diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp | 4364 |
1 files changed, 2147 insertions, 2217 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp index 85944e9f6a..afb06dedea 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp @@ -96,238 +96,232 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "1:" // Row loop "cmp %x[M], #0x6\n" - "bge 181f\n" + "bge 176f\n" "cmp %x[M], #0x4\n" - "bgt 145f\n" - "beq 109f\n" + "bgt 141f\n" + "beq 106f\n" "cmp %x[M], #0x2\n" - "bgt 73f\n" - "beq 37f\n" - "ldr x16, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[bias]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 2f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19, LSL #2\n" - "b 3f\n" - "2:" // Height 1: setup direct output - "mov x13, %x[output_ptr]\n" - "3:" // Height 1: Column loop - "cbz x14, 4f\n" - "ldr q8, [x14, #0x0]\n" - "ldr q9, [x14, #0x10]\n" - "ldr q10, [x14, #0x20]\n" - "ldr q11, [x14, #0x30]\n" - "add x14, x14, #0x40\n" - "b 15f\n" - "4:" // Height 1: no bias - "tbz %x[flags], #0, 14f\n" - "cmp x16, #0x10\n" - "bge 13f\n" - "tbz x16, #3, 8f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v9.4s }, [x13], #0x10\n" - "tbz x16, #2, 6f\n" - "ld1 { v10.4s }, [x13], #0x10\n" - "tbz x16, #1, 5f\n" + "bgt 71f\n" + "beq 36f\n" + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x9, %x[bias]\n" + "mov x28, %x[output_ptr]\n" + "2:" // Height 1: Column loop + "cbz x9, 3f\n" + "ldr q8, [x9, #0x0]\n" + "ldr q9, [x9, #0x10]\n" + "ldr q10, [x9, #0x20]\n" + "ldr q11, [x9, #0x30]\n" + "add x9, x9, #0x40\n" + "b 14f\n" + "3:" // Height 1: no bias + "tbz %x[flags], #0, 13f\n" + "cmp x11, #0x10\n" + "bge 12f\n" + "tbz x11, #3, 7f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v9.4s }, [x28], #0x10\n" + "tbz x11, #2, 5f\n" + "ld1 { v10.4s }, [x28], #0x10\n" + "tbz x11, #1, 4f\n" "mov x19, #0x38\n" - "ldr d11, [x13], #0x8\n" - "tbz x16, #0, 12f\n" - "ld1 { v11.s }[2], [x13]\n" - "b 12f\n" - "5:" // Height 1: Partial accumulate: partial_1_12 + "ldr d11, [x28], #0x8\n" + "tbz x11, #0, 11f\n" + "ld1 { v11.s }[2], [x28]\n" + "b 11f\n" + "4:" // Height 1: Partial accumulate: partial_1_12 "mov x19, #0x30\n" - "tbz x16, #0, 12f\n" - "ldr s11, [x13, #0x0]\n" - "b 12f\n" - "6:" // Height 1: Partial accumulate: partial_2_8 - "tbz x16, #1, 7f\n" - "ldr d10, [x13], #0x8\n" + "tbz x11, #0, 11f\n" + "ldr s11, [x28, #0x0]\n" + "b 11f\n" + "5:" // Height 1: Partial accumulate: partial_2_8 + "tbz x11, #1, 6f\n" + "ldr d10, [x28], #0x8\n" "mov x19, #0x28\n" - "tbz x16, #0, 12f\n" - "ld1 { v10.s }[2], [x13]\n" - "b 12f\n" - "7:" // Height 1: Partial accumulate: partial_1_8 + "tbz x11, #0, 11f\n" + "ld1 { v10.s }[2], [x28]\n" + "b 11f\n" + "6:" // Height 1: Partial accumulate: partial_1_8 "mov x19, #0x20\n" - "tbz x16, #0, 12f\n" - "ldr s10, [x13, #0x0]\n" - "b 12f\n" - "8:" // Height 1: Partial accumulate: partial_4_0 - "tbz x16, #2, 10f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "tbz x16, #1, 9f\n" + "tbz x11, #0, 11f\n" + "ldr s10, [x28, #0x0]\n" + "b 11f\n" + "7:" // Height 1: Partial accumulate: partial_4_0 + "tbz x11, #2, 9f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "tbz x11, #1, 8f\n" + "ldr d9, [x28], #0x8\n" "mov x19, #0x18\n" - "ldr d9, [x13], #0x8\n" - "tbz x16, #0, 12f\n" - "ld1 { v9.s }[2], [x13]\n" - "b 12f\n" - "9:" // Height 1: Partial accumulate: partial_1_4 + "tbz x11, #0, 11f\n" + "ld1 { v9.s }[2], [x28]\n" + "b 11f\n" + "8:" // Height 1: Partial accumulate: partial_1_4 "mov x19, #0x10\n" - "tbz x16, #0, 12f\n" - "ldr s9, [x13, #0x0]\n" - "b 12f\n" - "10:" // Height 1: Partial accumulate: partial_2_0 - "tbz x16, #1, 11f\n" - "ldr d8, [x13], #0x8\n" + "tbz x11, #0, 11f\n" + "ldr s9, [x28, #0x0]\n" + "b 11f\n" + "9:" // Height 1: Partial accumulate: partial_2_0 + "tbz x11, #1, 10f\n" + "ldr d8, [x28], #0x8\n" "mov x19, #0x8\n" - "tbz x16, #0, 12f\n" - "ld1 { v8.s }[2], [x13]\n" - "b 12f\n" - "11:" // Height 1: Partial accumulate: partial_1_0 + "tbz x11, #0, 11f\n" + "ld1 { v8.s }[2], [x28]\n" + "b 11f\n" + "10:" // Height 1: Partial accumulate: partial_1_0 + "ldr s8, [x28, #0x0]\n" "mov x19, #0x0\n" - "ldr s8, [x13, #0x0]\n" - "12:" // Height 1: Partial accumulate: Done - "sub x13, x13, x19\n" - "b 15f\n" - "13:" // Height 1: full accumulate - "ldr q8, [x13, #0x0]\n" - "ldr q9, [x13, #0x10]\n" - "ldr q10, [x13, #0x20]\n" - "ldr q11, [x13, #0x30]\n" - "b 15f\n" - "14:" // Height 1: no accumulate + "11:" // Height 1: Partial accumulate: Done + "sub x28, x28, x19\n" + "b 14f\n" + "12:" // Height 1: full accumulate + "ldr q8, [x28, #0x0]\n" + "ldr q9, [x28, #0x10]\n" + "ldr q10, [x28, #0x20]\n" + "ldr q11, [x28, #0x30]\n" + "b 14f\n" + "13:" // Height 1: no accumulate "movi v8.16b, #0x0\n" "movi v9.16b, #0x0\n" "movi v10.16b, #0x0\n" "movi v11.16b, #0x0\n" - "15:" // Height 1: setup done - "mov x12, #0x0\n" - "16:" // Height 1: String loop + "14:" // Height 1: setup done + "mov x27, #0x0\n" + "15:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 17f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 16f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "cbnz x12, 18f\n" + "ldr x25, [x20, #0x0]\n" + "cbnz x27, 17f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19, LSL #1\n" - "b 18f\n" - "17:" // Height 1: setup direct input - "mov x10, %x[input_ptr]\n" - "18:" // Height 1: input setup done - "cmp x11, #0x8\n" - "blt 21f\n" - "cmp x11, #0x10\n" + "add x25, x25, x19, LSL #1\n" + "b 17f\n" + "16:" // Height 1: setup direct input + "mov x25, %x[input_ptr]\n" + "17:" // Height 1: input setup done + "cmp x26, #0x8\n" "blt 20f\n" - "19:" // Height 1: Multiply loop: Main loop head - "ldr q0, [x10, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "ldr q0, [x25, #0x0]\n" + "ldr q6, [x10, #0x0]\n" + "cmp x26, #0x10\n" + "blt 19f\n" + "18:" // Height 1: Multiply loop: Main loop head ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q7, [x10, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q6, [x10, #0x20]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "ldr q6, [x15, #0x40]\n" - "add x10, x10, #0x10\n" + "ldr q7, [x10, #0x30]\n" + "cmp x26, #0x10\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q6, [x10, #0x40]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - "ldr q6, [x15, #0x60]\n" - "sub x11, x11, #0x8\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - "ldr q7, [x15, #0x70]\n" - "cmp x11, #0x10\n" + "ldr q7, [x10, #0x70]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - "ldr q7, [x15, #0xf0]\n" - "add x15, x15, #0x100\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" - "bge 19b\n" - "20:" // Height 1: Multiply loop: Single iteration only - "sub x11, x11, #0x8\n" - "ldr q0, [x10, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "ldr q0, [x25, #0x0]\n" + "bge 18b\n" + "19:" // Height 1: Multiply loop: Single iteration only ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q7, [x10, #0x10]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q6, [x10, #0x20]\n" + "add x25, x25, #0x10\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "ldr q6, [x15, #0x40]\n" - "add x10, x10, #0x10\n" + "ldr q7, [x10, #0x30]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - "ldr q7, [x15, #0xf0]\n" - "add x15, x15, #0x100\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" - "21:" // Height 1: Multiply loop: Main loop skip - "cbz x11, 25f\n" - "cmp x11, #0x2\n" - "blt 23f\n" - "22:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x10], #0x4\n" - "ldr q6, [x15, #0x0]\n" + "20:" // Height 1: Multiply loop: Main loop skip + "cbz x26, 24f\n" + "cmp x26, #0x2\n" + "blt 22f\n" + "21:" // Height 1: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x2\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q7, [x10, #0x10]\n" + "cmp x26, #0x2\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q6, [x10, #0x20]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "sub x11, x11, #0x2\n" - "add x15, x15, #0x40\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - "cmp x11, #0x2\n" - "bge 22b\n" - "cbz x11, 25f\n" - "23:" // Height 1: Multiply loop: Skip odd blocks - "ldr h0, [x10, #0x0]\n" - "24:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q6, [x15, #0x0]\n" + "bge 21b\n" + "cbz x26, 24f\n" + "22:" // Height 1: Multiply loop: Skip odd blocks + "ldr h0, [x25, #0x0]\n" + "23:" // Height 1: Multiply loop: Ragged operand read: Done + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q7, [x10, #0x10]\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - "25:" // Height 1: Multiply loop: No odd multiplies + "24:" // Height 1: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x12, x12, #0x1\n" - "cmp x12, x19\n" - "bne 16b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "tbz %x[flags], #1, 26f\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 15b\n" + "prfm pstl1keep, [x28, #0x0]\n" + "tbz %x[flags], #1, 25f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1r { v1.4s }, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -340,178 +334,170 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "fmax v9.4s, v9.4s, v1.4s\n" "fmax v10.4s, v10.4s, v1.4s\n" "fmax v11.4s, v11.4s, v1.4s\n" - "26:" // Height 1: No activation - "cmp x16, #0x10\n" - "bge 35f\n" - "tbz x16, #3, 30f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v9.4s }, [x13], #0x10\n" - "tbz x16, #2, 28f\n" - "st1 { v10.4s }, [x13], #0x10\n" - "tbz x16, #1, 27f\n" - "str d11, [x13], #0x8\n" - "tbz x16, #0, 34f\n" - "st1 { v11.s }[2], [x13]\n" - "b 34f\n" - "27:" // Height 1: Partial direct writeback: partial_1_12 - "tbz x16, #0, 34f\n" - "str s11, [x13, #0x0]\n" - "b 34f\n" - "28:" // Height 1: Partial direct writeback: partial_2_8 - "tbz x16, #1, 29f\n" - "str d10, [x13], #0x8\n" - "tbz x16, #0, 34f\n" - "st1 { v10.s }[2], [x13]\n" - "b 34f\n" - "29:" // Height 1: Partial direct writeback: partial_1_8 - "tbz x16, #0, 34f\n" - "str s10, [x13, #0x0]\n" - "b 34f\n" - "30:" // Height 1: Partial direct writeback: partial_4_0 - "tbz x16, #2, 32f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "tbz x16, #1, 31f\n" - "str d9, [x13], #0x8\n" - "tbz x16, #0, 34f\n" - "st1 { v9.s }[2], [x13]\n" - "b 34f\n" - "31:" // Height 1: Partial direct writeback: partial_1_4 - "tbz x16, #0, 34f\n" - "str s9, [x13, #0x0]\n" - "b 34f\n" - "32:" // Height 1: Partial direct writeback: partial_2_0 - "tbz x16, #1, 33f\n" - "str d8, [x13], #0x8\n" - "tbz x16, #0, 34f\n" - "st1 { v8.s }[2], [x13]\n" - "b 34f\n" - "33:" // Height 1: Partial direct writeback: partial_1_0 - "str s8, [x13, #0x0]\n" - "34:" // Height 1: Partial direct writeback: Done - "b 36f\n" - "35:" // Height 1: Full writeback - "str q8, [x13, #0x0]\n" - "str q9, [x13, #0x10]\n" - "str q10, [x13, #0x20]\n" - "str q11, [x13, #0x30]\n" - "add x13, x13, #0x40\n" - "36:" // Height 1: Writeback done - "subs x16, x16, #0x10\n" - "bgt 3b\n" - "b 218f\n" - "37:" // Height 2 - "ldr x16, [%x[args_ptr], %[offsetof_N]]\n" - "mov x14, %x[bias]\n" - "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 38f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "add x9, x9, x19, LSL #2\n" - "b 39f\n" - "38:" // Height 2: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19, LSL #2\n" - "39:" // Height 2: Column loop - "cbz x14, 40f\n" - "ldr q8, [x14, #0x0]\n" + "25:" // Height 1: No activation + "cmp x11, #0x10\n" + "bge 34f\n" + "tbz x11, #3, 29f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v9.4s }, [x28], #0x10\n" + "tbz x11, #2, 27f\n" + "st1 { v10.4s }, [x28], #0x10\n" + "tbz x11, #1, 26f\n" + "str d11, [x28], #0x8\n" + "tbz x11, #0, 33f\n" + "st1 { v11.s }[2], [x28]\n" + "b 33f\n" + "26:" // Height 1: Partial direct writeback: partial_1_12 + "tbz x11, #0, 33f\n" + "str s11, [x28, #0x0]\n" + "b 33f\n" + "27:" // Height 1: Partial direct writeback: partial_2_8 + "tbz x11, #1, 28f\n" + "str d10, [x28], #0x8\n" + "tbz x11, #0, 33f\n" + "st1 { v10.s }[2], [x28]\n" + "b 33f\n" + "28:" // Height 1: Partial direct writeback: partial_1_8 + "tbz x11, #0, 33f\n" + "str s10, [x28, #0x0]\n" + "b 33f\n" + "29:" // Height 1: Partial direct writeback: partial_4_0 + "tbz x11, #2, 31f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "tbz x11, #1, 30f\n" + "str d9, [x28], #0x8\n" + "tbz x11, #0, 33f\n" + "st1 { v9.s }[2], [x28]\n" + "b 33f\n" + "30:" // Height 1: Partial direct writeback: partial_1_4 + "tbz x11, #0, 33f\n" + "str s9, [x28, #0x0]\n" + "b 33f\n" + "31:" // Height 1: Partial direct writeback: partial_2_0 + "tbz x11, #1, 32f\n" + "str d8, [x28], #0x8\n" + "tbz x11, #0, 33f\n" + "st1 { v8.s }[2], [x28]\n" + "b 33f\n" + "32:" // Height 1: Partial direct writeback: partial_1_0 + "str s8, [x28, #0x0]\n" + "33:" // Height 1: Partial direct writeback: Done + "b 35f\n" + "34:" // Height 1: Full writeback + "str q8, [x28, #0x0]\n" + "str q9, [x28, #0x10]\n" + "str q10, [x28, #0x20]\n" + "str q11, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + "35:" // Height 1: Writeback done + "subs x11, x11, #0x10\n" + "bgt 2b\n" + "b 212f\n" + "36:" // Height 2 + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "mov x9, %x[bias]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x28, %x[output_ptr]\n" + "37:" // Height 2: Column loop + "cbz x9, 38f\n" + "ldr q8, [x9, #0x0]\n" "mov v12.16b, v8.16b\n" - "ldr q9, [x14, #0x10]\n" - "ldr q10, [x14, #0x20]\n" + "ldr q9, [x9, #0x10]\n" + "ldr q10, [x9, #0x20]\n" "mov v13.16b, v9.16b\n" - "ldr q11, [x14, #0x30]\n" + "ldr q11, [x9, #0x30]\n" + "add x9, x9, #0x40\n" "mov v14.16b, v10.16b\n" - "add x14, x14, #0x40\n" "mov v15.16b, v11.16b\n" - "b 51f\n" - "40:" // Height 2: no bias - "tbz %x[flags], #0, 50f\n" - "cmp x16, #0x10\n" - "bge 49f\n" - "tbz x16, #3, 44f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "ld1 { v9.4s }, [x13], #0x10\n" - "ld1 { v13.4s }, [x9], #0x10\n" - "tbz x16, #2, 42f\n" - "ld1 { v10.4s }, [x13], #0x10\n" - "ld1 { v14.4s }, [x9], #0x10\n" - "tbz x16, #1, 41f\n" + "b 49f\n" + "38:" // Height 2: no bias + "tbz %x[flags], #0, 48f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "cmp x11, #0x10\n" + "add x24, x28, x19, LSL #2\n" + "bge 47f\n" + "tbz x11, #3, 42f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "ld1 { v9.4s }, [x28], #0x10\n" + "ld1 { v13.4s }, [x24], #0x10\n" + "tbz x11, #2, 40f\n" + "ld1 { v10.4s }, [x28], #0x10\n" + "ld1 { v14.4s }, [x24], #0x10\n" + "tbz x11, #1, 39f\n" "mov x19, #0x38\n" - "ldr d11, [x13], #0x8\n" - "ldr d15, [x9], #0x8\n" - "tbz x16, #0, 48f\n" - "ld1 { v11.s }[2], [x13]\n" - "ld1 { v15.s }[2], [x9]\n" - "b 48f\n" - "41:" // Height 2: Partial accumulate: partial_1_12 + "ldr d11, [x28], #0x8\n" + "ldr d15, [x24], #0x8\n" + "tbz x11, #0, 46f\n" + "ld1 { v11.s }[2], [x28]\n" + "ld1 { v15.s }[2], [x24]\n" + "b 46f\n" + "39:" // Height 2: Partial accumulate: partial_1_12 "mov x19, #0x30\n" - "tbz x16, #0, 48f\n" - "ldr s11, [x13, #0x0]\n" - "ldr s15, [x9, #0x0]\n" - "b 48f\n" - "42:" // Height 2: Partial accumulate: partial_2_8 - "tbz x16, #1, 43f\n" - "ldr d10, [x13], #0x8\n" - "ldr d14, [x9], #0x8\n" + "tbz x11, #0, 46f\n" + "ldr s11, [x28, #0x0]\n" + "ldr s15, [x24, #0x0]\n" + "b 46f\n" + "40:" // Height 2: Partial accumulate: partial_2_8 + "tbz x11, #1, 41f\n" + "ldr d10, [x28], #0x8\n" + "ldr d14, [x24], #0x8\n" "mov x19, #0x28\n" - "tbz x16, #0, 48f\n" - "ld1 { v10.s }[2], [x13]\n" - "ld1 { v14.s }[2], [x9]\n" - "b 48f\n" - "43:" // Height 2: Partial accumulate: partial_1_8 + "tbz x11, #0, 46f\n" + "ld1 { v10.s }[2], [x28]\n" + "ld1 { v14.s }[2], [x24]\n" + "b 46f\n" + "41:" // Height 2: Partial accumulate: partial_1_8 "mov x19, #0x20\n" - "tbz x16, #0, 48f\n" - "ldr s10, [x13, #0x0]\n" - "ldr s14, [x9, #0x0]\n" - "b 48f\n" - "44:" // Height 2: Partial accumulate: partial_4_0 - "tbz x16, #2, 46f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "tbz x16, #1, 45f\n" + "tbz x11, #0, 46f\n" + "ldr s10, [x28, #0x0]\n" + "ldr s14, [x24, #0x0]\n" + "b 46f\n" + "42:" // Height 2: Partial accumulate: partial_4_0 + "tbz x11, #2, 44f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "tbz x11, #1, 43f\n" "mov