diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp | 3289 |
1 files changed, 1607 insertions, 1682 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp index 9847e6553b..5a4df161aa 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_dot_6x16/generic.cpp @@ -88,210 +88,204 @@ void a64_hybrid_s8qs_dot_6x16 ( "1:" // Row loop "cmp %x[M], #0x6\n" - "bge 141f\n" + "bge 136f\n" "cmp %x[M], #0x4\n" - "bgt 113f\n" - "beq 85f\n" + "bgt 109f\n" + "beq 82f\n" "cmp %x[M], #0x2\n" - "bgt 57f\n" - "beq 29f\n" - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\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\n" - "b 3f\n" - "2:" // Height 1: setup direct output - "mov x13, %x[output_ptr]\n" - "3:" // Height 1: Column loop + "bgt 55f\n" + "beq 28f\n" + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "2:" // Height 1: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" "movi v10.4s, #0x0\n" "movi v11.4s, #0x0\n" - "4:" // Height 1: setup done - "mov x12, #0x0\n" - "5:" // Height 1: String loop + "3:" // Height 1: setup done + "mov x27, #0x0\n" + "4:" // 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, 6f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 5f\n" + "ldr x20, [%x[input_ptr], x27, LSL #0x3]\n" "add x20, x20, x19, LSL #3\n" - "ldr x10, [x20, #0x0]\n" - "cbnz x12, 7f\n" + "ldr x25, [x20, #0x0]\n" + "cbnz x27, 6f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "b 7f\n" - "6:" // Height 1: setup direct input - "mov x10, %x[input_ptr]\n" - "7:" // Height 1: input setup done - "cmp x11, #0x10\n" - "blt 10f\n" - "cmp x11, #0x20\n" + "add x25, x25, x19\n" + "b 6f\n" + "5:" // Height 1: setup direct input + "mov x25, %x[input_ptr]\n" + "6:" // Height 1: input setup done + "cmp x26, #0x10\n" "blt 9f\n" - "8:" // Height 1: Multiply loop: Main loop head - "ldr q0, [x10, #0x0]\n" - "ldr q6, [x14, #0x0]\n" + "ldr q0, [x25, #0x0]\n" + "ldr q6, [x28, #0x0]\n" + "cmp x26, #0x20\n" + "blt 8f\n" + "7:" // Height 1: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q7, [x28, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q6, [x28, #0x20]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x14, #0x40]\n" - "add x10, x10, #0x10\n" + "ldr q7, [x28, #0x30]\n" + "cmp x26, #0x20\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q6, [x28, #0x40]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x14, #0x60]\n" - "sub x11, x11, #0x10\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x14, #0x70]\n" - "cmp x11, #0x20\n" + "ldr q7, [x28, #0x70]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" - "add x14, x14, #0x100\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" - "bge 8b\n" - "9:" // Height 1: Multiply loop: Single iteration only - "sub x11, x11, #0x10\n" - "ldr q0, [x10, #0x0]\n" - "ldr q6, [x14, #0x0]\n" + "ldr q0, [x25, #0x0]\n" + "bge 7b\n" + "8:" // Height 1: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q7, [x28, #0x10]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q6, [x28, #0x20]\n" + "add x25, x25, #0x10\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x14, #0x40]\n" - "add x10, x10, #0x10\n" + "ldr q7, [x28, #0x30]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" - "add x14, x14, #0x100\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" - "10:" // Height 1: Multiply loop: Main loop skip - "cbz x11, 15f\n" - "cmp x11, #0x4\n" - "blt 12f\n" - "11:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x10], #0x4\n" - "ldr q6, [x14, #0x0]\n" + "9:" // Height 1: Multiply loop: Main loop skip + "cbz x26, 14f\n" + "cmp x26, #0x4\n" + "blt 11f\n" + "10:" // Height 1: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x4\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q7, [x28, #0x10]\n" + "cmp x26, #0x4\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q6, [x28, #0x20]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "sub x11, x11, #0x4\n" - "add x14, x14, #0x40\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "cmp x11, #0x4\n" - "bge 11b\n" - "cbz x11, 15f\n" - "12:" // Height 1: Multiply loop: Skip odd blocks - "tbz x11, #1, 13f\n" - "ldr h0, [x10], #0x2\n" - "tbz x11, #0, 14f\n" - "ld1 { v0.b }[2], [x10]\n" - "b 14f\n" - "13:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x10, #0x0]\n" - "14:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q6, [x14, #0x0]\n" + "bge 10b\n" + "cbz x26, 14f\n" + "11:" // Height 1: Multiply loop: Skip odd blocks + "tbz x26, #1, 12f\n" + "ldr h0, [x25], #0x2\n" + "tbz x26, #0, 13f\n" + "ld1 { v0.b }[2], [x25]\n" + "b 13f\n" + "12:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 + "ldr b0, [x25, #0x0]\n" + "13:" // Height 1: Multiply loop: Ragged operand read: Done + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q7, [x28, #0x10]\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "15:" // Height 1: Multiply loop: No odd multiplies + "14:" // 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 5b\n" - "prfm pstl1keep, [x13, #0x0]\n" - "ldr q0, [x16, #0x0]\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 4b\n" + "prfm pstl1keep, [x9, #0x0]\n" + "ldr q0, [x11, #0x0]\n" "add v8.4s, v8.4s, v0.4s\n" - "ldr q1, [x16, #0x10]\n" - "ldr q2, [x16, #0x20]\n" + "ldr q1, [x11, #0x10]\n" + "ldr q2, [x11, #0x20]\n" "add v9.4s, v9.4s, v1.4s\n" - "ldr q3, [x16, #0x30]\n" + "ldr q3, [x11, #0x30]\n" + "add x11, x11, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" - "add x16, x16, #0x40\n" "add v11.4s, v11.4s, v3.4s\n" - "tbz %x[flags], #4, 16f\n" - "ldr q0, [x17, #0x0]\n" - "ldr q4, [x8, #0x0]\n" - "ldr q1, [x17, #0x10]\n" - "ldr q5, [x8, #0x10]\n" - "ldr q2, [x17, #0x20]\n" - "ldr q6, [x8, #0x20]\n" - "ldr q3, [x17, #0x30]\n" - "ldr q7, [x8, #0x30]\n" - "add x17, x17, #0x40\n" - "add x8, x8, #0x40\n" - "b 17f\n" - "16:" // Height 1: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x19]\n" + "tbz %x[flags], #4, 15f\n" + "ldr q0, [x12, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q1, [x12, #0x10]\n" + "ldr q5, [x13, #0x10]\n" + "ldr q2, [x12, #0x20]\n" + "ldr q6, [x13, #0x20]\n" + "ldr q3, [x12, #0x30]\n" + "add x12, x12, #0x40\n" + "ldr q7, [x13, #0x30]\n" + "add x13, x13, #0x40\n" + "b 16f\n" + "15:" // Height 1: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1r { v0.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1r { v4.4s }, [x24]\n" "mov v2.16b, v0.16b\n" "mov v3.16b, v0.16b\n" "mov v5.16b, v4.16b\n" "mov v6.16b, v4.16b\n" "mov v7.16b, v4.16b\n" - "17:" // Height 1: parameters loaded + "16:" // Height 1: parameters loaded "sqrdmulh v8.4s, v8.4s, v4.4s\n" "sqrdmulh v9.4s, v9.4s, v5.4s\n" "sqrdmulh v10.4s, v10.4s, v6.4s\n" "sqrdmulh v11.4s, v11.4s, v7.4s\n" - "tbz %x[flags], #5, 18f\n" + "tbz %x[flags], #5, 17f\n" "and v4.16b, v8.16b, v0.16b\n" "sshr v4.4s, v4.4s, #0x1f\n" "and v5.16b, v9.16b, v1.16b\n" @@ -304,18 +298,18 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqadd v9.4s, v9.4s, v5.4s\n" "sqadd v10.4s, v10.4s, v6.4s\n" "sqadd v11.4s, v11.4s, v7.4s\n" - "18:" // Height 1: no shift correction + "17:" // Height 1: no shift correction "srshl v8.4s, v8.4s, v0.4s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1r { v4.4s }, [x24]\n" "srshl v9.4s, v9.4s, v1.4s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" "srshl v10.4s, v10.4s, v2.4s\n" - "ld1r { v5.4s }, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1r { v5.4s }, [x24]\n" + "add x24, %x[qp], %[maxval]\n" "srshl v11.4s, v11.4s, v3.4s\n" - "ld1r { v6.4s }, [x19]\n" - "cmp x15, #0x10\n" + "ld1r { v6.4s }, [x24]\n" + "cmp x10, #0x10\n" "add v8.4s, v8.4s, v4.4s\n" "add v9.4s, v9.4s, v4.4s\n" "add v10.4s, v10.4s, v4.4s\n" @@ -331,76 +325,67 @@ void a64_hybrid_s8qs_dot_6x16 ( "smax v11.4s, v11.4s, v5.4s\n" "uzp1 v9.8h, v10.8h, v11.8h\n" "uzp1 v8.16b, v8.16b, v9.16b\n" - "bge 27f\n" - "tbz x15, #3, 22f\n" - "str d8, [x13], #0x8\n" - "tbz x15, #2, 20f\n" - "st1 { v8.s }[2], [x13], #0x4\n" - "tbz x15, #1, 19f\n" - "st1 { v8.h }[6], [x13], #0x2\n" - "tbz x15, #0, 26f\n" - "st1 { v8.b }[14], [x13]\n" - "b 26f\n" - "19:" // Height 1: Partial direct writeback: partial_1_12 - "tbz x15, #0, 26f\n" - "st1 { v8.b }[12], [x13]\n" - "b 26f\n" - "20:" // Height 1: Partial direct writeback: partial_2_8 - "tbz x15, #1, 21f\n" - "st1 { v8.h }[4], [x13], #0x2\n" - "tbz x15, #0, 26f\n" - "st1 { v8.b }[10], [x13]\n" - "b 26f\n" - "21:" // Height 1: Partial direct writeback: partial_1_8 - "tbz x15, #0, 26f\n" - "st1 { v8.b }[8], [x13]\n" - "b 26f\n" - "22:" // Height 1: Partial direct writeback: partial_4_0 - "tbz x15, #2, 24f\n" - "str s8, [x13], #0x4\n" - "tbz x15, #1, 23f\n" - "st1 { v8.h }[2], [x13], #0x2\n" - "tbz x15, #0, 26f\n" - "st1 { v8.b }[6], [x13]\n" - "b 26f\n" - "23:" // Height 1: Partial direct writeback: partial_1_4 - "tbz x15, #0, 26f\n" - "st1 { v8.b }[4], [x13]\n" - "b 26f\n" - "24:" // Height 1: Partial direct writeback: partial_2_0 - "tbz x15, #1, 25f\n" - "str h8, [x13], #0x2\n" - "tbz x15, #0, 26f\n" - "st1 { v8.b }[2], [x13]\n" - "b 26f\n" - "25:" // Height 1: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" - "26:" // Height 1: Partial direct writeback: Done - "b 28f\n" - "27:" // Height 1: Full writeback - "str q8, [x13, #0x0]\n" - "add x13, x13, #0x10\n" - "28:" // Height 1: Writeback done - "subs x15, x15, #0x10\n" - "bgt 3b\n" - "b 170f\n" - "29:" // Height 2 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 30f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "add x9, x9, x19\n" - "b 31f\n" - "30:" // Height 2: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "31:" // Height 2: Column loop + "bge 26f\n" + "tbz x10, #3, 21f\n" + "str d8, [x9], #0x8\n" + "tbz x10, #2, 19f\n" + "st1 { v8.s }[2], [x9], #0x4\n" + "tbz x10, #1, 18f\n" + "st1 { v8.h }[6], [x9], #0x2\n" + "tbz x10, #0, 25f\n" + "st1 { v8.b }[14], [x9]\n" + "b 25f\n" + "18:" // Height 1: Partial direct writeback: partial_1_12 + "tbz x10, #0, 25f\n" + "st1 { v8.b }[12], [x9]\n" + "b 25f\n" + "19:" // Height 1: Partial direct writeback: partial_2_8 + "tbz x10, #1, 20f\n" + "st1 { v8.h }[4], [x9], #0x2\n" + "tbz x10, #0, 25f\n" + "st1 { v8.b }[10], [x9]\n" + "b 25f\n" + "20:" // Height 1: Partial direct writeback: partial_1_8 + "tbz x10, #0, 25f\n" + "st1 { v8.b }[8], [x9]\n" + "b 25f\n" + "21:" // Height 1: Partial direct writeback: partial_4_0 + "tbz x10, #2, 23f\n" + "str s8, [x9], #0x4\n" + "tbz x10, #1, 22f\n" + "st1 { v8.h }[2], [x9], #0x2\n" + "tbz x10, #0, 25f\n" + "st1 { v8.b }[6], [x9]\n" + "b 25f\n" + "22:" // Height 1: Partial direct writeback: partial_1_4 + "tbz x10, #0, 25f\n" + "st1 { v8.b }[4], [x9]\n" + "b 25f\n" + "23:" // Height 1: Partial direct writeback: partial_2_0 + "tbz x10, #1, 24f\n" + "str h8, [x9], #0x2\n" + "tbz x10, #0, 25f\n" + "st1 { v8.b }[2], [x9]\n" + "b 25f\n" + "24:" // Height 1: Partial direct writeback: partial_1_0 + "str b8, [x9, #0x0]\n" + "25:" // Height 1: Partial direct writeback: Done + "b 27f\n" + "26:" // Height 1: Full writeback + "str q8, [x9, #0x0]\n" + "add x9, x9, #0x10\n" + "27:" // Height 1: Writeback done + "subs x10, x10, #0x10\n" + "bgt 2b\n" + "b 164f\n" + "28:" // Height 2 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "29:" // Height 2: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" "movi v10.4s, #0x0\n" @@ -409,239 +394,241 @@ void a64_hybrid_s8qs_dot_6x16 ( "movi v13.4s, #0x0\n" "movi v14.4s, #0x0\n" "movi v15.4s, #0x0\n" - "32:" // Height 2: setup done - "mov x12, #0x0\n" - "33:" // Height 2: String loop + "30:" // Height 2: setup done + "mov x27, #0x0\n" + "31:" // 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, 34f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 32f\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, 35f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "cbnz x27, 33f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "b 35f\n" - "34:" // Height 