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 | 201 |
1 files changed, 102 insertions, 99 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 2938639048..8485820c7c 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 @@ -29,8 +29,12 @@ namespace arm_gemm { void a64_interleaved_bf16fp32_mmla_8x12( - const bfloat16 *Apanel, const bfloat16 *Bpanel, - float *Cpanel, int ablocks, int bblocks, int K) { + const bfloat16 *Apanel, + const bfloat16 *Bpanel, + float *Cpanel, + int ablocks, + int bblocks, + int K) { struct KernelArgs { size_t K = {}; @@ -43,7 +47,6 @@ void a64_interleaved_bf16fp32_mmla_8x12( ka.bblocks = bblocks; __asm__ __volatile__( - "1:" // Height loop "ldr x23, [%x[args_ptr], %[offsetof_bblocks]]\n" "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n" @@ -85,31 +88,31 @@ void a64_interleaved_bf16fp32_mmla_8x12( "movi v31.16b, #0x0\n" "blt 4f\n" "3:" // main loop head - "ldr q3, [%x[Apanel], #0x0]\n" - "ldr q6, [x22, #0x0]\n" + "ldr q6, [%x[Apanel], #0x0]\n" + "ldr q7, [x22, #0x0]\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q7, [x22, #0x10]\n" + "ldr q3, [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" ".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" - ".inst 0x6e44ec7a // bfmmla v26.4s, v3.8h, v4.8h\n" + ".inst 0x6e44ecda // bfmmla v26.4s, v6.8h, v4.8h\n" "ldr q4, [x22, #0x20]\n" - ".inst 0x6e45ec7d // bfmmla v29.4s, v3.8h, v5.8h\n" + ".inst 0x6e45ecdd // bfmmla v29.4s, v6.8h, v5.8h\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" + ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + ".inst 0x6e43ec0c // bfmmla v12.4s, v0.8h, v3.8h\n" + ".inst 0x6e47ec2f // bfmmla v15.4s, v1.8h, v7.8h\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, [x22, #0x40]\n" - ".inst 0x6e47ec7e // bfmmla v30.4s, v3.8h, v7.8h\n" - "ldr q7, [x22, #0x50]\n" + ".inst 0x6e43ec32 // bfmmla v18.4s, v1.8h, v3.8h\n" + ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" + ".inst 0x6e43ec58 // bfmmla v24.4s, v2.8h, v3.8h\n" + ".inst 0x6e47ecdb // bfmmla v27.4s, v6.8h, v7.8h\n" + "ldr q7, [x22, #0x40]\n" + ".inst 0x6e43ecde // bfmmla v30.4s, v6.8h, v3.8h\n" + "ldr q3, [x22, #0x50]\n" ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n" "ldr q0, [%x[Apanel], #0x10]\n" @@ -119,42 +122,42 @@ void a64_interleaved_bf16fp32_mmla_8x12( ".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" + ".inst 0x6e44ecdc // bfmmla v28.4s, v6.8h, v4.8h\n" "ldr q4, [x22, #0x60]\n" - ".inst 0x6e45ec7f // bfmmla v31.4s, v3.8h, v5.8h\n" - "ldr q3, [%x[Apanel], #0x40]\n" + ".inst 0x6e45ecdf // bfmmla v31.4s, v6.8h, v5.8h\n" + "ldr q6, [%x[Apanel], #0x40]\n" "ldr q5, [x22, #0x70]\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\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" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\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 