x19, #0x18\n" - "ldr d9, [x13], #0x8\n" - "ldr d13, [x9], #0x8\n" - "tbz x16, #0, 48f\n" - "ld1 { v9.s }[2], [x13]\n" - "ld1 { v13.s }[2], [x9]\n" - "b 48f\n" - "45:" // Height 2: Partial accumulate: partial_1_4 + "ldr d9, [x28], #0x8\n" + "ldr d13, [x24], #0x8\n" + "tbz x11, #0, 46f\n" + "ld1 { v9.s }[2], [x28]\n" + "ld1 { v13.s }[2], [x24]\n" + "b 46f\n" + "43:" // Height 2: Partial accumulate: partial_1_4 "mov x19, #0x10\n" - "tbz x16, #0, 48f\n" - "ldr s9, [x13, #0x0]\n" - "ldr s13, [x9, #0x0]\n" - "b 48f\n" - "46:" // Height 2: Partial accumulate: partial_2_0 - "tbz x16, #1, 47f\n" - "ldr d8, [x13], #0x8\n" - "ldr d12, [x9], #0x8\n" + "tbz x11, #0, 46f\n" + "ldr s9, [x28, #0x0]\n" + "ldr s13, [x24, #0x0]\n" + "b 46f\n" + "44:" // Height 2: Partial accumulate: partial_2_0 + "tbz x11, #1, 45f\n" + "ldr d8, [x28], #0x8\n" + "ldr d12, [x24], #0x8\n" "mov x19, #0x8\n" - "tbz x16, #0, 48f\n" - "ld1 { v8.s }[2], [x13]\n" - "ld1 { v12.s }[2], [x9]\n" - "b 48f\n" - "47:" // Height 2: Partial accumulate: partial_1_0 + "tbz x11, #0, 46f\n" + "ld1 { v8.s }[2], [x28]\n" + "ld1 { v12.s }[2], [x24]\n" + "b 46f\n" + "45:" // Height 2: Partial accumulate: partial_1_0 + "ldr s8, [x28, #0x0]\n" "mov x19, #0x0\n" - "ldr s8, [x13, #0x0]\n" - "ldr s12, [x9, #0x0]\n" - "48:" // Height 2: Partial accumulate: Done - "sub x13, x13, x19\n" - "sub x9, x9, x19\n" - "b 51f\n" - "49:" // Height 2: full accumulate - "ldr q8, [x13, #0x0]\n" - "ldr q9, [x13, #0x10]\n" - "ldr q10, [x13, #0x20]\n" - "ldr q11, [x13, #0x30]\n" - "ldr q12, [x9, #0x0]\n" - "ldr q13, [x9, #0x10]\n" - "ldr q14, [x9, #0x20]\n" - "ldr q15, [x9, #0x30]\n" - "b 51f\n" - "50:" // Height 2: no accumulate + "ldr s12, [x24, #0x0]\n" + "46:" // Height 2: Partial accumulate: Done + "sub x28, x28, x19\n" + "b 49f\n" + "47:" // Height 2: full accumulate + "ldr q8, [x28, #0x0]\n" + "ldr q9, [x28, #0x10]\n" + "ldr q10, [x28, #0x20]\n" + "ldr q11, [x28, #0x30]\n" + "ldr q12, [x24, #0x0]\n" + "ldr q13, [x24, #0x10]\n" + "ldr q14, [x24, #0x20]\n" + "ldr q15, [x24, #0x30]\n" + "b 49f\n" + "48:" // Height 2: no accumulate "movi v8.16b, #0x0\n" "movi v9.16b, #0x0\n" "movi v10.16b, #0x0\n" @@ -520,195 +506,197 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "movi v13.16b, #0x0\n" "movi v14.16b, #0x0\n" "movi v15.16b, #0x0\n" - "51:" // Height 2: setup done - "mov x12, #0x0\n" - "52:" // Height 2: String loop + "49:" // Height 2: setup done + "mov x27, #0x0\n" + "50:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 53f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 51f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "cbnz x12, 54f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "cbnz x27, 52f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19, LSL #1\n" - "add x28, x28, x19, LSL #1\n" - "b 54f\n" - "53:" // Height 2: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19, LSL #1\n" - "54:" // Height 2: input setup done - "cmp x11, #0x8\n" - "blt 57f\n" - "cmp x11, #0x10\n" - "blt 56f\n" - "55:" // Height 2: Multiply loop: Main loop head - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "add x25, x25, x19, LSL #1\n" + "add x24, x24, x19, LSL #1\n" + "b 52f\n" + "51:" // Height 2: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19, LSL #1\n" + "52:" // Height 2: input setup done + "cmp x26, #0x8\n" + "blt 55f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x10\n" + "ldr q6, [x10, #0x0]\n" + "blt 54f\n" + "53:" // Height 2: Multiply loop: Main loop head ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x15, #0x20]\n" - "add x10, x10, #0x10\n" + "ldr q6, [x10, #0x20]\n" + "add x24, x24, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "cmp x26, #0x10\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "sub x11, x11, #0x8\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x15, #0x40]\n" - "cmp x11, #0x10\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - "ldr q7, [x15, #0xf0]\n" - "add x15, x15, #0x100\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" - "bge 55b\n" - "56:" // Height 2: Multiply loop: Single iteration only - "sub x11, x11, #0x8\n" - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "bge 53b\n" + "54:" // Height 2: Multiply loop: Single iteration only ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x15, #0x20]\n" - "add x10, x10, #0x10\n" + "ldr q6, [x10, #0x20]\n" + "add x25, x25, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x15, #0x40]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - "ldr q7, [x15, #0xf0]\n" - "add x15, x15, #0x100\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" - "57:" // Height 2: Multiply loop: Main loop skip - "cbz x11, 61f\n" - "cmp x11, #0x2\n" - "blt 59f\n" - "58:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x10], #0x4\n" - "ldr s1, [x28], #0x4\n" - "ldr q6, [x15, #0x0]\n" + "55:" // Height 2: Multiply loop: Main loop skip + "cbz x26, 59f\n" + "cmp x26, #0x2\n" + "blt 57f\n" + "56:" // Height 2: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x2\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x2\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x15, #0x20]\n" - "sub x11, x11, #0x2\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "cmp x11, #0x2\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - "bge 58b\n" - "cbz x11, 61f\n" - "59:" // Height 2: Multiply loop: Skip odd blocks - "ldr h0, [x10, #0x0]\n" - "ldr h1, [x28, #0x0]\n" - "60:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q6, [x15, #0x0]\n" + "bge 56b\n" + "cbz x26, 59f\n" + "57:" // Height 2: Multiply loop: Skip odd blocks + "ldr h0, [x25, #0x0]\n" + "ldr h1, [x24, #0x0]\n" + "58:" // Height 2: Multiply loop: Ragged operand read: Done + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - "61:" // Height 2: Multiply loop: No odd multiplies + "59:" // Height 2: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x12, x12, #0x1\n" - "cmp x12, x19\n" - "bne 52b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" - "tbz %x[flags], #1, 62f\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 50b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "prfm pstl1keep, [x28, #0x0]\n" + "add x24, x28, x19, LSL #2\n" + "prfm pstl1keep, [x24, #0x0]\n" + "tbz %x[flags], #1, 60f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1r { v1.4s }, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -729,227 +717,215 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "fmax v14.4s, v14.4s, v1.4s\n" "fmin v15.4s, v15.4s, v0.4s\n" "fmax v15.4s, v15.4s, v1.4s\n" - "62:" // Height 2: No activation - "cmp x16, #0x10\n" - "bge 71f\n" - "tbz x16, #3, 66f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v9.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "st1 { v13.4s }, [x9], #0x10\n" - "tbz x16, #2, 64f\n" - "st1 { v10.4s }, [x13], #0x10\n" - "st1 { v14.4s }, [x9], #0x10\n" - "tbz x16, #1, 63f\n" - "str d11, [x13], #0x8\n" - "str d15, [x9], #0x8\n" - "tbz x16, #0, 70f\n" - "st1 { v11.s }[2], [x13]\n" - "st1 { v15.s }[2], [x9]\n" - "b 70f\n" - "63:" // Height 2: Partial direct writeback: partial_1_12 - "tbz x16, #0, 70f\n" - "str s11, [x13, #0x0]\n" - "str s15, [x9, #0x0]\n" - "b 70f\n" - "64:" // Height 2: Partial direct writeback: partial_2_8 - "tbz x16, #1, 65f\n" - "str d10, [x13], #0x8\n" - "str d14, [x9], #0x8\n" - "tbz x16, #0, 70f\n" - "st1 { v10.s }[2], [x13]\n" - "st1 { v14.s }[2], [x9]\n" - "b 70f\n" - "65:" // Height 2: Partial direct writeback: partial_1_8 - "tbz x16, #0, 70f\n" - "str s10, [x13, #0x0]\n" - "str s14, [x9, #0x0]\n" - "b 70f\n" - "66:" // Height 2: Partial direct writeback: partial_4_0 - "tbz x16, #2, 68f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "tbz x16, #1, 67f\n" - "str d9, [x13], #0x8\n" - "str d13, [x9], #0x8\n" - "tbz x16, #0, 70f\n" - "st1 { v9.s }[2], [x13]\n" - "st1 { v13.s }[2], [x9]\n" - "b 70f\n" - "67:" // Height 2: Partial direct writeback: partial_1_4 - "tbz x16, #0, 70f\n" - "str s9, [x13, #0x0]\n" - "str s13, [x9, #0x0]\n" - "b 70f\n" - "68:" // Height 2: Partial direct writeback: partial_2_0 - "tbz x16, #1, 69f\n" - "str d8, [x13], #0x8\n" - "str d12, [x9], #0x8\n" - "tbz x16, #0, 70f\n" - "st1 { v8.s }[2], [x13]\n" - "st1 { v12.s }[2], [x9]\n" + "60:" // Height 2: No activation + "cmp x11, #0x10\n" + "bge 69f\n" + "tbz x11, #3, 64f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v9.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "st1 { v13.4s }, [x24], #0x10\n" + "tbz x11, #2, 62f\n" + "st1 { v10.4s }, [x28], #0x10\n" + "st1 { v14.4s }, [x24], #0x10\n" + "tbz x11, #1, 61f\n" + "str d11, [x28], #0x8\n" + "str d15, [x24], #0x8\n" + "tbz x11, #0, 68f\n" + "st1 { v11.s }[2], [x28]\n" + "st1 { v15.s }[2], [x24]\n" + "b 68f\n" + "61:" // Height 2: Partial direct writeback: partial_1_12 + "tbz x11, #0, 68f\n" + "str s11, [x28, #0x0]\n" + "str s15, [x24, #0x0]\n" + "b 68f\n" + "62:" // Height 2: Partial direct writeback: partial_2_8 + "tbz x11, #1, 63f\n" + "str d10, [x28], #0x8\n" + "str d14, [x24], #0x8\n" + "tbz x11, #0, 68f\n" + "st1 { v10.s }[2], [x28]\n" + "st1 { v14.s }[2], [x24]\n" + "b 68f\n" + "63:" // Height 2: Partial direct writeback: partial_1_8 + "tbz x11, #0, 68f\n" + "str s10, [x28, #0x0]\n" + "str s14, [x24, #0x0]\n" + "b 68f\n" + "64:" // Height 2: Partial direct writeback: partial_4_0 + "tbz x11, #2, 66f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "tbz x11, #1, 65f\n" + "str d9, [x28], #0x8\n" + "str d13, [x24], #0x8\n" + "tbz x11, #0, 68f\n" + "st1 { v9.s }[2], [x28]\n" + "st1 { v13.s }[2], [x24]\n" + "b 68f\n" + "65:" // Height 2: Partial direct writeback: partial_1_4 + "tbz x11, #0, 68f\n" + "str s9, [x28, #0x0]\n" + "str s13, [x24, #0x0]\n" + "b 68f\n" + "66:" // Height 2: Partial direct writeback: partial_2_0 + "tbz x11, #1, 67f\n" + "str d8, [x28], #0x8\n" + "str d12, [x24], #0x8\n" + "tbz x11, #0, 68f\n" + "st1 { v8.s }[2], [x28]\n" + "st1 { v12.s }[2], [x24]\n" + "b 68f\n" + "67:" // Height 2: Partial direct writeback: partial_1_0 + "str s8, [x28, #0x0]\n" + "str s12, [x24, #0x0]\n" + "68:" // Height 2: Partial direct writeback: Done "b 70f\n" - "69:" // Height 2: Partial direct writeback: partial_1_0 - "str s8, [x13, #0x0]\n" - "str s12, [x9, #0x0]\n" - "70:" // Height 2: Partial direct writeback: Done - "b 72f\n" - "71:" // Height 2: Full writeback - "str q8, [x13, #0x0]\n" - "str q9, [x13, #0x10]\n" - "str q10, [x13, #0x20]\n" - "str q11, [x13, #0x30]\n" - "str q12, [x9, #0x0]\n" - "str q13, [x9, #0x10]\n" - "str q14, [x9, #0x20]\n" - "str q15, [x9, #0x30]\n" - "add x13, x13, #0x40\n" - "add x9, x9, #0x40\n" - "72:" // Height 2: Writeback done - "subs x16, x16, #0x10\n" - "bgt 39b\n" - "b 218f\n" - "73:" // Height 3 - "ldr x16, [%x[args_ptr], %[offsetof_N]]\n" - "mov x14, %x[bias]\n" - "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 74f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19, LSL #2\n" - "add x27, x27, x19, LSL #2\n" - "b 75f\n" - "74:" // Height 3: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19, LSL #2\n" - "add x27, x9, x19, LSL #2\n" - "75:" // Height 3: Column loop - "cbz x14, 76f\n" - "ldr q8, [x14, #0x0]\n" + "69:" // Height 2: Full writeback + "str q8, [x28, #0x0]\n" + "str q9, [x28, #0x10]\n" + "str q10, [x28, #0x20]\n" + "str q11, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + "str q12, [x24, #0x0]\n" + "str q13, [x24, #0x10]\n" + "str q14, [x24, #0x20]\n" + "str q15, [x24, #0x30]\n" + "70:" // Height 2: Writeback done + "subs x11, x11, #0x10\n" + "bgt 37b\n" + "b 212f\n" + "71:" // Height 3 + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "mov x9, %x[bias]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x28, %x[output_ptr]\n" + "72:" // Height 3: Column loop + "cbz x9, 73f\n" + "ldr q8, [x9, #0x0]\n" "mov v12.16b, v8.16b\n" - "ldr q9, [x14, #0x10]\n" + "ldr q9, [x9, #0x10]\n" "mov v16.16b, v8.16b\n" - "ldr q10, [x14, #0x20]\n" - "ldr q11, [x14, #0x30]\n" + "ldr q10, [x9, #0x20]\n" + "ldr q11, [x9, #0x30]\n" "mov v13.16b, v9.16b\n" - "add x14, x14, #0x40\n" + "add x9, x9, #0x40\n" "mov v17.16b, v9.16b\n" "mov v14.16b, v10.16b\n" "mov v15.16b, v11.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" - "b 87f\n" - "76:" // Height 3: no bias - "tbz %x[flags], #0, 86f\n" - "cmp x16, #0x10\n" - "bge 85f\n" - "tbz x16, #3, 80f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "ld1 { v16.4s }, [x27], #0x10\n" - "ld1 { v9.4s }, [x13], #0x10\n" - "ld1 { v13.4s }, [x9], #0x10\n" - "ld1 { v17.4s }, [x27], #0x10\n" - "tbz x16, #2, 78f\n" - "ld1 { v10.4s }, [x13], #0x10\n" - "ld1 { v14.4s }, [x9], #0x10\n" - "ld1 { v18.4s }, [x27], #0x10\n" - "tbz x16, #1, 77f\n" - "mov x19, #0x38\n" - "ldr d11, [x13], #0x8\n" - "ldr d15, [x9], #0x8\n" - "ldr d19, [x27], #0x8\n" - "tbz x16, #0, 84f\n" - "ld1 { v11.s }[2], [x13]\n" - "ld1 { v15.s }[2], [x9]\n" - "ld1 { v19.s }[2], [x27]\n" "b 84f\n" - "77:" // Height 3: Partial accumulate: partial_1_12 + "73:" // Height 3: no bias + "tbz %x[flags], #0, 83f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "cmp x11, #0x10\n" + "add x24, x28, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "bge 82f\n" + "tbz x11, #3, 77f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "ld1 { v16.4s }, [x23], #0x10\n" + "ld1 { v9.4s }, [x28], #0x10\n" + "ld1 { v13.4s }, [x24], #0x10\n" + "ld1 { v17.4s }, [x23], #0x10\n" + "tbz x11, #2, 75f\n" + "ld1 { v10.4s }, [x28], #0x10\n" + "ld1 { v14.4s }, [x24], #0x10\n" + "ld1 { v18.4s }, [x23], #0x10\n" + "tbz x11, #1, 74f\n" + "mov x19, #0x38\n" + "ldr d11, [x28], #0x8\n" + "ldr d15, [x24], #0x8\n" + "ldr d19, [x23], #0x8\n" + "tbz x11, #0, 81f\n" + "ld1 { v11.s }[2], [x28]\n" + "ld1 { v15.s }[2], [x24]\n" + "ld1 { v19.s }[2], [x23]\n" + "b 81f\n" + "74:" // Height 3: Partial accumulate: partial_1_12 "mov x19, #0x30\n" - "tbz x16, #0, 84f\n" - "ldr s11, [x13, #0x0]\n" - "ldr s15, [x9, #0x0]\n" - "ldr s19, [x27, #0x0]\n" - "b 84f\n" - "78:" // Height 3: Partial accumulate: partial_2_8 - "tbz x16, #1, 79f\n" - "ldr d10, [x13], #0x8\n" - "ldr d14, [x9], #0x8\n" - "ldr d18, [x27], #0x8\n" + "tbz x11, #0, 81f\n" + "ldr s11, [x28, #0x0]\n" + "ldr s15, [x24, #0x0]\n" + "ldr s19, [x23, #0x0]\n" + "b 81f\n" + "75:" // Height 3: Partial accumulate: partial_2_8 + "tbz x11, #1, 76f\n" + "ldr d10, [x28], #0x8\n" + "ldr d14, [x24], #0x8\n" "mov x19, #0x28\n" - "tbz x16, #0, 84f\n" - "ld1 { v10.s }[2], [x13]\n" - "ld1 { v14.s }[2], [x9]\n" - "ld1 { v18.s }[2], [x27]\n" - "b 84f\n" - "79:" // Height 3: Partial accumulate: partial_1_8 + "ldr d18, [x23], #0x8\n" + "tbz x11, #0, 81f\n" + "ld1 { v10.s }[2], [x28]\n" + "ld1 { v14.s }[2], [x24]\n" + "ld1 { v18.s }[2], [x23]\n" + "b 81f\n" + "76:" // Height 3: Partial accumulate: partial_1_8 "mov x19, #0x20\n" - "tbz x16, #0, 84f\n" - "ldr s10, [x13, #0x0]\n" - "ldr s14, [x9, #0x0]\n" - "ldr s18, [x27, #0x0]\n" - "b 84f\n" - "80:" // Height 3: Partial accumulate: partial_4_0 - "tbz x16, #2, 82f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "ld1 { v16.4s }, [x27], #0x10\n" - "tbz x16, #1, 81f\n" + "tbz x11, #0, 81f\n" + "ldr s10, [x28, #0x0]\n" + "ldr s14, [x24, #0x0]\n" + "ldr s18, [x23, #0x0]\n" + "b 81f\n" + "77:" // Height 3: Partial accumulate: partial_4_0 + "tbz x11, #2, 79f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "ld1 { v16.4s }, [x23], #0x10\n" + "tbz x11, #1, 78f\n" "mov x19, #0x18\n" - "ldr d9, [x13], #0x8\n" - "ldr d13, [x9], #0x8\n" - "ldr d17, [x27], #0x8\n" - "tbz x16, #0, 84f\n" - "ld1 { v9.s }[2], [x13]\n" - "ld1 { v13.s }[2], [x9]\n" - "ld1 { v17.s }[2], [x27]\n" - "b 84f\n" - "81:" // Height 3: Partial accumulate: partial_1_4 + "ldr d9, [x28], #0x8\n" + "ldr d13, [x24], #0x8\n" + "ldr d17, [x23], #0x8\n" + "tbz x11, #0, 81f\n" + "ld1 { v9.s }[2], [x28]\n" + "ld1 { v13.s }[2], [x24]\n" + "ld1 { v17.s }[2], [x23]\n" + "b 81f\n" + "78:" // Height 3: Partial accumulate: partial_1_4 "mov x19, #0x10\n" - "tbz x16, #0, 84f\n" - "ldr s9, [x13, #0x0]\n" - "ldr s13, [x9, #0x0]\n" - "ldr s17, [x27, #0x0]\n" - "b 84f\n" - "82:" // Height 3: Partial accumulate: partial_2_0 - "tbz x16, #1, 83f\n" - "ldr