diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/interleaves')
4 files changed, 366 insertions, 366 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp index 3de4bdc1fb..f18208d6c4 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_s8q_3x3_dot.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -52,189 +52,189 @@ void interleave_a64_s8q_3x3_dot::pack_parameters(unsigned int n_channels, void * { __asm__ __volatile__( "cmp %x[ld_weight_col], XZR\n" - "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n" + "mov x22, #0x3\n" "movi v16.4s, #0x9\n" - "movi v31.16b, #0x0\n" - "mov x21, #0x3\n" - "mul x21, %x[ld_weight_col], x21\n" + "movi v0.16b, #0x0\n" + "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n" "add x20, %x[qp], %[offsetof_input_offset]\n" + "movi v31.16b, #0x1\n" "ld1r { v30.4s }, [x20]\n" "add x20, %x[qp], %[offsetof_weights_offset]\n" - "ld1r { v29.4s }, [x20]\n" "cmp %x[ld_weight_row], XZR\n" - "mul v29.4s, v29.4s, v30.4s\n" - "csel %x[ld_weight_row], %x[ld_weight_row], x21, NE\n" + "ld1r { v29.4s }, [x20]\n" "lsr x21, %x[n_channels], #0x2\n" - "movi v28.16b, #0x1\n" - "mul v29.4s, v29.4s, v16.4s\n" - "add x25, %x[weights], %x[ld_weight_row]\n" "add x20, %x[qp], %[offsetof_per_layer_mul]\n" - "ld1r { v27.4s }, [x20]\n" + "ld1r { v28.4s }, [x20]\n" + "mul x22, %x[ld_weight_col], x22\n" "add x20, %x[qp], %[offsetof_per_layer_right_shift]\n" - "ld1r { v26.4s }, [x20]\n" - "add x24, x25, %x[ld_weight_row]\n" - "add x23, %x[ld_weight_col], %x[ld_weight_col]\n" - "mov x22, #0x0\n" + "add x25, %x[ld_weight_col], %x[ld_weight_col]\n" + "ld1r { v27.4s }, [x20]\n" + "mul v29.4s, v29.4s, v30.4s\n" + "mov x24, #0x0\n" + "csel %x[ld_weight_row], %x[ld_weight_row], x22, NE\n" + "mul v29.4s, v29.4s, v16.4s\n" + "add x23, %x[weights], %x[ld_weight_row]\n" + "add x22, x23, %x[ld_weight_row]\n" "cbz x21, 4f\n" "1:" // Loop - "movi v25.4s, #0x0\n" + "movi v26.4s, #0x0\n" "cbz %x[bias], 2f\n" - "ldr q25, [%x[bias], x22]\n" + "ldr q26, [%x[bias], x24]\n" "2:" // Loop: Skip bias load - "ldr s19, [%x[weights], #0x0]\n" - "ldr s16, [%x[weights], %x[ld_weight_col]]\n" - "zip1 v17.16b, v16.16b, v31.16b\n" - "movi v21.4s, #0x0\n" - "ldr s16, [%x[weights], x23]\n" - "ldr s18, [x25, #0x0]\n" - "zip1 v16.16b, v19.16b, v16.16b\n" - "zip1 v20.16b, v16.16b, v17.16b\n" - "ldr s17, [x25, %x[ld_weight_col]]\n" - "ldr s16, [x25, x23]\n" - "zip1 v18.16b, v18.16b, v16.16b\n" - "zip1 v16.16b, v17.16b, v31.16b\n" - "ldr s17, [x24, #0x0]\n" - "ldr s19, [x24, %x[ld_weight_col]]\n" - ".inst 0x4e949795 // sdot v21.4s, v28.16b, v20.16b\n" - "zip1 v18.16b, v18.16b, v16.16b\n" - "ldr s16, [x24, x23]\n" - "zip1 v17.16b, v17.16b, v16.16b\n" - "zip1 v16.16b, v19.16b, v31.16b\n" - ".inst 0x4e929795 // sdot v21.4s, v28.16b, v18.16b\n" - "zip1 v16.16b, v17.16b, v16.16b\n" - ".inst 0x4e909795 // sdot v21.4s, v28.16b, v16.16b\n" + "ldr s25, [%x[weights], #0x0]\n" + "ldr s18, [%x[weights], %x[ld_weight_col]]\n" + "movi v24.4s, #0x0\n" + "ldr s16, [%x[weights], x25]\n" + "ldr s20, [x23, #0x0]\n" "add %x[weights], %x[weights], #0x4\n" - "add x25, x25, #0x4\n" - "mls v25.4s, v21.4s, v30.4s\n" - "add x24, x24, #0x4\n" - "add v25.4s, v25.4s, v29.4s\n" - "str q25, [%x[outptr], #0x0]\n" - "str q20, [%x[outptr], #0x10]\n" - "str q18, [%x[outptr], #0x20]\n" + "ldr s23, [x23, %x[ld_weight_col]]\n" + "ldr s17, [x23, x25]\n" + "add x23, x23, #0x4\n" + "ldr s22, [x22, #0x0]\n" + "ldr s19, [x22, %x[ld_weight_col]]\n" + "zip1 v18.16b, v18.16b, v0.16b\n" + "ldr s21, [x22, x25]\n" + "zip1 v16.16b, v25.16b, v16.16b\n" + "add x22, x22, #0x4\n" + "zip1 v20.16b, v20.16b, v17.16b\n" + "zip1 v17.16b, v23.16b, v0.16b\n" + "zip1 v19.16b, v19.16b, v0.16b\n" + "zip1 v18.16b, v16.16b, v18.16b\n" + "zip1 v16.16b, v22.16b, v21.16b\n" + "zip1 v17.16b, v20.16b, v17.16b\n" + ".inst 0x4e9297f8 // sdot v24.4s, v31.16b, v18.16b\n" + "zip1 v16.16b, v16.16b, v19.16b\n" + ".inst 0x4e9197f8 // sdot v24.4s, v31.16b, v17.16b\n" + ".inst 0x4e9097f8 // sdot v24.4s, v31.16b, v16.16b\n" + "mls v26.4s, v24.4s, v30.4s\n" + "add v26.4s, v26.4s, v29.4s\n" + "str q26, [%x[outptr], #0x0]\n" + "str q18, [%x[outptr], #0x10]\n" + "str q17, [%x[outptr], #0x20]\n" "str q16, [%x[outptr], #0x30]\n" "add %x[outptr], %x[outptr], #0x40\n" "cbz %x[rq_mul_perchannel], 3f\n" - "ldr q27, [%x[rq_mul_perchannel], x22]\n" - "ldr q26, [%x[rq_shift_perchannel], x22]\n" + "ldr q28, [%x[rq_mul_perchannel], x24]\n" + "ldr