diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp | 1529 | ||||
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp | 403 |
2 files changed, 967 insertions, 965 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp index eac0e7167e..4d23660942 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/a55.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021, 2023 Arm Limited. + * Copyright (c) 2019-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,18 +45,18 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; + void *output_ptr = {}; } ka; unsigned long flags=0; - void *output_ptr; void *input_ptr; if (output_arg.is_indirect) { - output_ptr=(void *)(output_arg.indirect.ptr); + ka.output_ptr=(void *)(output_arg.indirect.ptr); ka.output_offset=output_arg.indirect.offset; flags |= 0x4; } else { - output_ptr=(void *)(output_arg.direct.base); + ka.output_ptr=(void *)(output_arg.direct.base); ka.output_offset=output_arg.direct.stride; } @@ -84,133 +84,133 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "cmp %x[M], #0x2\n" "bgt 61f\n" "beq 31f\n" - "mov x15, %x[col_bias]\n" + "mov x16, %x[col_bias]\n" "movi v11.4s, #0x0\n" "movi v15.16b, #0x1\n" "bic %x[flags], %x[flags], #0x80000000\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "mov x13, %x[output_ptr]\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" "2:" // Height 1: Column loop "movi v16.4s, #0x0\n" "movi v17.4s, #0x0\n" "movi v18.4s, #0x0\n" "movi v19.4s, #0x0\n" "3:" // Height 1: setup done - "mov x11, #0x0\n" + "mov x12, #0x0\n" "4:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" "tbz %x[flags], #3, 5f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "cbnz x11, 6f\n" + "ldr x10, [x20, #0x0]\n" + "cbnz x12, 6f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x9, x9, x20\n" + "add x10, x10, x20\n" "b 6f\n" "5:" // Height 1: setup direct input - "mov x9, %x[input_ptr]\n" + "mov x10, %x[input_ptr]\n" "6:" // Height 1: input setup done - "cmp x10, #0x10\n" + "cmp x11, #0x10\n" "blt 11f\n" - "ldr q0, [x9, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "ldr q8, [x12, #0x40]\n" - "ldr q9, [x12, #0x50]\n" - "ldr q10, [x12, #0x60]\n" + "ldr q0, [x10, #0x0]\n" + "cmp x11, #0x20\n" + "ldr q4, [x14, #0x0]\n" + "ldr q5, [x14, #0x10]\n" + "ldr q6, [x14, #0x20]\n" + "ldr q7, [x14, #0x30]\n" + "ldr q8, [x14, #0x40]\n" + "ldr q9, [x14, #0x50]\n" + "ldr q10, [x14, #0x60]\n" "blt 9f\n" "7:" // Height 1: Multiply loop: Main loop head ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - "ldr d21, [x12, #0x70]\n" - "ldr x20, [x12, #0x78]\n" + "ldr d21, [x14, #0x70]\n" + "ldr x20, [x14, #0x78]\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" - "ldr d20, [x12, #0x80]\n" + "ldr d20, [x14, #0x80]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" - "ldr d26, [x12, #0x90]\n" + "ldr d26, [x14, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" - "ldr d25, [x12, #0xa0]\n" + "ldr d25, [x14, #0xa0]\n" "mov v21.d[1], x20\n" - "ldr x20, [x12, #0x88]\n" + "ldr x20, [x14, #0x88]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" - "ldr d24, [x12, #0xb0]\n" + "ldr d24, [x14, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" - "ldr d23, [x12, #0xc0]\n" + "ldr d23, [x14, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" - "ldr d22, [x12, #0xd0]\n" + "ldr d22, [x14, #0xd0]\n" ".inst 0x4fa0e2b3 // sdot v19.4s, v21.16b, v0.4b[1]\n" - "ldr d21, [x12, #0xe0]\n" + "ldr d21, [x14, #0xe0]\n" "mov v20.d[1], x20\n" - "ldr x20, [x12, #0x98]\n" - "mov v26.d[1], x20\n" - "ldr x20, [x12, #0xa8]\n" - "mov v25.d[1], x20\n" - "ldr x20, [x12, #0xb8]\n" - "mov v24.d[1], x20\n" - "ldr x23, [x12, #0xc8]\n" + "ldr x22, [x14, #0x98]\n" + "add x10, x10, #0x10\n" + "ldr x21, [x14, #0xa8]\n" ".inst 0x4f80ea90 // sdot v16.4s, v20.16b, v0.4b[2]\n" - "ldr d20, [x12, #0xf0]\n" + "ldr d20, [x14, #0xf0]\n" + "ldr x20, [x14, #0xb8]\n" + "mov v26.d[1], x22\n" + "mov v25.d[1], x21\n" + "ldr x23, [x14, #0xc8]\n" + "ldr x22, [x14, #0xd8]\n" ".inst 0x4f80eb51 // sdot v17.4s, v26.16b, v0.4b[2]\n" - "ldr x22, [x12, #0xd8]\n" + "mov v24.d[1], x20\n" + "ldr x21, [x14, #0xe8]\n" + "ldr x20, [x14, #0xf8]\n" ".inst 0x4f80eb32 // sdot v18.4s, v25.16b, v0.4b[2]\n" - "ldr x21, [x12, #0xe8]\n" ".inst 0x4f80eb13 // sdot v19.4s, v24.16b, v0.4b[2]\n" - "ldr x20, [x12, #0xf8]\n" "mov v23.d[1], x23\n" "mov v22.d[1], x22\n" - "add x9, x9, #0x10\n" + "add x14, x14, #0x100\n" "mov v21.d[1], x21\n" - "add x12, x12, #0x100\n" - "mov v20.d[1], x20\n" ".inst 0x4fa0eaf0 // sdot v16.4s, v23.16b, v0.4b[3]\n" + "mov v20.d[1], x20\n" ".inst 0x4fa0ead1 // sdot v17.4s, v22.16b, v0.4b[3]\n" ".inst 0x4fa0eab2 // sdot v18.4s, v21.16b, v0.4b[3]\n" ".inst 0x4fa0ea93 // sdot v19.4s, v20.16b, v0.4b[3]\n" "tbnz %x[flags], #31, 8f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" "8:" // Height 1: Multiply loop: unique 1: skip row sum - "ldr q0, [x9, #0x0]\n" - "sub x10, x10, #0x10\n" - "ldr q4, [x12, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q5, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "ldr q8, [x12, #0x40]\n" - "ldr q9, [x12, #0x50]\n" - "ldr q10, [x12, #0x60]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "ldr q0, [x10, #0x0]\n" + "sub x11, x11, #0x10\n" + "ldr q4, [x14, #0x0]\n" + "cmp x11, #0x20\n" + "ldr q5, [x14, #0x10]\n" + "ldr q6, [x14, #0x20]\n" + "ldr q7, [x14, #0x30]\n" + "ldr q8, [x14, #0x40]\n" + "ldr q9, [x14, #0x50]\n" + "ldr q10, [x14, #0x60]\n" + "prfm pldl1keep, [x10, #0x80]\n" "bge 7b\n" "9:" // Height 1: Multiply loop: Single iteration only ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - "ldr q21, [x12, #0x70]\n" + "ldr q21, [x14, #0x70]\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" - "ldr q20, [x12, #0x80]\n" + "ldr q20, [x14, #0x80]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" - "ldr q26, [x12, #0x90]\n" + "ldr q26, [x14, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" - "ldr q25, [x12, #0xa0]\n" + "ldr q25, [x14, #0xa0]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" - "ldr q24, [x12, #0xb0]\n" + "ldr q24, [x14, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" - "ldr q23, [x12, #0xc0]\n" + "ldr q23, [x14, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" - "ldr q22, [x12, #0xd0]\n" + "ldr q22, [x14, #0xd0]\n" ".inst 0x4fa0e2b3 // sdot v19.4s, v21.16b, v0.4b[1]\n" - "ldr q21, [x12, #0xe0]\n" + "ldr q21, [x14, #0xe0]\n" ".inst 0x4f80ea90 // sdot v16.4s, v20.16b, v0.4b[2]\n" - "ldr q20, [x12, #0xf0]\n" + "ldr q20, [x14, #0xf0]\n" ".inst 0x4f80eb51 // sdot v17.4s, v26.16b, v0.4b[2]\n" - "sub x10, x10, #0x10\n" + "sub x11, x11, #0x10\n" ".inst 0x4f80eb32 // sdot v18.4s, v25.16b, v0.4b[2]\n" - "add x9, x9, #0x10\n" + "add x10, x10, #0x10\n" ".inst 0x4f80eb13 // sdot v19.4s, v24.16b, v0.4b[2]\n" - "add x12, x12, #0x100\n" + "add x14, x14, #0x100\n" ".inst 0x4fa0eaf0 // sdot v16.4s, v23.16b, v0.4b[3]\n" ".inst 0x4fa0ead1 // sdot v17.4s, v22.16b, v0.4b[3]\n" ".inst 0x4fa0eab2 // sdot v18.4s, v21.16b, v0.4b[3]\n" @@ -218,54 +218,54 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "tbnz %x[flags], #31, 10f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" "10:" // Height 1: Multiply loop: unique 2: skip row sum - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x10, #0x80]\n" "11:" // Height 1: Multiply loop: Main loop skip - "cbz x10, 18f\n" - "cmp x10, #0x4\n" + "cbz x11, 18f\n" + "cmp x11, #0x4\n" "blt 14f\n" "12:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" + "ldr s0, [x10], #0x4\n" "tbnz %x[flags], #31, 13f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" "13:" // Height 1: Multiply loop: unique 3: skip row sum - "ldr q20, [x12, #0x0]\n" - "sub x10, x10, #0x4\n" - "ldr q22, [x12, #0x10]\n" - "cmp x10, #0x4\n" - "ldr q21, [x12, #0x20]\n" - ".inst 0x4f80e290 // sdot v16.4s, v20.16b, v0.4b[0]\n" - "ldr q20, [x12, #0x30]\n" + "ldr q23, [x14, #0x0]\n" + "sub x11, x11, #0x4\n" + "ldr q22, [x14, #0x10]\n" + "cmp x11, #0x4\n" + "ldr q21, [x14, #0x20]\n" + "ldr q20, [x14, #0x30]\n" + ".inst 0x4f80e2f0 // sdot v16.4s, v23.16b, v0.4b[0]\n" ".inst 0x4f80e2d1 // sdot v17.4s, v22.16b, v0.4b[0]\n" + "add x14, x14, #0x40\n" ".inst 0x4f80e2b2 // sdot v18.4s, v21.16b, v0.4b[0]\n" - "add x12, x12, #0x40\n" ".inst 0x4f80e293 // sdot v19.4s, v20.16b, v0.4b[0]\n" "bge 12b\n" "14:" // Height 1: Multiply loop: Skip odd blocks - "cbz x10, 18f\n" - "tbz x10, #1, 15f\n" - "ldr h0, [x9], #0x2\n" - "tbz x10, #0, 16f\n" - "ld1 { v0.b }[2], [x9]\n" + "cbz x11, 18f\n" + "tbz x11, #1, 15f\n" + "ldr h0, [x10], #0x2\n" + "tbz x11, #0, 16f\n" + "ld1 { v0.b }[2], [x10]\n" "b 16f\n" "15:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" + "ldr b0, [x10, #0x0]\n" "16:" // Height 1: Multiply loop: Ragged operand read: Done "tbnz %x[flags], #31, 17f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" "17:" // Height 1: Multiply loop: unique 4: skip row sum - "ldr q20, [x12, #0x0]\n" - ".inst 0x4f80e290 // sdot v16.4s, v20.16b, v0.4b[0]\n" - "ldr q20, [x12, #0x10]\n" - ".inst 0x4f80e291 // sdot v17.4s, v20.16b, v0.4b[0]\n" - "ldr q20, [x12, #0x20]\n" - ".inst 0x4f80e292 // sdot v18.4s, v20.16b, v0.4b[0]\n" - "ldr q20, [x12, #0x30]\n" + "ldr q23, [x14, #0x0]\n" + "ldr q22, [x14, #0x10]\n" + "ldr q21, [x14, #0x20]\n" + "ldr q20, [x14, #0x30]\n" + ".inst 0x4f80e2f0 // sdot v16.4s, v23.16b, v0.4b[0]\n" + ".inst 0x4f80e2d1 // sdot v17.4s, v22.16b, v0.4b[0]\n" + "add x14, x14, #0x40\n" + ".inst 0x4f80e2b2 // sdot v18.4s, v21.16b, v0.4b[0]\n" ".inst 0x4f80e293 // sdot v19.4s, v20.16b, v0.4b[0]\n" - "add x12, x12, #0x40\n" "18:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x20\n" + "add x12, x12, #0x1\n" + "cmp x12, x20\n" "bne 4b\n" "prfm pstl1keep, [x13, #0x0]\n" "tbnz %x[flags], #31, 19f\n" @@ -276,28 +276,28 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "addp v11.4s, v11.4s, v11.4s\n" "mul v11.4s, v11.4s, v20.4s\n" "19:" // Height 1: skip row sum fixup - "ldr q23, [x15, #0x0]\n" + "ldr q24, [x16, #0x0]\n" "add v16.4s, v16.4s, v11.4s\n" - "ldr q22, [x15, #0x10]\n" + "ldr q23, [x16, #0x10]\n" "add v17.4s, v17.4s, v11.4s\n" - "ldr q21, [x15, #0x20]\n" + "ldr q22, [x16, #0x20]\n" "add v18.4s, v18.4s, v11.4s\n" - "ldr q20, [x15, #0x30]\n" + "ldr q21, [x16, #0x30]\n" "add v19.4s, v19.4s, v11.4s\n" - "add v16.4s, v16.4s, v23.4s\n" - "add v17.4s, v17.4s, v22.4s\n" - "add v18.4s, v18.4s, v21.4s\n" - "add v19.4s, v19.4s, v20.4s\n" "add x20, %x[qp], %[per_layer_mul]\n" - "ld1r { v20.4s }, [x20]\n" "orr %x[flags], %x[flags], #0x80000000\n" + "ld1r { v20.4s }, [x20]\n" + "add v16.4s, v16.4s, v24.4s\n" + "add v17.4s, v17.4s, v23.4s\n" "add x20, %x[qp], %[per_layer_right_shift]\n" "ld1r { v0.4s }, [x20]\n" + "add v18.4s, v18.4s, v22.4s\n" + "add v19.4s, v19.4s, v21.4s\n" + "add x16, x16, #0x40\n" "sqrdmulh v16.4s, v16.4s, v20.4s\n" "sqrdmulh v17.4s, v17.4s, v20.4s\n" "sqrdmulh v18.4s, v18.4s, v20.4s\n" "sqrdmulh v19.4s, v19.4s, v20.4s\n" - "add x15, x15, #0x40\n" "tbz %x[flags], #5, 20f\n" "and v23.16b, v16.16b, v0.16b\n" "and v22.16b, v17.16b, v0.16b\n" @@ -317,67 +317,67 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "srshl v18.4s, v18.4s, v0.4s\n" "srshl v19.4s, v19.4s, v0.4s\n" "add x20, %x[qp], %[c_offset]\n" - "ld1r { v20.4s }, [x20]\n" - "add v16.4s, v16.4s, v20.4s\n" - "add v17.4s, v17.4s, v20.4s\n" - "add v18.4s, v18.4s, v20.4s\n" - "add v19.4s, v19.4s, v20.4s\n" - "add x20, %x[qp], %[maxval]\n" - "ld1r { v20.4s }, [x20]\n" - "smin v16.4s, v16.4s, v20.4s\n" - "smin v17.4s, v17.4s, v20.4s\n" - "smin v18.4s, v18.4s, v20.4s\n" - "smin v19.4s, v19.4s, v20.4s\n" + "add x21, %x[qp], %[maxval]\n" + "ld1r { v22.4s }, [x20]\n" "add x20, %x[qp], %[minval]\n" + "ld1r { v21.4s }, [x21]\n" + "cmp x15, #0x10\n" "ld1r { v20.4s }, [x20]\n" + "add v16.4s, v16.4s, v22.4s\n" + "add v17.4s, v17.4s, v22.4s\n" + "add v18.4s, v18.4s, v22.4s\n" + "add v19.4s, v19.4s, v22.4s\n" + "smin v16.4s, v16.4s, v21.4s\n" + "smin v17.4s, v17.4s, v21.4s\n" + "smin v18.4s, v18.4s, v21.4s\n" + "smin v19.4s, v19.4s, v21.4s\n" "smax v16.4s, v16.4s, v20.4s\n" "smax v17.4s, v17.4s, v20.4s\n" "smax v18.4s, v18.4s, v20.4s\n" "smax v19.4s, v19.4s, v20.4s\n" "uzp1 v16.8h, v16.8h, v17.8h\n" "uzp1 v17.8h, v18.8h, v19.8h\n" - "cmp x14, #0x10\n" "uzp1 v16.16b, v16.16b, v17.16b\n" "bge 29f\n" - "tbz x14, #3, 24f\n" + "tbz x15, #3, 24f\n" "str d16, [x13], #0x8\n" - "tbz x14, #2, 22f\n" + "tbz x15, #2, 22f\n" "st1 { v16.s }[2], [x13], #0x4\n" - "tbz x14, #1, 21f\n" + "tbz x15, #1, 21f\n" "st1 { v16.h }[6], [x13], #0x2\n" - "tbz x14, #0, 28f\n" + "tbz x15, #0, 28f\n" "st1 { v16.b }[14], [x13]\n" "b 28f\n" "21:" // Height 1: Partial direct writeback: partial_1_12 - "tbz x14, #0, 28f\n" + "tbz x15, #0, 28f\n" "st1 { v16.b }[12], [x13]\n" "b 28f\n" "22:" // Height 1: Partial direct writeback: partial_2_8 - "tbz x14, #1, 23f\n" + "tbz x15, #1, 23f\n" "st1 { v16.h }[4], [x13], #0x2\n" - "tbz x14, #0, 28f\n" + "tbz x15, #0, 28f\n" "st1 { v16.b }[10], [x13]\n" "b 28f\n" "23:" // Height 1: Partial direct writeback: partial_1_8 - "tbz x14, #0, 28f\n" + "tbz x15, #0, 28f\n" "st1 { v16.b }[8], [x13]\n" "b 28f\n" "24:" // Height 1: Partial direct writeback: partial_4_0 - "tbz x14, #2, 26f\n" + "tbz x15, #2, 26f\n" "str s16, [x13], #0x4\n" - "tbz x14, #1, 25f\n" + "tbz x15, #1, 25f\n" "st1 { v16.h }[2], [x13], #0x2\n" - "tbz x14, #0, 28f\n" + "tbz x15, #0, 28f\n" "st1 { v16.b }[6], [x13]\n" "b 28f\n" "25:" // Height 1: Partial direct writeback: partial_1_4 - "tbz x14, #0, 28f\n" + "tbz x15, #0, 28f\n" "st1 { v16.b }[4], [x13]\n" "b 28f\n" "26:" // Height 1: Partial direct writeback: partial_2_0 - "tbz x14, #1, 27f\n" + "tbz x15, #1, 27f\n" "str h16, [x13], #0x2\n" - "tbz x14, #0, 28f\n" + "tbz x15, #0, 28f\n" "st1 { v16.b }[2], [x13]\n" "b 28f\n" "27:" // Height 1: Partial direct writeback: partial_1_0 @@ -388,18 +388,18 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "str q16, [x13, #0x0]\n" "add x13, x13, #0x10\n" "30:" // Height 1: Writeback done - "subs x14, x14, #0x10\n" + "subs x15, x15, #0x10\n" "bgt 2b\n" "b 122f\n" "31:" // Height 2 - "mov x15, %x[col_bias]\n" + "mov x16, %x[col_bias]\n" "movi v11.4s, #0x0\n" "movi v12.4s, #0x0\n" "bic %x[flags], %x[flags], #0x80000000\n" "movi v15.16b, #0x1\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x13, %x[output_ptr]\n" + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" "32:" // Height 2: Column loop "movi v16.4s, #0x0\n" "movi v17.4s, #0x0\n" @@ -410,80 +410,80 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "movi v22.4s, #0x0\n" "movi v23.4s, #0x0\n" "33:" // Height 2: setup done - "mov x11, #0x0\n" + "mov x12, #0x0\n" "34:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" "tbz %x[flags], #3, 35f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "cbnz x11, 36f\n" + "ldr x10, [x20, #0x0]\n" + "ldr x9, [x20, #0x8]\n" + "cbnz x12, 36f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x10, x10, x20\n" "add x9, x9, x20\n" - "add x28, x28, x20\n" "b 36f\n" "35:" // Height 2: setup direct input - "mov x9, %x[input_ptr]\n" - "add x28, x9, x21\n" + "mov x10, %x[input_ptr]\n" + "add x9, x10, x21\n" "36:" // Height 2: input setup done - "cmp x10, #0x10\n" + "cmp x11, #0x10\n" "blt 41f\n" - "ldr q0, [x9, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q1, [x28, #0x0]\n" - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "ldr q8, [x12, #0x40]\n" - "ldr q9, [x12, #0x50]\n" - "ldr q10, [x12, #0x60]\n" + "ldr q0, [x10, #0x0]\n" + "cmp x11, #0x20\n" + "ldr q1, [x9, #0x0]\n" + "ldr q4, [x14, #0x0]\n" + "ldr q5, [x14, #0x10]\n" + "ldr q6, [x14, #0x20]\n" + "ldr q7, [x14, #0x30]\n" + "ldr q8, [x14, #0x40]\n" + "ldr q9, [x14, #0x50]\n" + "ldr q10, [x14, #0x60]\n" "blt 39f\n" "37:" // Height 2: Multiply loop: Main loop head ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - "ldr x20, [x12, #0x78]\n" + "ldr x20, [x14, #0x78]\n" ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - "ldr d25, [x12, #0x70]\n" + "ldr d25, [x14, #0x70]\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" - "mov v25.d[1], x20\n" + "ldr x23, [x14, #0x88]\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" - "ldr d24, [x12, #0x80]\n" + "ldr d24, [x14, #0x80]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" - "ldr x23, [x12, #0x88]\n" + "mov v25.d[1], x20\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" - "ldr d30, [x12, #0x90]\n" + "ldr d30, [x14, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" - "ldr x22, [x12, #0x98]\n" + "ldr x22, [x14, #0x98]\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" - "ldr d29, [x12, #0xa0]\n" - "ldr x21, [x12, #0xa8]\n" + "ldr d29, [x14, #0xa0]\n" + "ldr x21, [x14, #0xa8]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" - "ldr d28, [x12, #0xb0]\n" - "ldr x20, [x12, #0xb8]\n" + "ldr d28, [x14, #0xb0]\n" + "ldr x20, [x14, #0xb8]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" - "ldr d27, [x12, #0xc0]\n" + "ldr d27, [x14, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" "mov v24.d[1], x23\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" - "ldr d26, [x12, #0xd0]\n" + "ldr d26, [x14, #0xd0]\n" ".inst 0x4fa0e333 // sdot v19.4s, v25.16b, v0.4b[1]\n" "mov v30.d[1], x22\n" ".inst 0x4fa1e337 // sdot v23.4s, v25.16b, v1.4b[1]\n" - "ldr d25, [x12, #0xe0]\n" + "ldr d25, [x14, #0xe0]\n" "mov v29.d[1], x21\n" - "ldr x23, [x12, #0xc8]\n" + "ldr x23, [x14, #0xc8]\n" "mov v28.d[1], x20\n" - "ldr x22, [x12, #0xd8]\n" - "ldr x21, [x12, #0xe8]\n" + "ldr x22, [x14, #0xd8]\n" + "ldr x21, [x14, #0xe8]\n" ".inst 0x4f80eb10 // sdot v16.4s, v24.16b, v0.4b[2]\n" ".inst 0x4f81eb14 // sdot v20.4s, v24.16b, v1.4b[2]\n" - "ldr d24, [x12, #0xf0]\n" - "ldr x20, [x12, #0xf8]\n" + "ldr d24, [x14, #0xf0]\n" + "ldr x20, [x14, #0xf8]\n" ".inst 0x4f80ebd1 // sdot v17.4s, v30.16b, v0.4b[2]\n" ".inst 0x4f81ebd5 // sdot v21.4s, v30.16b, v1.4b[2]\n" "mov v27.d[1], x23\n" @@ -494,9 +494,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f80eb93 // sdot v19.4s, v28.16b, v0.4b[2]\n" "mov v24.d[1], x20\n" ".inst 0x4f81eb97 // sdot v23.4s, v28.16b, v1.4b[2]\n" + "add x10, x10, #0x10\n" "add x9, x9, #0x10\n" - "add x28, x28, #0x10\n" - "add x12, x12, #0x100\n" + "add x14, x14, #0x100\n" ".inst 0x4fa0eb70 // sdot v16.4s, v27.16b, v0.4b[3]\n" ".inst 0x4fa1eb74 // sdot v20.4s, v27.16b, v1.4b[3]\n" ".inst 0x4fa0eb51 // sdot v17.4s, v26.16b, v0.4b[3]\n" @@ -509,53 +509,53 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" "38:" // Height 2: Multiply loop: unique 5: skip row sum - "ldr q0, [x9, #0x0]\n" - "sub x10, x10, #0x10\n" - "ldr q1, [x28, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "ldr q8, [x12, #0x40]\n" - "ldr q9, [x12, #0x50]\n" - "ldr q10, [x12, #0x60]\n" + "ldr q0, [x10, #0x0]\n" + "sub x11, x11, #0x10\n" + "ldr q1, [x9, #0x0]\n" + "cmp x11, #0x20\n" + "ldr q4, [x14, #0x0]\n" + "ldr q5, [x14, #0x10]\n" + "ldr q6, [x14, #0x20]\n" + "ldr q7, [x14, #0x30]\n" + "ldr q8, [x14, #0x40]\n" + "ldr q9, [x14, #0x50]\n" + "ldr q10, [x14, #0x60]\n" + "prfm pldl1keep, [x10, #0x80]\n" "prfm pldl1keep, [x9, #0x80]\n" - "prfm pldl1keep, [x28, #0x80]\n" "bge 37b\n" "39:" // Height 2: Multiply loop: Single iteration only ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - "sub x10, x10, #0x10\n" + "sub x11, x11, #0x10\n" ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - "ldr q25, [x12, #0x70]\n" + "ldr q25, [x14, #0x70]\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" - "add x9, x9, #0x10\n" + "add x10, x10, #0x10\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" - "ldr q24, [x12, #0x80]\n" + "ldr q24, [x14, #0x80]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" - "add x28, x28, #0x10\n" + "add x9, x9, #0x10\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" - "ldr q30, [x12, #0x90]\n" + "ldr q30, [x14, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" - "ldr q29, [x12, #0xa0]\n" + "ldr q29, [x14, #0xa0]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" - "ldr q28, [x12, #0xb0]\n" + "ldr q28, [x14, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" - "ldr q27, [x12, #0xc0]\n" + "ldr q27, [x14, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" - "ldr q26, [x12, #0xd0]\n" + "ldr q26, [x14, #0xd0]\n" ".inst 0x4fa0e333 // sdot v19.4s, v25.16b, v0.4b[1]\n" ".inst 0x4fa1e337 // sdot v23.4s, v25.16b, v1.4b[1]\n" - "ldr q25, [x12, #0xe0]\n" + "ldr q25, [x14, #0xe0]\n" ".inst 0x4f80eb10 // sdot v16.4s, v24.16b, v0.4b[2]\n" ".inst 0x4f81eb14 // sdot v20.4s, v24.16b, v1.4b[2]\n" - "ldr q24, [x12, #0xf0]\n" + "ldr q24, [x14, #0xf0]\n" ".inst 0x4f80ebd1 // sdot v17.4s, v30.16b, v0.4b[2]\n" - "add x12, x12, #0x100\n" + "add x14, x14, #0x100\n" ".inst 0x4f81ebd5 // sdot v21.4s, v30.16b, v1.4b[2]\n" ".inst 0x4f80ebb2 // sdot v18.4s, v29.16b, v0.4b[2]\n" ".inst 0x4f81ebb6 // sdot v22.4s, v29.16b, v1.4b[2]\n" @@ -573,29 +573,29 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" "40:" // Height 2: Multiply loop: unique 6: skip row sum + "prfm pldl1keep, [x10, #0x80]\n" "prfm pldl1keep, [x9, #0x80]\n" - "prfm pldl1keep, [x28, #0x80]\n" "41:" // Height 2: Multiply loop: Main loop skip - "cbz x10, 48f\n" - "cmp x10, #0x4\n" + "cbz x11, 48f\n" + "cmp x11, #0x4\n" "blt 44f\n" "42:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "ldr s1, [x28], #0x4\n" + "ldr s0, [x10], #0x4\n" + "ldr s1, [x9], #0x4\n" "tbnz %x[flags], #31, 43f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" "43:" // Height 2: Multiply loop: unique 7: skip row sum - "ldr q27, [x12, #0x0]\n" - "sub x10, x10, #0x4\n" - "ldr q26, [x12, #0x10]\n" - "cmp x10, #0x4\n" - "ldr q25, [x12, #0x20]\n" + "ldr q27, [x14, #0x0]\n" + "sub x11, x11, #0x4\n" + "ldr q26, [x14, #0x10]\n" + "cmp x11, #0x4\n" + "ldr q25, [x14, #0x20]\n" + "ldr q24, [x14, #0x30]\n" ".inst 0x4f80e370 // sdot v16.4s, v27.16b, v0.4b[0]\n" - "ldr q24, [x12, #0x30]\n" ".inst 0x4f81e374 // sdot v20.4s, v27.16b, v1.4b[0]\n" + "add x14, x14, #0x40\n" ".inst 0x4f80e351 // sdot v17.4s, v26.16b, v0.4b[0]\n" - "add x12, x12, #0x40\n" ".inst 0x4f81e355 // sdot v21.4s, v26.16b, v1.4b[0]\n" ".inst 0x4f80e332 // sdot v18.4s, v25.16b, v0.4b[0]\n" ".inst 0x4f81e336 // sdot v22.4s, v25.16b, v1.4b[0]\n" @@ -603,44 +603,44 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f81e317 // sdot v23.4s, v24.16b, v1.4b[0]\n" "bge 42b\n" "44:" // Height 2: Multiply loop: Skip odd blocks - "cbz x10, 48f\n" - "tbz x10, #1, 45f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x28], #0x2\n" - "tbz x10, #0, 46f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x28]\n" + "cbz x11, 48f\n" + "tbz x11, #1, 45f\n" + "ldr h0, [x10], #0x2\n" + "ldr h1, [x9], #0x2\n" + "tbz x11, #0, 46f\n" + "ld1 { v0.b }[2], [x10]\n" + "ld1 { v1.b }[2], [x9]\n" "b 46f\n" "45:" // Height 2: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x28, #0x0]\n" + "ldr b0, [x10, #0x0]\n" + "ldr b1, [x9, #0x0]\n" "46:" // Height 2: Multiply loop: Ragged operand read: Done "tbnz %x[flags], #31, 47f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" "47:" // Height 2: Multiply loop: unique 8: skip row sum - "ldr q24, [x12, #0x0]\n" - ".inst 0x4f80e310 // sdot v16.4s, v24.16b, v0.4b[0]\n" - "ldr q26, [x12, #0x10]\n" - ".inst 0x4f81e314 // sdot v20.4s, v24.16b, v1.4b[0]\n" - "ldr q25, [x12, #0x20]\n" + "ldr q27, [x14, #0x0]\n" + "ldr q26, [x14, #0x10]\n" + "ldr q25, [x14, #0x20]\n" + "ldr q24, [x14, #0x30]\n" + ".inst 0x4f80e370 // sdot v16.4s, v27.16b, v0.4b[0]\n" + ".inst 0x4f81e374 // sdot v20.4s, v27.16b, v1.4b[0]\n" + "add x14, x14, #0x40\n" ".inst 0x4f80e351 // sdot v17.4s, v26.16b, v0.4b[0]\n" - "ldr q24, [x12, #0x30]\n" ".inst 0x4f81e355 // sdot v21.4s, v26.16b, v1.4b[0]\n" ".inst 0x4f80e332 // sdot v18.4s, v25.16b, v0.4b[0]\n" - "add x12, x12, #0x40\n" ".inst 0x4f81e336 // sdot v22.4s, v25.16b, v1.4b[0]\n" ".inst 0x4f80e313 // sdot v19.4s, v24.16b, v0.4b[0]\n" ".inst 0x4f81e317 // sdot v23.4s, v24.16b, v1.4b[0]\n" "48:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x20\n" + "add x12, x12, #0x1\n" + "cmp x12, x20\n" "bne 34b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x13, x20\n" "prfm pstl1keep, [x13, #0x0]\n" - "prfm pstl1keep, [x23, #0x0]\n" + "add x24, x13, x20\n" + "prfm pstl1keep, [x24, #0x0]\n" "tbnz %x[flags], #31, 49f\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" @@ -652,28 +652,28 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "mul v11.4s, v11.4s, v24.4s\n" "mul v12.4s, v12.4s, v24.4s\n" "49:" // Height 2: skip row sum fixup - "ldr q27, [x15, #0x0]\n" + "ldr q28, [x16, #0x0]\n" "add v16.4s, v16.4s, v11.4s\n" - "ldr q26, [x15, #0x10]\n" + "ldr q27, [x16, #0x10]\n" "add v17.4s, v17.4s, v11.4s\n" - "ldr q25, [x15, #0x20]\n" + "ldr q26, [x16, #0x20]\n" "add v18.4s, v18.4s, v11.4s\n" - "ldr q24, [x15, #0x30]\n" + "ldr q25, [x16, #0x30]\n" "add v19.4s, v19.4s, v11.4s\n" "add v20.4s, v20.4s, v12.4s\n" "add v21.4s, v21.4s, v12.4s\n" "add v22.4s, v22.4s, v12.4s\n" "add v23.4s, v23.4s, v12.4s\n" - "add v16.4s, v16.4s, v27.4s\n" - "add v17.4s, v17.4s, v26.4s\n" - "add v18.4s, v18.4s, v25.4s\n" - "add v19.4s, v19.4s, v24.4s\n" - "add v20.4s, v20.4s, v27.4s\n" - "add v21.4s, v21.4s, v26.4s\n" - "add v22.4s, v22.4s, v25.4s\n" - "add v23.4s, v23.4s, v24.4s\n" + "add v16.4s, v16.4s, v28.4s\n" + "add v17.4s, v17.4s, v27.4s\n" + "add v18.4s, v18.4s, v26.4s\n" "add x20, %x[qp], %[per_layer_mul]\n" "ld1r { v24.4s }, [x20]\n" + "add v19.4s, v19.4s, v25.4s\n" + "add v20.4s, v20.4s, v28.4s\n" + "add v21.4s, v21.4s, v27.4s\n" + "add v22.4s, v22.4s, v26.4s\n" + "add v23.4s, v23.4s, v25.4s\n" "orr %x[flags], %x[flags], #0x80000000\n" "add x20, %x[qp], %[per_layer_right_shift]\n" "ld1r { v0.4s }, [x20]\n" @@ -685,31 +685,31 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "sqrdmulh v21.4s, v21.4s, v24.4s\n" "sqrdmulh v22.4s, v22.4s, v24.4s\n" "sqrdmulh v23.4s, v23.4s, v24.4s\n" - "add x15, x15, #0x40\n" + "add x16, x16, #0x40\n" "tbz %x[flags], #5, 50f\n" "and v24.16b, v16.16b, v0.16b\n" - "sshr v24.4s, v24.4s, #0x1f\n" - "sqadd v16.4s, v16.4s, v24.4s\n" "and v30.16b, v17.16b, v0.16b\n" "and v29.16b, v18.16b, v0.16b\n" "and v28.16b, v19.16b, v0.16b\n" "and v27.16b, v20.16b, v0.16b\n" "and v26.16b, v21.16b, v0.16b\n" "and v25.16b, v22.16b, v0.16b\n" - "and v24.16b, v23.16b, v0.16b\n" + "sshr v24.4s, v24.4s, #0x1f\n" "sshr v30.4s, v30.4s, #0x1f\n" "sshr v29.4s, v29.4s, #0x1f\n" "sshr v28.4s, v28.4s, #0x1f\n" "sshr v27.4s, v27.4s, #0x1f\n" "sshr v26.4s, v26.4s, #0x1f\n" "sshr v25.4s, v25.4s, #0x1f\n" - "sshr v24.4s, v24.4s, #0x1f\n" + "sqadd v16.4s, v16.4s, v24.4s\n" + "and v24.16b, v23.16b, v0.16b\n" "sqadd v17.4s, v17.4s, v30.4s\n" "sqadd v18.4s, v18.4s, v29.4s\n" "sqadd v19.4s, v19.4s, v28.4s\n" "sqadd v20.4s, v20.4s, v27.4s\n" "sqadd v21.4s, v21.4s, v26.4s\n" "sqadd v22.4s, v22.4s, v25.4s\n" + "sshr v24.4s, v24.4s, #0x1f\n" "sqadd v23.4s, v23.4s, v24.4s\n" "50:" // Height 2: no shift correction "srshl v16.4s, v16.4s, v0.4s\n" @@ -721,27 +721,28 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "srshl v22.4s, v22.4s, v0.4s\n" "srshl v23.4s, v23.4s, v0.4s\n" "add x20, %x[qp], %[c_offset]\n" - "ld1r { v24.4s }, [x20]\n" - "add v16.4s, v16.4s, v24.4s\n" - "add v17.4s, v17.4s, v24.4s\n" - "add v18.4s, v18.4s, v24.4s\n" - "add v19.4s, v19.4s, v24.4s\n" - "add v20.4s, v20.4s, v24.4s\n" - "add v21.4s, v21.4s, v24.4s\n" - "add v22.4s, v22.4s, v24.4s\n" - "add v23.4s, v23.4s, v24.4s\n" - "add x20, %x[qp], %[maxval]\n" - "ld1r { v24.4s }, [x20]\n" - "smin v16.4s, v16.4s, v24.4s\n" - "smin v17.4s, v17.4s, v24.4s\n" - "smin v18.4s, v18.4s, v24.4s\n" - "smin v19.4s, v19.4s, v24.4s\n" - "smin v20.4s, v20.4s, v24.4s\n" - "smin v21.4s, v21.4s, v24.4s\n" - "smin v22.4s, v22.4s, v24.4s\n" - "smin v23.4s, v23.4s, v24.4s\n" + "add x21, %x[qp], %[maxval]\n" + "ld1r { v26.4s }, [x20]\n" "add x20, %x[qp], %[minval]\n" + "ld1r { v25.4s }, [x21]\n" + "cmp x15, #0x10\n" "ld1r { v24.4s }, [x20]\n" + "add v16.4s, v16.4s, v26.4s\n" + "add v17.4s, v17.4s, v26.4s\n" + "add v18.4s, v18.4s, v26.4s\n" + "add v19.4s, v19.4s, v26.4s\n" + "add v20.4s, v20.4s, v26.4s\n" + "add v21.4s, v21.4s, v26.4s\n" + "add v22.4s, v22.4s, v26.4s\n" + "add v23.4s, v23.4s, v26.4s\n" + "smin v16.4s, v16.4s, v25.4s\n" + "smin v17.4s, v17.4s, v25.4s\n" + "smin v18.4s, v18.4s, v25.4s\n" + "smin v19.4s, v19.4s, v25.4s\n" + "smin v20.4s, v20.4s, v25.4s\n" + "smin v21.4s, v21.4s, v25.4s\n" + "smin v22.4s, v22.4s, v25.4s\n" + "smin v23.4s, v23.4s, v25.4s\n" "smax v16.4s, v16.4s, v24.4s\n" "smax v17.4s, v17.4s, v24.4s\n" "smax v18.4s, v18.4s, v24.4s\n" @@ -754,88 +755,87 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "uzp1 v18.8h, v18.8h, v19.8h\n" "uzp1 v20.8h, v20.8h, v21.8h\n" "uzp1 v17.8h, v22.8h, v23.8h\n" - "cmp x14, #0x10\n" "uzp1 v16.16b, v16.16b, v18.16b\n" "uzp1 v20.16b, v20.16b, v17.16b\n" "bge 59f\n" - "tbz x14, #3, 54f\n" + "tbz x15, #3, 54f\n" "str d16, [x13], #0x8\n" - "str d20, [x23], #0x8\n" - "tbz x14, #2, 52f\n" + "str d20, [x24], #0x8\n" + "tbz x15, #2, 52f\n" "st1 { v16.s }[2], [x13], #0x4\n" - "st1 { v20.s }[2], [x23], #0x4\n" - "tbz x14, #1, 51f\n" + "st1 { v20.s }[2], [x24], #0x4\n" + "tbz x15, #1, 51f\n" "st1 { v16.h }[6], [x13], #0x2\n" - "st1 { v20.h }[6], [x23], #0x2\n" - "tbz x14, #0, 58f\n" + "st1 { v20.h }[6], [x24], #0x2\n" + "tbz x15, #0, 58f\n" "st1 { v16.b }[14], [x13]\n" - "st1 { v20.b }[14], [x23]\n" + "st1 { v20.b }[14], [x24]\n" "b 58f\n" "51:" // Height 2: Partial direct writeback: partial_1_12 - "tbz x14, #0, 58f\n" + "tbz x15, #0, 58f\n" "st1 { v16.b }[12], [x13]\n" - "st1 { v20.b }[12], [x23]\n" + "st1 { v20.b }[12], [x24]\n" "b 58f\n" "52:" // Height 2: Partial direct writeback: partial_2_8 - "tbz x14, #1, 53f\n" + "tbz x15, #1, 53f\n" "st1 { v16.h }[4], [x13], #0x2\n" - "st1 { v20.h }[4], [x23], #0x2\n" - "tbz x14, #0, 58f\n" + "st1 { v20.h }[4], [x24], #0x2\n" + "tbz x15, #0, 58f\n" "st1 { v16.b }[10], [x13]\n" - "st1 { v20.b }[10], [x23]\n" + "st1 { v20.b }[10], [x24]\n" "b 58f\n" "53:" // Height 2: Partial direct writeback: partial_1_8 - "tbz x14, #0, 58f\n" + "tbz x15, #0, 58f\n" "st1 { v16.b }[8], [x13]\n" - "st1 { v20.b }[8], [x23]\n" + "st1 { v20.b }[8], [x24]\n" "b 58f\n" "54:" // Height 2: Partial direct writeback: partial_4_0 - "tbz x14, #2, 56f\n" + "tbz x15, #2, 56f\n" "str s16, [x13], #0x4\n" - "str s20, [x23], #0x4\n" - "tbz x14, #1, 55f\n" + "str s20, [x24], #0x4\n" + "tbz x15, #1, 55f\n" "st1 { v16.h }[2], [x13], #0x2\n" - "st1 { v20.h }[2], [x23], #0x2\n" - "tbz x14, #0, 58f\n" + "st1 { v20.h }[2], [x24], #0x2\n" + "tbz x15, #0, 58f\n" "st1 { v16.b }[6], [x13]\n" - "st1 { v20.b }[6], [x23]\n" + "st1 { v20.b }[6], [x24]\n" "b 58f\n" "55:" // Height 2: Partial direct writeback: partial_1_4 - "tbz x14, #0, 58f\n" + "tbz x15, #0, 58f\n" "st1 { v16.b }[4], [x13]\n" - "st1 { v20.b }[4], [x23]\n" + "st1 { v20.b }[4], [x24]\n" "b 58f\n" "56:" // Height 2: Partial direct writeback: partial_2_0 - "tbz x14, #1, 57f\n" + "tbz x15, #1, 57f\n" "str h16, [x13], #0x2\n" - "str h20, [x23], #0x2\n" - "tbz x14, #0, 58f\n" + "str h20, [x24], #0x2\n" + "tbz x15, #0, 58f\n" "st1 { v16.b }[2], [x13]\n" - "st1 { v20.b }[2], [x23]\n" + "st1 { v20.b }[2], [x24]\n" "b 58f\n" "57:" // Height 2: Partial direct writeback: partial_1_0 "str b16, [x13, #0x0]\n" - "str b20, [x23, #0x0]\n" + "str b20, [x24, #0x0]\n" "58:" // Height 2: Partial direct writeback: Done "b 60f\n" "59:" // Height 2: Full writeback "str q16, [x13, #0x0]\n" "add x13, x13, #0x10\n" - "str q20, [x23, #0x0]\n" + "str q20, [x24, #0x0]\n" "60:" // Height 2: Writeback done - "subs x14, x14, #0x10\n" + "subs x15, x15, #0x10\n" "bgt 32b\n" "b 122f\n" "61:" // Height 3 - "mov x15, %x[col_bias]\n" + "mov x16, %x[col_bias]\n" "movi v11.4s, #0x0\n" "movi v12.4s, #0x0\n" "bic %x[flags], %x[flags], #0x80000000\n" "movi v13.4s, #0x0\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" "movi v15.16b, #0x1\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x13, %x[output_ptr]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" "62:" // Height 3: Column loop "movi v16.4s, #0x0\n" "movi v17.4s, #0x0\n" @@ -850,105 +850,105 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "movi v26.4s, #0x0\n" "movi v27.4s, #0x0\n" "63:" // Height 3: setup done - "mov x11, #0x0\n" + "mov x12, #0x0\n" "64:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" "tbz %x[flags], #3, 65f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x27, [x20, #0x10]\n" - "cbnz x11, 66f\n" + "ldr x10, [x20, #0x0]\n" + "ldr x9, [x20, #0x8]\n" + "ldr x28, [x20, #0x10]\n" + "cbnz x12, 66f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x10, x10, x20\n" "add x9, x9, x20\n" "add x28, x28, x20\n" - "add x27, x27, x20\n" "b 66f\n" "65:" // Height 3: setup direct input - "mov x9, %x[input_ptr]\n" + "mov x10, %x[input_ptr]\n" + "add x9, x10, x21\n" "add x28, x9, x21\n" - "add x27, x28, x21\n" "66:" // Height 3: input setup done - "cmp x10, #0x10\n" + "cmp x11, #0x10\n" "blt 71f\n" - "ldr q0, [x9, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x27, #0x0]\n" - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "ldr q8, [x12, #0x40]\n" - "ldr q9, [x12, #0x50]\n" - "ldr q10, [x12, #0x60]\n" + "ldr q0, [x10, #0x0]\n" + "cmp x11, #0x20\n" + "ldr q1, [x9, #0x0]\n" + "ldr q2, [x28, #0x0]\n" + "ldr q4, [x14, #0x0]\n" + "ldr q5, [x14, #0x10]\n" + "ldr q6, [x14, #0x20]\n" + "ldr q7, [x14, #0x30]\n" + "ldr q8, [x14, #0x40]\n" + "ldr q9, [x14, #0x50]\n" + "ldr q10, [x14, #0x60]\n" "blt 69f\n" "67:" // Height 3: Multiply loop: Main loop head ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - "ldr x20, [x12, #0x78]\n" + "ldr x20, [x14, #0x78]\n" ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - "ldr x23, [x12, #0x88]\n" + "ldr x23, [x14, #0x88]\n" ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" - "ldr d29, [x12, #0x70]\n" + "ldr d29, [x14, #0x70]\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" - "mov v29.d[1], x20\n" + "ldr x22, [x14, #0x98]\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" - "ldr x22, [x12, #0x98]\n" + "ldr x21, [x14, #0xa8]\n" ".inst 0x4f82e0b9 // sdot v25.4s, v5.16b, v2.4b[0]\n" - "ldr d28, [x12, #0x80]\n" + "ldr d28, [x14, #0x80]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" - "ldr x21, [x12, #0xa8]\n" + "mov v29.d[1], x20\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" - "ldr x20, [x12, #0xb8]\n" + "ldr x20, [x14, #0xb8]\n" ".inst 0x4f82e0da // sdot v26.4s, v6.16b, v2.4b[0]\n" - "ldr d5, [x12, #0x90]\n" + "ldr d5, [x14, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" "mov v28.d[1], x23\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" - "mov v5.d[1], x22\n" + "ldr x23, [x14, #0xc8]\n" ".inst 0x4f82e0fb // sdot v27.4s, v7.16b, v2.4b[0]\n" - "ldr d4, [x12, #0xa0]\n" + "ldr d4, [x14, #0xa0]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" - "mov v4.d[1], x21\n" + "mov v5.d[1], x22\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" - "ldr x23, [x12, #0xc8]\n" + "ldr x22, [x14, #0xd8]\n" ".inst 0x4fa2e118 // sdot v24.4s, v8.16b, v2.4b[1]\n" - "ldr d3, [x12, #0xb0]\n" + "ldr d3, [x14, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" - "mov v3.d[1], x20\n" + "mov v4.d[1], x21\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" - "ldr x22, [x12, #0xd8]\n" + "ldr x21, [x14, #0xe8]\n" ".inst 0x4fa2e139 // sdot v25.4s, v9.16b, v2.4b[1]\n" - "ldr d31, [x12, #0xc0]\n" + "ldr d31, [x14, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" - "ldr x21, [x12, #0xe8]\n" + "mov v3.d[1], x20\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" - "ldr x20, [x12, #0xf8]\n" + "ldr x20, [x14, #0xf8]\n" ".inst 0x4fa2e15a // sdot v26.4s, v10.16b, v2.4b[1]\n" - "ldr d30, [x12, #0xd0]\n" + "ldr d30, [x14, #0xd0]\n" ".inst 0x4fa0e3b3 // sdot v19.4s, v29.16b, v0.4b[1]\n" "mov v31.d[1], x23\n" ".inst 0x4fa1e3b7 // sdot v23.4s, v29.16b, v1.4b[1]\n" - "mov v30.d[1], x22\n" + "add x10, x10, #0x10\n" ".inst 0x4fa2e3bb // sdot v27.4s, v29.16b, v2.4b[1]\n" - "ldr d29, [x12, #0xe0]\n" + "ldr d29, [x14, #0xe0]\n" ".inst 0x4f80eb90 // sdot v16.4s, v28.16b, v0.4b[2]\n" - "mov v29.d[1], x21\n" + "mov v30.d[1], x22\n" ".inst 0x4f81eb94 // sdot v20.4s, v28.16b, v1.4b[2]\n" "add x9, x9, #0x10\n" ".inst 0x4f82eb98 // sdot v24.4s, v28.16b, v2.4b[2]\n" - "ldr d28, [x12, #0xf0]\n" + "ldr d28, [x14, #0xf0]\n" ".inst 0x4f80e8b1 // sdot v17.4s, v5.16b, v0.4b[2]\n" - "mov v28.d[1], x20\n" + "mov v29.d[1], x21\n" ".inst 0x4f81e8b5 // sdot v21.4s, v5.16b, v1.4b[2]\n" "add x28, x28, #0x10\n" ".inst 0x4f82e8b9 // sdot v25.4s, v5.16b, v2.4b[2]\n" - "add x27, x27, #0x10\n" + "mov v28.d[1], x20\n" ".inst 0x4f80e892 // sdot v18.4s, v4.16b, v0.4b[2]\n" - "add x12, x12, #0x100\n" + "add x14, x14, #0x100\n" ".inst 0x4f81e896 // sdot v22.4s, v4.16b, v1.4b[2]\n" ".inst 0x4f82e89a // sdot v26.4s, v4.16b, v2.4b[2]\n" ".inst 0x4f80e873 // sdot v19.4s, v3.16b, v0.4b[2]\n" @@ -971,65 +971,65 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n" "68:" // Height 3: Multiply loop: unique 9: skip row sum - "ldr q0, [x9, #0x0]\n" - "sub x10, x10, #0x10\n" - "ldr q1, [x28, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x27, #0x0]\n" - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "ldr q8, [x12, #0x40]\n" - "ldr q9, [x12, #0x50]\n" - "ldr q10, [x12, #0x60]\n" + "ldr q0, [x10, #0x0]\n" + "sub x11, x11, #0x10\n" + "ldr q1, [x9, #0x0]\n" + "cmp x11, #0x20\n" + "ldr q2, [x28, #0x0]\n" + "ldr q4, [x14, #0x0]\n" + "ldr q5, [x14, #0x10]\n" + "ldr q6, [x14, #0x20]\n" + "ldr q7, [x14, #0x30]\n" + "ldr q8, [x14, #0x40]\n" + "ldr q9, [x14, #0x50]\n" + "ldr q10, [x14, #0x60]\n" + "prfm pldl1keep, [x10, #0x80]\n" "prfm pldl1keep, [x9, #0x80]\n" "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x27, #0x80]\n" "bge 67b\n" "69:" // Height 3: Multiply loop: Single iteration only ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - "sub x10, x10, #0x10\n" + "sub x11, x11, #0x10\n" ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - "add x9, x9, #0x10\n" + "add x10, x10, #0x10\n" ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" - "ldr q29, [x12, #0x70]\n" + "ldr q29, [x14, #0x70]\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" - "add x28, x28, #0x10\n" + "add x9, x9, #0x10\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" - "add x27, x27, #0x10\n" + "add x28, x28, #0x10\n" ".inst 0x4f82e0b9 // sdot v25.4s, v5.16b, v2.4b[0]\n" - "ldr q28, [x12, #0x80]\n" + "ldr q28, [x14, #0x80]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0da // sdot v26.4s, v6.16b, v2.4b[0]\n" - "ldr q5, [x12, #0x90]\n" + "ldr q5, [x14, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0fb // sdot v27.4s, v7.16b, v2.4b[0]\n" - "ldr q4, [x12, #0xa0]\n" + "ldr q4, [x14, #0xa0]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" ".inst 0x4fa2e118 // sdot v24.4s, v8.16b, v2.4b[1]\n" - "ldr q3, [x12, #0xb0]\n" + "ldr q3, [x14, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" ".inst 0x4fa2e139 // sdot v25.4s, v9.16b, v2.4b[1]\n" - "ldr q31, [x12, #0xc0]\n" + "ldr q31, [x14, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" ".inst 0x4fa2e15a // sdot v26.4s, v10.16b, v2.4b[1]\n" - "ldr q30, [x12, #0xd0]\n" + "ldr q30, [x14, #0xd0]\n" ".inst 0x4fa0e3b3 // sdot v19.4s, v29.16b, v0.4b[1]\n" ".inst 0x4fa1e3b7 // sdot v23.4s, v29.16b, v1.4b[1]\n" ".inst 0x4fa2e3bb // sdot v27.4s, v29.16b, v2.4b[1]\n" - "ldr q29, [x12, #0xe0]\n" + "ldr q29, [x14, #0xe0]\n" ".inst 0x4f80eb90 // sdot v16.4s, v28.16b, v0.4b[2]\n" ".inst 0x4f81eb94 // sdot v20.4s, v28.16b, v1.4b[2]\n" ".inst 0x4f82eb98 // sdot v24.4s, v28.16b, v2.4b[2]\n" - "ldr q28, [x12, #0xf0]\n" + "ldr q28, [x14, #0xf0]\n" ".inst 0x4f80e8b1 // sdot v17.4s, v5.16b, v0.4b[2]\n" - "add x12, x12, #0x100\n" + "add x14, x14, #0x100\n" ".inst 0x4f81e8b5 // sdot v21.4s, v5.16b, v1.4b[2]\n" ".inst 0x4f82e8b9 // sdot v25.4s, v5.16b, v2.4b[2]\n" ".inst 0x4f80e892 // sdot v18.4s, v4.16b, v0.4b[2]\n" @@ -1055,32 +1055,32 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n" "70:" // Height 3: Multiply loop: unique 10: skip row sum + "prfm pldl1keep, [x10, #0x80]\n" "prfm pldl1keep, [x9, #0x80]\n" "prfm pldl1keep, [x28, #0x80]\n" - "prfm pldl1keep, [x27, #0x80]\n" "71:" // Height 3: Multiply loop: Main loop skip - "cbz x10, 78f\n" - "cmp x10, #0x4\n" + "cbz x11, 78f\n" + "cmp x11, #0x4\n" "blt 74f\n" "72:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "ldr s1, [x28], #0x4\n" - "ldr s2, [x27], #0x4\n" + "ldr s0, [x10], #0x4\n" + "ldr s1, [x9], #0x4\n" + "ldr s2, [x28], #0x4\n" "tbnz %x[flags], #31, 73f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n" "73:" // Height 3: Multiply loop: unique 11: skip row sum - "ldr q31, [x12, #0x0]\n" - "sub x10, x10, #0x4\n" - "ldr q30, [x12, #0x10]\n" - "cmp x10, #0x4\n" - "ldr q29, [x12, #0x20]\n" + "ldr q31, [x14, #0x0]\n" + "sub x11, x11, #0x4\n" + "ldr q30, [x14, #0x10]\n" + "cmp x11, #0x4\n" + "ldr q29, [x14, #0x20]\n" + "ldr q28, [x14, #0x30]\n" ".inst 0x4f80e3f0 // sdot v16.4s, v31.16b, v0.4b[0]\n" - "ldr q28, [x12, #0x30]\n" ".inst 0x4f81e3f4 // sdot v20.4s, v31.16b, v1.4b[0]\n" + "add x14, x14, #0x40\n" ".inst 0x4f82e3f8 // sdot v24.4s, v31.16b, v2.4b[0]\n" - "add x12, x12, #0x40\n" ".inst 0x4f80e3d1 // sdot v17.4s, v30.16b, v0.4b[0]\n" ".inst 0x4f81e3d5 // sdot v21.4s, v30.16b, v1.4b[0]\n" ".inst 0x4f82e3d9 // sdot v25.4s, v30.16b, v2.4b[0]\n" @@ -1092,36 +1092,36 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f82e39b // sdot v27.4s, v28.16b, v2.4b[0]\n" "bge 72b\n" "74:" // Height 3: Multiply loop: Skip odd blocks - "cbz x10, 78f\n" - "tbz x10, #1, 75f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x28], #0x2\n" - "ldr h2, [x27], #0x2\n" - "tbz x10, #0, 76f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x28]\n" - "ld1 { v2.b }[2], [x27]\n" + "cbz x11, 78f\n" + "tbz x11, #1, 75f\n" + "ldr h0, [x10], #0x2\n" + "ldr h1, [x9], #0x2\n" + "ldr h2, [x28], #0x2\n" + "tbz x11, #0, 76f\n" + "ld1 { v0.b }[2], [x10]\n" + "ld1 { v1.b }[2], [x9]\n" + "ld1 { v2.b }[2], [x28]\n" "b 76f\n" "75:" // Height 3: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x28, #0x0]\n" - "ldr b2, [x27, #0x0]\n" + "ldr b0, [x10, #0x0]\n" + "ldr b1, [x9, #0x0]\n" + "ldr b2, [x28, #0x0]\n" "76:" // Height 3: Multiply loop: Ragged operand read: Done "tbnz %x[flags], #31, 77f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n" "77:" // Height 3: Multiply loop: unique 12: skip row sum - "ldr q28, [x12, #0x0]\n" - ".inst 0x4f80e390 // sdot v16.4s, v28.16b, v0.4b[0]\n" - "ldr q30, [x12, #0x10]\n" - ".inst 0x4f81e394 // sdot v20.4s, v28.16b, v1.4b[0]\n" - "ldr q29, [x12, #0x20]\n" - ".inst 0x4f82e398 // sdot v24.4s, v28.16b, v2.4b[0]\n" - "ldr q28, [x12, #0x30]\n" + "ldr q31, [x14, #0x0]\n" + "ldr q30, [x14, #0x10]\n" + "ldr q29, [x14, #0x20]\n" + "ldr q28, [x14, #0x30]\n" + ".inst 0x4f80e3f0 // sdot v16.4s, v31.16b, v0.4b[0]\n" + ".inst 0x4f81e3f4 // sdot v20.4s, v31.16b, v1.4b[0]\n" + "add x14, x14, #0x40\n" + ".inst 0x4f82e3f8 // sdot v24.4s, v31.16b, v2.4b[0]\n" ".inst 0x4f80e3d1 // sdot v17.4s, v30.16b, v0.4b[0]\n" ".inst 0x4f81e3d5 // sdot v21.4s, v30.16b, v1.4b[0]\n" - "add x12, x12, #0x40\n" ".inst 0x4f82e3d9 // sdot v25.4s, v30.16b, v2.4b[0]\n" ".inst 0x4f80e3b2 // sdot v18.4s, v29.16b, v0.4b[0]\n" ".inst 0x4f81e3b6 // sdot v22.4s, v29.16b, v1.4b[0]\n" @@ -1131,15 +1131,15 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f82e39b // sdot v27.4s, v28.16b, v2.4b[0]\n" "78:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x20\n" + "add x12, x12, #0x1\n" + "cmp x12, x20\n" "bne 64b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x13, x20\n" - "add x22, x23, x20\n" "prfm pstl1keep, [x13, #0x0]\n" + "add x24, x13, x20\n" + "add x23, x24, x20\n" + "prfm pstl1keep, [x24, #0x0]\n" "prfm pstl1keep, [x23, #0x0]\n" - "prfm pstl1keep, [x22, #0x0]\n" "tbnz %x[flags], #31, 79f\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" @@ -1154,13 +1154,13 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "mul v12.4s, v12.4s, v28.4s\n" "mul v13.4s, v13.4s, v28.4s\n" "79:" // Height 3: skip row sum fixup - "ldr q31, [x15, #0x0]\n" + "ldr q31, [x16, #0x0]\n" "add v16.4s, v16.4s, v11.4s\n" - "ldr q30, [x15, #0x10]\n" + "ldr q30, [x16, #0x10]\n" "add v17.4s, v17.4s, v11.4s\n" - "ldr q29, [x15, #0x20]\n" + "ldr q29, [x16, #0x20]\n" "add v18.4s, v18.4s, v11.4s\n" - "ldr q28, [x15, #0x30]\n" + "ldr q28, [x16, #0x30]\n" "add v19.4s, v19.4s, v11.4s\n" "add v20.4s, v20.4s, v12.4s\n" "add v21.4s, v21.4s, v12.4s\n" @@ -1183,10 +1183,11 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "add v26.4s, v26.4s, v29.4s\n" "add v27.4s, v27.4s, v28.4s\n" "add x20, %x[qp], %[per_layer_mul]\n" - "ld1r { v28.4s }, [x20]\n" "orr %x[flags], %x[flags], #0x80000000\n" + "ld1r { v28.4s }, [x20]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" "ld1r { v0.4s }, [x20]\n" + "add x16, x16, #0x40\n" "sqrdmulh v16.4s, v16.4s, v28.4s\n" "sqrdmulh v17.4s, v17.4s, v28.4s\n" "sqrdmulh v18.4s, v18.4s, v28.4s\n" @@ -1199,39 +1200,38 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "sqrdmulh v25.4s, v25.4s, v28.4s\n" "sqrdmulh v26.4s, v26.4s, v28.4s\n" "sqrdmulh v27.4s, v27.4s, v28.4s\n" - "add x15, x15, #0x40\n" "tbz %x[flags], #5, 80f\n" "and v1.16b, v16.16b, v0.16b\n" "and v31.16b, v17.16b, v0.16b\n" "and v30.16b, v18.16b, v0.16b\n" "and v29.16b, v19.16b, v0.16b\n" "and v28.16b, v20.16b, v0.16b\n" + "and v3.16b, v21.16b, v0.16b\n" + "and v2.16b, v22.16b, v0.16b\n" "sshr v1.4s, v1.4s, #0x1f\n" "sshr v31.4s, v31.4s, #0x1f\n" "sshr v30.4s, v30.4s, #0x1f\n" "sshr v29.4s, v29.4s, #0x1f\n" "sshr v28.4s, v28.4s, #0x1f\n" + "sshr v3.4s, v3.4s, #0x1f\n" + "sshr v2.4s, v2.4s, #0x1f\n" "sqadd v16.4s, v16.4s, v1.4s\n" "sqadd v17.4s, v17.4s, v31.4s\n" "sqadd v18.4s, v18.4s, v30.4s\n" "sqadd v19.4s, v19.4s, v29.4s\n" "sqadd v20.4s, v20.4s, v28.4s\n" - "and v3.16b, v21.16b, v0.16b\n" - "and v2.16b, v22.16b, v0.16b\n" "and v1.16b, v23.16b, v0.16b\n" "and v31.16b, v24.16b, v0.16b\n" "and v30.16b, v25.16b, v0.16b\n" "and v29.16b, v26.16b, v0.16b\n" "and v28.16b, v27.16b, v0.16b\n" - "sshr v3.4s, v3.4s, #0x1f\n" - "sshr v2.4s, v2.4s, #0x1f\n" + "sqadd v21.4s, v21.4s, v3.4s\n" + "sqadd v22.4s, v22.4s, v2.4s\n" "sshr v1.4s, v1.4s, #0x1f\n" "sshr v31.4s, v31.4s, #0x1f\n" "sshr v30.4s, v30.4s, #0x1f\n" "sshr v29.4s, v29.4s, #0x1f\n" "sshr v28.4s, v28.4s, #0x1f\n" - "sqadd v21.4s, v21.4s, v3.4s\n" - "sqadd v22.4s, v22.4s, v2.4s\n" "sqadd v23.4s, v23.4s, v1.4s\n" "sqadd v24.4s, v24.4s, v31.4s\n" "sqadd v25.4s, v25.4s, v30.4s\n" @@ -1251,35 +1251,36 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "srshl v26.4s, v26.4s, v0.4s\n" "srshl v27.4s, v27.4s, v0.4s\n" "add x20, %x[qp], %[c_offset]\n" - "ld1r { v28.4s }, [x20]\n" - "add v16.4s, v16.4s, v28.4s\n" - "add v17.4s, v17.4s, v28.4s\n" - "add v18.4s, v18.4s, v28.4s\n" - "add v19.4s, v19.4s, v28.4s\n" - "add v20.4s, v20.4s, v28.4s\n" - "add v21.4s, v21.4s, v28.4s\n" - "add v22.4s, v22.4s, v28.4s\n" - "add v23.4s, v23.4s, v28.4s\n" - "add v24.4s, v24.4s, v28.4s\n" - "add v25.4s, v25.4s, v28.4s\n" - "add v26.4s, v26.4s, v28.4s\n" - "add v27.4s, v27.4s, v28.4s\n" - "add x20, %x[qp], %[maxval]\n" - "ld1r { v28.4s }, [x20]\n" - "smin v16.4s, v16.4s, v28.4s\n" - "smin v17.4s, v17.4s, v28.4s\n" - "smin v18.4s, v18.4s, v28.4s\n" - "smin v19.4s, v19.4s, v28.4s\n" - "smin v20.4s, v20.4s, v28.4s\n" - "smin v21.4s, v21.4s, v28.4s\n" - "smin v22.4s, v22.4s, v28.4s\n" - "smin v23.4s, v23.4s, v28.4s\n" - "smin v24.4s, v24.4s, v28.4s\n" - "smin v25.4s, v25.4s, v28.4s\n" - "smin v26.4s, v26.4s, v28.4s\n" - "smin v27.4s, v27.4s, v28.4s\n" + "add x21, %x[qp], %[maxval]\n" + "ld1r { v30.4s }, [x20]\n" "add x20, %x[qp], %[minval]\n" + "ld1r { v29.4s }, [x21]\n" + "cmp x15, #0x10\n" "ld1r { v28.4s }, [x20]\n" + "add v16.4s, v16.4s, v30.4s\n" + "add v17.4s, v17.4s, v30.4s\n" + "add v18.4s, v18.4s, v30.4s\n" + "add v19.4s, v19.4s, v30.4s\n" + "add v20.4s, v20.4s, v30.4s\n" + "add v21.4s, v21.4s, v30.4s\n" + "add v22.4s, v22.4s, v30.4s\n" + "add v23.4s, v23.4s, v30.4s\n" + "add v24.4s, v24.4s, v30.4s\n" + "add v25.4s, v25.4s, v30.4s\n" + "add v26.4s, v26.4s, v30.4s\n" + "add v27.4s, v27.4s, v30.4s\n" + "smin v16.4s, v16.4s, v29.4s\n" + "smin v17.4s, v17.4s, v29.4s\n" + "smin v18.4s, v18.4s, v29.4s\n" + "smin v19.4s, v19.4s, v29.4s\n" + "smin v20.4s, v20.4s, v29.4s\n" + "smin v21.4s, v21.4s, v29.4s\n" + "smin v22.4s, v22.4s, v29.4s\n" + "smin v23.4s, v23.4s, v29.4s\n" + "smin v24.4s, v24.4s, v29.4s\n" + "smin v25.4s, v25.4s, v29.4s\n" + "smin v26.4s, v26.4s, v29.4s\n" + "smin v27.4s, v27.4s, v29.4s\n" "smax v16.4s, v16.4s, v28.4s\n" "smax v17.4s, v17.4s, v28.4s\n" "smax v18.4s, v18.4s, v28.4s\n" @@ -1298,109 +1299,109 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "uzp1 v18.8h, v22.8h, v23.8h\n" "uzp1 v24.8h, v24.8h, v25.8h\n" "uzp1 v17.8h, v26.8h, v27.8h\n" - "cmp x14, #0x10\n" "uzp1 v16.16b, v16.16b, v19.16b\n" "uzp1 v20.16b, v20.16b, v18.16b\n" "uzp1 v24.16b, v24.16b, v17.16b\n" "bge 89f\n" - "tbz x14, #3, 84f\n" + "tbz x15, #3, 84f\n" "str d16, [x13], #0x8\n" - "str d20, [x23], #0x8\n" - "str d24, [x22], #0x8\n" - "tbz x14, #2, 82f\n" + "str d20, [x24], #0x8\n" + "str d24, [x23], #0x8\n" + "tbz x15, #2, 82f\n" "st1 { v16.s }[2], [x13], #0x4\n" - "st1 { v20.s }[2], [x23], #0x4\n" - "st1 { v24.s }[2], [x22], #0x4\n" - "tbz x14, #1, 81f\n" + "st1 { v20.s }[2], [x24], #0x4\n" + "st1 { v24.s }[2], [x23], #0x4\n" + "tbz x15, #1, 81f\n" "st1 { v16.h }[6], [x13], #0x2\n" - "st1 { v20.h }[6], [x23], #0x2\n" - "st1 { v24.h }[6], [x22], #0x2\n" - "tbz x14, #0, 88f\n" + "st1 { v20.h }[6], [x24], #0x2\n" + "st1 { v24.h }[6], [x23], #0x2\n" + "tbz x15, #0, 88f\n" "st1 { v16.b }[14], [x13]\n" - "st1 { v20.b }[14], [x23]\n" - "st1 { v24.b }[14], [x22]\n" + "st1 { v20.b }[14], [x24]\n" + "st1 { v24.b }[14], [x23]\n" "b 88f\n" "81:" // Height 3: Partial direct writeback: partial_1_12 - "tbz x14, #0, 88f\n" + "tbz x15, #0, 88f\n" "st1 { v16.b }[12], [x13]\n" - "st1 { v20.b }[12], [x23]\n" - "st1 { v24.b }[12], [x22]\n" + "st1 { v20.b }[12], [x24]\n" + "st1 { v24.b }[12], [x23]\n" "b 88f\n" "82:" // Height 3: Partial direct writeback: partial_2_8 - "tbz x14, #1, 83f\n" + "tbz x15, #1, 83f\n" "st1 { v16.h }[4], [x13], #0x2\n" - "st1 { v20.h }[4], [x23], #0x2\n" - "st1 { v24.h }[4], [x22], #0x2\n" - "tbz x14, #0, 88f\n" + "st1 { v20.h }[4], [x24], #0x2\n" + "st1 { v24.h }[4], [x23], #0x2\n" + "tbz x15, #0, 88f\n" "st1 { v16.b }[10], [x13]\n" - "st1 { v20.b }[10], [x23]\n" - "st1 { v24.b }[10], [x22]\n" + "st1 { v20.b }[10], [x24]\n" + "st1 { v24.b }[10], [x23]\n" "b 88f\n" "83:" // Height 3: Partial direct writeback: partial_1_8 - "tbz x14, #0, 88f\n" + "tbz x15, #0, 88f\n" "st1 { v16.b }[8], [x13]\n" - "st1 { v20.b }[8], [x23]\n" - "st1 { v24.b }[8], [x22]\n" + "st1 { v20.b }[8], [x24]\n" + "st1 { v24.b }[8], [x23]\n" "b 88f\n" "84:" // Height 3: Partial direct writeback: partial_4_0 - "tbz x14, #2, 86f\n" + "tbz x15, #2, 86f\n" "str s16, [x13], #0x4\n" - "str s20, [x23], #0x4\n" - "str s24, [x22], #0x4\n" - "tbz x14, #1, 85f\n" + "str s20, [x24], #0x4\n" + "str s24, [x23], #0x4\n" + "tbz x15, #1, 85f\n" "st1 { v16.h }[2], [x13], #0x2\n" - "st1 { v20.h }[2], [x23], #0x2\n" - "st1 { v24.h }[2], [x22], #0x2\n" - "tbz x14, #0, 88f\n" + "st1 { v20.h }[2], [x24], #0x2\n" + "st1 { v24.h }[2], [x23], #0x2\n" + "tbz x15, #0, 88f\n" "st1 { v16.b }[6], [x13]\n" - "st1 { v20.b }[6], [x23]\n" - "st1 { v24.b }[6], [x22]\n" + "st1 { v20.b }[6], [x24]\n" + "st1 { v24.b }[6], [x23]\n" "b 88f\n" "85:" // Height 3: Partial direct writeback: partial_1_4 - "tbz x14, #0, 88f\n" + "tbz x15, #0, 88f\n" "st1 { v16.b }[4], [x13]\n" - "st1 { v20.b }[4], [x23]\n" - "st1 { v24.b }[4], [x22]\n" + "st1 { v20.b }[4], [x24]\n" + "st1 { v24.b }[4], [x23]\n" "b 88f\n" "86:" // Height 3: Partial direct writeback: partial_2_0 - "tbz x14, #1, 87f\n" + "tbz x15, #1, 87f\n" "str h16, [x13], #0x2\n" - "str h20, [x23], #0x2\n" - "str h24, [x22], #0x2\n" - "tbz x14, #0, 88f\n" + "str h20, [x24], #0x2\n" + "str h24, [x23], #0x2\n" + "tbz x15, #0, 88f\n" "st1 { v16.b }[2], [x13]\n" - "st1 { v20.b }[2], [x23]\n" - "st1 { v24.b }[2], [x22]\n" + "st1 { v20.b }[2], [x24]\n" + "st1 { v24.b }[2], [x23]\n" "b 88f\n" "87:" // Height 3: Partial direct writeback: partial_1_0 "str b16, [x13, #0x0]\n" - "str b20, [x23, #0x0]\n" - "str b24, [x22, #0x0]\n" + "str b20, [x24, #0x0]\n" + "str b24, [x23, #0x0]\n" "88:" // Height 3: Partial direct writeback: Done "b 90f\n" "89:" // Height 3: Full writeback "str q16, [x13, #0x0]\n" "add x13, x13, #0x10\n" - "str q20, [x23, #0x0]\n" - "str q24, [x22, #0x0]\n" + "str q20, [x24, #0x0]\n" + "str q24, [x23, #0x0]\n" "90:" // Height 3: Writeback done - "subs x14, x14, #0x10\n" + "subs x15, x15, #0x10\n" "bgt 62b\n" "b 122f\n" "91:" // Height 4 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" "mov x20, #0x4\n" - "mov x15, %x[col_bias]\n" + "ldr x13, [%x[args_ptr], %[offsetof_output_ptr]]\n" + "mov x16, %x[col_bias]\n" "movi v11.4s, #0x0\n" - "movi v12.4s, #0x0\n" "bic %x[flags], %x[flags], #0x80000000\n" + "movi v12.4s, #0x0\n" + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" + "madd x20, x21, x20, x13\n" "movi v13.4s, #0x0\n" - "ldr x14, [%x[args_ptr], %[offsetof_N]]\n" "movi v14.4s, #0x0\n" - "ldr x12, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" "movi v15.16b, #0x1\n" - "mov x13, %x[output_ptr]\n" - "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" + "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n" "92:" // Height 4: Column loop "movi v16.4s, #0x0\n" "movi v17.4s, #0x0\n" @@ -1419,118 +1420,118 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "movi v30.4s, #0x0\n" "movi v31.4s, #0x0\n" "93:" // Height 4: setup done - "mov x11, #0x0\n" + "mov x12, #0x0\n" "94:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w10, [x20, x11, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" "tbz %x[flags], #3, 95f\n" - "ldr x20, [%x[input_ptr], x11, LSL #0x3]\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" - "ldr x9, [x20, #0x0]\n" - "ldr x28, [x20, #0x8]\n" - "ldr x27, [x20, #0x10]\n" - "ldr x26, [x20, #0x18]\n" - "cbnz x11, 96f\n" + "ldr x10, [x20, #0x0]\n" + "ldr x9, [x20, #0x8]\n" + "ldr x28, [x20, #0x10]\n" + "ldr x27, [x20, #0x18]\n" + "cbnz x12, 96f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x10, x10, x20\n" "add x9, x9, x20\n" "add x28, x28, x20\n" "add x27, x27, x20\n" - "add x26, x26, x20\n" "b 96f\n" "95:" // Height 4: setup direct input - "mov x9, %x[input_ptr]\n" + "mov x10, %x[input_ptr]\n" + "add x9, x10, x21\n" "add x28, x9, x21\n" "add x27, x28, x21\n" - "add x26, x27, x21\n" "96:" // Height 4: input setup done - "cmp x10, #0x10\n" + "cmp x11, #0x10\n" "blt 101f\n" - "ldr q0, [x9, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q1, [x28, #0x0]\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x26, #0x0]\n" - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "ldr q8, [x12, #0x40]\n" - "ldr q9, [x12, #0x50]\n" - "ldr q10, [x12, #0x60]\n" + "ldr q0, [x10, #0x0]\n" + "cmp x11, #0x20\n" + "ldr q1, [x9, #0x0]\n" + "ldr q2, [x28, #0x0]\n" + "ldr q3, [x27, #0x0]\n" + "ldr q4, [x14, #0x0]\n" + "ldr q5, [x14, #0x10]\n" + "ldr q6, [x14, #0x20]\n" + "ldr q7, [x14, #0x30]\n" + "ldr q8, [x14, #0x40]\n" + "ldr q9, [x14, #0x50]\n" + "ldr q10, [x14, #0x60]\n" "blt 99f\n" "97:" // Height 4: Multiply loop: Main loop head ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - "ldr x22, [x12, #0x78]\n" + "ldr x21, [x14, #0x78]\n" ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - "ldr x21, [x12, #0x88]\n" + "ldr x20, [x14, #0x88]\n" ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" - "ldr x20, [x12, #0x98]\n" + "ldr x26, [x14, #0x98]\n" ".inst 0x4f83e09c // sdot v28.4s, v4.16b, v3.4b[0]\n" - "ldr d4, [x12, #0x70]\n" + "ldr d4, [x14, #0x70]\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" - "mov v4.d[1], x22\n" + "ldr x25, [x14, #0xa8]\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" - "ldr x25, [x12, #0xa8]\n" + "ldr x24, [x14, #0xb8]\n" ".inst 0x4f82e0b9 // sdot v25.4s, v5.16b, v2.4b[0]\n" - "ldr x24, [x12, #0xb8]\n" + "mov v4.d[1], x21\n" ".inst 0x4f83e0bd // sdot v29.4s, v5.16b, v3.4b[0]\n" - "ldr d5, [x12, #0x80]\n" + "ldr d5, [x14, #0x80]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" - "mov v5.d[1], x21\n" + "ldr x23, [x14, #0xc8]\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" - "ldr x23, [x12, #0xc8]\n" + "ldr x22, [x14, #0xd8]\n" ".inst 0x4f82e0da // sdot v26.4s, v6.16b, v2.4b[0]\n" - "ldr x22, [x12, #0xd8]\n" + "mov v5.d[1], x20\n" ".inst 0x4f83e0de // sdot v30.4s, v6.16b, v3.4b[0]\n" - "ldr d6, [x12, #0x90]\n" + "ldr d6, [x14, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" - "mov v6.d[1], x20\n" + "ldr x21, [x14, #0xe8]\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" - "ldr x21, [x12, #0xe8]\n" + "ldr x20, [x14, #0xf8]\n" ".inst 0x4f82e0fb // sdot v27.4s, v7.16b, v2.4b[0]\n" - "ldr x20, [x12, #0xf8]\n" + "mov v6.d[1], x26\n" ".inst 0x4f83e0ff // sdot v31.4s, v7.16b, v3.4b[0]\n" - "ldr d7, [x12, #0xa0]\n" + "ldr d7, [x14, #0xa0]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" - "mov v7.d[1], x25\n" + "add x10, x10, #0x10\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" "add x9, x9, #0x10\n" ".inst 0x4fa2e118 // sdot v24.4s, v8.16b, v2.4b[1]\n" - "add x28, x28, #0x10\n" + "mov v7.d[1], x25\n" ".inst 0x4fa3e11c // sdot v28.4s, v8.16b, v3.4b[1]\n" - "ldr d8, [x12, #0xb0]\n" + "ldr d8, [x14, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" - "mov v8.d[1], x24\n" + "add x28, x28, #0x10\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" "add x27, x27, #0x10\n" ".inst 0x4fa2e139 // sdot v25.4s, v9.16b, v2.4b[1]\n" - "add x26, x26, #0x10\n" + "mov v8.d[1], x24\n" ".inst 0x4fa3e13d // sdot v29.4s, v9.16b, v3.4b[1]\n" - "ldr d9, [x12, #0xc0]\n" + "ldr d9, [x14, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" - "mov v9.d[1], x23\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" ".inst 0x4fa2e15a // sdot v26.4s, v10.16b, v2.4b[1]\n" + "mov v9.d[1], x23\n" ".inst 0x4fa3e15e // sdot v30.4s, v10.16b, v3.4b[1]\n" - "ldr d10, [x12, #0xd0]\n" + "ldr d10, [x14, #0xd0]\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" - "mov v10.d[1], x22\n" ".inst 0x4fa1e097 // sdot v23.4s, v4.16b, v1.4b[1]\n" ".inst 0x4fa2e09b // sdot v27.4s, v4.16b, v2.4b[1]\n" + "mov v10.d[1], x22\n" ".inst 0x4fa3e09f // sdot v31.4s, v4.16b, v3.4b[1]\n" - "ldr d4, [x12, #0xe0]\n" + "ldr d4, [x14, #0xe0]\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" - "mov v4.d[1], x21\n" ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" ".inst 0x4f82e8b8 // sdot v24.4s, v5.16b, v2.4b[2]\n" + "mov v4.d[1], x21\n" ".inst 0x4f83e8bc // sdot v28.4s, v5.16b, v3.4b[2]\n" - "ldr d5, [x12, #0xf0]\n" + "ldr d5, [x14, #0xf0]\n" ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" - "mov v5.d[1], x20\n" + "add x14, x14, #0x100\n" ".inst 0x4f81e8d5 // sdot v21.4s, v6.16b, v1.4b[2]\n" - "add x12, x12, #0x100\n" ".inst 0x4f82e8d9 // sdot v25.4s, v6.16b, v2.4b[2]\n" + "mov v5.d[1], x20\n" ".inst 0x4f83e8dd // sdot v29.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f80e8f2 // sdot v18.4s, v7.16b, v0.4b[2]\n" ".inst 0x4f81e8f6 // sdot v22.4s, v7.16b, v1.4b[2]\n" @@ -1562,77 +1563,77 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n" ".inst 0x4e8f946e // sdot v14.4s, v3.16b, v15.16b\n" "98:" // Height 4: Multiply loop: unique 13: skip row sum - "ldr q0, [x9, #0x0]\n" - "sub x10, x10, #0x10\n" - "ldr q1, [x28, #0x0]\n" - "cmp x10, #0x20\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x26, #0x0]\n" - "ldr q4, [x12, #0x0]\n" - "ldr q5, [x12, #0x10]\n" - "ldr q6, [x12, #0x20]\n" - "ldr q7, [x12, #0x30]\n" - "ldr q8, [x12, #0x40]\n" - "ldr q9, [x12, #0x50]\n" - "ldr q10, [x12, #0x60]\n" + "ldr q0, [x10, #0x0]\n" + "sub x11, x11, #0x10\n" + "ldr q1, [x9, #0x0]\n" + "cmp x11, #0x20\n" + "ldr q2, [x28, #0x0]\n" + "ldr q3, [x27, #0x0]\n" + "ldr q4, [x14, #0x0]\n" + "ldr q5, [x14, #0x10]\n" + "ldr q6, [x14, #0x20]\n" + "ldr q7, [x14, #0x30]\n" + "ldr q8, [x14, #0x40]\n" + "ldr q9, [x14, #0x50]\n" + "ldr q10, [x14, #0x60]\n" + "prfm pldl1keep, [x10, #0x80]\n" "prfm pldl1keep, [x9, #0x80]\n" "prfm pldl1keep, [x28, #0x80]\n" "prfm pldl1keep, [x27, #0x80]\n" - "prfm pldl1keep, [x26, #0x80]\n" "bge 97b\n" "99:" // Height 4: Multiply loop: Single iteration only ".inst 0x4f80e090 // sdot v16.4s, v4.16b, v0.4b[0]\n" - "sub x10, x10, #0x10\n" + "sub x11, x11, #0x10\n" ".inst 0x4f81e094 // sdot v20.4s, v4.16b, v1.4b[0]\n" - "add x9, x9, #0x10\n" + "add x10, x10, #0x10\n" ".inst 0x4f82e098 // sdot v24.4s, v4.16b, v2.4b[0]\n" - "add x28, x28, #0x10\n" + "add x9, x9, #0x10\n" ".inst 0x4f83e09c // sdot v28.4s, v4.16b, v3.4b[0]\n" - "ldr q4, [x12, #0x70]\n" + "ldr q4, [x14, #0x70]\n" ".inst 0x4f80e0b1 // sdot v17.4s, v5.16b, v0.4b[0]\n" - "add x27, x27, #0x10\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0b5 // sdot v21.4s, v5.16b, v1.4b[0]\n" - "add x26, x26, #0x10\n" + "add x27, x27, #0x10\n" ".inst 0x4f82e0b9 // sdot v25.4s, v5.16b, v2.4b[0]\n" ".inst 0x4f83e0bd // sdot v29.4s, v5.16b, v3.4b[0]\n" - "ldr q5, [x12, #0x80]\n" + "ldr q5, [x14, #0x80]\n" ".inst 0x4f80e0d2 // sdot v18.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0d6 // sdot v22.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0da // sdot v26.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0de // sdot v30.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x12, #0x90]\n" + "ldr q6, [x14, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0f7 // sdot v23.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0fb // sdot v27.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0ff // sdot v31.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x12, #0xa0]\n" + "ldr q7, [x14, #0xa0]\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" ".inst 0x4fa1e114 // sdot v20.4s, v8.16b, v1.4b[1]\n" ".inst 0x4fa2e118 // sdot v24.4s, v8.16b, v2.4b[1]\n" ".inst 0x4fa3e11c // sdot v28.4s, v8.16b, v3.4b[1]\n" - "ldr q8, [x12, #0xb0]\n" + "ldr q8, [x14, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" ".inst 0x4fa1e135 // sdot v21.4s, v9.16b, v1.4b[1]\n" ".inst 0x4fa2e139 // sdot v25.4s, v9.16b, v2.4b[1]\n" ".inst 0x4fa3e13d // sdot v29.4s, v9.16b, v3.4b[1]\n" - "ldr q9, [x12, #0xc0]\n" + "ldr q9, [x14, #0xc0]\n" ".inst 0x4fa0e152 // sdot v18.4s, v10.16b, v0.4b[1]\n" ".inst 0x4fa1e156 // sdot v22.4s, v10.16b, v1.4b[1]\n" ".inst 0x4fa2e15a // sdot v26.4s, v10.16b, v2.4b[1]\n" ".inst 0x4fa3e15e // sdot v30.4s, v10.16b, v3.4b[1]\n" - "ldr q10, [x12, #0xd0]\n" + "ldr q10, [x14, #0xd0]\n" ".inst 0x4fa0e093 // sdot v19.4s, v4.16b, v0.4b[1]\n" ".inst 0x4fa1e097 // sdot v23.4s, v4.16b, v1.4b[1]\n" ".inst 0x4fa2e09b // sdot v27.4s, v4.16b, v2.4b[1]\n" ".inst 0x4fa3e09f // sdot v31.4s, v4.16b, v3.4b[1]\n" - "ldr q4, [x12, #0xe0]\n" + "ldr q4, [x14, #0xe0]\n" ".inst 0x4f80e8b0 // sdot v16.4s, v5.16b, v0.4b[2]\n" ".inst 0x4f81e8b4 // sdot v20.4s, v5.16b, v1.4b[2]\n" ".inst 0x4f82e8b8 // sdot v24.4s, v5.16b, v2.4b[2]\n" ".inst 0x4f83e8bc // sdot v28.4s, v5.16b, v3.4b[2]\n" - "ldr q5, [x12, #0xf0]\n" + "ldr q5, [x14, #0xf0]\n" ".inst 0x4f80e8d1 // sdot v17.4s, v6.16b, v0.4b[2]\n" - "add x12, x12, #0x100\n" + "add x14, x14, #0x100\n" ".inst 0x4f81e8d5 // sdot v21.4s, v6.16b, v1.4b[2]\n" ".inst 0x4f82e8d9 // sdot v25.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8dd // sdot v29.4s, v6.16b, v3.4b[2]\n" @@ -1666,35 +1667,35 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n" ".inst 0x4e8f946e // sdot v14.4s, v3.16b, v15.16b\n" "100:" // Height 4: Multiply loop: unique 14: skip row sum + "prfm pldl1keep, [x10, #0x80]\n" "prfm pldl1keep, [x9, #0x80]\n" "prfm pldl1keep, [x28, #0x80]\n" "prfm pldl1keep, [x27, #0x80]\n" - "prfm pldl1keep, [x26, #0x80]\n" "101:" // Height 4: Multiply loop: Main loop skip - "cbz x10, 108f\n" - "cmp x10, #0x4\n" + "cbz x11, 108f\n" + "cmp x11, #0x4\n" "blt 104f\n" "102:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x9], #0x4\n" - "ldr s1, [x28], #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr s3, [x26], #0x4\n" + "ldr s0, [x10], #0x4\n" + "ldr s1, [x9], #0x4\n" + "ldr s2, [x28], #0x4\n" + "ldr s3, [x27], #0x4\n" "tbnz %x[flags], #31, 103f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n" ".inst 0x4e8f946e // sdot v14.4s, v3.16b, v15.16b\n" "103:" // Height 4: Multiply loop: unique 15: skip row sum - "ldr q7, [x12, #0x0]\n" - "sub x10, x10, #0x4\n" - "ldr q6, [x12, #0x10]\n" - "cmp x10, #0x4\n" - "ldr q5, [x12, #0x20]\n" + "ldr q7, [x14, #0x0]\n" + "sub x11, x11, #0x4\n" + "ldr q6, [x14, #0x10]\n" + "cmp x11, #0x4\n" + "ldr q5, [x14, #0x20]\n" + "ldr q4, [x14, #0x30]\n" ".inst 0x4f80e0f0 // sdot v16.4s, v7.16b, v0.4b[0]\n" - "ldr q4, [x12, #0x30]\n" ".inst 0x4f81e0f4 // sdot v20.4s, v7.16b, v1.4b[0]\n" + "add x14, x14, #0x40\n" ".inst 0x4f82e0f8 // sdot v24.4s, v7.16b, v2.4b[0]\n" - "add x12, x12, #0x40\n" ".inst 0x4f83e0fc // sdot v28.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f80e0d1 // sdot v17.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0d5 // sdot v21.4s, v6.16b, v1.4b[0]\n" @@ -1710,23 +1711,23 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f83e09f // sdot v31.4s, v4.16b, v3.4b[0]\n" "bge 102b\n" "104:" // Height 4: Multiply loop: Skip odd blocks - "cbz x10, 108f\n" - "tbz x10, #1, 105f\n" - "ldr h0, [x9], #0x2\n" - "ldr h1, [x28], #0x2\n" - "ldr h2, [x27], #0x2\n" - "ldr h3, [x26], #0x2\n" - "tbz x10, #0, 106f\n" - "ld1 { v0.b }[2], [x9]\n" - "ld1 { v1.b }[2], [x28]\n" - "ld1 { v2.b }[2], [x27]\n" - "ld1 { v3.b }[2], [x26]\n" + "cbz x11, 108f\n" + "tbz x11, #1, 105f\n" + "ldr h0, [x10], #0x2\n" + "ldr h1, [x9], #0x2\n" + "ldr h2, [x28], #0x2\n" + "ldr h3, [x27], #0x2\n" + "tbz x11, #0, 106f\n" + "ld1 { v0.b }[2], [x10]\n" + "ld1 { v1.b }[2], [x9]\n" + "ld1 { v2.b }[2], [x28]\n" + "ld1 { v3.b }[2], [x27]\n" "b 106f\n" "105:" // Height 4: Multiply loop: Ragged operand read: partial_1_0 - "ldr b0, [x9, #0x0]\n" - "ldr b1, [x28, #0x0]\n" - "ldr b2, [x27, #0x0]\n" - "ldr b3, [x26, #0x0]\n" + "ldr b0, [x10, #0x0]\n" + "ldr b1, [x9, #0x0]\n" + "ldr b2, [x28, #0x0]\n" + "ldr b3, [x27, #0x0]\n" "106:" // Height 4: Multiply loop: Ragged operand read: Done "tbnz %x[flags], #31, 107f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" @@ -1734,16 +1735,16 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n" ".inst 0x4e8f946e // sdot v14.4s, v3.16b, v15.16b\n" "107:" // Height 4: Multiply loop: unique 16: skip row sum - "ldr q7, [x12, #0x0]\n" + "ldr q7, [x14, #0x0]\n" + "ldr q6, [x14, #0x10]\n" + "ldr q5, [x14, #0x20]\n" + "ldr q4, [x14, #0x30]\n" ".inst 0x4f80e0f0 // sdot v16.4s, v7.16b, v0.4b[0]\n" - "ldr q6, [x12, #0x10]\n" ".inst 0x4f81e0f4 // sdot v20.4s, v7.16b, v1.4b[0]\n" - "ldr q5, [x12, #0x20]\n" + "add x14, x14, #0x40\n" ".inst 0x4f82e0f8 // sdot v24.4s, v7.16b, v2.4b[0]\n" - "ldr q4, [x12, #0x30]\n" ".inst 0x4f83e0fc // sdot v28.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f80e0d1 // sdot v17.4s, v6.16b, v0.4b[0]\n" - "add x12, x12, #0x40\n" ".inst 0x4f81e0d5 // sdot v21.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d9 // sdot v25.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0dd // sdot v29.4s, v6.16b, v3.4b[0]\n" @@ -1757,17 +1758,17 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( ".inst 0x4f83e09f // sdot v31.4s, v4.16b, v3.4b[0]\n" "108:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x11, x11, #0x1\n" - "cmp x11, x20\n" + "add x12, x12, #0x1\n" + "cmp x12, x20\n" "bne 94b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x13, x20\n" - "add x22, x23, x20\n" - "add x21, x22, x20\n" "prfm pstl1keep, [x13, #0x0]\n" + "add x24, x13, x20\n" + "add x23, x24, x20\n" + "prfm pstl1keep, [x24, #0x0]\n" + "add x22, x23, x20\n" "prfm pstl1keep, [x23, #0x0]\n" "prfm pstl1keep, [x22, #0x0]\n" - "prfm pstl1keep, [x21, #0x0]\n" "tbnz %x[flags], #31, 109f\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" @@ -1775,9 +1776,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "addp v14.4s, v14.4s, v14.4s\n" "add x20, %x[qp], %[b_offset]\n" "ld1r { v0.4s }, [x20]\n" - "neg v0.4s, v0.4s\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" + "neg v0.4s, v0.4s\n" "addp v13.4s, v13.4s, v13.4s\n" "addp v14.4s, v14.4s, v14.4s\n" "mul v11.4s, v11.4s, v0.4s\n" @@ -1785,13 +1786,13 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "mul v13.4s, v13.4s, v0.4s\n" "mul v14.4s, v14.4s, v0.4s\n" "109:" // Height 4: skip row sum fixup - "ldr q3, [x15, #0x0]\n" + "ldr q3, [x16, #0x0]\n" "add v16.4s, v16.4s, v11.4s\n" - "ldr q2, [x15, #0x10]\n" + "ldr q2, [x16, #0x10]\n" "add v17.4s, v17.4s, v11.4s\n" - "ldr q1, [x15, #0x20]\n" + "ldr q1, [x16, #0x20]\n" "add v18.4s, v18.4s, v11.4s\n" - "ldr q0, [x15, #0x30]\n" + "ldr q0, [x16, #0x30]\n" "add v19.4s, v19.4s, v11.4s\n" "add v20.4s, v20.4s, v12.4s\n" "add v21.4s, v21.4s, v12.4s\n" @@ -1822,10 +1823,11 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "add v30.4s, v30.4s, v1.4s\n" "add v31.4s, v31.4s, v0.4s\n" "add x20, %x[qp], %[per_layer_mul]\n" - "ld1r { v1.4s }, [x20]\n" "orr %x[flags], %x[flags], #0x80000000\n" + "ld1r { v1.4s }, [x20]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" "ld1r { v0.4s }, [x20]\n" + "add x16, x16, #0x40\n" "sqrdmulh v16.4s, v16.4s, v1.4s\n" "sqrdmulh v17.4s, v17.4s, v1.4s\n" "sqrdmulh v18.4s, v18.4s, v1.4s\n" @@ -1842,52 +1844,51 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "sqrdmulh v29.4s, v29.4s, v1.4s\n" "sqrdmulh v30.4s, v30.4s, v1.4s\n" "sqrdmulh v31.4s, v31.4s, v1.4s\n" - "add x15, x15, #0x40\n" "tbz %x[flags], #5, 110f\n" "and v2.16b, v16.16b, v0.16b\n" "and v1.16b, v17.16b, v0.16b\n" - "sshr v2.4s, v2.4s, #0x1f\n" - "sshr v1.4s, v1.4s, #0x1f\n" - "sqadd v16.4s, v16.4s, v2.4s\n" - "sqadd v17.4s, v17.4s, v1.4s\n" "and v7.16b, v18.16b, v0.16b\n" "and v6.16b, v19.16b, v0.16b\n" "and v5.16b, v20.16b, v0.16b\n" "and v4.16b, v21.16b, v0.16b\n" "and v3.16b, v22.16b, v0.16b\n" - "and v2.16b, v23.16b, v0.16b\n" - "and v1.16b, v24.16b, v0.16b\n" + "sshr v2.4s, v2.4s, #0x1f\n" + "sshr v1.4s, v1.4s, #0x1f\n" "sshr v7.4s, v7.4s, #0x1f\n" "sshr v6.4s, v6.4s, #0x1f\n" "sshr v5.4s, v5.4s, #0x1f\n" "sshr v4.4s, v4.4s, #0x1f\n" "sshr v3.4s, v3.4s, #0x1f\n" - "sshr v2.4s, v2.4s, #0x1f\n" - "sshr v1.4s, v1.4s, #0x1f\n" + "sqadd v16.4s, v16.4s, v2.4s\n" + "sqadd v17.4s, v17.4s, v1.4s\n" + "and v2.16b, v23.16b, v0.16b\n" + "and v1.16b, v24.16b, v0.16b\n" "sqadd v18.4s, v18.4s, v7.4s\n" "sqadd v19.4s, v19.4s, v6.4s\n" "sqadd v20.4s, v20.4s, v5.4s\n" "sqadd v21.4s, v21.4s, v4.4s\n" "sqadd v22.4s, v22.4s, v3.4s\n" - "sqadd v23.4s, v23.4s, v2.4s\n" - "sqadd v24.4s, v24.4s, v1.4s\n" "and v7.16b, v25.16b, v0.16b\n" "and v6.16b, v26.16b, v0.16b\n" "and v5.16b, v27.16b, v0.16b\n" + "sshr v2.4s, v2.4s, #0x1f\n" + "sshr v1.4s, v1.4s, #0x1f\n" "and v4.16b, v28.16b, v0.16b\n" "and v3.16b, v29.16b, v0.16b\n" - "and v2.16b, v30.16b, v0.16b\n" - "and v1.16b, v31.16b, v0.16b\n" "sshr v7.4s, v7.4s, #0x1f\n" "sshr v6.4s, v6.4s, #0x1f\n" "sshr v5.4s, v5.4s, #0x1f\n" + "sqadd v23.4s, v23.4s, v2.4s\n" + "sqadd v24.4s, v24.4s, v1.4s\n" + "and v2.16b, v30.16b, v0.16b\n" + "and v1.16b, v31.16b, v0.16b\n" "sshr v4.4s, v4.4s, #0x1f\n" "sshr v3.4s, v3.4s, #0x1f\n" - "sshr v2.4s, v2.4s, #0x1f\n" - "sshr v1.4s, v1.4s, #0x1f\n" "sqadd v25.4s, v25.4s, v7.4s\n" "sqadd v26.4s, v26.4s, v6.4s\n" "sqadd v27.4s, v27.4s, v5.4s\n" + "sshr v2.4s, v2.4s, #0x1f\n" + "sshr v1.4s, v1.4s, #0x1f\n" "sqadd v28.4s, v28.4s, v4.4s\n" "sqadd v29.4s, v29.4s, v3.4s\n" "sqadd v30.4s, v30.4s, v2.4s\n" @@ -1910,43 +1911,44 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "srshl v30.4s, v30.4s, v0.4s\n" "srshl v31.4s, v31.4s, v0.4s\n" "add x20, %x[qp], %[c_offset]\n" - "ld1r { v0.4s }, [x20]\n" - "add v16.4s, v16.4s, v0.4s\n" - "add v17.4s, v17.4s, v0.4s\n" - "add v18.4s, v18.4s, v0.4s\n" - "add v19.4s, v19.4s, v0.4s\n" - "add v20.4s, v20.4s, v0.4s\n" - "add v21.4s, v21.4s, v0.4s\n" - "add v22.4s, v22.4s, v0.4s\n" - "add v23.4s, v23.4s, v0.4s\n" - "add v24.4s, v24.4s, v0.4s\n" - "add v25.4s, v25.4s, v0.4s\n" - "add v26.4s, v26.4s, v0.4s\n" - "add v27.4s, v27.4s, v0.4s\n" - "add v28.4s, v28.4s, v0.4s\n" - "add v29.4s, v29.4s, v0.4s\n" - "add v30.4s, v30.4s, v0.4s\n" - "add v31.4s, v31.4s, v0.4s\n" - "add x20, %x[qp], %[maxval]\n" - "ld1r { v0.4s }, [x20]\n" - "smin v16.4s, v16.4s, v0.4s\n" - "smin v17.4s, v17.4s, v0.4s\n" - "smin v18.4s, v18.4s, v0.4s\n" - "smin v19.4s, v19.4s, v0.4s\n" - "smin v20.4s, v20.4s, v0.4s\n" - "smin v21.4s, v21.4s, v0.4s\n" - "smin v22.4s, v22.4s, v0.4s\n" - "smin v23.4s, v23.4s, v0.4s\n" - "smin v24.4s, v24.4s, v0.4s\n" - "smin v25.4s, v25.4s, v0.4s\n" - "smin v26.4s, v26.4s, v0.4s\n" - "smin v27.4s, v27.4s, v0.4s\n" - "smin v28.4s, v28.4s, v0.4s\n" - "smin v29.4s, v29.4s, v0.4s\n" - "smin v30.4s, v30.4s, v0.4s\n" - "smin v31.4s, v31.4s, v0.4s\n" + "add x21, %x[qp], %[maxval]\n" + "ld1r { v2.4s }, [x20]\n" "add x20, %x[qp], %[minval]\n" + "ld1r { v1.4s }, [x21]\n" + "cmp x15, #0x10\n" "ld1r { v0.4s }, [x20]\n" + "add v16.4s, v16.4s, v2.4s\n" + "add v17.4s, v17.4s, v2.4s\n" + "add v18.4s, v18.4s, v2.4s\n" + "add v19.4s, v19.4s, v2.4s\n" + "add v20.4s, v20.4s, v2.4s\n" + "add v21.4s, v21.4s, v2.4s\n" + "add v22.4s, v22.4s, v2.4s\n" + "add v23.4s, v23.4s, v2.4s\n" + "add v24.4s, v24.4s, v2.4s\n" + "add v25.4s, v25.4s, v2.4s\n" + "add v26.4s, v26.4s, v2.4s\n" + "add v27.4s, v27.4s, v2.4s\n" + "add v28.4s, v28.4s, v2.4s\n" + "add v29.4s, v29.4s, v2.4s\n" + "add v30.4s, v30.4s, v2.4s\n" + "add v31.4s, v31.4s, v2.4s\n" + "smin v16.4s, v16.4s, v1.4s\n" + "smin v17.4s, v17.4s, v1.4s\n" + "smin v18.4s, v18.4s, v1.4s\n" + "smin v19.4s, v19.4s, v1.4s\n" + "smin v20.4s, v20.4s, v1.4s\n" + "smin v21.4s, v21.4s, v1.4s\n" + "smin v22.4s, v22.4s, v1.4s\n" + "smin v23.4s, v23.4s, v1.4s\n" + "smin v24.4s, v24.4s, v1.4s\n" + "smin v25.4s, v25.4s, v1.4s\n" + "smin v26.4s, v26.4s, v1.4s\n" + "smin v27.4s, v27.4s, v1.4s\n" + "smin v28.4s, v28.4s, v1.4s\n" + "smin v29.4s, v29.4s, v1.4s\n" + "smin v30.4s, v30.4s, v1.4s\n" + "smin v31.4s, v31.4s, v1.4s\n" "smax v16.4s, v16.4s, v0.4s\n" "smax v17.4s, v17.4s, v0.4s\n" "smax v18.4s, v18.4s, v0.4s\n" @@ -1971,110 +1973,109 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "uzp1 v18.8h, v26.8h, v27.8h\n" "uzp1 v28.8h, v28.8h, v29.8h\n" "uzp1 v17.8h, v30.8h, v31.8h\n" - "cmp x14, #0x10\n" "uzp1 v16.16b, v16.16b, v0.16b\n" "uzp1 v20.16b, v20.16b, v19.16b\n" "uzp1 v24.16b, v24.16b, v18.16b\n" "uzp1 v28.16b, v28.16b, v17.16b\n" "bge 119f\n" - "tbz x14, #3, 114f\n" + "tbz x15, #3, 114f\n" "str d16, [x13], #0x8\n" - "str d20, [x23], #0x8\n" - "str d24, [x22], #0x8\n" - "str d28, [x21], #0x8\n" - "tbz x14, #2, 112f\n" + "str d20, [x24], #0x8\n" + "str d24, [x23], #0x8\n" + "str d28, [x22], #0x8\n" + "tbz x15, #2, 112f\n" "st1 { v16.s }[2], [x13], #0x4\n" - "st1 { v20.s }[2], [x23], #0x4\n" - "st1 { v24.s }[2], [x22], #0x4\n" - "st1 { v28.s }[2], [x21], #0x4\n" - "tbz x14, #1, 111f\n" + "st1 { v20.s }[2], [x24], #0x4\n" + "st1 { v24.s }[2], [x23], #0x4\n" + "st1 { v28.s }[2], [x22], #0x4\n" + "tbz x15, #1, 111f\n" "st1 { v16.h }[6], [x13], #0x2\n" - "st1 { v20.h }[6], [x23], #0x2\n" - "st1 { v24.h }[6], [x22], #0x2\n" - "st1 { v28.h }[6], [x21], #0x2\n" - "tbz x14, #0, 118f\n" + "st1 { v20.h }[6], [x24], #0x2\n" + "st1 { v24.h }[6], [x23], #0x2\n" + "st1 { v28.h }[6], [x22], #0x2\n" + "tbz x15, #0, 118f\n" "st1 { v16.b }[14], [x13]\n" - "st1 { v20.b }[14], [x23]\n" - "st1 { v24.b }[14], [x22]\n" - "st1 { v28.b }[14], [x21]\n" + "st1 { v20.b }[14], [x24]\n" + "st1 { v24.b }[14], [x23]\n" + "st1 { v28.b }[14], [x22]\n" "b 118f\n" "111:" // Height 4: Partial direct writeback: partial_1_12 - "tbz x14, #0, 118f\n" + "tbz x15, #0, 118f\n" "st1 { v16.b }[12], [x13]\n" - "st1 { v20.b }[12], [x23]\n" - "st1 { v24.b }[12], [x22]\n" - "st1 { v28.b }[12], [x21]\n" + "st1 { v20.b }[12], [x24]\n" + "st1 { v24.b }[12], [x23]\n" + "st1 { v28.b }[12], [x22]\n" "b 118f\n" "112:" // Height 4: Partial direct writeback: partial_2_8 - "tbz x14, #1, 113f\n" + "tbz x15, #1, 113f\n" "st1 { v16.h }[4], [x13], #0x2\n" - "st1 { v20.h }[4], [x23], #0x2\n" - "st1 { v24.h }[4], [x22], #0x2\n" - "st1 { v28.h }[4], [x21], #0x2\n" - "tbz x14, #0, 118f\n" + "st1 { v20.h }[4], [x24], #0x2\n" + "st1 { v24.h }[4], [x23], #0x2\n" + "st1 { v28.h }[4], [x22], #0x2\n" + "tbz x15, #0, 118f\n" "st1 { v16.b }[10], [x13]\n" - "st1 { v20.b }[10], [x23]\n" - "st1 { v24.b }[10], [x22]\n" - "st1 { v28.b }[10], [x21]\n" + "st1 { v20.b }[10], [x24]\n" + "st1 { v24.b }[10], [x23]\n" + "st1 { v28.b }[10], [x22]\n" "b 118f\n" "113:" // Height 4: Partial direct writeback: partial_1_8 - "tbz x14, #0, 118f\n" + "tbz x15, #0, 118f\n" "st1 { v16.b }[8], [x13]\n" - "st1 { v20.b }[8], [x23]\n" - "st1 { v24.b }[8], [x22]\n" - "st1 { v28.b }[8], [x21]\n" + "st1 { v20.b }[8], [x24]\n" + "st1 { v24.b }[8], [x23]\n" + "st1 { v28.b }[8], [x22]\n" "b 118f\n" "114:" // Height 4: Partial direct writeback: partial_4_0 - "tbz x14, #2, 116f\n" + "tbz x15, #2, 116f\n" "str s16, [x13], #0x4\n" - "str s20, [x23], #0x4\n" - "str s24, [x22], #0x4\n" - "str s28, [x21], #0x4\n" - "tbz x14, #1, 115f\n" + "str s20, [x24], #0x4\n" + "str s24, [x23], #0x4\n" + "str s28, [x22], #0x4\n" + "tbz x15, #1, 115f\n" "st1 { v16.h }[2], [x13], #0x2\n" - "st1 { v20.h }[2], [x23], #0x2\n" - "st1 { v24.h }[2], [x22], #0x2\n" - "st1 { v28.h }[2], [x21], #0x2\n" - "tbz x14, #0, 118f\n" + "st1 { v20.h }[2], [x24], #0x2\n" + "st1 { v24.h }[2], [x23], #0x2\n" + "st1 { v28.h }[2], [x22], #0x2\n" + "tbz x15, #0, 118f\n" "st1 { v16.b }[6], [x13]\n" - "st1 { v20.b }[6], [x23]\n" - "st1 { v24.b }[6], [x22]\n" - "st1 { v28.b }[6], [x21]\n" + "st1 { v20.b }[6], [x24]\n" + "st1 { v24.b }[6], [x23]\n" + "st1 { v28.b }[6], [x22]\n" "b 118f\n" "115:" // Height 4: Partial direct writeback: partial_1_4 - "tbz x14, #0, 118f\n" + "tbz x15, #0, 118f\n" "st1 { v16.b }[4], [x13]\n" - "st1 { v20.b }[4], [x23]\n" - "st1 { v24.b }[4], [x22]\n" - "st1 { v28.b }[4], [x21]\n" + "st1 { v20.b }[4], [x24]\n" + "st1 { v24.b }[4], [x23]\n" + "st1 { v28.b }[4], [x22]\n" "b 118f\n" "116:" // Height 4: Partial direct writeback: partial_2_0 - "tbz x14, #1, 117f\n" + "tbz x15, #1, 117f\n" "str h16, [x13], #0x2\n" - "str h20, [x23], #0x2\n" - "str h24, [x22], #0x2\n" - "str h28, [x21], #0x2\n" - "tbz x14, #0, 118f\n" + "str h20, [x24], #0x2\n" + "str h24, [x23], #0x2\n" + "str h28, [x22], #0x2\n" + "tbz x15, #0, 118f\n" "st1 { v16.b }[2], [x13]\n" - "st1 { v20.b }[2], [x23]\n" - "st1 { v24.b }[2], [x22]\n" - "st1 { v28.b }[2], [x21]\n" + "st1 { v20.b }[2], [x24]\n" + "st1 { v24.b }[2], [x23]\n" + "st1 { v28.b }[2], [x22]\n" "b 118f\n" "117:" // Height 4: Partial direct writeback: partial_1_0 "str b16, [x13, #0x0]\n" - "str b20, [x23, #0x0]\n" - "str b24, [x22, #0x0]\n" - "str b28, [x21, #0x0]\n" + "str b20, [x24, #0x0]\n" + "str b24, [x23, #0x0]\n" + "str b28, [x22, #0x0]\n" "118:" // Height 4: Partial direct writeback: Done "b 120f\n" "119:" // Height 4: Full writeback "str q16, [x13, #0x0]\n" "add x13, x13, #0x10\n" - "str q20, [x23, #0x0]\n" - "str q24, [x22, #0x0]\n" - "str q28, [x21, #0x0]\n" + "str q20, [x24, #0x0]\n" + "str q24, [x23, #0x0]\n" + "str q28, [x22, #0x0]\n" "120:" // Height 4: Writeback done - "subs x14, x14, #0x10\n" + "subs x15, x15, #0x10\n" "bgt 92b\n" "subs %x[M], %x[M], #0x4\n" "beq 122f\n" @@ -2088,9 +2089,9 @@ void a64_hybrid_s8qa_dot_4x16_a55 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "122:" // Exit - : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) - : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [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_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [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", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr) + : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [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_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_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", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp index 3b773a6827..75e35a3e98 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021, 2023 Arm Limited. + * Copyright (c) 2019-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,18 +45,18 @@ void a64_hybrid_s8qa_dot_4x16 ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; + void *output_ptr = {}; } ka; unsigned long flags=0; - void *output_ptr; void *input_ptr; if (output_arg.is_indirect) { - output_ptr=(void *)(output_arg.indirect.ptr); + ka.output_ptr=(void *)(output_arg.indirect.ptr); ka.output_offset=output_arg.indirect.offset; flags |= 0x4; } else { - output_ptr=(void *)(output_arg.direct.base); + ka.output_ptr=(void *)(output_arg.direct.base); ka.output_offset=output_arg.direct.stride; } @@ -90,7 +90,7 @@ void a64_hybrid_s8qa_dot_4x16 ( "bic %x[flags], %x[flags], #0x80000000\n" "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x27, %x[output_ptr]\n" + "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n" "2:" // Height 1: Column loop "movi v16.4s, #0x0\n" "movi v17.4s, #0x0\n" @@ -100,8 +100,8 @@ void a64_hybrid_s8qa_dot_4x16 ( "mov x26, #0x0\n" "4:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w25, [x20, x26, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w25, [x20, x26, LSL #0x2]\n" "tbz %x[flags], #3, 5f\n" "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -134,6 +134,7 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q26, [x28, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" "ldr q25, [x28, #0xa0]\n" + "add x24, x24, #0x10\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" "ldr q24, [x28, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" @@ -144,11 +145,10 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q21, [x28, #0xe0]\n" ".inst 0x4f80ea90 // sdot v16.4s, v20.16b, v0.4b[2]\n" "ldr q20, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4f80eb51 // sdot v17.4s, v26.16b, v0.4b[2]\n" - "add x24, x24, #0x10\n" ".inst 0x4f80eb32 // sdot v18.4s, v25.16b, v0.4b[2]\n" ".inst 0x4f80eb13 // sdot v19.4s, v24.16b, v0.4b[2]\n" - "add x28, x28, #0x100\n" ".inst 0x4fa0eaf0 // sdot v16.4s, v23.16b, v0.4b[3]\n" ".inst 0x4fa0ead1 // sdot v17.4s, v22.16b, v0.4b[3]\n" ".inst 0x4fa0eab2 // sdot v18.4s, v21.16b, v0.4b[3]\n" @@ -159,9 +159,9 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q0, [x24, #0x0]\n" "ldr q4, [x28, #0x0]\n" "sub x25, x25, #0x10\n" - "cmp x25, #0x20\n" "ldr q5, [x28, #0x10]\n" "ldr q6, [x28, #0x20]\n" + "cmp x25, #0x20\n" "ldr q7, [x28, #0x30]\n" "ldr q8, [x28, #0x40]\n" "ldr q9, [x28, #0x50]\n" @@ -177,6 +177,8 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q26, [x28, #0x90]\n" ".inst 0x4f80e0f3 // sdot v19.4s, v7.16b, v0.4b[0]\n" "ldr q25, [x28, #0xa0]\n" + "sub x25, x25, #0x10\n" + "add x24, x24, #0x10\n" ".inst 0x4fa0e110 // sdot v16.4s, v8.16b, v0.4b[1]\n" "ldr q24, [x28, #0xb0]\n" ".inst 0x4fa0e131 // sdot v17.4s, v9.16b, v0.4b[1]\n" @@ -187,12 +189,10 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q21, [x28, #0xe0]\n" ".inst 0x4f80ea90 // sdot v16.4s, v20.16b, v0.4b[2]\n" "ldr q20, [x28, #0xf0]\n" + "add x28, x28, #0x100\n" ".inst 0x4f80eb51 // sdot v17.4s, v26.16b, v0.4b[2]\n" - "sub x25, x25, #0x10\n" ".inst 0x4f80eb32 // sdot v18.4s, v25.16b, v0.4b[2]\n" ".inst 0x4f80eb13 // sdot v19.4s, v24.16b, v0.4b[2]\n" - "add x24, x24, #0x10\n" - "add x28, x28, #0x100\n" ".inst 0x4fa0eaf0 // sdot v16.4s, v23.16b, v0.4b[3]\n" ".inst 0x4fa0ead1 // sdot v17.4s, v22.16b, v0.4b[3]\n" ".inst 0x4fa0eab2 // sdot v18.4s, v21.16b, v0.4b[3]\n" @@ -213,14 +213,14 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q23, [x28, #0x0]\n" "ldr q22, [x28, #0x10]\n" "sub x25, x25, #0x4\n" - "cmp x25, #0x4\n" "ldr q21, [x28, #0x20]\n" "ldr q20, [x28, #0x30]\n" + "cmp x25, #0x4\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e2f0 // sdot v16.4s, v23.16b, v0.4b[0]\n" ".inst 0x4f80e2d1 // sdot v17.4s, v22.16b, v0.4b[0]\n" ".inst 0x4f80e2b2 // sdot v18.4s, v21.16b, v0.4b[0]\n" ".inst 0x4f80e293 // sdot v19.4s, v20.16b, v0.4b[0]\n" - "add x28, x28, #0x40\n" "bge 12b\n" "14:" // Height 1: Multiply loop: Skip odd blocks "cbz x25, 18f\n" @@ -235,15 +235,15 @@ void a64_hybrid_s8qa_dot_4x16 ( "tbnz %x[flags], #31, 17f\n" ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" "17:" // Height 1: Multiply loop: unique 4: skip row sum - "ldr q21, [x28, #0x0]\n" - "ldr q20, [x28, #0x10]\n" - ".inst 0x4f80e2b0 // sdot v16.4s, v21.16b, v0.4b[0]\n" - ".inst 0x4f80e291 // sdot v17.4s, v20.16b, v0.4b[0]\n" + "ldr q23, [x28, #0x0]\n" + "ldr q22, [x28, #0x10]\n" "ldr q21, [x28, #0x20]\n" "ldr q20, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + ".inst 0x4f80e2f0 // sdot v16.4s, v23.16b, v0.4b[0]\n" + ".inst 0x4f80e2d1 // sdot v17.4s, v22.16b, v0.4b[0]\n" ".inst 0x4f80e2b2 // sdot v18.4s, v21.16b, v0.4b[0]\n" ".inst 0x4f80e293 // sdot v19.4s, v20.16b, v0.4b[0]\n" - "add x28, x28, #0x40\n" "18:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x26, x26, #0x1\n" @@ -252,8 +252,8 @@ void a64_hybrid_s8qa_dot_4x16 ( "prfm pstl1keep, [x27, #0x0]\n" "tbnz %x[flags], #31, 19f\n" "add x20, %x[qp], %[b_offset]\n" - "ld1r { v20.4s }, [x20]\n" "addp v11.4s, v11.4s, v11.4s\n" + "ld1r { v20.4s }, [x20]\n" "neg v20.4s, v20.4s\n" "addp v11.4s, v11.4s, v11.4s\n" "mul v11.4s, v11.4s, v20.4s\n" @@ -267,16 +267,16 @@ void a64_hybrid_s8qa_dot_4x16 ( "add v18.4s, v18.4s, v11.4s\n" "add v19.4s, v19.4s, v11.4s\n" "add x20, %x[qp], %[per_layer_mul]\n" - "ld1r { v20.4s }, [x20]\n" "orr %x[flags], %x[flags], #0x80000000\n" + "ld1r { v20.4s }, [x20]\n" "add v16.4s, v16.4s, v24.4s\n" "add v17.4s, v17.4s, v23.4s\n" - "add v18.4s, v18.4s, v22.4s\n" "add x20, %x[qp], %[per_layer_right_shift]\n" + "add x10, x10, #0x40\n" "ld1r { v0.4s }, [x20]\n" + "add v18.4s, v18.4s, v22.4s\n" "add v19.4s, v19.4s, v21.4s\n" "sqrdmulh v16.4s, v16.4s, v20.4s\n" - "add x10, x10, #0x40\n" "sqrdmulh v17.4s, v17.4s, v20.4s\n" "sqrdmulh v18.4s, v18.4s, v20.4s\n" "sqrdmulh v19.4s, v19.4s, v20.4s\n" @@ -294,21 +294,21 @@ void a64_hybrid_s8qa_dot_4x16 ( "sqadd v18.4s, v18.4s, v21.4s\n" "sqadd v19.4s, v19.4s, v20.4s\n" "20:" // Height 1: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v22.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v16.4s, v16.4s, v0.4s\n" "srshl v17.4s, v17.4s, v0.4s\n" - "srshl v18.4s, v18.4s, v0.4s\n" - "srshl v19.4s, v19.4s, v0.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v22.4s }, [x21]\n" "ld1r { v21.4s }, [x20]\n" - "add v16.4s, v16.4s, v22.4s\n" - "add v17.4s, v17.4s, v22.4s\n" + "srshl v18.4s, v18.4s, v0.4s\n" + "srshl v19.4s, v19.4s, v0.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x9, #0x10\n" "ld1r { v20.4s }, [x20]\n" + "add v16.4s, v16.4s, v22.4s\n" + "add v17.4s, v17.4s, v22.4s\n" "add v18.4s, v18.4s, v22.4s\n" "add v19.4s, v19.4s, v22.4s\n" - "cmp x9, #0x10\n" "smin v16.4s, v16.4s, v21.4s\n" "smin v17.4s, v17.4s, v21.4s\n" "smin v18.4s, v18.4s, v21.4s\n" @@ -381,7 +381,7 @@ void a64_hybrid_s8qa_dot_4x16 ( "movi v15.16b, #0x1\n" "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x27, %x[output_ptr]\n" + "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n" "32:" // Height 2: Column loop "movi v16.4s, #0x0\n" "movi v17.4s, #0x0\n" @@ -395,8 +395,8 @@ void a64_hybrid_s8qa_dot_4x16 ( "mov x26, #0x0\n" "34:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w25, [x20, x26, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w25, [x20, x26, LSL #0x2]\n" "tbz %x[flags], #3, 35f\n" "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -476,9 +476,9 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q0, [x24, #0x0]\n" "ldr q1, [x23, #0x0]\n" "sub x25, x25, #0x10\n" - "cmp x25, #0x20\n" "ldr q4, [x28, #0x0]\n" "ldr q5, [x28, #0x10]\n" + "cmp x25, #0x20\n" "ldr q6, [x28, #0x20]\n" "ldr q7, [x28, #0x30]\n" "ldr q8, [x28, #0x40]\n" @@ -553,14 +553,14 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q27, [x28, #0x0]\n" "ldr q26, [x28, #0x10]\n" "sub x25, x25, #0x4\n" - "cmp x25, #0x4\n" "ldr q25, [x28, #0x20]\n" "ldr q24, [x28, #0x30]\n" + "cmp x25, #0x4\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e370 // sdot v16.4s, v27.16b, v0.4b[0]\n" ".inst 0x4f81e374 // sdot v20.4s, v27.16b, v1.4b[0]\n" ".inst 0x4f80e351 // sdot v17.4s, v26.16b, v0.4b[0]\n" ".inst 0x4f81e355 // sdot v21.4s, v26.16b, v1.4b[0]\n" - "add x28, x28, #0x40\n" ".inst 0x4f80e332 // sdot v18.4s, v25.16b, v0.4b[0]\n" ".inst 0x4f81e336 // sdot v22.4s, v25.16b, v1.4b[0]\n" ".inst 0x4f80e313 // sdot v19.4s, v24.16b, v0.4b[0]\n" @@ -583,17 +583,17 @@ void a64_hybrid_s8qa_dot_4x16 ( ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n" ".inst 0x4e8f942c // sdot v12.4s, v1.16b, v15.16b\n" "47:" // Height 2: Multiply loop: unique 8: skip row sum - "ldr q24, [x28, #0x0]\n" + "ldr q27, [x28, #0x0]\n" "ldr q26, [x28, #0x10]\n" - ".inst 0x4f80e310 // sdot v16.4s, v24.16b, v0.4b[0]\n" - ".inst 0x4f81e314 // sdot v20.4s, v24.16b, v1.4b[0]\n" "ldr q25, [x28, #0x20]\n" "ldr q24, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + ".inst 0x4f80e370 // sdot v16.4s, v27.16b, v0.4b[0]\n" + ".inst 0x4f81e374 // sdot v20.4s, v27.16b, v1.4b[0]\n" ".inst 0x4f80e351 // sdot v17.4s, v26.16b, v0.4b[0]\n" ".inst 0x4f81e355 // sdot v21.4s, v26.16b, v1.4b[0]\n" ".inst 0x4f80e332 // sdot v18.4s, v25.16b, v0.4b[0]\n" ".inst 0x4f81e336 // sdot v22.4s, v25.16b, v1.4b[0]\n" - "add x28, x28, #0x40\n" ".inst 0x4f80e313 // sdot v19.4s, v24.16b, v0.4b[0]\n" ".inst 0x4f81e317 // sdot v23.4s, v24.16b, v1.4b[0]\n" "48:" // Height 2: Multiply loop: No odd multiplies @@ -602,14 +602,14 @@ void a64_hybrid_s8qa_dot_4x16 ( "cmp x26, x20\n" "bne 34b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x27, x20\n" "prfm pstl1keep, [x27, #0x0]\n" - "prfm pstl1keep, [x23, #0x0]\n" + "add x24, x27, x20\n" + "prfm pstl1keep, [x24, #0x0]\n" "tbnz %x[flags], #31, 49f\n" "add x20, %x[qp], %[b_offset]\n" - "ld1r { v24.4s }, [x20]\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" + "ld1r { v24.4s }, [x20]\n" "neg v24.4s, v24.4s\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" @@ -627,10 +627,10 @@ void a64_hybrid_s8qa_dot_4x16 ( "add v20.4s, v20.4s, v12.4s\n" "add v21.4s, v21.4s, v12.4s\n" "add x20, %x[qp], %[per_layer_mul]\n" + "orr %x[flags], %x[flags], #0x80000000\n" "ld1r { v24.4s }, [x20]\n" "add v22.4s, v22.4s, v12.4s\n" "add v23.4s, v23.4s, v12.4s\n" - "orr %x[flags], %x[flags], #0x80000000\n" "add x20, %x[qp], %[per_layer_right_shift]\n" "add v16.4s, v16.4s, v28.4s\n" "add v17.4s, v17.4s, v27.4s\n" @@ -652,45 +652,45 @@ void a64_hybrid_s8qa_dot_4x16 ( "sqrdmulh v23.4s, v23.4s, v24.4s\n" "tbz %x[flags], #5, 50f\n" "and v24.16b, v16.16b, v0.16b\n" - "sshr v24.4s, v24.4s, #0x1f\n" - "sqadd v16.4s, v16.4s, v24.4s\n" "and v30.16b, v17.16b, v0.16b\n" "and v29.16b, v18.16b, v0.16b\n" "and v28.16b, v19.16b, v0.16b\n" "and v27.16b, v20.16b, v0.16b\n" "and v26.16b, v21.16b, v0.16b\n" + "sshr v24.4s, v24.4s, #0x1f\n" "and v25.16b, v22.16b, v0.16b\n" - "and v24.16b, v23.16b, v0.16b\n" "sshr v30.4s, v30.4s, #0x1f\n" "sshr v29.4s, v29.4s, #0x1f\n" "sshr v28.4s, v28.4s, #0x1f\n" "sshr v27.4s, v27.4s, #0x1f\n" + "sqadd v16.4s, v16.4s, v24.4s\n" + "and v24.16b, v23.16b, v0.16b\n" "sshr v26.4s, v26.4s, #0x1f\n" "sshr v25.4s, v25.4s, #0x1f\n" - "sshr v24.4s, v24.4s, #0x1f\n" "sqadd v17.4s, v17.4s, v30.4s\n" "sqadd v18.4s, v18.4s, v29.4s\n" + "sshr v24.4s, v24.4s, #0x1f\n" "sqadd v19.4s, v19.4s, v28.4s\n" "sqadd v20.4s, v20.4s, v27.4s\n" "sqadd v21.4s, v21.4s, v26.4s\n" "sqadd v22.4s, v22.4s, v25.4s\n" "sqadd v23.4s, v23.4s, v24.4s\n" "50:" // Height 2: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v26.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v16.4s, v16.4s, v0.4s\n" "srshl v17.4s, v17.4s, v0.4s\n" - "srshl v18.4s, v18.4s, v0.4s\n" - "srshl v19.4s, v19.4s, v0.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v26.4s }, [x21]\n" "ld1r { v25.4s }, [x20]\n" + "srshl v18.4s, v18.4s, v0.4s\n" + "srshl v19.4s, v19.4s, v0.4s\n" "srshl v20.4s, v20.4s, v0.4s\n" "srshl v21.4s, v21.4s, v0.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x9, #0x10\n" "ld1r { v24.4s }, [x20]\n" "srshl v22.4s, v22.4s, v0.4s\n" "srshl v23.4s, v23.4s, v0.4s\n" - "cmp x9, #0x10\n" "add v16.4s, v16.4s, v26.4s\n" "add v17.4s, v17.4s, v26.4s\n" "add v18.4s, v18.4s, v26.4s\n" @@ -724,68 +724,68 @@ void a64_hybrid_s8qa_dot_4x16 ( "bge 59f\n" "tbz x9, #3, 54f\n" "str d16, [x27], #0x8\n" - "str d20, [x23], #0x8\n" + "str d20, [x24], #0x8\n" "tbz x9, #2, 52f\n" "st1 { v16.s }[2], [x27], #0x4\n" - "st1 { v20.s }[2], [x23], #0x4\n" + "st1 { v20.s }[2], [x24], #0x4\n" "tbz x9, #1, 51f\n" "st1 { v16.h }[6], [x27], #0x2\n" - "st1 { v20.h }[6], [x23], #0x2\n" + "st1 { v20.h }[6], [x24], #0x2\n" "tbz x9, #0, 58f\n" "st1 { v16.b }[14], [x27]\n" - "st1 { v20.b }[14], [x23]\n" + "st1 { v20.b }[14], [x24]\n" "b 58f\n" "51:" // Height 2: Partial direct writeback: partial_1_12 "tbz x9, #0, 58f\n" "st1 { v16.b }[12], [x27]\n" - "st1 { v20.b }[12], [x23]\n" + "st1 { v20.b }[12], [x24]\n" "b 58f\n" "52:" // Height 2: Partial direct writeback: partial_2_8 "tbz x9, #1, 53f\n" "st1 { v16.h }[4], [x27], #0x2\n" - "st1 { v20.h }[4], [x23], #0x2\n" + "st1 { v20.h }[4], [x24], #0x2\n" "tbz x9, #0, 58f\n" "st1 { v16.b }[10], [x27]\n" - "st1 { v20.b }[10], [x23]\n" + "st1 { v20.b }[10], [x24]\n" "b 58f\n" "53:" // Height 2: Partial direct writeback: partial_1_8 "tbz x9, #0, 58f\n" "st1 { v16.b }[8], [x27]\n" - "st1 { v20.b }[8], [x23]\n" + "st1 { v20.b }[8], [x24]\n" "b 58f\n" "54:" // Height 2: Partial direct writeback: partial_4_0 "tbz x9, #2, 56f\n" "str s16, [x27], #0x4\n" - "str s20, [x23], #0x4\n" + "str s20, [x24], #0x4\n" "tbz x9, #1, 55f\n" "st1 { v16.h }[2], [x27], #0x2\n" - "st1 { v20.h }[2], [x23], #0x2\n" + "st1 { v20.h }[2], [x24], #0x2\n" "tbz x9, #0, 58f\n" "st1 { v16.b }[6], [x27]\n" - "st1 { v20.b }[6], [x23]\n" + "st1 { v20.b }[6], [x24]\n" "b 58f\n" "55:" // Height 2: Partial direct writeback: partial_1_4 "tbz x9, #0, 58f\n" "st1 { v16.b }[4], [x27]\n" - "st1 { v20.b }[4], [x23]\n" + "st1 { v20.b }[4], [x24]\n" "b 58f\n" "56:" // Height 2: Partial direct writeback: partial_2_0 "tbz x9, #1, 57f\n" "str h16, [x27], #0x2\n" - "str h20, [x23], #0x2\n" + "str h20, [x24], #0x2\n" "tbz x9, #0, 58f\n" "st1 { v16.b }[2], [x27]\n" - "st1 { v20.b }[2], [x23]\n" + "st1 { v20.b }[2], [x24]\n" "b 58f\n" "57:" // Height 2: Partial direct writeback: partial_1_0 "str b16, [x27, #0x0]\n" - "str b20, [x23, #0x0]\n" + "str b20, [x24, #0x0]\n" "58:" // Height 2: Partial direct writeback: Done "b 60f\n" "59:" // Height 2: Full writeback "str q16, [x27, #0x0]\n" "add x27, x27, #0x10\n" - "str q20, [x23, #0x0]\n" + "str q20, [x24, #0x0]\n" "60:" // Height 2: Writeback done "subs x9, x9, #0x10\n" "bgt 32b\n" @@ -799,7 +799,7 @@ void a64_hybrid_s8qa_dot_4x16 ( "movi v15.16b, #0x1\n" "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x27, %x[output_ptr]\n" + "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n" "62:" // Height 3: Column loop "movi v16.4s, #0x0\n" "movi v17.4s, #0x0\n" @@ -817,8 +817,8 @@ void a64_hybrid_s8qa_dot_4x16 ( "mov x26, #0x0\n" "64:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w25, [x20, x26, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w25, [x20, x26, LSL #0x2]\n" "tbz %x[flags], #3, 65f\n" "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -920,9 +920,9 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q0, [x24, #0x0]\n" "ldr q1, [x23, #0x0]\n" "sub x25, x25, #0x10\n" - "cmp x25, #0x20\n" "ldr q2, [x22, #0x0]\n" "ldr q4, [x28, #0x0]\n" + "cmp x25, #0x20\n" "ldr q5, [x28, #0x10]\n" "ldr q6, [x28, #0x20]\n" "ldr q7, [x28, #0x30]\n" @@ -1020,14 +1020,14 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q31, [x28, #0x0]\n" "ldr q30, [x28, #0x10]\n" "sub x25, x25, #0x4\n" - "cmp x25, #0x4\n" "ldr q29, [x28, #0x20]\n" "ldr q28, [x28, #0x30]\n" + "cmp x25, #0x4\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e3f0 // sdot v16.4s, v31.16b, v0.4b[0]\n" ".inst 0x4f81e3f4 // sdot v20.4s, v31.16b, v1.4b[0]\n" ".inst 0x4f82e3f8 // sdot v24.4s, v31.16b, v2.4b[0]\n" ".inst 0x4f80e3d1 // sdot v17.4s, v30.16b, v0.4b[0]\n" - "add x28, x28, #0x40\n" ".inst 0x4f81e3d5 // sdot v21.4s, v30.16b, v1.4b[0]\n" ".inst 0x4f82e3d9 // sdot v25.4s, v30.16b, v2.4b[0]\n" ".inst 0x4f80e3b2 // sdot v18.4s, v29.16b, v0.4b[0]\n" @@ -1060,15 +1060,15 @@ void a64_hybrid_s8qa_dot_4x16 ( "77:" // Height 3: Multiply loop: unique 12: skip row sum "ldr q31, [x28, #0x0]\n" "ldr q30, [x28, #0x10]\n" - ".inst 0x4f80e3f0 // sdot v16.4s, v31.16b, v0.4b[0]\n" - ".inst 0x4f81e3f4 // sdot v20.4s, v31.16b, v1.4b[0]\n" "ldr q29, [x28, #0x20]\n" "ldr q28, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + ".inst 0x4f80e3f0 // sdot v16.4s, v31.16b, v0.4b[0]\n" + ".inst 0x4f81e3f4 // sdot v20.4s, v31.16b, v1.4b[0]\n" ".inst 0x4f82e3f8 // sdot v24.4s, v31.16b, v2.4b[0]\n" ".inst 0x4f80e3d1 // sdot v17.4s, v30.16b, v0.4b[0]\n" ".inst 0x4f81e3d5 // sdot v21.4s, v30.16b, v1.4b[0]\n" ".inst 0x4f82e3d9 // sdot v25.4s, v30.16b, v2.4b[0]\n" - "add x28, x28, #0x40\n" ".inst 0x4f80e3b2 // sdot v18.4s, v29.16b, v0.4b[0]\n" ".inst 0x4f81e3b6 // sdot v22.4s, v29.16b, v1.4b[0]\n" ".inst 0x4f82e3ba // sdot v26.4s, v29.16b, v2.4b[0]\n" @@ -1081,16 +1081,16 @@ void a64_hybrid_s8qa_dot_4x16 ( "cmp x26, x20\n" "bne 64b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x27, x20\n" - "add x22, x23, x20\n" "prfm pstl1keep, [x27, #0x0]\n" + "add x24, x27, x20\n" + "prfm pstl1keep, [x24, #0x0]\n" + "add x23, x24, x20\n" "prfm pstl1keep, [x23, #0x0]\n" - "prfm pstl1keep, [x22, #0x0]\n" "tbnz %x[flags], #31, 79f\n" "add x20, %x[qp], %[b_offset]\n" - "ld1r { v28.4s }, [x20]\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" + "ld1r { v28.4s }, [x20]\n" "addp v13.4s, v13.4s, v13.4s\n" "neg v28.4s, v28.4s\n" "addp v11.4s, v11.4s, v11.4s\n" @@ -1111,10 +1111,10 @@ void a64_hybrid_s8qa_dot_4x16 ( "add v20.4s, v20.4s, v12.4s\n" "add v21.4s, v21.4s, v12.4s\n" "add x20, %x[qp], %[per_layer_mul]\n" + "orr %x[flags], %x[flags], #0x80000000\n" "ld1r { v28.4s }, [x20]\n" "add v22.4s, v22.4s, v12.4s\n" "add v23.4s, v23.4s, v12.4s\n" - "orr %x[flags], %x[flags], #0x80000000\n" "add x20, %x[qp], %[per_layer_right_shift]\n" "add v24.4s, v24.4s, v13.4s\n" "add v25.4s, v25.4s, v13.4s\n" @@ -1152,18 +1152,18 @@ void a64_hybrid_s8qa_dot_4x16 ( "and v30.16b, v18.16b, v0.16b\n" "and v29.16b, v19.16b, v0.16b\n" "and v28.16b, v20.16b, v0.16b\n" + "and v3.16b, v21.16b, v0.16b\n" "sshr v1.4s, v1.4s, #0x1f\n" "sshr v31.4s, v31.4s, #0x1f\n" "sshr v30.4s, v30.4s, #0x1f\n" "sshr v29.4s, v29.4s, #0x1f\n" "sshr v28.4s, v28.4s, #0x1f\n" + "and v2.16b, v22.16b, v0.16b\n" "sqadd v16.4s, v16.4s, v1.4s\n" "sqadd v17.4s, v17.4s, v31.4s\n" "sqadd v18.4s, v18.4s, v30.4s\n" "sqadd v19.4s, v19.4s, v29.4s\n" "sqadd v20.4s, v20.4s, v28.4s\n" - "and v3.16b, v21.16b, v0.16b\n" - "and v2.16b, v22.16b, v0.16b\n" "and v1.16b, v23.16b, v0.16b\n" "and v31.16b, v24.16b, v0.16b\n" "and v30.16b, v25.16b, v0.16b\n" @@ -1184,21 +1184,21 @@ void a64_hybrid_s8qa_dot_4x16 ( "sqadd v26.4s, v26.4s, v29.4s\n" "sqadd v27.4s, v27.4s, v28.4s\n" "80:" // Height 3: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v30.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v16.4s, v16.4s, v0.4s\n" "srshl v17.4s, v17.4s, v0.4s\n" - "srshl v18.4s, v18.4s, v0.4s\n" - "srshl v19.4s, v19.4s, v0.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v30.4s }, [x21]\n" "ld1r { v29.4s }, [x20]\n" + "srshl v18.4s, v18.4s, v0.4s\n" + "srshl v19.4s, v19.4s, v0.4s\n" "srshl v20.4s, v20.4s, v0.4s\n" "srshl v21.4s, v21.4s, v0.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x9, #0x10\n" "ld1r { v28.4s }, [x20]\n" "srshl v22.4s, v22.4s, v0.4s\n" "srshl v23.4s, v23.4s, v0.4s\n" - "cmp x9, #0x10\n" "srshl v24.4s, v24.4s, v0.4s\n" "srshl v25.4s, v25.4s, v0.4s\n" "srshl v26.4s, v26.4s, v0.4s\n" @@ -1251,102 +1251,103 @@ void a64_hybrid_s8qa_dot_4x16 ( "bge 89f\n" "tbz x9, #3, 84f\n" "str d16, [x27], #0x8\n" - "str d20, [x23], #0x8\n" - "str d24, [x22], #0x8\n" + "str d20, [x24], #0x8\n" + "str d24, [x23], #0x8\n" "tbz x9, #2, 82f\n" "st1 { v16.s }[2], [x27], #0x4\n" - "st1 { v20.s }[2], [x23], #0x4\n" - "st1 { v24.s }[2], [x22], #0x4\n" + "st1 { v20.s }[2], [x24], #0x4\n" + "st1 { v24.s }[2], [x23], #0x4\n" "tbz x9, #1, 81f\n" "st1 { v16.h }[6], [x27], #0x2\n" - "st1 { v20.h }[6], [x23], #0x2\n" - "st1 { v24.h }[6], [x22], #0x2\n" + "st1 { v20.h }[6], [x24], #0x2\n" + "st1 { v24.h }[6], [x23], #0x2\n" "tbz x9, #0, 88f\n" "st1 { v16.b }[14], [x27]\n" - "st1 { v20.b }[14], [x23]\n" - "st1 { v24.b }[14], [x22]\n" + "st1 { v20.b }[14], [x24]\n" + "st1 { v24.b }[14], [x23]\n" "b 88f\n" "81:" // Height 3: Partial direct writeback: partial_1_12 "tbz x9, #0, 88f\n" "st1 { v16.b }[12], [x27]\n" - "st1 { v20.b }[12], [x23]\n" - "st1 { v24.b }[12], [x22]\n" + "st1 { v20.b }[12], [x24]\n" + "st1 { v24.b }[12], [x23]\n" "b 88f\n" "82:" // Height 3: Partial direct writeback: partial_2_8 "tbz x9, #1, 83f\n" "st1 { v16.h }[4], [x27], #0x2\n" - "st1 { v20.h }[4], [x23], #0x2\n" - "st1 { v24.h }[4], [x22], #0x2\n" + "st1 { v20.h }[4], [x24], #0x2\n" + "st1 { v24.h }[4], [x23], #0x2\n" "tbz x9, #0, 88f\n" "st1 { v16.b }[10], [x27]\n" - "st1 { v20.b }[10], [x23]\n" - "st1 { v24.b }[10], [x22]\n" + "st1 { v20.b }[10], [x24]\n" + "st1 { v24.b }[10], [x23]\n" "b 88f\n" "83:" // Height 3: Partial direct writeback: partial_1_8 "tbz x9, #0, 88f\n" "st1 { v16.b }[8], [x27]\n" - "st1 { v20.b }[8], [x23]\n" - "st1 { v24.b }[8], [x22]\n" + "st1 { v20.b }[8], [x24]\n" + "st1 { v24.b }[8], [x23]\n" "b 88f\n" "84:" // Height 3: Partial direct writeback: partial_4_0 "tbz x9, #2, 86f\n" "str s16, [x27], #0x4\n" - "str s20, [x23], #0x4\n" - "str s24, [x22], #0x4\n" + "str s20, [x24], #0x4\n" + "str s24, [x23], #0x4\n" "tbz x9, #1, 85f\n" "st1 { v16.h }[2], [x27], #0x2\n" - "st1 { v20.h }[2], [x23], #0x2\n" - "st1 { v24.h }[2], [x22], #0x2\n" + "st1 { v20.h }[2], [x24], #0x2\n" + "st1 { v24.h }[2], [x23], #0x2\n" "tbz x9, #0, 88f\n" "st1 { v16.b }[6], [x27]\n" - "st1 { v20.b }[6], [x23]\n" - "st1 { v24.b }[6], [x22]\n" + "st1 { v20.b }[6], [x24]\n" + "st1 { v24.b }[6], [x23]\n" "b 88f\n" "85:" // Height 3: Partial direct writeback: partial_1_4 "tbz x9, #0, 88f\n" "st1 { v16.b }[4], [x27]\n" - "st1 { v20.b }[4], [x23]\n" - "st1 { v24.b }[4], [x22]\n" + "st1 { v20.b }[4], [x24]\n" + "st1 { v24.b }[4], [x23]\n" "b 88f\n" "86:" // Height 3: Partial direct writeback: partial_2_0 "tbz x9, #1, 87f\n" "str h16, [x27], #0x2\n" - "str h20, [x23], #0x2\n" - "str h24, [x22], #0x2\n" + "str h20, [x24], #0x2\n" + "str h24, [x23], #0x2\n" "tbz x9, #0, 88f\n" "st1 { v16.b }[2], [x27]\n" - "st1 { v20.b }[2], [x23]\n" - "st1 { v24.b }[2], [x22]\n" + "st1 { v20.b }[2], [x24]\n" + "st1 { v24.b }[2], [x23]\n" "b 88f\n" "87:" // Height 3: Partial direct writeback: partial_1_0 "str b16, [x27, #0x0]\n" - "str b20, [x23, #0x0]\n" - "str b24, [x22, #0x0]\n" + "str b20, [x24, #0x0]\n" + "str b24, [x23, #0x0]\n" "88:" // Height 3: Partial direct writeback: Done "b 90f\n" "89:" // Height 3: Full writeback "str q16, [x27, #0x0]\n" "add x27, x27, #0x10\n" - "str q20, [x23, #0x0]\n" - "str q24, [x22, #0x0]\n" + "str q20, [x24, #0x0]\n" + "str q24, [x23, #0x0]\n" "90:" // Height 3: Writeback done "subs x9, x9, #0x10\n" "bgt 62b\n" "b 122f\n" "91:" // Height 4 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ldr x27, [%x[args_ptr], %[offsetof_output_ptr]]\n" "mov x20, #0x4\n" "mov x10, %x[col_bias]\n" "movi v11.4s, #0x0\n" "movi v12.4s, #0x0\n" - "movi v13.4s, #0x0\n" "bic %x[flags], %x[flags], #0x80000000\n" "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" + "movi v13.4s, #0x0\n" "movi v14.4s, #0x0\n" - "movi v15.16b, #0x1\n" "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x27, %x[output_ptr]\n" - "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" + "madd x20, x21, x20, x27\n" + "movi v15.16b, #0x1\n" + "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n" "92:" // Height 4: Column loop "movi v16.4s, #0x0\n" "movi v17.4s, #0x0\n" @@ -1368,8 +1369,8 @@ void a64_hybrid_s8qa_dot_4x16 ( "mov x26, #0x0\n" "94:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w25, [x20, x26, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w25, [x20, x26, LSL #0x2]\n" "tbz %x[flags], #3, 95f\n" "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1493,9 +1494,9 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q0, [x24, #0x0]\n" "ldr q1, [x23, #0x0]\n" "sub x25, x25, #0x10\n" - "cmp x25, #0x20\n" "ldr q2, [x22, #0x0]\n" "ldr q3, [x21, #0x0]\n" + "cmp x25, #0x20\n" "ldr q4, [x28, #0x0]\n" "ldr q5, [x28, #0x10]\n" "ldr q6, [x28, #0x20]\n" @@ -1616,14 +1617,14 @@ void a64_hybrid_s8qa_dot_4x16 ( "ldr q7, [x28, #0x0]\n" "ldr q6, [x28, #0x10]\n" "sub x25, x25, #0x4\n" - "cmp x25, #0x4\n" "ldr q5, [x28, #0x20]\n" "ldr q4, [x28, #0x30]\n" + "cmp x25, #0x4\n" + "add x28, x28, #0x40\n" ".inst 0x4f80e0f0 // sdot v16.4s, v7.16b, v0.4b[0]\n" ".inst 0x4f81e0f4 // sdot v20.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f8 // sdot v24.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0fc // sdot v28.4s, v7.16b, v3.4b[0]\n" - "add x28, x28, #0x40\n" ".inst 0x4f80e0d1 // sdot v17.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0d5 // sdot v21.4s, v6.16b, v1.4b[0]\n" ".inst 0x4f82e0d9 // sdot v25.4s, v6.16b, v2.4b[0]\n" @@ -1664,15 +1665,15 @@ void a64_hybrid_s8qa_dot_4x16 ( "107:" // Height 4: Multiply loop: unique 16: skip row sum "ldr q7, [x28, #0x0]\n" "ldr q6, [x28, #0x10]\n" - ".inst 0x4f80e0f0 // sdot v16.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0f4 // sdot v20.4s, v7.16b, v1.4b[0]\n" "ldr q5, [x28, #0x20]\n" "ldr q4, [x28, #0x30]\n" + "add x28, x28, #0x40\n" + ".inst 0x4f80e0f0 // sdot v16.4s, v7.16b, v0.4b[0]\n" + ".inst 0x4f81e0f4 // sdot v20.4s, v7.16b, v1.4b[0]\n" ".inst 0x4f82e0f8 // sdot v24.4s, v7.16b, v2.4b[0]\n" ".inst 0x4f83e0fc // sdot v28.4s, v7.16b, v3.4b[0]\n" ".inst 0x4f80e0d1 // sdot v17.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0d5 // sdot v21.4s, v6.16b, v1.4b[0]\n" - "add x28, x28, #0x40\n" ".inst 0x4f82e0d9 // sdot v25.4s, v6.16b, v2.4b[0]\n" ".inst 0x4f83e0dd // sdot v29.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f80e0b2 // sdot v18.4s, v5.16b, v0.4b[0]\n" @@ -1689,18 +1690,18 @@ void a64_hybrid_s8qa_dot_4x16 ( "cmp x26, x20\n" "bne 94b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x23, x27, x20\n" - "add x22, x23, x20\n" "prfm pstl1keep, [x27, #0x0]\n" - "add x21, x22, x20\n" + "add x24, x27, x20\n" + "prfm pstl1keep, [x24, #0x0]\n" + "add x23, x24, x20\n" "prfm pstl1keep, [x23, #0x0]\n" + "add x22, x23, x20\n" "prfm pstl1keep, [x22, #0x0]\n" - "prfm pstl1keep, [x21, #0x0]\n" "tbnz %x[flags], #31, 109f\n" "add x20, %x[qp], %[b_offset]\n" - "ld1r { v0.4s }, [x20]\n" "addp v11.4s, v11.4s, v11.4s\n" "addp v12.4s, v12.4s, v12.4s\n" + "ld1r { v0.4s }, [x20]\n" "addp v13.4s, v13.4s, v13.4s\n" "addp v14.4s, v14.4s, v14.4s\n" "neg v0.4s, v0.4s\n" @@ -1724,10 +1725,10 @@ void a64_hybrid_s8qa_dot_4x16 ( "add v20.4s, v20.4s, v12.4s\n" "add v21.4s, v21.4s, v12.4s\n" "add x20, %x[qp], %[per_layer_mul]\n" + "orr %x[flags], %x[flags], #0x80000000\n" "ld1r { v1.4s }, [x20]\n" "add v22.4s, v22.4s, v12.4s\n" "add v23.4s, v23.4s, v12.4s\n" - "orr %x[flags], %x[flags], #0x80000000\n" "add x20, %x[qp], %[per_layer_right_shift]\n" "add v24.4s, v24.4s, v13.4s\n" "add v25.4s, v25.4s, v13.4s\n" @@ -1774,32 +1775,32 @@ void a64_hybrid_s8qa_dot_4x16 ( "tbz %x[flags], #5, 110f\n" "and v2.16b, v16.16b, v0.16b\n" "and v1.16b, v17.16b, v0.16b\n" - "sshr v2.4s, v2.4s, #0x1f\n" - "sshr v1.4s, v1.4s, #0x1f\n" - "sqadd v16.4s, v16.4s, v2.4s\n" - "sqadd v17.4s, v17.4s, v1.4s\n" "and v7.16b, v18.16b, v0.16b\n" "and v6.16b, v19.16b, v0.16b\n" "and v5.16b, v20.16b, v0.16b\n" "and v4.16b, v21.16b, v0.16b\n" + "sshr v2.4s, v2.4s, #0x1f\n" + "sshr v1.4s, v1.4s, #0x1f\n" "and v3.16b, v22.16b, v0.16b\n" - "and v2.16b, v23.16b, v0.16b\n" - "and v1.16b, v24.16b, v0.16b\n" "sshr v7.4s, v7.4s, #0x1f\n" "sshr v6.4s, v6.4s, #0x1f\n" "sshr v5.4s, v5.4s, #0x1f\n" + "sqadd v16.4s, v16.4s, v2.4s\n" + "sqadd v17.4s, v17.4s, v1.4s\n" + "and v2.16b, v23.16b, v0.16b\n" + "and v1.16b, v24.16b, v0.16b\n" "sshr v4.4s, v4.4s, #0x1f\n" "sshr v3.4s, v3.4s, #0x1f\n" - "sshr v2.4s, v2.4s, #0x1f\n" - "sshr v1.4s, v1.4s, #0x1f\n" "sqadd v18.4s, v18.4s, v7.4s\n" "sqadd v19.4s, v19.4s, v6.4s\n" + "sshr v2.4s, v2.4s, #0x1f\n" + "sshr v1.4s, v1.4s, #0x1f\n" "sqadd v20.4s, v20.4s, v5.4s\n" "sqadd v21.4s, v21.4s, v4.4s\n" "sqadd v22.4s, v22.4s, v3.4s\n" + "and v7.16b, v25.16b, v0.16b\n" "sqadd v23.4s, v23.4s, v2.4s\n" "sqadd v24.4s, v24.4s, v1.4s\n" - "and v7.16b, v25.16b, v0.16b\n" "and v6.16b, v26.16b, v0.16b\n" "and v5.16b, v27.16b, v0.16b\n" "and v4.16b, v28.16b, v0.16b\n" @@ -1821,21 +1822,21 @@ void a64_hybrid_s8qa_dot_4x16 ( "sqadd v30.4s, v30.4s, v2.4s\n" "sqadd v31.4s, v31.4s, v1.4s\n" "110:" // Height 4: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v3.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v16.4s, v16.4s, v0.4s\n" "srshl v17.4s, v17.4s, v0.4s\n" - "srshl v18.4s, v18.4s, v0.4s\n" - "srshl v19.4s, v19.4s, v0.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v3.4s }, [x21]\n" "ld1r { v2.4s }, [x20]\n" + "srshl v18.4s, v18.4s, v0.4s\n" + "srshl v19.4s, v19.4s, v0.4s\n" "srshl v20.4s, v20.4s, v0.4s\n" "srshl v21.4s, v21.4s, v0.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x9, #0x10\n" "ld1r { v1.4s }, [x20]\n" "srshl v22.4s, v22.4s, v0.4s\n" "srshl v23.4s, v23.4s, v0.4s\n" - "cmp x9, #0x10\n" "srshl v24.4s, v24.4s, v0.4s\n" "srshl v25.4s, v25.4s, v0.4s\n" "srshl v26.4s, v26.4s, v0.4s\n" @@ -1907,100 +1908,100 @@ void a64_hybrid_s8qa_dot_4x16 ( "bge 119f\n" "tbz x9, #3, 114f\n" "str d16, [x27], #0x8\n" - "str d20, [x23], #0x8\n" - "str d24, [x22], #0x8\n" - "str d28, [x21], #0x8\n" + "str d20, [x24], #0x8\n" + "str d24, [x23], #0x8\n" + "str d28, [x22], #0x8\n" "tbz x9, #2, 112f\n" "st1 { v16.s }[2], [x27], #0x4\n" - "st1 { v20.s }[2], [x23], #0x4\n" - "st1 { v24.s }[2], [x22], #0x4\n" - "st1 { v28.s }[2], [x21], #0x4\n" + "st1 { v20.s }[2], [x24], #0x4\n" + "st1 { v24.s }[2], [x23], #0x4\n" + "st1 { v28.s }[2], [x22], #0x4\n" "tbz x9, #1, 111f\n" "st1 { v16.h }[6], [x27], #0x2\n" - "st1 { v20.h }[6], [x23], #0x2\n" - "st1 { v24.h }[6], [x22], #0x2\n" - "st1 { v28.h }[6], [x21], #0x2\n" + "st1 { v20.h }[6], [x24], #0x2\n" + "st1 { v24.h }[6], [x23], #0x2\n" + "st1 { v28.h }[6], [x22], #0x2\n" "tbz x9, #0, 118f\n" "st1 { v16.b }[14], [x27]\n" - "st1 { v20.b }[14], [x23]\n" - "st1 { v24.b }[14], [x22]\n" - "st1 { v28.b }[14], [x21]\n" + "st1 { v20.b }[14], [x24]\n" + "st1 { v24.b }[14], [x23]\n" + "st1 { v28.b }[14], [x22]\n" "b 118f\n" "111:" // Height 4: Partial direct writeback: partial_1_12 "tbz x9, #0, 118f\n" "st1 { v16.b }[12], [x27]\n" - "st1 { v20.b }[12], [x23]\n" - "st1 { v24.b }[12], [x22]\n" - "st1 { v28.b }[12], [x21]\n" + "st1 { v20.b }[12], [x24]\n" + "st1 { v24.b }[12], [x23]\n" + "st1 { v28.b }[12], [x22]\n" "b 118f\n" "112:" // Height 4: Partial direct writeback: partial_2_8 "tbz x9, #1, 113f\n" "st1 { v16.h }[4], [x27], #0x2\n" - "st1 { v20.h }[4], [x23], #0x2\n" - "st1 { v24.h }[4], [x22], #0x2\n" - "st1 { v28.h }[4], [x21], #0x2\n" + "st1 { v20.h }[4], [x24], #0x2\n" + "st1 { v24.h }[4], [x23], #0x2\n" + "st1 { v28.h }[4], [x22], #0x2\n" "tbz x9, #0, 118f\n" "st1 { v16.b }[10], [x27]\n" - "st1 { v20.b }[10], [x23]\n" - "st1 { v24.b }[10], [x22]\n" - "st1 { v28.b }[10], [x21]\n" + "st1 { v20.b }[10], [x24]\n" + "st1 { v24.b }[10], [x23]\n" + "st1 { v28.b }[10], [x22]\n" "b 118f\n" "113:" // Height 4: Partial direct writeback: partial_1_8 "tbz x9, #0, 118f\n" "st1 { v16.b }[8], [x27]\n" - "st1 { v20.b }[8], [x23]\n" - "st1 { v24.b }[8], [x22]\n" - "st1 { v28.b }[8], [x21]\n" + "st1 { v20.b }[8], [x24]\n" + "st1 { v24.b }[8], [x23]\n" + "st1 { v28.b }[8], [x22]\n" "b 118f\n" "114:" // Height 4: Partial direct writeback: partial_4_0 "tbz x9, #2, 116f\n" "str s16, [x27], #0x4\n" - "str s20, [x23], #0x4\n" - "str s24, [x22], #0x4\n" - "str s28, [x21], #0x4\n" + "str s20, [x24], #0x4\n" + "str s24, [x23], #0x4\n" + "str s28, [x22], #0x4\n" "tbz x9, #1, 115f\n" "st1 { v16.h }[2], [x27], #0x2\n" - "st1 { v20.h }[2], [x23], #0x2\n" - "st1 { v24.h }[2], [x22], #0x2\n" - "st1 { v28.h }[2], [x21], #0x2\n" + "st1 { v20.h }[2], [x24], #0x2\n" + "st1 { v24.h }[2], [x23], #0x2\n" + "st1 { v28.h }[2], [x22], #0x2\n" "tbz x9, #0, 118f\n" "st1 { v16.b }[6], [x27]\n" - "st1 { v20.b }[6], [x23]\n" - "st1 { v24.b }[6], [x22]\n" - "st1 { v28.b }[6], [x21]\n" + "st1 { v20.b }[6], [x24]\n" + "st1 { v24.b }[6], [x23]\n" + "st1 { v28.b }[6], [x22]\n" "b 118f\n" "115:" // Height 4: Partial direct writeback: partial_1_4 "tbz x9, #0, 118f\n" "st1 { v16.b }[4], [x27]\n" - "st1 { v20.b }[4], [x23]\n" - "st1 { v24.b }[4], [x22]\n" - "st1 { v28.b }[4], [x21]\n" + "st1 { v20.b }[4], [x24]\n" + "st1 { v24.b }[4], [x23]\n" + "st1 { v28.b }[4], [x22]\n" "b 118f\n" "116:" // Height 4: Partial direct writeback: partial_2_0 "tbz x9, #1, 117f\n" "str h16, [x27], #0x2\n" - "str h20, [x23], #0x2\n" - "str h24, [x22], #0x2\n" - "str h28, [x21], #0x2\n" + "str h20, [x24], #0x2\n" + "str h24, [x23], #0x2\n" + "str h28, [x22], #0x2\n" "tbz x9, #0, 118f\n" "st1 { v16.b }[2], [x27]\n" - "st1 { v20.b }[2], [x23]\n" - "st1 { v24.b }[2], [x22]\n" - "st1 { v28.b }[2], [x21]\n" + "st1 { v20.b }[2], [x24]\n" + "st1 { v24.b }[2], [x23]\n" + "st1 { v28.b }[2], [x22]\n" "b 118f\n" "117:" // Height 4: Partial direct writeback: partial_1_0 "str b16, [x27, #0x0]\n" - "str b20, [x23, #0x0]\n" - "str b24, [x22, #0x0]\n" - "str b28, [x21, #0x0]\n" + "str b20, [x24, #0x0]\n" + "str b24, [x23, #0x0]\n" + "str b28, [x22, #0x0]\n" "118:" // Height 4: Partial direct writeback: Done "b 120f\n" "119:" // Height 4: Full writeback "str q16, [x27, #0x0]\n" "add x27, x27, #0x10\n" - "str q20, [x23, #0x0]\n" - "str q24, [x22, #0x0]\n" - "str q28, [x21, #0x0]\n" + "str q20, [x24, #0x0]\n" + "str q24, [x23, #0x0]\n" + "str q28, [x22, #0x0]\n" "120:" // Height 4: Writeback done "subs x9, x9, #0x10\n" "bgt 92b\n" @@ -2016,8 +2017,8 @@ void a64_hybrid_s8qa_dot_4x16 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "122:" // Exit - : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) - : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [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_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [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) + : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr) + : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [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_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_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", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |