diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp | 602 |
1 files changed, 300 insertions, 302 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp index a0ea96822a..0e468b196a 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_4x24/generic.cpp @@ -93,7 +93,6 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( break; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x4\n" "bge 130f\n" @@ -255,11 +254,11 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "20:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 21f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" "cbnz x26, 22f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -279,31 +278,31 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "23:" // Height 1: Multiply loop: Main loop head ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x40]\n" + "ldr q24, [x28, #0x40]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x50]\n" + "ldr q23, [x28, #0x50]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x28, #0x60]\n" + "ldr q22, [x28, #0x60]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x80]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x90]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q6, [x28, #0xa0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q7, [x28, #0xb0]\n" + "ldr q21, [x28, #0x70]\n" + ".inst 0x6e58ec0a // bfmmla v10.4s, v0.8h, v24.8h\n" + "ldr q24, [x28, #0x80]\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + "ldr q23, [x28, #0x90]\n" + ".inst 0x6e56ec0b // bfmmla v11.4s, v0.8h, v22.8h\n" + "ldr q22, [x28, #0xa0]\n" + ".inst 0x6e55ec11 // bfmmla v17.4s, v0.8h, v21.8h\n" + "ldr q21, [x28, #0xb0]\n" "sub x25, x25, #0x4\n" "cmp x25, #0x8\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" + ".inst 0x6e58ec0c // bfmmla v12.4s, v0.8h, v24.8h\n" "add x28, x28, #0xc0\n" "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" "ldr q5, [x28, #0x10]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" "ldr q6, [x28, #0x20]\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "ldr q7, [x28, #0x30]\n" "prfm pldl1keep, [x24, #0x80]\n" "ld1 { v0.4s }, [x24], #0x10\n" @@ -311,28 +310,28 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "24:" // Height 1: Multiply loop: Single iteration only ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x40]\n" + "ldr q23, [x28, #0x40]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x50]\n" + "ldr q25, [x28, #0x50]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x28, #0x60]\n" + "ldr q21, [x28, #0x60]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x80]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x90]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q6, [x28, #0xa0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q7, [x28, #0xb0]\n" + "ldr q24, [x28, #0x70]\n" + ".inst 0x6e57ec0a // bfmmla v10.4s, v0.8h, v23.8h\n" + "ldr q23, [x28, #0x80]\n" + ".inst 0x6e59ec10 // bfmmla v16.4s, v0.8h, v25.8h\n" + "ldr q22, [x28, #0x90]\n" + ".inst 0x6e55ec0b // bfmmla v11.4s, v0.8h, v21.8h\n" + "ldr q21, [x28, #0xa0]\n" + ".inst 0x6e58ec11 // bfmmla v17.4s, v0.8h, v24.8h\n" + "ldr q5, [x28, #0xb0]\n" "sub x25, x25, #0x4\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" + ".inst 0x6e57ec0c // bfmmla v12.4s, v0.8h, v23.8h\n" + ".inst 0x6e56ec12 // bfmmla v18.4s, v0.8h, v22.8h\n" "prfm pldl1keep, [x24, #0x80]\n" "add x28, x28, #0xc0\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e55ec0d // bfmmla v13.4s, v0.8h, v21.8h\n" + ".inst 0x6e45ec13 // bfmmla v19.4s, v0.8h, v5.8h\n" "25:" // Height 1: Multiply loop: Main loop skip "cbz x25, 28f\n" "cbz x25, 28f\n" @@ -344,31 +343,31 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "26:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 "ldr s0, [x24, #0x0]\n" "27:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q4, [x28, #0x0]\n" - "ldr q5, [x28, #0x10]\n" + "ldr q21, [x28, #0x0]\n" + "ldr q1, [x28, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" - ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q6, [x28, #0x20]\n" - "ldr q7, [x28, #0x30]\n" - ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q4, [x28, #0x40]\n" - "ldr q5, [x28, #0x50]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q6, [x28, #0x60]\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q4, [x28, #0x80]\n" - "ldr q5, [x28, #0x90]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - "ldr q6, [x28, #0xa0]\n" - "ldr q7, [x28, #0xb0]\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e55ec08 // bfmmla v8.4s, v0.8h, v21.8h\n" + "ldr q21, [x28, #0x20]\n" + "ldr q22, [x28, #0x30]\n" + ".inst 0x6e41ec0e // bfmmla v14.4s, v0.8h, v1.8h\n" + ".inst 0x6e55ec09 // bfmmla v9.4s, v0.8h, v21.8h\n" + "ldr q21, [x28, #0x40]\n" + "ldr q23, [x28, #0x50]\n" + ".inst 0x6e56ec0f // bfmmla v15.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec0a // bfmmla v10.4s, v0.8h, v21.8h\n" + "ldr q21, [x28, #0x60]\n" + "ldr q22, [x28, #0x70]\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + ".inst 0x6e55ec0b // bfmmla v11.4s, v0.8h, v21.8h\n" + "ldr q21, [x28, #0x80]\n" + "ldr q23, [x28, #0x90]\n" + ".inst 0x6e56ec11 // bfmmla v17.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec0c // bfmmla v12.4s, v0.8h, v21.8h\n" + "ldr q22, [x28, #0xa0]\n" + "ldr q21, [x28, #0xb0]\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "add x28, x28, #0xc0\n" "28:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -384,21 +383,21 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "uzp1 v13.2d, v13.2d, v19.2d\n" "tbz %x[flags], #1, 29f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v22.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" + "ld1r { v21.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v22.4s\n" + "fmin v9.4s, v9.4s, v22.4s\n" + "fmin v10.4s, v10.4s, v22.4s\n" + "fmin v11.4s, v11.4s, v22.4s\n" + "fmin v12.4s, v12.4s, v22.4s\n" + "fmin v13.4s, v13.4s, v22.4s\n" + "fmax v8.4s, v8.4s, v21.4s\n" + "fmax v9.4s, v9.4s, v21.4s\n" + "fmax v10.4s, v10.4s, v21.4s\n" + "fmax v11.4s, v11.4s, v21.4s\n" + "fmax v12.4s, v12.4s, v21.4s\n" + "fmax v13.4s, v13.4s, v21.4s\n" "29:" // Height 1: No activation "cmp x9, #0x18\n" "bge 42f\n" @@ -678,12 +677,12 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "63:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 64f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" "cbnz x26, 65f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -691,7 +690,7 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "b 65f\n" "64:" // Height 2: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "65:" // Height 2: input setup done "cmp x25, #0x4\n" "blt 68f\n" @@ -707,31 +706,31 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x40]\n" + "ldr q3, [x28, #0x40]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x50]\n" + "ldr q23, [x28, #0x50]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x28, #0x60]\n" + "ldr q22, [x28, #0x60]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x80]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x90]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q6, [x28, #0xa0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q7, [x28, #0xb0]\n" + "ldr q21, [x28, #0x70]\n" + ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n" + "ldr q1, [x28, #0x80]\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + "ldr q23, [x28, #0x90]\n" + ".inst 0x6e56ec0b // bfmmla v11.4s, v0.8h, v22.8h\n" + "ldr q22, [x28, #0xa0]\n" + ".inst 0x6e55ec11 // bfmmla v17.4s, v0.8h, v21.8h\n" + "ldr q21, [x28, #0xb0]\n" "sub x25, x25, #0x4\n" "cmp x25, #0x8\n" "add x28, x28, #0xc0\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" + ".inst 0x6e41ec0c // bfmmla v12.4s, v0.8h, v1.8h\n" "ldr q4, [x28, #0x0]\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" "ldr q5, [x28, #0x10]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" "ldr q6, [x28, #0x20]\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "ldr q7, [x28, #0x30]\n" "prfm pldl1keep, [x24, #0x80]\n" "ld1 { v0.4s }, [x24], #0x10\n" @@ -742,28 +741,28 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x40]\n" + "ldr