diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp | 1876 |
1 files changed, 938 insertions, 938 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp index 74791f8d30..377daddae9 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_bf16fp32_mmla_6x16/generic.cpp @@ -231,11 +231,11 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "17:" // 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, 18f\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, 19f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -251,41 +251,41 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr q6, [x12, #0x10]\n" "blt 21f\n" "20:" // 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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x10]\n" - "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - "ldr q7, [x12, #0x20]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - "ldr q6, [x12, #0x30]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - "ldr q7, [x11, #0x20]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - "ldr q6, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - "ldr q6, [x9, #0x30]\n" + "trn1 v20.2d, v1.2d, v21.2d\n" + ".inst 0x6e47ee88 // bfmmla v8.4s, v20.8h, v7.8h\n" + "ldr q17, [x11, #0x0]\n" + ".inst 0x6e46ee8c // bfmmla v12.4s, v20.8h, v6.8h\n" + "ldr q19, [x11, #0x10]\n" + ".inst 0x6e51ee89 // bfmmla v9.4s, v20.8h, v17.8h\n" + "ldr q18, [x10, #0x0]\n" + ".inst 0x6e53ee8d // bfmmla v13.4s, v20.8h, v19.8h\n" + "ldr q17, [x10, #0x10]\n" + ".inst 0x6e52ee8a // bfmmla v10.4s, v20.8h, v18.8h\n" + "ldr q18, [x9, #0x0]\n" + ".inst 0x6e51ee8e // bfmmla v14.4s, v20.8h, v17.8h\n" + "ldr q17, [x9, #0x10]\n" + "trn2 v1.2d, v1.2d, v21.2d\n" + ".inst 0x6e52ee8b // bfmmla v11.4s, v20.8h, v18.8h\n" + "ldr q18, [x12, #0x20]\n" + ".inst 0x6e51ee8f // bfmmla v15.4s, v20.8h, v17.8h\n" + "ldr q17, [x12, #0x30]\n" + ".inst 0x6e52ec28 // bfmmla v8.4s, v1.8h, v18.8h\n" + "ldr q18, [x11, #0x20]\n" + ".inst 0x6e51ec2c // bfmmla v12.4s, v1.8h, v17.8h\n" + "ldr q17, [x11, #0x30]\n" + ".inst 0x6e52ec29 // bfmmla v9.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e51ec2d // bfmmla v13.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ec2a // bfmmla v10.4s, v1.8h, v18.8h\n" + "ldr q18, [x9, #0x20]\n" + ".inst 0x6e51ec2e // bfmmla v14.4s, v1.8h, v17.8h\n" + "ldr q17, [x9, #0x30]\n" "sub x27, x27, #0x8\n" "cmp x27, #0x10\n" "add x26, x26, #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 x12, x12, #0x40\n" "ldr q7, [x12, #0x0]\n" @@ -295,39 +295,39 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "add x9, x9, #0x40\n" "bge 20b\n" "21:" // 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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x10]\n" - "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - "ldr q7, [x12, #0x20]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - "ldr q6, [x12, #0x30]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - "ldr q7, [x11, #0x20]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - "ldr q6, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - "ldr q6, [x9, #0x30]\n" + "trn1 v19.2d, v1.2d, v20.2d\n" + ".inst 0x6e47ee68 // bfmmla v8.4s, v19.8h, v7.8h\n" + "ldr q17, [x11, #0x0]\n" + ".inst 0x6e46ee6c // bfmmla v12.4s, v19.8h, v6.8h\n" + "ldr q18, [x11, #0x10]\n" + ".inst 0x6e51ee69 // bfmmla v9.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x0]\n" + ".inst 0x6e52ee6d // bfmmla v13.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x10]\n" + ".inst 0x6e51ee6a // bfmmla v10.4s, v19.8h, v17.8h\n" + "ldr q17, [x9, #0x0]\n" + ".inst 0x6e52ee6e // bfmmla v14.4s, v19.8h, v18.8h\n" + "ldr q24, [x9, #0x10]\n" + "trn2 v1.2d, v1.2d, v20.2d\n" + ".inst 0x6e51ee6b // bfmmla v11.4s, v19.8h, v17.8h\n" + "ldr q18, [x12, #0x20]\n" + ".inst 0x6e58ee6f // bfmmla v15.4s, v19.8h, v24.8h\n" + "ldr q17, [x12, #0x30]\n" + ".inst 0x6e52ec28 // bfmmla v8.4s, v1.8h, v18.8h\n" + "ldr q19, [x11, #0x20]\n" + ".inst 0x6e51ec2c // bfmmla v12.4s, v1.8h, v17.8h\n" + "ldr q17, [x11, #0x30]\n" + ".inst 0x6e53ec29 // bfmmla v9.4s, v1.8h, v19.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e51ec2d // bfmmla v13.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ec2a // bfmmla v10.4s, v1.8h, v18.8h\n" + "ldr q18, [x9, #0x20]\n" + ".inst 0x6e51ec2e // bfmmla v14.4s, v1.8h, v17.8h\n" + "ldr q17, [x9, #0x30]\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" "add x26, x26, #0x10\n" "add x12, x12, #0x40\n" "add x11, x11, #0x40\n" @@ -338,26 +338,26 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "cmp x27, #0x4\n" "blt 24f\n" "23:" // Height 1: Multiply loop: Odd block loop - "ldr d1, [x26], #0x8\n" - "ldr q6, [x12, #0x0]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - "ldr q7, [x12, #0x10]\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q6, [x11, #0x0]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - "ldr q7, [x11, #0x10]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x0]\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x0]\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" + "ldr d19, [x26], #0x8\n" + "ldr q18, [x12, #0x0]\n" + "trn1 v19.2d, v19.2d, v17.2d\n" + "ldr q17, [x12, #0x10]\n" + ".inst 0x6e52ee68 // bfmmla v8.4s, v19.8h, v18.8h\n" + "ldr q18, [x11, #0x0]\n" + ".inst 0x6e51ee6c // bfmmla v12.4s, v19.8h, v17.8h\n" + "ldr q17, [x11, #0x10]\n" + ".inst 0x6e52ee69 // bfmmla v9.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x0]\n" + ".inst 0x6e51ee6d // bfmmla v13.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x10]\n" + ".inst 0x6e52ee6a // bfmmla v10.4s, v19.8h, v18.8h\n" + "ldr q18, [x9, #0x0]\n" + ".inst 0x6e51ee6e // bfmmla v14.4s, v19.8h, v17.8h\n" + "ldr q17, [x9, #0x10]\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 x12, x12, #0x20\n" "add x11, x11, #0x20\n" "add x10, x10, #0x20\n" @@ -373,23 +373,23 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "25:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 "ldr h1, [x26, #0x0]\n" "26:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q7, [x12, #0x0]\n" - "ldr q6, [x12, #0x10]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "ldr q7, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + "ldr q20, [x12, #0x0]\n" + "ldr q18, [x12, #0x10]\n" + "trn1 v19.2d, v1.2d, v17.2d\n" + ".inst 0x6e54ee68 // bfmmla v8.4s, v19.8h, v20.8h\n" + "ldr q17, [x11, #0x0]\n" + ".inst 0x6e52ee6c // bfmmla v12.4s, v19.8h, v18.8h\n" + "ldr q18, [x11, #0x10]\n" + ".inst 0x6e51ee69 // bfmmla v9.