q7, [x22, #0x90]\n" + ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" + ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n" + ".inst 0x6e47ec2e // bfmmla v14.4s, v1.8h, v7.8h\n" + ".inst 0x6e43ec31 // bfmmla v17.4s, v1.8h, v3.8h\n" + ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" + ".inst 0x6e43ec57 // bfmmla v23.4s, v2.8h, v3.8h\n" + ".inst 0x6e47ecda // bfmmla v26.4s, v6.8h, v7.8h\n" + "ldr q7, [x22, #0x80]\n" + ".inst 0x6e43ecdd // bfmmla v29.4s, v6.8h, v3.8h\n" + "ldr q3, [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" + ".inst 0x6e44ecdb // bfmmla v27.4s, v6.8h, v4.8h\n" "ldr q4, [x22, #0xa0]\n" - ".inst 0x6e45ec7e // bfmmla v30.4s, v3.8h, v5.8h\n" + ".inst 0x6e45ecde // bfmmla v30.4s, v6.8h, 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 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e43ec0d // bfmmla v13.4s, v0.8h, v3.8h\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" + ".inst 0x6e47ec30 // bfmmla v16.4s, v1.8h, v7.8h\n" + ".inst 0x6e43ec33 // bfmmla v19.4s, v1.8h, v3.8h\n" "ldr q1, [%x[Apanel], #0x60]\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec59 // bfmmla v25.4s, v2.8h, v7.8h\n" + ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n" + ".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\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" + ".inst 0x6e47ecdc // bfmmla v28.4s, v6.8h, v7.8h\n" + ".inst 0x6e43ecdf // bfmmla v31.4s, v6.8h, v3.8h\n" "add %x[Apanel], %x[Apanel], #0x80\n" "add x22, x22, #0xc0\n" "bge 3b\n" @@ -191,89 +194,89 @@ void a64_interleaved_bf16fp32_mmla_8x12( ".inst 0x6e44ec7c // bfmmla v28.4s, v3.8h, v4.8h\n" ".inst 0x6e45ec7f // bfmmla v31.4s, v3.8h, v5.8h\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, [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 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" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\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" + "ldr q3, [x22, #0x20]\n" + "ldr q2, [x22, #0x30]\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 0x6e46ec7a // bfmmla v26.4s, v3.8h, v6.8h\n" - "ldr q6, [x22, #0x40]\n" - ".inst 0x6e47ec7d // bfmmla v29.4s, v3.8h, v7.8h\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" - ".inst 0x6e45ec32 // bfmmla v18.4s, v1.8h, v5.8h\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" + "ldr q0, [x22, #0x50]\n" + ".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 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" - ".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 0x6e46ec30 // bfmmla v16.4s, v1.8h, v6.8h\n" - ".inst 0x6e47ec33 // bfmmla v19.4s, v1.8h, v7.8h\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" + ".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" + ".inst 0x6e42ec9e // bfmmla v30.4s, v4.8h, v2.8h\n" + ".inst 0x6e41ecea // bfmmla v10.4s, v7.8h, v1.8h\n" + ".inst 0x6e40eced // bfmmla v13.4s, v7.8h, v0.8h\n" + ".inst 0x6e41ecd0 // bfmmla v16.4s, v6.8h, v1.8h\n" + ".inst 0x6e40ecd3 // bfmmla v19.4s, v6.8h, v0.8h\n" + ".inst 0x6e41ecb6 // bfmmla v22.4s, v5.8h, v1.8h\n" + ".inst 0x6e40ecb9 // bfmmla v25.4s, v5.8h, v0.8h\n" + ".inst 0x6e41ec9c // bfmmla v28.4s, v4.8h, v1.8h\n" + ".inst 0x6e40ec9f // bfmmla v31.4s, v4.8h, v0.8h\n" "5:" // multiply loop done "subs x23, x23, #0x1\n" - "uzp1 v4.2d, v8.2d, v11.2d\n" + "uzp1 v0.2d, v8.2d, v11.2d\n" "uzp2 v8.2d, v8.2d, v11.2d\n" - "uzp1 v11.2d, v9.2d, v12.2d\n" + "uzp1 v1.2d, v9.2d, v12.2d\n" "uzp2 v9.2d, v9.2d, v12.2d\n" - "str q4, [%x[Cpanel], #0x0]\n" - "uzp1 v12.2d, v10.2d, v13.2d\n" + "str q0, [%x[Cpanel], #0x0]\n" + "uzp1 v0.2d, v10.2d, v13.2d\n" "uzp2 v10.2d, v10.2d, v13.2d\n" - "str q11, [%x[Cpanel], #0x10]\n" - "str q12, [%x[Cpanel], #0x20]\n" - "uzp1 v13.2d, v14.2d, v17.2d\n" + "str q1, [%x[Cpanel], #0x10]\n" + "str q0, [%x[Cpanel], #0x20]\n" + "uzp1 v0.2d, v14.2d, v17.2d\n" "uzp2 v14.2d, v14.2d, v17.2d\n" "str q8, [%x[Cpanel], #0x30]\n" - "uzp1 v17.2d, v15.2d, v18.2d\n" + "uzp1 v2.2d, v15.2d, v18.2d\n" "uzp2 v15.2d, v15.2d, v18.2d\n" "str q9, [%x[Cpanel], #0x40]\n" - "uzp1 v18.2d, v16.2d, v19.2d\n" + "uzp1 v17.2d, v16.2d, v19.2d\n" "uzp2 v16.2d, v16.2d, v19.2d\n" "str q10, [%x[Cpanel], #0x50]\n" - "uzp1 v19.2d, v20.2d, v23.2d\n" + "uzp1 v1.2d, v20.2d, v23.2d\n" "uzp2 v20.2d, v20.2d, v23.2d\n" - "str q13, [%x[Cpanel], #0x60]\n" - "uzp1 v23.2d, v21.2d, v24.2d\n" + "str q0, [%x[Cpanel], #0x60]\n" + "uzp1 v0.2d, v21.2d, v24.2d\n" "uzp2 v21.2d, v21.2d, v24.2d\n" - "str q17, [%x[Cpanel], #0x70]\n" - "uzp1 v24.2d, v22.2d, v25.2d\n" + "str q2, [%x[Cpanel], #0x70]\n" + "uzp1 v23.2d, v22.2d, v25.2d\n" "uzp2 v22.2d, v22.2d, v25.2d\n" - "str q18, [%x[Cpanel], #0x80]\n" - "uzp1 v25.2d, v26.2d, v29.2d\n" + "str q17, [%x[Cpanel], #0x80]\n" + "uzp1 v19.2d, v26.2d, v29.2d\n" "uzp2 v26.2d, v26.2d, v29.2d\n" "str q14, [%x[Cpanel], #0x90]\n" - "uzp1 v29.2d, v27.2d, v30.2d\n" + "uzp1 v18.2d, v27.2d, v30.2d\n" "uzp2 v27.2d, v27.2d, v30.2d\n" "str q15, [%x[Cpanel], #0xa0]\n" - "uzp1 v30.2d, v28.2d, v31.2d\n" + "uzp1 v17.2d, v28.2d, v31.2d\n" "uzp2 v28.2d, v28.2d, v31.2d\n" "str q16, [%x[Cpanel], #0xb0]\n" - "str q19, [%x[Cpanel], #0xc0]\n" - "str q23, [%x[Cpanel], #0xd0]\n" - "str q24, [%x[Cpanel], #0xe0]\n" + "str q1, [%x[Cpanel], #0xc0]\n" + "str q0, [%x[Cpanel], #0xd0]\n" + "str q23, [%x[Cpanel], #0xe0]\n" "str q20, [%x[Cpanel], #0xf0]\n" "str q21, [%x[Cpanel], #0x100]\n" "str q22, [%x[Cpanel], #0x110]\n" - "str q25, [%x[Cpanel], #0x120]\n" - "str q29, [%x[Cpanel], #0x130]\n" - "str q30, [%x[Cpanel], #0x140]\n" + "str q19, [%x[Cpanel], #0x120]\n" + "str q18, [%x[Cpanel], #0x130]\n" + "str q17, [%x[Cpanel], #0x140]\n" "str q26, [%x[Cpanel], #0x150]\n" "str q27, [%x[Cpanel], #0x160]\n" "str q28, [%x[Cpanel], #0x170]\n" |