q24, [x28, #0x40]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x50]\n" + "ldr q23, [x28, #0x50]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x28, #0x60]\n" + "ldr q22, [x28, #0x60]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - "ldr q4, [x28, #0x80]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q5, [x28, #0x90]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - "ldr q6, [x28, #0xa0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q7, [x28, #0xb0]\n" + "ldr q21, [x28, #0x70]\n" + ".inst 0x6e58ec0a // bfmmla v10.4s, v0.8h, v24.8h\n" + "ldr q24, [x28, #0x80]\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + "ldr q23, [x28, #0x90]\n" + ".inst 0x6e56ec0b // bfmmla v11.4s, v0.8h, v22.8h\n" + "ldr q22, [x28, #0xa0]\n" + ".inst 0x6e55ec11 // bfmmla v17.4s, v0.8h, v21.8h\n" + "ldr q21, [x28, #0xb0]\n" "sub x25, x25, #0x4\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" + ".inst 0x6e58ec0c // bfmmla v12.4s, v0.8h, v24.8h\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" "prfm pldl1keep, [x24, #0x80]\n" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "add x28, x28, #0xc0\n" "68:" // Height 2: Multiply loop: Main loop skip "cbz x25, 71f\n" @@ -779,32 +778,32 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "ldr s0, [x24, #0x0]\n" "ldr s1, [x23, #0x0]\n" "70:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q4, [x28, #0x0]\n" - "ldr q5, [x28, #0x10]\n" + "ldr q24, [x28, #0x0]\n" + "ldr q23, [x28, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" - "ldr q6, [x28, #0x20]\n" - "ldr q7, [x28, #0x30]\n" - ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - "ldr q4, [x28, #0x40]\n" - "ldr q5, [x28, #0x50]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - "ldr q6, [x28, #0x60]\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - "ldr q4, [x28, #0x80]\n" - "ldr q5, [x28, #0x90]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - "ldr q6, [x28, #0xa0]\n" - "ldr q7, [x28, #0xb0]\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + "ldr q22, [x28, #0x20]\n" + "ldr q21, [x28, #0x30]\n" + ".inst 0x6e58ec08 // bfmmla v8.4s, v0.8h, v24.8h\n" + ".inst 0x6e57ec0e // bfmmla v14.4s, v0.8h, v23.8h\n" + "ldr q24, [x28, #0x40]\n" + "ldr q23, [x28, #0x50]\n" + ".inst 0x6e56ec09 // bfmmla v9.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec0f // bfmmla v15.4s, v0.8h, v21.8h\n" + "ldr q22, [x28, #0x60]\n" + "ldr q21, [x28, #0x70]\n" + ".inst 0x6e58ec0a // bfmmla v10.4s, v0.8h, v24.8h\n" + ".inst 0x6e57ec10 // bfmmla v16.4s, v0.8h, v23.8h\n" + "ldr q24, [x28, #0x80]\n" + "ldr q23, [x28, #0x90]\n" + ".inst 0x6e56ec0b // bfmmla v11.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec11 // bfmmla v17.4s, v0.8h, v21.8h\n" + "ldr q22, [x28, #0xa0]\n" + "ldr q21, [x28, #0xb0]\n" + ".inst 0x6e58ec0c // bfmmla v12.4s, v0.8h, v24.8h\n" + ".inst 0x6e57ec12 // bfmmla v18.4s, v0.8h, v23.8h\n" + ".inst 0x6e56ec0d // bfmmla v13.4s, v0.8h, v22.8h\n" + ".inst 0x6e55ec13 // bfmmla v19.4s, v0.8h, v21.8h\n" "add x28, x28, #0xc0\n" "71:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -829,33 +828,33 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "uzp2 v13.2d, v13.2d, v19.2d\n" "tbz %x[flags], #1, 72f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v22.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v4.4s, v4.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmax v4.4s, v4.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" + "ld1r { v21.4s }, [x20]\n" + "fmin v4.4s, v4.4s, v22.4s\n" + "fmin v14.4s, v14.4s, v22.4s\n" + "fmin v15.4s, v15.4s, v22.4s\n" + "fmin v16.4s, v16.4s, v22.4s\n" + "fmin v17.4s, v17.4s, v22.4s\n" + "fmin v18.4s, v18.4s, v22.4s\n" + "fmin v8.4s, v8.4s, v22.4s\n" + "fmin v9.4s, v9.4s, v22.4s\n" + "fmin v10.4s, v10.4s, v22.4s\n" + "fmin v11.4s, v11.4s, v22.4s\n" + "fmin v12.4s, v12.4s, v22.4s\n" + "fmin v13.4s, v13.4s, v22.4s\n" + "fmax v4.4s, v4.4s, v21.4s\n" + "fmax v14.4s, v14.4s, v21.4s\n" + "fmax v15.4s, v15.4s, v21.4s\n" + "fmax v16.4s, v16.4s, v21.4s\n" + "fmax v17.4s, v17.4s, v21.4s\n" + "fmax v18.4s, v18.4s, v21.4s\n" + "fmax