| /* |
| * Copyright (c) 2021, 2023 Arm Limited. |
| * |
| * SPDX-License-Identifier: MIT |
| * |
| * Permission is hereby granted, free of charge, to any person obtaining a copy |
| * of this software and associated documentation files (the "Software"), to |
| * deal in the Software without restriction, including without limitation the |
| * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| * sell copies of the Software, and to permit persons to whom the Software is |
| * furnished to do so, subject to the following conditions: |
| * |
| * The above copyright notice and this permission notice shall be included in all |
| * copies or substantial portions of the Software. |
| * |
| * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| * SOFTWARE. |
| */ |
| |
| #if defined(__aarch64__) |
| |
| #include "arm_gemm.hpp" |
| #include <cstddef> |
| #include <cstdint> |
| |
| namespace arm_conv { |
| namespace depthwise { |
| |
| void a64_s8q_nhwc_generic_output9_mla_depthfirst_impl( |
| const int8_t *const *const inptrs, |
| int8_t *const *const outptrs, |
| const void *params, |
| const arm_gemm::Requantize32& qp, |
| const unsigned int n_points, |
| const unsigned int n_channels |
| ) |
| { |
| __asm__ __volatile__( |
| "lsr x12, %x[n_channels], #0x2\n" |
| "add x20, %x[qp], %[offsetof_Requantize32_minval]\n" |
| "ld1r { v8.4s }, [x20]\n" |
| "add x20, %x[qp], %[offsetof_Requantize32_maxval]\n" |
| "ld1r { v7.4s }, [x20]\n" |
| "add x20, %x[qp], %[offsetof_Requantize32_a_offset]\n" |
| "ld1r { v6.16b }, [x20]\n" |
| "add x20, %x[qp], %[offsetof_Requantize32_b_offset]\n" |
| "ld1r { v5.16b }, [x20]\n" |
| "add x20, %x[qp], %[offsetof_Requantize32_c_offset]\n" |
| "ld1r { v4.4s }, [x20]\n" |
| "add x20, %x[qp], %[offsetof_Requantize32_per_layer_left_shift]\n" |
| "ld1r { v3.4s }, [x20]\n" |
| "add x20, %x[qp], %[offsetof_Requantize32_per_layer_mul]\n" |
| "ld1r { v2.4s }, [x20]\n" |
| "add x20, %x[qp], %[offsetof_Requantize32_per_layer_right_shift]\n" |
| "ld1r { v1.4s }, [x20]\n" |
| "mov x11, #0x0\n" |
| "cbz x12, 6f\n" |
| "1:" // Channel loop |
| "movi v23.4s, #0x0\n" |
| "cbz %x[bias], 2f\n" |
| "lsl x20, x11, #0x2\n" |
| "ldr q23, [%x[bias], x20]\n" |
| "2:" // Channel loop: Load bias: Done |
| "ldr s0, [%x[params]], #0x4\n" |
| "mov x21, %x[inptrs]\n" |
| "ldp x10, x9, [x21], #0x10\n" |
| "subs x20, %x[n_points], #0x1\n" |
| "ldr s14, [x10, x11]\n" |
| "ldr s15, [x9, x11]\n" |
| "mov v24.16b, v23.16b\n" |
| "mov v25.16b, v23.16b\n" |
| "ldp x28, x27, [x21], #0x10\n" |
| "ldr s16, [x28, x11]\n" |
| "mov v26.16b, v23.16b\n" |
| "mov v27.16b, v23.16b\n" |
| "ldr s17, [x27, x11]\n" |
| "ldp x26, x25, [x21], #0x10\n" |
| "mov v28.16b, v23.16b\n" |
| "mov v29.16b, v23.16b\n" |
| "ldr s18, [x26, x11]\n" |
| "ldr s19, [x25, x11]\n" |
| "mov v30.16b, v23.16b\n" |
| "mov v31.16b, v23.16b\n" |
| "ldp x24, x23, [x21], #0x10\n" |
| "ldr s20, [x24, x11]\n" |
| "ssubl v0.8h, v0.8b, v5.8b\n" |
| "ssubl v14.8h, v14.8b, v6.8b\n" |
| "ldr s21, [x23, x11]\n" |
| "ldr x22, [x21], #0x8\n" |
