diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp | 1880 |
1 files changed, 939 insertions, 941 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp index 5a000c69af..f6389e27d1 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_mmla_6x16/generic.cpp @@ -93,7 +93,6 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( break; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x6\n" "bge 186f\n" @@ -211,11 +210,11 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "16:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 17f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" "cbnz x28, 18f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -231,41 +230,41 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr q6, [x10, #0x10]\n" "blt 20f\n" "19:" // Height 1: Multiply loop: Main loop head - "trn1 v0.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + "trn1 v20.2d, v1.2d, v21.2d\n" + ".inst 0x6e47ee88 // bfmmla v8.4s, v20.8h, v7.8h\n" + "ldr q17, [x10, #0x20]\n" + ".inst 0x6e46ee8c // bfmmla v12.4s, v20.8h, v6.8h\n" + "ldr q19, [x10, #0x30]\n" + ".inst 0x6e51ee89 // bfmmla v9.4s, v20.8h, v17.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e53ee8d // bfmmla v13.4s, v20.8h, v19.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ee8a // bfmmla v10.4s, v20.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ee8e // bfmmla v14.4s, v20.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" + "trn2 v1.2d, v1.2d, v21.2d\n" + ".inst 0x6e52ee8b // bfmmla v11.4s, v20.8h, v18.8h\n" + "ldr q18, [x10, #0x80]\n" + ".inst 0x6e51ee8f // bfmmla v15.4s, v20.8h, v17.8h\n" + "ldr q17, [x10, #0x90]\n" + ".inst 0x6e52ec28 // bfmmla v8.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xa0]\n" + ".inst 0x6e51ec2c // bfmmla v12.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xb0]\n" + ".inst 0x6e52ec29 // bfmmla v9.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xc0]\n" + ".inst 0x6e51ec2d // bfmmla v13.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xd0]\n" + ".inst 0x6e52ec2a // bfmmla v10.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xe0]\n" + ".inst 0x6e51ec2e // bfmmla v14.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xf0]\n" "sub x27, x27, #0x8\n" "add x26, x26, #0x10\n" "cmp x27, #0x10\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + ".inst 0x6e52ec2b // bfmmla v11.4s, v1.8h, v18.8h\n" + ".inst 0x6e51ec2f // bfmmla v15.4s, v1.8h, v17.8h\n" "ldr q1, [x26, #0x0]\n" "add x10, x10, #0x100\n" "ldr q7, [x10, #0x0]\n" @@ -273,40 +272,40 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "prfm pldl1keep, [x26, #0x80]\n" "bge 19b\n" "20:" // Height 1: Multiply loop: Single iteration only - "trn1 v0.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + "trn1 v19.2d, v1.2d, v20.2d\n" + ".inst 0x6e47ee68 // bfmmla v8.4s, v19.8h, v7.8h\n" + "ldr q17, [x10, #0x20]\n" + ".inst 0x6e46ee6c // bfmmla v12.4s, v19.8h, v6.8h\n" + "ldr q18, [x10, #0x30]\n" + ".inst 0x6e51ee69 // bfmmla v9.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x40]\n" + ".inst 0x6e52ee6d // bfmmla v13.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x50]\n" + ".inst 0x6e51ee6a // bfmmla v10.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x60]\n" + ".inst 0x6e52ee6e // bfmmla v14.4s, v19.8h, v18.8h\n" + "ldr q24, [x10, #0x70]\n" + "trn2 v1.2d, v1.2d, v20.2d\n" + ".inst 0x6e51ee6b // bfmmla v11.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x80]\n" + ".inst 0x6e58ee6f // bfmmla v15.4s, v19.8h, v24.8h\n" + "ldr q2, [x10, #0x90]\n" + ".inst 0x6e51ec28 // bfmmla v8.4s, v1.8h, v17.8h\n" + "ldr q18, [x10, #0xa0]\n" + ".inst 0x6e42ec2c // bfmmla v12.4s, v1.8h, v2.8h\n" + "ldr q17, [x10, #0xb0]\n" + ".inst 0x6e52ec29 // bfmmla v9.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xc0]\n" + ".inst 0x6e51ec2d // bfmmla v13.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xd0]\n" + ".inst 0x6e52ec2a // bfmmla v10.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xe0]\n" + ".inst 0x6e51ec2e // bfmmla v14.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xf0]\n" "add x26, x26, #0x10\n" "sub x27, x27, #0x8\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + ".inst 0x6e52ec2b // bfmmla v11.4s, v1.8h, v18.8h\n" + ".inst 0x6e51ec2f // bfmmla v15.4s, v1.8h, v17.8h\n" "prfm pldl1keep, [x26, #0x80]\n" "add x10, x10, #0x100\n" "21:" // Height 1: Multiply loop: Main loop skip @@ -314,26 +313,26 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "cmp x27, #0x4\n" "blt 23f\n" "22:" // Height 1: Multiply loop: Odd block loop - "ldr d1, [x26], #0x8\n" - "ldr q6, [x10, #0x0]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x70]\n" + "ldr d19, [x26], #0x8\n" + "ldr q18, [x10, #0x0]\n" + "trn1 v19.2d, v19.2d, v17.2d\n" + "ldr q17, [x10, #0x10]\n" + ".inst 0x6e52ee68 // bfmmla v8.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e51ee6c // bfmmla v12.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ee69 // bfmmla v9.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e51ee6d // bfmmla v13.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ee6a // bfmmla v10.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ee6e // bfmmla v14.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" "add x10, x10, #0x80\n" "bge 22b\n" "23:" // Height 1: Multiply loop: Skip odd blocks @@ -346,23 +345,23 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "24:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 "ldr h1, [x26, #0x0]\n" "25:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q7, [x10, #0x0]\n" - "ldr q6, [x10, #0x10]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + "ldr q20, [x10, #0x0]\n" + "ldr q18, [x10, #0x10]\n" + "trn1 v19.2d, v1.2d, v17.2d\n" + ".inst 0x6e54ee68 // bfmmla v8.4s, v19.8h, v20.8h\n" + "ldr q17, [x10, #0x20]\n" + ".inst 0x6e52ee6c // bfmmla v12.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x30]\n" + ".inst 0x6e51ee69 // bfmmla v9.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x40]\n" + ".inst 0x6e52ee6d // bfmmla v13.4s, v19.8h, v18.8h\n" + "ldr q2, [x10, #0x50]\n" + ".inst 0x6e51ee6a // bfmmla v10.4s, v19.8h, v17.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e42ee6e // bfmmla v14.4s, v19.8h, v2.8h\n" + "ldr q17, [x10, #0x70]\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" "add x10, x10, #0x80\n" "26:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -376,17 +375,17 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "uzp1 v11.2d, v11.2d, v15.2d\n" "tbz %x[flags], #1, 27f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v18.4s }, [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" - "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" + "ld1r { v17.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v18.4s\n" + "fmin v9.4s, v9.4s, v18.4s\n" + "fmin v10.4s, v10.4s, v18.4s\n" + "fmin v11.4s, v11.4s, v18.4s\n" + "fmax v8.4s, v8.4s, v17.4s\n" + "fmax v9.4s, v9.4s, v17.4s\n" + "fmax v10.4s, v10.4s, v17.4s\n" + "fmax v11.4s, v11.4s, v17.4s\n" "27:" // Height 1: No activation "cmp x11, #0x10\n" "bge 36f\n" @@ -577,12 +576,12 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "53:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 54f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" "cbnz x28, 55f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -590,7 +589,7 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "b 55f\n" "54:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" "55:" // Height 2: input setup done "cmp x27, #0x8\n" "blt 58f\n" @@ -601,85 +600,85 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr q6, [x10, #0x10]\n" "blt 57f\n" "56:" // Height 2: Multiply loop: Main loop head - "trn1 v0.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" + "trn1 v19.2d, v1.2d, v2.2d\n" + ".inst 0x6e47ee68 // bfmmla v8.4s, v19.8h, v7.