2: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "35:" // Height 2: input setup done - "cmp x11, #0x10\n" - "blt 38f\n" - "cmp x11, #0x20\n" - "blt 37f\n" - "36:" // Height 2: Multiply loop: Main loop head - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q6, [x14, #0x0]\n" + "add x25, x25, x19\n" + "add x24, x24, x19\n" + "b 33f\n" + "32:" // Height 2: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "33:" // Height 2: input setup done + "cmp x26, #0x10\n" + "blt 36f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x20\n" + "ldr q6, [x28, #0x0]\n" + "blt 35f\n" + "34:" // Height 2: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x14, #0x20]\n" - "add x10, x10, #0x10\n" + "ldr q6, [x28, #0x20]\n" + "add x24, x24, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "cmp x26, #0x20\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "sub x11, x11, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x14, #0x40]\n" - "cmp x11, #0x20\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" - "add x14, x14, #0x100\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "bge 36b\n" - "37:" // Height 2: Multiply loop: Single iteration only - "sub x11, x11, #0x10\n" - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q6, [x14, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "bge 34b\n" + "35:" // Height 2: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x14, #0x20]\n" - "add x10, x10, #0x10\n" + "ldr q6, [x28, #0x20]\n" + "add x25, x25, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "prfm pldl1keep, [x10, #0x80]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x14, #0x40]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" - "add x14, x14, #0x100\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "38:" // Height 2: Multiply loop: Main loop skip - "cbz x11, 43f\n" - "cmp x11, #0x4\n" - "blt 40f\n" - "39:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x10], #0x4\n" - "ldr s1, [x28], #0x4\n" - "ldr q6, [x14, #0x0]\n" + "36:" // Height 2: Multiply loop: Main loop skip + "cbz x26, 41f\n" + "cmp x26, #0x4\n" + "blt 38f\n" + "37:" // Height 2: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x4\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x4\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x14, #0x20]\n" - "sub x11, x11, #0x4\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "cmp x11, #0x4\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "bge 39b\n" - "cbz x11, 43f\n" - "40:" // Height 2: Multiply loop: Skip odd blocks - "tbz x11, #1, 41f\n" - "ldr h0, [x10], #0x2\n" - "ldr h1, [x28], #0x2\n" - "tbz x11, #0, 42f\n" - "ld1 { v0.b }[2], [x10]\n" - "ld1 { v1.b }[2], [x28]\n" - "b 42f\n" - "41:" // Height 2: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x10, #0x0]\n" - "ldr b1, [x28, #0x0]\n" - "42:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q6, [x14, #0x0]\n" + "bge 37b\n" + "cbz x26, 41f\n" + "38:" // Height 2: Multiply loop: Skip odd blocks + "tbz x26, #1, 39f\n" + "ldr h0, [x25], #0x2\n" + "ldr h1, [x24], #0x2\n" + "tbz x26, #0, 40f\n" + "ld1 { v0.b }[2], [x25]\n" + "ld1 { v1.b }[2], [x24]\n" + "b 40f\n" + "39:" // Height 2: Multiply loop: Ragged operand read: partial_1_0 + "ldr b0, [x25, #0x0]\n" + "ldr b1, [x24, #0x0]\n" + "40:" // Height 2: Multiply loop: Ragged operand read: Done + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "43:" // Height 2: Multiply loop: No odd multiplies + "41:" // 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 33b\n" - "prfm pstl1keep, [x13, #0x0]\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 31b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "prfm pstl1keep, [x9, #0x0]\n" - "ldr q0, [x16, #0x0]\n" + "add x23, x9, x19\n" + "ldr q0, [x11, #0x0]\n" "add v8.4s, v8.4s, v0.4s\n" - "ldr q1, [x16, #0x10]\n" + "prfm pstl1keep, [x23, #0x0]\n" "add v12.4s, v12.4s, v0.4s\n" - "ldr q2, [x16, #0x20]\n" - "ldr q3, [x16, #0x30]\n" + "ldr q1, [x11, #0x10]\n" + "ldr q2, [x11, #0x20]\n" "add v9.4s, v9.4s, v1.4s\n" - "add x16, x16, #0x40\n" - "add v13.4s, v13.4s, v1.4s\n" + "ldr q3, [x11, #0x30]\n" + "add x11, x11, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" - "add v11.4s, v11.4s, v3.4s\n" + "add v13.4s, v13.4s, v1.4s\n" "add v14.4s, v14.4s, v2.4s\n" + "add v11.4s, v11.4s, v3.4s\n" "add v15.4s, v15.4s, v3.4s\n" - "tbz %x[flags], #4, 44f\n" - "ldr q0, [x17, #0x0]\n" - "ldr q4, [x8, #0x0]\n" - "ldr q1, [x17, #0x10]\n" - "ldr q5, [x8, #0x10]\n" - "ldr q2, [x17, #0x20]\n" - "ldr q6, [x8, #0x20]\n" - "ldr q3, [x17, #0x30]\n" - "ldr q7, [x8, #0x30]\n" - "add x17, x17, #0x40\n" - "add x8, x8, #0x40\n" - "b 45f\n" - "44:" // Height 2: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x19]\n" + "tbz %x[flags], #4, 42f\n" + "ldr q0, [x12, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q1, [x12, #0x10]\n" + "ldr q5, [x13, #0x10]\n" + "ldr q2, [x12, #0x20]\n" + "ldr q6, [x13, #0x20]\n" + "ldr q3, [x12, #0x30]\n" + "add x12, x12, #0x40\n" + "ldr q7, [x13, #0x30]\n" + "add x13, x13, #0x40\n" + "b 43f\n" + "42:" // Height 2: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1r { v0.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1r { v4.4s }, [x24]\n" "mov v2.16b, v0.16b\n" "mov v3.16b, v0.16b\n" "mov v5.16b, v4.16b\n" "mov v6.16b, v4.16b\n" "mov v7.16b, v4.16b\n" - "45:" // Height 2: parameters loaded + "43:" // Height 2: parameters loaded "sqrdmulh v8.4s, v8.4s, v4.4s\n" "sqrdmulh v9.4s, v9.4s, v5.4s\n" "sqrdmulh v10.4s, v10.4s, v6.4s\n" @@ -650,7 +637,7 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqrdmulh v13.4s, v13.4s, v5.4s\n" "sqrdmulh v14.4s, v14.4s, v6.4s\n" "sqrdmulh v15.4s, v15.4s, v7.4s\n" - "tbz %x[flags], #5, 46f\n" + "tbz %x[flags], #5, 44f\n" "and v4.16b, v8.16b, v0.16b\n" "sshr v4.4s, v4.4s, #0x1f\n" "and v5.16b, v9.16b, v1.16b\n" @@ -675,18 +662,18 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqadd v13.4s, v13.4s, v5.4s\n" "sqadd v14.4s, v14.4s, v6.4s\n" "sqadd v15.4s, v15.4s, v7.4s\n" - "46:" // Height 2: no shift correction + "44:" // Height 2: no shift correction "srshl v8.4s, v8.4s, v0.4s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1r { v4.4s }, [x24]\n" "srshl v9.4s, v9.4s, v1.4s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" "srshl v10.4s, v10.4s, v2.4s\n" - "ld1r { v5.4s }, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1r { v5.4s }, [x24]\n" + "add x24, %x[qp], %[maxval]\n" "srshl v11.4s, v11.4s, v3.4s\n" - "ld1r { v6.4s }, [x19]\n" - "cmp x15, #0x10\n" + "ld1r { v6.4s }, [x24]\n" + "cmp x10, #0x10\n" "srshl v12.4s, v12.4s, v0.4s\n" "srshl v13.4s, v13.4s, v1.4s\n" "srshl v14.4s, v14.4s, v2.4s\n" @@ -721,96 +708,83 @@ void a64_hybrid_s8qs_dot_6x16 ( "uzp1 v8.16b, v8.16b, v9.16b\n" "uzp1 v13.8h, v14.8h, v15.8h\n" "uzp1 v12.16b, v12.16b, v13.16b\n" - "bge 55f\n" - "tbz x15, #3, 50f\n" - "str d8, [x13], #0x8\n" - "str d12, [x9], #0x8\n" - "tbz x15, #2, 48f\n" - "st1 { v8.s }[2], [x13], #0x4\n" - "st1 { v12.s }[2], [x9], #0x4\n" - "tbz x15, #1, 47f\n" - "st1 { v8.h }[6], [x13], #0x2\n" - "st1 { v12.h }[6], [x9], #0x2\n" - "tbz x15, #0, 54f\n" - "st1 { v8.b }[14], [x13]\n" - "st1 { v12.b }[14], [x9]\n" - "b 54f\n" - "47:" // Height 2: Partial direct writeback: partial_1_12 - "tbz x15, #0, 54f\n" - "st1 { v8.b }[12], [x13]\n" - "st1 { v12.b }[12], [x9]\n" - "b 54f\n" - "48:" // Height 2: Partial direct writeback: partial_2_8 - "tbz x15, #1, 49f\n" - "st1 { v8.h }[4], [x13], #0x2\n" - "st1 { v12.h }[4], [x9], #0x2\n" - "tbz x15, #0, 54f\n" - "st1 { v8.b }[10], [x13]\n" - "st1 { v12.b }[10], [x9]\n" - "b 54f\n" - "49:" // Height 2: Partial direct writeback: partial_1_8 - "tbz x15, #0, 54f\n" - "st1 { v8.b }[8], [x13]\n" - "st1 { v12.b }[8], [x9]\n" - "b 54f\n" - "50:" // Height 2: Partial direct writeback: partial_4_0 - "tbz x15, #2, 52f\n" - "str s8, [x13], #0x4\n" - "str s12, [x9], #0x4\n" - "tbz x15, #1, 51f\n" - "st1 { v8.h }[2], [x13], #0x2\n" - "st1 { v12.h }[2], [x9], #0x2\n" - "tbz x15, #0, 54f\n" - "st1 { v8.b }[6], [x13]\n" - "st1 { v12.b }[6], [x9]\n" - "b 54f\n" - "51:" // Height 2: Partial direct writeback: partial_1_4 - "tbz x15, #0, 54f\n" - "st1 { v8.b }[4], [x13]\n" - "st1 { v12.b }[4], [x9]\n" + "bge 53f\n" + "tbz x10, #3, 48f\n" + "str d8, [x9], #0x8\n" + "str d12, [x23], #0x8\n" + "tbz x10, #2, 46f\n" + "st1 { v8.s }[2], [x9], #0x4\n" + "st1 { v12.s }[2], [x23], #0x4\n" + "tbz x10, #1, 45f\n" + "st1 { v8.h }[6], [x9], #0x2\n" + "st1 { v12.h }[6], [x23], #0x2\n" + "tbz x10, #0, 52f\n" + "st1 { v8.b }[14], [x9]\n" + "st1 { v12.b }[14], [x23]\n" + "b 52f\n" + "45:" // Height 2: Partial direct writeback: partial_1_12 + "tbz x10, #0, 52f\n" + "st1 { v8.b }[12], [x9]\n" + "st1 { v12.b }[12], [x23]\n" + "b 52f\n" + "46:" // Height 2: Partial direct writeback: partial_2_8 + "tbz x10, #1, 47f\n" + "st1 { v8.h }[4], [x9], #0x2\n" + "st1 { v12.h }[4], [x23], #0x2\n" + "tbz x10, #0, 52f\n" + "st1 { v8.b }[10], [x9]\n" + "st1 { v12.b }[10], [x23]\n" + "b 52f\n" + "47:" // Height 2: Partial direct writeback: partial_1_8 + "tbz x10, #0, 52f\n" + "st1 { v8.b }[8], [x9]\n" + "st1 { v12.b }[8], [x23]\n" + "b 52f\n" + "48:" // Height 2: Partial direct writeback: partial_4_0 + "tbz x10, #2, 50f\n" + "str s8, [x9], #0x4\n" + "str s12, [x23], #0x4\n" + "tbz x10, #1, 49f\n" + "st1 { v8.h }[2], [x9], #0x2\n" + "st1 { v12.h }[2], [x23], #0x2\n" + "tbz x10, #0, 52f\n" + "st1 { v8.b }[6], [x9]\n" + "st1 { v12.b }[6], [x23]\n" + "b 52f\n" + "49:" // Height 2: Partial direct writeback: partial_1_4 + "tbz x10, #0, 52f\n" + "st1 { v8.b }[4], [x9]\n" + "st1 { v12.b }[4], [x23]\n" + "b 52f\n" + "50:" // Height 2: Partial direct writeback: partial_2_0 + "tbz x10, #1, 51f\n" + "str h8, [x9], #0x2\n" + "str h12, [x23], #0x2\n" + "tbz x10, #0, 52f\n" + "st1 { v8.b }[2], [x9]\n" + "st1 { v12.b }[2], [x23]\n" + "b 52f\n" + "51:" // Height 2: Partial direct writeback: partial_1_0 + "str b8, [x9, #0x0]\n" + "str b12, [x23, #0x0]\n" + "52:" // Height 2: Partial direct writeback: Done "b 54f\n" - "52:" // Height 2: Partial direct writeback: partial_2_0 - "tbz x15, #1, 53f\n" - "str h8, [x13], #0x2\n" - "str h12, [x9], #0x2\n" - "tbz x15, #0, 54f\n" - "st1 { v8.b }[2], [x13]\n" - "st1 { v12.b }[2], [x9]\n" - "b 54f\n" - "53:" // Height 2: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" - "str b12, [x9, #0x0]\n" - "54:" // Height 2: Partial direct writeback: Done - "b 56f\n" - "55:" // Height 2: Full writeback - "str q8, [x13, #0x0]\n" - "str q12, [x9, #0x0]\n" - "add x13, x13, #0x10\n" + "53:" // Height 2: Full writeback + "str q8, [x9, #0x0]\n" "add x9, x9, #0x10\n" - "56:" // Height 2: Writeback done - "subs x15, x15, #0x10\n" - "bgt 31b\n" - "b 170f\n" - "57:" // Height 3 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 58f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19\n" - "add x27, x27, x19\n" - "b 59f\n" - "58:" // Height 3: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "add x27, x9, x19\n" - "59:" // Height 3: Column loop + "str q12, [x23, #0x0]\n" + "54:" // Height 2: Writeback done + "subs x10, x10, #0x10\n" + "bgt 29b\n" + "b 164f\n" + "55:" // Height 3 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "56:" // Height 3: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" "movi v10.4s, #0x0\n" @@ -823,297 +797,300 @@ void a64_hybrid_s8qs_dot_6x16 ( "movi v17.4s, #0x0\n" "movi v18.4s, #0x0\n" "movi v19.4s, #0x0\n" - "60:" // Height 3: setup done - "mov x12, #0x0\n" - "61:" // Height 3: String loop + "57:" // Height 3: setup done + "mov x27, #0x0\n" + "58:" // 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, 62f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 59f\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, 63f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "cbnz x27, 60f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "add x26, x26, x19\n" - "b 63f\n" - "62:" // Height 3: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "add x26, x28, x19\n" - "63:" // Height 3: input setup done - "cmp x11, #0x10\n" - "blt 66f\n" - "cmp x11, #0x20\n" - "blt 65f\n" - "64:" // Height 3: Multiply loop: Main loop head - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q6, [x14, #0x0]\n" + "add x25, x25, x19\n" + "add x24, x24, x19\n" + "add x23, x23, x19\n" + "b 60f\n" + "59:" // Height 3: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "add x23, x24, x19\n" + "60:" // Height 3: input setup done + "cmp x26, #0x10\n" + "blt 63f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x20\n" + "ldr q2, [x23, #0x0]\n" + "ldr q6, [x28, #0x0]\n" + "blt 62f\n" + "61:" // Height 3: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" + "add x23, x23, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x28, x28, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" + "cmp x26, #0x20\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "sub x11, x11, #0x10\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "cmp x11, #0x20\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x14, #0x40]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" - "add x14, x14, #0x100\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" + "ldr q1, [x24, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - "bge 64b\n" - "65:" // Height 3: Multiply loop: Single iteration only - "sub x11, x11, #0x10\n" - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q6, [x14, #0x0]\n" + "ldr q2, [x23, #0x0]\n" + "bge 61b\n" + "62:" // Height 3: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "add x24, x24, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x28, x28, #0x10\n" - "prfm pldl1keep, [x28, #0x80]\n" + "ldr q6, [x28, #0x20]\n" + "add x23, x23, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q7, [x14, #0x30]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x14, #0x40]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" - "add x14, x14, #0x100\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - "66:" // Height 3: Multiply loop: Main loop skip - "cbz x11, 71f\n" - "cmp x11, #0x4\n" - "blt 68f\n" - "67:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x10], #0x4\n" - "ldr s1, [x28], #0x4\n" - "ldr s2, [x26], #0x4\n" - "ldr q6, [x14, #0x0]\n" + "63:" // Height 3: Multiply loop: Main loop skip + "cbz x26, 68f\n" + "cmp x26, #0x4\n" + "blt 65f\n" + "64:" // Height 3: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x4\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x4\n" + "ldr s2, [x23], #0x4\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "sub x11, x11, #0x4\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x14, #0x20]\n" - "cmp x11, #0x4\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "bge 67b\n" - "cbz x11, 71f\n" - "68:" // Height 3: Multiply loop: Skip odd blocks - "tbz x11, #1, 69f\n" - "ldr h0, [x10], #0x2\n" - "ldr h1, [x28], #0x2\n" - "ldr h2, [x26], #0x2\n" - "tbz x11, #0, 70f\n" - "ld1 { v0.b }[2], [x10]\n" - "ld1 { v1.b }[2], [x28]\n" - "ld1 { v2.b }[2], [x26]\n" - "b 70f\n" - "69:" // Height 3: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x10, #0x0]\n" - "ldr b1, [x28, #0x0]\n" - "ldr b2, [x26, #0x0]\n" - "70:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q6, [x14, #0x0]\n" + "bge 64b\n" + "cbz x26, 68f\n" + "65:" // Height 3: Multiply loop: Skip odd blocks + "tbz x26, #1, 66f\n" + "ldr h0, [x25], #0x2\n" + "ldr h1, [x24], #0x2\n" + "ldr h2, [x23], #0x2\n" + "tbz x26, #0, 67f\n" + "ld1 { v0.b }[2], [x25]\n" + "ld1 { v1.b }[2], [x24]\n" + "ld1 { v2.b }[2], [x23]\n" + "b 67f\n" + "66:" // Height 3: Multiply loop: Ragged operand read: partial_1_0 + "ldr b0, [x25, #0x0]\n" + "ldr b1, [x24, #0x0]\n" + "ldr b2, [x23, #0x0]\n" + "67:" // Height 3: Multiply loop: Ragged operand read: Done + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "71:" // Height 3: Multiply loop: No odd multiplies + "68:" // 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 61b\n" - "prfm pstl1keep, [x13, #0x0]\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 58b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "prfm pstl1keep, [x9, #0x0]\n" - "ldr q0, [x16, #0x0]\n" + "add x23, x9, x19\n" + "ldr q0, [x11, #0x0]\n" "add v8.4s, v8.4s, v0.4s\n" - "prfm pstl1keep, [x27, #0x0]\n" - "ldr q1, [x16, #0x10]\n" + "prfm pstl1keep, [x23, #0x0]\n" + "add x22, x23, x19\n" "add v12.4s, v12.4s, v0.4s\n" - "ldr q2, [x16, #0x20]\n" + "prfm pstl1keep, [x22, #0x0]\n" "add v16.4s, v16.4s, v0.4s\n" - "ldr q3, [x16, #0x30]\n" - "add x16, x16, #0x40\n" + "ldr q1, [x11, #0x10]\n" + "ldr q2, [x11, #0x20]\n" "add v9.4s, v9.4s, v1.4s\n" - "add v13.4s, v13.4s, v1.4s\n" + "ldr q3, [x11, #0x30]\n" + "add x11, x11, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" - "add v11.4s, v11.4s, v3.4s\n" + "add v13.4s, v13.4s, v1.4s\n" "add v14.4s, v14.4s, v2.4s\n" - "add v15.4s, v15.4s, v3.4s\n" "add v17.4s, v17.4s, v1.4s\n" + "add v11.4s, v11.4s, v3.4s\n" + "add v15.4s, v15.4s, v3.4s\n" "add v18.4s, v18.4s, v2.4s\n" "add v19.4s, v19.4s, v3.4s\n" - "tbz %x[flags], #4, 72f\n" - "ldr q0, [x17, #0x0]\n" - "ldr q4, [x8, #0x0]\n" - "ldr q1, [x17, #0x10]\n" - "ldr q5, [x8, #0x10]\n" - "ldr q2, [x17, #0x20]\n" - "ldr q6, [x8, #0x20]\n" - "ldr q3, [x17, #0x30]\n" - "ldr q7, [x8, #0x30]\n" - "add x17, x17, #0x40\n" - "add x8, x8, #0x40\n" - "b 73f\n" - "72:" // Height 3: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x19]\n" + "tbz %x[flags], #4, 69f\n" + "ldr q0, [x12, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q1, [x12, #0x10]\n" + "ldr q5, [x13, #0x10]\n" + "ldr q2, [x12, #0x20]\n" + "ldr q6, [x13, #0x20]\n" + "ldr q3, [x12, #0x30]\n" + "add x12, x12, #0x40\n" + "ldr q7, [x13, #0x30]\n" + "add x13, x13, #0x40\n" + "b 70f\n" + "69:" // Height 3: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1r { v0.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1r { v4.4s }, [x24]\n" "mov v2.16b, v0.16b\n" "mov v3.16b, v0.16b\n" "mov v5.16b, v4.16b\n" "mov v6.16b, v4.16b\n" "mov v7.16b, v4.16b\n" - "73:" // Height 3: parameters loaded + "70:" // Height 3: parameters loaded "sqrdmulh v8.4s, v8.4s, v4.4s\n" "sqrdmulh v9.4s, v9.4s, v5.4s\n" "sqrdmulh v10.4s, v10.4s, v6.4s\n" @@ -1126,7 +1103,7 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqrdmulh v17.4s, v17.4s, v5.4s\n" "sqrdmulh v18.4s, v18.4s, v6.4s\n" "sqrdmulh v19.4s, v19.4s, v7.4s\n" - "tbz %x[flags], #5, 74f\n" + "tbz %x[flags], #5, 71f\n" "and v4.16b, v8.16b, v0.16b\n" "sshr v4.4s, v4.4s, #0x1f\n" "and v5.16b, v9.16b, v1.16b\n" @@ -1163,18 +1140,18 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqadd v17.4s, v17.4s, v5.4s\n" "sqadd v18.4s, v18.4s, v6.4s\n" "sqadd v19.4s, v19.4s, v7.4s\n" - "74:" // Height 3: no shift correction + "71:" // Height 3: no shift correction "srshl v8.4s, v8.4s, v0.4s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1r { v4.4s }, [x24]\n" "srshl v9.4s, v9.4s, v1.4s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" "srshl v10.4s, v10.4s, v2.4s\n" - "ld1r { v5.4s }, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1r { v5.4s }, [x24]\n" + "add x24, %x[qp], %[maxval]\n" "srshl v11.4s, v11.4s, v3.4s\n" - "ld1r { v6.4s }, [x19]\n" - "cmp x15, #0x10\n" + "ld1r { v6.4s }, [x24]\n" + "cmp x10, #0x10\n" "srshl v12.4s, v12.4s, v0.4s\n" "srshl v13.4s, v13.4s, v1.4s\n" "srshl v14.4s, v14.4s, v2.4s\n" @@ -1228,116 +1205,99 @@ void a64_hybrid_s8qs_dot_6x16 ( "uzp1 v8.16b, v8.16b, v9.16b\n" "uzp1 v12.16b, v12.16b, v13.16b\n" "uzp1 v16.16b, v16.16b, v17.16b\n" - "bge 83f\n" - "tbz x15, #3, 78f\n" - "str d8, [x13], #0x8\n" - "str d12, [x9], #0x8\n" - "str d16, [x27], #0x8\n" - "tbz x15, #2, 76f\n" - "st1 { v8.s }[2], [x13], #0x4\n" - "st1 { v12.s }[2], [x9], #0x4\n" - "st1 { v16.s }[2], [x27], #0x4\n" - "tbz x15, #1, 75f\n" - "st1 { v8.h }[6], [x13], #0x2\n" - "st1 { v12.h }[6], [x9], #0x2\n" - "st1 { v16.h }[6], [x27], #0x2\n" - "tbz x15, #0, 82f\n" - "st1 { v8.b }[14], [x13]\n" - "st1 { v12.b }[14], [x9]\n" - "st1 { v16.b }[14], [x27]\n" - "b 82f\n" - "75:" // Height 3: Partial direct writeback: partial_1_12 - "tbz x15, #0, 82f\n" - "st1 { v8.b }[12], [x13]\n" - "st1 { v12.b }[12], [x9]\n" - "st1 { v16.b }[12], [x27]\n" - "b 82f\n" - "76:" // Height 3: Partial direct writeback: partial_2_8 - "tbz x15, #1, 77f\n" - "st1 { v8.h }[4], [x13], #0x2\n" - "st1 { v12.h }[4], [x9], #0x2\n" - "st1 { v16.h }[4], [x27], #0x2\n" - "tbz x15, #0, 82f\n" - "st1 { v8.b }[10], [x13]\n" - "st1 { v12.b }[10], [x9]\n" - "st1 { v16.b }[10], [x27]\n" - "b 82f\n" - "77:" // Height 3: Partial direct writeback: partial_1_8 - "tbz x15, #0, 82f\n" - "st1 { v8.b }[8], [x13]\n" - "st1 { v12.b }[8], [x9]\n" - "st1 { v16.b }[8], [x27]\n" - "b 82f\n" - "78:" // Height 3: Partial direct writeback: partial_4_0 - "tbz x15, #2, 80f\n" - "str s8, [x13], #0x4\n" - "str s12, [x9], #0x4\n" - "str s16, [x27], #0x4\n" - "tbz x15, #1, 79f\n" - "st1 { v8.h }[2], [x13], #0x2\n" - "st1 { v12.h }[2], [x9], #0x2\n" - "st1 { v16.h }[2], [x27], #0x2\n" - "tbz x15, #0, 82f\n" - "st1 { v8.b }[6], [x13]\n" - "st1 { v12.b }[6], [x9]\n" - "st1 { v16.b }[6], [x27]\n" - "b 82f\n" - "79:" // Height 3: Partial direct writeback: partial_1_4 - "tbz x15, #0, 82f\n" - "st1 { v8.b }[4], [x13]\n" - "st1 { v12.b }[4], [x9]\n" - "st1 { v16.b }[4], [x27]\n" - "b 82f\n" - "80:" // Height 3: Partial direct writeback: partial_2_0 - "tbz x15, #1, 81f\n" - "str h8, [x13], #0x2\n" - "str h12, [x9], #0x2\n" - "str h16, [x27], #0x2\n" - "tbz x15, #0, 82f\n" - "st1 { v8.b }[2], [x13]\n" - "st1 { v12.b }[2], [x9]\n" - "st1 { v16.b }[2], [x27]\n" - "b 82f\n" - "81:" // Height 3: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" - "str b12, [x9, #0x0]\n" - "str b16, [x27, #0x0]\n" - "82:" // Height 3: Partial direct writeback: Done - "b 84f\n" - "83:" // Height 3: Full writeback - "str q8, [x13, #0x0]\n" - "str q12, [x9, #0x0]\n" - "str q16, [x27, #0x0]\n" - "add x13, x13, #0x10\n" + "bge 80f\n" + "tbz x10, #3, 75f\n" + "str d8, [x9], #0x8\n" + "str d12, [x23], #0x8\n" + "str d16, [x22], #0x8\n" + "tbz x10, #2, 73f\n" + "st1 { v8.s }[2], [x9], #0x4\n" + "st1 { v12.s }[2], [x23], #0x4\n" + "st1 { v16.s }[2], [x22], #0x4\n" + "tbz x10, #1, 72f\n" + "st1 { v8.h }[6], [x9], #0x2\n" + "st1 { v12.h }[6], [x23], #0x2\n" + "st1 { v16.h }[6], [x22], #0x2\n" + "tbz x10, #0, 79f\n" + "st1 { v8.b }[14], [x9]\n" + "st1 { v12.b }[14], [x23]\n" + "st1 { v16.b }[14], [x22]\n" + "b 79f\n" + "72:" // Height 3: Partial direct writeback: partial_1_12 + "tbz x10, #0, 79f\n" + "st1 { v8.b }[12], [x9]\n" + "st1 { v12.b }[12], [x23]\n" + "st1 { v16.b }[12], [x22]\n" + "b 79f\n" + "73:" // Height 3: Partial direct writeback: partial_2_8 + "tbz x10, #1, 74f\n" + "st1 { v8.h }[4], [x9], #0x2\n" + "st1 { v12.h }[4], [x23], #0x2\n" + "st1 { v16.h }[4], [x22], #0x2\n" + "tbz x10, #0, 79f\n" + "st1 { v8.b }[10], [x9]\n" + "st1 { v12.b }[10], [x23]\n" + "st1 { v16.b }[10], [x22]\n" + "b 79f\n" + "74:" // Height 3: Partial direct writeback: partial_1_8 + "tbz x10, #0, 79f\n" + "st1 { v8.b }[8], [x9]\n" + "st1 { v12.b }[8], [x23]\n" + "st1 { v16.b }[8], [x22]\n" + "b 79f\n" + "75:" // Height 3: Partial direct writeback: partial_4_0 + "tbz x10, #2, 77f\n" + "str s8, [x9], #0x4\n" + "str s12, [x23], #0x4\n" + "str s16, [x22], #0x4\n" + "tbz x10, #1, 76f\n" + "st1 { v8.h }[2], [x9], #0x2\n" + "st1 { v12.h }[2], [x23], #0x2\n" + "st1 { v16.h }[2], [x22], #0x2\n" + "tbz x10, #0, 79f\n" + "st1 { v8.b }[6], [x9]\n" + "st1 { v12.b }[6], [x23]\n" + "st1 { v16.b }[6], [x22]\n" + "b 79f\n" + "76:" // Height 3: Partial direct writeback: partial_1_4 + "tbz x10, #0, 79f\n" + "st1 { v8.b }[4], [x9]\n" + "st1 { v12.b }[4], [x23]\n" + "st1 { v16.b }[4], [x22]\n" + "b 79f\n" + "77:" // Height 3: Partial direct writeback: partial_2_0 + "tbz x10, #1, 78f\n" + "str h8, [x9], #0x2\n" + "str h12, [x23], #0x2\n" + "str h16, [x22], #0x2\n" + "tbz x10, #0, 79f\n" + "st1 { v8.b }[2], [x9]\n" + "st1 { v12.b }[2], [x23]\n" + "st1 { v16.b }[2], [x22]\n" + "b 79f\n" + "78:" // Height 3: Partial direct writeback: partial_1_0 + "str b8, [x9, #0x0]\n" + "str b12, [x23, #0x0]\n" + "str b16, [x22, #0x0]\n" + "79:" // Height 3: Partial direct writeback: Done + "b 81f\n" + "80:" // Height 3: Full writeback + "str q8, [x9, #0x0]\n" "add x9, x9, #0x10\n" - "add x27, x27, #0x10\n" - "84:" // Height 3: Writeback done - "subs x15, x15, #0x10\n" - "bgt 59b\n" - "b 170f\n" - "85:" // Height 4 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 86f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19\n" - "ldr x25, [%x[output_ptr], #0x18]\n" - "add x27, x27, x19\n" - "add x25, x25, x19\n" - "b 87f\n" - "86:" // Height 4: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "add x27, x9, x19\n" - "add x25, x27, x19\n" - "87:" // Height 4: Column loop + "str q12, [x23, #0x0]\n" + "str q16, [x22, #0x0]\n" + "81:" // Height 3: Writeback done + "subs x10, x10, #0x10\n" + "bgt 56b\n" + "b 164f\n" + "82:" // Height 4 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "83:" // Height 4: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" "movi v10.4s, #0x0\n" @@ -1354,220 +1314,220 @@ void a64_hybrid_s8qs_dot_6x16 ( "movi v21.4s, #0x0\n" "movi v22.4s, #0x0\n" "movi v23.4s, #0x0\n" - "88:" // Height 4: setup done - "mov x12, #0x0\n" - "89:" // Height 4: String loop + "84:" // Height 4: setup done + "mov x27, #0x0\n" + "85:" // 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, 90f\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" - "ldr x24, [x20, #0x18]\n" - "cbnz x12, 91f\n" + "ldr x25, [x20, #0x0]\n" + "ldr x24, [x20, #0x8]\n" + "ldr x23, [x20, #0x10]\n" + "ldr x22, [x20, #0x18]\n" + "cbnz x27, 87f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "add x26, x26, x19\n" + "add x25, x25, x19\n" "add x24, x24, x19\n" - "b 91f\n" - "90:" // Height 4: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "add x26, x28, x19\n" - "add x24, x26, x19\n" - "91:" // Height 4: input setup done - "cmp x11, #0x10\n" - "blt 94f\n" - "cmp x11, #0x20\n" - "blt 93f\n" - "92:" // 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, [x14, #0x0]\n" + "add x23, x23, x19\n" + "add x22, x22, x19\n" + "b 87f\n" + "86:" // Height 4: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "add x23, x24, x19\n" + "add x22, x23, x19\n" + "87:" // Height 4: input setup done + "cmp x26, #0x10\n" + "blt 90f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x20\n" + "ldr q2, [x23, #0x0]\n" + "ldr q3, [x22, #0x0]\n" + "ldr q6, [x28, #0x0]\n" + "blt 89f\n" + "88:" // Height 4: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" + "add x22, x22, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x22, #0x80]\n" + "cmp x26, #0x20\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "sub x11, x11, #0x10\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "cmp x11, #0x20\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x14, #0x40]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" - "add x14, x14, #0x100\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" + "ldr q1, [x24, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + "ldr q2, [x23, #0x0]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" - "bge 92b\n" - "93:" // Height 4: Multiply loop: Single iteration only - "sub x11, x11, #0x10\n" - "ldr q0, [x10, #0x0]\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x26, #0x0]\n" - "ldr q3, [x24, #0x0]\n" - "ldr q6, [x14, #0x0]\n" + "ldr q3, [x22, #0x0]\n" + "bge 88b\n" + "89:" // Height 4: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x28, x28, #0x10\n" + "add x24, x24, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "ldr q6, [x14, #0x20]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x26, x26, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" + "ldr q6, [x28, #0x20]\n" + "add x22, x22, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x14, #0x40]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" - "add x14, x14, #0x100\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" @@ -1576,31 +1536,31 @@ void a64_hybrid_s8qs_dot_6x16 ( ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" - "94:" // Height 4: Multiply loop: Main loop skip - "cbz x11, 99f\n" - "cmp x11, #0x4\n" - "blt 96f\n" - "95:" // 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, [x14, #0x0]\n" + "90:" // Height 4: Multiply loop: Main loop skip + "cbz x26, 95f\n" + "cmp x26, #0x4\n" + "blt 92f\n" + "91:" // Height 4: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x4\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x4\n" + "ldr s2, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "sub x11, x11, #0x4\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "cmp x11, #0x4\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -1608,40 +1568,40 @@ void a64_hybrid_s8qs_dot_6x16 ( ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "bge 95b\n" - "cbz x11, 99f\n" - "96:" // Height 4: Multiply loop: Skip odd blocks - "tbz x11, #1, 97f\n" - "ldr h0, [x10], #0x2\n" - "ldr h1, [x28], #0x2\n" - "ldr h2, [x26], #0x2\n" - "ldr h3, [x24], #0x2\n" - "tbz x11, #0, 98f\n" - "ld1 { v0.b }[2], [x10]\n" - "ld1 { v1.b }[2], [x28]\n" - "ld1 { v2.b }[2], [x26]\n" - "ld1 { v3.b }[2], [x24]\n" - "b 98f\n" - "97:" // Height 4: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x10, #0x0]\n" - "ldr b1, [x28, #0x0]\n" - "ldr b2, [x26, #0x0]\n" - "ldr b3, [x24, #0x0]\n" - "98:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q6, [x14, #0x0]\n" + "bge 91b\n" + "cbz x26, 95f\n" + "92:" // Height 4: Multiply loop: Skip odd blocks + "tbz x26, #1, 93f\n" + "ldr h0, [x25], #0x2\n" + "ldr h1, [x24], #0x2\n" + "ldr h2, [x23], #0x2\n" + "ldr h3, [x22], #0x2\n" + "tbz x26, #0, 94f\n" + "ld1 { v0.b }[2], [x25]\n" + "ld1 { v1.b }[2], [x24]\n" + "ld1 { v2.b }[2], [x23]\n" + "ld1 { v3.b }[2], [x22]\n" + "b 94f\n" + "93:" // Height 4: Multiply loop: Ragged operand read: partial_1_0 + "ldr b0, [x25, #0x0]\n" + "ldr b1, [x24, #0x0]\n" + "ldr b2, [x23, #0x0]\n" + "ldr b3, [x22, #0x0]\n" + "94:" // Height 4: Multiply loop: Ragged operand read: Done + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -1649,60 +1609,64 @@ void a64_hybrid_s8qs_dot_6x16 ( ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "99:" // Height 4: Multiply loop: No odd multiplies + "95:" // 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 89b\n" - "prfm pstl1keep, [x13, #0x0]\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 85b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "prfm pstl1keep, [x9, #0x0]\n" - "ldr q0, [x16, #0x0]\n" + "add x23, x9, x19\n" + "ldr q0, [x11, #0x0]\n" "add v8.4s, v8.4s, v0.4s\n" - "prfm pstl1keep, [x27, #0x0]\n" - "ldr q1, [x16, #0x10]\n" + "prfm pstl1keep, [x23, #0x0]\n" + "add x22, x23, x19\n" "add v12.4s, v12.4s, v0.4s\n" - "prfm pstl1keep, [x25, #0x0]\n" - "ldr q2, [x16, #0x20]\n" + "prfm pstl1keep, [x22, #0x0]\n" + "add x21, x22, x19\n" "add v16.4s, v16.4s, v0.4s\n" - "ldr q3, [x16, #0x30]\n" + "prfm pstl1keep, [x21, #0x0]\n" "add v20.4s, v20.4s, v0.4s\n" - "add x16, x16, #0x40\n" + "ldr q1, [x11, #0x10]\n" + "ldr q2, [x11, #0x20]\n" "add v9.4s, v9.4s, v1.4s\n" - "add v13.4s, v13.4s, v1.4s\n" + "ldr q3, [x11, #0x30]\n" + "add x11, x11, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" - "add v11.4s, v11.4s, v3.4s\n" + "add v13.4s, v13.4s, v1.4s\n" "add v14.4s, v14.4s, v2.4s\n" - "add v15.4s, v15.4s, v3.4s\n" "add v17.4s, v17.4s, v1.4s\n" + "add v11.4s, v11.4s, v3.4s\n" + "add v15.4s, v15.4s, v3.4s\n" "add v18.4s, v18.4s, v2.4s\n" "add v19.4s, v19.4s, v3.4s\n" "add v21.4s, v21.4s, v1.4s\n" "add v22.4s, v22.4s, v2.4s\n" "add v23.4s, v23.4s, v3.4s\n" - "tbz %x[flags], #4, 100f\n" - "ldr q0, [x17, #0x0]\n" - "ldr q4, [x8, #0x0]\n" - "ldr q1, [x17, #0x10]\n" - "ldr q5, [x8, #0x10]\n" - "ldr q2, [x17, #0x20]\n" - "ldr q6, [x8, #0x20]\n" - "ldr q3, [x17, #0x30]\n" - "ldr q7, [x8, #0x30]\n" - "add x17, x17, #0x40\n" - "add x8, x8, #0x40\n" - "b 101f\n" - "100:" // Height 4: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x19]\n" + "tbz %x[flags], #4, 96f\n" + "ldr q0, [x12, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q1, [x12, #0x10]\n" + "ldr q5, [x13, #0x10]\n" + "ldr q2, [x12, #0x20]\n" + "ldr q6, [x13, #0x20]\n" + "ldr q3, [x12, #0x30]\n" + "add x12, x12, #0x40\n" + "ldr q7, [x13, #0x30]\n" + "add x13, x13, #0x40\n" + "b 97f\n" + "96:" // Height 4: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1r { v0.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1r { v4.4s }, [x24]\n" "mov v2.16b, v0.16b\n" "mov v3.16b, v0.16b\n" "mov v5.16b, v4.16b\n" "mov v6.16b, v4.16b\n" "mov v7.16b, v4.16b\n" - "101:" // Height 4: parameters loaded + "97:" // Height 4: parameters loaded "sqrdmulh v8.4s, v8.4s, v4.4s\n" "sqrdmulh v9.4s, v9.4s, v5.4s\n" "sqrdmulh v10.4s, v10.4s, v6.4s\n" @@ -1719,7 +1683,7 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqrdmulh v21.4s, v21.4s, v5.4s\n" "sqrdmulh v22.4s, v22.4s, v6.4s\n" "sqrdmulh v23.4s, v23.4s, v7.4s\n" - "tbz %x[flags], #5, 102f\n" + "tbz %x[flags], #5, 98f\n" "and v4.16b, v8.16b, v0.16b\n" "sshr v4.4s, v4.4s, #0x1f\n" "and v5.16b, v9.16b, v1.16b\n" @@ -1768,18 +1732,18 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqadd v21.4s, v21.4s, v5.4s\n" "sqadd v22.4s, v22.4s, v6.4s\n" "sqadd v23.4s, v23.4s, v7.4s\n" - "102:" // Height 4: no shift correction + "98:" // Height 4: no shift correction "srshl v8.4s, v8.4s, v0.4s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1r { v4.4s }, [x24]\n" "srshl v9.4s, v9.4s, v1.4s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" "srshl v10.4s, v10.4s, v2.4s\n" - "ld1r { v5.4s }, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1r { v5.4s }, [x24]\n" + "add x24, %x[qp], %[maxval]\n" "srshl v11.4s, v11.4s, v3.4s\n" - "ld1r { v6.4s }, [x19]\n" - "cmp x15, #0x10\n" + "ld1r { v6.4s }, [x24]\n" + "cmp x10, #0x10\n" "srshl v12.4s, v12.4s, v0.4s\n" "srshl v13.4s, v13.4s, v1.4s\n" "srshl v14.4s, v14.4s, v2.4s\n" @@ -1852,136 +1816,115 @@ void a64_hybrid_s8qs_dot_6x16 ( "uzp1 v12.16b, v12.16b, v13.16b\n" "uzp1 v16.16b, v16.16b, v17.16b\n" "uzp1 v20.16b, v20.16b, v21.16b\n" - "bge 111f\n" - "tbz x15, #3, 106f\n" - "str d8, [x13], #0x8\n" - "str d12, [x9], #0x8\n" - "str d16, [x27], #0x8\n" - "str d20, [x25], #0x8\n" - "tbz x15, #2, 104f\n" - "st1 { v8.s }[2], [x13], #0x4\n" - "st1 { v12.s }[2], [x9], #0x4\n" - "st1 { v16.s }[2], [x27], #0x4\n" - "st1 { v20.s }[2], [x25], #0x4\n" - "tbz x15, #1, 103f\n" - "st1 { v8.h }[6], [x13], #0x2\n" - "st1 { v12.h }[6], [x9], #0x2\n" - "st1 { v16.h }[6], [x27], #0x2\n" - "st1 { v20.h }[6], [x25], #0x2\n" - "tbz x15, #0, 110f\n" - "st1 { v8.b }[14], [x13]\n" - "st1 { v12.b }[14], [x9]\n" - "st1 { v16.b }[14], [x27]\n" - "st1 { v20.b }[14], [x25]\n" - "b 110f\n" - "103:" // Height 4: Partial direct writeback: partial_1_12 - "tbz x15, #0, 110f\n" - "st1 { v8.b }[12], [x13]\n" - "st1 { v12.b }[12], [x9]\n" - "st1 { v16.b }[12], [x27]\n" - "st1 { v20.b }[12], [x25]\n" - "b 110f\n" - "104:" // Height 4: Partial direct writeback: partial_2_8 - "tbz x15, #1, 105f\n" - "st1 { v8.h }[4], [x13], #0x2\n" - "st1 { v12.h }[4], [x9], #0x2\n" - "st1 { v16.h }[4], [x27], #0x2\n" - "st1 { v20.h }[4], [x25], #0x2\n" - "tbz x15, #0, 110f\n" - "st1 { v8.b }[10], [x13]\n" - "st1 { v12.b }[10], [x9]\n" - "st1 { v16.b }[10], [x27]\n" - "st1 { v20.b }[10], [x25]\n" - "b 110f\n" - "105:" // Height 4: Partial direct writeback: partial_1_8 - "tbz x15, #0, 110f\n" - "st1 { v8.b }[8], [x13]\n" - "st1 { v12.b }[8], [x9]\n" - "st1 { v16.b }[8], [x27]\n" - "st1 { v20.b }[8], [x25]\n" - "b 110f\n" - "106:" // Height 4: Partial