d8, [x13], #0x8\n" - "ldr d12, [x9], #0x8\n" - "ldr d16, [x27], #0x8\n" + "tbz x11, #0, 81f\n" + "ldr s9, [x28, #0x0]\n" + "ldr s13, [x24, #0x0]\n" + "ldr s17, [x23, #0x0]\n" + "b 81f\n" + "79:" // Height 3: Partial accumulate: partial_2_0 + "tbz x11, #1, 80f\n" + "ldr d8, [x28], #0x8\n" + "ldr d12, [x24], #0x8\n" "mov x19, #0x8\n" - "tbz x16, #0, 84f\n" - "ld1 { v8.s }[2], [x13]\n" - "ld1 { v12.s }[2], [x9]\n" - "ld1 { v16.s }[2], [x27]\n" - "b 84f\n" - "83:" // Height 3: Partial accumulate: partial_1_0 + "ldr d16, [x23], #0x8\n" + "tbz x11, #0, 81f\n" + "ld1 { v8.s }[2], [x28]\n" + "ld1 { v12.s }[2], [x24]\n" + "ld1 { v16.s }[2], [x23]\n" + "b 81f\n" + "80:" // Height 3: Partial accumulate: partial_1_0 + "ldr s8, [x28, #0x0]\n" "mov x19, #0x0\n" - "ldr s8, [x13, #0x0]\n" - "ldr s12, [x9, #0x0]\n" - "ldr s16, [x27, #0x0]\n" - "84:" // Height 3: Partial accumulate: Done - "sub x13, x13, x19\n" - "sub x9, x9, x19\n" - "sub x27, x27, x19\n" - "b 87f\n" - "85:" // Height 3: full accumulate - "ldr q8, [x13, #0x0]\n" - "ldr q9, [x13, #0x10]\n" - "ldr q10, [x13, #0x20]\n" - "ldr q11, [x13, #0x30]\n" - "ldr q12, [x9, #0x0]\n" - "ldr q13, [x9, #0x10]\n" - "ldr q14, [x9, #0x20]\n" - "ldr q15, [x9, #0x30]\n" - "ldr q16, [x27, #0x0]\n" - "ldr q17, [x27, #0x10]\n" - "ldr q18, [x27, #0x20]\n" - "ldr q19, [x27, #0x30]\n" - "b 87f\n" - "86:" // Height 3: no accumulate + "ldr s12, [x24, #0x0]\n" + "ldr s16, [x23, #0x0]\n" + "81:" // Height 3: Partial accumulate: Done + "sub x28, x28, x19\n" + "b 84f\n" + "82:" // Height 3: full accumulate + "ldr q8, [x28, #0x0]\n" + "ldr q9, [x28, #0x10]\n" + "ldr q10, [x28, #0x20]\n" + "ldr q11, [x28, #0x30]\n" + "ldr q12, [x24, #0x0]\n" + "ldr q13, [x24, #0x10]\n" + "ldr q14, [x24, #0x20]\n" + "ldr q15, [x24, #0x30]\n" + "ldr q16, [x23, #0x0]\n" + "ldr q17, [x23, #0x10]\n" + "ldr q18, [x23, #0x20]\n" + "ldr q19, [x23, #0x30]\n" + "b 84f\n" + "83:" // Height 3: no accumulate "movi v8.16b, #0x0\n" "movi v9.16b, #0x0\n" "movi v10.16b, #0x0\n" @@ -962,247 +938,250 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "movi v17.16b, #0x0\n" "movi v18.16b, #0x0\n" "movi v19.16b, #0x0\n" - "87:" // Height 3: setup done - "mov x12, #0x0\n" - "88:" // Height 3: String loop + "84:" // Height 3: setup done + "mov x27, #0x0\n" + "85:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 89f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 86f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x26, [x20, #0x10]\n" - "cbnz x12, 90f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "cbnz x27, 87f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19, LSL #1\n" - "add x28, x28, x19, LSL #1\n" - "add x26, x26, x19, LSL #1\n" - "b 90f\n" - "89:" // Height 3: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19, LSL #1\n" - "add x26, x28, x19, LSL #1\n" - "90:" // Height 3: input setup done - "cmp x11, #0x8\n" - "blt 93f\n" - "cmp x11, #0x10\n" - "blt 92f\n" - "91:" // Height 3: Multiply loop: Main loop head - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "add x25, x25, x19, LSL #1\n" + "add x24, x24, x19, LSL #1\n" + "add x23, x23, x19, LSL #1\n" + "b 87f\n" + "86:" // Height 3: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19, LSL #1\n" + "add x23, x24, x19, LSL #1\n" + "87:" // Height 3: input setup done + "cmp x26, #0x8\n" + "blt 90f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x10\n" + "ldr q2, [x23, #0x0]\n" + "ldr q6, [x10, #0x0]\n" + "blt 89f\n" + "88:" // Height 3: Multiply loop: Main loop head ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" + "add x23, x23, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "add x28, x28, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" + "cmp x26, #0x10\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "sub x11, x11, #0x8\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - "cmp x11, #0x10\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x15, #0x40]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" - "ldr q7, [x15, #0xf0]\n" - "add x15, x15, #0x100\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" + "ldr q1, [x24, #0x0]\n" ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" - "bge 91b\n" - "92:" // Height 3: Multiply loop: Single iteration only - "sub x11, x11, #0x8\n" - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "ldr q2, [x23, #0x0]\n" + "bge 88b\n" + "89:" // Height 3: Multiply loop: Single iteration only ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "add x24, x24, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "add x28, x28, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" + "ldr q6, [x10, #0x20]\n" + "add x23, x23, #0x10\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q7, [x15, #0x30]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x15, #0x40]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" - "ldr q7, [x15, #0xf0]\n" - "add x15, x15, #0x100\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" - "93:" // Height 3: Multiply loop: Main loop skip - "cbz x11, 97f\n" - "cmp x11, #0x2\n" - "blt 95f\n" - "94:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x10], #0x4\n" - "ldr s1, [x28], #0x4\n" - "ldr s2, [x26], #0x4\n" - "ldr q6, [x15, #0x0]\n" + "90:" // Height 3: Multiply loop: Main loop skip + "cbz x26, 94f\n" + "cmp x26, #0x2\n" + "blt 92f\n" + "91:" // Height 3: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x2\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x2\n" + "ldr s2, [x23], #0x4\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "sub x11, x11, #0x2\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x15, #0x20]\n" - "cmp x11, #0x2\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - "bge 94b\n" - "cbz x11, 97f\n" - "95:" // Height 3: Multiply loop: Skip odd blocks - "ldr h0, [x10, #0x0]\n" - "ldr h1, [x28, #0x0]\n" - "ldr h2, [x26, #0x0]\n" - "96:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q6, [x15, #0x0]\n" + "bge 91b\n" + "cbz x26, 94f\n" + "92:" // Height 3: Multiply loop: Skip odd blocks + "ldr h0, [x25, #0x0]\n" + "ldr h1, [x24, #0x0]\n" + "ldr h2, [x23, #0x0]\n" + "93:" // Height 3: Multiply loop: Ragged operand read: Done + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - "97:" // Height 3: Multiply loop: No odd multiplies + "94:" // Height 3: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x12, x12, #0x1\n" - "cmp x12, x19\n" - "bne 88b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x27, #0x0]\n" - "tbz %x[flags], #1, 98f\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 85b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "prfm pstl1keep, [x28, #0x0]\n" + "add x24, x28, x19, LSL #2\n" + "prfm pstl1keep, [x24, #0x0]\n" + "add x23, x24, x19, LSL #2\n" + "prfm pstl1keep, [x23, #0x0]\n" + "tbz %x[flags], #1, 95f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1r { v1.4s }, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -1231,137 +1210,120 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "fmin v19.4s, v19.4s, v0.4s\n" "fmax v18.4s, v18.4s, v1.4s\n" "fmax v19.4s, v19.4s, v1.4s\n" - "98:" // Height 3: No activation - "cmp x16, #0x10\n" - "bge 107f\n" - "tbz x16, #3, 102f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v9.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "st1 { v13.4s }, [x9], #0x10\n" - "st1 { v16.4s }, [x27], #0x10\n" - "st1 { v17.4s }, [x27], #0x10\n" - "tbz x16, #2, 100f\n" - "st1 { v10.4s }, [x13], #0x10\n" - "st1 { v14.4s }, [x9], #0x10\n" - "st1 { v18.4s }, [x27], #0x10\n" - "tbz x16, #1, 99f\n" - "str d11, [x13], #0x8\n" - "str d15, [x9], #0x8\n" - "str d19, [x27], #0x8\n" - "tbz x16, #0, 106f\n" - "st1 { v11.s }[2], [x13]\n" - "st1 { v15.s }[2], [x9]\n" - "st1 { v19.s }[2], [x27]\n" - "b 106f\n" - "99:" // Height 3: Partial direct writeback: partial_1_12 - "tbz x16, #0, 106f\n" - "str s11, [x13, #0x0]\n" - "str s15, [x9, #0x0]\n" - "str s19, [x27, #0x0]\n" - "b 106f\n" - "100:" // Height 3: Partial direct writeback: partial_2_8 - "tbz x16, #1, 101f\n" - "str d10, [x13], #0x8\n" - "str d14, [x9], #0x8\n" - "str d18, [x27], #0x8\n" - "tbz x16, #0, 106f\n" - "st1 { v10.s }[2], [x13]\n" - "st1 { v14.s }[2], [x9]\n" - "st1 { v18.s }[2], [x27]\n" - "b 106f\n" - "101:" // Height 3: Partial direct writeback: partial_1_8 - "tbz x16, #0, 106f\n" - "str s10, [x13, #0x0]\n" - "str s14, [x9, #0x0]\n" - "str s18, [x27, #0x0]\n" - "b 106f\n" - "102:" // Height 3: Partial direct writeback: partial_4_0 - "tbz x16, #2, 104f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "st1 { v16.4s }, [x27], #0x10\n" - "tbz x16, #1, 103f\n" - "str d9, [x13], #0x8\n" - "str d13, [x9], #0x8\n" - "str d17, [x27], #0x8\n" - "tbz x16, #0, 106f\n" - "st1 { v9.s }[2], [x13]\n" - "st1 { v13.s }[2], [x9]\n" - "st1 { v17.s }[2], [x27]\n" - "b 106f\n" - "103:" // Height 3: Partial direct writeback: partial_1_4 - "tbz x16, #0, 106f\n" - "str s9, [x13, #0x0]\n" - "str s13, [x9, #0x0]\n" - "str s17, [x27, #0x0]\n" - "b 106f\n" - "104:" // Height 3: Partial direct writeback: partial_2_0 - "tbz x16, #1, 105f\n" - "str d8, [x13], #0x8\n" - "str d12, [x9], #0x8\n" - "str d16, [x27], #0x8\n" - "tbz x16, #0, 106f\n" - "st1 { v8.s }[2], [x13]\n" - "st1 { v12.s }[2], [x9]\n" - "st1 { v16.s }[2], [x27]\n" - "b 106f\n" - "105:" // Height 3: Partial direct writeback: partial_1_0 - "str s8, [x13, #0x0]\n" - "str s12, [x9, #0x0]\n" - "str s16, [x27, #0x0]\n" - "106:" // Height 3: Partial direct writeback: Done - "b 108f\n" - "107:" // Height 3: Full writeback - "str q8, [x13, #0x0]\n" - "str q9, [x13, #0x10]\n" - "str q10, [x13, #0x20]\n" - "str q11, [x13, #0x30]\n" - "str q12, [x9, #0x0]\n" - "str q13, [x9, #0x10]\n" - "str q14, [x9, #0x20]\n" - "str q15, [x9, #0x30]\n" - "str q16, [x27, #0x0]\n" - "str q17, [x27, #0x10]\n" - "str q18, [x27, #0x20]\n" - "str q19, [x27, #0x30]\n" - "add x13, x13, #0x40\n" - "add x9, x9, #0x40\n" - "add x27, x27, #0x40\n" - "108:" // Height 3: Writeback done - "subs x16, x16, #0x10\n" - "bgt 75b\n" - "b 218f\n" - "109:" // Height 4 - "ldr x16, [%x[args_ptr], %[offsetof_N]]\n" - "mov x14, %x[bias]\n" - "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 110f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19, LSL #2\n" - "ldr x25, [%x[output_ptr], #0x18]\n" - "add x27, x27, x19, LSL #2\n" - "add x25, x25, x19, LSL #2\n" - "b 111f\n" - "110:" // Height 4: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19, LSL #2\n" - "add x27, x9, x19, LSL #2\n" - "add x25, x27, x19, LSL #2\n" - "111:" // Height 4: Column loop - "cbz x14, 112f\n" - "ldr q8, [x14, #0x0]\n" + "95:" // Height 3: No activation + "cmp x11, #0x10\n" + "bge 104f\n" + "tbz x11, #3, 99f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v9.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "st1 { v13.4s }, [x24], #0x10\n" + "st1 { v16.4s }, [x23], #0x10\n" + "st1 { v17.4s }, [x23], #0x10\n" + "tbz x11, #2, 97f\n" + "st1 { v10.4s }, [x28], #0x10\n" + "st1 { v14.4s }, [x24], #0x10\n" + "st1 { v18.4s }, [x23], #0x10\n" + "tbz x11, #1, 96f\n" + "str d11, [x28], #0x8\n" + "str d15, [x24], #0x8\n" + "str d19, [x23], #0x8\n" + "tbz x11, #0, 103f\n" + "st1 { v11.s }[2], [x28]\n" + "st1 { v15.s }[2], [x24]\n" + "st1 { v19.s }[2], [x23]\n" + "b 103f\n" + "96:" // Height 3: Partial direct writeback: partial_1_12 + "tbz x11, #0, 103f\n" + "str s11, [x28, #0x0]\n" + "str s15, [x24, #0x0]\n" + "str s19, [x23, #0x0]\n" + "b 103f\n" + "97:" // Height 3: Partial direct writeback: partial_2_8 + "tbz x11, #1, 98f\n" + "str d10, [x28], #0x8\n" + "str d14, [x24], #0x8\n" + "str d18, [x23], #0x8\n" + "tbz x11, #0, 103f\n" + "st1 { v10.s }[2], [x28]\n" + "st1 { v14.s }[2], [x24]\n" + "st1 { v18.s }[2], [x23]\n" + "b 103f\n" + "98:" // Height 3: Partial direct writeback: partial_1_8 + "tbz x11, #0, 103f\n" + "str s10, [x28, #0x0]\n" + "str s14, [x24, #0x0]\n" + "str s18, [x23, #0x0]\n" + "b 103f\n" + "99:" // Height 3: Partial direct writeback: partial_4_0 + "tbz x11, #2, 101f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "st1 { v16.4s }, [x23], #0x10\n" + "tbz x11, #1, 100f\n" + "str d9, [x28], #0x8\n" + "str d13, [x24], #0x8\n" + "str d17, [x23], #0x8\n" + "tbz x11, #0, 103f\n" + "st1 { v9.s }[2], [x28]\n" + "st1 { v13.s }[2], [x24]\n" + "st1 { v17.s }[2], [x23]\n" + "b 103f\n" + "100:" // Height 3: Partial direct writeback: partial_1_4 + "tbz x11, #0, 103f\n" + "str s9, [x28, #0x0]\n" + "str s13, [x24, #0x0]\n" + "str s17, [x23, #0x0]\n" + "b 103f\n" + "101:" // Height 3: Partial direct writeback: partial_2_0 + "tbz x11, #1, 102f\n" + "str d8, [x28], #0x8\n" + "str d12, [x24], #0x8\n" + "str d16, [x23], #0x8\n" + "tbz x11, #0, 103f\n" + "st1 { v8.s }[2], [x28]\n" + "st1 { v12.s }[2], [x24]\n" + "st1 { v16.s }[2], [x23]\n" + "b 103f\n" + "102:" // Height 3: Partial direct writeback: partial_1_0 + "str s8, [x28, #0x0]\n" + "str s12, [x24, #0x0]\n" + "str s16, [x23, #0x0]\n" + "103:" // Height 3: Partial direct writeback: Done + "b 105f\n" + "104:" // Height 3: Full writeback + "str q8, [x28, #0x0]\n" + "str q9, [x28, #0x10]\n" + "str q10, [x28, #0x20]\n" + "str q11, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + "str q12, [x24, #0x0]\n" + "str q13, [x24, #0x10]\n" + "str q14, [x24, #0x20]\n" + "str q15, [x24, #0x30]\n" + "str q16, [x23, #0x0]\n" + "str q17, [x23, #0x10]\n" + "str q18, [x23, #0x20]\n" + "str q19, [x23, #0x30]\n" + "105:" // Height 3: Writeback done + "subs x11, x11, #0x10\n" + "bgt 72b\n" + "b 212f\n" + "106:" // Height 4 + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "mov x9, %x[bias]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x28, %x[output_ptr]\n" + "107:" // Height 4: Column loop + "cbz x9, 108f\n" + "ldr q8, [x9, #0x0]\n" "mov v12.16b, v8.16b\n" - "ldr q9, [x14, #0x10]\n" + "ldr q9, [x9, #0x10]\n" "mov v16.16b, v8.16b\n" - "ldr q10, [x14, #0x20]\n" + "ldr q10, [x9, #0x20]\n" "mov v20.16b, v8.16b\n" - "ldr q11, [x14, #0x30]\n" - "add x14, x14, #0x40\n" + "ldr q11, [x9, #0x30]\n" + "add x9, x9, #0x40\n" "mov v13.16b, v9.16b\n" "mov v17.16b, v9.16b\n" "mov v14.16b, v10.16b\n" @@ -1371,136 +1333,137 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "mov v21.16b, v9.16b\n" "mov v22.16b, v10.16b\n" "mov v23.16b, v11.16b\n" - "b 123f\n" - "112:" // Height 4: no bias - "tbz %x[flags], #0, 122f\n" - "cmp x16, #0x10\n" - "bge 121f\n" - "tbz x16, #3, 116f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "ld1 { v16.4s }, [x27], #0x10\n" - "ld1 { v20.4s }, [x25], #0x10\n" - "ld1 { v9.4s }, [x13], #0x10\n" - "ld1 { v13.4s }, [x9], #0x10\n" - "ld1 { v17.4s }, [x27], #0x10\n" - "ld1 { v21.4s }, [x25], #0x10\n" - "tbz x16, #2, 114f\n" - "ld1 { v10.4s }, [x13], #0x10\n" - "ld1 { v14.4s }, [x9], #0x10\n" - "ld1 { v18.4s }, [x27], #0x10\n" - "ld1 { v22.4s }, [x25], #0x10\n" - "tbz x16, #1, 113f\n" + "b 119f\n" + "108:" // Height 4: no bias + "tbz %x[flags], #0, 118f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "cmp x11, #0x10\n" + "add x24, x28, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "add x22, x23, x19, LSL #2\n" + "bge 117f\n" + "tbz x11, #3, 112f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "ld1 { v16.4s }, [x23], #0x10\n" + "ld1 { v20.4s }, [x22], #0x10\n" + "ld1 { v9.4s }, [x28], #0x10\n" + "ld1 { v13.4s }, [x24], #0x10\n" + "ld1 { v17.4s }, [x23], #0x10\n" + "ld1 { v21.4s }, [x22], #0x10\n" + "tbz x11, #2, 110f\n" + "ld1 { v10.4s }, [x28], #0x10\n" + "ld1 { v14.4s }, [x24], #0x10\n" + "ld1 { v18.4s }, [x23], #0x10\n" + "ld1 { v22.4s }, [x22], #0x10\n" + "tbz x11, #1, 109f\n" "mov x19, #0x38\n" - "ldr d11, [x13], #0x8\n" - "ldr d15, [x9], #0x8\n" - "ldr d19, [x27], #0x8\n" - "ldr d23, [x25], #0x8\n" - "tbz x16, #0, 120f\n" - "ld1 { v11.s }[2], [x13]\n" - "ld1 { v15.s }[2], [x9]\n" - "ld1 { v19.s }[2], [x27]\n" - "ld1 { v23.s }[2], [x25]\n" - "b 120f\n" - "113:" // Height 4: Partial accumulate: partial_1_12 + "ldr d11, [x28], #0x8\n" + "ldr d15, [x24], #0x8\n" + "ldr d19, [x23], #0x8\n" + "ldr d23, [x22], #0x8\n" + "tbz x11, #0, 116f\n" + "ld1 { v11.s }[2], [x28]\n" + "ld1 { v15.s }[2], [x24]\n" + "ld1 { v19.s }[2], [x23]\n" + "ld1 { v23.s }[2], [x22]\n" + "b 116f\n" + "109:" // Height 4: Partial accumulate: partial_1_12 "mov x19, #0x30\n" - "tbz x16, #0, 120f\n" - "ldr s11, [x13, #0x0]\n" - "ldr s15, [x9, #0x0]\n" - "ldr s19, [x27, #0x0]\n" - "ldr s23, [x25, #0x0]\n" - "b 120f\n" - "114:" // Height 4: Partial accumulate: partial_2_8 - "tbz x16, #1, 115f\n" - "ldr d10, [x13], #0x8\n" - "ldr d14, [x9], #0x8\n" - "ldr d18, [x27], #0x8\n" - "ldr d22, [x25], #0x8\n" + "tbz x11, #0, 116f\n" + "ldr s11, [x28, #0x0]\n" + "ldr s15, [x24, #0x0]\n" + "ldr s19, [x23, #0x0]\n" + "ldr s23, [x22, #0x0]\n" + "b 116f\n" + "110:" // Height 4: Partial accumulate: partial_2_8 + "tbz x11, #1, 111f\n" + "ldr d10, [x28], #0x8\n" + "ldr d14, [x24], #0x8\n" "mov x19, #0x28\n" - "tbz x16, #0, 120f\n" - "ld1 { v10.s }[2], [x13]\n" - "ld1 { v14.s }[2], [x9]\n" - "ld1 { v18.s }[2], [x27]\n" - "ld1 { v22.s }[2], [x25]\n" - "b 120f\n" - "115:" // Height 4: Partial accumulate: partial_1_8 + "ldr d18, [x23], #0x8\n" + "ldr d22, [x22], #0x8\n" + "tbz x11, #0, 116f\n" + "ld1 { v10.s }[2], [x28]\n" + "ld1 { v14.s }[2], [x24]\n" + "ld1 { v18.s }[2], [x23]\n" + "ld1 { v22.s }[2], [x22]\n" + "b 116f\n" + "111:" // Height 4: Partial accumulate: partial_1_8 "mov x19, #0x20\n" - "tbz x16, #0, 120f\n" - "ldr s10, [x13, #0x0]\n" - "ldr s14, [x9, #0x0]\n" - "ldr s18, [x27, #0x0]\n" - "ldr s22, [x25, #0x0]\n" - "b 120f\n" - "116:" // Height 4: Partial accumulate: partial_4_0 - "tbz x16, #2, 118f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "ld1 { v16.4s }, [x27], #0x10\n" - "ld1 { v20.4s }, [x25], #0x10\n" - "tbz x16, #1, 117f\n" + "tbz x11, #0, 116f\n" + "ldr s10, [x28, #0x0]\n" + "ldr s14, [x24, #0x0]\n" + "ldr s18, [x23, #0x0]\n" + "ldr s22, [x22, #0x0]\n" + "b 116f\n" + "112:" // Height 4: Partial accumulate: partial_4_0 + "tbz x11, #2, 114f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "ld1 { v16.4s }, [x23], #0x10\n" + "ld1 { v20.4s }, [x22], #0x10\n" + "tbz x11, #1, 113f\n" "mov x19, #0x18\n" - "ldr d9, [x13], #0x8\n" - "ldr d13, [x9], #0x8\n" - "ldr d17, [x27], #0x8\n" - "ldr d21, [x25], #0x8\n" - "tbz x16, #0, 120f\n" - "ld1 { v9.s }[2], [x13]\n" - "ld1 { v13.s }[2], [x9]\n" - "ld1 { v17.s }[2], [x27]\n" - "ld1 { v21.s }[2], [x25]\n" - "b 120f\n" - "117:" // Height 4: Partial accumulate: partial_1_4 + "ldr d9, [x28], #0x8\n" + "ldr d13, [x24], #0x8\n" + "ldr d17, [x23], #0x8\n" + "ldr d21, [x22], #0x8\n" + "tbz x11, #0, 116f\n" + "ld1 { v9.s }[2], [x28]\n" + "ld1 { v13.s }[2], [x24]\n" + "ld1 { v17.s }[2], [x23]\n" + "ld1 { v21.s }[2], [x22]\n" + "b 116f\n" + "113:" // Height 4: Partial accumulate: partial_1_4 "mov x19, #0x10\n" - "tbz x16, #0, 120f\n" - "ldr s9, [x13, #0x0]\n" - "ldr s13, [x9, #0x0]\n" - "ldr s17, [x27, #0x0]\n" - "ldr s21, [x25, #0x0]\n" - "b 120f\n" - "118:" // Height 4: Partial accumulate: partial_2_0 - "tbz x16, #1, 119f\n" - "ldr d8, [x13], #0x8\n" - "ldr d12, [x9], #0x8\n" - "ldr d16, [x27], #0x8\n" - "ldr d20, [x25], #0x8\n" + "tbz x11, #0, 116f\n" + "ldr s9, [x28, #0x0]\n" + "ldr s13, [x24, #0x0]\n" + "ldr s17, [x23, #0x0]\n" + "ldr s21, [x22, #0x0]\n" + "b 116f\n" + "114:" // Height 4: Partial accumulate: partial_2_0 + "tbz x11, #1, 115f\n" + "ldr d8, [x28], #0x8\n" + "ldr d12, [x24], #0x8\n" "mov x19, #0x8\n" - "tbz x16, #0, 120f\n" - "ld1 { v8.s }[2], [x13]\n" - "ld1 { v12.s }[2], [x9]\n" - "ld1 { v16.s }[2], [x27]\n" - "ld1 { v20.s }[2], [x25]\n" - "b 120f\n" - "119:" // Height 4: Partial accumulate: partial_1_0 + "ldr d16, [x23], #0x8\n" + "ldr d20, [x22], #0x8\n" + "tbz x11, #0, 116f\n" + "ld1 { v8.s }[2], [x28]\n" + "ld1 { v12.s }[2], [x24]\n" + "ld1 { v16.s }[2], [x23]\n" + "ld1 { v20.s }[2], [x22]\n" + "b 116f\n" + "115:" // Height 4: Partial accumulate: partial_1_0 + "ldr s8, [x28, #0x0]\n" "mov x19, #0x0\n" - "ldr s8, [x13, #0x0]\n" - "ldr s12, [x9, #0x0]\n" - "ldr s16, [x27, #0x0]\n" - "ldr s20, [x25, #0x0]\n" - "120:" // Height 4: Partial accumulate: Done - "sub x13, x13, x19\n" - "sub x9, x9, x19\n" - "sub x27, x27, x19\n" - "sub x25, x25, x19\n" - "b 123f\n" - "121:" // Height 4: full accumulate - "ldr q8, [x13, #0x0]\n" - "ldr q9, [x13, #0x10]\n" - "ldr q10, [x13, #0x20]\n" - "ldr q11, [x13, #0x30]\n" - "ldr q12, [x9, #0x0]\n" - "ldr q13, [x9, #0x10]\n" - "ldr q14, [x9, #0x20]\n" - "ldr q15, [x9, #0x30]\n" - "ldr q16, [x27, #0x0]\n" - "ldr q17, [x27, #0x10]\n" - "ldr q18, [x27, #0x20]\n" - "ldr q19, [x27, #0x30]\n" - "ldr q20, [x25, #0x0]\n" - "ldr q21, [x25, #0x10]\n" - "ldr q22, [x25, #0x20]\n" - "ldr q23, [x25, #0x30]\n" - "b 123f\n" - "122:" // Height 4: no accumulate + "ldr s12, [x24, #0x0]\n" + "ldr s16, [x23, #0x0]\n" + "ldr s20, [x22, #0x0]\n" + "116:" // Height 4: Partial accumulate: Done + "sub x28, x28, x19\n" + "b 119f\n" + "117:" // Height 4: full accumulate + "ldr q8, [x28, #0x0]\n" + "ldr q9, [x28, #0x10]\n" + "ldr q10, [x28, #0x20]\n" + "ldr q11, [x28, #0x30]\n" + "ldr q12, [x24, #0x0]\n" + "ldr q13, [x24, #0x10]\n" + "ldr q14, [x24, #0x20]\n" + "ldr q15, [x24, #0x30]\n" + "ldr q16, [x23, #0x0]\n" + "ldr q17, [x23, #0x10]\n" + "ldr q18, [x23, #0x20]\n" + "ldr q19, [x23, #0x30]\n" + "ldr q20, [x22, #0x0]\n" + "ldr q21, [x22, #0x10]\n" + "ldr q22, [x22, #0x20]\n" + "ldr q23, [x22, #0x30]\n" + "b 119f\n" + "118:" // Height 4: no accumulate "movi v8.16b, #0x0\n" "movi v9.16b, #0x0\n" "movi v10.16b, #0x0\n" @@ -1517,220 +1480,220 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "movi v21.16b, #0x0\n" "movi v22.16b, #0x0\n" "movi v23.16b, #0x0\n" - "123:" // Height 4: setup done - "mov x12, #0x0\n" - "124:" // Height 4: String loop + "119:" // Height 4: setup done + "mov x27, #0x0\n" + "120:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 125f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 121f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x26, [x20, #0x10]\n" - "ldr x24, [x20, #0x18]\n" - "cbnz x12, 126f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "ldr x22, [x20, #0x18]\n" + "cbnz x27, 122f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19, LSL #1\n" - "add x28, x28, x19, LSL #1\n" - "add x26, x26, x19, LSL #1\n" + "add x25, x25, x19, LSL #1\n" "add x24, x24, x19, LSL #1\n" - "b 126f\n" - "125:" // Height 4: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19, LSL #1\n" - "add x26, x28, x19, LSL #1\n" - "add x24, x26, x19, LSL #1\n" - "126:" // Height 4: input setup done - "cmp x11, #0x8\n" - "blt 129f\n" - "cmp x11, #0x10\n" - "blt 128f\n" - "127:" // Height 4: Multiply loop: Main loop head - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q3, [x24, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "add x23, x23, x19, LSL #1\n" + "add x22, x22, x19, LSL #1\n" + "b 122f\n" + "121:" // Height 4: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19, LSL #1\n" + "add x23, x24, x19, LSL #1\n" + "add x22, x23, x19, LSL #1\n" + "122:" // Height 4: input setup done + "cmp x26, #0x8\n" + "blt 125f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x10\n" + "ldr q2, [x23, #0x0]\n" + "ldr q3, [x22, #0x0]\n" + "ldr q6, [x10, #0x0]\n" + "blt 124f\n" + "123:" // Height 4: Multiply loop: Main loop head ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" + "add x22, x22, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x22, #0x80]\n" + "cmp x26, #0x10\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "sub x11, x11, #0x8\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "cmp x11, #0x10\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x15, #0x40]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" - "ldr q7, [x15, #0xf0]\n" - "add x15, x15, #0x100\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d6 // bfdot v22.4s, v6.8h, v3.h[3]\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" + "ldr q1, [x24, #0x0]\n" ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" + "ldr q2, [x23, #0x0]\n" ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" - "bge 127b\n" - "128:" // Height 4: Multiply loop: Single iteration only - "sub x11, x11, #0x8\n" - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q3, [x24, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "ldr q3, [x22, #0x0]\n" + "bge 123b\n" + "124:" // Height 4: Multiply loop: Single iteration only ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "add x28, x28, #0x10\n" + "add x24, x24, #0x10\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "ldr q6, [x15, #0x20]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" + "ldr q6, [x10, #0x20]\n" + "add x22, x22, #0x10\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x15, #0x40]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" - "ldr q7, [x15, #0xf0]\n" - "add x15, x15, #0x100\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" @@ -1739,31 +1702,31 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" - "129:" // Height 4: Multiply loop: Main loop skip - "cbz x11, 133f\n" - "cmp x11, #0x2\n" - "blt 131f\n" - "130:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x10], #0x4\n" - "ldr s1, [x28], #0x4\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x24], #0x4\n" - "ldr q6, [x15, #0x0]\n" + "125:" // Height 4: Multiply loop: Main loop skip + "cbz x26, 129f\n" + "cmp x26, #0x2\n" + "blt 127f\n" + "126:" // Height 4: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x2\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x2\n" + "ldr s2, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "sub x11, x11, #0x2\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "cmp x11, #0x2\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" @@ -1771,28 +1734,28 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - "bge 130b\n" - "cbz x11, 133f\n" - "131:" // Height 4: Multiply loop: Skip odd blocks - "ldr h0, [x10, #0x0]\n" - "ldr h1, [x28, #0x0]\n" - "ldr h2, [x26, #0x0]\n" - "ldr h3, [x24, #0x0]\n" - "132:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q6, [x15, #0x0]\n" + "bge 126b\n" + "cbz x26, 129f\n" + "127:" // Height 4: Multiply loop: Skip odd blocks + "ldr h0, [x25, #0x0]\n" + "ldr h1, [x24, #0x0]\n" + "ldr h2, [x23, #0x0]\n" + "ldr h3, [x22, #0x0]\n" + "128:" // Height 4: Multiply loop: Ragged operand read: Done + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" @@ -1800,16 +1763,20 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - "133:" // Height 4: Multiply loop: No odd multiplies + "129:" // Height 4: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x12, x12, #0x1\n" - "cmp x12, x19\n" - "bne 124b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x27, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" - "tbz %x[flags], #1, 134f\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 120b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "prfm pstl1keep, [x28, #0x0]\n" + "add x24, x28, x19, LSL #2\n" + "prfm pstl1keep, [x24, #0x0]\n" + "add x23, x24, x19, LSL #2\n" + "prfm pstl1keep, [x23, #0x0]\n" + "add x22, x23, x19, LSL #2\n" + "prfm pstl1keep, [x22, #0x0]\n" + "tbz %x[flags], #1, 130f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1r { v1.4s }, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -1846,162 +1813,141 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "fmax v21.4s, v21.4s, v1.4s\n" "fmax v22.4s, v22.4s, v1.4s\n" "fmax v23.4s, v23.4s, v1.4s\n" - "134:" // Height 4: No activation - "cmp x16, #0x10\n" - "bge 143f\n" - "tbz x16, #3, 138f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v9.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "st1 { v13.4s }, [x9], #0x10\n" - "st1 { v16.4s }, [x27], #0x10\n" - "st1 { v17.4s }, [x27], #0x10\n" - "st1 { v20.4s }, [x25], #0x10\n" - "st1 { v21.4s }, [x25], #0x10\n" - "tbz x16, #2, 136f\n" - "st1 { v10.4s }, [x13], #0x10\n" - "st1 { v14.4s }, [x9], #0x10\n" - "st1 { v18.4s }, [x27], #0x10\n" - "st1 { v22.4s }, [x25], #0x10\n" - "tbz x16, #1, 135f\n" - "str d11, [x13], #0x8\n" - "str d15, [x9], #0x8\n" - "str d19, [x27], #0x8\n" - "str d23, [x25], #0x8\n" - "tbz x16, #0, 142f\n" - "st1 { v11.s }[2], [x13]\n" - "st1 { v15.s }[2], [x9]\n" - "st1 { v19.s }[2], [x27]\n" - "st1 { v23.s }[2], [x25]\n" - "b 142f\n" - "135:" // Height 4: Partial direct writeback: partial_1_12 - "tbz x16, #0, 142f\n" - "str s11, [x13, #0x0]\n" - "str s15, [x9, #0x0]\n" - "str s19, [x27, #0x0]\n" - "str s23, [x25, #0x0]\n" - "b 142f\n" - "136:" // Height 4: Partial direct writeback: partial_2_8 - "tbz x16, #1, 137f\n" - "str d10, [x13], #0x8\n" - "str d14, [x9], #0x8\n" - "str d18, [x27], #0x8\n" - "str d22, [x25], #0x8\n" - "tbz x16, #0, 142f\n" - "st1 { v10.s }[2], [x13]\n" - "st1 { v14.s }[2], [x9]\n" - "st1 { v18.s }[2], [x27]\n" - "st1 { v22.s }[2], [x25]\n" - "b 142f\n" - "137:" // Height 4: Partial direct writeback: partial_1_8 - "tbz x16, #0, 142f\n" - "str s10, [x13, #0x0]\n" - "str s14, [x9, #0x0]\n" - "str s18, [x27, #0x0]\n" - "str s22, [x25, #0x0]\n" - "b 142f\n" - "138:" // Height 4: Partial direct writeback: partial_4_0 - "tbz x16, #2, 140f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "st1 { v16.4s }, [x27], #0x10\n" - "st1 { v20.4s }, [x25], #0x10\n" - "tbz x16, #1, 139f\n" - "str d9, [x13], #0x8\n" - "str d13, [x9], #0x8\n" - "str d17, [x27], #0x8\n" - "str d21, [x25], #0x8\n" - "tbz x16, #0, 142f\n" - "st1 { v9.s }[2], [x13]\n" - "st1 { v13.s }[2], [x9]\n" - "st1 { v17.s }[2], [x27]\n" - "st1 { v21.s }[2], [x25]\n" - "b 142f\n" - "139:" // Height 4: Partial direct writeback: partial_1_4 - "tbz x16, #0, 142f\n" - "str s9, [x13, #0x0]\n" - "str s13, [x9, #0x0]\n" - "str s17, [x27, #0x0]\n" - "str s21, [x25, #0x0]\n" - "b 142f\n" - "140:" // Height 4: Partial direct writeback: partial_2_0 - "tbz x16, #1, 141f\n" - "str d8, [x13], #0x8\n" - "str d12, [x9], #0x8\n" - "str d16, [x27], #0x8\n" - "str d20, [x25], #0x8\n" - "tbz x16, #0, 142f\n" - "st1 { v8.s }[2], [x13]\n" - "st1 { v12.s }[2], [x9]\n" - "st1 { v16.s }[2], [x27]\n" - "st1 { v20.s }[2], [x25]\n" - "b 142f\n" - "141:" // Height 4: Partial direct writeback: partial_1_0 - "str s8, [x13, #0x0]\n" - "str s12, [x9, #0x0]\n" - "str s16, [x27, #0x0]\n" - "str s20, [x25, #0x0]\n" - "142:" // Height 4: Partial direct writeback: Done - "b 144f\n" - "143:" // Height 4: Full writeback - "str q8, [x13, #0x0]\n" - "str q9, [x13, #0x10]\n" - "str q10, [x13, #0x20]\n" - "str q11, [x13, #0x30]\n" - "str q12, [x9, #0x0]\n" - "str q13, [x9, #0x10]\n" - "str q14, [x9, #0x20]\n" - "str q15, [x9, #0x30]\n" - "str q16, [x27, #0x0]\n" - "str q17, [x27, #0x10]\n" - "str q18, [x27, #0x20]\n" - "str q19, [x27, #0x30]\n" - "str q20, [x25, #0x0]\n" - "str q21, [x25, #0x10]\n" - "str q22, [x25, #0x20]\n" - "str q23, [x25, #0x30]\n" - "add x13, x13, #0x40\n" - "add x9, x9, #0x40\n" - "add x27, x27, #0x40\n" - "add x25, x25, #0x40\n" - "144:" // Height 4: Writeback done - "subs x16, x16, #0x10\n" - "bgt 111b\n" - "b 218f\n" - "145:" // Height 5 - "ldr x16, [%x[args_ptr], %[offsetof_N]]\n" - "mov x14, %x[bias]\n" - "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 146f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19, LSL #2\n" - "ldr x25, [%x[output_ptr], #0x18]\n" - "ldr x23, [%x[output_ptr], #0x20]\n" - "add x27, x27, x19, LSL #2\n" - "add x25, x25, x19, LSL #2\n" - "add x23, x23, x19, LSL #2\n" - "b 147f\n" - "146:" // Height 5: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19, LSL #2\n" - "add x27, x9, x19, LSL #2\n" - "add x25, x27, x19, LSL #2\n" - "add x23, x25, x19, LSL #2\n" - "147:" // Height 5: Column loop - "cbz x14, 148f\n" - "ldr q8, [x14, #0x0]\n" + "130:" // Height 4: No activation + "cmp x11, #0x10\n" + "bge 139f\n" + "tbz x11, #3, 134f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v9.