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x0]\n" + ".inst 0x6e52ee6d // bfmmla v13.4s, v19.8h, v18.8h\n" "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x10]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + ".inst 0x6e51ee6a // bfmmla v10.4s, v19.8h, v17.8h\n" + "ldr q18, [x9, #0x0]\n" + ".inst 0x6e46ee6e // bfmmla v14.4s, v19.8h, v6.8h\n" + "ldr q17, [x9, #0x10]\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" "add x12, x12, #0x20\n" "add x11, x11, #0x20\n" "add x10, x10, #0x20\n" @@ -405,17 +405,17 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "uzp1 v11.2d, v11.2d, v15.2d\n" "tbz %x[flags], #1, 28f\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" "28:" // Height 1: No activation "cmp x14, #0x10\n" "bge 37f\n" @@ -624,12 +624,12 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "55:" // 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, 56f\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, 57f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -637,7 +637,7 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "b 57f\n" "56:" // 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" "57:" // Height 2: input setup done "cmp x27, #0x8\n" "blt 60f\n" @@ -648,45 +648,45 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr q6, [x12, #0x10]\n" "blt 59f\n" "58:" // 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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x10]\n" + "trn1 v19.2d, v1.2d, v2.2d\n" + ".inst 0x6e47ee68 // bfmmla v8.4s, v19.8h, v7.8h\n" + "ldr q18, [x11, #0x0]\n" + ".inst 0x6e46ee6c // bfmmla v12.4s, v19.8h, v6.8h\n" + "ldr q17, [x11, #0x10]\n" + ".inst 0x6e52ee69 // bfmmla v9.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x0]\n" + ".inst 0x6e51ee6d // bfmmla v13.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x10]\n" + ".inst 0x6e52ee6a // bfmmla v10.4s, v19.8h, v18.8h\n" + "ldr q18, [x9, #0x0]\n" + ".inst 0x6e51ee6e // bfmmla v14.4s, v19.8h, v17.8h\n" + "ldr q17, [x9, #0x10]\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - "ldr q7, [x12, #0x20]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - "ldr q6, [x12, #0x30]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - "ldr q7, [x11, #0x20]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - "ldr q6, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - "ldr q6, [x9, #0x30]\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + "ldr q18, [x12, #0x20]\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" + "ldr q17, [x12, #0x30]\n" + ".inst 0x6e52ec28 // bfmmla v8.4s, v1.8h, v18.8h\n" + "ldr q18, [x11, #0x20]\n" + ".inst 0x6e51ec2c // bfmmla v12.4s, v1.8h, v17.8h\n" + "ldr q17, [x11, #0x30]\n" + ".inst 0x6e52ec29 // bfmmla v9.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e51ec2d // bfmmla v13.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ec2a // bfmmla v10.4s, v1.8h, v18.8h\n" + "ldr q18, [x9, #0x20]\n" + ".inst 0x6e51ec2e // bfmmla v14.4s, v1.8h, v17.8h\n" + "ldr q17, [x9, #0x30]\n" "sub x27, x27, #0x8\n" "cmp x27, #0x10\n" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "ldr q2, [x25, #0x0]\n" - ".inst 0x6e47ec2b // bfmmla v11.4s, v1.8h, v7.8h\n" + ".inst 0x6e52ec2b // bfmmla v11.4s, v1.8h, v18.8h\n" "add x12, x12, #0x40\n" "ldr q7, [x12, #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, [x12, #0x10]\n" "add x11, x11, #0x40\n" @@ -694,39 +694,39 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "add x9, x9, #0x40\n" "bge 58b\n" "59:" // 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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x10]\n" + "trn1 v19.2d, v1.2d, v2.2d\n" + ".inst 0x6e47ee68 // bfmmla v8.4s, v19.8h, v7.8h\n" + "ldr q18, [x11, #0x0]\n" + ".inst 0x6e46ee6c // bfmmla v12.4s, v19.8h, v6.8h\n" + "ldr q17, [x11, #0x10]\n" + ".inst 0x6e52ee69 // bfmmla v9.4s, v19.8h, v18.8h\n" + "ldr q18, [x10, #0x0]\n" + ".inst 0x6e51ee6d // bfmmla v13.4s, v19.8h, v17.8h\n" + "ldr q17, [x10, #0x10]\n" + ".inst 0x6e52ee6a // bfmmla v10.4s, v19.8h, v18.8h\n" + "ldr q18, [x9, #0x0]\n" + ".inst 0x6e51ee6e // bfmmla v14.4s, v19.8h, v17.8h\n" + "ldr q17, [x9, #0x10]\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - "ldr q7, [x12, #0x20]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - "ldr q6, [x12, #0x30]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" - "ldr q7, [x11, #0x20]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - "ldr q6, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - "ldr q6, [x9, #0x30]\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + "ldr q18, [x12, #0x20]\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" + "ldr q17, [x12, #0x30]\n" + ".inst 0x6e52ec28 // bfmmla v8.4s, v1.8h, v18.8h\n" + "ldr q18, [x11, #0x20]\n" + ".inst 0x6e51ec2c // bfmmla v12.4s, v1.8h, v17.8h\n" + "ldr q17, [x11, #0x30]\n" + ".inst 0x6e52ec29 // bfmmla v9.4s, v1.8h, v18.8h\n" + "ldr q18, [x10, #0x20]\n" + ".inst 0x6e51ec2d // bfmmla v13.4s, v1.8h, v17.8h\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x6e52ec2a // bfmmla v10.4s, v1.8h, v18.8h\n" + "ldr q18, [x9, #0x20]\n" + ".inst 0x6e51ec2e // bfmmla v14.4s, v1.8h, v17.8h\n" + "ldr q17, [x9, #0x30]\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" "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "add x12, x12, #0x40\n" @@ -738,27 +738,27 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "cmp x27, #0x4\n" "blt 62f\n" "61:" // 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, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - "ldr q6, [x11, #0x0]\n" - "ldr q7, [x11, #0x10]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" - "ldr q6, [x9, #0x0]\n" - "ldr q7, [x9, #0x10]\n" + "ldr q18, [x12, #0x0]\n" + "ldr q17, [x12, #0x10]\n" + ".inst 0x6e52ee68 // bfmmla v8.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6c // bfmmla v12.4s, v19.8h, v17.8h\n" + "ldr q26, [x11, #0x0]\n" + "ldr q6, [x11, #0x10]\n" + ".inst 0x6e5aee69 // bfmmla v9.4s, v19.8h, v26.8h\n" + ".inst 0x6e46ee6d // bfmmla v13.4s, v19.8h, v6.8h\n" + "ldr q18, [x10, #0x0]\n" + "ldr q17, [x10, #0x10]\n" + ".inst 0x6e52ee6a // bfmmla v10.