q27, [%x[rq_shift_perchannel], x24]\n" "3:" // Loop: Quantisation parameters: Store "subs x21, x21, #0x1\n" - "str q27, [%x[outptr], #0x0]\n" - "add x22, x22, #0x10\n" - "str q26, [%x[outptr], #0x10]\n" + "str q28, [%x[outptr], #0x0]\n" + "add x24, x24, #0x10\n" + "str q27, [%x[outptr], #0x10]\n" "add %x[outptr], %x[outptr], #0x20\n" "bgt 1b\n" "tst %x[n_channels], #0x3\n" "beq 13f\n" "4:" // Oddments - "movi v25.4s, #0x0\n" + "movi v26.4s, #0x0\n" "cbz %x[bias], 7f\n" - "add %x[bias], %x[bias], x22\n" + "add %x[bias], %x[bias], x24\n" "tbz %x[n_channels], #1, 5f\n" - "ld1 { v25.d }[0], [%x[bias]], #0x8\n" + "ld1 { v26.d }[0], [%x[bias]], #0x8\n" "tbz %x[n_channels], #0, 6f\n" - "ld1 { v25.s }[2], [%x[bias]], #0x4\n" + "ld1 { v26.s }[2], [%x[bias]], #0x4\n" "b 6f\n" "5:" // Oddments: Load bias: Bit 1: Unset - "ld1 { v25.s }[0], [%x[bias]], #0x4\n" + "ld1 { v26.s }[0], [%x[bias]], #0x4\n" "6:" // Oddments: Load bias: Bit 1: End "7:" // Oddments: Skip bias load "tbz %x[n_channels], #1, 8f\n" - "ld1 { v17.h }[0], [%x[weights]]\n" - "ld1 { v24.h }[0], [x25]\n" + "ld1 { v18.h }[0], [%x[weights]]\n" + "ld1 { v21.h }[0], [x23]\n" "add x21, %x[weights], %x[ld_weight_col]\n" - "add x20, %x[weights], x23\n" - "ld1 { v20.h }[0], [x21]\n" + "add x20, %x[weights], x25\n" + "ld1 { v24.h }[0], [x21]\n" "ld1 { v16.h }[0], [x20]\n" - "add x21, x25, %x[ld_weight_col]\n" - "add x20, x25, x23\n" - "ld1 { v19.h }[0], [x21]\n" - "ld1 { v18.h }[0], [x20]\n" - "add x21, x24, %x[ld_weight_col]\n" - "add x20, x24, x23\n" - "ld1 { v23.h }[0], [x24]\n" + "add x21, x23, %x[ld_weight_col]\n" + "add x20, x23, x25\n" + "ld1 { v20.h }[0], [x21]\n" + "ld1 { v17.h }[0], [x20]\n" + "add x21, x22, %x[ld_weight_col]\n" + "add x20, x22, x25\n" + "ld1 { v23.h }[0], [x22]\n" "ld1 { v22.h }[0], [x21]\n" "add %x[weights], %x[weights], #0x2\n" - "add x25, x25, #0x2\n" - "ld1 { v21.h }[0], [x20]\n" - "add x24, x24, #0x2\n" + "add x23, x23, #0x2\n" + "ld1 { v19.h }[0], [x20]\n" + "add x22, x22, #0x2\n" "tbz %x[n_channels], #0, 9f\n" - "ld1 { v17.b }[2], [%x[weights]]\n" - "ld1 { v24.b }[2], [x25]\n" + "ld1 { v18.b }[2], [%x[weights]]\n" + "ld1 { v21.b }[2], [x23]\n" "add x21, %x[weights], %x[ld_weight_col]\n" - "add x20, %x[weights], x23\n" - "ld1 { v20.b }[2], [x21]\n" + "add x20, %x[weights], x25\n" + "ld1 { v24.b }[2], [x21]\n" "ld1 { v16.b }[2], [x20]\n" - "add x21, x25, %x[ld_weight_col]\n" - "add x20, x25, x23\n" - "ld1 { v19.b }[2], [x21]\n" - "ld1 { v18.b }[2], [x20]\n" - "add x21, x24, %x[ld_weight_col]\n" - "add x20, x24, x23\n" - "ld1 { v23.b }[2], [x24]\n" + "add x21, x23, %x[ld_weight_col]\n" + "add x20, x23, x25\n" + "ld1 { v20.b }[2], [x21]\n" + "ld1 { v17.b }[2], [x20]\n" + "add x21, x22, %x[ld_weight_col]\n" + "add x20, x22, x25\n" + "ld1 { v23.b }[2], [x22]\n" "ld1 { v22.b }[2], [x21]\n" "add %x[weights], %x[weights], #0x1\n" - "ld1 { v21.b }[2], [x20]\n" + "ld1 { v19.b }[2], [x20]\n" "b 9f\n" "8:" // Oddments: Load weights: Bit 1: Unset - "ld1 { v17.b }[0], [%x[weights]]\n" - "ld1 { v24.b }[0], [x25]\n" + "ld1 { v18.b }[0], [%x[weights]]\n" + "ld1 { v21.b }[0], [x23]\n" "add x21, %x[weights], %x[ld_weight_col]\n" - "add x20, %x[weights], x23\n" - "ld1 { v20.b }[0], [x21]\n" + "add x20, %x[weights], x25\n" + "ld1 { v24.b }[0], [x21]\n" "ld1 { v16.b }[0], [x20]\n" - "add x21, x25, %x[ld_weight_col]\n" - "add x20, x25, x23\n" - "ld1 { v19.b }[0], [x21]\n" - "ld1 { v18.b }[0], [x20]\n" - "add x21, x24, %x[ld_weight_col]\n" - "add x20, x24, x23\n" - "ld1 { v23.b }[0], [x24]\n" + "add x21, x23, %x[ld_weight_col]\n" + "add x20, x23, x25\n" + "ld1 { v20.b }[0], [x21]\n" + "ld1 { v17.b }[0], [x20]\n" + "add x21, x22, %x[ld_weight_col]\n" + "add x20, x22, x25\n" + "ld1 { v23.b }[0], [x22]\n" "ld1 { v22.b }[0], [x21]\n" "add %x[weights], %x[weights], #0x1\n" - "ld1 { v21.b }[0], [x20]\n" + "ld1 { v19.b }[0], [x20]\n" "9:" // Oddments: Load weights: Bit 1: End - "zip1 v17.16b, v17.16b, v16.16b\n" - "zip1 v16.16b, v20.16b, v31.16b\n" - "zip1 v20.16b, v17.16b, v16.16b\n" - "zip1 v17.16b, v24.16b, v18.16b\n" - "zip1 v16.16b, v19.16b, v31.16b\n" - "movi v19.4s, #0x0\n" - ".inst 0x4e949793 // sdot v19.4s, v28.16b, v20.16b\n" - "zip1 v18.16b, v17.16b, v16.16b\n" - "zip1 v17.16b, v23.16b, v21.16b\n" - ".inst 0x4e929793 // sdot v19.4s, v28.16b, v18.16b\n" - "zip1 v16.16b, v22.16b, v31.16b\n" - "zip1 v16.16b, v17.16b, v16.16b\n" - ".inst 0x4e909793 // sdot v19.4s, v28.16b, v16.16b\n" - "mls