v8.4s, v8.4s, v21.4s\n" + "fmax v9.4s, v9.4s, v21.4s\n" + "fmax v10.4s, v10.4s, v21.4s\n" + "fmax v11.4s, v11.4s, v21.4s\n" + "fmax v12.4s, v12.4s, v21.4s\n" + "fmax v13.4s, v13.4s, v21.4s\n" "72:" // Height 2: No activation "cmp x9, #0x18\n" "bge 85f\n" @@ -1238,13 +1237,13 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "106:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 107f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" - "ldr x22, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" + "ldr x22, [x20, #0x10]\n" "cbnz x26, 108f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -1253,8 +1252,8 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "b 108f\n" "107:" // Height 3: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "108:" // Height 3: input setup done "cmp x25, #0x4\n" "blt 111f\n" @@ -1285,7 +1284,7 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0x70]\n" + "ldr q3, [x28, #0x70]\n" ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" "prfm pldl1keep, [x23, #0x80]\n" "ld1 { v1.4s }, [x23], #0x10\n" @@ -1298,9 +1297,9 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" "ldr q6, [x28, #0xa0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0xb0]\n" + ".inst 0x6e43ec11 // bfmmla v17.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec5d // bfmmla v29.4s, v2.8h, v3.8h\n" + "ldr q3, [x28, #0xb0]\n" "add x28, x28, #0xc0\n" ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" @@ -1311,9 +1310,9 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" "ldr q6, [x28, #0x20]\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" + ".inst 0x6e43ec13 // bfmmla v19.4s, v0.8h, v3.8h\n" "ld1 { v0.4s }, [x24], #0x10\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e43ec5f // bfmmla v31.4s, v2.8h, v3.8h\n" "ld1 { v2.4s }, [x22], #0x10\n" "ldr q7, [x28, #0x30]\n" "bge 109b\n" @@ -1324,10 +1323,10 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "sub x25, x25, #0x4\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x40]\n" + "ldr q3, [x28, #0x40]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x50]\n" + "ldr q4, [x28, #0x50]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" @@ -1335,29 +1334,29 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" + "ldr q1, [x28, #0x70]\n" + ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x80]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x90]\n" + ".inst 0x6e43ec56 // bfmmla v22.4s, v2.8h, v3.8h\n" + "ldr q5, [x28, #0x80]\n" + ".inst 0x6e44ec10 // bfmmla v16.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5c // bfmmla v28.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x90]\n" ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x28, #0xa0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0xb0]\n" + "ldr q3, [x28, #0xa0]\n" + ".inst 0x6e41ec11 // bfmmla v17.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x28, #0xb0]\n" "add x28, x28, #0xc0\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" + ".inst 0x6e44ec12 // bfmmla v18.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5e // bfmmla v30.4s, v2.8h, v4.8h\n" + ".inst 0x6e43ec0d // bfmmla v13.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\n" + ".inst 0x6e41ec13 // bfmmla v19.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5f // bfmmla v31.4s, v2.8h, v1.8h\n" "111:" // Height 3: Multiply loop: Main loop skip "cbz x25, 114f\n" "cbz x25, 114f\n" @@ -1375,46 +1374,46 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "ldr s1, [x23, #0x0]\n" "ldr s2, [x22, #0x0]\n" "113:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q4, [x28, #0x0]\n" - "ldr q5, [x28, #0x10]\n" + "ldr q5, [x28, #0x0]\n" + "ldr q4, [x28, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" - "ldr q6, [x28, #0x20]\n" - "ldr q7, [x28, #0x30]\n" + "ldr q3, [x28, #0x20]\n" + "ldr q1, [x28, #0x30]\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" - ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x40]\n" - ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x50]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x28, #0x60]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x80]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x90]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x28, #0xa0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0xb0]\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" + ".inst 0x6e45ec08 // bfmmla v8.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec54 // bfmmla v20.4s, v2.8h, v5.8h\n" + "ldr q5, [x28, #0x40]\n" + ".inst 0x6e44ec0e // bfmmla v14.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5a // bfmmla v26.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x50]\n" + ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec55 // bfmmla v21.4s, v2.8h, v3.8h\n" + "ldr q3, [x28, #0x60]\n" + ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5b // bfmmla v27.4s, v2.8h, v1.8h\n" + "ldr q1, [x28, #0x70]\n" + ".inst 0x6e45ec0a // bfmmla v10.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec56 // bfmmla v22.4s, v2.8h, v5.8h\n" + "ldr q5, [x28, #0x80]\n" + ".inst 0x6e44ec10 // bfmmla v16.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5c // bfmmla v28.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x90]\n" + ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec57 // bfmmla v23.4s, v2.8h, v3.8h\n" + "ldr q3, [x28, #0xa0]\n" + ".inst 0x6e41ec11 // bfmmla v17.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x28, #0xb0]\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" "add x28, x28, #0xc0\n" - ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" + ".inst 0x6e44ec12 // bfmmla v18.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5e // bfmmla v30.4s, v2.8h, v4.8h\n" + ".inst 0x6e43ec0d // bfmmla v13.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\n" + ".inst 0x6e41ec13 // bfmmla v19.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5f // bfmmla v31.4s, v2.8h, v1.8h\n" "114:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" @@ -1937,14 +1936,14 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "149:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w25, [x20, x26, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 150f\n" - "ldr x21, [%x[input_ptr], x26, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x24, [x21, #0x0]\n" - "ldr x23, [x21, #0x8]\n" - "ldr x22, [x21, #0x10]\n" - "ldr x21, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x24, [x20, #0x0]\n" + "ldr x23, [x20, #0x8]\n" + "ldr x22, [x20, #0x10]\n" + "ldr x21, [x20, #0x18]\n" "cbnz x26, 151f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x24, x24, x20, LSL #2\n" @@ -1954,9 +1953,9 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "b 151f\n" "150:" // Height 4: setup direct input "mov x24, %x[input_ptr]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "151:" // Height 4: input setup done "cmp x25, #0x4\n" "blt 154f\n" @@ -2033,39 +2032,39 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x40]\n" + "ldr q3, [x28, #0x40]\n" ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x50]\n" + "ldr q4, [x28, #0x50]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" "prfm pldl1keep, [x21, #0x80]\n" ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" "ldr q6, [x28, #0x60]\n" ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x80]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x90]\n" + "ldr q1, [x28, #0x70]\n" + ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec56 // bfmmla v22.4s, v2.8h, v3.8h\n" + "ldr q5, [x28, #0x80]\n" + ".inst 0x6e44ec10 // bfmmla v16.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5c // bfmmla v28.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x90]\n" ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x28, #0xa0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0xb0]\n" + "ldr q3, [x28, #0xa0]\n" + ".inst 0x6e41ec11 // bfmmla v17.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x28, #0xb0]\n" "add x28, x28, #0xc0\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" + ".inst 0x6e44ec12 // bfmmla v18.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5e // bfmmla v30.4s, v2.8h, v4.8h\n" + ".inst 0x6e43ec0d // bfmmla v13.