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6e // bfmmla v14.4s, v19.8h, v17.8h\n" + "ldr q18, [x9, #0x0]\n" + "ldr q17, [x9, #0x10]\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 x12, x12, #0x20\n" "add x11, x11, #0x20\n" "add x10, x10, #0x20\n" @@ -777,23 +777,23 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr h1, [x26, #0x0]\n" "ldr h2, [x25, #0x0]\n" "64:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q7, [x12, #0x0]\n" - "ldr q6, [x12, #0x10]\n" - "trn1 v0.2d, v1.2d, v2.2d\n" - ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" - "ldr q7, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - "ldr q6, [x9, #0x10]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + "ldr q18, [x12, #0x0]\n" + "ldr q17, [x12, #0x10]\n" + "trn1 v19.2d, v1.2d, v2.2d\n" + ".inst 0x6e52ee68 // bfmmla v8.4s, v19.8h, v18.8h\n" + "ldr q18, [x11, #0x0]\n" + ".inst 0x6e51ee6c // bfmmla v12.4s, v19.8h, v17.8h\n" + "ldr q17, [x11, #0x10]\n" + ".inst 0x6e52ee69 // bfmmla v9.4s, v19.8h, v18.8h\n" + "ldr q3, [x10, #0x0]\n" + ".inst 0x6e51ee6d // bfmmla v13.4s, v19.8h, v17.8h\n" + "ldr q27, [x10, #0x10]\n" + ".inst 0x6e43ee6a // bfmmla v10.4s, v19.8h, v3.8h\n" + "ldr q18, [x9, #0x0]\n" + ".inst 0x6e5bee6e // bfmmla v14.4s, v19.8h, v27.8h\n" + "ldr q17, [x9, #0x10]\n" + ".inst 0x6e52ee6b // bfmmla v11.4s, v19.8h, v18.8h\n" + ".inst 0x6e51ee6f // bfmmla v15.4s, v19.8h, v17.8h\n" "add x12, x12, #0x20\n" "add x11, x11, #0x20\n" "add x10, x10, #0x20\n" @@ -815,25 +815,25 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "uzp2 v11.2d, v11.2d, v15.2d\n" "tbz %x[flags], #1, 66f\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" "66:" // Height 2: No activation "cmp x14, #0x10\n" "bge 75f\n" @@ -1107,13 +1107,13 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "93:" // 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, 94f\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, 95f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -1122,8 +1122,8 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "b 95f\n" "94:" // 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" "95:" // Height 3: input setup done "cmp x27, #0x8\n" "blt 98f\n" @@ -1135,170 +1135,170 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr q6, [x12, #0x10]\n" "blt 97f\n" "96:" // 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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\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, #0x0]\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, [x11, #0x0]\n" + ".inst 0x6e46ef8c // bfmmla v12.4s, v28.8h, v6.8h\n" + ".inst 0x6e46ef74 // bfmmla v20.4s, v27.8h, v6.8h\n" + "ldr q25, [x11, #0x10]\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, #0x0]\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, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x10]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" "cmp x27, #0x10\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x9, #0x0]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" "add x26, x26, #0x10\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - "ldr q6, [x9, #0x10]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x9, #0x10]\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" "add x25, x25, #0x10\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - "ldr q7, [x12, #0x20]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + "ldr q26, [x12, #0x20]\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" "add x24, x24, #0x10\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x12, #0x30]\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" + "ldr q25, [x12, #0x30]\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, [x11, #0x20]\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, [x11, #0x20]\n" + ".inst 0x6e59ec2c // bfmmla v12.4s, v1.8h, v25.8h\n" "add x12, x12, #0x40\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - "ldr q6, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" + ".inst 0x6e59ec74 // bfmmla v20.4s, v3.8h, v25.8h\n" + "ldr q25, [x11, #0x30]\n" + ".inst 0x6e5aec29 // bfmmla v9.4s, v1.8h, v26.8h\n" "add x11, x11, #0x40\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" + ".inst 0x6e5aec71 // bfmmla v17.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e59ec2d // bfmmla v13.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec75 // bfmmla v21.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aec2a // bfmmla v10.4s, v1.8h, v26.8h\n" "add x10, x10, #0x40\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - "ldr q6, [x9, #0x30]\n" + ".inst 0x6e5aec72 // bfmmla v18.4s, v3.8h, v26.8h\n" + "ldr q26, [x9, #0x20]\n" + ".inst 0x6e59ec2e // bfmmla v14.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec76 // bfmmla v22.4s, v3.8h, v25.8h\n" + "ldr q25, [x9, #0x30]\n" "add x9, x9, #0x40\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, [x12, #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, [x12, #0x10]\n" "bge 96b\n" "97:" // 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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\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, #0x0]\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, [x11, #0x0]\n" + ".inst 0x6e46ef8c // bfmmla v12.4s, v28.8h, v6.8h\n" + ".inst 0x6e46ef74 // bfmmla v20.4s, v27.8h, v6.8h\n" + "ldr q25, [x11, #0x10]\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, #0x0]\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, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x10]\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, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x9, #0x0]\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, [x9, #0x10]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x9, #0x10]\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, [x12, #0x20]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x12, #0x30]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + "ldr q26, [x12, #0x20]\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" + "ldr q25, [x12, #0x30]\n" + ".inst 0x6e5aec28 // bfmmla v8.4s, v1.8h, v26.8h\n" "add x12, x12, #0x40\n" - ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" - "ldr q7, [x11, #0x20]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - "ldr q6, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" + ".inst 0x6e5aec70 // bfmmla v16.4s, v3.8h, v26.8h\n" + "ldr q26, [x11, #0x20]\n" + ".inst 0x6e59ec2c // bfmmla v12.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec74 // bfmmla v20.4s, v3.8h, v25.8h\n" + "ldr q25, [x11, #0x30]\n" + ".inst 0x6e5aec29 // bfmmla v9.4s, v1.8h, v26.8h\n" "add x11, x11, #0x40\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" + ".inst 0x6e5aec71 // bfmmla v17.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e59ec2d // bfmmla v13.