direct writeback: partial_4_0 - "tbz x15, #2, 108f\n" - "str s8, [x13], #0x4\n" - "str s12, [x9], #0x4\n" - "str s16, [x27], #0x4\n" - "str s20, [x25], #0x4\n" - "tbz x15, #1, 107f\n" - "st1 { v8.h }[2], [x13], #0x2\n" - "st1 { v12.h }[2], [x9], #0x2\n" - "st1 { v16.h }[2], [x27], #0x2\n" - "st1 { v20.h }[2], [x25], #0x2\n" - "tbz x15, #0, 110f\n" - "st1 { v8.b }[6], [x13]\n" - "st1 { v12.b }[6], [x9]\n" - "st1 { v16.b }[6], [x27]\n" - "st1 { v20.b }[6], [x25]\n" - "b 110f\n" - "107:" // Height 4: Partial direct writeback: partial_1_4 - "tbz x15, #0, 110f\n" - "st1 { v8.b }[4], [x13]\n" - "st1 { v12.b }[4], [x9]\n" - "st1 { v16.b }[4], [x27]\n" - "st1 { v20.b }[4], [x25]\n" - "b 110f\n" - "108:" // Height 4: Partial direct writeback: partial_2_0 - "tbz x15, #1, 109f\n" - "str h8, [x13], #0x2\n" - "str h12, [x9], #0x2\n" - "str h16, [x27], #0x2\n" - "str h20, [x25], #0x2\n" - "tbz x15, #0, 110f\n" - "st1 { v8.b }[2], [x13]\n" - "st1 { v12.b }[2], [x9]\n" - "st1 { v16.b }[2], [x27]\n" - "st1 { v20.b }[2], [x25]\n" - "b 110f\n" - "109:" // Height 4: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" - "str b12, [x9, #0x0]\n" - "str b16, [x27, #0x0]\n" - "str b20, [x25, #0x0]\n" - "110:" // Height 4: Partial direct writeback: Done - "b 112f\n" - "111:" // Height 4: Full writeback - "str q8, [x13, #0x0]\n" - "str q12, [x9, #0x0]\n" - "str q16, [x27, #0x0]\n" - "str q20, [x25, #0x0]\n" - "add x13, x13, #0x10\n" + "bge 107f\n" + "tbz x10, #3, 102f\n" + "str d8, [x9], #0x8\n" + "str d12, [x23], #0x8\n" + "str d16, [x22], #0x8\n" + "str d20, [x21], #0x8\n" + "tbz x10, #2, 100f\n" + "st1 { v8.s }[2], [x9], #0x4\n" + "st1 { v12.s }[2], [x23], #0x4\n" + "st1 { v16.s }[2], [x22], #0x4\n" + "st1 { v20.s }[2], [x21], #0x4\n" + "tbz x10, #1, 99f\n" + "st1 { v8.h }[6], [x9], #0x2\n" + "st1 { v12.h }[6], [x23], #0x2\n" + "st1 { v16.h }[6], [x22], #0x2\n" + "st1 { v20.h }[6], [x21], #0x2\n" + "tbz x10, #0, 106f\n" + "st1 { v8.b }[14], [x9]\n" + "st1 { v12.b }[14], [x23]\n" + "st1 { v16.b }[14], [x22]\n" + "st1 { v20.b }[14], [x21]\n" + "b 106f\n" + "99:" // Height 4: Partial direct writeback: partial_1_12 + "tbz x10, #0, 106f\n" + "st1 { v8.b }[12], [x9]\n" + "st1 { v12.b }[12], [x23]\n" + "st1 { v16.b }[12], [x22]\n" + "st1 { v20.b }[12], [x21]\n" + "b 106f\n" + "100:" // Height 4: Partial direct writeback: partial_2_8 + "tbz x10, #1, 101f\n" + "st1 { v8.h }[4], [x9], #0x2\n" + "st1 { v12.h }[4], [x23], #0x2\n" + "st1 { v16.h }[4], [x22], #0x2\n" + "st1 { v20.h }[4], [x21], #0x2\n" + "tbz x10, #0, 106f\n" + "st1 { v8.b }[10], [x9]\n" + "st1 { v12.b }[10], [x23]\n" + "st1 { v16.b }[10], [x22]\n" + "st1 { v20.b }[10], [x21]\n" + "b 106f\n" + "101:" // Height 4: Partial direct writeback: partial_1_8 + "tbz x10, #0, 106f\n" + "st1 { v8.b }[8], [x9]\n" + "st1 { v12.b }[8], [x23]\n" + "st1 { v16.b }[8], [x22]\n" + "st1 { v20.b }[8], [x21]\n" + "b 106f\n" + "102:" // Height 4: Partial direct writeback: partial_4_0 + "tbz x10, #2, 104f\n" + "str s8, [x9], #0x4\n" + "str s12, [x23], #0x4\n" + "str s16, [x22], #0x4\n" + "str s20, [x21], #0x4\n" + "tbz x10, #1, 103f\n" + "st1 { v8.h }[2], [x9], #0x2\n" + "st1 { v12.h }[2], [x23], #0x2\n" + "st1 { v16.h }[2], [x22], #0x2\n" + "st1 { v20.h }[2], [x21], #0x2\n" + "tbz x10, #0, 106f\n" + "st1 { v8.b }[6], [x9]\n" + "st1 { v12.b }[6], [x23]\n" + "st1 { v16.b }[6], [x22]\n" + "st1 { v20.b }[6], [x21]\n" + "b 106f\n" + "103:" // Height 4: Partial direct writeback: partial_1_4 + "tbz x10, #0, 106f\n" + "st1 { v8.b }[4], [x9]\n" + "st1 { v12.b }[4], [x23]\n" + "st1 { v16.b }[4], [x22]\n" + "st1 { v20.b }[4], [x21]\n" + "b 106f\n" + "104:" // Height 4: Partial direct writeback: partial_2_0 + "tbz x10, #1, 105f\n" + "str h8, [x9], #0x2\n" + "str h12, [x23], #0x2\n" + "str h16, [x22], #0x2\n" + "str h20, [x21], #0x2\n" + "tbz x10, #0, 106f\n" + "st1 { v8.b }[2], [x9]\n" + "st1 { v12.b }[2], [x23]\n" + "st1 { v16.b }[2], [x22]\n" + "st1 { v20.b }[2], [x21]\n" + "b 106f\n" + "105:" // Height 4: Partial direct writeback: partial_1_0 + "str b8, [x9, #0x0]\n" + "str b12, [x23, #0x0]\n" + "str b16, [x22, #0x0]\n" + "str b20, [x21, #0x0]\n" + "106:" // Height 4: Partial direct writeback: Done + "b 108f\n" + "107:" // Height 4: Full writeback + "str q8, [x9, #0x0]\n" "add x9, x9, #0x10\n" - "add x27, x27, #0x10\n" - "add x25, x25, #0x10\n" - "112:" // Height 4: Writeback done - "subs x15, x15, #0x10\n" - "bgt 87b\n" - "b 170f\n" - "113:" // Height 5 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 114f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19\n" - "ldr x25, [%x[output_ptr], #0x18]\n" - "ldr x23, [%x[output_ptr], #0x20]\n" - "add x27, x27, x19\n" - "add x25, x25, x19\n" - "add x23, x23, x19\n" - "b 115f\n" - "114:" // Height 5: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "add x27, x9, x19\n" - "add x25, x27, x19\n" - "add x23, x25, x19\n" - "115:" // Height 5: Column loop + "str q12, [x23, #0x0]\n" + "str q16, [x22, #0x0]\n" + "str q20, [x21, #0x0]\n" + "108:" // Height 4: Writeback done + "subs x10, x10, #0x10\n" + "bgt 83b\n" + "b 164f\n" + "109:" // Height 5 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "110:" // Height 5: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" "movi v10.4s, #0x0\n" @@ -2002,260 +1945,260 @@ void a64_hybrid_s8qs_dot_6x16 ( "movi v25.4s, #0x0\n" "movi v26.4s, #0x0\n" "movi v27.4s, #0x0\n" - "116:" // Height 5: setup done - "mov x12, #0x0\n" - "117:" // Height 5: String loop + "111:" // Height 5: setup done + "mov x27, #0x0\n" + "112:" // 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, 118f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 113f\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, 119f\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, 114f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "add x26, x26, x19\n" + "add x25, x25, x19\n" "add x24, x24, x19\n" + "add x23, x23, x19\n" "add x22, x22, x19\n" - "b 119f\n" - "118:" // Height 5: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "add x26, x28, x19\n" - "add x24, x26, x19\n" - "add x22, x24, x19\n" - "119:" // Height 5: input setup done - "cmp x11, #0x10\n" - "blt 122f\n" - "cmp x11, #0x20\n" - "blt 121f\n" - "120:" // 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, [x14, #0x0]\n" + "add x21, x21, x19\n" + "b 114f\n" + "113:" // Height 5: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "add x23, x24, x19\n" + "add x22, x23, x19\n" + "add x21, x22, x19\n" + "114:" // Height 5: input setup done + "cmp x26, #0x10\n" + "blt 117f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x20\n" + "ldr q2, [x23, #0x0]\n" + "ldr q3, [x22, #0x0]\n" + "ldr q4, [x21, #0x0]\n" + "ldr q6, [x28, #0x0]\n" + "blt 116f\n" + "115:" // Height 5: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" + "add x22, x22, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" + "add x21, x21, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x24, x24, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "add x22, x22, #0x10\n" + "prfm pldl1keep, [x21, #0x80]\n" + "cmp x26, #0x20\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "sub x11, x11, #0x10\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "cmp x11, #0x20\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x14, #0x40]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "add x14, x14, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" + "ldr q1, [x24, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + "ldr q2, [x23, #0x0]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" + "ldr q3, [x22, #0x0]\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" - "bge 120b\n" - "121:" // Height 5: Multiply loop: Single iteration only - "sub x11, x11, #0x10\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, [x14, #0x0]\n" + "ldr q4, [x21, #0x0]\n" + "bge 115b\n" + "116:" // Height 5: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x28, x28, #0x10\n" + "add x24, x24, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "ldr q6, [x14, #0x20]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "add x22, x22, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x24, x24, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" + "ldr q6, [x28, #0x20]\n" + "add x21, x21, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "add x22, x22, #0x10\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" "prfm pldl1keep, [x22, #0x80]\n" + ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "prfm pldl1keep, [x21, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x14, #0x40]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "add x14, x14, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" @@ -2265,34 +2208,34 @@ void a64_hybrid_s8qs_dot_6x16 ( ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" - "122:" // Height 5: Multiply loop: Main loop skip - "cbz x11, 127f\n" - "cmp x11, #0x4\n" - "blt 124f\n" - "123:" // 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, [x14, #0x0]\n" + "117:" // Height 5: Multiply loop: Main loop skip + "cbz x26, 122f\n" + "cmp x26, #0x4\n" + "blt 119f\n" + "118:" // Height 5: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x4\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x4\n" + "ldr s2, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr s4, [x21], #0x4\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "sub x11, x11, #0x4\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "cmp x11, #0x4\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -2302,45 +2245,45 @@ void a64_hybrid_s8qs_dot_6x16 ( ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" - "bge 123b\n" - "cbz x11, 127f\n" - "124:" // Height 5: Multiply loop: Skip odd blocks - "tbz x11, #1, 125f\n" - "ldr h0, [x10], #0x2\n" - "ldr h1, [x28], #0x2\n" - "ldr h2, [x26], #0x2\n" - "ldr h3, [x24], #0x2\n" - "ldr h4, [x22], #0x2\n" - "tbz x11, #0, 126f\n" - "ld1 { v0.b }[2], [x10]\n" - "ld1 { v1.b }[2], [x28]\n" - "ld1 { v2.b }[2], [x26]\n" - "ld1 { v3.b }[2], [x24]\n" - "ld1 { v4.b }[2], [x22]\n" - "b 126f\n" - "125:" // Height 5: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x10, #0x0]\n" - "ldr b1, [x28, #0x0]\n" - "ldr b2, [x26, #0x0]\n" - "ldr b3, [x24, #0x0]\n" - "ldr b4, [x22, #0x0]\n" - "126:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q6, [x14, #0x0]\n" + "bge 118b\n" + "cbz x26, 122f\n" + "119:" // Height 5: Multiply loop: Skip odd blocks + "tbz x26, #1, 120f\n" + "ldr h0, [x25], #0x2\n" + "ldr h1, [x24], #0x2\n" + "ldr h2, [x23], #0x2\n" + "ldr h3, [x22], #0x2\n" + "ldr h4, [x21], #0x2\n" + "tbz x26, #0, 121f\n" + "ld1 { v0.b }[2], [x25]\n" + "ld1 { v1.b }[2], [x24]\n" + "ld1 { v2.b }[2], [x23]\n" + "ld1 { v3.b }[2], [x22]\n" + "ld1 { v4.b }[2], [x21]\n" + "b 121f\n" + "120:" // Height 5: Multiply loop: Ragged operand read: partial_1_0 + "ldr b0, [x25, #0x0]\n" + "ldr b1, [x24, #0x0]\n" + "ldr b2, [x23, #0x0]\n" + "ldr b3, [x22, #0x0]\n" + "ldr b4, [x21, #0x0]\n" + "121:" // Height 5: Multiply loop: Ragged operand read: Done + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -2350,33 +2293,38 @@ void a64_hybrid_s8qs_dot_6x16 ( ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" - "127:" // Height 5: Multiply loop: No odd multiplies + "122:" // 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 117b\n" - "prfm pstl1keep, [x13, #0x0]\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 112b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "prfm pstl1keep, [x9, #0x0]\n" - "ldr q0, [x16, #0x0]\n" + "add x23, x9, x19\n" + "ldr q0, [x11, #0x0]\n" "add v8.4s, v8.4s, v0.4s\n" - "prfm pstl1keep, [x27, #0x0]\n" - "ldr q1, [x16, #0x10]\n" + "prfm pstl1keep, [x23, #0x0]\n" + "add x22, x23, x19\n" "add v12.4s, v12.4s, v0.4s\n" - "prfm pstl1keep, [x25, #0x0]\n" - "ldr q2, [x16, #0x20]\n" + "prfm pstl1keep, [x22, #0x0]\n" + "add x21, x22, x19\n" "add v16.4s, v16.4s, v0.4s\n" - "prfm pstl1keep, [x23, #0x0]\n" - "ldr q3, [x16, #0x30]\n" + "prfm pstl1keep, [x21, #0x0]\n" + "add x20, x21, x19\n" "add v20.4s, v20.4s, v0.4s\n" - "add x16, x16, #0x40\n" + "prfm pstl1keep, [x20, #0x0]\n" "add v24.4s, v24.4s, v0.4s\n" + "ldr q1, [x11, #0x10]\n" + "ldr q2, [x11, #0x20]\n" "add v9.4s, v9.4s, v1.4s\n" - "add v13.4s, v13.4s, v1.4s\n" + "ldr q3, [x11, #0x30]\n" + "add x11, x11, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" - "add v11.4s, v11.4s, v3.4s\n" + "add v13.4s, v13.4s, v1.4s\n" "add v14.4s, v14.4s, v2.4s\n" - "add v15.4s, v15.4s, v3.4s\n" "add v17.4s, v17.4s, v1.4s\n" + "add v11.4s, v11.4s, v3.4s\n" + "add v15.4s, v15.4s, v3.4s\n" "add v18.4s, v18.4s, v2.4s\n" "add v19.4s, v19.4s, v3.4s\n" "add v21.4s, v21.4s, v1.4s\n" @@ -2385,30 +2333,30 @@ void a64_hybrid_s8qs_dot_6x16 ( "add v25.4s, v25.4s, v1.4s\n" "add v26.4s, v26.4s, v2.4s\n" "add v27.4s, v27.4s, v3.4s\n" - "tbz %x[flags], #4, 128f\n" - "ldr q0, [x17, #0x0]\n" - "ldr q4, [x8, #0x0]\n" - "ldr q1, [x17, #0x10]\n" - "ldr q5, [x8, #0x10]\n" - "ldr q2, [x17, #0x20]\n" - "ldr q6, [x8, #0x20]\n" - "ldr q3, [x17, #0x30]\n" - "ldr q7, [x8, #0x30]\n" - "add x17, x17, #0x40\n" - "add x8, x8, #0x40\n" - "b 129f\n" - "128:" // Height 5: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x19]\n" + "tbz %x[flags], #4, 123f\n" + "ldr q0, [x12, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q1, [x12, #0x10]\n" + "ldr q5, [x13, #0x10]\n" + "ldr q2, [x12, #0x20]\n" + "ldr q6, [x13, #0x20]\n" + "ldr q3, [x12, #0x30]\n" + "add x12, x12, #0x40\n" + "ldr q7, [x13, #0x30]\n" + "add x13, x13, #0x40\n" + "b 124f\n" + "123:" // Height 5: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1r { v0.