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e46ee6c // bfmmla v12.4s, v19.8h, v6.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ee69 // bfmmla v9.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e51ee6d // bfmmla v13.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ee6a // bfmmla v10.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ee6e // bfmmla v14.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x80]\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x90]\n" + ".inst 0x6e52ec28 // bfmmla v8.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xa0]\n" + ".inst 0x6e51ec2c // bfmmla v12.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xb0]\n" + ".inst 0x6e52ec29 // bfmmla v9.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xc0]\n" + ".inst 0x6e51ec2d // bfmmla v13.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xd0]\n" + ".inst 0x6e52ec2a // bfmmla v10.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xe0]\n" + ".inst 0x6e51ec2e // bfmmla v14.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xf0]\n" "sub x27, x27, #0x8\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "ldr q2, [x25, #0x0]\n" "cmp x27, #0x10\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" + ".inst 0x6e52ec2b // bfmmla v11.4s, v1.8h, v18.8h\n" "add x10, x10, #0x100\n" "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + ".inst 0x6e51ec2f // bfmmla v15.4s, v1.8h, v17.8h\n" "ldr q1, [x26, #0x0]\n" "ldr q6, [x10, #0x10]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" "bge 56b\n" "57:" // Height 2: Multiply loop: Single iteration only - "trn1 v0.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" + "trn1 v19.2d, v1.2d, v2.2d\n" + ".inst 0x6e47ee68 // bfmmla v8.4s, v19.8h, v7.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e46ee6c // bfmmla v12.4s, v19.8h, v6.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ee69 // bfmmla v9.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e51ee6d // bfmmla v13.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ee6a // bfmmla v10.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ee6e // bfmmla v14.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x80]\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x90]\n" + ".inst 0x6e52ec28 // bfmmla v8.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xa0]\n" + ".inst 0x6e51ec2c // bfmmla v12.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xb0]\n" + ".inst 0x6e52ec29 // bfmmla v9.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xc0]\n" + ".inst 0x6e51ec2d // bfmmla v13.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xd0]\n" + ".inst 0x6e52ec2a // bfmmla v10.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0xe0]\n" + ".inst 0x6e51ec2e // bfmmla v14.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0xf0]\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + ".inst 0x6e52ec2b // bfmmla v11.4s, v1.8h, v18.8h\n" + ".inst 0x6e51ec2f // bfmmla v15.4s, v1.8h, v17.8h\n" "sub x27, x27, #0x8\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" @@ -689,27 +688,27 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "cmp x27, #0x4\n" "blt 60f\n" "59:" // Height 2: Multiply loop: Odd block loop - "ldr d1, [x26], #0x8\n" - "ldr d2, [x25], #0x8\n" - "trn1 v0.2d, v1.2d, v2.2d\n" + "ldr d18, [x26], #0x8\n" + "ldr d17, [x25], #0x8\n" + "trn1 v19.2d, v18.2d, v17.2d\n" "sub x27, x27, #0x4\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - "ldr q6, [x10, #0x20]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - "ldr q6, [x10, #0x40]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" - "ldr q6, [x10, #0x60]\n" - "ldr q7, [x10, #0x70]\n" + "ldr q18, [x10, #0x0]\n" + "ldr q17, [x10, #0x10]\n" + ".inst 0x6e52ee68 // bfmmla v8.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6c // bfmmla v12.4s, v19.8h, v17.8h\n" + "ldr q26, [x10, #0x20]\n" + "ldr q5, [x10, #0x30]\n" + ".inst 0x6e5aee69 // bfmmla v9.4s, v19.8h, v26.8h\n" + ".inst 0x6e45ee6d // bfmmla v13.4s, v19.8h, v5.8h\n" + "ldr q18, [x10, #0x40]\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ee6a // bfmmla v10.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6e // bfmmla v14.4s, v19.8h, v17.8h\n" + "ldr q18, [x10, #0x60]\n" + "ldr q17, [x10, #0x70]\n" "cmp x27, #0x4\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" "add x10, x10, #0x80\n" "bge 59b\n" "60:" // Height 2: Multiply loop: Skip odd blocks @@ -725,23 +724,23 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr h1, [x26, #0x0]\n" "ldr h2, [x25, #0x0]\n" "62:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q7, [x10, #0x0]\n" - "ldr q6, [x10, #0x10]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + "ldr q18, [x10, #0x0]\n" + "ldr q17, [x10, #0x10]\n" + "trn1 v19.2d, v1.2d, v2.2d\n" + ".inst 0x6e52ee68 // bfmmla v8.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e51ee6c // bfmmla v12.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ee69 // bfmmla v9.4s, v19.8h, v18.8h\n" + "ldr q30, [x10, #0x40]\n" + ".inst 0x6e51ee6d // bfmmla v13.4s, v19.8h, v17.8h\n" + "ldr q26, [x10, #0x50]\n" + ".inst 0x6e5eee6a // bfmmla v10.4s, v19.8h, v30.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e5aee6e // bfmmla v14.4s, v19.8h, v26.8h\n" + "ldr q17, [x10, #0x70]\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" "add x10, x10, #0x80\n" "63:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -762,25 +761,25 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "uzp2 v11.2d, v11.2d, v15.2d\n" "tbz %x[flags], #1, 64f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v18.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v7.4s, v7.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, 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" - "fmax v7.4s, v7.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, 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" + "ld1r { v17.4s }, [x20]\n" + "fmin v7.4s, v7.4s, v18.4s\n" + "fmin v12.4s, v12.4s, v18.4s\n" + "fmin v13.4s, v13.4s, v18.4s\n" + "fmin v14.4s, v14.4s, v18.4s\n" + "fmin v8.4s, v8.4s, v18.4s\n" + "fmin v9.4s, v9.4s, v18.4s\n" + "fmin v10.4s, v10.4s, v18.4s\n" + "fmin v11.4s, v11.4s, v18.4s\n" + "fmax v7.4s, v7.4s, v17.4s\n" + "fmax v12.4s, v12.4s, v17.4s\n" + "fmax v13.4s, v13.4s, v17.4s\n" + "fmax v14.4s, v14.4s, v17.4s\n" + "fmax v8.4s, v8.4s, v17.4s\n" + "fmax v9.4s, v9.4s, v17.4s\n" + "fmax v10.4s, v10.4s, v17.4s\n" + "fmax v11.4s, v11.4s, v17.4s\n" "64:" // Height 2: No activation "cmp x11, #0x10\n" "bge 73f\n" @@ -1036,13 +1035,13 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "90:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 91f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" "cbnz x28, 92f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -1051,8 +1050,8 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "b 92f\n" "91:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" "92:" // Height 3: input setup done "cmp x27, #0x8\n" "blt 95f\n" @@ -1064,167 +1063,167 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr q6, [x10, #0x10]\n" "blt 94f\n" "93:" // Height 3: Multiply loop: Main loop head - "trn1 v0.2d, v1.2d, v2.2d\n" + "trn1 v28.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "trn2 v3.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e47ef88 // bfmmla v8.4s, v28.8h, v7.8h\n" + "trn1 v27.2d, v3.2d, v29.2d\n" + ".inst 0x6e47ef70 // bfmmla v16.4s, v27.8h, v7.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e46ef8c // bfmmla v12.4s, v28.8h, v6.8h\n" + ".inst 0x6e46ef74 // bfmmla v20.4s, v27.8h, v6.