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec75 // bfmmla v21.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aec2a // bfmmla v10.4s, v1.8h, v26.8h\n" "add x10, x10, #0x40\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - "ldr q6, [x9, #0x30]\n" + ".inst 0x6e5aec72 // bfmmla v18.4s, v3.8h, v26.8h\n" + "ldr q26, [x9, #0x20]\n" + ".inst 0x6e59ec2e // bfmmla v14.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec76 // bfmmla v22.4s, v3.8h, v25.8h\n" + "ldr q25, [x9, #0x30]\n" "add x9, x9, #0x40\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" "98:" // Height 3: Multiply loop: Main loop skip "cbz x27, 103f\n" "cmp x27, #0x4\n" "blt 100f\n" "99:" // 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, [x12, #0x0]\n" - "trn1 v2.2d, v3.2d, v4.2d\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q7, [x12, #0x10]\n" - ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - "ldr q6, [x11, #0x0]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" - "ldr q7, [x11, #0x10]\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, [x12, #0x0]\n" + "trn1 v27.2d, v25.2d, v27.2d\n" + ".inst 0x6e5aef88 // bfmmla v8.4s, v28.8h, v26.8h\n" + "ldr q25, [x12, #0x10]\n" + ".inst 0x6e5aef70 // bfmmla v16.4s, v27.8h, v26.8h\n" + "ldr q26, [x11, #0x0]\n" + ".inst 0x6e59ef8c // bfmmla v12.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef74 // bfmmla v20.4s, v27.8h, v25.8h\n" + "ldr q25, [x11, #0x10]\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, #0x0]\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x0]\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, #0x10]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x10]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" "add x12, x12, #0x20\n" - ".inst 0x6e46ec52 // bfmmla v18.4s, v2.8h, v6.8h\n" - "ldr q6, [x9, #0x0]\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x9, #0x0]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" "add x11, x11, #0x20\n" - ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" - ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x9, #0x10]\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" "add x10, x10, #0x20\n" - ".inst 0x6e46ec53 // bfmmla v19.4s, v2.8h, v6.8h\n" - ".inst 0x6e47ec0f // bfmmla v15.4s, v0.8h, v7.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" "add x9, x9, #0x20\n" - ".inst 0x6e47ec57 // bfmmla v23.4s, v2.8h, v7.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" "bge 99b\n" "100:" // Height 3: Multiply loop: Skip odd blocks "cbz x27, 103f\n" @@ -1316,36 +1316,36 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr h2, [x25, #0x0]\n" "ldr h3, [x24, #0x0]\n" "102:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q7, [x12, #0x0]\n" - "ldr q6, [x12, #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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" + "ldr q26, [x12, #0x0]\n" + "ldr q29, [x12, #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, [x11, #0x0]\n" + ".inst 0x6e5def8c // bfmmla v12.4s, v28.8h, v29.8h\n" + ".inst 0x6e5def74 // bfmmla v20.4s, v27.8h, v29.8h\n" + "ldr q25, [x11, #0x10]\n" "add x12, x12, #0x20\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x0]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x0]\n" "add x11, x11, #0x20\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x10]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x10]\n" "add x10, x10, #0x20\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - "ldr q6, [x9, #0x10]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x9, #0x0]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x9, #0x10]\n" "add x9, x9, #0x20\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" "103:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1368,33 +1368,33 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "uzp1 v19.2d, v19.2d, v23.2d\n" "tbz %x[flags], #1, 104f\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" "104:" // Height 3: No activation "cmp x14, #0x10\n" "bge 113f\n" @@ -1709,14 +1709,14 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "131:" // 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, 132f\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, 133f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -1726,9 +1726,9 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "b 133f\n" "132:" // 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" "133:" // Height 4: input setup done "cmp x27, #0x8\n" "blt 136f\n" @@ -1741,174 +1741,174 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr q6, [x12, #0x10]\n" "blt 135f\n" "134:" // 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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\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, [x11, #0x0]\n" + ".inst 0x6e46ef8c // bfmmla v12.4s, v28.8h, v6.8h\n" + ".inst 0x6e46ef74 // bfmmla v20.4s, v27.8h, v6.8h\n" + "ldr q25, [x11, #0x10]\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, #0x0]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x0]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" "cmp x27, #0x10\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x10]\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, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x9, #0x0]\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, [x9, #0x10]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x9, #0x10]\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, [x12, #0x20]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + "ldr q26, [x12, #0x20]\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" "add x23, x23, #0x10\n" "ldr q4, [x23, #0x0]\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x12, #0x30]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" + "ldr q25, [x12, #0x30]\n" + ".inst 0x6e5aec28 // bfmmla v8.4s, v1.8h, v26.8h\n" "ldr q2, [x25, #0x0]\n" - ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" - "ldr q7, [x11, #0x20]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - "ldr q6, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" + ".inst 0x6e5aec70 // bfmmla v16.4s, v3.8h, v26.8h\n" + "ldr q26, [x11, #0x20]\n" + ".inst 0x6e59ec2c // bfmmla v12.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec74 // bfmmla v20.4s, v3.8h, v25.8h\n" + "ldr q25, [x11, #0x30]\n" + ".inst 0x6e5aec29 // bfmmla v9.4s, v1.8h, v26.8h\n" "add x12, x12, #0x40\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" + ".inst 0x6e5aec71 // bfmmla v17.