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "st1 { v13.4s }, [x24], #0x10\n" + "st1 { v16.4s }, [x23], #0x10\n" + "st1 { v17.4s }, [x23], #0x10\n" + "st1 { v20.4s }, [x22], #0x10\n" + "st1 { v21.4s }, [x22], #0x10\n" + "tbz x11, #2, 132f\n" + "st1 { v10.4s }, [x28], #0x10\n" + "st1 { v14.4s }, [x24], #0x10\n" + "st1 { v18.4s }, [x23], #0x10\n" + "st1 { v22.4s }, [x22], #0x10\n" + "tbz x11, #1, 131f\n" + "str d11, [x28], #0x8\n" + "str d15, [x24], #0x8\n" + "str d19, [x23], #0x8\n" + "str d23, [x22], #0x8\n" + "tbz x11, #0, 138f\n" + "st1 { v11.s }[2], [x28]\n" + "st1 { v15.s }[2], [x24]\n" + "st1 { v19.s }[2], [x23]\n" + "st1 { v23.s }[2], [x22]\n" + "b 138f\n" + "131:" // Height 4: Partial direct writeback: partial_1_12 + "tbz x11, #0, 138f\n" + "str s11, [x28, #0x0]\n" + "str s15, [x24, #0x0]\n" + "str s19, [x23, #0x0]\n" + "str s23, [x22, #0x0]\n" + "b 138f\n" + "132:" // Height 4: Partial direct writeback: partial_2_8 + "tbz x11, #1, 133f\n" + "str d10, [x28], #0x8\n" + "str d14, [x24], #0x8\n" + "str d18, [x23], #0x8\n" + "str d22, [x22], #0x8\n" + "tbz x11, #0, 138f\n" + "st1 { v10.s }[2], [x28]\n" + "st1 { v14.s }[2], [x24]\n" + "st1 { v18.s }[2], [x23]\n" + "st1 { v22.s }[2], [x22]\n" + "b 138f\n" + "133:" // Height 4: Partial direct writeback: partial_1_8 + "tbz x11, #0, 138f\n" + "str s10, [x28, #0x0]\n" + "str s14, [x24, #0x0]\n" + "str s18, [x23, #0x0]\n" + "str s22, [x22, #0x0]\n" + "b 138f\n" + "134:" // Height 4: Partial direct writeback: partial_4_0 + "tbz x11, #2, 136f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "st1 { v16.4s }, [x23], #0x10\n" + "st1 { v20.4s }, [x22], #0x10\n" + "tbz x11, #1, 135f\n" + "str d9, [x28], #0x8\n" + "str d13, [x24], #0x8\n" + "str d17, [x23], #0x8\n" + "str d21, [x22], #0x8\n" + "tbz x11, #0, 138f\n" + "st1 { v9.s }[2], [x28]\n" + "st1 { v13.s }[2], [x24]\n" + "st1 { v17.s }[2], [x23]\n" + "st1 { v21.s }[2], [x22]\n" + "b 138f\n" + "135:" // Height 4: Partial direct writeback: partial_1_4 + "tbz x11, #0, 138f\n" + "str s9, [x28, #0x0]\n" + "str s13, [x24, #0x0]\n" + "str s17, [x23, #0x0]\n" + "str s21, [x22, #0x0]\n" + "b 138f\n" + "136:" // Height 4: Partial direct writeback: partial_2_0 + "tbz x11, #1, 137f\n" + "str d8, [x28], #0x8\n" + "str d12, [x24], #0x8\n" + "str d16, [x23], #0x8\n" + "str d20, [x22], #0x8\n" + "tbz x11, #0, 138f\n" + "st1 { v8.s }[2], [x28]\n" + "st1 { v12.s }[2], [x24]\n" + "st1 { v16.s }[2], [x23]\n" + "st1 { v20.s }[2], [x22]\n" + "b 138f\n" + "137:" // Height 4: Partial direct writeback: partial_1_0 + "str s8, [x28, #0x0]\n" + "str s12, [x24, #0x0]\n" + "str s16, [x23, #0x0]\n" + "str s20, [x22, #0x0]\n" + "138:" // Height 4: Partial direct writeback: Done + "b 140f\n" + "139:" // Height 4: Full writeback + "str q8, [x28, #0x0]\n" + "str q9, [x28, #0x10]\n" + "str q10, [x28, #0x20]\n" + "str q11, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + "str q12, [x24, #0x0]\n" + "str q13, [x24, #0x10]\n" + "str q14, [x24, #0x20]\n" + "str q15, [x24, #0x30]\n" + "str q16, [x23, #0x0]\n" + "str q17, [x23, #0x10]\n" + "str q18, [x23, #0x20]\n" + "str q19, [x23, #0x30]\n" + "str q20, [x22, #0x0]\n" + "str q21, [x22, #0x10]\n" + "str q22, [x22, #0x20]\n" + "str q23, [x22, #0x30]\n" + "140:" // Height 4: Writeback done + "subs x11, x11, #0x10\n" + "bgt 107b\n" + "b 212f\n" + "141:" // Height 5 + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "mov x9, %x[bias]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x28, %x[output_ptr]\n" + "142:" // Height 5: Column loop + "cbz x9, 143f\n" + "ldr q8, [x9, #0x0]\n" "mov v12.16b, v8.16b\n" - "ldr q9, [x14, #0x10]\n" + "ldr q9, [x9, #0x10]\n" "mov v16.16b, v8.16b\n" - "ldr q10, [x14, #0x20]\n" + "ldr q10, [x9, #0x20]\n" "mov v20.16b, v8.16b\n" - "ldr q11, [x14, #0x30]\n" + "ldr q11, [x9, #0x30]\n" + "add x9, x9, #0x40\n" "mov v24.16b, v8.16b\n" - "add x14, x14, #0x40\n" "mov v13.16b, v9.16b\n" "mov v17.16b, v9.16b\n" "mov v14.16b, v10.16b\n" @@ -2014,157 +1960,158 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "mov v25.16b, v9.16b\n" "mov v26.16b, v10.16b\n" "mov v27.16b, v11.16b\n" - "b 159f\n" - "148:" // Height 5: no bias - "tbz %x[flags], #0, 158f\n" - "cmp x16, #0x10\n" - "bge 157f\n" - "tbz x16, #3, 152f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "ld1 { v16.4s }, [x27], #0x10\n" - "ld1 { v20.4s }, [x25], #0x10\n" - "ld1 { v24.4s }, [x23], #0x10\n" - "ld1 { v9.4s }, [x13], #0x10\n" - "ld1 { v13.4s }, [x9], #0x10\n" - "ld1 { v17.4s }, [x27], #0x10\n" - "ld1 { v21.4s }, [x25], #0x10\n" - "ld1 { v25.4s }, [x23], #0x10\n" - "tbz x16, #2, 150f\n" - "ld1 { v10.4s }, [x13], #0x10\n" - "ld1 { v14.4s }, [x9], #0x10\n" - "ld1 { v18.4s }, [x27], #0x10\n" - "ld1 { v22.4s }, [x25], #0x10\n" - "ld1 { v26.4s }, [x23], #0x10\n" - "tbz x16, #1, 149f\n" + "b 154f\n" + "143:" // Height 5: no bias + "tbz %x[flags], #0, 153f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "cmp x11, #0x10\n" + "add x24, x28, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "add x22, x23, x19, LSL #2\n" + "add x21, x22, x19, LSL #2\n" + "bge 152f\n" + "tbz x11, #3, 147f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "ld1 { v16.4s }, [x23], #0x10\n" + "ld1 { v20.4s }, [x22], #0x10\n" + "ld1 { v24.4s }, [x21], #0x10\n" + "ld1 { v9.4s }, [x28], #0x10\n" + "ld1 { v13.4s }, [x24], #0x10\n" + "ld1 { v17.4s }, [x23], #0x10\n" + "ld1 { v21.4s }, [x22], #0x10\n" + "ld1 { v25.4s }, [x21], #0x10\n" + "tbz x11, #2, 145f\n" + "ld1 { v10.4s }, [x28], #0x10\n" + "ld1 { v14.4s }, [x24], #0x10\n" + "ld1 { v18.4s }, [x23], #0x10\n" + "ld1 { v22.4s }, [x22], #0x10\n" + "ld1 { v26.4s }, [x21], #0x10\n" + "tbz x11, #1, 144f\n" "mov x19, #0x38\n" - "ldr d11, [x13], #0x8\n" - "ldr d15, [x9], #0x8\n" - "ldr d19, [x27], #0x8\n" - "ldr d23, [x25], #0x8\n" - "ldr d27, [x23], #0x8\n" - "tbz x16, #0, 156f\n" - "ld1 { v11.s }[2], [x13]\n" - "ld1 { v15.s }[2], [x9]\n" - "ld1 { v19.s }[2], [x27]\n" - "ld1 { v23.s }[2], [x25]\n" - "ld1 { v27.s }[2], [x23]\n" - "b 156f\n" - "149:" // Height 5: Partial accumulate: partial_1_12 + "ldr d11, [x28], #0x8\n" + "ldr d15, [x24], #0x8\n" + "ldr d19, [x23], #0x8\n" + "ldr d23, [x22], #0x8\n" + "ldr d27, [x21], #0x8\n" + "tbz x11, #0, 151f\n" + "ld1 { v11.s }[2], [x28]\n" + "ld1 { v15.s }[2], [x24]\n" + "ld1 { v19.s }[2], [x23]\n" + "ld1 { v23.s }[2], [x22]\n" + "ld1 { v27.s }[2], [x21]\n" + "b 151f\n" + "144:" // Height 5: Partial accumulate: partial_1_12 "mov x19, #0x30\n" - "tbz x16, #0, 156f\n" - "ldr s11, [x13, #0x0]\n" - "ldr s15, [x9, #0x0]\n" - "ldr s19, [x27, #0x0]\n" - "ldr s23, [x25, #0x0]\n" - "ldr s27, [x23, #0x0]\n" - "b 156f\n" - "150:" // Height 5: Partial accumulate: partial_2_8 - "tbz x16, #1, 151f\n" - "ldr d10, [x13], #0x8\n" - "ldr d14, [x9], #0x8\n" - "ldr d18, [x27], #0x8\n" - "ldr d22, [x25], #0x8\n" - "ldr d26, [x23], #0x8\n" + "tbz x11, #0, 151f\n" + "ldr s11, [x28, #0x0]\n" + "ldr s15, [x24, #0x0]\n" + "ldr s19, [x23, #0x0]\n" + "ldr s23, [x22, #0x0]\n" + "ldr s27, [x21, #0x0]\n" + "b 151f\n" + "145:" // Height 5: Partial accumulate: partial_2_8 + "tbz x11, #1, 146f\n" + "ldr d10, [x28], #0x8\n" + "ldr d14, [x24], #0x8\n" "mov x19, #0x28\n" - "tbz x16, #0, 156f\n" - "ld1 { v10.s }[2], [x13]\n" - "ld1 { v14.s }[2], [x9]\n" - "ld1 { v18.s }[2], [x27]\n" - "ld1 { v22.s }[2], [x25]\n" - "ld1 { v26.s }[2], [x23]\n" - "b 156f\n" - "151:" // Height 5: Partial accumulate: partial_1_8 + "ldr d18, [x23], #0x8\n" + "ldr d22, [x22], #0x8\n" + "ldr d26, [x21], #0x8\n" + "tbz x11, #0, 151f\n" + "ld1 { v10.s }[2], [x28]\n" + "ld1 { v14.s }[2], [x24]\n" + "ld1 { v18.s }[2], [x23]\n" + "ld1 { v22.s }[2], [x22]\n" + "ld1 { v26.s }[2], [x21]\n" + "b 151f\n" + "146:" // Height 5: Partial accumulate: partial_1_8 "mov x19, #0x20\n" - "tbz x16, #0, 156f\n" - "ldr s10, [x13, #0x0]\n" - "ldr s14, [x9, #0x0]\n" - "ldr s18, [x27, #0x0]\n" - "ldr s22, [x25, #0x0]\n" - "ldr s26, [x23, #0x0]\n" - "b 156f\n" - "152:" // Height 5: Partial accumulate: partial_4_0 - "tbz x16, #2, 154f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "ld1 { v16.4s }, [x27], #0x10\n" - "ld1 { v20.4s }, [x25], #0x10\n" - "ld1 { v24.4s }, [x23], #0x10\n" - "tbz x16, #1, 153f\n" + "tbz x11, #0, 151f\n" + "ldr s10, [x28, #0x0]\n" + "ldr s14, [x24, #0x0]\n" + "ldr s18, [x23, #0x0]\n" + "ldr s22, [x22, #0x0]\n" + "ldr s26, [x21, #0x0]\n" + "b 151f\n" + "147:" // Height 5: Partial accumulate: partial_4_0 + "tbz x11, #2, 149f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "ld1 { v16.4s }, [x23], #0x10\n" + "ld1 { v20.4s }, [x22], #0x10\n" + "ld1 { v24.4s }, [x21], #0x10\n" + "tbz x11, #1, 148f\n" "mov x19, #0x18\n" - "ldr d9, [x13], #0x8\n" - "ldr d13, [x9], #0x8\n" - "ldr d17, [x27], #0x8\n" - "ldr d21, [x25], #0x8\n" - "ldr d25, [x23], #0x8\n" - "tbz x16, #0, 156f\n" - "ld1 { v9.s }[2], [x13]\n" - "ld1 { v13.s }[2], [x9]\n" - "ld1 { v17.s }[2], [x27]\n" - "ld1 { v21.s }[2], [x25]\n" - "ld1 { v25.s }[2], [x23]\n" - "b 156f\n" - "153:" // Height 5: Partial accumulate: partial_1_4 + "ldr d9, [x28], #0x8\n" + "ldr d13, [x24], #0x8\n" + "ldr d17, [x23], #0x8\n" + "ldr d21, [x22], #0x8\n" + "ldr d25, [x21], #0x8\n" + "tbz x11, #0, 151f\n" + "ld1 { v9.s }[2], [x28]\n" + "ld1 { v13.s }[2], [x24]\n" + "ld1 { v17.s }[2], [x23]\n" + "ld1 { v21.s }[2], [x22]\n" + "ld1 { v25.s }[2], [x21]\n" + "b 151f\n" + "148:" // Height 5: Partial accumulate: partial_1_4 "mov x19, #0x10\n" - "tbz x16, #0, 156f\n" - "ldr s9, [x13, #0x0]\n" - "ldr s13, [x9, #0x0]\n" - "ldr s17, [x27, #0x0]\n" - "ldr s21, [x25, #0x0]\n" - "ldr s25, [x23, #0x0]\n" - "b 156f\n" - "154:" // Height 5: Partial accumulate: partial_2_0 - "tbz x16, #1, 155f\n" - "ldr d8, [x13], #0x8\n" - "ldr d12, [x9], #0x8\n" - "ldr d16, [x27], #0x8\n" - "ldr d20, [x25], #0x8\n" - "ldr d24, [x23], #0x8\n" + "tbz x11, #0, 151f\n" + "ldr s9, [x28, #0x0]\n" + "ldr s13, [x24, #0x0]\n" + "ldr s17, [x23, #0x0]\n" + "ldr s21, [x22, #0x0]\n" + "ldr s25, [x21, #0x0]\n" + "b 151f\n" + "149:" // Height 5: Partial accumulate: partial_2_0 + "tbz x11, #1, 150f\n" + "ldr d8, [x28], #0x8\n" + "ldr d12, [x24], #0x8\n" "mov x19, #0x8\n" - "tbz x16, #0, 156f\n" - "ld1 { v8.s }[2], [x13]\n" - "ld1 { v12.s }[2], [x9]\n" - "ld1 { v16.s }[2], [x27]\n" - "ld1 { v20.s }[2], [x25]\n" - "ld1 { v24.s }[2], [x23]\n" - "b 156f\n" - "155:" // Height 5: Partial accumulate: partial_1_0 + "ldr d16, [x23], #0x8\n" + "ldr d20, [x22], #0x8\n" + "ldr d24, [x21], #0x8\n" + "tbz x11, #0, 151f\n" + "ld1 { v8.s }[2], [x28]\n" + "ld1 { v12.s }[2], [x24]\n" + "ld1 { v16.s }[2], [x23]\n" + "ld1 { v20.s }[2], [x22]\n" + "ld1 { v24.s }[2], [x21]\n" + "b 151f\n" + "150:" // Height 5: Partial accumulate: partial_1_0 + "ldr s8, [x28, #0x0]\n" "mov x19, #0x0\n" - "ldr s8, [x13, #0x0]\n" - "ldr s12, [x9, #0x0]\n" - "ldr s16, [x27, #0x0]\n" - "ldr s20, [x25, #0x0]\n" - "ldr s24, [x23, #0x0]\n" - "156:" // Height 5: Partial accumulate: Done - "sub x13, x13, x19\n" - "sub x9, x9, x19\n" - "sub x27, x27, x19\n" - "sub x25, x25, x19\n" - "sub x23, x23, x19\n" - "b 159f\n" - "157:" // Height 5: full accumulate - "ldr q8, [x13, #0x0]\n" - "ldr q9, [x13, #0x10]\n" - "ldr q10, [x13, #0x20]\n" - "ldr q11, [x13, #0x30]\n" - "ldr q12, [x9, #0x0]\n" - "ldr q13, [x9, #0x10]\n" - "ldr q14, [x9, #0x20]\n" - "ldr q15, [x9, #0x30]\n" - "ldr q16, [x27, #0x0]\n" - "ldr q17, [x27, #0x10]\n" - "ldr q18, [x27, #0x20]\n" - "ldr q19, [x27, #0x30]\n" - "ldr q20, [x25, #0x0]\n" - "ldr q21, [x25, #0x10]\n" - "ldr q22, [x25, #0x20]\n" - "ldr q23, [x25, #0x30]\n" - "ldr q24, [x23, #0x0]\n" - "ldr q25, [x23, #0x10]\n" - "ldr q26, [x23, #0x20]\n" - "ldr q27, [x23, #0x30]\n" - "b 159f\n" - "158:" // Height 5: no accumulate + "ldr s12, [x24, #0x0]\n" + "ldr s16, [x23, #0x0]\n" + "ldr s20, [x22, #0x0]\n" + "ldr s24, [x21, #0x0]\n" + "151:" // Height 5: Partial accumulate: Done + "sub x28, x28, x19\n" + "b 154f\n" + "152:" // Height 5: full accumulate + "ldr q8, [x28, #0x0]\n" + "ldr q9, [x28, #0x10]\n" + "ldr q10, [x28, #0x20]\n" + "ldr q11, [x28, #0x30]\n" + "ldr q12, [x24, #0x0]\n" + "ldr q13, [x24, #0x10]\n" + "ldr q14, [x24, #0x20]\n" + "ldr q15, [x24, #0x30]\n" + "ldr q16, [x23, #0x0]\n" + "ldr q17, [x23, #0x10]\n" + "ldr q18, [x23, #0x20]\n" + "ldr q19, [x23, #0x30]\n" + "ldr q20, [x22, #0x0]\n" + "ldr q21, [x22, #0x10]\n" + "ldr q22, [x22, #0x20]\n" + "ldr q23, [x22, #0x30]\n" + "ldr q24, [x21, #0x0]\n" + "ldr q25, [x21, #0x10]\n" + "ldr q26, [x21, #0x20]\n" + "ldr q27, [x21, #0x30]\n" + "b 154f\n" + "153:" // Height 5: no accumulate "movi v8.16b, #0x0\n" "movi v9.16b, #0x0\n" "movi v10.16b, #0x0\n" @@ -2185,260 +2132,260 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "movi v25.16b, #0x0\n" "movi v26.16b, #0x0\n" "movi v27.16b, #0x0\n" - "159:" // Height 5: setup done - "mov x12, #0x0\n" - "160:" // Height 5: String loop + "154:" // Height 5: setup done + "mov x27, #0x0\n" + "155:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 161f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 156f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x26, [x20, #0x10]\n" - "ldr x24, [x20, #0x18]\n" - "ldr x22, [x20, #0x20]\n" - "cbnz x12, 162f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "ldr x22, [x20, #0x18]\n" + "ldr x21, [x20, #0x20]\n" + "cbnz x27, 157f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19, LSL #1\n" - "add x28, x28, x19, LSL #1\n" - "add x26, x26, x19, LSL #1\n" + "add x25, x25, x19, LSL #1\n" "add x24, x24, x19, LSL #1\n" + "add x23, x23, x19, LSL #1\n" "add x22, x22, x19, LSL #1\n" - "b 162f\n" - "161:" // Height 5: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19, LSL #1\n" - "add x26, x28, x19, LSL #1\n" - "add x24, x26, x19, LSL #1\n" - "add x22, x24, x19, LSL #1\n" - "162:" // Height 5: input setup done - "cmp x11, #0x8\n" - "blt 165f\n" - "cmp x11, #0x10\n" - "blt 164f\n" - "163:" // Height 5: Multiply loop: Main loop head - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q3, [x24, #0x0]\n" - "ldr q4, [x22, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "add x21, x21, x19, LSL #1\n" + "b 157f\n" + "156:" // Height 5: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19, LSL #1\n" + "add x23, x24, x19, LSL #1\n" + "add x22, x23, x19, LSL #1\n" + "add x21, x22, x19, LSL #1\n" + "157:" // Height 5: input setup done + "cmp x26, #0x8\n" + "blt 160f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x10\n" + "ldr q2, [x23, #0x0]\n" + "ldr q3, [x22, #0x0]\n" + "ldr q4, [x21, #0x0]\n" + "ldr q6, [x10, #0x0]\n" + "blt 159f\n" + "158:" // Height 5: Multiply loop: Main loop head ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" + "add x22, x22, #0x10\n" ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" + "add x21, x21, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "add x24, x24, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "add x22, x22, #0x10\n" + "prfm pldl1keep, [x21, #0x80]\n" + "cmp x26, #0x10\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "sub x11, x11, #0x8\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "cmp x11, #0x10\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x15, #0x40]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" ".inst 0x4f64f0d8 // bfdot v24.4s, v6.8h, v4.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" ".inst 0x4f64f0f9 // bfdot v25.4s, v7.8h, v4.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" ".inst 0x4f64f0da // bfdot v26.4s, v6.8h, v4.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" ".inst 0x4f64f0fb // bfdot v27.4s, v7.8h, v4.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" ".inst 0x4f44f8d8 // bfdot v24.4s, v6.8h, v4.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" ".inst 0x4f44f8f9 // bfdot v25.4s, v7.8h, v4.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" ".inst 0x4f44f8da // bfdot v26.4s, v6.8h, v4.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" ".inst 0x4f44f8fb // bfdot v27.4s, v7.8h, v4.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" ".inst 0x4f64f8d8 // bfdot v24.4s, v6.8h, v4.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" ".inst 0x4f64f8f9 // bfdot v25.4s, v7.8h, v4.h[3]\n" - "ldr q7, [x15, #0xf0]\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - "add x15, x15, #0x100\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d6 // bfdot v22.4s, v6.8h, v3.h[3]\n" ".inst 0x4f64f8da // bfdot v26.4s, v6.8h, v4.h[3]\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" + "ldr q1, [x24, #0x0]\n" ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" + "ldr q2, [x23, #0x0]\n" ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" + "ldr q3, [x22, #0x0]\n" ".