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" + "trn2 v3.2d, v3.2d, v29.2d\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" "sub x27, x27, #0x8\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" "add x26, x26, #0x10\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" "add x25, x25, #0x10\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" "add x24, x24, #0x10\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x80]\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" "cmp x27, #0x10\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x90]\n" "ldr q2, [x25, #0x0]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" + ".inst 0x6e5aec28 // bfmmla v8.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec70 // bfmmla v16.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xa0]\n" + ".inst 0x6e59ec2c // bfmmla v12.4s, v1.8h, v25.8h\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" + ".inst 0x6e59ec74 // bfmmla v20.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xb0]\n" + ".inst 0x6e5aec29 // bfmmla v9.4s, v1.8h, v26.8h\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" + ".inst 0x6e5aec71 // bfmmla v17.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xc0]\n" + ".inst 0x6e59ec2d // bfmmla v13.4s, v1.8h, v25.8h\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + ".inst 0x6e59ec75 // bfmmla v21.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xd0]\n" + ".inst 0x6e5aec2a // bfmmla v10.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec72 // bfmmla v18.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xe0]\n" + ".inst 0x6e59ec2e // bfmmla v14.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec76 // bfmmla v22.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec73 // bfmmla v19.4s, v3.8h, v7.8h\n" + ".inst 0x6e5aec2b // bfmmla v11.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec73 // bfmmla v19.4s, v3.8h, v26.8h\n" "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + ".inst 0x6e59ec2f // bfmmla v15.4s, v1.8h, v25.8h\n" "ldr q1, [x26, #0x0]\n" - ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" + ".inst 0x6e59ec77 // bfmmla v23.4s, v3.8h, v25.8h\n" "ldr q3, [x24, #0x0]\n" "ldr q6, [x10, #0x10]\n" "bge 93b\n" "94:" // Height 3: Multiply loop: Single iteration only - "trn1 v0.2d, v1.2d, v2.2d\n" + "trn1 v28.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "trn2 v3.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e47ef88 // bfmmla v8.4s, v28.8h, v7.8h\n" + "trn1 v27.2d, v3.2d, v29.2d\n" + ".inst 0x6e47ef70 // bfmmla v16.4s, v27.8h, v7.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e46ef8c // bfmmla v12.4s, v28.8h, v6.8h\n" + ".inst 0x6e46ef74 // bfmmla v20.4s, v27.8h, v6.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" + "trn2 v3.2d, v3.2d, v29.2d\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" "add x26, x26, #0x10\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" "add x25, x25, #0x10\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" "add x24, x24, #0x10\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" "sub x27, x27, #0x8\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x80]\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x90]\n" + ".inst 0x6e5aec28 // bfmmla v8.4s, v1.8h, v26.8h\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" + ".inst 0x6e5aec70 // bfmmla v16.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xa0]\n" + ".inst 0x6e59ec2c // bfmmla v12.4s, v1.8h, v25.8h\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + ".inst 0x6e59ec74 // bfmmla v20.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xb0]\n" + ".inst 0x6e5aec29 // bfmmla v9.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec71 // bfmmla v17.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xc0]\n" + ".inst 0x6e59ec2d // bfmmla v13.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec75 // bfmmla v21.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xd0]\n" + ".inst 0x6e5aec2a // bfmmla v10.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec72 // bfmmla v18.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xe0]\n" + ".inst 0x6e59ec2e // bfmmla v14.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec76 // bfmmla v22.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec73 // bfmmla v19.4s, v3.8h, v7.8h\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" + ".inst 0x6e5aec2b // bfmmla v11.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec73 // bfmmla v19.4s, v3.8h, v26.8h\n" + ".inst 0x6e59ec2f // bfmmla v15.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec77 // bfmmla v23.4s, v3.8h, v25.8h\n" "95:" // Height 3: Multiply loop: Main loop skip "cbz x27, 100f\n" "cmp x27, #0x4\n" "blt 97f\n" "96:" // Height 3: Multiply loop: Odd block loop - "ldr d1, [x26], #0x8\n" - "ldr d2, [x25], #0x8\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - "ldr d3, [x24], #0x8\n" - "ldr q6, [x10, #0x0]\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" + "ldr d26, [x26], #0x8\n" + "ldr d25, [x25], #0x8\n" + "trn1 v28.2d, v26.2d, v25.2d\n" + "ldr d25, [x24], #0x8\n" + "ldr q26, [x10, #0x0]\n" + "trn1 v27.2d, v25.2d, v27.2d\n" + ".inst 0x6e5aef88 // bfmmla v8.4s, v28.8h, v26.8h\n" + "ldr q25, [x10, #0x10]\n" + ".inst 0x6e5aef70 // bfmmla v16.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e59ef8c // bfmmla v12.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef74 // bfmmla v20.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" "sub x27, x27, #0x4\n" - ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" "cmp x27, #0x4\n" - ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec52 // bfmmla v18.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" "add x10, x10, #0x80\n" - ".inst 0x6e46ec53 // bfmmla v19.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" "bge 96b\n" "97:" // Height 3: Multiply loop: Skip odd blocks "cbz x27, 100f\n" @@ -1242,33 +1241,33 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr h2, [x25, #0x0]\n" "ldr h3, [x24, #0x0]\n" "99:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q7, [x10, #0x0]\n" - "ldr q6, [x10, #0x10]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" + "ldr q26, [x10, #0x0]\n" + "ldr q29, [x10, #0x10]\n" + "trn1 v28.2d, v1.2d, v2.2d\n" + "trn1 v27.2d, v3.2d, v25.2d\n" + ".inst 0x6e5aef88 // bfmmla v8.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef70 // bfmmla v16.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e5def8c // bfmmla v12.4s, v28.8h, v29.8h\n" + ".inst 0x6e5def74 // bfmmla v20.4s, v27.8h, v29.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" "add x10, x10, #0x80\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" "100:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1294,33 +1293,33 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "uzp1 v19.2d, v19.2d, v23.2d\n" "tbz %x[flags], #1, 101f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v26.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v7.4s, v7.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, 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 v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v19.4s, v19.4s, v1.4s\n" - "fmax v7.4s, v7.