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e59ec2d // bfmmla v13.4s, v1.8h, v25.8h\n" "add x11, x11, #0x40\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" + ".inst 0x6e59ec75 // bfmmla v21.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aec2a // bfmmla v10.4s, v1.8h, v26.8h\n" "add x10, x10, #0x40\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - "ldr q6, [x9, #0x30]\n" + ".inst 0x6e5aec72 // bfmmla v18.4s, v3.8h, v26.8h\n" + "ldr q26, [x9, #0x20]\n" + ".inst 0x6e59ec2e // bfmmla v14.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec76 // bfmmla v22.4s, v3.8h, v25.8h\n" + "ldr q25, [x9, #0x30]\n" "add x9, x9, #0x40\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, [x12, #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, [x12, #0x10]\n" "bge 134b\n" "135:" // 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" "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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\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, [x11, #0x0]\n" + ".inst 0x6e46ef8c // bfmmla v12.4s, v28.8h, v6.8h\n" + ".inst 0x6e46ef74 // bfmmla v20.4s, v27.8h, v6.8h\n" + "ldr q25, [x11, #0x10]\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, #0x0]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x0]\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, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x10]\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, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x9, #0x0]\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, [x9, #0x10]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x9, #0x10]\n" + ".inst 0x6e5aef8b // bfmmla v11.4s, v28.8h, v26.8h\n" "add x23, x23, #0x10\n" - ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" - "ldr q7, [x12, #0x20]\n" - ".inst 0x6e46ec0f // bfmmla v15.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" - "ldr q6, [x12, #0x30]\n" - ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" + ".inst 0x6e5aef73 // bfmmla v19.4s, v27.8h, v26.8h\n" + "ldr q26, [x12, #0x20]\n" + ".inst 0x6e59ef8f // bfmmla v15.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef77 // bfmmla v23.4s, v27.8h, v25.8h\n" + "ldr q25, [x12, #0x30]\n" + ".inst 0x6e5aec28 // bfmmla v8.4s, v1.8h, v26.8h\n" "add x12, x12, #0x40\n" - ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" - "ldr q7, [x11, #0x20]\n" - ".inst 0x6e46ec2c // bfmmla v12.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec74 // bfmmla v20.4s, v3.8h, v6.8h\n" - "ldr q6, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" + ".inst 0x6e5aec70 // bfmmla v16.4s, v3.8h, v26.8h\n" + "ldr q26, [x11, #0x20]\n" + ".inst 0x6e59ec2c // bfmmla v12.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec74 // bfmmla v20.4s, v3.8h, v25.8h\n" + "ldr q25, [x11, #0x30]\n" + ".inst 0x6e5aec29 // bfmmla v9.4s, v1.8h, v26.8h\n" "add x11, x11, #0x40\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\n" - ".inst 0x6e46ec2d // bfmmla v13.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec75 // bfmmla v21.4s, v3.8h, v6.8h\n" - "ldr q6, [x10, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" + ".inst 0x6e5aec71 // bfmmla v17.4s, v3.8h, v26.8h\n" + "ldr q26, [x10, #0x20]\n" + ".inst 0x6e59ec2d // bfmmla v13.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec75 // bfmmla v21.4s, v3.8h, v25.8h\n" + "ldr q25, [x10, #0x30]\n" + ".inst 0x6e5aec2a // bfmmla v10.4s, v1.8h, v26.8h\n" "add x10, x10, #0x40\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\n" - ".inst 0x6e46ec2e // bfmmla v14.4s, v1.8h, v6.8h\n" - ".inst 0x6e46ec76 // bfmmla v22.4s, v3.8h, v6.8h\n" - "ldr q6, [x9, #0x30]\n" + ".inst 0x6e5aec72 // bfmmla v18.4s, v3.8h, v26.8h\n" + "ldr q26, [x9, #0x20]\n" + ".inst 0x6e59ec2e // bfmmla v14.4s, v1.8h, v25.8h\n" + ".inst 0x6e59ec76 // bfmmla v22.4s, v3.8h, v25.8h\n" + "ldr q25, [x9, #0x30]\n" "add x9, x9, #0x40\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" "136:" // Height 4: Multiply loop: Main loop skip "cbz x27, 141f\n" "cmp x27, #0x4\n" "blt 138f\n" "137:" // 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, [x12, #0x0]\n" - "ldr q7, [x12, #0x10]\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - "ldr q6, [x11, #0x0]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" - "ldr q7, [x11, #0x10]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x0]\n" + "ldr q26, [x12, #0x0]\n" + "ldr q25, [x12, #0x10]\n" + ".inst 0x6e5aef88 // bfmmla v8.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef70 // bfmmla v16.4s, v27.8h, v26.8h\n" + "ldr q26, [x11, #0x0]\n" + ".inst 0x6e59ef8c // bfmmla v12.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef74 // bfmmla v20.4s, v27.8h, v25.8h\n" + "ldr q25, [x11, #0x10]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x0]\n" "add x12, x12, #0x20\n" - ".inst 0x6e47ec0d // bfmmla v13.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x10]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x10]\n" "add x11, x11, #0x20\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec52 // bfmmla v18.4s, v2.8h, v6.8h\n" - "ldr q6, [x9, #0x0]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x9, #0x0]\n" "add x10, x10, #0x20\n" - ".inst 0x6e47ec0e // bfmmla v14.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x10]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x9, #0x10]\n" "add x9, x9, #0x20\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 137b\n" "138:" // Height 4: Multiply loop: Skip odd blocks "cbz x27, 141f\n" @@ -1929,36 +1929,36 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr h3, [x24, #0x0]\n" "ldr h4, [x23, #0x0]\n" "140:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q7, [x12, #0x0]\n" - "ldr q6, [x12, #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, [x11, #0x0]\n" + "ldr q26, [x12, #0x0]\n" + "ldr q25, [x12, #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, [x11, #0x0]\n" "add x12, x12, #0x20\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" + ".inst 0x6e59ef8c // bfmmla v12.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef74 // bfmmla v20.4s, v27.8h, v25.8h\n" + "ldr q25, [x11, #0x10]\n" "add x11, x11, #0x20\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - "ldr q7, [x10, #0x0]\n" - ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec55 // bfmmla v21.4s, v2.8h, v6.8h\n" - "ldr q6, [x10, #0x10]\n" + ".inst 0x6e5aef89 // bfmmla v9.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef71 // bfmmla v17.4s, v27.8h, v26.8h\n" + "ldr q26, [x10, #0x0]\n" + ".inst 0x6e59ef8d // bfmmla v13.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef75 // bfmmla v21.4s, v27.8h, v25.8h\n" + "ldr q25, [x10, #0x10]\n" "add x10, x10, #0x20\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\n" - ".inst 0x6e46ec0e // bfmmla v14.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec56 // bfmmla v22.4s, v2.8h, v6.8h\n" - "ldr q6, [x9, #0x10]\n" + ".inst 0x6e5aef8a // bfmmla v10.4s, v28.8h, v26.8h\n" + ".inst 0x6e5aef72 // bfmmla v18.4s, v27.8h, v26.8h\n" + "ldr q26, [x9, #0x0]\n" + ".inst 0x6e59ef8e // bfmmla v14.4s, v28.8h, v25.8h\n" + ".inst 0x6e59ef76 // bfmmla v22.4s, v27.8h, v25.8h\n" + "ldr q25, [x9, #0x10]\n" "add x9, x9, #0x20\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" "141:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1986,41 +1986,41 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "uzp2 v19.2d, v19.2d, v23.2d\n" "tbz %x[flags], #1, 142f\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" "142:" // Height 4: No activation "cmp x14, #0x10\n" "bge 151f\n" @@ -2400,15 +2400,15 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "169:" // 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, 170f\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, 171f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -2419,10 +2419,10 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "b 171f\n" "170:" // 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" "171:" // Height 5: input setup done "cmp x27, #0x8\n" "blt 174f\n" @@ -2435,170 +2435,170 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr q7, [x12, #0x0]\n" "blt 173f\n" "172:" // 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, [x12, #0x10]\n" + "trn1 v4.2d, v5.2d, v0.2d\n" + "trn2 v5.2d, v5.2d, v0.2d\n" + "ldr q0, [x12, #0x10]\n" ".inst 0x6e47ec98 // bfmmla v24.4s, v4.8h, v7.8h\n" "ldr q7, [x11, #0x0]\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" "cmp x27, #0x10\n" - ".inst 0x6e46ec9c // bfmmla v28.4s, v4.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9c // bfmmla v28.4s, v4.8h, v0.8h\n" + "ldr q0, [x11, #0x10]\n" + ".inst 0x6e47ecc9 // bfmmla v9.4s, v6.8h, v7.8h\n" "add x26, x26, #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, #0x0]\n" "add x25, x25, #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 x24, x24, #0x10\n" "add x23, x23, #0x10\n" - ".inst 0x6e46ec9d // bfmmla v29.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9d // bfmmla v29.4s, v4.8h, v0.8h\n" + "ldr q0, [x10, #0x10]\n" + ".inst 0x6e47ecca // bfmmla v10.4s, v6.8h, v7.8h\n" "add x22, x22, #0x10\n" ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" ".inst 0x6e47ec9a // bfmmla v26.4s, v4.8h, v7.8h\n" "ldr q7, [x9, #0x0]\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, [x9, #0x10]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ecce // bfmmla v14.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec56 // bfmmla v22.4s, v2.8h, v0.8h\n" + ".inst 0x6e40ec9e // bfmmla v30.4s, v4.8h, v0.8h\n" + "ldr q0, [x9, #0x10]\n" + ".inst 0x6e47eccb // bfmmla v11.4s, v6.8h, v7.8h\n" ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" ".inst 0x6e47ec9b // bfmmla v27.4s, v4.8h, v7.8h\n" "ldr q7, [x12, #0x20]\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, [x12, #0x30]\n" + ".inst 0x6e40ec9f // bfmmla v31.4s, v4.8h, v0.8h\n" + "ldr q0, [x12, #0x30]\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, [x11, #0x20]\n" + "ldr q6, [x11, #0x20]\n" "add x12, x12, #0x40\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, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\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, [x11, #0x30]\n" + ".inst 0x6e46ec29 // bfmmla v9.4s, v1.8h, v6.8h\n" "add x11, x11, #0x40\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecb9 // bfmmla v25.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\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, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" + ".inst 0x6e46ec71 // bfmmla v17.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecb9 // bfmmla v25.4s, v5.8h, v6.8h\n" + "ldr q6, [x10, #0x20]\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, #0x30]\n" + ".inst 0x6e46ec2a // bfmmla v10.4s, v1.8h, v6.8h\n" "add x10, x10, #0x40\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecba // bfmmla v26.4s, v5.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\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, [x9, #0x30]\n" + ".inst 0x6e46ec72 // bfmmla v18.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecba // bfmmla v26.4s, v5.8h, v6.8h\n" + "ldr q6, [x9, #0x20]\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, [x9, #0x30]\n" "add x9, x9, #0x40\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, [x12, #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 172b\n" "173:" // 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" "sub x27, x27, #0x8\n" - "trn1 v4.2d, v5.2d, v6.2d\n" - "trn2 v5.2d, v5.2d, v6.2d\n" - "ldr q6, [x12, #0x10]\n" + "trn1 v4.2d, v5.2d, v0.2d\n" + "trn2 v5.2d, v5.2d, v0.2d\n" + "ldr q0, [x12, #0x10]\n" ".inst 0x6e47ec98 // bfmmla v24.4s, v4.8h, v7.8h\n" "ldr q7, [x11, #0x0]\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, [x11, #0x10]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9c // bfmmla v28.4s, v4.8h, v0.8h\n" + "ldr q0, [x11, #0x10]\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, #0x0]\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, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec9d // bfmmla v29.4s, v4.8h, v0.8h\n" + "ldr q0, [x10, #0x10]\n" + ".inst 0x6e47ecca // bfmmla v10.4s, v6.8h, 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, [x9, #0x0]\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, [x9, #0x10]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ecce // bfmmla v14.4s, v6.8h, v0.8h\n" + ".inst 0x6e40ec56 // bfmmla v22.4s, v2.8h, v0.8h\n" + ".inst 0x6e40ec9e // bfmmla v30.4s, v4.8h, v0.8h\n" + "ldr q0, [x9, #0x10]\n" + ".inst 0x6e47eccb // bfmmla v11.4s, v6.8h, v7.8h\n" ".inst 0x6e47ec53 // bfmmla v19.4s, v2.8h, v7.8h\n" ".inst 0x6e47ec9b // bfmmla v27.4s, v4.8h, v7.8h\n" "ldr q7, [x12, #0x20]\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, [x12, #0x30]\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, [x12, #0x30]\n" ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" "add x12, x12, #0x40\n" ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" ".inst 0x6e47ecb8 // bfmmla v24.4s, v5.8h, v7.8h\n" - "ldr q7, [x11, #0x20]\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, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" + "ldr q0, [x11, #0x20]\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, [x11, #0x30]\n" + ".inst 0x6e40ec29 // bfmmla v9.4s, v1.8h, v0.8h\n" "add x11, x11, #0x40\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecb9 // bfmmla v25.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\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, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" + ".inst 0x6e40ec71 // bfmmla v17.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecb9 // bfmmla v25.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0x20]\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, #0x30]\n" + ".inst 0x6e40ec2a // bfmmla v10.4s, v1.8h, v0.8h\n" "add x10, x10, #0x40\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecba // bfmmla v26.4s, v5.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\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" + ".inst 0x6e40ec72 // bfmmla v18.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecba // bfmmla v26.4s, v5.8h, v0.8h\n" + "ldr q0, [x9, #0x20]\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, [x9, #0x30]\n" "add x9, x9, #0x40\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" @@ -2608,51 +2608,51 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "blt 176f\n" "175:" // 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, [x12, #0x0]\n" - "trn1 v4.2d, v5.2d, v7.2d\n" - ".inst 0x6e46ec08 // bfmmla v8.4s, v0.8h, v6.8h\n" - "ldr q7, [x12, #0x10]\n" - ".inst 0x6e46ec50 // bfmmla v16.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec98 // bfmmla v24.4s, v4.8h, v6.8h\n" - "ldr q6, [x11, #0x0]\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, [x12, #0x0]\n" + "trn1 v2.2d, v0.2d, v2.2d\n" + ".inst 0x6e41ec88 // bfmmla v8.4s, v4.8h, v1.8h\n" + "ldr q0, [x12, #0x10]\n" + ".inst 0x6e41ec70 // bfmmla v16.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec58 // bfmmla v24.4s, v2.8h, v1.8h\n" + "ldr q1, [x11, #0x0]\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" "add x12, x12, #0x20\n" - ".inst 0x6e47ec9c // bfmmla v28.4s, v4.8h, v7.8h\n" - "ldr q7, [x11, #0x10]\n" - ".inst 0x6e46ec09 // bfmmla v9.4s, v0.8h, v6.8h\n" + ".inst 0x6e40ec5c // bfmmla v28.4s, v2.8h, v0.8h\n" + "ldr q0, [x11, #0x10]\n" + ".inst 0x6e41ec89 // bfmmla v9.4s, v4.8h, v1.8h\n" "add x11, x11, #0x20\n" - ".inst 0x6e46ec51 // bfmmla v17.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec99 // bfmmla v25.4s, v4.8h, v6.8h\n" - "ldr q6, [x10, #0x0]\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, #0x10]\n" - ".inst 0x6e46ec0a // bfmmla v10.4s, v0.8h, v6.8h\n" + ".inst 0x6e41ec71 // bfmmla v17.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec59 // bfmmla v25.4s, v2.8h, v1.8h\n" + "ldr q1, [x10, #0x0]\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, #0x10]\n" + ".inst 0x6e41ec8a // bfmmla v10.4s, v4.8h, v1.8h\n" "add x10, x10, #0x20\n" - ".inst 0x6e46ec52 // bfmmla v18.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9a // bfmmla v26.4s, v4.8h, v6.8h\n" + ".inst 0x6e41ec72 // bfmmla v18.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5a // bfmmla v26.4s, v2.8h, v1.8h\n" "ldr q6, [x9, #0x0]\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, [x9, #0x10]\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, [x9, #0x10]\n" + ".inst 0x6e46ec8b // bfmmla v11.4s, v4.8h, v6.8h\n" "add x9, x9, #0x20\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 175b\n" "176:" // Height 5: Multiply loop: Skip odd blocks "cbz x27, 179f\n" @@ -2676,45 +2676,45 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr h4, [x23, #0x0]\n" "ldr h5, [x22, #0x0]\n" "178:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q7, [x12, #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, [x12, #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, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" + "ldr q6, [x12, #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, [x12, #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, [x11, #0x0]\n" + ".inst 0x6e41ecec // bfmmla v12.4s, v7.8h, v1.8h\n" "add x12, x12, #0x20\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" - ".inst 0x6e46ec9c // bfmmla v28.4s, v4.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" + ".inst 0x6e41ec74 // bfmmla v20.4s, v3.8h, v1.8h\n" + ".inst 0x6e41ec5c // bfmmla v28.4s, v2.8h, v1.8h\n" + "ldr q1, [x11, #0x10]\n" "add x11, x11, #0x20\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, #0x0]\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, #0x10]\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, #0x0]\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, #0x10]\n" "add x10, x10, #0x20\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, [x9, #0x0]\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" + ".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, [x9, #0x0]\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, [x9, #0x10]\n" "add x9, x9, #0x20\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" "179:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -3213,16 +3213,16 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "207:" // 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, 208f\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, 209f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -3234,11 +3234,11 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "b 209f\n" "208:" // 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" "209:" // Height 6: input setup done "cmp x27, #0x8\n" "blt 212f\n" @@ -3299,45 +3299,45 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" "ldr q2, [x25, #0x0]\n" ".inst 0x6e46ec9f // bfmmla v31.4s, v4.8h, v6.8h\n" - "ldr q6, [x12, #0x30]\n" + "ldr q0, [x12, #0x30]\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, [x11, #0x20]\n" + "ldr q6, [x11, #0x20]\n" "add x12, x12, #0x40\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, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\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, [x11, #0x30]\n" + ".inst 0x6e46ec29 // bfmmla v9.4s, v1.8h, v6.8h\n" "add x11, x11, #0x40\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecb9 // bfmmla v25.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\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, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" + ".inst 0x6e46ec71 // bfmmla v17.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecb9 // bfmmla v25.4s, v5.8h, v6.8h\n" + "ldr q6, [x10, #0x20]\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, #0x30]\n" + ".inst 0x6e46ec2a // bfmmla v10.4s, v1.8h, v6.8h\n" "add x10, x10, #0x40\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecba // bfmmla v26.4s, v5.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\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, [x9, #0x30]\n" + ".inst 0x6e46ec72 // bfmmla v18.4s, v3.8h, v6.8h\n" + ".inst 0x6e46ecba // bfmmla v26.4s, v5.8h, v6.8h\n" + "ldr q6, [x9, #0x20]\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, [x9, #0x30]\n" "add x9, x9, #0x40\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, [x12, #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 210b\n" @@ -3387,38 +3387,38 @@ void a64_ffhybrid_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, [x12, #0x30]\n" + "ldr q2, [x12, #0x30]\n" ".inst 0x6e47ec28 // bfmmla v8.4s, v1.8h, v7.8h\n" "add x12, x12, #0x40\n" ".inst 0x6e47ec70 // bfmmla v16.4s, v3.8h, v7.8h\n" ".inst 0x6e47ecb8 // bfmmla v24.4s, v5.8h, v7.8h\n" - "ldr q7, [x11, #0x20]\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, [x11, #0x30]\n" - ".inst 0x6e47ec29 // bfmmla v9.4s, v1.8h, v7.8h\n" + "ldr q0, [x11, #0x20]\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, [x11, #0x30]\n" + ".inst 0x6e40ec29 // bfmmla v9.4s, v1.8h, v0.8h\n" "add x11, x11, #0x40\n" - ".inst 0x6e47ec71 // bfmmla v17.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecb9 // bfmmla v25.4s, v5.8h, v7.8h\n" - "ldr q7, [x10, #0x20]\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, #0x30]\n" - ".inst 0x6e47ec2a // bfmmla v10.4s, v1.8h, v7.8h\n" + ".inst 0x6e40ec71 // bfmmla v17.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecb9 // bfmmla v25.4s, v5.8h, v0.8h\n" + "ldr q0, [x10, #0x20]\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, #0x30]\n" + ".inst 0x6e40ec2a // bfmmla v10.4s, v1.8h, v0.8h\n" "add x10, x10, #0x40\n" - ".inst 0x6e47ec72 // bfmmla v18.4s, v3.8h, v7.8h\n" - ".inst 0x6e47ecba // bfmmla v26.4s, v5.8h, v7.8h\n" - "ldr q7, [x9, #0x20]\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" + ".inst 0x6e40ec72 // bfmmla v18.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ecba // bfmmla v26.4s, v5.8h, v0.8h\n" + "ldr q0, [x9, #0x20]\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, [x9, #0x30]\n" "add x9, x9, #0x40\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" @@ -3428,52 +3428,52 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "blt 214f\n" "213:" // 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, [x12, #0x0]\n" - "ldr q7, [x12, #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, [x11, #0x0]\n" - ".inst 0x6e47ec0c // bfmmla v12.4s, v0.8h, v7.8h\n" + "ldr d1, [x22], #0x8\n" + "ldr d0, [x21], #0x8\n" + "trn1 v2.2d, v1.2d, v0.2d\n" + "ldr q1, [x12, #0x0]\n" + "ldr q0, [x12, #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, [x11, #0x0]\n" + ".inst 0x6e40ec8c // bfmmla v12.4s, v4.8h, v0.8h\n" "add x12, x12, #0x20\n" - ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9c // bfmmla v28.4s, v4.8h, v7.8h\n" - "ldr q7, [x11, #0x10]\n" + ".inst 0x6e40ec74 // bfmmla v20.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5c // bfmmla v28.4s, v2.8h, v0.8h\n" + "ldr q0, [x11, #0x10]\n" "add x11, x11, #0x20\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, #0x0]\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, #0x10]\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, #0x0]\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, #0x10]\n" "add x10, x10, #0x20\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 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, [x9, #0x0]\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, [x9, #0x10]\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, [x9, #0x10]\n" "add x9, x9, #0x20\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 213b\n" "214:" // Height 6: Multiply loop: Skip odd blocks "cbz x27, 217f\n" @@ -3500,45 +3500,45 @@ void a64_ffhybrid_bf16fp32_mmla_6x16 ( "ldr h5, [x22, #0x0]\n" "ldr h6, [x21, #0x0]\n" "216:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q7, [x12, #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, [x12, #0x10]\n" - ".inst 0x6e47ec50 // bfmmla v16.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec98 // bfmmla v24.4s, v4.8h, v7.8h\n" - "ldr q7, [x11, #0x0]\n" - ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" - ".inst 0x6e46ec54 // bfmmla v20.4s, v2.8h, v6.8h\n" + "ldr q0, [x12, #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, [x12, #0x10]\n" + ".inst 0x6e40ec70 // bfmmla v16.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec58 // bfmmla v24.4s, v2.8h, v0.8h\n" + "ldr q0, [x11, #0x0]\n" + ".inst 0x6e41ecec // bfmmla v12.4s, v7.8h, v1.8h\n" + ".inst 0x6e41ec74 // bfmmla v20.4s, v3.8h, v1.8h\n" "add x12, x12, #0x20\n" - ".inst 0x6e46ec9c // bfmmla v28.4s, v4.8h, v6.8h\n" - "ldr q6, [x11, #0x10]\n" - ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" + ".inst 0x6e41ec5c // bfmmla v28.4s, v2.8h, v1.8h\n" + "ldr q1, [x11, #0x10]\n" + ".inst 0x6e40ece9 // bfmmla v9.4s, v7.8h, v0.8h\n" "add x11, x11, #0x20\n" - ".inst 0x6e47ec51 // bfmmla v17.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec99 // bfmmla v25.4s, v4.8h, v7.8h\n" - "ldr q7, [x10, #0x0]\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, #0x10]\n" - ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" + ".inst 0x6e40ec71 // bfmmla v17.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec59 // bfmmla v25.4s, v2.8h, v0.8h\n" + "ldr q0, [x10, #0x0]\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, #0x10]\n" + ".inst 0x6e40ecea // bfmmla v10.4s, v7.8h, v0.8h\n" "add x10, x10, #0x20\n" - ".inst 0x6e47ec52 // bfmmla v18.4s, v2.8h, v7.8h\n" - ".inst 0x6e47ec9a // bfmmla v26.4s, v4.8h, v7.8h\n" - "ldr q7, [x9, #0x0]\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" + ".inst 0x6e40ec72 // bfmmla v18.4s, v3.8h, v0.8h\n" + ".inst 0x6e40ec5a // bfmmla v26.4s, v2.8h, v0.8h\n" + "ldr q0, [x9, #0x0]\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, [x9, #0x10]\n" - ".inst 0x6e47ec0b // bfmmla v11.4s, v0.8h, v7.8h\n" + ".inst 0x6e40eceb // bfmmla v11.4s, v7.8h, v0.8h\n" "add x9, x9, #0x20\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" "217:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" |