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\n" + ".inst 0x6e41ec13 // bfmmla v19.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5f // bfmmla v31.4s, v2.8h, v1.8h\n" "154:" // Height 4: Multiply loop: Main loop skip "cbz x25, 157f\n" "cbz x25, 157f\n" @@ -2086,47 +2085,47 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "ldr s2, [x22, #0x0]\n" "ldr s3, [x21, #0x0]\n" "156:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q4, [x28, #0x0]\n" - "ldr q5, [x28, #0x10]\n" + "ldr q5, [x28, #0x0]\n" + "ldr q4, [x28, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" - "ldr q6, [x28, #0x20]\n" - "ldr q7, [x28, #0x30]\n" + "ldr q7, [x28, #0x20]\n" + "ldr q6, [x28, #0x30]\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n" - ".inst 0x6e44ec08 // bfmmla v8.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec54 // bfmmla v20.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x40]\n" - ".inst 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5a // bfmmla v26.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x50]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x28, #0x60]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5b // bfmmla v27.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0x70]\n" - ".inst 0x6e44ec0a // bfmmla v10.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec56 // bfmmla v22.4s, v2.8h, v4.8h\n" - "ldr q4, [x28, #0x80]\n" - ".inst 0x6e45ec10 // bfmmla v16.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5c // bfmmla v28.4s, v2.8h, v5.8h\n" - "ldr q5, [x28, #0x90]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x28, #0xa0]\n" - ".inst 0x6e47ec11 // bfmmla v17.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5d // bfmmla v29.4s, v2.8h, v7.8h\n" - "ldr q7, [x28, #0xb0]\n" + ".inst 0x6e45ec08 // bfmmla v8.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec54 // bfmmla v20.4s, v2.8h, v5.8h\n" + "ldr q5, [x28, #0x40]\n" + ".inst 0x6e44ec0e // bfmmla v14.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5a // bfmmla v26.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x50]\n" + ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" + "ldr q3, [x28, #0x60]\n" + ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + ".inst 0x6e46ec5b // bfmmla v27.4s, v2.8h, v6.8h\n" + "ldr q1, [x28, #0x70]\n" + ".inst 0x6e45ec0a // bfmmla v10.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec56 // bfmmla v22.4s, v2.8h, v5.8h\n" + "ldr q5, [x28, #0x80]\n" + ".inst 0x6e44ec10 // bfmmla v16.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5c // bfmmla v28.4s, v2.8h, v4.8h\n" + "ldr q4, [x28, #0x90]\n" + ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec57 // bfmmla v23.4s, v2.8h, v3.8h\n" + "ldr q3, [x28, #0xa0]\n" + ".inst 0x6e41ec11 // bfmmla v17.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x28, #0xb0]\n" "add x28, x28, #0xc0\n" - ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" - ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" - ".inst 0x6e45ec12 // bfmmla v18.4s, v0.8h, v5.8h\n" - ".inst 0x6e45ec5e // bfmmla v30.4s, v2.8h, v5.8h\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec13 // bfmmla v19.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec5f // bfmmla v31.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec58 // bfmmla v24.4s, v2.8h, v5.8h\n" + ".inst 0x6e44ec12 // bfmmla v18.4s, v0.8h, v4.8h\n" + ".inst 0x6e44ec5e // bfmmla v30.4s, v2.8h, v4.8h\n" + ".inst 0x6e43ec0d // bfmmla v13.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec59 // bfmmla v25.4s, v2.8h, v3.8h\n" + ".inst 0x6e41ec13 // bfmmla v19.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec5f // bfmmla v31.4s, v2.8h, v1.8h\n" "157:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" @@ -2415,7 +2414,6 @@ void a64_hybrid_fp32bf16fp32_mmla_4x24 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "174:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "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", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" |