v25.4s, v19.4s, v30.4s\n" - "add v25.4s, v25.4s, v29.4s\n" - "str q25, [%x[outptr], #0x0]\n" - "str q20, [%x[outptr], #0x10]\n" - "str q18, [%x[outptr], #0x20]\n" + "zip1 v18.16b, v18.16b, v16.16b\n" + "zip1 v16.16b, v24.16b, v0.16b\n" + "zip1 v21.16b, v21.16b, v17.16b\n" + "zip1 v17.16b, v20.16b, v0.16b\n" + "movi v20.4s, #0x0\n" + "zip1 v19.16b, v23.16b, v19.16b\n" + "zip1 v18.16b, v18.16b, v16.16b\n" + "zip1 v16.16b, v22.16b, v0.16b\n" + "zip1 v17.16b, v21.16b, v17.16b\n" + ".inst 0x4e9297f4 // sdot v20.4s, v31.16b, v18.16b\n" + "zip1 v16.16b, v19.16b, v16.16b\n" + ".inst 0x4e9197f4 // sdot v20.4s, v31.16b, v17.16b\n" + ".inst 0x4e9097f4 // sdot v20.4s, v31.16b, v16.16b\n" + "mls v26.4s, v20.4s, v30.4s\n" + "add v26.4s, v26.4s, v29.4s\n" + "str q26, [%x[outptr], #0x0]\n" + "str q18, [%x[outptr], #0x10]\n" + "str q17, [%x[outptr], #0x20]\n" "str q16, [%x[outptr], #0x30]\n" "add %x[outptr], %x[outptr], #0x40\n" "cbz %x[rq_mul_perchannel], 12f\n" - "add x21, %x[rq_mul_perchannel], x22\n" - "add x20, %x[rq_shift_perchannel], x22\n" + "add x21, %x[rq_mul_perchannel], x24\n" + "add x20, %x[rq_shift_perchannel], x24\n" "tbz %x[n_channels], #1, 10f\n" - "ld1 { v27.d }[0], [x21], #0x8\n" - "ld1 { v26.d }[0], [x20], #0x8\n" + "ld1 { v28.d }[0], [x21], #0x8\n" + "ld1 { v27.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 11f\n" - "ld1 { v27.s }[2], [x21], #0x4\n" - "ld1 { v26.s }[2], [x20], #0x4\n" + "ld1 { v28.s }[2], [x21], #0x4\n" + "ld1 { v27.s }[2], [x20], #0x4\n" "b 11f\n" "10:" // Oddments: Quantisation parameters: Load quant params: Bit 1: Unset - "ld1 { v27.s }[0], [x21], #0x4\n" - "ld1 { v26.s }[0], [x20], #0x4\n" + "ld1 { v28.s }[0], [x21], #0x4\n" + "ld1 { v27.s }[0], [x20], #0x4\n" "11:" // Oddments: Quantisation parameters: Load quant params: Bit 1: End "12:" // Oddments: Quantisation parameters: Store - "str q27, [%x[outptr], #0x0]\n" - "str q26, [%x[outptr], #0x10]\n" + "str q28, [%x[outptr], #0x0]\n" + "str q27, [%x[outptr], #0x10]\n" "add %x[outptr], %x[outptr], #0x20\n" "13:" // End : [bias] "+&r" (bias), [ld_weight_col] "+&r" (ld_weight_col), [ld_weight_row] "+&r" (ld_weight_row), [outptr] "+&r" (outptr), [weights] "+&r" (weights) : [n_channels] "r" (n_channels), [offsetof_input_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_per_layer_mul] "I" (offsetof(arm_gemm::Requantize32, per_layer_mul)), [offsetof_per_layer_right_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_right_shift)), [offsetof_weights_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [qp] "r" (&qp), [rq_mul_perchannel] "r" (qp.per_channel_muls), [rq_shift_perchannel] "r" (qp.per_channel_right_shifts) - : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25" + : "cc", "memory", "v0", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25" ); } diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp index 19264c9fce..0ebf6ac10f 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/a64_u8q_3x3_dot.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -52,189 +52,189 @@ void interleave_a64_u8q_3x3_dot::pack_parameters(unsigned int n_channels, void * { __asm__ __volatile__( "cmp %x[ld_weight_col], XZR\n" - "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n" + "mov x22, #0x3\n" "movi v16.4s, #0x9\n" - "movi v31.16b, #0x0\n" - "mov x21, #0x3\n" - "mul x21, %x[ld_weight_col], x21\n" + "movi v0.16b, #0x0\n" + "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n" "add x20, %x[qp], %[offsetof_input_offset]\n" + "movi v31.16b, #0x1\n" "ld1r { v30.4s }, [x20]\n" "add x20, %x[qp], %[offsetof_weights_offset]\n" - "ld1r { v29.4s }, [x20]\n" "cmp %x[ld_weight_row], XZR\n" - "mul v29.4s, v29.4s, v30.4s\n" - "csel %x[ld_weight_row], %x[ld_weight_row], x21, NE\n" + "ld1r { v29.4s }, [x20]\n" "lsr x21, %x[n_channels], #0x2\n" - "movi v28.16b, #0x1\n" - "mul v29.4s, v29.4s, v16.4s\n" - "add x25, %x[weights], %x[ld_weight_row]\n" "add x20, %x[qp], %[offsetof_per_layer_mul]\n" - "ld1r { v27.4s }, [x20]\n" + "ld1r { v28.4s }, [x20]\n" + "mul x22, %x[ld_weight_col], x22\n" "add x20, %x[qp], %[offsetof_per_layer_right_shift]\n" - "ld1r { v26.4s }, [x20]\n" - "add x24, x25, %x[ld_weight_row]\n" - "add x23, %x[ld_weight_col], %x[ld_weight_col]\n" - "mov x22, #0x0\n" + "add x25, %x[ld_weight_col], %x[ld_weight_col]\n" + "ld1r { v27.4s }, [x20]\n" + "mul v29.4s, v29.4s, v30.4s\n" + "mov x24, #0x0\n" + "csel %x[ld_weight_row], %x[ld_weight_row], x22, NE\n" + "mul