diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp | 942 |
1 files changed, 470 insertions, 472 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp index 4993777d62..5693c3f397 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32bf16fp32_mmla_6x16/generic.cpp @@ -93,7 +93,6 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( break; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x6\n" "bge 176f\n" @@ -211,11 +210,11 @@ void a64_hybrid_fp32bf16fp32_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 #2\n" @@ -233,23 +232,23 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "19:" // Height 1: Multiply loop: Main loop head ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" + "ldr q18, [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 q17, [x10, #0x30]\n" + ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" "sub x27, x27, #0x4\n" "cmp x27, #0x8\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" + ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n" "add x10, x10, #0x80\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n" "ldr q7, [x10, #0x10]\n" "prfm pldl1keep, [x26, #0x80]\n" "ld1 { v0.4s }, [x26], #0x10\n" @@ -257,20 +256,20 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "20:" // Height 1: Multiply loop: Single iteration only ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" + "ldr q18, [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 q17, [x10, #0x30]\n" + ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" "sub x27, x27, #0x4\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n" + ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n" "prfm pldl1keep, [x26, #0x80]\n" "add x10, x10, #0x80\n" "21:" // Height 1: Multiply loop: Main loop skip @@ -284,23 +283,23 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "22:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 "ldr s0, [x26, #0x0]\n" "23:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" + "ldr q18, [x10, #0x0]\n" + "ldr q17, [x10, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\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" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e52ec08 // bfmmla v8.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e51ec0c // bfmmla v12.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" + ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n" + ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n" "add x10, x10, #0x80\n" "24:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -314,17 +313,17 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "uzp1 v11.2d, v11.2d, v15.2d\n" "tbz %x[flags], #1, 25f\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" "25:" // Height 1: No activation "cmp x11, #0x10\n" "bge 34f\n" @@ -515,12 +514,12 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "51:" // 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, 52f\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, 53f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -528,7 +527,7 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "b 53f\n" "52:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" "53:" // Height 2: input setup done "cmp x27, #0x4\n" "blt 56f\n" @@ -542,23 +541,23 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" + "ldr q18, [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 q17, [x10, #0x30]\n" + ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" "sub x27, x27, #0x4\n" "cmp x27, #0x8\n" "add x10, x10, #0x80\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" + ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n" "ldr q7, [x10, #0x10]\n" "prfm pldl1keep, [x26, #0x80]\n" "ld1 { v0.4s }, [x26], #0x10\n" @@ -569,20 +568,20 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" + "ldr q18, [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 q17, [x10, #0x30]\n" + ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" "sub x27, x27, #0x4\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n" + ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" "add x10, x10, #0x80\n" @@ -600,24 +599,24 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "ldr s0, [x26, #0x0]\n" "ldr s1, [x25, #0x0]\n" "58:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" + "ldr q18, [x10, #0x0]\n" + "ldr q17, [x10, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\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" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e52ec08 // bfmmla v8.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e51ec0c // bfmmla v12.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ec09 // bfmmla v9.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x40]\n" + ".inst 0x6e51ec0d // bfmmla v13.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x6e52ec0a // bfmmla v10.4s, v0.8h, v18.8h\n" + "ldr q18, [x10, #0x60]\n" + ".inst 0x6e51ec0e // bfmmla v14.4s, v0.8h, v17.8h\n" + "ldr q17, [x10, #0x70]\n" + ".inst 0x6e52ec0b // bfmmla v11.4s, v0.8h, v18.8h\n" + ".inst 0x6e51ec0f // bfmmla v15.4s, v0.8h, v17.8h\n" "add x10, x10, #0x80\n" "59:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -638,25 +637,25 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "uzp2 v11.2d, v11.2d, v15.2d\n" "tbz %x[flags], #1, 60f\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 v6.4s, v6.4s, 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 v6.4s, v6.4s, 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 v6.4s, v6.4s, 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 v6.4s, v6.4s, 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" "60:" // Height 2: No activation "cmp x11, #0x10\n" "bge 69f\n" @@ -912,13 +911,13 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "86:" // 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, 87f\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, 88f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -927,8 +926,8 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "b 88f\n" "87:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" "88:" // Height 3: input setup done "cmp x27, #0x4\n" "blt 91f\n" @@ -946,34 +945,34 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "sub