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, 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 v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v19.4s, v19.4s, v0.4s\n" + "ld1r { v25.4s }, [x20]\n" + "fmin v7.4s, v7.4s, v26.4s\n" + "fmin v12.4s, v12.4s, v26.4s\n" + "fmin v13.4s, v13.4s, v26.4s\n" + "fmin v14.4s, v14.4s, v26.4s\n" + "fmin v8.4s, v8.4s, v26.4s\n" + "fmin v9.4s, v9.4s, v26.4s\n" + "fmin v10.4s, v10.4s, v26.4s\n" + "fmin v11.4s, v11.4s, v26.4s\n" + "fmin v16.4s, v16.4s, v26.4s\n" + "fmin v17.4s, v17.4s, v26.4s\n" + "fmin v18.4s, v18.4s, v26.4s\n" + "fmin v19.4s, v19.4s, v26.4s\n" + "fmax v7.4s, v7.4s, v25.4s\n" + "fmax v12.4s, v12.4s, v25.4s\n" + "fmax v13.4s, v13.4s, v25.4s\n" + "fmax v14.4s, v14.4s, v25.4s\n" + "fmax v8.4s, v8.4s, v25.4s\n" + "fmax v9.4s, v9.4s, v25.4s\n" + "fmax v10.4s, v10.4s, v25.4s\n" + "fmax v11.4s, v11.4s, v25.4s\n" + "fmax v16.4s, v16.4s, v25.4s\n" + "fmax v17.4s, v17.4s, v25.4s\n" + "fmax v18.4s, v18.4s, v25.4s\n" + "fmax v19.4s, v19.4s, v25.4s\n" "101:" // Height 3: No activation "cmp x11, #0x10\n" "bge 110f\n" @@ -1617,14 +1616,14 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "127:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 128f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" "cbnz x28, 129f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -1634,9 +1633,9 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "b 129f\n" "128:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" "129:" // Height 4: input setup done "cmp x27, #0x8\n" "blt 132f\n" @@ -1649,173 +1648,173 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr q6, [x10, #0x10]\n" "blt 131f\n" "130:" // Height 4: Multiply loop: Main loop head - "trn1 v0.2d, v1.2d, v2.2d\n" + "trn1 v28.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" + ".inst 0x6e47ef88 // bfmmla v8.4s, v28.8h, v7.8h\n" "sub x27, x27, #0x8\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + "trn1 v27.2d, v3.2d, v4.2d\n" + ".inst 0x6e47ef70 // bfmmla v16.4s, v27.8h, v7.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e46ef8c // bfmmla v12.4s, v28.8h, v6.8h\n" + ".inst 0x6e46ef74 // bfmmla v20.4s, v27.8h, v6.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" "trn2 v3.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" "add x26, x26, #0x10\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" "add x25, x25, #0x10\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" "add x24, x24, #0x10\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" "add x23, x23, #0x10\n" "ldr q4, [x23, #0x0]\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x80]\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x90]\n" "ldr q2, [x25, #0x0]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" + ".inst 0x6e5aec28 // bfmmla v8.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec70 // bfmmla v16.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xa0]\n" + ".inst 0x6e59ec2c // bfmmla v12.4s, v1.8h, v25.8h\n" "cmp x27, #0x10\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" + ".inst 0x6e59ec74 // bfmmla v20.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xb0]\n" + ".inst 0x6e5aec29 // bfmmla v9.4s, v1.8h, v26.8h\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" + ".inst 0x6e5aec71 // bfmmla v17.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xc0]\n" + ".inst 0x6e59ec2d // bfmmla v13.4s, v1.8h, v25.8h\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" + ".inst 0x6e59ec75 // bfmmla v21.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xd0]\n" + ".inst 0x6e5aec2a // bfmmla v10.4s, v1.8h, v26.8h\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" + ".inst 0x6e5aec72 // bfmmla v18.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xe0]\n" + ".inst 0x6e59ec2e // bfmmla v14.4s, v1.8h, v25.8h\n" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + ".inst 0x6e59ec76 // bfmmla v22.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec73 // bfmmla v19.4s, v3.8h, v7.8h\n" + ".inst 0x6e5aec2b // bfmmla v11.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec73 // bfmmla v19.4s, v3.8h, v26.8h\n" "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + ".inst 0x6e59ec2f // bfmmla v15.4s, v1.8h, v25.8h\n" "ldr q1, [x26, #0x0]\n" - ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" + ".inst 0x6e59ec77 // bfmmla v23.4s, v3.8h, v25.8h\n" "ldr q3, [x24, #0x0]\n" "ldr q6, [x10, #0x10]\n" "bge 130b\n" "131:" // Height 4: Multiply loop: Single iteration only - "trn1 v0.2d, v1.2d, v2.2d\n" + "trn1 v28.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" + ".inst 0x6e47ef88 // bfmmla v8.4s, v28.8h, v7.8h\n" "add x26, x26, #0x10\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + "trn1 v27.2d, v3.2d, v4.2d\n" + ".inst 0x6e47ef70 // bfmmla v16.4s, v27.8h, v7.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e46ef8c // bfmmla v12.4s, v28.8h, v6.8h\n" + ".inst 0x6e46ef74 // bfmmla v20.4s, v27.8h, v6.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" "trn2 v3.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" "add x25, x25, #0x10\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" "add x24, x24, #0x10\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" "add x23, x23, #0x10\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" "sub x27, x27, #0x8\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x80]\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x90]\n" + ".inst 0x6e5aec28 // bfmmla v8.4s, v1.8h, v26.8h\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" + ".inst 0x6e5aec70 // bfmmla v16.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xa0]\n" + ".inst 0x6e59ec2c // bfmmla v12.4s, v1.8h, v25.8h\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" + ".inst 0x6e59ec74 // bfmmla v20.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xb0]\n" + ".inst 0x6e5aec29 // bfmmla v9.4s, v1.8h, v26.8h\n" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + ".inst 0x6e5aec71 // bfmmla v17.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xc0]\n" + ".inst 0x6e59ec2d // bfmmla v13.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec75 // bfmmla v21.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xd0]\n" + ".inst 0x6e5aec2a // bfmmla v10.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec72 // bfmmla v18.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0xe0]\n" + ".inst 0x6e59ec2e // bfmmla v14.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec76 // bfmmla v22.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec73 // bfmmla v19.4s, v3.8h, v7.8h\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" + ".inst 0x6e5aec2b // bfmmla v11.4s, v1.8h, v26.8h\n" + ".inst 0x6e5aec73 // bfmmla v19.4s, v3.8h, v26.8h\n" + ".inst 0x6e59ec2f // bfmmla v15.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec77 // bfmmla v23.4s, v3.8h, v25.8h\n" "132:" // Height 4: Multiply loop: Main loop skip "cbz x27, 137f\n" "cmp x27, #0x4\n" "blt 134f\n" "133:" // Height 4: Multiply loop: Odd block loop - "ldr d1, [x26], #0x8\n" - "ldr d2, [x25], #0x8\n" - "trn1 v0.2d, v1.2d, v2.2d\n" + "ldr d26, [x26], #0x8\n" + "ldr d25, [x25], #0x8\n" + "trn1 v28.2d, v26.2d, v25.2d\n" "sub x27, x27, #0x4\n" - "ldr d3, [x24], #0x8\n" - "ldr d4, [x23], #0x8\n" - "trn1 v2.2d, v3.2d, v4.2d\n" + "ldr d26, [x24], #0x8\n" + "ldr d25, [x23], #0x8\n" + "trn1 v27.2d, v26.2d, v25.2d\n" "cmp x27, #0x4\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec52 // bfmmla v18.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x70]\n" + "ldr q26, [x10, #0x0]\n" + "ldr q25, [x10, #0x10]\n" + ".inst 0x6e5aef88 // bfmmla v8.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef70 // bfmmla v16.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e59ef8c // bfmmla v12.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef74 // bfmmla v20.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" "add x10, x10, #0x80\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec53 // bfmmla v19.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" "bge 133b\n" "134:" // Height 4: Multiply loop: Skip odd blocks "cbz x27, 137f\n" @@ -1836,33 +1835,33 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr h3, [x24, #0x0]\n" "ldr h4, [x23, #0x0]\n" "136:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q7, [x10, #0x0]\n" - "ldr q6, [x10, #0x10]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" + "ldr q26, [x10, #0x0]\n" + "ldr q25, [x10, #0x10]\n" + "trn1 v28.2d, v1.2d, v2.2d\n" + "trn1 v27.2d, v3.2d, v4.2d\n" + ".inst 0x6e5aef88 // bfmmla v8.