v29.4s, v29.4s, v16.4s\n" + "add x23, %x[weights], %x[ld_weight_row]\n" + "add x22, x23, %x[ld_weight_row]\n" "cbz x21, 4f\n" "1:" // Loop - "movi v25.4s, #0x0\n" + "movi v26.4s, #0x0\n" "cbz %x[bias], 2f\n" - "ldr q25, [%x[bias], x22]\n" + "ldr q26, [%x[bias], x24]\n" "2:" // Loop: Skip bias load - "ldr s19, [%x[weights], #0x0]\n" - "ldr s16, [%x[weights], %x[ld_weight_col]]\n" - "zip1 v17.16b, v16.16b, v31.16b\n" - "movi v21.4s, #0x0\n" - "ldr s16, [%x[weights], x23]\n" - "ldr s18, [x25, #0x0]\n" - "zip1 v16.16b, v19.16b, v16.16b\n" - "zip1 v20.16b, v16.16b, v17.16b\n" - "ldr s17, [x25, %x[ld_weight_col]]\n" - "ldr s16, [x25, x23]\n" - "zip1 v18.16b, v18.16b, v16.16b\n" - "zip1 v16.16b, v17.16b, v31.16b\n" - "ldr s17, [x24, #0x0]\n" - "ldr s19, [x24, %x[ld_weight_col]]\n" - ".inst 0x6e949795 // udot v21.4s, v28.16b, v20.16b\n" - "zip1 v18.16b, v18.16b, v16.16b\n" - "ldr s16, [x24, x23]\n" - "zip1 v17.16b, v17.16b, v16.16b\n" - "zip1 v16.16b, v19.16b, v31.16b\n" - ".inst 0x6e929795 // udot v21.4s, v28.16b, v18.16b\n" - "zip1 v16.16b, v17.16b, v16.16b\n" - ".inst 0x6e909795 // udot v21.4s, v28.16b, v16.16b\n" + "ldr s25, [%x[weights], #0x0]\n" + "ldr s18, [%x[weights], %x[ld_weight_col]]\n" + "movi v24.4s, #0x0\n" + "ldr s16, [%x[weights], x25]\n" + "ldr s20, [x23, #0x0]\n" "add %x[weights], %x[weights], #0x4\n" - "add x25, x25, #0x4\n" - "mls v25.4s, v21.4s, v30.4s\n" - "add x24, x24, #0x4\n" - "add v25.4s, v25.4s, v29.4s\n" - "str q25, [%x[outptr], #0x0]\n" - "str q20, [%x[outptr], #0x10]\n" - "str q18, [%x[outptr], #0x20]\n" + "ldr s23, [x23, %x[ld_weight_col]]\n" + "ldr s17, [x23, x25]\n" + "add x23, x23, #0x4\n" + "ldr s22, [x22, #0x0]\n" + "ldr s19, [x22, %x[ld_weight_col]]\n" + "zip1 v18.16b, v18.16b, v0.16b\n" + "ldr s21, [x22, x25]\n" + "zip1 v16.16b, v25.16b, v16.16b\n" + "add x22, x22, #0x4\n" + "zip1 v20.16b, v20.16b, v17.16b\n" + "zip1 v17.16b, v23.16b, v0.16b\n" + "zip1 v19.16b, v19.16b, v0.16b\n" + "zip1 v18.16b, v16.16b, v18.16b\n" + "zip1 v16.16b, v22.16b, v21.16b\n" + "zip1 v17.16b, v20.16b, v17.16b\n" + ".inst 0x6e9297f8 // udot v24.4s, v31.16b, v18.16b\n" + "zip1 v16.16b, v16.16b, v19.16b\n" + ".inst 0x6e9197f8 // udot v24.4s, v31.16b, v17.16b\n" + ".inst 0x6e9097f8 // udot v24.4s, v31.16b, v16.16b\n" + "mls v26.4s, v24.4s, v30.4s\n" + "add v26.4s, v26.4s, v29.4s\n" + "str q26, [%x[outptr], #0x0]\n" + "str q18, [%x[outptr], #0x10]\n" + "str q17, [%x[outptr], #0x20]\n" "str q16, [%x[outptr], #0x30]\n" "add %x[outptr], %x[outptr], #0x40\n" "cbz %x[rq_mul_perchannel], 3f\n" - "ldr q27, [%x[rq_mul_perchannel], x22]\n" - "ldr q26, [%x[rq_shift_perchannel], x22]\n" + "ldr q28, [%x[rq_mul_perchannel], x24]\n" + "ldr q27, [%x[rq_shift_perchannel], x24]\n" "3:" // Loop: Quantisation parameters: Store "subs x21, x21, #0x1\n" - "str q27, [%x[outptr], #0x0]\n" - "add x22, x22, #0x10\n" - "str q26, [%x[outptr], #0x10]\n" + "str q28, [%x[outptr], #0x0]\n" + "add x24, x24, #0x10\n" + "str q27, [%x[outptr], #0x10]\n" "add %x[outptr], %x[outptr], #0x20\n" "bgt 1b\n" "tst %x[n_channels], #0x3\n" "beq 13f\n" "4:" // Oddments - "movi v25.4s, #0x0\n" + "movi v26.4s, #0x0\n" "cbz %x[bias], 7f\n" - "add %x[bias], %x[bias], x22\n" + "add %x[bias], %x[bias], x24\n" "tbz %x[n_channels], #1, 5f\n" - "ld1 { v25.d }[0], [%x[bias]], #0x8\n" + "ld1 { v26.d }[0], [%x[bias]], #0x8\n" "tbz %x[n_channels], #0, 6f\n" - "ld1 { v25.s }[2], [%x[bias]], #0x4\n" + "ld1 { v26.s }[2], [%x[bias]], #0x4\n" "b 6f\n" "5:" // Oddments: Load bias: Bit 1: Unset - "ld1 { v25.s }[0], [%x[bias]], #0x4\n" + "ld1 { v26.s }[0], [%x[bias]], #0x4\n" "6:" // Oddments: Load bias: Bit 1: End "7:" // Oddments: Skip bias load "tbz %x[n_channels], #1, 8f\n" - "ld1 { v17.h }[0], [%x[weights]]\n" - "ld1 { v24.h }[0], [x25]\n" + "ld1 { v18.h }[0], [%x[weights]]\n" + "ld1 { v21.h }[0], [x23]\n" "add x21, %x[weights], %x[ld_weight_col]\n" - "add x20, %x[weights], x23\n" - "ld1 { v20.h }[0], [x21]\n" + "add x20, %x[weights], x25\n" + "ld1 { v24.h }[0], [x21]\n" "ld1 { v16.h }[0], [x20]\n" - "add x21, x25, %x[ld_weight_col]\n" - "add x20, x25, x23\n" - "ld1 { v19.h }[0], [x21]\n" - "ld1 { v18.h }[0], [x20]\n" - "add x21, x24, %x[ld_weight_col]\n" - "add x20, x24, x23\n" - "ld1 { v23.h }[0], [x24]\n" + "add x21, x23, %x[ld_weight_col]\n" + "add x20, x23, x25\n" + "ld1 { v20.h }[0], [x21]\n" + "ld1 { v17.h }[0], [x20]\n" + "add x21, x22, %x[ld_weight_col]\n" + "add x20, x22, x25\n" + "ld1 { v23.h }[0], [x22]\n" "ld1 { v22.h }[0], [x21]\n" "add %x[weights], %x[weights], #0x2\n" - "add x25, x25, #0x2\n" - "ld1 { v21.h }[0], [x20]\n" - "add x24, x24, #0x2\n" + "add x23, x23, #0x2\n" + "ld1 { v19.h }[0], [x20]\n" + "add x22, x22, #0x2\n" "tbz %x[n_channels], #0, 9f\n" - "ld1 { v17.b }[2], [%x[weights]]\n" - "ld1 { v24.b }[2], [x25]\n" + "ld1 { v18.b }[2], [%x[weights]]\n" + "ld1 { v21.b }[2], [x23]\n" "add x21, %x[weights], %x[ld_weight_col]\n" - "add x20, %x[weights], x23\n" - "ld1 { v20.b }[2], [x21]\n" + "add x20, %x[weights], x25\n" + "ld1 { v24.b }[2], [x21]\n" "ld1 { v16.b }[2], [x20]\n" - "add x21, x25, %x[ld_weight_col]\n" - "add x20, x25, x23\n" - "ld1 { v19.b }[2], [x21]\n" - "ld1 { v18.b }[2], [x20]\n" - "add x21, x24, %x[ld_weight_col]\n" - "add x20, x24, x23\n" - "ld1 { v23.b }[2], [x24]\n" + "add x21, x23, %x[ld_weight_col]\n" + "add x20, x23, x25\n" + "ld1 { v20.b }[2], [x21]\n" + "ld1 { v17.b }[2], [x20]\n" + "add x21, x22, %x[ld_weight_col]\n" + "add x20, x22, x25\n" + "ld1 { v23.b }[2], [x22]\n" "ld1 { v22.b }[2], [x21]\n" "add %x[weights], %x[weights], #0x1\n" - "ld1 { v21.b }[2], [x20]\n" + "ld1 { v19.b }[2], [x20]\n" "b 9f\n" "8:" // Oddments: Load weights: Bit 1: Unset - "ld1 { v17.b }[0], [%x[weights]]\n" - "ld1 { v24.b }[0], [x25]\n" + "ld1 { v18.b }[0], [%x[weights]]\n" + "ld1 { v21.b }[0], [x23]\n" "add x21, %x[weights], %x[ld_weight_col]\n" - "add x20, %x[weights], x23\n" - "ld1 { v20.b }[0], [x21]\n" + "add x20, %x[weights], x25\n" + "ld1 { v24.b }[0], [x21]\n" "ld1 { v16.b }[0], [x20]\n" - "add x21, x25, %x[ld_weight_col]\n" - "add x20, x25, x23\n" - "ld1 { v19.b }[0], [x21]\n" - "ld1 { v18.b }[0], [x20]\n" - "add x21, x24, %x[ld_weight_col]\n" - "add x20, x24, x23\n" - "ld1 { v23.b }[0], [x24]\n" + "add x21, x23, %x[ld_weight_col]\n" + "add x20, x23, x25\n" + "ld1 { v20.b }[0], [x21]\n" + "ld1 { v17.b }[0], [x20]\n" + "add x21, x22, %x[ld_weight_col]\n" + "add x20, x22, x25\n" + "ld1 { v23.b }[0], [x22]\n" "ld1 { v22.b }[0], [x21]\n" "add %x[weights], %x[weights], #0x1\n" - "ld1 { v21.b }[0], [x20]\n" + "ld1 { v19.b }[0], [x20]\n" "9:" // Oddments: Load weights: Bit 1: End - "zip1 v17.16b, v17.16b, v16.16b\n" - "zip1 v16.16b, v20.16b, v31.16b\n" - "zip1 v20.16b, v17.16b, v16.16b\n" - "zip1 v17.16b, v24.16b, v18.16b\n" - "zip1 v16.16b, v19.16b, v31.16b\n" - "movi v19.4s, #0x0\n" - ".inst 0x6e949793 // udot v19.4s, v28.16b, v20.16b\n" - "zip1 v18.16b, v17.16b, v16.16b\n" - "zip1 v17.16b, v23.16b, v21.16b\n" - ".inst 0x6e929793 // udot v19.4s, v28.16b, v18.16b\n" - "zip1 v16.16b, v22.16b, v31.16b\n" - "zip1 v16.16b, v17.16b, v16.16b\n" - ".inst 0x6e909793 // udot v19.4s, v28.16b, v16.16b\n" - "mls v25.4s, v19.4s, v30.4s\n" - "add v25.4s, v25.4s, v29.4s\n" - "str q25, [%x[outptr], #0x0]\n" - "str q20, [%x[outptr], #0x10]\n" - "str q18, [%x[outptr], #0x20]\n" + "zip1 v18.16b, v18.16b, v16.16b\n" + "zip1 v16.16b, v24.16b, v0.16b\n" + "zip1 v21.16b, v21.16b, v17.16b\n" + "zip1 v17.16b, v20.16b, v0.16b\n" + "movi v20.4s, #0x0\n" + "zip1 v19.16b, v23.16b, v19.16b\n" + "zip1 v18.16b, v18.16b, v16.16b\n" + "zip1 v16.16b, v22.16b, v0.16b\n" + "zip1 v17.16b, v21.16b, v17.16b\n" + ".inst 0x6e9297f4 // udot v20.4s, v31.16b, v18.16b\n" + "zip1 v16.16b, v19.16b, v16.16b\n" + ".inst 0x6e9197f4 // udot v20.4s, v31.16b, v17.16b\n" + ".inst 0x6e9097f4 // udot v20.4s, v31.16b, v16.16b\n" + "mls v26.4s, v20.4s, v30.4s\n" + "add v26.4s, v26.4s, v29.4s\n" + "str q26, [%x[outptr], #0x0]\n" + "str q18, [%x[outptr], #0x10]\n" + "str q17, [%x[outptr], #0x20]\n" "str q16, [%x[outptr], #0x30]\n" "add %x[outptr], %x[outptr], #0x40\n" "cbz %x[rq_mul_perchannel], 12f\n" - "add x21, %x[rq_mul_perchannel], x22\n" - "add x20, %x[rq_shift_perchannel], x22\n" + "add x21, %x[rq_mul_perchannel], x24\n" + "add x20, %x[rq_shift_perchannel], x24\n" "tbz %x[n_channels], #1, 10f\n" - "ld1 { v27.d }[0], [x21], #0x8\n" - "ld1 { v26.d }[0], [x20], #0x8\n" + "ld1 { v28.d }[0], [x21], #0x8\n" + "ld1 { v27.