x27, x27, #0x4\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" + "ldr q26, [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 q25, [x10, #0x30]\n" + ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n" "cmp x27, #0x8\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 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n" "prfm pldl1keep, [x26, #0x80]\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 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n" "prfm pldl1keep, [x25, #0x80]\n" "ld1 { v1.4s }, [x25], #0x10\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 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" "add x10, x10, #0x80\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec53 // bfmmla v19.4s, v2.8h, v6.8h\n" + ".inst 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n" "ld1 { v0.4s }, [x26], #0x10\n" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" + ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n" "ld1 { v2.4s }, [x24], #0x10\n" "ldr q7, [x10, #0x10]\n" "bge 89b\n" @@ -984,30 +983,30 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "sub x27, x27, #0x4\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" + "ldr q26, [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 q25, [x10, #0x30]\n" + ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n" "prfm pldl1keep, [x26, #0x80]\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 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n" "prfm pldl1keep, [x25, #0x80]\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 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n" "prfm pldl1keep, [x24, #0x80]\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 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, 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 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n" + ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n" "91:" // Height 3: Multiply loop: Main loop skip "cbz x27, 94f\n" "cbz x27, 94f\n" @@ -1025,34 +1024,34 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "ldr s1, [x25, #0x0]\n" "ldr s2, [x24, #0x0]\n" "93:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" + "ldr q26, [x10, #0x0]\n" + "ldr q25, [x10, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\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" + ".inst 0x6e5aec08 // bfmmla v8.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec50 // bfmmla v16.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e59ec0c // bfmmla v12.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec54 // bfmmla v20.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, 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 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n" + ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n" "94:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1078,33 +1077,33 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "uzp1 v19.2d, v19.2d, v23.2d\n" "tbz %x[flags], #1, 95f\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 v6.4s, v6.4s, 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 v6.4s, v6.4s, 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 v6.4s, v6.4s, 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 v6.4s, v6.4s, 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" "95:" // Height 3: No activation "cmp x11, #0x10\n" "bge 104f\n" @@ -1401,14 +1400,14 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "121:" // 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, 122f\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, 123f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1418,9 +1417,9 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "b 123f\n" "122:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "123:" // Height 4: input setup done "cmp x27, #0x4\n" "blt 126f\n" @@ -1442,34 +1441,34 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" + "ldr q26, [x10, #0x20]\n" ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" "prfm pldl1keep, [x25, #0x80]\n" "ld1 { v1.4s }, [x25], #0x10\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" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n" "prfm pldl1keep, [x24, #0x80]\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 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n" "prfm pldl1keep, [x23, #0x80]\n" "ld1 { v3.4s }, [x23], #0x10\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 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, 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 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n" "ld1 { v0.4s }, [x26], #0x10\n" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" + ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n" "ld1 { v2.4s }, [x24], #0x10\n" "ldr q7, [x10, #0x10]\n" "bge 124b\n" @@ -1483,29 +1482,29 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" + "ldr q26, [x10, #0x20]\n" ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" "prfm pldl1keep, [x24, #0x80]\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 q25, [x10, #0x30]\n" + ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n" "prfm pldl1keep, [x23, #0x80]\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" + ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, 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 0x6e5aec0b // bfmmla v11.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n" + ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n" "126:" // Height 4: Multiply loop: Main loop skip "cbz x27, 129f\n" "cbz x27, 129f\n" @@ -1526,35 +1525,35 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "ldr s2, [x24, #0x0]\n" "ldr s3, [x23, #0x0]\n" "128:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" + "ldr q26, [x10, #0x0]\n" + "ldr q25, [x10, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\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" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aec08 // bfmmla v8.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec50 // bfmmla v16.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e59ec0c // bfmmla v12.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec54 // bfmmla v20.