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef70 // bfmmla v16.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e59ef8c // bfmmla v12.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef74 // bfmmla v20.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" "add x10, x10, #0x80\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" "137:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1894,41 +1893,41 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "uzp2 v19.2d, v19.2d, v23.2d\n" "tbz %x[flags], #1, 138f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v26.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v7.4s, v7.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, 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 v15.4s, v15.4s, v1.4s\n" - "fmin v20.4s, v20.4s, v1.4s\n" - "fmin v21.4s, v21.4s, v1.4s\n" - "fmin v22.4s, v22.4s, 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 v19.4s, v19.4s, v1.4s\n" - "fmax v7.4s, v7.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, 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 v15.4s, v15.4s, v0.4s\n" - "fmax v20.4s, v20.4s, v0.4s\n" - "fmax v21.4s, v21.4s, v0.4s\n" - "fmax v22.4s, v22.4s, 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 v19.4s, v19.4s, v0.4s\n" + "ld1r { v25.4s }, [x20]\n" + "fmin v7.4s, v7.4s, v26.4s\n" + "fmin v12.4s, v12.4s, v26.4s\n" + "fmin v13.4s, v13.4s, v26.4s\n" + "fmin v14.4s, v14.4s, v26.4s\n" + "fmin v8.4s, v8.4s, v26.4s\n" + "fmin v9.4s, v9.4s, v26.4s\n" + "fmin v10.4s, v10.4s, v26.4s\n" + "fmin v11.4s, v11.4s, v26.4s\n" + "fmin v15.4s, v15.4s, v26.4s\n" + "fmin v20.4s, v20.4s, v26.4s\n" + "fmin v21.4s, v21.4s, v26.4s\n" + "fmin v22.4s, v22.4s, v26.4s\n" + "fmin v16.4s, v16.4s, v26.4s\n" + "fmin v17.4s, v17.4s, v26.4s\n" + "fmin v18.4s, v18.4s, v26.4s\n" + "fmin v19.4s, v19.4s, v26.4s\n" + "fmax v7.4s, v7.4s, v25.4s\n" + "fmax v12.4s, v12.4s, v25.4s\n" + "fmax v13.4s, v13.4s, v25.4s\n" + "fmax v14.4s, v14.4s, v25.4s\n" + "fmax v8.4s, v8.4s, v25.4s\n" + "fmax v9.4s, v9.4s, v25.4s\n" + "fmax v10.4s, v10.4s, v25.4s\n" + "fmax v11.4s, v11.4s, v25.4s\n" + "fmax v15.4s, v15.4s, v25.4s\n" + "fmax v20.4s, v20.4s, v25.4s\n" + "fmax v21.4s, v21.4s, v25.4s\n" + "fmax v22.4s, v22.4s, v25.4s\n" + "fmax v16.4s, v16.4s, v25.4s\n" + "fmax v17.4s, v17.4s, v25.4s\n" + "fmax v18.4s, v18.4s, v25.4s\n" + "fmax v19.4s, v19.4s, v25.4s\n" "138:" // Height 4: No activation "cmp x11, #0x10\n" "bge 147f\n" @@ -2290,15 +2289,15 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "164:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 165f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" "cbnz x28, 166f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -2309,10 +2308,10 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "b 166f\n" "165:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" + "add x22, x23, x21, LSL #1\n" "166:" // Height 5: input setup done "cmp x27, #0x8\n" "blt 169f\n" @@ -2325,174 +2324,174 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr q7, [x10, #0x0]\n" "blt 168f\n" "167:" // Height 5: Multiply loop: Main loop head - "trn1 v0.2d, v1.2d, v2.2d\n" + "trn1 v6.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" + ".inst 0x6e47ecc8 // bfmmla v8.4s, v6.8h, v7.8h\n" "trn1 v2.2d, v3.2d, v4.2d\n" "trn2 v3.2d, v3.2d, v4.2d\n" ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" "sub x27, x27, #0x8\n" - "trn1 v4.2d, v5.2d, v6.2d\n" - "trn2 v5.2d, v5.2d, v6.2d\n" - "ldr q6, [x10, #0x10]\n" + "trn1 v4.2d, v5.2d, v0.2d\n" + "trn2 v5.2d, v5.2d, v0.2d\n" + "ldr q0, [x10, #0x10]\n" ".inst 0x6e47ec98 // bfmmla v24.4s, v4.8h, v7.8h\n" "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" + ".inst 0x6e40eccc // bfmmla v12.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec54 // bfmmla v20.4s, v2.8h, v0.8h\n" "add x26, x26, #0x10\n" - ".inst 0x6e46ec9c // bfmmla v28.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9c // bfmmla v28.4s, v4.8h, v0.8h\n" + "ldr q0, [x10, #0x30]\n" + ".inst 0x6e47ecc9 // bfmmla v9.4s, v6.8h, v7.8h\n" "add x25, x25, #0x10\n" ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" ".inst 0x6e47ec99 // bfmmla v25.4s, v4.8h, v7.8h\n" "ldr q7, [x10, #0x40]\n" "add x24, x24, #0x10\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" + ".inst 0x6e40eccd // bfmmla v13.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec55 // bfmmla v21.4s, v2.8h, v0.8h\n" "add x23, x23, #0x10\n" "add x22, x22, #0x10\n" - ".inst 0x6e46ec9d // bfmmla v29.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9d // bfmmla v29.4s, v4.8h, v0.8h\n" + "ldr q0, [x10, #0x50]\n" + ".inst 0x6e47ecca // bfmmla v10.4s, v6.8h, v7.8h\n" "cmp x27, #0x10\n" ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" ".inst 0x6e47ec9a // bfmmla v26.4s, v4.8h, v7.8h\n" "ldr q7, [x10, #0x60]\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" + ".inst 0x6e40ecce // bfmmla v14.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec56 // bfmmla v22.4s, v2.8h, v0.8h\n" "prfm pldl1keep, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6e46ec9e // bfmmla v30.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9e // bfmmla v30.4s, v4.8h, v0.8h\n" + "ldr q0, [x10, #0x70]\n" + ".inst 0x6e47eccb // bfmmla v11.4s, v6.8h, v7.8h\n" "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" ".inst 0x6e47ec9b // bfmmla v27.4s, v4.8h, v7.8h\n" "ldr q7, [x10, #0x80]\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" + ".inst 0x6e40eccf // bfmmla v15.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec57 // bfmmla v23.4s, v2.8h, v0.8h\n" "ldr q2, [x25, #0x0]\n" - ".inst 0x6e46ec9f // bfmmla v31.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" + ".inst 0x6e40ec9f // bfmmla v31.4s, v4.8h, v0.8h\n" + "ldr q0, [x10, #0x90]\n" "ldr q4, [x23, #0x0]\n" ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" ".inst 0x6e47ecb8 // bfmmla v24.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbc // bfmmla v28.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecb9 // bfmmla v25.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbd // bfmmla v29.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecba // bfmmla v26.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbe // bfmmla v30.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + "ldr q6, [x10, #0xa0]\n" + ".inst 0x6e40ec2c // bfmmla v12.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec74 // bfmmla v20.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecbc // bfmmla v28.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xb0]\n" + ".inst 0x6e46ec29 // bfmmla v9.4s, v1.8h, v6.8h\n" + ".inst 0x6e46ec71 // bfmmla v17.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecb9 // bfmmla v25.4s, v5.8h, v6.8h\n" + "ldr q6, [x10, #0xc0]\n" + ".inst 0x6e40ec2d // bfmmla v13.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec75 // bfmmla v21.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecbd // bfmmla v29.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xd0]\n" + ".inst 0x6e46ec2a // bfmmla v10.4s, v1.8h, v6.8h\n" + ".inst 0x6e46ec72 // bfmmla v18.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecba // bfmmla v26.4s, v5.8h, v6.8h\n" + "ldr q6, [x10, #0xe0]\n" + ".inst 0x6e40ec2e // bfmmla v14.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec76 // bfmmla v22.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecbe // bfmmla v30.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec73 // bfmmla v19.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecbb // bfmmla v27.4s, v5.8h, v7.8h\n" + ".inst 0x6e46ec2b // bfmmla v11.4s, v1.8h, v6.8h\n" + ".inst 0x6e46ec73 // bfmmla v19.