From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../a64_interleaved_s8s32_mmla_8x12/generic.cpp | 201 +++++++++++---------- 1 file changed, 102 insertions(+), 99 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp index fc20c2fc9d..a097dc358a 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_s8s32_mmla_8x12/generic.cpp @@ -29,8 +29,12 @@ namespace arm_gemm { void a64_interleaved_s8s32_mmla_8x12( - const int8_t *Apanel, const int8_t *Bpanel, - int32_t *Cpanel, int ablocks, int bblocks, int K) { + const int8_t *Apanel, + const int8_t *Bpanel, + int32_t *Cpanel, + int ablocks, + int bblocks, + int K) { struct KernelArgs { size_t K = {}; @@ -43,7 +47,6 @@ void a64_interleaved_s8s32_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_s8s32_mmla_8x12( "movi v31.4s, #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 0x4e84a408 // smmla v8.4s, v0.16b, v4.16b\n" - "ldr q7, [x22, #0x10]\n" + "ldr q3, [x22, #0x10]\n" ".inst 0x4e85a40b // smmla v11.4s, v0.16b, v5.16b\n" ".inst 0x4e84a42e // smmla v14.4s, v1.16b, v4.16b\n" ".inst 0x4e85a431 // smmla v17.4s, v1.16b, v5.16b\n" ".inst 0x4e84a454 // smmla v20.4s, v2.16b, v4.16b\n" "sub x20, x20, #0x2\n" ".inst 0x4e85a457 // smmla v23.4s, v2.16b, v5.16b\n" - ".inst 0x4e84a47a // smmla v26.4s, v3.16b, v4.16b\n" + ".inst 0x4e84a4da // smmla v26.4s, v6.16b, v4.16b\n" "ldr q4, [x22, #0x20]\n" - ".inst 0x4e85a47d // smmla v29.4s, v3.16b, v5.16b\n" + ".inst 0x4e85a4dd // smmla v29.4s, v6.16b, v5.16b\n" "ldr q5, [x22, #0x30]\n" - ".inst 0x4e86a409 // smmla v9.4s, v0.16b, v6.16b\n" - ".inst 0x4e87a40c // smmla v12.4s, v0.16b, v7.16b\n" - ".inst 0x4e86a42f // smmla v15.4s, v1.16b, v6.16b\n" + ".inst 0x4e87a409 // smmla v9.4s, v0.16b, v7.16b\n" + ".inst 0x4e83a40c // smmla v12.4s, v0.16b, v3.16b\n" + ".inst 0x4e87a42f // smmla v15.4s, v1.16b, v7.16b\n" "cmp x20, #0x2\n" - ".inst 0x4e87a432 // smmla v18.4s, v1.16b, v7.16b\n" - ".inst 0x4e86a455 // smmla v21.4s, v2.16b, v6.16b\n" - ".inst 0x4e87a458 // smmla v24.4s, v2.16b, v7.16b\n" - ".inst 0x4e86a47b // smmla v27.4s, v3.16b, v6.16b\n" - "ldr q6, [x22, #0x40]\n" - ".inst 0x4e87a47e // smmla v30.4s, v3.16b, v7.16b\n" - "ldr q7, [x22, #0x50]\n" + ".inst 0x4e83a432 // smmla v18.4s, v1.16b, v3.16b\n" + ".inst 0x4e87a455 // smmla v21.4s, v2.16b, v7.16b\n" + ".inst 0x4e83a458 // smmla v24.4s, v2.16b, v3.16b\n" + ".inst 