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n" - "bge 163b\n" - "164:" // Height 5: Multiply loop: Single iteration only - "sub x11, x11, #0x8\n" - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q3, [x24, #0x0]\n" - "ldr q4, [x22, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "ldr q4, [x21, #0x0]\n" + "bge 158b\n" + "159:" // Height 5: Multiply loop: Single iteration only ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "add x28, x28, #0x10\n" + "add x24, x24, #0x10\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x15, #0x20]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "add x22, x22, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "add x24, x24, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" + "ldr q6, [x10, #0x20]\n" + "add x21, x21, #0x10\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "add x22, x22, #0x10\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" "prfm pldl1keep, [x22, #0x80]\n" + ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" + "prfm pldl1keep, [x21, #0x80]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x15, #0x40]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" ".inst 0x4f64f0d8 // bfdot v24.4s, v6.8h, v4.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" ".inst 0x4f64f0f9 // bfdot v25.4s, v7.8h, v4.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" ".inst 0x4f64f0da // bfdot v26.4s, v6.8h, v4.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" ".inst 0x4f64f0fb // bfdot v27.4s, v7.8h, v4.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" ".inst 0x4f44f8d8 // bfdot v24.4s, v6.8h, v4.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" ".inst 0x4f44f8f9 // bfdot v25.4s, v7.8h, v4.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" ".inst 0x4f44f8da // bfdot v26.4s, v6.8h, v4.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" ".inst 0x4f44f8fb // bfdot v27.4s, v7.8h, v4.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" ".inst 0x4f64f8d8 // bfdot v24.4s, v6.8h, v4.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" ".inst 0x4f64f8f9 // bfdot v25.4s, v7.8h, v4.h[3]\n" - "ldr q7, [x15, #0xf0]\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - "add x15, x15, #0x100\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d6 // bfdot v22.4s, v6.8h, v3.h[3]\n" @@ -2448,34 +2395,34 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" ".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n" - "165:" // Height 5: Multiply loop: Main loop skip - "cbz x11, 169f\n" - "cmp x11, #0x2\n" - "blt 167f\n" - "166:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x10], #0x4\n" - "ldr s1, [x28], #0x4\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x24], #0x4\n" - "ldr s4, [x22], #0x4\n" - "ldr q6, [x15, #0x0]\n" + "160:" // Height 5: Multiply loop: Main loop skip + "cbz x26, 164f\n" + "cmp x26, #0x2\n" + "blt 162f\n" + "161:" // Height 5: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x2\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x2\n" + "ldr s2, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr s4, [x21], #0x4\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "sub x11, x11, #0x2\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "cmp x11, #0x2\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" @@ -2485,31 +2432,31 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" - "bge 166b\n" - "cbz x11, 169f\n" - "167:" // Height 5: Multiply loop: Skip odd blocks - "ldr h0, [x10, #0x0]\n" - "ldr h1, [x28, #0x0]\n" - "ldr h2, [x26, #0x0]\n" - "ldr h3, [x24, #0x0]\n" - "ldr h4, [x22, #0x0]\n" - "168:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q6, [x15, #0x0]\n" + "bge 161b\n" + "cbz x26, 164f\n" + "162:" // Height 5: Multiply loop: Skip odd blocks + "ldr h0, [x25, #0x0]\n" + "ldr h1, [x24, #0x0]\n" + "ldr h2, [x23, #0x0]\n" + "ldr h3, [x22, #0x0]\n" + "ldr h4, [x21, #0x0]\n" + "163:" // Height 5: Multiply loop: Ragged operand read: Done + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" @@ -2519,17 +2466,22 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" - "169:" // Height 5: Multiply loop: No odd multiplies + "164:" // Height 5: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x12, x12, #0x1\n" - "cmp x12, x19\n" - "bne 160b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x27, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 155b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "prfm pstl1keep, [x28, #0x0]\n" + "add x24, x28, x19, LSL #2\n" + "prfm pstl1keep, [x24, #0x0]\n" + "add x23, x24, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" - "tbz %x[flags], #1, 170f\n" + "add x22, x23, x19, LSL #2\n" + "prfm pstl1keep, [x22, #0x0]\n" + "add x21, x22, x19, LSL #2\n" + "prfm pstl1keep, [x21, #0x0]\n" + "tbz %x[flags], #1, 165f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1r { v1.4s }, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -2574,193 +2526,169 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "fmax v26.4s, v26.4s, v1.4s\n" "fmin v27.4s, v27.4s, v0.4s\n" "fmax v27.4s, v27.4s, v1.4s\n" - "170:" // Height 5: No activation - "cmp x16, #0x10\n" - "bge 179f\n" - "tbz x16, #3, 174f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v9.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "st1 { v13.4s }, [x9], #0x10\n" - "st1 { v16.4s }, [x27], #0x10\n" - "st1 { v17.4s }, [x27], #0x10\n" - "st1 { v20.4s }, [x25], #0x10\n" - "st1 { v21.4s }, [x25], #0x10\n" - "st1 { v24.4s }, [x23], #0x10\n" - "st1 { v25.4s }, [x23], #0x10\n" - "tbz x16, #2, 172f\n" - "st1 { v10.4s }, [x13], #0x10\n" - "st1 { v14.4s }, [x9], #0x10\n" - "st1 { v18.4s }, [x27], #0x10\n" - "st1 { v22.4s }, [x25], #0x10\n" - "st1 { v26.4s }, [x23], #0x10\n" - "tbz x16, #1, 171f\n" - "str d11, [x13], #0x8\n" - "str d15, [x9], #0x8\n" - "str d19, [x27], #0x8\n" - "str d23, [x25], #0x8\n" - "str d27, [x23], #0x8\n" - "tbz x16, #0, 178f\n" - "st1 { v11.s }[2], [x13]\n" - "st1 { v15.s }[2], [x9]\n" - "st1 { v19.s }[2], [x27]\n" - "st1 { v23.s }[2], [x25]\n" - "st1 { v27.s }[2], [x23]\n" - "b 178f\n" - "171:" // Height 5: Partial direct writeback: partial_1_12 - "tbz x16, #0, 178f\n" - "str s11, [x13, #0x0]\n" - "str s15, [x9, #0x0]\n" - "str s19, [x27, #0x0]\n" - "str s23, [x25, #0x0]\n" - "str s27, [x23, #0x0]\n" - "b 178f\n" - "172:" // Height 5: Partial direct writeback: partial_2_8 - "tbz x16, #1, 173f\n" - "str d10, [x13], #0x8\n" - "str d14, [x9], #0x8\n" - "str d18, [x27], #0x8\n" - "str d22, [x25], #0x8\n" - "str d26, [x23], #0x8\n" - "tbz x16, #0, 178f\n" - "st1 { v10.s }[2], [x13]\n" - "st1 { v14.s }[2], [x9]\n" - "st1 { v18.s }[2], [x27]\n" - "st1 { v22.s }[2], [x25]\n" - "st1 { v26.s }[2], [x23]\n" - "b 178f\n" - "173:" // Height 5: Partial direct writeback: partial_1_8 - "tbz x16, #0, 178f\n" - "str s10, [x13, #0x0]\n" - "str s14, [x9, #0x0]\n" - "str s18, [x27, #0x0]\n" - "str s22, [x25, #0x0]\n" - "str s26, [x23, #0x0]\n" - "b 178f\n" - "174:" // Height 5: Partial direct writeback: partial_4_0 - "tbz x16, #2, 176f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "st1 { v16.4s }, [x27], #0x10\n" - "st1 { v20.4s }, [x25], #0x10\n" - "st1 { v24.4s }, [x23], #0x10\n" - "tbz x16, #1, 175f\n" - "str d9, [x13], #0x8\n" - "str d13, [x9], #0x8\n" - "str d17, [x27], #0x8\n" - "str d21, [x25], #0x8\n" - "str d25, [x23], #0x8\n" - "tbz x16, #0, 178f\n" - "st1 { v9.s }[2], [x13]\n" - "st1 { v13.s }[2], [x9]\n" - "st1 { v17.s }[2], [x27]\n" - "st1 { v21.s }[2], [x25]\n" - "st1 { v25.s }[2], [x23]\n" - "b 178f\n" - "175:" // Height 5: Partial direct writeback: partial_1_4 - "tbz x16, #0, 178f\n" - "str s9, [x13, #0x0]\n" - "str s13, [x9, #0x0]\n" - "str s17, [x27, #0x0]\n" - "str s21, [x25, #0x0]\n" - "str s25, [x23, #0x0]\n" - "b 178f\n" - "176:" // Height 5: Partial direct writeback: partial_2_0 - "tbz x16, #1, 177f\n" - "str d8, [x13], #0x8\n" - "str d12, [x9], #0x8\n" - "str d16, [x27], #0x8\n" - "str d20, [x25], #0x8\n" - "str d24, [x23], #0x8\n" - "tbz x16, #0, 178f\n" - "st1 { v8.s }[2], [x13]\n" - "st1 { v12.s }[2], [x9]\n" - "st1 { v16.s }[2], [x27]\n" - "st1 { v20.s }[2], [x25]\n" - "st1 { v24.s }[2], [x23]\n" - "b 178f\n" - "177:" // Height 5: Partial direct writeback: partial_1_0 - "str s8, [x13, #0x0]\n" - "str s12, [x9, #0x0]\n" - "str s16, [x27, #0x0]\n" - "str s20, [x25, #0x0]\n" - "str s24, [x23, #0x0]\n" - "178:" // Height 5: Partial direct writeback: Done - "b 180f\n" - "179:" // Height 5: Full writeback - "str q8, [x13, #0x0]\n" - "str q9, [x13, #0x10]\n" - "str q10, [x13, #0x20]\n" - "str q11, [x13, #0x30]\n" - "str q12, [x9, #0x0]\n" - "str q13, [x9, #0x10]\n" - "str q14, [x9, #0x20]\n" - "str q15, [x9, #0x30]\n" - "str q16, [x27, #0x0]\n" - "str q17, [x27, #0x10]\n" - "str q18, [x27, #0x20]\n" - "str q19, [x27, #0x30]\n" - "str q20, [x25, #0x0]\n" - "str q21, [x25, #0x10]\n" - "str q22, [x25, #0x20]\n" - "str q23, [x25, #0x30]\n" - "str q24, [x23, #0x0]\n" - "str q25, [x23, #0x10]\n" - "str q26, [x23, #0x20]\n" - "str q27, [x23, #0x30]\n" - "add x13, x13, #0x40\n" - "add x9, x9, #0x40\n" - "add x27, x27, #0x40\n" - "add x25, x25, #0x40\n" - "add x23, x23, #0x40\n" - "180:" // Height 5: Writeback done - "subs x16, x16, #0x10\n" - "bgt 147b\n" - "b 218f\n" - "181:" // Height 6 - "ldr x16, [%x[args_ptr], %[offsetof_N]]\n" - "mov x14, %x[bias]\n" - "ldr x15, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 182f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19, LSL #2\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19, LSL #2\n" - "ldr x25, [%x[output_ptr], #0x18]\n" - "ldr x23, [%x[output_ptr], #0x20]\n" - "add x27, x27, x19, LSL #2\n" - "ldr x21, [%x[output_ptr], #0x28]\n" - "add %x[output_ptr], %x[output_ptr], #0x30\n" - "add x25, x25, x19, LSL #2\n" - "add x23, x23, x19, LSL #2\n" - "add x21, x21, x19, LSL #2\n" - "b 183f\n" - "182:" // Height 6: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19, LSL #2\n" - "add x27, x9, x19, LSL #2\n" - "add x25, x27, x19, LSL #2\n" - "add x23, x25, x19, LSL #2\n" - "add x21, x23, x19, LSL #2\n" - "add %x[output_ptr], x21, x19, LSL #2\n" - "183:" // Height 6: Column loop - "cbz x14, 184f\n" - "ldr q8, [x14, #0x0]\n" + "165:" // Height 5: No activation + "cmp x11, #0x10\n" + "bge 174f\n" + "tbz x11, #3, 169f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v9.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "st1 { v13.4s }, [x24], #0x10\n" + "st1 { v16.4s }, [x23], #0x10\n" + "st1 { v17.4s }, [x23], #0x10\n" + "st1 { v20.4s }, [x22], #0x10\n" + "st1 { v21.4s }, [x22], #0x10\n" + "st1 { v24.4s }, [x21], #0x10\n" + "st1 { v25.4s }, [x21], #0x10\n" + "tbz x11, #2, 167f\n" + "st1 { v10.4s }, [x28], #0x10\n" + "st1 { v14.4s }, [x24], #0x10\n" + "st1 { v18.4s }, [x23], #0x10\n" + "st1 { v22.4s }, [x22], #0x10\n" + "st1 { v26.4s }, [x21], #0x10\n" + "tbz x11, #1, 166f\n" + "str d11, [x28], #0x8\n" + "str d15, [x24], #0x8\n" + "str d19, [x23], #0x8\n" + "str d23, [x22], #0x8\n" + "str d27, [x21], #0x8\n" + "tbz x11, #0, 173f\n" + "st1 { v11.s }[2], [x28]\n" + "st1 { v15.s }[2], [x24]\n" + "st1 { v19.s }[2], [x23]\n" + "st1 { v23.s }[2], [x22]\n" + "st1 { v27.s }[2], [x21]\n" + "b 173f\n" + "166:" // Height 5: Partial direct writeback: partial_1_12 + "tbz x11, #0, 173f\n" + "str s11, [x28, #0x0]\n" + "str s15, [x24, #0x0]\n" + "str s19, [x23, #0x0]\n" + "str s23, [x22, #0x0]\n" + "str s27, [x21, #0x0]\n" + "b 173f\n" + "167:" // Height 5: Partial direct writeback: partial_2_8 + "tbz x11, #1, 168f\n" + "str d10, [x28], #0x8\n" + "str d14, [x24], #0x8\n" + "str d18, [x23], #0x8\n" + "str d22, [x22], #0x8\n" + "str d26, [x21], #0x8\n" + "tbz x11, #0, 173f\n" + "st1 { v10.s }[2], [x28]\n" + "st1 { v14.s }[2], [x24]\n" + "st1 { v18.s }[2], [x23]\n" + "st1 { v22.s }[2], [x22]\n" + "st1 { v26.s }[2], [x21]\n" + "b 173f\n" + "168:" // Height 5: Partial direct writeback: partial_1_8 + "tbz x11, #0, 173f\n" + "str s10, [x28, #0x0]\n" + "str s14, [x24, #0x0]\n" + "str s18, [x23, #0x0]\n" + "str s22, [x22, #0x0]\n" + "str s26, [x21, #0x0]\n" + "b 173f\n" + "169:" // Height 5: Partial direct writeback: partial_4_0 + "tbz x11, #2, 171f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "st1 { v16.4s }, [x23], #0x10\n" + "st1 { v20.4s }, [x22], #0x10\n" + "st1 { v24.4s }, [x21], #0x10\n" + "tbz x11, #1, 170f\n" + "str d9, [x28], #0x8\n" + "str d13, [x24], #0x8\n" + "str d17, [x23], #0x8\n" + "str d21, [x22], #0x8\n" + "str d25, [x21], #0x8\n" + "tbz x11, #0, 173f\n" + "st1 { v9.s }[2], [x28]\n" + "st1 { v13.s }[2], [x24]\n" + "st1 { v17.s }[2], [x23]\n" + "st1 { v21.s }[2], [x22]\n" + "st1 { v25.s }[2], [x21]\n" + "b 173f\n" + "170:" // Height 5: Partial direct writeback: partial_1_4 + "tbz x11, #0, 173f\n" + "str s9, [x28, #0x0]\n" + "str s13, [x24, #0x0]\n" + "str s17, [x23, #0x0]\n" + "str s21, [x22, #0x0]\n" + "str s25, [x21, #0x0]\n" + "b 173f\n" + "171:" // Height 5: Partial direct writeback: partial_2_0 + "tbz x11, #1, 172f\n" + "str d8, [x28], #0x8\n" + "str d12, [x24], #0x8\n" + "str d16, [x23], #0x8\n" + "str d20, [x22], #0x8\n" + "str d24, [x21], #0x8\n" + "tbz x11, #0, 173f\n" + "st1 { v8.s }[2], [x28]\n" + "st1 { v12.s }[2], [x24]\n" + "st1 { v16.s }[2], [x23]\n" + "st1 { v20.s }[2], [x22]\n" + "st1 { v24.s }[2], [x21]\n" + "b 173f\n" + "172:" // Height 5: Partial direct writeback: partial_1_0 + "str s8, [x28, #0x0]\n" + "str s12, [x24, #0x0]\n" + "str s16, [x23, #0x0]\n" + "str s20, [x22, #0x0]\n" + "str s24, [x21, #0x0]\n" + "173:" // Height 5: Partial direct writeback: Done + "b 175f\n" + "174:" // Height 5: Full writeback + "str q8, [x28, #0x0]\n" + "str q9, [x28, #0x10]\n" + "str q10, [x28, #0x20]\n" + "str q11, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + "str q12, [x24, #0x0]\n" + "str q13, [x24, #0x10]\n" + "str q14, [x24, #0x20]\n" + "str q15, [x24, #0x30]\n" + "str q16, [x23, #0x0]\n" + "str q17, [x23, #0x10]\n" + "str q18, [x23, #0x20]\n" + "str q19, [x23, #0x30]\n" + "str q20, [x22, #0x0]\n" + "str q21, [x22, #0x10]\n" + "str q22, [x22, #0x20]\n" + "str q23, [x22, #0x30]\n" + "str q24, [x21, #0x0]\n" + "str q25, [x21, #0x10]\n" + "str q26, [x21, #0x20]\n" + "str q27, [x21, #0x30]\n" + "175:" // Height 5: Writeback done + "subs x11, x11, #0x10\n" + "bgt 142b\n" + "b 212f\n" + "176:" // Height 6 + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "mov x9, %x[bias]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x28, %x[output_ptr]\n" + "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" + "mov x19, #0x18\n" + "madd %x[output_ptr], x20, x19, %x[output_ptr]\n" + "177:" // Height 6: Column loop + "cbz x9, 178f\n" + "ldr q8, [x9, #0x0]\n" "mov v12.16b, v8.16b\n" - "ldr q9, [x14, #0x10]\n" + "ldr q9, [x9, #0x10]\n" "mov v16.16b, v8.16b\n" - "ldr q10, [x14, #0x20]\n" + "ldr q10, [x9, #0x20]\n" "mov v20.16b, v8.16b\n" - "ldr q11, [x14, #0x30]\n" + "ldr q11, [x9, #0x30]\n" + "add x9, x9, #0x40\n" "mov v24.16b, v8.16b\n" - "add x14, x14, #0x40\n" "mov v28.16b, v8.16b\n" "mov v13.16b, v9.16b\n" - "mov v17.16b, v9.16b\n" "mov v14.16b, v10.16b\n" "mov v15.16b, v11.16b\n" + "mov v17.16b, v9.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" "mov v21.16b, v9.16b\n" @@ -2772,178 +2700,179 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "mov v29.16b, v9.16b\n" "mov v30.16b, v10.16b\n" "mov v31.16b, v11.16b\n" - "b 195f\n" - "184:" // Height 6: no bias - "tbz %x[flags], #0, 194f\n" - "cmp x16, #0x10\n" - "bge 193f\n" - "tbz x16, #3, 188f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "ld1 { v16.4s }, [x27], #0x10\n" - "ld1 { v20.4s }, [x25], #0x10\n" - "ld1 { v24.4s }, [x23], #0x10\n" - "ld1 { v28.4s }, [x21], #0x10\n" - "ld1 { v9.4s }, [x13], #0x10\n" - "ld1 { v13.4s }, [x9], #0x10\n" - "ld1 { v17.4s }, [x27], #0x10\n" - "ld1 { v21.4s }, [x25], #0x10\n" - "ld1 { v25.4s }, [x23], #0x10\n" - "ld1 { v29.4s }, [x21], #0x10\n" - "tbz x16, #2, 186f\n" - "ld1 { v10.4s }, [x13], #0x10\n" - "ld1 { v14.4s }, [x9], #0x10\n" - "ld1 { v18.4s }, [x27], #0x10\n" - "ld1 { v22.4s }, [x25], #0x10\n" - "ld1 { v26.4s }, [x23], #0x10\n" - "ld1 { v30.4s }, [x21], #0x10\n" - "tbz x16, #1, 185f\n" + "b 189f\n" + "178:" // Height 6: no bias + "tbz %x[flags], #0, 188f\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "cmp x11, #0x10\n" + "add x24, x28, x19, LSL #2\n" + "add x23, x24, x19, LSL #2\n" + "add x22, x23, x19, LSL #2\n" + "add x21, x22, x19, LSL #2\n" + "add x20, x21, x19, LSL #2\n" + "bge 187f\n" + "tbz x11, #3, 182f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "ld1 { v16.4s }, [x23], #0x10\n" + "ld1 { v20.4s }, [x22], #0x10\n" + "ld1 { v24.4s }, [x21], #0x10\n" + "ld1 { v28.4s }, [x20], #0x10\n" + "ld1 { v9.4s }, [x28], #0x10\n" + "ld1 { v13.4s }, [x24], #0x10\n" + "ld1 { v17.4s }, [x23], #0x10\n" + "ld1 { v21.4s }, [x22], #0x10\n" + "ld1 { v25.4s }, [x21], #0x10\n" + "ld1 { v29.4s }, [x20], #0x10\n" + "tbz x11, #2, 180f\n" + "ld1 { v10.4s }, [x28], #0x10\n" + "ld1 { v14.4s }, [x24], #0x10\n" + "ld1 { v18.4s }, [x23], #0x10\n" + "ld1 { v22.4s }, [x22], #0x10\n" + "ld1 { v26.4s }, [x21], #0x10\n" + "ld1 { v30.4s }, [x20], #0x10\n" + "tbz x11, #1, 179f\n" "mov x19, #0x38\n" - "ldr d11, [x13], #0x8\n" - "ldr d15, [x9], #0x8\n" - "ldr d19, [x27], #0x8\n" - "ldr d23, [x25], #0x8\n" - "ldr d27, [x23], #0x8\n" - "ldr d31, [x21], #0x8\n" - "tbz x16, #0, 192f\n" - "ld1 { v11.s }[2], [x13]\n" - "ld1 { v15.s }[2], [x9]\n" - "ld1 { v19.s }[2], [x27]\n" - "ld1 { v23.s }[2], [x25]\n" - "ld1 { v27.s }[2], [x23]\n" - "ld1 { v31.s }[2], [x21]\n" - "b 192f\n" - "185:" // Height 6: Partial accumulate: partial_1_12 + "ldr d11, [x28], #0x8\n" + "ldr d15, [x24], #0x8\n" + "ldr d19, [x23], #0x8\n" + "ldr d23, [x22], #0x8\n" + "ldr d27, [x21], #0x8\n" + "ldr d31, [x20], #0x8\n" + "tbz x11, #0, 186f\n" + "ld1 { v11.s }[2], [x28]\n" + "ld1 { v15.s }[2], [x24]\n" + "ld1 { v19.s }[2], [x23]\n" + "ld1 { v23.s }[2], [x22]\n" + "ld1 { v27.s }[2], [x21]\n" + "ld1 { v31.s }[2], [x20]\n" + "b 186f\n" + "179:" // Height 6: Partial accumulate: partial_1_12 "mov x19, #0x30\n" - "tbz x16, #0, 192f\n" - "ldr s11, [x13, #0x0]\n" - "ldr s15, [x9, #0x0]\n" - "ldr s19, [x27, #0x0]\n" - "ldr s23, [x25, #0x0]\n" - "ldr s27, [x23, #0x0]\n" - "ldr s31, [x21, #0x0]\n" - "b 192f\n" - "186:" // Height 6: Partial accumulate: partial_2_8 - "tbz x16, #1, 187f\n" - "ldr d10, [x13], #0x8\n" - "ldr d14, [x9], #0x8\n" - "ldr d18, [x27], #0x8\n" - "ldr d22, [x25], #0x8\n" - "ldr d26, [x23], #0x8\n" - "ldr d30, [x21], #0x8\n" + "tbz x11, #0, 186f\n" + "ldr s11, [x28, #0x0]\n" + "ldr s15, [x24, #0x0]\n" + "ldr s19, [x23, #0x0]\n" + "ldr s23, [x22, #0x0]\n" + "ldr s27, [x21, #0x0]\n" + "ldr s31, [x20, #0x0]\n" + "b 186f\n" + "180:" // Height 6: Partial accumulate: partial_2_8 + "tbz x11, #1, 181f\n" + "ldr d10, [x28], #0x8\n" + "ldr d14, [x24], #0x8\n" "mov x19, #0x28\n" - "tbz x16, #0, 192f\n" - "ld1 { v10.s }[2], [x13]\n" - "ld1 { v14.s }[2], [x9]\n" - "ld1 { v18.s }[2], [x27]\n" - "ld1 { v22.s }[2], [x25]\n" - "ld1 { v26.s }[2], [x23]\n" - "ld1 { v30.s }[2], [x21]\n" - "b 192f\n" - "187:" // Height 6: Partial accumulate: partial_1_8 + "ldr d18, [x23], #0x8\n" + "ldr d22, [x22], #0x8\n" + "ldr d26, [x21], #0x8\n" + "ldr d30, [x20], #0x8\n" + "tbz x11, #0, 186f\n" + "ld1 { v10.s }[2], [x28]\n" + "ld1 { v14.s }[2], [x24]\n" + "ld1 { v18.s }[2], [x23]\n" + "ld1 { v22.s }[2], [x22]\n" + "ld1 { v26.s }[2], [x21]\n" + "ld1 { v30.s }[2], [x20]\n" + "b 186f\n" + "181:" // Height 6: Partial accumulate: partial_1_8 "mov x19, #0x20\n" - "tbz x16, #0, 192f\n" - "ldr s10, [x13, #0x0]\n" - "ldr s14, [x9, #0x0]\n" - "ldr s18, [x27, #0x0]\n" - "ldr s22, [x25, #0x0]\n" - "ldr s26, [x23, #0x0]\n" - "ldr s30, [x21, #0x0]\n" - "b 192f\n" - "188:" // Height 6: Partial accumulate: partial_4_0 - "tbz x16, #2, 190f\n" - "ld1 { v8.4s }, [x13], #0x10\n" - "ld1 { v12.4s }, [x9], #0x10\n" - "ld1 { v16.4s }, [x27], #0x10\n" - "ld1 { v20.4s }, [x25], #0x10\n" - "ld1 { v24.4s }, [x23], #0x10\n" - "ld1 { v28.4s }, [x21], #0x10\n" - "tbz x16, #1, 189f\n" + "tbz x11, #0, 186f\n" + "ldr s10, [x28, #0x0]\n" + "ldr s14, [x24, #0x0]\n" + "ldr s18, [x23, #0x0]\n" + "ldr s22, [x22, #0x0]\n" + "ldr s26, [x21, #0x0]\n" + "ldr s30, [x20, #0x0]\n" + "b 186f\n" + "182:" // Height 6: Partial accumulate: partial_4_0 + "tbz x11, #2, 184f\n" + "ld1 { v8.4s }, [x28], #0x10\n" + "ld1 { v12.4s }, [x24], #0x10\n" + "ld1 { v16.4s }, [x23], #0x10\n" + "ld1 { v20.4s }, [x22], #0x10\n" + "ld1 { v24.4s }, [x21], #0x10\n" + "ld1 { v28.4s }, [x20], #0x10\n" + "tbz x11, #1, 183f\n" "mov x19, #0x18\n" - "ldr d9, [x13], #0x8\n" - "ldr d13, [x9], #0x8\n" - "ldr d17, [x27], #0x8\n" - "ldr d21, [x25], #0x8\n" - "ldr d25, [x23], #0x8\n" - "ldr d29, [x21], #0x8\n" - "tbz x16, #0, 192f\n" - "ld1 { v9.s }[2], [x13]\n" - "ld1 { v13.s }[2], [x9]\n" - "ld1 { v17.s }[2], [x27]\n" - "ld1 { v21.s }[2], [x25]\n" - "ld1 { v25.s }[2], [x23]\n" - "ld1 { v29.s }[2], [x21]\n" - "b 192f\n" - "189:" // Height 6: Partial accumulate: partial_1_4 + "ldr d9, [x28], #0x8\n" + "ldr d13, [x24], #0x8\n" + "ldr d17, [x23], #0x8\n" + "ldr d21, [x22], #0x8\n" + "ldr d25, [x21], #0x8\n" + "ldr d29, [x20], #0x8\n" + "tbz x11, #0, 186f\n" + "ld1 { v9.s }[2], [x28]\n" + "ld1 { v13.s }[2], [x24]\n" + "ld1 { v17.s }[2], [x23]\n" + "ld1 { v21.s }[2], [x22]\n" + "ld1 { v25.s }[2], [x21]\n" + "ld1 { v29.s }[2], [x20]\n" + "b 186f\n" + "183:" // Height 6: Partial accumulate: partial_1_4 "mov x19, #0x10\n" - "tbz x16, #0, 192f\n" - "ldr s9, [x13, #0x0]\n" - "ldr s13, [x9, #0x0]\n" - "ldr s17, [x27, #0x0]\n" - "ldr s21, [x25, #0x0]\n" - "ldr s25, [x23, #0x0]\n" - "ldr s29, [x21, #0x0]\n" - "b 192f\n" - "190:" // Height 6: Partial accumulate: partial_2_0 - "tbz x16, #1, 191f\n" - "ldr d8, [x13], #0x8\n" - "ldr d12, [x9], #0x8\n" - "ldr d16, [x27], #0x8\n" - "ldr d20, [x25], #0x8\n" - "ldr d24, [x23], #0x8\n" - "ldr d28, [x21], #0x8\n" + "tbz x11, #0, 186f\n" + "ldr s9, [x28, #0x0]\n" + "ldr s13, [x24, #0x0]\n" + "ldr s17, [x23, #0x0]\n" + "ldr s21, [x22, #0x0]\n" + "ldr s25, [x21, #0x0]\n" + "ldr s29, [x20, #0x0]\n" + "b 186f\n" + "184:" // Height 6: Partial accumulate: partial_2_0 + "tbz x11, #1, 185f\n" + "ldr d8, [x28], #0x8\n" + "ldr d12, [x24], #0x8\n" "mov x19, #0x8\n" - "tbz x16, #0, 192f\n" - "ld1 { v8.s }[2], [x13]\n" - "ld1 { v12.s }[2], [x9]\n" - "ld1 { v16.s }[2], [x27]\n" - "ld1 { v20.s }[2], [x25]\n" - "ld1 { v24.s }[2], [x23]\n" - "ld1 { v28.s }[2], [x21]\n" - "b 192f\n" - "191:" // Height 6: Partial accumulate: partial_1_0 + "ldr d16, [x23], #0x8\n" + "ldr d20, [x22], #0x8\n" + "ldr d24, [x21], #0x8\n" + "ldr d28, [x20], #0x8\n" + "tbz x11, #0, 186f\n" + "ld1 { v8.s }[2], [x28]\n" + "ld1 { v12.s }[2], [x24]\n" + "ld1 { v16.s }[2], [x23]\n" + "ld1 { v20.s }[2], [x22]\n" + "ld1 { v24.s }[2], [x21]\n" + "ld1 { v28.s }[2], [x20]\n" + "b 186f\n" + "185:" // Height 6: Partial accumulate: partial_1_0 + "ldr s8, [x28, #0x0]\n" "mov x19, #0x0\n" - "ldr s8, [x13, #0x0]\n" - "ldr s12, [x9, #0x0]\n" - "ldr s16, [x27, #0x0]\n" - "ldr s20, [x25, #0x0]\n" - "ldr s24, [x23, #0x0]\n" - "ldr s28, [x21, #0x0]\n" - "192:" // Height 6: Partial accumulate: Done - "sub x13, x13, x19\n" - "sub x9, x9, x19\n" - "sub x27, x27, x19\n" - "sub x25, x25, x19\n" - "sub x23, x23, x19\n" - "sub x21, x21, x19\n" - "b 195f\n" - "193:" // Height 6: full accumulate - "ldr q8, [x13, #0x0]\n" - "ldr q9, [x13, #0x10]\n" - "ldr q10, [x13, #0x20]\n" - "ldr q11, [x13, #0x30]\n" - "ldr q12, [x9, #0x0]\n" - "ldr q13, [x9, #0x10]\n" - "ldr q14, [x9, #0x20]\n" - "ldr q15, [x9, #0x30]\n" - "ldr q16, [x27, #0x0]\n" - "ldr q17, [x27, #0x10]\n" - "ldr q18, [x27, #0x20]\n" - "ldr q19, [x27, #0x30]\n" - "ldr q20, [x25, #0x0]\n" - "ldr q21, [x25, #0x10]\n" - "ldr q22, [x25, #0x20]\n" - "ldr q23, [x25, #0x30]\n" - "ldr q24, [x23, #0x0]\n" - "ldr q25, [x23, #0x10]\n" - "ldr q26, [x23, #0x20]\n" - "ldr q27, [x23, #0x30]\n" - "ldr q28, [x21, #0x0]\n" - "ldr q29, [x21, #0x10]\n" - "ldr q30, [x21, #0x20]\n" - "ldr q31, [x21, #0x30]\n" - "b 195f\n" - "194:" // Height 6: no accumulate + "ldr s12, [x24, #0x0]\n" + "ldr s16, [x23, #0x0]\n" + "ldr s20, [x22, #0x0]\n" + "ldr s24, [x21, #0x0]\n" + "ldr s28, [x20, #0x0]\n" + "186:" // Height 6: Partial accumulate: Done + "sub x28, x28, x19\n" + "b 189f\n" + "187:" // Height 6: full accumulate + "ldr q8, [x28, #0x0]\n" + "ldr q9, [x28, #0x10]\n" + "ldr q10, [x28, #0x20]\n" + "ldr q11, [x28, #0x30]\n" + "ldr q12, [x24, #0x0]\n" + "ldr q13, [x24, #0x10]\n" + "ldr q14, [x24, #0x20]\n" + "ldr q15, [x24, #0x30]\n" + "ldr q16, [x23, #0x0]\n" + "ldr q17, [x23, #0x10]\n" + "ldr q18, [x23, #0x20]\n" + "ldr q19, [x23, #0x30]\n" + "ldr q20, [x22, #0x0]\n" + "ldr q21, [x22, #0x10]\n" + "ldr q22, [x22, #0x20]\n" + "ldr q23, [x22, #0x30]\n" + "ldr q24, [x21, #0x0]\n" + "ldr q25, [x21, #0x10]\n" + "ldr q26, [x21, #0x20]\n" + "ldr q27, [x21, #0x30]\n" + "ldr q28, [x20, #0x0]\n" + "ldr q29, [x20, #0x10]\n" + "ldr q30, [x20, #0x20]\n" + "ldr q31, [x20, #0x30]\n" + "b 189f\n" + "188:" // Height 6: no accumulate "movi v8.16b, #0x0\n" "movi v9.16b, #0x0\n" "movi v10.16b, #0x0\n" @@ -2968,299 +2897,299 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "movi v29.16b, #0x0\n" "movi v30.16b, #0x0\n" "movi v31.16b, #0x0\n" - "195:" // Height 6: setup done - "mov x12, #0x0\n" - "196:" // Height 6: String loop + "189:" // Height 6: setup done + "mov x27, #0x0\n" + "190:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w11, [x20, x12, LSL #0x2]\n" - "tbz %x[flags], #3, 197f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 191f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x26, [x20, #0x10]\n" - "ldr x24, [x20, #0x18]\n" - "ldr x22, [x20, #0x20]\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "ldr x22, [x20, #0x18]\n" + "ldr x21, [x20, #0x20]\n" "ldr x20, [x20, #0x28]\n" - "cbnz x12, 198f\n" + "cbnz x27, 192f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19, LSL #1\n" - "add x28, x28, x19, LSL #1\n" - "add x26, x26, x19, LSL #1\n" + "add x25, x25, x19, LSL #1\n" "add x24, x24, x19, LSL #1\n" + "add x23, x23, x19, LSL #1\n" "add x22, x22, x19, LSL #1\n" + "add x21, x21, x19, LSL #1\n" "add x20, x20, x19, LSL #1\n" - "b 198f\n" - "197:" // Height 6: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19, LSL #1\n" - "add x26, x28, x19, LSL #1\n" - "add x24, x26, x19, LSL #1\n" - "add x22, x24, x19, LSL #1\n" - "add x20, x22, x19, LSL #1\n" - "198:" // Height 6: input setup done - "cmp x11, #0x8\n" - "blt 201f\n" - "cmp x11, #0x10\n" - "blt 200f\n" - "199:" // Height 6: Multiply loop: Main loop head - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q3, [x24, #0x0]\n" - "ldr q4, [x22, #0x0]\n" + "b 192f\n" + "191:" // Height 6: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19, LSL #1\n" + "add x23, x24, x19, LSL #1\n" + "add x22, x23, x19, LSL #1\n" + "add x21, x22, x19, LSL #1\n" + "add x20, x21, x19, LSL #1\n" + "192:" // Height 6: input setup done + "cmp x26, #0x8\n" + "blt 195f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x10\n" + "ldr q2, [x23, #0x0]\n" + "ldr q3, [x22, #0x0]\n" + "ldr q4, [x21, #0x0]\n" "ldr q5, [x20, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "ldr q6, [x10, #0x0]\n" + "blt 194f\n" + "193:" // Height 6: Multiply loop: Main loop head ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" + "add x22, x22, #0x10\n" ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x22, #0x80]\n" + "add x21, x21, #0x10\n" ".inst 0x4f45f0dc // bfdot v28.4s, v6.8h, v5.h[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" + "add x20, x20, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "add x22, x22, #0x10\n" - "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x21, #0x80]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "add x20, x20, #0x10\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" "prfm pldl1keep, [x20, #0x80]\n" - "sub x11, x11, #0x8\n" + "cmp x26, #0x10\n" + ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "cmp x11, #0x10\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" ".inst 0x4f45f0fd // bfdot v29.4s, v7.8h, v5.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" ".inst 0x4f45f0de // bfdot v30.4s, v6.8h, v5.h[0]\n" - "ldr q6, [x15, #0x40]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" ".inst 0x4f45f0ff // bfdot v31.4s, v7.8h, v5.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" ".inst 0x4f64f0d8 // bfdot v24.4s, v6.8h, v4.h[1]\n" ".inst 0x4f65f0dc // bfdot v28.4s, v6.8h, v5.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" ".inst 0x4f64f0f9 // bfdot v25.4s, v7.8h, v4.h[1]\n" ".inst 0x4f65f0fd // bfdot v29.4s, v7.8h, v5.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" ".inst 0x4f64f0da // bfdot v26.4s, v6.8h, v4.h[1]\n" ".inst 0x4f65f0de // bfdot v30.4s, v6.8h, v5.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" ".inst 0x4f64f0fb // bfdot v27.4s, v7.8h, v4.h[1]\n" ".inst 0x4f65f0ff // bfdot v31.4s, v7.8h, v5.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" ".inst 0x4f44f8d8 // bfdot v24.4s, v6.8h, v4.h[2]\n" ".inst 0x4f45f8dc // bfdot v28.4s, v6.8h, v5.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" ".inst 0x4f44f8f9 // bfdot v25.4s, v7.8h, v4.h[2]\n" ".inst 0x4f45f8fd // bfdot v29.4s, v7.8h, v5.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" ".inst 0x4f44f8da // bfdot v26.4s, v6.8h, v4.h[2]\n" ".inst 0x4f45f8de // bfdot v30.4s, v6.8h, v5.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" ".inst 0x4f44f8fb // bfdot v27.4s, v7.8h, v4.h[2]\n" ".inst 0x4f45f8ff // bfdot v31.4s, v7.8h, v5.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" ".inst 0x4f64f8d8 // bfdot v24.4s, v6.8h, v4.h[3]\n" ".inst 0x4f65f8dc // bfdot v28.4s, v6.8h, v5.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" ".inst 0x4f64f8f9 // bfdot v25.4s, v7.8h, v4.h[3]\n" ".inst 0x4f65f8fd // bfdot v29.4s, v7.8h, v5.h[3]\n" - "ldr q7, [x15, #0xf0]\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - "add x15, x15, #0x100\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d6 // bfdot v22.4s, v6.8h, v3.h[3]\n" ".inst 0x4f64f8da // bfdot v26.4s, v6.8h, v4.h[3]\n" ".inst 0x4f65f8de // bfdot v30.4s, v6.8h, v5.h[3]\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" + "ldr q1, [x24, #0x0]\n" ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" + "ldr q2, [x23, #0x0]\n" ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" + "ldr q3, [x22, #0x0]\n" ".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n" + "ldr q4, [x21, #0x0]\n" ".inst 0x4f65f8ff // bfdot v31.4s, v7.8h, v5.h[3]\n" - "bge 199b\n" - "200:" // Height 6: Multiply loop: Single iteration only - "sub x11, x11, #0x8\n" - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q3, [x24, #0x0]\n" - "ldr q4, [x22, #0x0]\n" "ldr q5, [x20, #0x0]\n" - "ldr q6, [x15, #0x0]\n" + "bge 193b\n" + "194:" // Height 6: Multiply loop: Single iteration only ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" + "sub x26, x26, #0x8\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "add x28, x28, #0x10\n" - ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "add x26, x26, #0x10\n" - ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" "add x24, x24, #0x10\n" - ".inst 0x4f45f0dc // bfdot v28.4s, v6.8h, v5.h[0]\n" + ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x15, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" + "add x23, x23, #0x10\n" + ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" + "prfm pldl1keep, [x23, #0x80]\n" "add x22, x22, #0x10\n" + ".inst 0x4f45f0dc // bfdot v28.4s, v6.8h, v5.h[0]\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" + "add x21, x21, #0x10\n" + ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" + "ldr q6, [x10, #0x20]\n" "add x20, x20, #0x10\n" + ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" + "prfm pldl1keep, [x21, #0x80]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" "prfm pldl1keep, [x20, #0x80]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" ".inst 0x4f45f0fd // bfdot v29.4s, v7.8h, v5.