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aec09 // bfmmla v9.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec51 // bfmmla v17.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x40]\n" + ".inst 0x6e59ec0d // bfmmla v13.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec55 // bfmmla v21.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x50]\n" + ".inst 0x6e5aec0a // bfmmla v10.4s, v0.8h, v26.8h\n" + ".inst 0x6e5aec52 // bfmmla v18.4s, v2.8h, v26.8h\n" + "ldr q26, [x10, #0x60]\n" + ".inst 0x6e59ec0e // bfmmla v14.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec56 // bfmmla v22.4s, v2.8h, v25.8h\n" + "ldr q25, [x10, #0x70]\n" + ".inst 0x6e5aec0b // bfmmla v11.4s, v0.8h, 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 0x6e5aec53 // bfmmla v19.4s, v2.8h, v26.8h\n" + ".inst 0x6e59ec0f // bfmmla v15.4s, v0.8h, v25.8h\n" + ".inst 0x6e59ec57 // bfmmla v23.4s, v2.8h, v25.8h\n" "129:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1586,41 +1585,41 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "uzp2 v19.2d, v19.2d, v23.2d\n" "tbz %x[flags], #1, 130f\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 v6.4s, v6.4s, 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 v6.4s, v6.4s, 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 v6.4s, v6.4s, 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 v6.4s, v6.4s, 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" "130:" // Height 4: No activation "cmp x11, #0x10\n" "bge 139f\n" @@ -1982,15 +1981,15 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "156:" // 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, 157f\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, 158f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -2001,10 +2000,10 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "b 158f\n" "157:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "158:" // Height 5: input setup done "cmp x27, #0x4\n" "blt 161f\n" @@ -2029,43 +2028,43 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\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" + "ldr q3, [x10, #0x20]\n" ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" "prfm pldl1keep, [x25, #0x80]\n" "ld1 { v1.4s }, [x25], #0x10\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" + "ldr q5, [x10, #0x30]\n" + ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec99 // bfmmla v25.4s, v4.8h, v6.8h\n" + ".inst 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n" "ldr q6, [x10, #0x40]\n" "prfm pldl1keep, [x23, #0x80]\n" "ld1 { v3.4s }, [x23], #0x10\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x6e47ec9d // bfmmla v29.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x50]\n" + ".inst 0x6e45ec9d // bfmmla v29.4s, v4.8h, v5.8h\n" + "ldr q5, [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 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 0x6e45ec0e // bfmmla v14.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec56 // bfmmla v22.4s, v2.8h, v5.8h\n" + ".inst 0x6e45ec9e // bfmmla v30.4s, v4.8h, v5.8h\n" + "ldr q5, [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" "ldr q6, [x10, #0x0]\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e45ec0f // bfmmla v15.4s, v0.8h, v5.8h\n" "ld1 { v0.4s }, [x26], #0x10\n" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" + ".inst 0x6e45ec57 // bfmmla v23.4s, v2.8h, v5.8h\n" "ld1 { v2.4s }, [x24], #0x10\n" - ".inst 0x6e47ec9f // bfmmla v31.4s, v4.8h, v7.8h\n" + ".inst 0x6e45ec9f // bfmmla v31.4s, v4.8h, v5.8h\n" "ld1 { v4.4s }, [x22], #0x10\n" "ldr q7, [x10, #0x10]\n" "bge 159b\n" @@ -2081,37 +2080,37 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\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" + "ldr q3, [x10, #0x20]\n" ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" "prfm pldl1keep, [x24, #0x80]\n" "prfm pldl1keep, [x23, #0x80]\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" + "ldr q1, [x10, #0x30]\n" + ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n" "prfm pldl1keep, [x22, #0x80]\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 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 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n" + "ldr q3, [x10, #0x40]\n" + ".inst 0x6e41ec0d // bfmmla v13.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec55 // bfmmla v21.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9d // bfmmla v29.4s, v4.8h, v1.8h\n" + "ldr q1, [x10, #0x50]\n" + ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec52 // bfmmla v18.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec9a // bfmmla v26.4s, v4.8h, v3.8h\n" + "ldr q3, [x10, #0x60]\n" + ".inst 0x6e41ec0e // bfmmla v14.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec56 // bfmmla v22.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9e // bfmmla v30.4s, v4.8h, v1.8h\n" + "ldr q1, [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 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec53 // bfmmla v19.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec9b // bfmmla v27.4s, v4.8h, v3.8h\n" + ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec57 // bfmmla v23.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9f // bfmmla v31.4s, v4.8h, v1.8h\n" "161:" // Height 5: Multiply loop: Main loop skip "cbz x27, 164f\n" "cbz x27, 164f\n" @@ -2136,7 +2135,7 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "ldr s4, [x22, #0x0]\n" "163:" // Height 5: Multiply loop: Ragged operand read: Done "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" + "ldr q5, [x10, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" @@ -2145,34 +2144,34 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\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 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" + "ldr q3, [x10, #0x20]\n" + ".inst 0x6e45ec0c // bfmmla v12.4s, v0.8h, v5.8h\n" + ".inst 0x6e45ec54 // bfmmla v20.4s, v2.8h, v5.8h\n" + ".inst 0x6e45ec9c // bfmmla v28.4s, v4.8h, v5.8h\n" + "ldr q1, [x10, #0x30]\n" + ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n" + "ldr q3, [x10, #0x40]\n" + ".inst 0x6e41ec0d // bfmmla v13.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec55 // bfmmla v21.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9d // bfmmla v29.4s, v4.8h, v1.8h\n" + "ldr q1, [x10, #0x50]\n" + ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec52 // bfmmla v18.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec9a // bfmmla v26.4s, v4.8h, v3.8h\n" + "ldr q3, [x10, #0x60]\n" + ".inst 0x6e41ec0e // bfmmla v14.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec56 // bfmmla v22.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9e // bfmmla v30.4s, v4.8h, v1.8h\n" + "ldr q1, [x10, #0x70]\n" + ".inst 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\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 0x6e43ec53 // bfmmla v19.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec9b // bfmmla v27.4s, v4.8h, v3.8h\n" + ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec57 // bfmmla v23.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9f // bfmmla v31.4s, v4.8h, v1.8h\n" "164:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -2658,16 +2657,16 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "191:" // 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, 192f\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, 193f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -2679,11 +2678,11 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "b 193f\n" "192:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "193:" // Height 6: input setup done "cmp x27, #0x4\n" "blt 196f\n" @@ -2716,7 +2715,7 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( ".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" + "ldr q5, [x10, #0x30]\n" "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n" @@ -2724,10 +2723,10 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "ld1 { v3.4s }, [x23], #0x10\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 0x6e45ec0d // bfmmla v13.4s, v0.8h, v5.8h\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9d // bfmmla v29.4s, v4.8h, v7.8h\n" + ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n" + ".inst 0x6e45ec9d // bfmmla v29.4s, v4.8h, v5.8h\n" "ldr q7, [x10, #0x50]\n" "prfm pldl1keep, [x21, #0x80]\n" "ld1 { v5.4s }, [x21], #0x10\n" @@ -2766,37 +2765,37 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x6e46ec98 // bfmmla v24.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x20]\n" + "ldr q3, [x10, #0x20]\n" ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" "prfm pldl1keep, [x23, #0x80]\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" + "ldr q1, [x10, #0x30]\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n" + ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n" "prfm pldl1keep, [x21, #0x80]\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 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 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n" + "ldr q3, [x10, #0x40]\n" + ".inst 0x6e41ec0d // bfmmla v13.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec55 // bfmmla v21.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9d // bfmmla v29.4s, v4.8h, v1.8h\n" + "ldr q1, [x10, #0x50]\n" + ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec52 // bfmmla v18.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec9a // bfmmla v26.4s, v4.8h, v3.8h\n" + "ldr q3, [x10, #0x60]\n" + ".inst 0x6e41ec0e // bfmmla v14.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec56 // bfmmla v22.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9e // bfmmla v30.4s, v4.8h, v1.8h\n" + "ldr q1, [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 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec53 // bfmmla v19.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec9b // bfmmla v27.4s, v4.8h, v3.8h\n" + ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec57 // bfmmla v23.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9f // bfmmla v31.4s, v4.8h, v1.8h\n" "196:" // Height 6: Multiply loop: Main loop skip "cbz x27, 199f\n" "cbz x27, 199f\n" @@ -2823,45 +2822,45 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "ldr s4, [x22, #0x0]\n" "ldr s5, [x21, #0x0]\n" "198:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" + "ldr q7, [x10, #0x0]\n" + "ldr q6, [x10, #0x10]\n" ".inst 0x0ea16800 // bfcvtn v0.4h, v0.4s\n" ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\n" ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" + ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n" ".inst 0x4ea168a4 // bfcvtn2 v4.8h, v5.4s\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 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 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" + ".inst 0x6e47ec98 // bfmmla v24.4s, v4.8h, v7.8h\n" + "ldr q3, [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 q1, [x10, #0x30]\n" + ".inst 0x6e43ec09 // bfmmla v9.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec51 // bfmmla v17.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec99 // bfmmla v25.4s, v4.8h, v3.8h\n" + "ldr q3, [x10, #0x40]\n" + ".inst 0x6e41ec0d // bfmmla v13.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec55 // bfmmla v21.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9d // bfmmla v29.4s, v4.8h, v1.8h\n" + "ldr q1, [x10, #0x50]\n" + ".inst 0x6e43ec0a // bfmmla v10.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec52 // bfmmla v18.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec9a // bfmmla v26.4s, v4.8h, v3.8h\n" + "ldr q3, [x10, #0x60]\n" + ".inst 0x6e41ec0e // bfmmla v14.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec56 // bfmmla v22.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9e // bfmmla v30.4s, v4.8h, v1.8h\n" + "ldr q1, [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 0x6e43ec0b // bfmmla v11.4s, v0.8h, v3.8h\n" + ".inst 0x6e43ec53 // bfmmla v19.4s, v2.8h, v3.8h\n" + ".inst 0x6e43ec9b // bfmmla v27.4s, v4.8h, v3.8h\n" + ".inst 0x6e41ec0f // bfmmla v15.4s, v0.8h, v1.8h\n" + ".inst 0x6e41ec57 // bfmmla v23.4s, v2.8h, v1.8h\n" + ".inst 0x6e41ec9f // bfmmla v31.4s, v4.8h, v1.8h\n" "199:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -3126,7 +3125,6 @@ void a64_hybrid_fp32bf16fp32_mmla_6x16 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "212:" // 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" |