0x4e87a4db // smmla v27.4s, v6.16b, v7.16b\n" + "ldr q7, [x22, #0x40]\n" + ".inst 0x4e83a4de // smmla v30.4s, v6.16b, v3.16b\n" + "ldr q3, [x22, #0x50]\n" ".inst 0x4e84a40a // smmla v10.4s, v0.16b, v4.16b\n" ".inst 0x4e85a40d // smmla v13.4s, v0.16b, v5.16b\n" "ldr q0, [%x[Apanel], #0x10]\n" @@ -119,42 +122,42 @@ void a64_interleaved_s8s32_mmla_8x12( ".inst 0x4e84a456 // smmla v22.4s, v2.16b, v4.16b\n" ".inst 0x4e85a459 // smmla v25.4s, v2.16b, v5.16b\n" "ldr q2, [%x[Apanel], #0x30]\n" - ".inst 0x4e84a47c // smmla v28.4s, v3.16b, v4.16b\n" + ".inst 0x4e84a4dc // smmla v28.4s, v6.16b, v4.16b\n" "ldr q4, [x22, #0x60]\n" - ".inst 0x4e85a47f // smmla v31.4s, v3.16b, v5.16b\n" - "ldr q3, [%x[Apanel], #0x40]\n" + ".inst 0x4e85a4df // smmla v31.4s, v6.16b, v5.16b\n" + "ldr q6, [%x[Apanel], #0x40]\n" "ldr q5, [x22, #0x70]\n" - ".inst 0x4e86a408 // smmla v8.4s, v0.16b, v6.16b\n" - ".inst 0x4e87a40b // smmla v11.4s, v0.16b, v7.16b\n" - ".inst 0x4e86a42e // smmla v14.4s, v1.16b, v6.16b\n" - ".inst 0x4e87a431 // smmla v17.4s, v1.16b, v7.16b\n" - ".inst 0x4e86a454 // smmla v20.4s, v2.16b, v6.16b\n" - ".inst 0x4e87a457 // smmla v23.4s, v2.16b, v7.16b\n" - ".inst 0x4e86a47a // smmla v26.4s, v3.16b, v6.16b\n" - "ldr q6, [x22, #0x80]\n" - ".inst 0x4e87a47d // smmla v29.4s, v3.16b, v7.16b\n" - "ldr q7, [x22, #0x90]\n" + ".inst 0x4e87a408 // smmla v8.4s, v0.16b, v7.16b\n" + ".inst 0x4e83a40b // smmla v11.4s, v0.16b, v3.16b\n" + ".inst 0x4e87a42e // smmla v14.4s, v1.16b, v7.16b\n" + ".inst 0x4e83a431 // smmla v17.4s, v1.16b, v3.16b\n" + ".inst 0x4e87a454 // smmla v20.4s, v2.16b, v7.16b\n" + ".inst 0x4e83a457 // smmla v23.4s, v2.16b, v3.16b\n" + ".inst 0x4e87a4da // smmla v26.4s, v6.16b, v7.16b\n" + "ldr q7, [x22, #0x80]\n" + ".inst 0x4e83a4dd // smmla v29.4s, v6.16b, v3.16b\n" + "ldr q3, [x22, #0x90]\n" ".inst 0x4e84a409 // smmla v9.4s, v0.16b, v4.16b\n" ".inst 0x4e85a40c // smmla v12.4s, v0.16b, v5.16b\n" ".inst 0x4e84a42f // smmla v15.4s, v1.16b, v4.16b\n" ".inst 0x4e85a432 // smmla v18.4s, v1.16b, v5.16b\n" ".inst 0x4e84a455 // smmla v21.4s, v2.16b, v4.16b\n" ".inst 0x4e85a458 // smmla v24.4s, v2.16b, v5.16b\n" - ".inst 0x4e84a47b // smmla v27.4s, v3.16b, v4.16b\n" + ".inst 0x4e84a4db // smmla v27.4s, v6.16b, v4.16b\n" "ldr q4, [x22, #0xa0]\n" - ".inst 0x4e85a47e // smmla v30.4s, v3.16b, v5.16b\n" + ".inst 0x4e85a4de // smmla v30.4s, v6.16b, v5.16b\n" "ldr q5, [x22, #0xb0]\n" - ".inst 0x4e86a40a // smmla v10.4s, v0.16b, v6.16b\n" - ".inst 0x4e87a40d // smmla v13.4s, v0.16b, v7.16b\n" + ".inst 0x4e87a40a // smmla v10.4s, v0.16b, v7.16b\n" + ".inst 0x4e83a40d // smmla v13.4s, v0.16b, v3.16b\n" "ldr q0, [%x[Apanel], #0x50]\n" - ".inst 0x4e86a430 // smmla v16.4s, v1.16b, v6.16b\n" - ".inst 0x4e87a433 // smmla v19.4s, v1.16b, v7.16b\n" + ".inst 0x4e87a430 // smmla v16.4s, v1.16b, v7.16b\n" + ".inst 0x4e83a433 // smmla v19.4s, v1.16b, v3.16b\n" "ldr q1, [%x[Apanel], #0x60]\n" - ".inst 0x4e86a456 // smmla v22.4s, v2.16b, v6.16b\n" - ".inst 0x4e87a459 // smmla v25.4s, v2.16b, v7.16b\n" + ".inst 0x4e87a456 // smmla v22.4s, v2.16b, v7.16b\n" + ".inst 0x4e83a459 // smmla v25.4s, v2.16b, v3.16b\n" "ldr q2, [%x[Apanel], #0x70]\n" - ".inst 0x4e86a47c // smmla v28.4s, v3.16b, v6.16b\n" - ".inst 0x4e87a47f // smmla v31.4s, v3.16b, v7.16b\n" + ".inst 0x4e87a4dc // smmla v28.4s, v6.16b, v7.16b\n" + ".inst 0x4e83a4df // smmla v31.4s, v6.16b, v3.16b\n" "add %x[Apanel], %x[Apanel], #0x80\n" "add x22, x22, #0xc0\n" "bge 3b\n" @@ -191,89 +194,89 @@ void a64_interleaved_s8s32_mmla_8x12( ".inst 0x4e84a47c // smmla v28.4s, v3.16b, v4.16b\n" ".inst 0x4e85a47f // smmla v31.4s, v3.16b, v5.16b\n" "cbz x20, 5f\n" - "ldr q6, [x22, #0x0]\n" - "ldr q0, [%x[Apanel], #0x0]\n" - ".inst 0x4e86a408 // smmla v8.4s, v0.16b, v6.16b\n" - "ldr q1, [%x[Apanel], #0x10]\n" - "ldr q7, [x22, #0x10]\n" - ".inst 0x4e87a40b // smmla v11.4s, v0.16b, v7.16b\n" - "ldr q2, [%x[Apanel], #0x20]\n" - "ldr q3, [%x[Apanel], #0x30]\n" - ".inst 0x4e86a42e // smmla v14.4s, v1.16b, v6.16b\n" - "ldr q4, [x22, #0x20]\n" - "ldr q5, [x22, #0x30]\n" - ".inst 0x4e87a431 // smmla v17.4s, v1.16b, v7.16b\n" - ".inst 0x4e86a454 // smmla v20.4s, v2.16b, v6.16b\n" - ".inst 0x4e87a457 // smmla v23.4s, v2.16b, v7.16b\n" + "ldr q1, [x22, #0x0]\n" + "ldr q7, [%x[Apanel], #0x0]\n" + ".inst 0x4e81a4e8 // smmla v8.4s, v7.16b, v1.16b\n" + "ldr q6, [%x[Apanel], #0x10]\n" + "ldr q0, [x22, #0x10]\n" + ".inst 0x4e80a4eb // smmla v11.4s, v7.16b, v0.16b\n" + "ldr q5, [%x[Apanel], #0x20]\n" + "ldr q4, [%x[Apanel], #0x30]\n" + ".inst 0x4e81a4ce // smmla v14.4s, v6.16b, v1.16b\n" + "ldr q3, [x22, #0x20]\n" + "ldr q2, [x22, #0x30]\n" + ".inst 0x4e80a4d1 // smmla v17.4s, v6.16b, v0.16b\n" + ".inst 0x4e81a4b4 // smmla v20.4s, v5.16b, v1.16b\n" + ".inst 0x4e80a4b7 // smmla v23.4s, v5.16b, v0.16b\n" "add %x[Apanel], %x[Apanel], #0x40\n" - ".inst 0x4e86a47a // smmla v26.4s, v3.16b, v6.16b\n" - "ldr q6, [x22, #0x40]\n" - ".inst 0x4e87a47d // smmla v29.4s, v3.16b, v7.16b\n" - "ldr q7, [x22, #0x50]\n" - ".inst 0x4e84a409 // smmla v9.4s, v0.16b, v4.16b\n" - ".inst 0x4e85a40c // smmla v12.4s, v0.16b, v5.16b\n" - ".inst 0x4e84a42f // smmla v15.4s, v1.16b, v4.16b\n" - ".inst 0x4e85a432 // smmla v18.4s, v1.16b, v5.16b\n" + ".inst 0x4e81a49a // smmla v26.4s, v4.16b, v1.16b\n" + "ldr q1, [x22, #0x40]\n" + ".inst 0x4e80a49d // smmla v29.4s, v4.16b, v0.16b\n" + "ldr q0, [x22, #0x50]\n" + ".inst 0x4e83a4e9 // smmla v9.4s, v7.16b, v3.16b\n" + ".inst 0x4e82a4ec // smmla v12.4s, v7.16b, v2.16b\n" + ".inst 0x4e83a4cf // smmla v15.4s, v6.16b, v3.16b\n" + ".inst 0x4e82a4d2 // smmla v18.4s, v6.16b, v2.16b\n" "add x22, x22, #0x60\n" - ".inst 0x4e84a455 // smmla v21.4s, v2.16b, v4.16b\n" - ".inst 0x4e85a458 // smmla v24.4s, v2.16b, v5.16b\n" - ".inst 0x4e84a47b // smmla v27.4s, v3.16b, v4.16b\n" - ".inst 0x4e85a47e // smmla v30.4s, v3.16b, v5.16b\n" - ".inst 0x4e86a40a // smmla v10.4s, v0.16b, v6.16b\n" - ".inst 0x4e87a40d // smmla v13.4s, v0.16b, v7.16b\n" - ".inst 0x4e86a430 // smmla v16.4s, v1.16b, v6.16b\n" - ".inst 0x4e87a433 // smmla v19.4s, v1.16b, v7.16b\n" - ".inst 0x4e86a456 // smmla v22.4s, v2.16b, v6.16b\n" - ".inst 0x4e87a459 // smmla v25.4s, v2.16b, v7.16b\n" - ".inst 0x4e86a47c // smmla v28.4s, v3.16b, v6.16b\n" - ".inst 0x4e87a47f // smmla v31.4s, v3.16b, v7.16b\n" + ".inst 0x4e83a4b5 // smmla v21.4s, v5.16b, v3.16b\n" + ".inst 0x4e82a4b8 // smmla v24.4s, v5.16b, v2.16b\n" + ".inst 0x4e83a49b // smmla v27.4s, v4.16b, v3.16b\n" + ".inst 0x4e82a49e // smmla v30.4s, v4.16b, v2.16b\n" + ".inst 0x4e81a4ea // smmla v10.4s, v7.16b, v1.16b\n" + ".inst 0x4e80a4ed // smmla v13.4s, v7.16b, v0.16b\n" + ".inst 0x4e81a4d0 // smmla v16.4s, v6.16b, v1.16b\n" + ".inst 0x4e80a4d3 // smmla v19.4s, v6.16b, v0.16b\n" + ".inst 0x4e81a4b6 // smmla v22.4s, v5.16b, v1.16b\n" + ".inst 0x4e80a4b9 // smmla v25.4s, v5.16b, v0.16b\n" + ".inst 0x4e81a49c // smmla v28.4s, v4.16b, v1.16b\n" + ".inst 0x4e80a49f // smmla v31.4s, v4.16b, v0.16b\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" -- cgit v1.2.1