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecbb // bfmmla v27.4s, v5.8h, v6.8h\n" "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + ".inst 0x6e40ec2f // bfmmla v15.4s, v1.8h, v0.8h\n" "ldr q1, [x26, #0x0]\n" - ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" + ".inst 0x6e40ec77 // bfmmla v23.4s, v3.8h, v0.8h\n" "ldr q3, [x24, #0x0]\n" - ".inst 0x6e46ecbf // bfmmla v31.4s, v5.8h, v6.8h\n" + ".inst 0x6e40ecbf // bfmmla v31.4s, v5.8h, v0.8h\n" "ldr q5, [x22, #0x0]\n" "bge 167b\n" "168:" // Height 5: Multiply loop: Single iteration only - "trn1 v0.2d, v1.2d, v2.2d\n" + "trn1 v6.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" + ".inst 0x6e47ecc8 // bfmmla v8.4s, v6.8h, v7.8h\n" "trn1 v2.2d, v3.2d, v4.2d\n" "trn2 v3.2d, v3.2d, v4.2d\n" ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" "add x26, x26, #0x10\n" - "trn1 v4.2d, v5.2d, v6.2d\n" - "trn2 v5.2d, v5.2d, v6.2d\n" - "ldr q6, [x10, #0x10]\n" + "trn1 v4.2d, v5.2d, v0.2d\n" + "trn2 v5.2d, v5.2d, v0.2d\n" + "ldr q0, [x10, #0x10]\n" ".inst 0x6e47ec98 // bfmmla v24.4s, v4.8h, v7.8h\n" "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" + ".inst 0x6e40eccc // bfmmla v12.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec54 // bfmmla v20.4s, v2.8h, v0.8h\n" "add x25, x25, #0x10\n" - ".inst 0x6e46ec9c // bfmmla v28.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9c // bfmmla v28.4s, v4.8h, v0.8h\n" + "ldr q0, [x10, #0x30]\n" + ".inst 0x6e47ecc9 // bfmmla v9.4s, v6.8h, v7.8h\n" "add x24, x24, #0x10\n" ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" ".inst 0x6e47ec99 // bfmmla v25.4s, v4.8h, v7.8h\n" "ldr q7, [x10, #0x40]\n" "add x23, x23, #0x10\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" + ".inst 0x6e40eccd // bfmmla v13.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec55 // bfmmla v21.4s, v2.8h, v0.8h\n" "add x22, x22, #0x10\n" "sub x27, x27, #0x8\n" - ".inst 0x6e46ec9d // bfmmla v29.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9d // bfmmla v29.4s, v4.8h, v0.8h\n" + "ldr q0, [x10, #0x50]\n" + ".inst 0x6e47ecca // bfmmla v10.4s, v6.8h, v7.8h\n" "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" ".inst 0x6e47ec9a // bfmmla v26.4s, v4.8h, v7.8h\n" "ldr q7, [x10, #0x60]\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" + ".inst 0x6e40ecce // bfmmla v14.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec56 // bfmmla v22.4s, v2.8h, v0.8h\n" "prfm pldl1keep, [x24, #0x80]\n" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x6e46ec9e // bfmmla v30.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x70]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9e // bfmmla v30.4s, v4.8h, v0.8h\n" + "ldr q0, [x10, #0x70]\n" + ".inst 0x6e47eccb // bfmmla v11.4s, v6.8h, v7.8h\n" "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" ".inst 0x6e47ec9b // bfmmla v27.4s, v4.8h, v7.8h\n" "ldr q7, [x10, #0x80]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9f // bfmmla v31.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" + ".inst 0x6e40eccf // bfmmla v15.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec57 // bfmmla v23.4s, v2.8h, v0.8h\n" + ".inst 0x6e40ec9f // bfmmla v31.4s, v4.8h, v0.8h\n" + "ldr q2, [x10, #0x90]\n" ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" ".inst 0x6e47ecb8 // bfmmla v24.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbc // bfmmla v28.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecb9 // bfmmla v25.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbd // bfmmla v29.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecba // bfmmla v26.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbe // bfmmla v30.4s, v5.8h, v6.8h\n" + "ldr q0, [x10, #0xa0]\n" + ".inst 0x6e42ec2c // bfmmla v12.4s, v1.8h, v2.8h\n" + ".inst 0x6e42ec74 // bfmmla v20.4s, v3.8h, v2.8h\n" + ".inst 0x6e42ecbc // bfmmla v28.4s, v5.8h, v2.8h\n" + "ldr q2, [x10, #0xb0]\n" + ".inst 0x6e40ec29 // bfmmla v9.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec71 // bfmmla v17.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecb9 // bfmmla v25.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xc0]\n" + ".inst 0x6e42ec2d // bfmmla v13.4s, v1.8h, v2.8h\n" + ".inst 0x6e42ec75 // bfmmla v21.4s, v3.8h, v2.8h\n" + ".inst 0x6e42ecbd // bfmmla v29.4s, v5.8h, v2.8h\n" + "ldr q2, [x10, #0xd0]\n" + ".inst 0x6e40ec2a // bfmmla v10.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec72 // bfmmla v18.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecba // bfmmla v26.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xe0]\n" + ".inst 0x6e42ec2e // bfmmla v14.4s, v1.8h, v2.8h\n" + ".inst 0x6e42ec76 // bfmmla v22.4s, v3.8h, v2.8h\n" + ".inst 0x6e42ecbe // bfmmla v30.4s, v5.8h, v2.8h\n" "ldr q6, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec73 // bfmmla v19.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecbb // bfmmla v27.4s, v5.8h, v7.8h\n" + ".inst 0x6e40ec2b // bfmmla v11.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec73 // bfmmla v19.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecbb // bfmmla v27.4s, v5.8h, v0.8h\n" ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" ".inst 0x6e46ecbf // bfmmla v31.4s, v5.8h, v6.8h\n" @@ -2502,48 +2501,48 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "blt 171f\n" "170:" // Height 5: Multiply loop: Odd block loop "ldr d1, [x26], #0x8\n" - "ldr d2, [x25], #0x8\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - "ldr d3, [x24], #0x8\n" - "ldr d4, [x23], #0x8\n" - "trn1 v2.2d, v3.2d, v4.2d\n" + "ldr d0, [x25], #0x8\n" + "trn1 v4.2d, v1.2d, v0.2d\n" + "ldr d1, [x24], #0x8\n" + "ldr d0, [x23], #0x8\n" + "trn1 v3.2d, v1.2d, v0.2d\n" "sub x27, x27, #0x4\n" - "ldr d5, [x22], #0x8\n" - "ldr q6, [x10, #0x0]\n" - "trn1 v4.2d, v5.2d, v7.2d\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec98 // bfmmla v24.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" + "ldr d0, [x22], #0x8\n" + "ldr q1, [x10, #0x0]\n" + "trn1 v2.2d, v0.2d, v2.2d\n" + ".inst 0x6e41ec88 // bfmmla v8.4s, v4.8h, v1.8h\n" + "ldr q0, [x10, #0x10]\n" + ".inst 0x6e41ec70 // bfmmla v16.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec58 // bfmmla v24.4s, v2.8h, v1.8h\n" + "ldr q1, [x10, #0x20]\n" + ".inst 0x6e40ec8c // bfmmla v12.4s, v4.8h, v0.8h\n" + ".inst 0x6e40ec74 // bfmmla v20.4s, v3.8h, v0.8h\n" "cmp x27, #0x4\n" - ".inst 0x6e47ec9c // bfmmla v28.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec99 // bfmmla v25.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9d // bfmmla v29.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec52 // bfmmla v18.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9a // bfmmla v26.4s, v4.8h, v6.8h\n" + ".inst 0x6e40ec5c // bfmmla v28.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x30]\n" + ".inst 0x6e41ec89 // bfmmla v9.4s, v4.8h, v1.8h\n" + ".inst 0x6e41ec71 // bfmmla v17.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec59 // bfmmla v25.4s, v2.8h, v1.8h\n" + "ldr q1, [x10, #0x40]\n" + ".inst 0x6e40ec8d // bfmmla v13.4s, v4.8h, v0.8h\n" + ".inst 0x6e40ec75 // bfmmla v21.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5d // bfmmla v29.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x50]\n" + ".inst 0x6e41ec8a // bfmmla v10.4s, v4.8h, v1.8h\n" + ".inst 0x6e41ec72 // bfmmla v18.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5a // bfmmla v26.4s, v2.8h, v1.8h\n" "ldr q6, [x10, #0x60]\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9e // bfmmla v30.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" + ".inst 0x6e40ec8e // bfmmla v14.4s, v4.8h, v0.8h\n" + ".inst 0x6e40ec76 // bfmmla v22.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5e // bfmmla v30.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x70]\n" + ".inst 0x6e46ec8b // bfmmla v11.4s, v4.8h, v6.8h\n" "add x10, x10, #0x80\n" - ".inst 0x6e46ec53 // bfmmla v19.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9b // bfmmla v27.4s, v4.