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" ".inst 0x4f45f0de // bfdot v30.4s, v6.8h, v5.h[0]\n" - "ldr q6, [x15, #0x40]\n" + "ldr q6, [x10, #0x40]\n" ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" ".inst 0x4f45f0ff // bfdot v31.4s, v7.8h, v5.h[0]\n" - "ldr q7, [x15, #0x50]\n" + "ldr q7, [x10, #0x50]\n" ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" ".inst 0x4f64f0d8 // bfdot v24.4s, v6.8h, v4.h[1]\n" ".inst 0x4f65f0dc // bfdot v28.4s, v6.8h, v5.h[1]\n" - "ldr q6, [x15, #0x60]\n" + "ldr q6, [x10, #0x60]\n" ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" ".inst 0x4f64f0f9 // bfdot v25.4s, v7.8h, v4.h[1]\n" ".inst 0x4f65f0fd // bfdot v29.4s, v7.8h, v5.h[1]\n" - "ldr q7, [x15, #0x70]\n" + "ldr q7, [x10, #0x70]\n" ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" ".inst 0x4f64f0da // bfdot v26.4s, v6.8h, v4.h[1]\n" ".inst 0x4f65f0de // bfdot v30.4s, v6.8h, v5.h[1]\n" - "ldr q6, [x15, #0x80]\n" + "ldr q6, [x10, #0x80]\n" ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" ".inst 0x4f64f0fb // bfdot v27.4s, v7.8h, v4.h[1]\n" ".inst 0x4f65f0ff // bfdot v31.4s, v7.8h, v5.h[1]\n" - "ldr q7, [x15, #0x90]\n" + "ldr q7, [x10, #0x90]\n" ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" ".inst 0x4f44f8d8 // bfdot v24.4s, v6.8h, v4.h[2]\n" ".inst 0x4f45f8dc // bfdot v28.4s, v6.8h, v5.h[2]\n" - "ldr q6, [x15, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" ".inst 0x4f44f8f9 // bfdot v25.4s, v7.8h, v4.h[2]\n" ".inst 0x4f45f8fd // bfdot v29.4s, v7.8h, v5.h[2]\n" - "ldr q7, [x15, #0xb0]\n" + "ldr q7, [x10, #0xb0]\n" ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" ".inst 0x4f44f8da // bfdot v26.4s, v6.8h, v4.h[2]\n" ".inst 0x4f45f8de // bfdot v30.4s, v6.8h, v5.h[2]\n" - "ldr q6, [x15, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" ".inst 0x4f44f8fb // bfdot v27.4s, v7.8h, v4.h[2]\n" ".inst 0x4f45f8ff // bfdot v31.4s, v7.8h, v5.h[2]\n" - "ldr q7, [x15, #0xd0]\n" + "ldr q7, [x10, #0xd0]\n" ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" ".inst 0x4f64f8d8 // bfdot v24.4s, v6.8h, v4.h[3]\n" ".inst 0x4f65f8dc // bfdot v28.4s, v6.8h, v5.h[3]\n" - "ldr q6, [x15, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" ".inst 0x4f64f8f9 // bfdot v25.4s, v7.8h, v4.h[3]\n" ".inst 0x4f65f8fd // bfdot v29.4s, v7.8h, v5.h[3]\n" - "ldr q7, [x15, #0xf0]\n" + "ldr q7, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - "add x15, x15, #0x100\n" ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" ".inst 0x4f63f8d6 // bfdot v22.4s, v6.8h, v3.h[3]\n" @@ -3272,37 +3201,37 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" ".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n" ".inst 0x4f65f8ff // bfdot v31.4s, v7.8h, v5.h[3]\n" - "201:" // Height 6: Multiply loop: Main loop skip - "cbz x11, 205f\n" - "cmp x11, #0x2\n" - "blt 203f\n" - "202:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x10], #0x4\n" - "ldr s1, [x28], #0x4\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x24], #0x4\n" - "ldr s4, [x22], #0x4\n" + "195:" // Height 6: Multiply loop: Main loop skip + "cbz x26, 199f\n" + "cmp x26, #0x2\n" + "blt 197f\n" + "196:" // Height 6: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x2\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x2\n" + "ldr s2, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr s4, [x21], #0x4\n" "ldr s5, [x20], #0x4\n" - "ldr q6, [x15, #0x0]\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "sub x11, x11, #0x2\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "cmp x11, #0x2\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" ".inst 0x4f45f0dc // bfdot v28.4s, v6.8h, v5.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" ".inst 0x4f45f0fd // bfdot v29.4s, v7.8h, v5.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" @@ -3314,34 +3243,34 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" ".inst 0x4f45f0ff // bfdot v31.4s, v7.8h, v5.h[0]\n" - "bge 202b\n" - "cbz x11, 205f\n" - "203:" // Height 6: Multiply loop: Skip odd blocks - "ldr h0, [x10, #0x0]\n" - "ldr h1, [x28, #0x0]\n" - "ldr h2, [x26, #0x0]\n" - "ldr h3, [x24, #0x0]\n" - "ldr h4, [x22, #0x0]\n" + "bge 196b\n" + "cbz x26, 199f\n" + "197:" // Height 6: Multiply loop: Skip odd blocks + "ldr h0, [x25, #0x0]\n" + "ldr h1, [x24, #0x0]\n" + "ldr h2, [x23, #0x0]\n" + "ldr h3, [x22, #0x0]\n" + "ldr h4, [x21, #0x0]\n" "ldr h5, [x20, #0x0]\n" - "204:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q6, [x15, #0x0]\n" + "198:" // Height 6: Multiply loop: Ragged operand read: Done + "ldr q6, [x10, #0x0]\n" ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q7, [x15, #0x10]\n" + "ldr q7, [x10, #0x10]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" ".inst 0x4f45f0dc // bfdot v28.4s, v6.8h, v5.h[0]\n" - "ldr q6, [x15, #0x20]\n" + "ldr q6, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" ".inst 0x4f45f0fd // bfdot v29.4s, v7.8h, v5.h[0]\n" - "ldr q7, [x15, #0x30]\n" + "ldr q7, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "add x15, x15, #0x40\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" @@ -3353,18 +3282,24 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" ".inst 0x4f45f0ff // bfdot v31.4s, v7.8h, v5.h[0]\n" - "205:" // Height 6: Multiply loop: No odd multiplies + "199:" // Height 6: Multiply loop: No odd multiplies "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x12, x12, #0x1\n" - "cmp x12, x19\n" - "bne 196b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x27, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 190b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "prfm pstl1keep, [x28, #0x0]\n" + "add x24, x28, x19, LSL #2\n" + "prfm pstl1keep, [x24, #0x0]\n" + "add x23, x24, x19, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" + "add x22, x23, x19, LSL #2\n" + "prfm pstl1keep, [x22, #0x0]\n" + "add x21, x22, x19, LSL #2\n" "prfm pstl1keep, [x21, #0x0]\n" - "tbz %x[flags], #1, 206f\n" + "add x20, x21, x19, LSL #2\n" + "prfm pstl1keep, [x20, #0x0]\n" + "tbz %x[flags], #1, 200f\n" "add x19, %x[args_ptr], %[offset_min]\n" "ld1r { v1.4s }, [x19]\n" "add x19, %x[args_ptr], %[offset_max]\n" @@ -3417,185 +3352,180 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "fmin v31.4s, v31.4s, v0.4s\n" "fmax v30.4s, v30.4s, v1.4s\n" "fmax v31.4s, v31.4s, v1.4s\n" - "206:" // Height 6: No activation - "cmp x16, #0x10\n" - "bge 215f\n" - "tbz x16, #3, 210f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v9.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "st1 { v13.4s }, [x9], #0x10\n" - "st1 { v16.4s }, [x27], #0x10\n" - "st1 { v17.4s }, [x27], #0x10\n" - "st1 { v20.4s }, [x25], #0x10\n" - "st1 { v21.4s }, [x25], #0x10\n" - "st1 { v24.4s }, [x23], #0x10\n" - "st1 { v25.4s }, [x23], #0x10\n" - "st1 { v28.4s }, [x21], #0x10\n" - "st1 { v29.4s }, [x21], #0x10\n" - "tbz x16, #2, 208f\n" - "st1 { v10.4s }, [x13], #0x10\n" - "st1 { v14.4s }, [x9], #0x10\n" - "st1 { v18.4s }, [x27], #0x10\n" - "st1 { v22.4s }, [x25], #0x10\n" - "st1 { v26.4s }, [x23], #0x10\n" - "st1 { v30.4s }, [x21], #0x10\n" - "tbz x16, #1, 207f\n" - "str d11, [x13], #0x8\n" - "str d15, [x9], #0x8\n" - "str d19, [x27], #0x8\n" - "str d23, [x25], #0x8\n" - "str d27, [x23], #0x8\n" - "str d31, [x21], #0x8\n" - "tbz x16, #0, 214f\n" - "st1 { v11.s }[2], [x13]\n" - "st1 { v15.s }[2], [x9]\n" - "st1 { v19.s }[2], [x27]\n" - "st1 { v23.s }[2], [x25]\n" - "st1 { v27.s }[2], [x23]\n" - "st1 { v31.s }[2], [x21]\n" - "b 214f\n" - "207:" // Height 6: Partial direct writeback: partial_1_12 - "tbz x16, #0, 214f\n" - "str s11, [x13, #0x0]\n" - "str s15, [x9, #0x0]\n" - "str s19, [x27, #0x0]\n" - "str s23, [x25, #0x0]\n" - "str s27, [x23, #0x0]\n" - "str s31, [x21, #0x0]\n" - "b 214f\n" - "208:" // Height 6: Partial direct writeback: partial_2_8 - "tbz x16, #1, 209f\n" - "str d10, [x13], #0x8\n" - "str d14, [x9], #0x8\n" - "str d18, [x27], #0x8\n" - "str d22, [x25], #0x8\n" - "str d26, [x23], #0x8\n" - "str d30, [x21], #0x8\n" - "tbz x16, #0, 214f\n" - "st1 { v10.s }[2], [x13]\n" - "st1 { v14.s }[2], [x9]\n" - "st1 { v18.s }[2], [x27]\n" - "st1 { v22.s }[2], [x25]\n" - "st1 { v26.s }[2], [x23]\n" - "st1 { v30.s }[2], [x21]\n" - "b 214f\n" - "209:" // Height 6: Partial direct writeback: partial_1_8 - "tbz x16, #0, 214f\n" - "str s10, [x13, #0x0]\n" - "str s14, [x9, #0x0]\n" - "str s18, [x27, #0x0]\n" - "str s22, [x25, #0x0]\n" - "str s26, [x23, #0x0]\n" - "str s30, [x21, #0x0]\n" - "b 214f\n" - "210:" // Height 6: Partial direct writeback: partial_4_0 - "tbz x16, #2, 212f\n" - "st1 { v8.4s }, [x13], #0x10\n" - "st1 { v12.4s }, [x9], #0x10\n" - "st1 { v16.4s }, [x27], #0x10\n" - "st1 { v20.4s }, [x25], #0x10\n" - "st1 { v24.4s }, [x23], #0x10\n" - "st1 { v28.4s }, [x21], #0x10\n" - "tbz x16, #1, 211f\n" - "str d9, [x13], #0x8\n" - "str d13, [x9], #0x8\n" - "str d17, [x27], #0x8\n" - "str d21, [x25], #0x8\n" - "str d25, [x23], #0x8\n" - "str d29, [x21], #0x8\n" - "tbz x16, #0, 214f\n" - "st1 { v9.s }[2], [x13]\n" - "st1 { v13.s }[2], [x9]\n" - "st1 { v17.s }[2], [x27]\n" - "st1 { v21.s }[2], [x25]\n" - "st1 { v25.s }[2], [x23]\n" - "st1 { v29.s }[2], [x21]\n" - "b 214f\n" - "211:" // Height 6: Partial direct writeback: partial_1_4 - "tbz x16, #0, 214f\n" - "str s9, [x13, #0x0]\n" - "str s13, [x9, #0x0]\n" - "str s17, [x27, #0x0]\n" - "str s21, [x25, #0x0]\n" - "str s25, [x23, #0x0]\n" - "str s29, [x21, #0x0]\n" - "b 214f\n" - "212:" // Height 6: Partial direct writeback: partial_2_0 - "tbz x16, #1, 213f\n" - "str d8, [x13], #0x8\n" - "str d12, [x9], #0x8\n" - "str d16, [x27], #0x8\n" - "str d20, [x25], #0x8\n" - "str d24, [x23], #0x8\n" - "str d28, [x21], #0x8\n" - "tbz x16, #0, 214f\n" - "st1 { v8.s }[2], [x13]\n" - "st1 { v12.s }[2], [x9]\n" - "st1 { v16.s }[2], [x27]\n" - "st1 { v20.s }[2], [x25]\n" - "st1 { v24.s }[2], [x23]\n" - "st1 { v28.s }[2], [x21]\n" - "b 214f\n" - "213:" // Height 6: Partial direct writeback: partial_1_0 - "str s8, [x13, #0x0]\n" - "str s12, [x9, #0x0]\n" - "str s16, [x27, #0x0]\n" - "str s20, [x25, #0x0]\n" - "str s24, [x23, #0x0]\n" - "str s28, [x21, #0x0]\n" - "214:" // Height 6: Partial direct writeback: Done - "b 216f\n" - "215:" // Height 6: Full writeback - "str q8, [x13, #0x0]\n" - "str q9, [x13, #0x10]\n" - "str q10, [x13, #0x20]\n" - "str q11, [x13, #0x30]\n" - "str q12, [x9, #0x0]\n" - "str q13, [x9, #0x10]\n" - "str q14, [x9, #0x20]\n" - "str q15, [x9, #0x30]\n" - "str q16, [x27, #0x0]\n" - "str q17, [x27, #0x10]\n" - "str q18, [x27, #0x20]\n" - "str q19, [x27, #0x30]\n" - "str q20, [x25, #0x0]\n" - "str q21, [x25, #0x10]\n" - "str q22, [x25, #0x20]\n" - "str q23, [x25, #0x30]\n" - "str q24, [x23, #0x0]\n" - "str q25, [x23, #0x10]\n" - "str q26, [x23, #0x20]\n" - "str q27, [x23, #0x30]\n" - "str q28, [x21, #0x0]\n" - "str q29, [x21, #0x10]\n" - "str q30, [x21, #0x20]\n" - "str q31, [x21, #0x30]\n" - "add x13, x13, #0x40\n" - "add x9, x9, #0x40\n" - "add x27, x27, #0x40\n" - "add x25, x25, #0x40\n" - "add x23, x23, #0x40\n" - "add x21, x21, #0x40\n" - "216:" // Height 6: Writeback done - "subs x16, x16, #0x10\n" - "bgt 183b\n" + "200:" // Height 6: No activation + "cmp x11, #0x10\n" + "bge 209f\n" + "tbz x11, #3, 204f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v9.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "st1 { v13.4s }, [x24], #0x10\n" + "st1 { v16.4s }, [x23], #0x10\n" + "st1 { v17.4s }, [x23], #0x10\n" + "st1 { v20.4s }, [x22], #0x10\n" + "st1 { v21.4s }, [x22], #0x10\n" + "st1 { v24.4s }, [x21], #0x10\n" + "st1 { v25.4s }, [x21], #0x10\n" + "st1 { v28.4s }, [x20], #0x10\n" + "st1 { v29.4s }, [x20], #0x10\n" + "tbz x11, #2, 202f\n" + "st1 { v10.4s }, [x28], #0x10\n" + "st1 { v14.4s }, [x24], #0x10\n" + "st1 { v18.4s }, [x23], #0x10\n" + "st1 { v22.4s }, [x22], #0x10\n" + "st1 { v26.4s }, [x21], #0x10\n" + "st1 { v30.4s }, [x20], #0x10\n" + "tbz x11, #1, 201f\n" + "str d11, [x28], #0x8\n" + "str d15, [x24], #0x8\n" + "str d19, [x23], #0x8\n" + "str d23, [x22], #0x8\n" + "str d27, [x21], #0x8\n" + "str d31, [x20], #0x8\n" + "tbz x11, #0, 208f\n" + "st1 { v11.s }[2], [x28]\n" + "st1 { v15.s }[2], [x24]\n" + "st1 { v19.s }[2], [x23]\n" + "st1 { v23.s }[2], [x22]\n" + "st1 { v27.s }[2], [x21]\n" + "st1 { v31.s }[2], [x20]\n" + "b 208f\n" + "201:" // Height 6: Partial direct writeback: partial_1_12 + "tbz x11, #0, 208f\n" + "str s11, [x28, #0x0]\n" + "str s15, [x24, #0x0]\n" + "str s19, [x23, #0x0]\n" + "str s23, [x22, #0x0]\n" + "str s27, [x21, #0x0]\n" + "str s31, [x20, #0x0]\n" + "b 208f\n" + "202:" // Height 6: Partial direct writeback: partial_2_8 + "tbz x11, #1, 203f\n" + "str d10, [x28], #0x8\n" + "str d14, [x24], #0x8\n" + "str d18, [x23], #0x8\n" + "str d22, [x22], #0x8\n" + "str d26, [x21], #0x8\n" + "str d30, [x20], #0x8\n" + "tbz x11, #0, 208f\n" + "st1 { v10.s }[2], [x28]\n" + "st1 { v14.s }[2], [x24]\n" + "st1 { v18.s }[2], [x23]\n" + "st1 { v22.s }[2], [x22]\n" + "st1 { v26.s }[2], [x21]\n" + "st1 { v30.s }[2], [x20]\n" + "b 208f\n" + "203:" // Height 6: Partial direct writeback: partial_1_8 + "tbz x11, #0, 208f\n" + "str s10, [x28, #0x0]\n" + "str s14, [x24, #0x0]\n" + "str s18, [x23, #0x0]\n" + "str s22, [x22, #0x0]\n" + "str s26, [x21, #0x0]\n" + "str s30, [x20, #0x0]\n" + "b 208f\n" + "204:" // Height 6: Partial direct writeback: partial_4_0 + "tbz x11, #2, 206f\n" + "st1 { v8.4s }, [x28], #0x10\n" + "st1 { v12.4s }, [x24], #0x10\n" + "st1 { v16.4s }, [x23], #0x10\n" + "st1 { v20.4s }, [x22], #0x10\n" + "st1 { v24.4s }, [x21], #0x10\n" + "st1 { v28.4s }, [x20], #0x10\n" + "tbz x11, #1, 205f\n" + "str d9, [x28], #0x8\n" + "str d13, [x24], #0x8\n" + "str d17, [x23], #0x8\n" + "str d21, [x22], #0x8\n" + "str d25, [x21], #0x8\n" + "str d29, [x20], #0x8\n" + "tbz x11, #0, 208f\n" + "st1 { v9.s }[2], [x28]\n" + "st1 { v13.s }[2], [x24]\n" + "st1 { v17.s }[2], [x23]\n" + "st1 { v21.s }[2], [x22]\n" + "st1 { v25.s }[2], [x21]\n" + "st1 { v29.s }[2], [x20]\n" + "b 208f\n" + "205:" // Height 6: Partial direct writeback: partial_1_4 + "tbz x11, #0, 208f\n" + "str s9, [x28, #0x0]\n" + "str s13, [x24, #0x0]\n" + "str s17, [x23, #0x0]\n" + "str s21, [x22, #0x0]\n" + "str s25, [x21, #0x0]\n" + "str s29, [x20, #0x0]\n" + "b 208f\n" + "206:" // Height 6: Partial direct writeback: partial_2_0 + "tbz x11, #1, 207f\n" + "str d8, [x28], #0x8\n" + "str d12, [x24], #0x8\n" + "str d16, [x23], #0x8\n" + "str d20, [x22], #0x8\n" + "str d24, [x21], #0x8\n" + "str d28, [x20], #0x8\n" + "tbz x11, #0, 208f\n" + "st1 { v8.s }[2], [x28]\n" + "st1 { v12.s }[2], [x24]\n" + "st1 { v16.s }[2], [x23]\n" + "st1 { v20.s }[2], [x22]\n" + "st1 { v24.s }[2], [x21]\n" + "st1 { v28.s }[2], [x20]\n" + "b 208f\n" + "207:" // Height 6: Partial direct writeback: partial_1_0 + "str s8, [x28, #0x0]\n" + "str s12, [x24, #0x0]\n" + "str s16, [x23, #0x0]\n" + "str s20, [x22, #0x0]\n" + "str s24, [x21, #0x0]\n" + "str s28, [x20, #0x0]\n" + "208:" // Height 6: Partial direct writeback: Done + "b 210f\n" + "209:" // Height 6: Full writeback + "str q8, [x28, #0x0]\n" + "str q9, [x28, #0x10]\n" + "str q10, [x28, #0x20]\n" + "str q11, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + "str q12, [x24, #0x0]\n" + "str q13, [x24, #0x10]\n" + "str q14, [x24, #0x20]\n" + "str q15, [x24, #0x30]\n" + "str q16, [x23, #0x0]\n" + "str q17, [x23, #0x10]\n" + "str q18, [x23, #0x20]\n" + "str q19, [x23, #0x30]\n" + "str q20, [x22, #0x0]\n" + "str q21, [x22, #0x10]\n" + "str q22, [x22, #0x20]\n" + "str q23, [x22, #0x30]\n" + "str q24, [x21, #0x0]\n" + "str q25, [x21, #0x10]\n" + "str q26, [x21, #0x20]\n" + "str q27, [x21, #0x30]\n" + "str q28, [x20, #0x0]\n" + "str q29, [x20, #0x10]\n" + "str q30, [x20, #0x20]\n" + "str q31, [x20, #0x30]\n" + "210:" // Height 6: Writeback done + "subs x11, x11, #0x10\n" + "bgt 177b\n" "subs %x[M], %x[M], #0x6\n" - "beq 218f\n" + "beq 212f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" - "tbz %x[flags], #3, 217f\n" + "tbz %x[flags], #3, 211f\n" "add x20, x20, #0x6\n" "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "b 1b\n" - "217:" // Update direct input + "211:" // Update direct input "mov x19, #0xc\n" "madd %x[input_ptr], x19, x20, %x[input_ptr]\n" "b 1b\n" - "218:" // Exit + "212:" // Exit : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |