" - "ldr q6, [x20, #0x0]\n" ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" + "add %x[Apanel], %x[Apanel], #0x10\n" ".inst 0x6e45ec57 // bfmmla v23.4s, v2.8h, v5.8h\n" - "ldr q7, [x20, #0x10]\n" ".inst 0x6e44ec7a // bfmmla v26.4s, v3.8h, v4.8h\n" + "ldr q4, [x22, #0x20]\n" ".inst 0x6e45ec7d // bfmmla v29.4s, v3.8h, v5.8h\n" - "ldr q4, [x20, #0x20]\n" - "ldr q5, [x20, #0x30]\n" + "ldr q5, [x22, #0x30]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" + ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + "add x22, x22, #0x40\n" + ".inst 0x6e47ec32 // bfmmla v18.4s, v1.8h, v7.8h\n" ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" + ".inst 0x6e47ec58 // bfmmla v24.4s, v2.8h, v7.8h\n" ".inst 0x6e46ec7b // bfmmla v27.4s, v3.8h, v6.8h\n" - "add %x[Apanel], %x[Apanel], #0x10\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" + ".inst 0x6e47ec7e // bfmmla v30.4s, v3.8h, v7.8h\n" ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "add x20, x20, #0x40\n" ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n" - ".inst 0x6e47ec32 // bfmmla v18.4s, v1.8h, v7.8h\n" ".inst 0x6e44ec30 // bfmmla v16.4s, v1.8h, v4.8h\n" ".inst 0x6e45ec33 // bfmmla v19.4s, v1.8h, v5.8h\n" - ".inst 0x6e47ec58 // bfmmla v24.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec7e // bfmmla v30.4s, v3.8h, v7.8h\n" ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" ".inst 0x6e45ec59 // bfmmla v25.4s, v2.8h, v5.8h\n" ".inst 0x6e44ec7c // bfmmla v28.4s, v3.8h, v4.8h\n" ".inst 0x6e45ec7f // bfmmla v31.4s, v3.8h, v5.8h\n" - "cbz x19, 5f\n" - "ldr q6, [x20, #0x0]\n" + "cbz x20, 5f\n" + "ldr q6, [x22, #0x0]\n" "ldr q0, [%x[Apanel], #0x0]\n" ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" "ldr q1, [%x[Apanel], #0x10]\n" - "ldr q7, [x20, #0x10]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" + "ldr q7, [x22, #0x10]\n" + ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" "ldr q2, [%x[Apanel], #0x20]\n" "ldr q3, [%x[Apanel], #0x30]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" + "ldr q4, [x22, #0x20]\n" + "ldr q5, [x22, #0x30]\n" ".inst 0x6e47ec31 // bfmmla v17.4s, v1.8h, v7.8h\n" ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q4, [x20, #0x20]\n" ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" + "add %x[Apanel], %x[Apanel], #0x40\n" ".inst 0x6e46ec7a // bfmmla v26.4s, v3.8h, v6.8h\n" - "ldr q5, [x20, #0x30]\n" + "ldr q6, [x22, #0x40]\n" ".inst 0x6e47ec7d // bfmmla v29.4s, v3.8h, v7.8h\n" - "ldr q6, [x20, #0x40]\n" - "ldr q7, [x20, #0x50]\n" + "ldr q7, [x22, #0x50]\n" ".inst 0x6e44ec09 // bfmmla v9.4s, v0.8h, v4.8h\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" ".inst 0x6e44ec2f // bfmmla v15.4s, v1.8h, v4.8h\n" - "add x20, x20, #0x60\n" + ".inst 0x6e45ec32 // bfmmla v18.4s, v1.8h, v5.8h\n" + "add x22, x22, #0x60\n" ".inst 0x6e44ec55 // bfmmla v21.4s, v2.8h, v4.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" ".inst 0x6e44ec7b // bfmmla v27.4s, v3.8h, v4.8h\n" - "add %x[Apanel], %x[Apanel], #0x40\n" - ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec7e // bfmmla v30.4s, v3.8h, v5.8h\n" ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - ".inst 0x6e45ec32 // bfmmla v18.4s, v1.8h, v5.8h\n" ".inst 0x6e46ec30 // bfmmla v16.4s, v1.8h, v6.8h\n" ".inst 0x6e47ec33 // bfmmla v19.4s, v1.8h, v7.8h\n" - ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" - ".inst 0x6e45ec7e // bfmmla v30.4s, v3.8h, v5.8h\n" ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" ".inst 0x6e47ec59 // bfmmla v25.4s, v2.8h, v7.8h\n" ".inst 0x6e46ec7c // bfmmla v28.4s, v3.8h, v6.8h\n" ".inst 0x6e47ec7f // bfmmla v31.4s, v3.8h, v7.8h\n" "5:" // multiply loop done - "subs x22, x22, #0x1\n" + "subs x23, x23, #0x1\n" "uzp1 v4.2d, v8.2d, v11.2d\n" "uzp2 v8.2d, v8.2d, v11.2d\n" "uzp1 v11.2d, v9.2d, v12.2d\n" @@ -283,7 +283,7 @@ void a64_interleaved_bf16fp32_mmla_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", "x19", "x20", "x21", "x22" + : "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" ); } |