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 11f\n" - "ld1 { v27.s }[2], [x21], #0x4\n" - "ld1 { v26.s }[2], [x20], #0x4\n" + "ld1 { v28.s }[2], [x21], #0x4\n" + "ld1 { v27.s }[2], [x20], #0x4\n" "b 11f\n" "10:" // Oddments: Quantisation parameters: Load quant params: Bit 1: Unset - "ld1 { v27.s }[0], [x21], #0x4\n" - "ld1 { v26.s }[0], [x20], #0x4\n" + "ld1 { v28.s }[0], [x21], #0x4\n" + "ld1 { v27.s }[0], [x20], #0x4\n" "11:" // Oddments: Quantisation parameters: Load quant params: Bit 1: End "12:" // Oddments: Quantisation parameters: Store - "str q27, [%x[outptr], #0x0]\n" - "str q26, [%x[outptr], #0x10]\n" + "str q28, [%x[outptr], #0x0]\n" + "str q27, [%x[outptr], #0x10]\n" "add %x[outptr], %x[outptr], #0x20\n" "13:" // End : [bias] "+&r" (bias), [ld_weight_col] "+&r" (ld_weight_col), [ld_weight_row] "+&r" (ld_weight_row), [outptr] "+&r" (outptr), [weights] "+&r" (weights) : [n_channels] "r" (n_channels), [offsetof_input_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_per_layer_mul] "I" (offsetof(arm_gemm::Requantize32, per_layer_mul)), [offsetof_per_layer_right_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_right_shift)), [offsetof_weights_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [qp] "r" (&qp), [rq_mul_perchannel] "r" (qp.per_channel_muls), [rq_shift_perchannel] "r" (qp.per_channel_right_shifts) - : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25" + : "cc", "memory", "v0", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25" ); } diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp index 5d7b54f235..7364963477 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_s8q_3x3_dot.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -52,80 +52,80 @@ void interleave_sve_s8q_3x3_dot::pack_parameters(unsigned int n_channels, void * { __asm__ __volatile__( "cmp %x[ld_weight_col], XZR\n" - "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n" - "mov z16.s, #0x9\n" - "mov z28.b, #0x0\n" "mov x20, #0x3\n" - "ptrue p2.b\n" - "mul x20, %x[ld_weight_col], x20\n" - "ld1rw { z27.s }, p2/Z, [%x[qp], %[offsetof_input_offset]]\n" - "ld1rw { z26.s }, p2/Z, [%x[qp], %[offsetof_weights_offset]]\n" + "mov z16.s, #0x9\n" + "mov z31.b, #0x0\n" + "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n" + "ptrue p3.b\n" + "mov z30.b, #0x1\n" "cmp %x[ld_weight_row], XZR\n" - "csel %x[ld_weight_row], %x[ld_weight_row], x20, NE\n" - "mov z25.b, #0x1\n" - "mul z26.s, p2/M, z26.s, z27.s\n" - "add x24, %x[weights], %x[ld_weight_row]\n" - "ld1rw { z24.s }, p2/Z, [%x[qp], %[offsetof_per_layer_mul]]\n" - "ld1rw { z23.s }, p2/Z, [%x[qp], %[offsetof_per_layer_right_shift]]\n" - "add x23, x24, %x[ld_weight_row]\n" - "add x22, %x[ld_weight_col], %x[ld_weight_col]\n" - "whilelt p1.s, XZR, %x[n_channels]\n" - "mov x21, #0x0\n" - "mul z26.s, p2/M, z26.s, z16.s\n" + "mov x24, #0x0\n" "pfalse p8.b\n" + "mul x20, %x[ld_weight_col], x20\n" + "ld1rw { z29.s }, p3/Z, [%x[qp], %[offsetof_input_offset]]\n" + "ld1rw { z28.s }, p3/Z, [%x[qp], %[offsetof_weights_offset]]\n" + "add x23, %x[ld_weight_col], %x[ld_weight_col]\n" + "ld1rw { z27.s }, p3/Z, [%x[qp], %[offsetof_per_layer_mul]]\n" + "ld1rw { z26.s }, p3/Z, [%x[qp], %[offsetof_per_layer_right_shift]]\n" + "csel %x[ld_weight_row], %x[ld_weight_row], x20, NE\n" + "whilelt p2.s, XZR, %x[n_channels]\n" + "mul z28.s, p3/M, z28.s, z29.s\n" + "add x22, %x[weights], %x[ld_weight_row]\n" + "add x21, x22, %x[ld_weight_row]\n" + "mul z28.s, p3/M, z28.s, z16.s\n" "cbz %x[bias], 1f\n" "ptrue p8.s\n" "1:" // No bias "2:" // Loop - "cntp x20, p2, p1.s\n" + "cntp x20, p3, p2.s\n" + "mov z25.s, #0x0\n" + "and p1.b, p3/Z, p8.b, p2.b\n" "whilelt p0.b, XZR, x20\n" - "ld1b { z18.b }, p0/Z, [%x[weights]]\n" - "ld1b { z17.b }, p0/Z, [%x[weights], %x[ld_weight_col]]\n" - "ld1b { z16.b }, p0/Z, [%x[weights], x22]\n" - "zip1 z20.b, z18.b, z16.b\n" - "zip1 z19.b, z17.b, z28.b\n" - "ld1b { z18.b }, p0/Z, [x24]\n" - "ld1b { z17.b }, p0/Z, [x24, %x[ld_weight_col]]\n" - "ld1b { z16.b }, p0/Z, [x24, x22]\n" - "zip1 z22.b, z20.b, z19.b\n" - "zip1 z21.b, z18.b, z16.b\n" - "zip1 z19.b, z17.b, z28.b\n" - "mov z20.s, #0x0\n" - "ld1b { z18.b }, p0/Z, [x23]\n" - "ld1b { z17.b }, p0/Z, [x23, %x[ld_weight_col]]\n" - "ld1b { z16.b }, p0/Z, [x23, x22]\n" - "sdot z20.s, z25.b, z22.b\n" - "zip1 z19.b, z21.b, z19.b\n" - "sdot z20.s, z25.b, z19.b\n" - "zip1 z18.b, z18.b, z16.b\n" - "zip1 z16.b, z17.b, z28.b\n" - "and p0.b, p2/Z, p8.b, p1.b\n" - "ld1w { z17.s }, p0/Z, [%x[bias], x21, LSL #2]\n" - "zip1 z16.b, z18.b, z16.b\n" - "sdot z20.s, z25.b, z16.b\n" - "mls z17.s, p2/M, z20.s, z27.s\n" + "ld1w { z24.s }, p1/Z, [%x[bias], x24, LSL #2]\n" + "ld1b { z19.b }, p0/Z, [%x[weights]]\n" + "ld1b { z18.b }, p0/Z, [%x[weights], %x[ld_weight_col]]\n" + "ld1b { z16.b }, p0/Z, [%x[weights], x23]\n" "add %x[weights], %x[weights], x20\n" - "add x24, x24, x20\n" - "add x23, x23, x20\n" - "add z17.s, z17.s, z26.s\n" - "st1w { z17.s }, p2, [%x[outptr]]\n" - "st1b { z22.b }, p2, [%x[outptr], #1, MUL VL]\n" - "st1b { z19.b }, p2, [%x[outptr], #2, MUL VL]\n" - "st1b { z16.b }, p2, [%x[outptr], #3, MUL VL]\n" + "ld1b { z23.b }, p0/Z, [x22]\n" + "ld1b { z20.b }, p0/Z, [x22, %x[ld_weight_col]]\n" + "ld1b { z17.b }, p0/Z, [x22, x23]\n" + "ld1b { z22.b }, p0/Z, [x21]\n" + "add x22, x22, x20\n" + "zip1 z19.b, z19.b, z16.b\n" + "zip1 z18.b, z18.b, z31.b\n" + "ld1b { z21.b }, p0/Z, [x21, %x[ld_weight_col]]\n" + "ld1b { z16.b }, p0/Z, [x21, x23]\n" + "add x21, x21, x20\n" + "zip1 z20.b, z20.b, z31.b\n" + "zip1 z17.b, z23.b, z17.b\n" + "zip1 z19.b, z19.b, z18.b\n" + "zip1 z18.b, z22.b, z16.b\n" + "zip1 z16.b, z21.b, z31.b\n" + "zip1 z17.b, z17.b, z20.b\n" + "sdot z25.s, z30.b, z19.b\n" + "zip1 z16.b, z18.b, z16.b\n" + "sdot z25.s, z30.b, z17.b\n" + "sdot z25.s, z30.b, z16.b\n" + "mls z24.s, p3/M, z25.s, z29.s\n" + "add z24.s, z24.s, z28.s\n" + "st1w { z24.s }, p3, [%x[outptr]]\n" + "st1b { z19.b }, p3, [%x[outptr], #1, MUL VL]\n" + "st1b { z17.b }, p3, [%x[outptr], #2, MUL VL]\n" + "st1b { z16.b }, p3, [%x[outptr], #3, MUL VL]\n" "addvl %x[outptr], %x[outptr], #4\n" "cbz %x[rq_mul_perchannel], 3f\n" - "ld1w { z24.s }, p1/Z, [%x[rq_mul_perchannel], x21, LSL #2]\n" - "ld1w { z23.s }, p1/Z, [%x[rq_shift_perchannel], x21, LSL #2]\n" + "ld1w { z27.s }, p2/Z, [%x[rq_mul_perchannel], x24, LSL #2]\n" + "ld1w { z26.s }, p2/Z, [%x[rq_shift_perchannel], x24, LSL #2]\n" "3:" // Loop: Quantisation parameters: Store - "incw x21\n" - "whilelt p1.s, x21, %x[n_channels]\n" - "st1w { z24.s }, p2, [%x[outptr]]\n" - "st1w { z23.s }, p2, [%x[outptr], #1, MUL VL]\n" + "incw x24\n" + "st1w { z27.s }, p3, [%x[outptr]]\n" + "st1w { z26.s }, p3, [%x[outptr], #1, MUL VL]\n" "addvl %x[outptr], %x[outptr], #2\n" + "whilelt p2.s, x24, %x[n_channels]\n" "b.any 2b\n" : [ld_weight_col] "+&r" (ld_weight_col), [ld_weight_row] "+&r" (ld_weight_row), [outptr] "+&r" (outptr), [weights] "+&r" (weights) : [bias] "r" (bias), [n_channels] "r" (n_channels), [offsetof_input_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_per_layer_mul] "I" (offsetof(arm_gemm::Requantize32, per_layer_mul)), [offsetof_per_layer_right_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_right_shift)), [offsetof_weights_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [qp] "r" (&qp), [rq_mul_perchannel] "r" (qp.per_channel_muls), [rq_shift_perchannel] "r" (qp.per_channel_right_shifts) - : "cc", "memory", "p0", "p1", "p2", "p8", "x20", "x21", "x22", "x23", "x24", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28" + : "cc", "memory", "p0", "p1", "p2", "p3", "p8", "x20", "x21", "x22", "x23", "x24", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp index c3da81448b..e1b01663f6 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/sve_u8q_3x3_dot.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -52,80 +52,80 @@ void interleave_sve_u8q_3x3_dot::pack_parameters(unsigned int n_channels, void * { __asm__ __volatile__( "cmp %x[ld_weight_col], XZR\n" - "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n" - "mov z16.s, #0x9\n" - "mov z28.b, #0x0\n" "mov x20, #0x3\n" - "ptrue p2.b\n" - "mul x20, %x[ld_weight_col], x20\n" - "ld1rw { z27.s }, p2/Z, [%x[qp], %[offsetof_input_offset]]\n" - "ld1rw { z26.s }, p2/Z, [%x[qp], %[offsetof_weights_offset]]\n" + "mov z16.s, #0x9\n" + "mov z31.b, #0x0\n" + "csel %x[ld_weight_col], %x[ld_weight_col], %x[n_channels], NE\n" + "ptrue p3.b\n" + "mov z30.b, #0x1\n" "cmp %x[ld_weight_row], XZR\n" - "csel %x[ld_weight_row], %x[ld_weight_row], x20, NE\n" - "mov z25.b, #0x1\n" - "mul z26.s, p2/M, z26.s, z27.s\n" - "add x24, %x[weights], %x[ld_weight_row]\n" - "ld1rw { z24.s }, p2/Z, [%x[qp], %[offsetof_per_layer_mul]]\n" - "ld1rw { z23.s }, p2/Z, [%x[qp], %[offsetof_per_layer_right_shift]]\n" - "add x23, x24, %x[ld_weight_row]\n" - "add x22, %x[ld_weight_col], %x[ld_weight_col]\n" - "whilelt p1.s, XZR, %x[n_channels]\n" - "mov x21, #0x0\n" - "mul z26.s, p2/M, z26.s, z16.s\n" + "mov x24, #0x0\n" "pfalse p8.b\n" + "mul x20, %x[ld_weight_col], x20\n" + "ld1rw { z29.s }, p3/Z, [%x[qp], %[offsetof_input_offset]]\n" + "ld1rw { z28.s }, p3/Z, [%x[qp], %[offsetof_weights_offset]]\n" + "add x23, %x[ld_weight_col], %x[ld_weight_col]\n" + "ld1rw { z27.s }, p3/Z, [%x[qp], %[offsetof_per_layer_mul]]\n" + "ld1rw { z26.s }, p3/Z, [%x[qp], %[offsetof_per_layer_right_shift]]\n" + "csel %x[ld_weight_row], %x[ld_weight_row], x20, NE\n" + "whilelt p2.s, XZR, %x[n_channels]\n" + "mul z28.s, p3/M, z28.s, z29.s\n" + "add x22, %x[weights], %x[ld_weight_row]\n" + "add x21, x22, %x[ld_weight_row]\n" + "mul z28.s, p3/M, z28.s, z16.s\n" "cbz %x[bias], 1f\n" "ptrue p8.s\n" "1:" // No bias "2:" // Loop - "cntp x20, p2, p1.s\n" + "cntp x20, p3, p2.s\n" + "mov z25.s, #0x0\n" + "and p1.b, p3/Z, p8.b, p2.b\n" "whilelt p0.b, XZR, x20\n" - "ld1b { z18.b }, p0/Z, [%x[weights]]\n" - "ld1b { z17.b }, p0/Z, [%x[weights], %x[ld_weight_col]]\n" - "ld1b { z16.b }, p0/Z, [%x[weights], x22]\n" - "zip1 z20.b, z18.b, z16.b\n" - "zip1 z19.b, z17.b, z28.b\n" - "ld1b { z18.b }, p0/Z, [x24]\n" - "ld1b { z17.b }, p0/Z, [x24, %x[ld_weight_col]]\n" - "ld1b { z16.b }, p0/Z, [x24, x22]\n" - "zip1 z22.b, z20.b, z19.b\n" - "zip1 z21.b, z18.b, z16.b\n" - "zip1 z19.b, z17.b, z28.b\n" - "mov z20.s, #0x0\n" - "ld1b { z18.b }, p0/Z, [x23]\n" - "ld1b { z17.b }, p0/Z, [x23, %x[ld_weight_col]]\n" - "ld1b { z16.b }, p0/Z, [x23, x22]\n" - "udot z20.s, z25.b, z22.b\n" - "zip1 z19.b, z21.b, z19.b\n" - "udot z20.s, z25.b, z19.b\n" - "zip1 z18.b, z18.b, z16.b\n" - "zip1 z16.b, z17.b, z28.b\n" - "and p0.b, p2/Z, p8.b, p1.b\n" - "ld1w { z17.s }, p0/Z, [%x[bias], x21, LSL #2]\n" - "zip1 z16.b, z18.b, z16.b\n" - "udot z20.s, z25.b, z16.b\n" - "mls z17.s, p2/M, z20.s, z27.s\n" + "ld1w { z24.s }, p1/Z, [%x[bias], x24, LSL #2]\n" + "ld1b { z19.b }, p0/Z, [%x[weights]]\n" + "ld1b { z18.b }, p0/Z, [%x[weights], %x[ld_weight_col]]\n" + "ld1b { z16.b }, p0/Z, [%x[weights], x23]\n" "add %x[weights], %x[weights], x20\n" - "add x24, x24, x20\n" - "add x23, x23, x20\n" - "add z17.s, z17.s, z26.s\n" - "st1w { z17.s }, p2, [%x[outptr]]\n" - "st1b { z22.b }, p2, [%x[outptr], #1, MUL VL]\n" - "st1b { z19.b }, p2, [%x[outptr], #2, MUL VL]\n" - "st1b { z16.b }, p2, [%x[outptr], #3, MUL VL]\n" + "ld1b { z23.b }, p0/Z, [x22]\n" + "ld1b { z20.b }, p0/Z, [x22, %x[ld_weight_col]]\n" + "ld1b { z17.b }, p0/Z, [x22, x23]\n" + "ld1b { z22.b }, p0/Z, [x21]\n" + "add x22, x22, x20\n" + "zip1 z19.b, z19.b, z16.b\n" + "zip1 z18.b, z18.b, z31.b\n" + "ld1b { z21.b }, p0/Z, [x21, %x[ld_weight_col]]\n" + "ld1b { z16.b }, p0/Z, [x21, x23]\n" + "add x21, x21, x20\n" + "zip1 z20.b, z20.b, z31.b\n" + "zip1 z17.b, z23.b, z17.b\n" + "zip1 z19.b, z19.b, z18.b\n" + "zip1 z18.b, z22.b, z16.b\n" + "zip1 z16.b, z21.b, z31.b\n" + "zip1 z17.b, z17.b, z20.b\n" + "udot z25.s, z30.b, z19.b\n" + "zip1 z16.b, z18.b, z16.b\n" + "udot z25.s, z30.b, z17.b\n" + "udot z25.s, z30.b, z16.b\n" + "mls z24.s, p3/M, z25.s, z29.s\n" + "add z24.s, z24.s, z28.s\n" + "st1w { z24.s }, p3, [%x[outptr]]\n" + "st1b { z19.b }, p3, [%x[outptr], #1, MUL VL]\n" + "st1b { z17.b }, p3, [%x[outptr], #2, MUL VL]\n" + "st1b { z16.b }, p3, [%x[outptr], #3, MUL VL]\n" "addvl %x[outptr], %x[outptr], #4\n" "cbz %x[rq_mul_perchannel], 3f\n" - "ld1w { z24.s }, p1/Z, [%x[rq_mul_perchannel], x21, LSL #2]\n" - "ld1w { z23.s }, p1/Z, [%x[rq_shift_perchannel], x21, LSL #2]\n" + "ld1w { z27.s }, p2/Z, [%x[rq_mul_perchannel], x24, LSL #2]\n" + "ld1w { z26.s }, p2/Z, [%x[rq_shift_perchannel], x24, LSL #2]\n" "3:" // Loop: Quantisation parameters: Store - "incw x21\n" - "whilelt p1.s, x21, %x[n_channels]\n" - "st1w { z24.s }, p2, [%x[outptr]]\n" - "st1w { z23.s }, p2, [%x[outptr], #1, MUL VL]\n" + "incw x24\n" + "st1w { z27.s }, p3, [%x[outptr]]\n" + "st1w { z26.s }, p3, [%x[outptr], #1, MUL VL]\n" "addvl %x[outptr], %x[outptr], #2\n" + "whilelt p2.s, x24, %x[n_channels]\n" "b.any 2b\n" : [ld_weight_col] "+&r" (ld_weight_col), [ld_weight_row] "+&r" (ld_weight_row), [outptr] "+&r" (outptr), [weights] "+&r" (weights) : [bias] "r" (bias), [n_channels] "r" (n_channels), [offsetof_input_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_per_layer_mul] "I" (offsetof(arm_gemm::Requantize32, per_layer_mul)), [offsetof_per_layer_right_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_right_shift)), [offsetof_weights_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [qp] "r" (&qp), [rq_mul_perchannel] "r" (qp.per_channel_muls), [rq_shift_perchannel] "r" (qp.per_channel_right_shifts) - : "cc", "memory", "p0", "p1", "p2", "p8", "x20", "x21", "x22", "x23", "x24", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28" + : "cc", "memory", "p0", "p1", "p2", "p3", "p8", "x20", "x21", "x22", "x23", "x24", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" ); } |