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9f // bfmmla v31.4s, v4.8h, v7.8h\n" + ".inst 0x6e46ec73 // bfmmla v19.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ec5b // bfmmla v27.4s, v2.8h, v6.8h\n" + ".inst 0x6e40ec8f // bfmmla v15.4s, v4.8h, v0.8h\n" + ".inst 0x6e40ec77 // bfmmla v23.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5f // bfmmla v31.4s, v2.8h, v0.8h\n" "bge 170b\n" "171:" // Height 5: Multiply loop: Skip odd blocks "cbz x27, 174f\n" @@ -2567,42 +2566,42 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr h4, [x23, #0x0]\n" "ldr h5, [x22, #0x0]\n" "173:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q7, [x10, #0x0]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - "trn1 v4.2d, v5.2d, v6.2d\n" - "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec98 // bfmmla v24.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9c // bfmmla v28.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec99 // bfmmla v25.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9d // bfmmla v29.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9a // bfmmla v26.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9e // bfmmla v30.4s, v4.8h, v6.8h\n" + "ldr q6, [x10, #0x0]\n" + "trn1 v7.2d, v1.2d, v2.2d\n" + "trn1 v3.2d, v3.2d, v4.2d\n" + "trn1 v2.2d, v5.2d, v0.2d\n" + "ldr q1, [x10, #0x10]\n" + ".inst 0x6e46ece8 // bfmmla v8.4s, v7.8h, v6.8h\n" + ".inst 0x6e46ec70 // bfmmla v16.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ec58 // bfmmla v24.4s, v2.8h, v6.8h\n" + "ldr q0, [x10, #0x20]\n" + ".inst 0x6e41ecec // bfmmla v12.4s, v7.8h, v1.8h\n" + ".inst 0x6e41ec74 // bfmmla v20.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5c // bfmmla v28.4s, v2.8h, v1.8h\n" + "ldr q1, [x10, #0x30]\n" + ".inst 0x6e40ece9 // bfmmla v9.4s, v7.8h, v0.8h\n" + ".inst 0x6e40ec71 // bfmmla v17.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec59 // bfmmla v25.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x40]\n" + ".inst 0x6e41eced // bfmmla v13.4s, v7.8h, v1.8h\n" + ".inst 0x6e41ec75 // bfmmla v21.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x10, #0x50]\n" + ".inst 0x6e40ecea // bfmmla v10.4s, v7.8h, v0.8h\n" + ".inst 0x6e40ec72 // bfmmla v18.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5a // bfmmla v26.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x60]\n" + ".inst 0x6e41ecee // bfmmla v14.4s, v7.8h, v1.8h\n" + ".inst 0x6e41ec76 // bfmmla v22.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5e // bfmmla v30.4s, v2.8h, v1.8h\n" "ldr q6, [x10, #0x70]\n" "add x10, x10, #0x80\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9b // bfmmla v27.4s, v4.8h, v7.8h\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9f // bfmmla v31.4s, v4.8h, v6.8h\n" + ".inst 0x6e40eceb // bfmmla v11.4s, v7.8h, v0.8h\n" + ".inst 0x6e40ec73 // bfmmla v19.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5b // bfmmla v27.4s, v2.8h, v0.8h\n" + ".inst 0x6e46ecef // bfmmla v15.4s, v7.8h, v6.8h\n" + ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ec5f // bfmmla v31.4s, v2.8h, v6.8h\n" "174:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -3088,16 +3087,16 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "201:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, 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, 202f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "ldr x21, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "ldr x21, [x20, #0x28]\n" "cbnz x28, 203f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -3109,11 +3108,11 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "b 203f\n" "202:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" - "add x21, x22, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" + "add x22, x23, x21, LSL #1\n" + "add x21, x22, x21, LSL #1\n" "203:" // Height 6: input setup done "cmp x27, #0x8\n" "blt 206f\n" @@ -3180,42 +3179,42 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr q2, [x25, #0x0]\n" "prfm pldl1keep, [x21, #0x80]\n" ".inst 0x6e46ec9f // bfmmla v31.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" + "ldr q0, [x10, #0x90]\n" "ldr q4, [x23, #0x0]\n" ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" ".inst 0x6e47ecb8 // bfmmla v24.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbc // bfmmla v28.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecb9 // bfmmla v25.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbd // bfmmla v29.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecba // bfmmla v26.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbe // bfmmla v30.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xf0]\n" + "ldr q6, [x10, #0xa0]\n" + ".inst 0x6e40ec2c // bfmmla v12.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec74 // bfmmla v20.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecbc // bfmmla v28.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xb0]\n" + ".inst 0x6e46ec29 // bfmmla v9.4s, v1.8h, v6.8h\n" + ".inst 0x6e46ec71 // bfmmla v17.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecb9 // bfmmla v25.4s, v5.8h, v6.8h\n" + "ldr q6, [x10, #0xc0]\n" + ".inst 0x6e40ec2d // bfmmla v13.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec75 // bfmmla v21.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecbd // bfmmla v29.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xd0]\n" + ".inst 0x6e46ec2a // bfmmla v10.4s, v1.8h, v6.8h\n" + ".inst 0x6e46ec72 // bfmmla v18.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecba // bfmmla v26.4s, v5.8h, v6.8h\n" + "ldr q6, [x10, #0xe0]\n" + ".inst 0x6e40ec2e // bfmmla v14.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec76 // bfmmla v22.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecbe // bfmmla v30.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec73 // bfmmla v19.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecbb // bfmmla v27.4s, v5.8h, v7.8h\n" + ".inst 0x6e46ec2b // bfmmla v11.4s, v1.8h, v6.8h\n" + ".inst 0x6e46ec73 // bfmmla v19.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecbb // bfmmla v27.4s, v5.8h, v6.8h\n" "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" + ".inst 0x6e40ec2f // bfmmla v15.4s, v1.8h, v0.8h\n" "ldr q1, [x26, #0x0]\n" - ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" + ".inst 0x6e40ec77 // bfmmla v23.4s, v3.8h, v0.8h\n" "ldr q3, [x24, #0x0]\n" - ".inst 0x6e46ecbf // bfmmla v31.4s, v5.8h, v6.8h\n" + ".inst 0x6e40ecbf // bfmmla v31.4s, v5.8h, v0.8h\n" "ldr q5, [x22, #0x0]\n" "ldr q6, [x21, #0x0]\n" "bge 204b\n" @@ -3271,35 +3270,35 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" ".inst 0x6e46ec9f // bfmmla v31.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x90]\n" + "ldr q2, [x10, #0x90]\n" ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" ".inst 0x6e47ecb8 // bfmmla v24.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xa0]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbc // bfmmla v28.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xb0]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecb9 // bfmmla v25.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xc0]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbd // bfmmla v29.4s, v5.8h, v6.8h\n" - "ldr q6, [x10, #0xd0]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecba // bfmmla v26.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0xe0]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - ".inst 0x6e46ecbe // bfmmla v30.4s, v5.8h, v6.8h\n" + "ldr q0, [x10, #0xa0]\n" + ".inst 0x6e42ec2c // bfmmla v12.4s, v1.8h, v2.8h\n" + ".inst 0x6e42ec74 // bfmmla v20.4s, v3.8h, v2.8h\n" + ".inst 0x6e42ecbc // bfmmla v28.4s, v5.8h, v2.8h\n" + "ldr q2, [x10, #0xb0]\n" + ".inst 0x6e40ec29 // bfmmla v9.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec71 // bfmmla v17.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecb9 // bfmmla v25.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xc0]\n" + ".inst 0x6e42ec2d // bfmmla v13.4s, v1.8h, v2.8h\n" + ".inst 0x6e42ec75 // bfmmla v21.4s, v3.8h, v2.8h\n" + ".inst 0x6e42ecbd // bfmmla v29.4s, v5.8h, v2.8h\n" + "ldr q2, [x10, #0xd0]\n" + ".inst 0x6e40ec2a // bfmmla v10.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec72 // bfmmla v18.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecba // bfmmla v26.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0xe0]\n" + ".inst 0x6e42ec2e // bfmmla v14.4s, v1.8h, v2.8h\n" + ".inst 0x6e42ec76 // bfmmla v22.4s, v3.8h, v2.8h\n" + ".inst 0x6e42ecbe // bfmmla v30.4s, v5.8h, v2.8h\n" "ldr q6, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" - ".inst 0x6e47ec73 // bfmmla v19.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecbb // bfmmla v27.4s, v5.8h, v7.8h\n" + ".inst 0x6e40ec2b // bfmmla v11.4s, v1.8h, v0.8h\n" + ".inst 0x6e40ec73 // bfmmla v19.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecbb // bfmmla v27.4s, v5.8h, v0.8h\n" ".inst 0x6e46ec2f // bfmmla v15.4s, v1.8h, v6.8h\n" ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" ".inst 0x6e46ecbf // bfmmla v31.4s, v5.8h, v6.8h\n" @@ -3309,49 +3308,49 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "blt 208f\n" "207:" // Height 6: Multiply loop: Odd block loop "ldr d1, [x26], #0x8\n" - "ldr d2, [x25], #0x8\n" - "trn1 v0.2d, v1.2d, v2.2d\n" + "ldr d0, [x25], #0x8\n" + "trn1 v4.2d, v1.2d, v0.2d\n" "sub x27, x27, #0x4\n" - "ldr d3, [x24], #0x8\n" - "ldr d4, [x23], #0x8\n" - "trn1 v2.2d, v3.2d, v4.2d\n" + "ldr d1, [x24], #0x8\n" + "ldr d0, [x23], #0x8\n" + "trn1 v3.2d, v1.2d, v0.2d\n" "cmp x27, #0x4\n" - "ldr d5, [x22], #0x8\n" - "ldr d7, [x21], #0x8\n" - "trn1 v4.2d, v5.2d, v7.2d\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec98 // bfmmla v24.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9c // bfmmla v28.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec99 // bfmmla v25.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9d // bfmmla v29.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec52 // bfmmla v18.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9a // bfmmla v26.4s, v4.8h, v6.8h\n" + "ldr d1, [x22], #0x8\n" + "ldr d0, [x21], #0x8\n" + "trn1 v2.2d, v1.2d, v0.2d\n" + "ldr q1, [x10, #0x0]\n" + "ldr q0, [x10, #0x10]\n" + ".inst 0x6e41ec88 // bfmmla v8.4s, v4.8h, v1.8h\n" + ".inst 0x6e41ec70 // bfmmla v16.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec58 // bfmmla v24.4s, v2.8h, v1.8h\n" + "ldr q1, [x10, #0x20]\n" + ".inst 0x6e40ec8c // bfmmla v12.4s, v4.8h, v0.8h\n" + ".inst 0x6e40ec74 // bfmmla v20.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5c // bfmmla v28.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x30]\n" + ".inst 0x6e41ec89 // bfmmla v9.4s, v4.8h, v1.8h\n" + ".inst 0x6e41ec71 // bfmmla v17.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec59 // bfmmla v25.4s, v2.8h, v1.8h\n" + "ldr q1, [x10, #0x40]\n" + ".inst 0x6e40ec8d // bfmmla v13.4s, v4.8h, v0.8h\n" + ".inst 0x6e40ec75 // bfmmla v21.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5d // bfmmla v29.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x50]\n" + ".inst 0x6e41ec8a // bfmmla v10.4s, v4.8h, v1.8h\n" + ".inst 0x6e41ec72 // bfmmla v18.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5a // bfmmla v26.4s, v2.8h, v1.8h\n" "ldr q6, [x10, #0x60]\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9e // bfmmla v30.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x70]\n" + ".inst 0x6e40ec8e // bfmmla v14.4s, v4.8h, v0.8h\n" + ".inst 0x6e40ec76 // bfmmla v22.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5e // bfmmla v30.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x70]\n" "add x10, x10, #0x80\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec53 // bfmmla v19.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9b // bfmmla v27.4s, v4.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9f // bfmmla v31.4s, v4.8h, v7.8h\n" + ".inst 0x6e46ec8b // bfmmla v11.4s, v4.8h, v6.8h\n" + ".inst 0x6e46ec73 // bfmmla v19.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ec5b // bfmmla v27.4s, v2.8h, v6.8h\n" + ".inst 0x6e40ec8f // bfmmla v15.4s, v4.8h, v0.8h\n" + ".inst 0x6e40ec77 // bfmmla v23.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5f // bfmmla v31.4s, v2.8h, v0.8h\n" "bge 207b\n" "208:" // Height 6: Multiply loop: Skip odd blocks "cbz x27, 211f\n" @@ -3378,42 +3377,42 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "ldr h5, [x22, #0x0]\n" "ldr h6, [x21, #0x0]\n" "210:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q7, [x10, #0x0]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "trn1 v4.2d, v5.2d, v6.2d\n" - "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec98 // bfmmla v24.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9c // bfmmla v28.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec99 // bfmmla v25.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x40]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9d // bfmmla v29.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x50]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9a // bfmmla v26.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x60]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9e // bfmmla v30.4s, v4.8h, v6.8h\n" + "ldr q0, [x10, #0x0]\n" + "trn1 v7.2d, v1.2d, v2.2d\n" + "trn1 v3.2d, v3.2d, v4.2d\n" + ".inst 0x6e40ece8 // bfmmla v8.4s, v7.8h, v0.8h\n" + "trn1 v2.2d, v5.2d, v6.2d\n" + "ldr q1, [x10, #0x10]\n" + ".inst 0x6e40ec70 // bfmmla v16.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec58 // bfmmla v24.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x20]\n" + ".inst 0x6e41ecec // bfmmla v12.4s, v7.8h, v1.8h\n" + ".inst 0x6e41ec74 // bfmmla v20.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5c // bfmmla v28.4s, v2.8h, v1.8h\n" + "ldr q1, [x10, #0x30]\n" + ".inst 0x6e40ece9 // bfmmla v9.4s, v7.8h, v0.8h\n" + ".inst 0x6e40ec71 // bfmmla v17.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec59 // bfmmla v25.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x40]\n" + ".inst 0x6e41eced // bfmmla v13.4s, v7.8h, v1.8h\n" + ".inst 0x6e41ec75 // bfmmla v21.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5d // bfmmla v29.4s, v2.8h, v1.8h\n" + "ldr q1, [x10, #0x50]\n" + ".inst 0x6e40ecea // bfmmla v10.4s, v7.8h, v0.8h\n" + ".inst 0x6e40ec72 // bfmmla v18.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5a // bfmmla v26.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x60]\n" + ".inst 0x6e41ecee // bfmmla v14.4s, v7.8h, v1.8h\n" + ".inst 0x6e41ec76 // bfmmla v22.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5e // bfmmla v30.4s, v2.8h, v1.8h\n" "ldr q6, [x10, #0x70]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e40eceb // bfmmla v11.4s, v7.8h, v0.8h\n" "add x10, x10, #0x80\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9b // bfmmla v27.4s, v4.8h, v7.8h\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9f // bfmmla v31.4s, v4.8h, v6.8h\n" + ".inst 0x6e40ec73 // bfmmla v19.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5b // bfmmla v27.4s, v2.8h, v0.8h\n" + ".inst 0x6e46ecef // bfmmla v15.4s, v7.8h, v6.8h\n" + ".inst 0x6e46ec77 // bfmmla v23.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ec5f // bfmmla v31.4s, v2.8h, v6.8h\n" "211:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -3678,7 +3677,6 @@ void a64_hybrid_bf16fp32_mmla_6x16 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "224:" // 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", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" |