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1r { v4.4s }, [x24]\n" "mov v2.16b, v0.16b\n" "mov v3.16b, v0.16b\n" "mov v5.16b, v4.16b\n" "mov v6.16b, v4.16b\n" "mov v7.16b, v4.16b\n" - "129:" // Height 5: parameters loaded + "124:" // Height 5: parameters loaded "sqrdmulh v8.4s, v8.4s, v4.4s\n" "sqrdmulh v9.4s, v9.4s, v5.4s\n" "sqrdmulh v10.4s, v10.4s, v6.4s\n" @@ -2429,7 +2377,7 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqrdmulh v25.4s, v25.4s, v5.4s\n" "sqrdmulh v26.4s, v26.4s, v6.4s\n" "sqrdmulh v27.4s, v27.4s, v7.4s\n" - "tbz %x[flags], #5, 130f\n" + "tbz %x[flags], #5, 125f\n" "and v4.16b, v8.16b, v0.16b\n" "sshr v4.4s, v4.4s, #0x1f\n" "and v5.16b, v9.16b, v1.16b\n" @@ -2490,18 +2438,18 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqadd v25.4s, v25.4s, v5.4s\n" "sqadd v26.4s, v26.4s, v6.4s\n" "sqadd v27.4s, v27.4s, v7.4s\n" - "130:" // Height 5: no shift correction + "125:" // Height 5: no shift correction "srshl v8.4s, v8.4s, v0.4s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1r { v4.4s }, [x24]\n" "srshl v9.4s, v9.4s, v1.4s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" "srshl v10.4s, v10.4s, v2.4s\n" - "ld1r { v5.4s }, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1r { v5.4s }, [x24]\n" + "add x24, %x[qp], %[maxval]\n" "srshl v11.4s, v11.4s, v3.4s\n" - "ld1r { v6.4s }, [x19]\n" - "cmp x15, #0x10\n" + "ld1r { v6.4s }, [x24]\n" + "cmp x10, #0x10\n" "srshl v12.4s, v12.4s, v0.4s\n" "srshl v13.4s, v13.4s, v1.4s\n" "srshl v14.4s, v14.4s, v2.4s\n" @@ -2593,158 +2541,134 @@ void a64_hybrid_s8qs_dot_6x16 ( "uzp1 v16.16b, v16.16b, v17.16b\n" "uzp1 v20.16b, v20.16b, v21.16b\n" "uzp1 v24.16b, v24.16b, v25.16b\n" - "bge 139f\n" - "tbz x15, #3, 134f\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 x15, #2, 132f\n" - "st1 { v8.s }[2], [x13], #0x4\n" - "st1 { v12.s }[2], [x9], #0x4\n" - "st1 { v16.s }[2], [x27], #0x4\n" - "st1 { v20.s }[2], [x25], #0x4\n" - "st1 { v24.s }[2], [x23], #0x4\n" - "tbz x15, #1, 131f\n" - "st1 { v8.h }[6], [x13], #0x2\n" - "st1 { v12.h }[6], [x9], #0x2\n" - "st1 { v16.h }[6], [x27], #0x2\n" - "st1 { v20.h }[6], [x25], #0x2\n" - "st1 { v24.h }[6], [x23], #0x2\n" - "tbz x15, #0, 138f\n" - "st1 { v8.b }[14], [x13]\n" - "st1 { v12.b }[14], [x9]\n" - "st1 { v16.b }[14], [x27]\n" - "st1 { v20.b }[14], [x25]\n" - "st1 { v24.b }[14], [x23]\n" - "b 138f\n" - "131:" // Height 5: Partial direct writeback: partial_1_12 - "tbz x15, #0, 138f\n" - "st1 { v8.b }[12], [x13]\n" - "st1 { v12.b }[12], [x9]\n" - "st1 { v16.b }[12], [x27]\n" - "st1 { v20.b }[12], [x25]\n" - "st1 { v24.b }[12], [x23]\n" - "b 138f\n" - "132:" // Height 5: Partial direct writeback: partial_2_8 - "tbz x15, #1, 133f\n" - "st1 { v8.h }[4], [x13], #0x2\n" - "st1 { v12.h }[4], [x9], #0x2\n" - "st1 { v16.h }[4], [x27], #0x2\n" - "st1 { v20.h }[4], [x25], #0x2\n" - "st1 { v24.h }[4], [x23], #0x2\n" - "tbz x15, #0, 138f\n" - "st1 { v8.b }[10], [x13]\n" - "st1 { v12.b }[10], [x9]\n" - "st1 { v16.b }[10], [x27]\n" - "st1 { v20.b }[10], [x25]\n" - "st1 { v24.b }[10], [x23]\n" - "b 138f\n" - "133:" // Height 5: Partial direct writeback: partial_1_8 - "tbz x15, #0, 138f\n" - "st1 { v8.b }[8], [x13]\n" - "st1 { v12.b }[8], [x9]\n" - "st1 { v16.b }[8], [x27]\n" - "st1 { v20.b }[8], [x25]\n" - "st1 { v24.b }[8], [x23]\n" - "b 138f\n" - "134:" // Height 5: Partial direct writeback: partial_4_0 - "tbz x15, #2, 136f\n" - "str s8, [x13], #0x4\n" - "str s12, [x9], #0x4\n" - "str s16, [x27], #0x4\n" - "str s20, [x25], #0x4\n" - "str s24, [x23], #0x4\n" - "tbz x15, #1, 135f\n" - "st1 { v8.h }[2], [x13], #0x2\n" - "st1 { v12.h }[2], [x9], #0x2\n" - "st1 { v16.h }[2], [x27], #0x2\n" - "st1 { v20.h }[2], [x25], #0x2\n" - "st1 { v24.h }[2], [x23], #0x2\n" - "tbz x15, #0, 138f\n" - "st1 { v8.b }[6], [x13]\n" - "st1 { v12.b }[6], [x9]\n" - "st1 { v16.b }[6], [x27]\n" - "st1 { v20.b }[6], [x25]\n" - "st1 { v24.b }[6], [x23]\n" - "b 138f\n" - "135:" // Height 5: Partial direct writeback: partial_1_4 - "tbz x15, #0, 138f\n" - "st1 { v8.b }[4], [x13]\n" - "st1 { v12.b }[4], [x9]\n" - "st1 { v16.b }[4], [x27]\n" - "st1 { v20.b }[4], [x25]\n" - "st1 { v24.b }[4], [x23]\n" - "b 138f\n" - "136:" // Height 5: Partial direct writeback: partial_2_0 - "tbz x15, #1, 137f\n" - "str h8, [x13], #0x2\n" - "str h12, [x9], #0x2\n" - "str h16, [x27], #0x2\n" - "str h20, [x25], #0x2\n" - "str h24, [x23], #0x2\n" - "tbz x15, #0, 138f\n" - "st1 { v8.b }[2], [x13]\n" - "st1 { v12.b }[2], [x9]\n" - "st1 { v16.b }[2], [x27]\n" - "st1 { v20.b }[2], [x25]\n" - "st1 { v24.b }[2], [x23]\n" - "b 138f\n" - "137:" // Height 5: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" - "str b12, [x9, #0x0]\n" - "str b16, [x27, #0x0]\n" - "str b20, [x25, #0x0]\n" - "str b24, [x23, #0x0]\n" - "138:" // Height 5: Partial direct writeback: Done - "b 140f\n" - "139:" // Height 5: Full writeback - "str q8, [x13, #0x0]\n" - "str q12, [x9, #0x0]\n" - "str q16, [x27, #0x0]\n" - "str q20, [x25, #0x0]\n" - "str q24, [x23, #0x0]\n" - "add x13, x13, #0x10\n" + "bge 134f\n" + "tbz x10, #3, 129f\n" + "str d8, [x9], #0x8\n" + "str d12, [x23], #0x8\n" + "str d16, [x22], #0x8\n" + "str d20, [x21], #0x8\n" + "str d24, [x20], #0x8\n" + "tbz x10, #2, 127f\n" + "st1 { v8.s }[2], [x9], #0x4\n" + "st1 { v12.s }[2], [x23], #0x4\n" + "st1 { v16.s }[2], [x22], #0x4\n" + "st1 { v20.s }[2], [x21], #0x4\n" + "st1 { v24.s }[2], [x20], #0x4\n" + "tbz x10, #1, 126f\n" + "st1 { v8.h }[6], [x9], #0x2\n" + "st1 { v12.h }[6], [x23], #0x2\n" + "st1 { v16.h }[6], [x22], #0x2\n" + "st1 { v20.h }[6], [x21], #0x2\n" + "st1 { v24.h }[6], [x20], #0x2\n" + "tbz x10, #0, 133f\n" + "st1 { v8.b }[14], [x9]\n" + "st1 { v12.b }[14], [x23]\n" + "st1 { v16.b }[14], [x22]\n" + "st1 { v20.b }[14], [x21]\n" + "st1 { v24.b }[14], [x20]\n" + "b 133f\n" + "126:" // Height 5: Partial direct writeback: partial_1_12 + "tbz x10, #0, 133f\n" + "st1 { v8.b }[12], [x9]\n" + "st1 { v12.b }[12], [x23]\n" + "st1 { v16.b }[12], [x22]\n" + "st1 { v20.b }[12], [x21]\n" + "st1 { v24.b }[12], [x20]\n" + "b 133f\n" + "127:" // Height 5: Partial direct writeback: partial_2_8 + "tbz x10, #1, 128f\n" + "st1 { v8.h }[4], [x9], #0x2\n" + "st1 { v12.h }[4], [x23], #0x2\n" + "st1 { v16.h }[4], [x22], #0x2\n" + "st1 { v20.h }[4], [x21], #0x2\n" + "st1 { v24.h }[4], [x20], #0x2\n" + "tbz x10, #0, 133f\n" + "st1 { v8.b }[10], [x9]\n" + "st1 { v12.b }[10], [x23]\n" + "st1 { v16.b }[10], [x22]\n" + "st1 { v20.b }[10], [x21]\n" + "st1 { v24.b }[10], [x20]\n" + "b 133f\n" + "128:" // Height 5: Partial direct writeback: partial_1_8 + "tbz x10, #0, 133f\n" + "st1 { v8.b }[8], [x9]\n" + "st1 { v12.b }[8], [x23]\n" + "st1 { v16.b }[8], [x22]\n" + "st1 { v20.b }[8], [x21]\n" + "st1 { v24.b }[8], [x20]\n" + "b 133f\n" + "129:" // Height 5: Partial direct writeback: partial_4_0 + "tbz x10, #2, 131f\n" + "str s8, [x9], #0x4\n" + "str s12, [x23], #0x4\n" + "str s16, [x22], #0x4\n" + "str s20, [x21], #0x4\n" + "str s24, [x20], #0x4\n" + "tbz x10, #1, 130f\n" + "st1 { v8.h }[2], [x9], #0x2\n" + "st1 { v12.h }[2], [x23], #0x2\n" + "st1 { v16.h }[2], [x22], #0x2\n" + "st1 { v20.h }[2], [x21], #0x2\n" + "st1 { v24.h }[2], [x20], #0x2\n" + "tbz x10, #0, 133f\n" + "st1 { v8.b }[6], [x9]\n" + "st1 { v12.b }[6], [x23]\n" + "st1 { v16.b }[6], [x22]\n" + "st1 { v20.b }[6], [x21]\n" + "st1 { v24.b }[6], [x20]\n" + "b 133f\n" + "130:" // Height 5: Partial direct writeback: partial_1_4 + "tbz x10, #0, 133f\n" + "st1 { v8.b }[4], [x9]\n" + "st1 { v12.b }[4], [x23]\n" + "st1 { v16.b }[4], [x22]\n" + "st1 { v20.b }[4], [x21]\n" + "st1 { v24.b }[4], [x20]\n" + "b 133f\n" + "131:" // Height 5: Partial direct writeback: partial_2_0 + "tbz x10, #1, 132f\n" + "str h8, [x9], #0x2\n" + "str h12, [x23], #0x2\n" + "str h16, [x22], #0x2\n" + "str h20, [x21], #0x2\n" + "str h24, [x20], #0x2\n" + "tbz x10, #0, 133f\n" + "st1 { v8.b }[2], [x9]\n" + "st1 { v12.b }[2], [x23]\n" + "st1 { v16.b }[2], [x22]\n" + "st1 { v20.b }[2], [x21]\n" + "st1 { v24.b }[2], [x20]\n" + "b 133f\n" + "132:" // Height 5: Partial direct writeback: partial_1_0 + "str b8, [x9, #0x0]\n" + "str b12, [x23, #0x0]\n" + "str b16, [x22, #0x0]\n" + "str b20, [x21, #0x0]\n" + "str b24, [x20, #0x0]\n" + "133:" // Height 5: Partial direct writeback: Done + "b 135f\n" + "134:" // Height 5: Full writeback + "str q8, [x9, #0x0]\n" "add x9, x9, #0x10\n" - "add x27, x27, #0x10\n" - "add x25, x25, #0x10\n" - "add x23, x23, #0x10\n" - "140:" // Height 5: Writeback done - "subs x15, x15, #0x10\n" - "bgt 115b\n" - "b 170f\n" - "141:" // Height 6 - "ldr x8, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" - "mov x16, %x[col_bias]\n" - "ldr x17, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "str q12, [x23, #0x0]\n" + "str q16, [x22, #0x0]\n" + "str q20, [x21, #0x0]\n" + "str q24, [x20, #0x0]\n" + "135:" // Height 5: Writeback done + "subs x10, x10, #0x10\n" + "bgt 110b\n" + "b 164f\n" + "136:" // Height 6 + "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" + "mov x11, %x[col_bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" + "mov x9, %x[output_ptr]\n" + "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" + "mov x20, #0x6\n" + "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" - "tbz %x[flags], #2, 142f\n" - "ldr x13, [%x[output_ptr], #0x0]\n" - "add x13, x13, x19\n" - "ldr x9, [%x[output_ptr], #0x8]\n" - "ldr x27, [%x[output_ptr], #0x10]\n" - "add x9, x9, x19\n" - "ldr x25, [%x[output_ptr], #0x18]\n" - "ldr x23, [%x[output_ptr], #0x20]\n" - "add x27, x27, x19\n" - "ldr x21, [%x[output_ptr], #0x28]\n" - "add %x[output_ptr], %x[output_ptr], #0x30\n" - "add x25, x25, x19\n" - "add x23, x23, x19\n" - "add x21, x21, x19\n" - "b 143f\n" - "142:" // Height 6: setup direct output - "mov x13, %x[output_ptr]\n" - "add x9, x13, x19\n" - "add x27, x9, x19\n" - "add x25, x27, x19\n" - "add x23, x25, x19\n" - "add x21, x23, x19\n" - "add %x[output_ptr], x21, x19\n" - "143:" // Height 6: Column loop + "madd %x[output_ptr], x19, x20, %x[output_ptr]\n" + "137:" // Height 6: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" "movi v10.4s, #0x0\n" @@ -2769,299 +2693,299 @@ void a64_hybrid_s8qs_dot_6x16 ( "movi v29.4s, #0x0\n" "movi v30.4s, #0x0\n" "movi v31.4s, #0x0\n" - "144:" // Height 6: setup done - "mov x12, #0x0\n" - "145:" // Height 6: String loop + "138:" // Height 6: setup done + "mov x27, #0x0\n" + "139:" // 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, 146f\n" - "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "ldr w26, [x20, x27, LSL #0x2]\n" + "tbz %x[flags], #3, 140f\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, 147f\n" + "cbnz x27, 141f\n" "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x10, x10, x19\n" - "add x28, x28, x19\n" - "add x26, x26, x19\n" + "add x25, x25, x19\n" "add x24, x24, x19\n" + "add x23, x23, x19\n" "add x22, x22, x19\n" + "add x21, x21, x19\n" "add x20, x20, x19\n" - "b 147f\n" - "146:" // Height 6: setup direct input - "mov x10, %x[input_ptr]\n" - "add x28, x10, x19\n" - "add x26, x28, x19\n" - "add x24, x26, x19\n" - "add x22, x24, x19\n" - "add x20, x22, x19\n" - "147:" // Height 6: input setup done - "cmp x11, #0x10\n" - "blt 150f\n" - "cmp x11, #0x20\n" - "blt 149f\n" - "148:" // 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 141f\n" + "140:" // Height 6: setup direct input + "mov x25, %x[input_ptr]\n" + "add x24, x25, x19\n" + "add x23, x24, x19\n" + "add x22, x23, x19\n" + "add x21, x22, x19\n" + "add x20, x21, x19\n" + "141:" // Height 6: input setup done + "cmp x26, #0x10\n" + "blt 144f\n" + "ldr q0, [x25, #0x0]\n" + "ldr q1, [x24, #0x0]\n" + "cmp x26, #0x20\n" + "ldr q2, [x23, #0x0]\n" + "ldr q3, [x22, #0x0]\n" + "ldr q4, [x21, #0x0]\n" "ldr q5, [x20, #0x0]\n" - "ldr q6, [x14, #0x0]\n" + "ldr q6, [x28, #0x0]\n" + "blt 143f\n" + "142:" // Height 6: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "add x25, x25, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "add x24, x24, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x28, x28, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "add x23, x23, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "add x26, x26, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" + "add x22, x22, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x22, #0x80]\n" + "add x21, x21, #0x10\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" + "add x20, x20, #0x10\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x22, x22, #0x10\n" - "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x21, #0x80]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "add x20, x20, #0x10\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" "prfm pldl1keep, [x20, #0x80]\n" - "sub x11, x11, #0x10\n" + "cmp x26, #0x20\n" + ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "cmp x11, #0x20\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" - "ldr q6, [x14, #0x40]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" ".inst 0x4fa5e0dc // sdot v28.4s, v6.16b, v5.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0fd // sdot v29.4s, v7.16b, v5.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" ".inst 0x4fa5e0de // sdot v30.4s, v6.16b, v5.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0ff // sdot v31.4s, v7.16b, v5.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8dc // sdot v28.4s, v6.16b, v5.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8fd // sdot v29.4s, v7.16b, v5.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8de // sdot v30.4s, v6.16b, v5.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8ff // sdot v31.4s, v7.16b, v5.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8dc // sdot v28.4s, v6.16b, v5.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8fd // sdot v29.4s, v7.16b, v5.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "add x14, x14, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8de // sdot v30.4s, v6.16b, v5.4b[3]\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr q0, [x25, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" + "ldr q1, [x24, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + "ldr q2, [x23, #0x0]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" + "ldr q3, [x22, #0x0]\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" + "ldr q4, [x21, #0x0]\n" ".inst 0x4fa5e8ff // sdot v31.4s, v7.16b, v5.4b[3]\n" - "bge 148b\n" - "149:" // Height 6: Multiply loop: Single iteration only - "sub x11, x11, #0x10\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, [x14, #0x0]\n" + "bge 142b\n" + "143:" // Height 6: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" + "sub x26, x26, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x10, x10, #0x10\n" - "prfm pldl1keep, [x10, #0x80]\n" + "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x28, x28, #0x10\n" - ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "prfm pldl1keep, [x28, #0x80]\n" - "add x26, x26, #0x10\n" - ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" "add x24, x24, #0x10\n" - ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" + ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" "prfm pldl1keep, [x24, #0x80]\n" - "ldr q6, [x14, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "add x23, x23, #0x10\n" + ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "prfm pldl1keep, [x23, #0x80]\n" "add x22, x22, #0x10\n" + ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "add x21, x21, #0x10\n" + ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x28, #0x20]\n" "add x20, x20, #0x10\n" + ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "prfm pldl1keep, [x21, #0x80]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" "prfm pldl1keep, [x20, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" - "ldr q6, [x14, #0x40]\n" + "ldr q6, [x28, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x14, #0x50]\n" + "ldr q7, [x28, #0x50]\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" ".inst 0x4fa5e0dc // sdot v28.4s, v6.16b, v5.4b[1]\n" - "ldr q6, [x14, #0x60]\n" + "ldr q6, [x28, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0fd // sdot v29.4s, v7.16b, v5.4b[1]\n" - "ldr q7, [x14, #0x70]\n" + "ldr q7, [x28, #0x70]\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" ".inst 0x4fa5e0de // sdot v30.4s, v6.16b, v5.4b[1]\n" - "ldr q6, [x14, #0x80]\n" + "ldr q6, [x28, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0ff // sdot v31.4s, v7.16b, v5.4b[1]\n" - "ldr q7, [x14, #0x90]\n" + "ldr q7, [x28, #0x90]\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8dc // sdot v28.4s, v6.16b, v5.4b[2]\n" - "ldr q6, [x14, #0xa0]\n" + "ldr q6, [x28, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8fd // sdot v29.4s, v7.16b, v5.4b[2]\n" - "ldr q7, [x14, #0xb0]\n" + "ldr q7, [x28, #0xb0]\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8de // sdot v30.4s, v6.16b, v5.4b[2]\n" - "ldr q6, [x14, #0xc0]\n" + "ldr q6, [x28, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8ff // sdot v31.4s, v7.16b, v5.4b[2]\n" - "ldr q7, [x14, #0xd0]\n" + "ldr q7, [x28, #0xd0]\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8dc // sdot v28.4s, v6.16b, v5.4b[3]\n" - "ldr q6, [x14, #0xe0]\n" + "ldr q6, [x28, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8fd // sdot v29.4s, v7.16b, v5.4b[3]\n" - "ldr q7, [x14, #0xf0]\n" + "ldr q7, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "add x14, x14, #0x100\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" @@ -3073,37 +2997,37 @@ void a64_hybrid_s8qs_dot_6x16 ( ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8ff // sdot v31.4s, v7.16b, v5.4b[3]\n" - "150:" // Height 6: Multiply loop: Main loop skip - "cbz x11, 155f\n" - "cmp x11, #0x4\n" - "blt 152f\n" - "151:" // 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" + "144:" // Height 6: Multiply loop: Main loop skip + "cbz x26, 149f\n" + "cmp x26, #0x4\n" + "blt 146f\n" + "145:" // Height 6: Multiply loop: Odd block loop + "ldr s0, [x25], #0x4\n" + "sub x26, x26, #0x4\n" + "ldr s1, [x24], #0x4\n" + "cmp x26, #0x4\n" + "ldr s2, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr s4, [x21], #0x4\n" "ldr s5, [x20], #0x4\n" - "ldr q6, [x14, #0x0]\n" + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "sub x11, x11, #0x4\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "cmp x11, #0x4\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -3115,50 +3039,50 @@ void a64_hybrid_s8qs_dot_6x16 ( ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" - "bge 151b\n" - "cbz x11, 155f\n" - "152:" // Height 6: Multiply loop: Skip odd blocks - "tbz x11, #1, 153f\n" - "ldr h0, [x10], #0x2\n" - "ldr h1, [x28], #0x2\n" - "ldr h2, [x26], #0x2\n" - "ldr h3, [x24], #0x2\n" - "ldr h4, [x22], #0x2\n" + "bge 145b\n" + "cbz x26, 149f\n" + "146:" // Height 6: Multiply loop: Skip odd blocks + "tbz x26, #1, 147f\n" + "ldr h0, [x25], #0x2\n" + "ldr h1, [x24], #0x2\n" + "ldr h2, [x23], #0x2\n" + "ldr h3, [x22], #0x2\n" + "ldr h4, [x21], #0x2\n" "ldr h5, [x20], #0x2\n" - "tbz x11, #0, 154f\n" - "ld1 { v0.b }[2], [x10]\n" - "ld1 { v1.b }[2], [x28]\n" - "ld1 { v2.b }[2], [x26]\n" - "ld1 { v3.b }[2], [x24]\n" - "ld1 { v4.b }[2], [x22]\n" + "tbz x26, #0, 148f\n" + "ld1 { v0.b }[2], [x25]\n" + "ld1 { v1.b }[2], [x24]\n" + "ld1 { v2.b }[2], [x23]\n" + "ld1 { v3.b }[2], [x22]\n" + "ld1 { v4.b }[2], [x21]\n" "ld1 { v5.b }[2], [x20]\n" - "b 154f\n" - "153:" // Height 6: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x10, #0x0]\n" - "ldr b1, [x28, #0x0]\n" - "ldr b2, [x26, #0x0]\n" - "ldr b3, [x24, #0x0]\n" - "ldr b4, [x22, #0x0]\n" + "b 148f\n" + "147:" // Height 6: Multiply loop: Ragged operand read: partial_1_0 + "ldr b0, [x25, #0x0]\n" + "ldr b1, [x24, #0x0]\n" + "ldr b2, [x23, #0x0]\n" + "ldr b3, [x22, #0x0]\n" + "ldr b4, [x21, #0x0]\n" "ldr b5, [x20, #0x0]\n" - "154:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q6, [x14, #0x0]\n" + "148:" // Height 6: Multiply loop: Ragged operand read: Done + "ldr q6, [x28, #0x0]\n" ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x14, #0x10]\n" + "ldr q7, [x28, #0x10]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" - "ldr q6, [x14, #0x20]\n" + "ldr q6, [x28, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x14, #0x30]\n" + "ldr q7, [x28, #0x30]\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "add x14, x14, #0x40\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" @@ -3170,35 +3094,41 @@ void a64_hybrid_s8qs_dot_6x16 ( ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" - "155:" // Height 6: Multiply loop: No odd multiplies + "149:" // 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 145b\n" - "prfm pstl1keep, [x13, #0x0]\n" + "add x27, x27, #0x1\n" + "cmp x27, x19\n" + "bne 139b\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" "prfm pstl1keep, [x9, #0x0]\n" - "ldr q0, [x16, #0x0]\n" + "add x23, x9, x19\n" + "ldr q0, [x11, #0x0]\n" "add v8.4s, v8.4s, v0.4s\n" - "prfm pstl1keep, [x27, #0x0]\n" - "ldr q1, [x16, #0x10]\n" + "prfm pstl1keep, [x23, #0x0]\n" + "add x22, x23, x19\n" "add v12.4s, v12.4s, v0.4s\n" - "prfm pstl1keep, [x25, #0x0]\n" - "ldr q2, [x16, #0x20]\n" + "prfm pstl1keep, [x22, #0x0]\n" + "add x21, x22, x19\n" "add v16.4s, v16.4s, v0.4s\n" - "prfm pstl1keep, [x23, #0x0]\n" - "ldr q3, [x16, #0x30]\n" - "add v20.4s, v20.4s, v0.4s\n" "prfm pstl1keep, [x21, #0x0]\n" - "add x16, x16, #0x40\n" + "add x20, x21, x19\n" + "add v20.4s, v20.4s, v0.4s\n" + "prfm pstl1keep, [x20, #0x0]\n" + "add x19, x20, x19\n" "add v24.4s, v24.4s, v0.4s\n" + "prfm pstl1keep, [x19, #0x0]\n" "add v28.4s, v28.4s, v0.4s\n" + "ldr q1, [x11, #0x10]\n" + "ldr q2, [x11, #0x20]\n" "add v9.4s, v9.4s, v1.4s\n" + "ldr q3, [x11, #0x30]\n" + "add x11, x11, #0x40\n" "add v10.4s, v10.4s, v2.4s\n" - "add v11.4s, v11.4s, v3.4s\n" "add v13.4s, v13.4s, v1.4s\n" "add v14.4s, v14.4s, v2.4s\n" - "add v15.4s, v15.4s, v3.4s\n" "add v17.4s, v17.4s, v1.4s\n" + "add v11.4s, v11.4s, v3.4s\n" + "add v15.4s, v15.4s, v3.4s\n" "add v18.4s, v18.4s, v2.4s\n" "add v19.4s, v19.4s, v3.4s\n" "add v21.4s, v21.4s, v1.4s\n" @@ -3210,30 +3140,30 @@ void a64_hybrid_s8qs_dot_6x16 ( "add v29.4s, v29.4s, v1.4s\n" "add v30.4s, v30.4s, v2.4s\n" "add v31.4s, v31.4s, v3.4s\n" - "tbz %x[flags], #4, 156f\n" - "ldr q0, [x17, #0x0]\n" - "ldr q4, [x8, #0x0]\n" - "ldr q1, [x17, #0x10]\n" - "ldr q5, [x8, #0x10]\n" - "ldr q2, [x17, #0x20]\n" - "ldr q6, [x8, #0x20]\n" - "ldr q3, [x17, #0x30]\n" - "ldr q7, [x8, #0x30]\n" - "add x17, x17, #0x40\n" - "add x8, x8, #0x40\n" - "b 157f\n" - "156:" // Height 6: per layer parameters - "add x19, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x19]\n" + "tbz %x[flags], #4, 150f\n" + "ldr q0, [x12, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q1, [x12, #0x10]\n" + "ldr q5, [x13, #0x10]\n" + "ldr q2, [x12, #0x20]\n" + "ldr q6, [x13, #0x20]\n" + "ldr q3, [x12, #0x30]\n" + "add x12, x12, #0x40\n" + "ldr q7, [x13, #0x30]\n" + "add x13, x13, #0x40\n" + "b 151f\n" + "150:" // Height 6: per layer parameters + "add x24, %x[qp], %[per_layer_right_shift]\n" + "ld1r { v0.4s }, [x24]\n" "mov v1.16b, v0.16b\n" - "add x19, %x[qp], %[per_layer_mul]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[per_layer_mul]\n" + "ld1r { v4.4s }, [x24]\n" "mov v2.16b, v0.16b\n" "mov v3.16b, v0.16b\n" "mov v5.16b, v4.16b\n" "mov v6.16b, v4.16b\n" "mov v7.16b, v4.16b\n" - "157:" // Height 6: parameters loaded + "151:" // Height 6: parameters loaded "sqrdmulh v8.4s, v8.4s, v4.4s\n" "sqrdmulh v9.4s, v9.4s, v5.4s\n" "sqrdmulh v10.4s, v10.4s, v6.4s\n" @@ -3258,7 +3188,7 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqrdmulh v29.4s, v29.4s, v5.4s\n" "sqrdmulh v30.4s, v30.4s, v6.4s\n" "sqrdmulh v31.4s, v31.4s, v7.4s\n" - "tbz %x[flags], #5, 158f\n" + "tbz %x[flags], #5, 152f\n" "and v4.16b, v8.16b, v0.16b\n" "sshr v4.4s, v4.4s, #0x1f\n" "and v5.16b, v9.16b, v1.16b\n" @@ -3331,18 +3261,18 @@ void a64_hybrid_s8qs_dot_6x16 ( "sqadd v29.4s, v29.4s, v5.4s\n" "sqadd v30.4s, v30.4s, v6.4s\n" "sqadd v31.4s, v31.4s, v7.4s\n" - "158:" // Height 6: no shift correction + "152:" // Height 6: no shift correction "srshl v8.4s, v8.4s, v0.4s\n" - "add x19, %x[qp], %[c_offset]\n" - "ld1r { v4.4s }, [x19]\n" + "add x24, %x[qp], %[c_offset]\n" + "ld1r { v4.4s }, [x24]\n" "srshl v9.4s, v9.4s, v1.4s\n" - "add x19, %x[qp], %[minval]\n" + "add x24, %x[qp], %[minval]\n" "srshl v10.4s, v10.4s, v2.4s\n" - "ld1r { v5.4s }, [x19]\n" - "add x19, %x[qp], %[maxval]\n" + "ld1r { v5.4s }, [x24]\n" + "add x24, %x[qp], %[maxval]\n" "srshl v11.4s, v11.4s, v3.4s\n" - "ld1r { v6.4s }, [x19]\n" - "cmp x15, #0x10\n" + "ld1r { v6.4s }, [x24]\n" + "cmp x10, #0x10\n" "srshl v12.4s, v12.4s, v0.4s\n" "srshl v13.4s, v13.4s, v1.4s\n" "srshl v14.4s, v14.4s, v2.4s\n" @@ -3453,159 +3383,154 @@ void a64_hybrid_s8qs_dot_6x16 ( "uzp1 v20.16b, v20.16b, v21.16b\n" "uzp1 v24.16b, v24.16b, v25.16b\n" "uzp1 v28.16b, v28.16b, v29.16b\n" - "bge 167f\n" - "tbz x15, #3, 162f\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 x15, #2, 160f\n" - "st1 { v8.s }[2], [x13], #0x4\n" - "st1 { v12.s }[2], [x9], #0x4\n" - "st1 { v16.s }[2], [x27], #0x4\n" - "st1 { v20.s }[2], [x25], #0x4\n" - "st1 { v24.s }[2], [x23], #0x4\n" - "st1 { v28.s }[2], [x21], #0x4\n" - "tbz x15, #1, 159f\n" - "st1 { v8.h }[6], [x13], #0x2\n" - "st1 { v12.h }[6], [x9], #0x2\n" - "st1 { v16.h }[6], [x27], #0x2\n" - "st1 { v20.h }[6], [x25], #0x2\n" - "st1 { v24.h }[6], [x23], #0x2\n" - "st1 { v28.h }[6], [x21], #0x2\n" - "tbz x15, #0, 166f\n" - "st1 { v8.b }[14], [x13]\n" - "st1 { v12.b }[14], [x9]\n" - "st1 { v16.b }[14], [x27]\n" - "st1 { v20.b }[14], [x25]\n" - "st1 { v24.b }[14], [x23]\n" - "st1 { v28.b }[14], [x21]\n" - "b 166f\n" - "159:" // Height 6: Partial direct writeback: partial_1_12 - "tbz x15, #0, 166f\n" - "st1 { v8.b }[12], [x13]\n" - "st1 { v12.b }[12], [x9]\n" - "st1 { v16.b }[12], [x27]\n" - "st1 { v20.b }[12], [x25]\n" - "st1 { v24.b }[12], [x23]\n" - "st1 { v28.b }[12], [x21]\n" - "b 166f\n" - "160:" // Height 6: Partial direct writeback: partial_2_8 - "tbz x15, #1, 161f\n" - "st1 { v8.h }[4], [x13], #0x2\n" - "st1 { v12.h }[4], [x9], #0x2\n" - "st1 { v16.h }[4], [x27], #0x2\n" - "st1 { v20.h }[4], [x25], #0x2\n" - "st1 { v24.h }[4], [x23], #0x2\n" - "st1 { v28.h }[4], [x21], #0x2\n" - "tbz x15, #0, 166f\n" - "st1 { v8.b }[10], [x13]\n" - "st1 { v12.b }[10], [x9]\n" - "st1 { v16.b }[10], [x27]\n" - "st1 { v20.b }[10], [x25]\n" - "st1 { v24.b }[10], [x23]\n" - "st1 { v28.b }[10], [x21]\n" - "b 166f\n" - "161:" // Height 6: Partial direct writeback: partial_1_8 - "tbz x15, #0, 166f\n" - "st1 { v8.b }[8], [x13]\n" - "st1 { v12.b }[8], [x9]\n" - "st1 { v16.b }[8], [x27]\n" - "st1 { v20.b }[8], [x25]\n" - "st1 { v24.b }[8], [x23]\n" - "st1 { v28.b }[8], [x21]\n" - "b 166f\n" - "162:" // Height 6: Partial direct writeback: partial_4_0 - "tbz x15, #2, 164f\n" - "str s8, [x13], #0x4\n" - "str s12, [x9], #0x4\n" - "str s16, [x27], #0x4\n" - "str s20, [x25], #0x4\n" - "str s24, [x23], #0x4\n" - "str s28, [x21], #0x4\n" - "tbz x15, #1, 163f\n" - "st1 { v8.h }[2], [x13], #0x2\n" - "st1 { v12.h }[2], [x9], #0x2\n" - "st1 { v16.h }[2], [x27], #0x2\n" - "st1 { v20.h }[2], [x25], #0x2\n" - "st1 { v24.h }[2], [x23], #0x2\n" - "st1 { v28.h }[2], [x21], #0x2\n" - "tbz x15, #0, 166f\n" - "st1 { v8.b }[6], [x13]\n" - "st1 { v12.b }[6], [x9]\n" - "st1 { v16.b }[6], [x27]\n" - "st1 { v20.b }[6], [x25]\n" - "st1 { v24.b }[6], [x23]\n" - "st1 { v28.b }[6], [x21]\n" - "b 166f\n" - "163:" // Height 6: Partial direct writeback: partial_1_4 - "tbz x15, #0, 166f\n" - "st1 { v8.b }[4], [x13]\n" - "st1 { v12.b }[4], [x9]\n" - "st1 { v16.b }[4], [x27]\n" - "st1 { v20.b }[4], [x25]\n" - "st1 { v24.b }[4], [x23]\n" - "st1 { v28.b }[4], [x21]\n" - "b 166f\n" - "164:" // Height 6: Partial direct writeback: partial_2_0 - "tbz x15, #1, 165f\n" - "str h8, [x13], #0x2\n" - "str h12, [x9], #0x2\n" - "str h16, [x27], #0x2\n" - "str h20, [x25], #0x2\n" - "str h24, [x23], #0x2\n" - "str h28, [x21], #0x2\n" - "tbz x15, #0, 166f\n" - "st1 { v8.b }[2], [x13]\n" - "st1 { v12.b }[2], [x9]\n" - "st1 { v16.b }[2], [x27]\n" - "st1 { v20.b }[2], [x25]\n" - "st1 { v24.b }[2], [x23]\n" - "st1 { v28.b }[2], [x21]\n" - "b 166f\n" - "165:" // Height 6: Partial direct writeback: partial_1_0 - "str b8, [x13, #0x0]\n" - "str b12, [x9, #0x0]\n" - "str b16, [x27, #0x0]\n" - "str b20, [x25, #0x0]\n" - "str b24, [x23, #0x0]\n" - "str b28, [x21, #0x0]\n" - "166:" // Height 6: Partial direct writeback: Done - "b 168f\n" - "167:" // Height 6: Full writeback - "str q8, [x13, #0x0]\n" - "str q12, [x9, #0x0]\n" - "str q16, [x27, #0x0]\n" - "str q20, [x25, #0x0]\n" - "str q24, [x23, #0x0]\n" - "str q28, [x21, #0x0]\n" - "add x13, x13, #0x10\n" + "bge 161f\n" + "tbz x10, #3, 156f\n" + "str d8, [x9], #0x8\n" + "str d12, [x23], #0x8\n" + "str d16, [x22], #0x8\n" + "str d20, [x21], #0x8\n" + "str d24, [x20], #0x8\n" + "str d28, [x19], #0x8\n" + "tbz x10, #2, 154f\n" + "st1 { v8.s }[2], [x9], #0x4\n" + "st1 { v12.s }[2], [x23], #0x4\n" + "st1 { v16.s }[2], [x22], #0x4\n" + "st1 { v20.s }[2], [x21], #0x4\n" + "st1 { v24.s }[2], [x20], #0x4\n" + "st1 { v28.s }[2], [x19], #0x4\n" + "tbz x10, #1, 153f\n" + "st1 { v8.h }[6], [x9], #0x2\n" + "st1 { v12.h }[6], [x23], #0x2\n" + "st1 { v16.h }[6], [x22], #0x2\n" + "st1 { v20.h }[6], [x21], #0x2\n" + "st1 { v24.h }[6], [x20], #0x2\n" + "st1 { v28.h }[6], [x19], #0x2\n" + "tbz x10, #0, 160f\n" + "st1 { v8.b }[14], [x9]\n" + "st1 { v12.b }[14], [x23]\n" + "st1 { v16.b }[14], [x22]\n" + "st1 { v20.b }[14], [x21]\n" + "st1 { v24.b }[14], [x20]\n" + "st1 { v28.b }[14], [x19]\n" + "b 160f\n" + "153:" // Height 6: Partial direct writeback: partial_1_12 + "tbz x10, #0, 160f\n" + "st1 { v8.b }[12], [x9]\n" + "st1 { v12.b }[12], [x23]\n" + "st1 { v16.b }[12], [x22]\n" + "st1 { v20.b }[12], [x21]\n" + "st1 { v24.b }[12], [x20]\n" + "st1 { v28.b }[12], [x19]\n" + "b 160f\n" + "154:" // Height 6: Partial direct writeback: partial_2_8 + "tbz x10, #1, 155f\n" + "st1 { v8.h }[4], [x9], #0x2\n" + "st1 { v12.h }[4], [x23], #0x2\n" + "st1 { v16.h }[4], [x22], #0x2\n" + "st1 { v20.h }[4], [x21], #0x2\n" + "st1 { v24.h }[4], [x20], #0x2\n" + "st1 { v28.h }[4], [x19], #0x2\n" + "tbz x10, #0, 160f\n" + "st1 { v8.b }[10], [x9]\n" + "st1 { v12.b }[10], [x23]\n" + "st1 { v16.b }[10], [x22]\n" + "st1 { v20.b }[10], [x21]\n" + "st1 { v24.b }[10], [x20]\n" + "st1 { v28.b }[10], [x19]\n" + "b 160f\n" + "155:" // Height 6: Partial direct writeback: partial_1_8 + "tbz x10, #0, 160f\n" + "st1 { v8.b }[8], [x9]\n" + "st1 { v12.b }[8], [x23]\n" + "st1 { v16.b }[8], [x22]\n" + "st1 { v20.b }[8], [x21]\n" + "st1 { v24.b }[8], [x20]\n" + "st1 { v28.b }[8], [x19]\n" + "b 160f\n" + "156:" // Height 6: Partial direct writeback: partial_4_0 + "tbz x10, #2, 158f\n" + "str s8, [x9], #0x4\n" + "str s12, [x23], #0x4\n" + "str s16, [x22], #0x4\n" + "str s20, [x21], #0x4\n" + "str s24, [x20], #0x4\n" + "str s28, [x19], #0x4\n" + "tbz x10, #1, 157f\n" + "st1 { v8.h }[2], [x9], #0x2\n" + "st1 { v12.h }[2], [x23], #0x2\n" + "st1 { v16.h }[2], [x22], #0x2\n" + "st1 { v20.h }[2], [x21], #0x2\n" + "st1 { v24.h }[2], [x20], #0x2\n" + "st1 { v28.h }[2], [x19], #0x2\n" + "tbz x10, #0, 160f\n" + "st1 { v8.b }[6], [x9]\n" + "st1 { v12.b }[6], [x23]\n" + "st1 { v16.b }[6], [x22]\n" + "st1 { v20.b }[6], [x21]\n" + "st1 { v24.b }[6], [x20]\n" + "st1 { v28.b }[6], [x19]\n" + "b 160f\n" + "157:" // Height 6: Partial direct writeback: partial_1_4 + "tbz x10, #0, 160f\n" + "st1 { v8.b }[4], [x9]\n" + "st1 { v12.b }[4], [x23]\n" + "st1 { v16.b }[4], [x22]\n" + "st1 { v20.b }[4], [x21]\n" + "st1 { v24.b }[4], [x20]\n" + "st1 { v28.b }[4], [x19]\n" + "b 160f\n" + "158:" // Height 6: Partial direct writeback: partial_2_0 + "tbz x10, #1, 159f\n" + "str h8, [x9], #0x2\n" + "str h12, [x23], #0x2\n" + "str h16, [x22], #0x2\n" + "str h20, [x21], #0x2\n" + "str h24, [x20], #0x2\n" + "str h28, [x19], #0x2\n" + "tbz x10, #0, 160f\n" + "st1 { v8.b }[2], [x9]\n" + "st1 { v12.b }[2], [x23]\n" + "st1 { v16.b }[2], [x22]\n" + "st1 { v20.b }[2], [x21]\n" + "st1 { v24.b }[2], [x20]\n" + "st1 { v28.b }[2], [x19]\n" + "b 160f\n" + "159:" // Height 6: Partial direct writeback: partial_1_0 + "str b8, [x9, #0x0]\n" + "str b12, [x23, #0x0]\n" + "str b16, [x22, #0x0]\n" + "str b20, [x21, #0x0]\n" + "str b24, [x20, #0x0]\n" + "str b28, [x19, #0x0]\n" + "160:" // Height 6: Partial direct writeback: Done + "b 162f\n" + "161:" // Height 6: Full writeback + "str q8, [x9, #0x0]\n" "add x9, x9, #0x10\n" - "add x27, x27, #0x10\n" - "add x25, x25, #0x10\n" - "add x23, x23, #0x10\n" - "add x21, x21, #0x10\n" - "168:" // Height 6: Writeback done - "subs x15, x15, #0x10\n" - "bgt 143b\n" + "str q12, [x23, #0x0]\n" + "str q16, [x22, #0x0]\n" + "str q20, [x21, #0x0]\n" + "str q24, [x20, #0x0]\n" + "str q28, [x19, #0x0]\n" + "162:" // Height 6: Writeback done + "subs x10, x10, #0x10\n" + "bgt 137b\n" "subs %x[M], %x[M], #0x6\n" - "beq 170f\n" + "beq 164f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" - "tbz %x[flags], #3, 169f\n" + "tbz %x[flags], #3, 163f\n" "add x20, x20, #0x6\n" "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n" "b 1b\n" - "169:" // Update direct input + "163:" // Update direct input "mov x19, #0x6\n" "madd %x[input_ptr], x19, x20, %x[input_ptr]\n" "b 1b\n" - "170:" // Exit + "164:" // Exit : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, 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_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp) - : "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", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "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", "x12", "x13", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |