diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_dot_4x16/generic.cpp | 403 |
1 files changed, 202 insertions, 201 deletions
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" ); } |