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_u8u32_mmla_8x12/a510.cpp | 201 +++++++++++---------- 1 file changed, 102 insertions(+), 99 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp index 83301d80bb..54c51954c8 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_u8u32_mmla_8x12/a510.cpp @@ -29,8 +29,12 @@ namespace arm_gemm { void a64_interleaved_u8u32_mmla_8x12_a510( - const uint8_t *Apanel, const uint8_t *Bpanel, - uint32_t *Cpanel, int ablocks, int bblocks, int K) { + const uint8_t *Apanel, + const uint8_t *Bpanel, + uint32_t *Cpanel, + int ablocks, + int bblocks, + int K) { struct KernelArgs { size_t K = {}; @@ -43,7 +47,6 @@ void a64_interleaved_u8u32_mmla_8x12_a510( ka.bblocks = bblocks; __asm__ __volatile__( - "1:" // Height loop "ldr x23, [%x[args_ptr], %[offsetof_bblocks]]\n" "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n" @@ -82,28 +85,28 @@ void a64_interleaved_u8u32_mmla_8x12_a510( "movi v31.4s, #0x0\n" "blt 4f\n" "3:" // main loop head - "ld1 { v3.16b }, [%x[Apanel]], #0x10\n" + "ld1 { v6.16b }, [%x[Apanel]], #0x10\n" ".inst 0x6e84a408 // ummla v8.4s, v0.16b, v4.16b\n" ".inst 0x6e85a40b // ummla v11.4s, v0.16b, v5.16b\n" - "ldp q6, q7, [x22], #0x20\n" + "ldp q3, q7, [x22], #0x20\n" ".inst 0x6e84a42e // ummla v14.4s, v1.16b, v4.16b\n" ".inst 0x6e85a431 // ummla v17.4s, v1.16b, v5.16b\n" ".inst 0x6e84a454 // ummla v20.4s, v2.16b, v4.16b\n" "sub x20, x20, #0x2\n" ".inst 0x6e85a457 // ummla v23.4s, v2.16b, v5.16b\n" - ".inst 0x6e84a47a // ummla v26.4s, v3.16b, v4.16b\n" + ".inst 0x6e84a4da // ummla v26.4s, v6.16b, v4.16b\n" "cmp x20, #0x2\n" - ".inst 0x6e85a47d // ummla v29.4s, v3.16b, v5.16b\n" + ".inst 0x6e85a4dd // ummla v29.4s, v6.16b, v5.16b\n" "ldp q4, q5, [x22], #0x20\n" - ".inst 0x6e86a409 // ummla v9.4s, v0.16b, v6.16b\n" + ".inst 0x6e83a409 // ummla v9.4s, v0.16b, v3.16b\n" ".inst 0x6e87a40c // ummla v12.4s, v0.16b, v7.16b\n" - ".inst 0x6e86a42f // ummla v15.4s, v1.16b, v6.16b\n" + ".inst 0x6e83a42f // ummla v15.4s, v1.16b, v3.16b\n" ".inst 0x6e87a432 // ummla v18.4s, v1.16b, v7.16b\n" - ".inst 0x6e86a455 // ummla v21.4s, v2.16b, v6.16b\n" + ".inst 0x6e83a455 // ummla v21.4s, v2.16b, v3.16b\n" ".inst 0x6e87a458 // ummla v24.4s, v2.16b, v7.16b\n" - ".inst 0x6e86a47b // ummla v27.4s, v3.16b, v6.16b\n" - ".inst 0x6e87a47e // ummla v30.4s, v3.16b, v7.16b\n" - "ldp q6, q7, [x22], #0x20\n" + ".inst 0x6e83a4db // ummla v27.4s, v6.16b, v3.16b\n" + ".inst 0x6e87a4de // ummla v30.4s, v6.16b, v7.16b\n" + "ldp q7, q3, [x22], #0x20\n" ".inst 0x6e84a40a // ummla v10.4s, v0.16b, v4.16b\n" ".inst 0x6e85a40d // ummla v13.4s, v0.16b, v5.16b\n" "ld1 { v0.16b }, [%x[Apanel]], #0x10\n" @@ -113,39 +116,39 @@ void a64_interleaved_u8u32_mmla_8x12_a510( ".inst 0x6e84a456 // ummla v22.4s, v2.16b, v4.16b\n" ".inst 0x6e85a459 // ummla v25.4s, v2.16b, v5.16b\n" "ld1 { v2.16b }, [%x[Apanel]], #0x10\n" - ".inst 0x6e84a47c // ummla v28.4s, v3.16b, v4.16b\n" - ".inst 0x6e85a47f // ummla v31.4s, v3.16b, v5.16b\n" - "ld1 { v3.16b }, [%x[Apanel]], #0x10\n" - ".inst 0x6e86a408 // ummla v8.4s, v0.16b, v6.16b\n" + ".inst 0x6e84a4dc // ummla v28.4s, v6.16b, v4.16b\n" + ".inst 0x6e85a4df // ummla v31.4s, v6.16b, v5.16b\n" + "ld1 { v6.16b }, [%x[Apanel]], #0x10\n" + ".inst 0x6e87a408 // ummla v8.4s, v0.16b, v7.16b\n" "ldp q4, q5, [x22], #0x20\n" - ".inst 0x6e87a40b // ummla v11.4s, v0.16b, v7.16b\n" - ".inst 0x6e86a42e // ummla v14.4s, v1.16b, v6.16b\n" - ".inst 0x6e87a431 // ummla v17.4s, v1.16b, v7.16b\n" - ".inst 0x6e86a454 // ummla v20.4s, v2.16b, v6.16b\n" - ".inst 0x6e87a457 // ummla v23.4s, v2.16b, v7.16b\n" - ".inst 0x6e86a47a // ummla v26.4s, v3.16b, v6.16b\n" - ".inst 0x6e87a47d // ummla v29.4s, v3.16b, v7.16b\n" - "ldp q6, q7, [x22], #0x20\n" + ".inst 0x6e83a40b // ummla v11.4s, v0.16b, v3.16b\n" + ".inst 0x6e87a42e // ummla v14.4s, v1.16b, v7.16b\n" + ".inst 0x6e83a431 // ummla v17.4s, v1.16b, v3.16b\n" + ".inst 0x6e87a454 // ummla v20.4s, v2.16b, v7.16b\n" + ".inst 0x6e83a457 // ummla v23.4s, v2.16b, v3.16b\n" + ".inst 0x6e87a4da // ummla v26.4s, v6.16b, v7.16b\n" + ".inst 0x6e83a4dd // ummla v29.4s, v6.16b, v3.16b\n" + "ldp q7, q3, [x22], #0x20\n" ".inst 0x6e84a409 // ummla v9.4s, v0.16b, v4.16b\n" ".inst 0x6e85a40c // ummla v12.4s, v0.16b, v5.16b\n" ".inst 0x6e84a42f // ummla v15.4s, v1.16b, v4.16b\n" ".inst 0x6e85a432 // ummla v18.4s, v1.16b, v5.16b\n" ".inst 0x6e84a455 // ummla v21.4s, v2.16b, v4.16b\n" ".inst 0x6e85a458 // ummla v24.4s, v2.16b, v5.16b\n" - ".inst 0x6e84a47b // ummla v27.4s, v3.16b, v4.16b\n" - ".inst 0x6e85a47e // ummla v30.4s, v3.16b, v5.16b\n" + ".inst 0x6e84a4db // ummla v27.4s, v6.16b, v4.16b\n" + ".inst 0x6e85a4de // ummla v30.4s, v6.16b, v5.16b\n" "ldp q4, q5, [x22], #0x20\n" - ".inst 0x6e86a40a // ummla v10.4s, v0.16b, v6.16b\n" - ".inst 0x6e87a40d // ummla v13.4s, v0.16b, v7.16b\n" + ".inst 0x6e87a40a // ummla v10.4s, v0.16b, v7.16b\n" + ".inst 0x6e83a40d // ummla v13.4s, v0.16b, v3.16b\n" "ld1 { v0.16b }, [%x[Apanel]], #0x10\n" - ".inst 0x6e86a430 // ummla v16.4s, v1.16b, v6.16b\n" - ".inst 0x6e87a433 // ummla v19.4s, v1.16b, v7.16b\n" + ".inst 0x6e87a430 // ummla v16.4s, v1.16b, v7.16b\n" + ".inst 0x6e83a433 // ummla v19.4s, v1.16b, v3.16b\n" "ld1 { v1.16b }, [%x[Apanel]], #0x10\n" - ".inst 0x6e86a456 // ummla v22.4s, v2.16b, v6.16b\n" - ".inst 0x6e87a459 // ummla v25.4s, v2.16b, v7.16b\n" + ".inst 0x6e87a456 // ummla v22.4s, v2.16b, v7.16b\n" + ".inst 0x6e83a459 // ummla v25.4s, v2.16b, v3.16b\n" "ld1 { v2.16b }, [%x[Apanel]], #0x10\n" - ".inst 0x6e86a47c // ummla v28.4s, v3.16b, v6.16b\n" - ".inst 0x6e87a47f // ummla v31.4s, v3.16b, v7.16b\n" + ".inst 0x6e87a4dc // ummla v28.4s, v6.16b, v7.16b\n" + ".inst 0x6e83a4df // ummla v31.4s, v6.16b, v3.16b\n" "bge 3b\n" "4:" // main loop skip "ld1 { v3.16b }, [%x[Apanel]], #0x10\n" @@ -158,7 +161,7 @@ void a64_interleaved_u8u32_mmla_8x12_a510( ".inst 0x6e85a457 // ummla v23.4s, v2.16b, v5.16b\n" ".inst 0x6e84a47a // ummla v26.4s, v3.16b, v4.16b\n" ".inst 0x6e85a47d // ummla v29.4s, v3.16b, v5.16b\n" - "ldp q4, q5, [x22], #0x20\n" + "ldp q5, q4, [x22], #0x20\n" ".inst 0x6e86a409 // ummla v9.4s, v0.16b, v6.16b\n" ".inst 0x6e87a40c // ummla v12.4s, v0.16b, v7.16b\n" ".inst 0x6e86a42f // ummla v15.4s, v1.16b, v6.16b\n" @@ -167,93 +170,93 @@ void a64_interleaved_u8u32_mmla_8x12_a510( ".inst 0x6e87a458 // ummla v24.4s, v2.16b, v7.16b\n" ".inst 0x6e86a47b // ummla v27.4s, v3.16b, v6.16b\n" ".inst 0x6e87a47e // ummla v30.4s, v3.16b, v7.16b\n" - ".inst 0x6e84a40a // ummla v10.4s, v0.16b, v4.16b\n" - ".inst 0x6e85a40d // ummla v13.4s, v0.16b, v5.16b\n" - ".inst 0x6e84a430 // ummla v16.4s, v1.16b, v4.16b\n" - ".inst 0x6e85a433 // ummla v19.4s, v1.16b, v5.16b\n" - ".inst 0x6e84a456 // ummla v22.4s, v2.16b, v4.16b\n" - ".inst 0x6e85a459 // ummla v25.4s, v2.16b, v5.16b\n" - ".inst 0x6e84a47c // ummla v28.4s, v3.16b, v4.16b\n" - ".inst 0x6e85a47f // ummla v31.4s, v3.16b, v5.16b\n" + ".inst 0x6e85a40a // ummla v10.4s, v0.16b, v5.16b\n" + ".inst 0x6e84a40d // ummla v13.4s, v0.16b, v4.16b\n" + ".inst 0x6e85a430 // ummla v16.4s, v1.16b, v5.16b\n" + ".inst 0x6e84a433 // ummla v19.4s, v1.16b, v4.16b\n" + ".inst 0x6e85a456 // ummla v22.4s, v2.16b, v5.16b\n" + ".inst 0x6e84a459 // ummla v25.4s, v2.16b, v4.16b\n" + ".inst 0x6e85a47c // ummla v28.4s, v3.16b, v5.16b\n" + ".inst 0x6e84a47f // ummla v31.4s, v3.16b, v4.16b\n" "cbz x20, 5f\n" - "ldp q6, q7, [x22], #0x20\n" - "ld1 { v0.16b }, [%x[Apanel]], #0x10\n" - "ld1 { v1.16b }, [%x[Apanel]], #0x10\n" - ".inst 0x6e86a408 // ummla v8.4s, v0.16b, v6.16b\n" - "ld1 { v2.16b }, [%x[Apanel]], #0x10\n" - "ld1 { v3.16b }, [%x[Apanel]], #0x10\n" - ".inst 0x6e87a40b // ummla v11.4s, v0.16b, v7.16b\n" - "ldp q4, q5, [x22], #0x20\n" - ".inst 0x6e86a42e // ummla v14.4s, v1.16b, v6.16b\n" - ".inst 0x6e87a431 // ummla v17.4s, v1.16b, v7.16b\n" - ".inst 0x6e86a454 // ummla v20.4s, v2.16b, v6.16b\n" - ".inst 0x6e87a457 // ummla v23.4s, v2.16b, v7.16b\n" - ".inst 0x6e86a47a // ummla v26.4s, v3.16b, v6.16b\n" - ".inst 0x6e87a47d // ummla v29.4s, v3.16b, v7.16b\n" - "ldp q6, q7, [x22], #0x20\n" - ".inst 0x6e84a409 // ummla v9.4s, v0.16b, v4.16b\n" - ".inst 0x6e85a40c // ummla v12.4s, v0.16b, v5.16b\n" - ".inst 0x6e84a42f // ummla v15.4s, v1.16b, v4.16b\n" - ".inst 0x6e85a432 // ummla v18.4s, v1.16b, v5.16b\n" - ".inst 0x6e84a455 // ummla v21.4s, v2.16b, v4.16b\n" - ".inst 0x6e85a458 // ummla v24.4s, v2.16b, v5.16b\n" - ".inst 0x6e84a47b // ummla v27.4s, v3.16b, v4.16b\n" - ".inst 0x6e85a47e // ummla v30.4s, v3.16b, v5.16b\n" - ".inst 0x6e86a40a // ummla v10.4s, v0.16b, v6.16b\n" - ".inst 0x6e87a40d // ummla v13.4s, v0.16b, v7.16b\n" - ".inst 0x6e86a430 // ummla v16.4s, v1.16b, v6.16b\n" - ".inst 0x6e87a433 // ummla v19.4s, v1.16b, v7.16b\n" - ".inst 0x6e86a456 // ummla v22.4s, v2.16b, v6.16b\n" - ".inst 0x6e87a459 // ummla v25.4s, v2.16b, v7.16b\n" - ".inst 0x6e86a47c // ummla v28.4s, v3.16b, v6.16b\n" - ".inst 0x6e87a47f // ummla v31.4s, v3.16b, v7.16b\n" + "ldp q1, q0, [x22], #0x20\n" + "ld1 { v7.16b }, [%x[Apanel]], #0x10\n" + "ld1 { v6.16b }, [%x[Apanel]], #0x10\n" + ".inst 0x6e81a4e8 // ummla v8.4s, v7.16b, v1.16b\n" + "ld1 { v5.16b }, [%x[Apanel]], #0x10\n" + "ld1 { v4.16b }, [%x[Apanel]], #0x10\n" + ".inst 0x6e80a4eb // ummla v11.4s, v7.16b, v0.16b\n" + "ldp q3, q2, [x22], #0x20\n" + ".inst 0x6e81a4ce // ummla v14.4s, v6.16b, v1.16b\n" + ".inst 0x6e80a4d1 // ummla v17.4s, v6.16b, v0.16b\n" + ".inst 0x6e81a4b4 // ummla v20.4s, v5.16b, v1.16b\n" + ".inst 0x6e80a4b7 // ummla v23.4s, v5.16b, v0.16b\n" + ".inst 0x6e81a49a // ummla v26.4s, v4.16b, v1.16b\n" + ".inst 0x6e80a49d // ummla v29.4s, v4.16b, v0.16b\n" + "ldp q1, q0, [x22], #0x20\n" + ".inst 0x6e83a4e9 // ummla v9.4s, v7.16b, v3.16b\n" + ".inst 0x6e82a4ec // ummla v12.4s, v7.16b, v2.16b\n" + ".inst 0x6e83a4cf // ummla v15.4s, v6.16b, v3.16b\n" + ".inst 0x6e82a4d2 // ummla v18.4s, v6.16b, v2.16b\n" + ".inst 0x6e83a4b5 // ummla v21.4s, v5.16b, v3.16b\n" + ".inst 0x6e82a4b8 // ummla v24.4s, v5.16b, v2.16b\n" + ".inst 0x6e83a49b // ummla v27.4s, v4.16b, v3.16b\n" + ".inst 0x6e82a49e // ummla v30.4s, v4.16b, v2.16b\n" + ".inst 0x6e81a4ea // ummla v10.4s, v7.16b, v1.16b\n" + ".inst 0x6e80a4ed // ummla v13.4s, v7.16b, v0.16b\n" + ".inst 0x6e81a4d0 // ummla v16.4s, v6.16b, v1.16b\n" + ".inst 0x6e80a4d3 // ummla v19.4s, v6.16b, v0.16b\n" + ".inst 0x6e81a4b6 // ummla v22.4s, v5.16b, v1.16b\n" + ".inst 0x6e80a4b9 // ummla v25.4s, v5.16b, v0.16b\n" + ".inst 0x6e81a49c // ummla v28.4s, v4.16b, v1.16b\n" + ".inst 0x6e80a49f // ummla 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