| "ssubl v15.8h, v15.8b, v6.8b\n" |
| "ssubl v16.8h, v16.8b, v6.8b\n" |
| "ldr s22, [x22, x11]\n" |
| "ssubl v17.8h, v17.8b, v6.8b\n" |
| "ssubl v18.8h, v18.8b, v6.8b\n" |
| "ssubl v19.8h, v19.8b, v6.8b\n" |
| "ssubl v20.8h, v20.8b, v6.8b\n" |
| "ssubl v21.8h, v21.8b, v6.8b\n" |
| "ssubl v22.8h, v22.8b, v6.8b\n" |
| "ble 4f\n" |
| "3:" // Channel loop: Planar loop |
| "ldp x10, x9, [x21], #0x10\n" |
| "ldp x28, x27, [x21], #0x10\n" |
| "smlal v23.4s, v14.4h, v0.4h\n" |
| "smlal v24.4s, v15.4h, v0.4h\n" |
| "ldr s14, [x10, x11]\n" |
| "ldr s15, [x9, x11]\n" |
| "smlal v25.4s, v16.4h, v0.4h\n" |
| "smlal v26.4s, v17.4h, v0.4h\n" |
| "ldr s16, [x28, x11]\n" |
| "ldr s17, [x27, x11]\n" |
| "smlal v27.4s, v18.4h, v0.4h\n" |
| "smlal v28.4s, v19.4h, v0.4h\n" |
| "ldp x26, x25, [x21], #0x10\n" |
| "ldr s18, [x26, x11]\n" |
| "smlal v29.4s, v20.4h, v0.4h\n" |
| "smlal v30.4s, v21.4h, v0.4h\n" |
| "ldr s19, [x25, x11]\n" |
| "ldp x24, x23, [x21], #0x10\n" |
| "smlal v31.4s, v22.4h, v0.4h\n" |
| "subs x20, x20, #0x1\n" |
| "ldr s0, [%x[params]], #0x4\n" |
| "ldr s20, [x24, x11]\n" |
| "ssubl v0.8h, v0.8b, v5.8b\n" |
| "ssubl v14.8h, v14.8b, v6.8b\n" |
| "ldr s21, [x23, x11]\n" |
| "ldr x22, [x21], #0x8\n" |
| "ssubl v15.8h, v15.8b, v6.8b\n" |
| "ssubl v16.8h, v16.8b, v6.8b\n" |
| "ldr s22, [x22, x11]\n" |
| "ssubl v17.8h, v17.8b, v6.8b\n" |
| "ssubl v18.8h, v18.8b, v6.8b\n" |
| "ssubl v19.8h, v19.8b, v6.8b\n" |
| "ssubl v20.8h, v20.8b, v6.8b\n" |
| "ssubl v21.8h, v21.8b, v6.8b\n" |
| "ssubl v22.8h, v22.8b, v6.8b\n" |
| "bgt 3b\n" |
| "4:" // Channel loop: Planar tail |
| "smlal v23.4s, v14.4h, v0.4h\n" |
| "smlal v24.4s, v15.4h, v0.4h\n" |
| "smlal v25.4s, v16.4h, v0.4h\n" |
| "smlal v26.4s, v17.4h, v0.4h\n" |
| "smlal v27.4s, v18.4h, v0.4h\n" |
| "smlal v28.4s, v19.4h, v0.4h\n" |
| "smlal v29.4s, v20.4h, v0.4h\n" |
| "smlal v30.4s, v21.4h, v0.4h\n" |
| "smlal v31.4s, v22.4h, v0.4h\n" |
| "cbz %x[rq_mul_ptr], 5f\n" |
| "lsl x20, x11, #0x2\n" |
| "ldr q2, [%x[rq_mul_ptr], x20]\n" |
| "ldr q1, [%x[rq_right_shift_ptr], x20]\n" |
| "cbz %x[rq_left_shift_ptr], 5f\n" |
| "ldr q3, [%x[rq_left_shift_ptr], x20]\n" |
| "5:" // Channel loop: Load quantisation parameters: Done |
| "sshl v23.4s, v23.4s, v3.4s\n" |
| "sshl v24.4s, v24.4s, v3.4s\n" |
| "ldp x28, x27, [%x[outptrs], #0x0]\n" |
| "ldp x26, x25, [%x[outptrs], #0x10]\n" |
| "sshl v25.4s, v25.4s, v3.4s\n" |
| "sqrdmulh v23.4s, v23.4s, v2.4s\n" |
| "ldp x24, x23, [%x[outptrs], #0x20]\n" |
| "ldp x22, x21, [%x[outptrs], #0x30]\n" |
| "sqrdmulh v24.4s, v24.4s, v2.4s\n" |
| "sqrdmulh v25.4s, v25.4s, v2.4s\n" |
| "ldr x20, [%x[outptrs], #0x40]\n" |
| "and v21.16b, v23.16b, v1.16b\n" |
| "and v20.16b, v24.16b, v1.16b\n" |
| "and v19.16b, v25.16b, v1.16b\n" |
| "sshl v26.4s, v26.4s, v3.4s\n" |
| "sshl v27.4s, v27.4s, v3.4s\n" |
| "sshl v28.4s, v28.4s, v3.4s\n" |
| "sshl v29.4s, v29.4s, v3.4s\n" |
| "sshl v30.4s, v30.4s, v3.4s\n" |
| "sshl v31.4s, v31.4s, v3.4s\n" |
| "sshr v21.4s, v21.4s, #0x1f\n" |
| "sshr v20.4s, v20.4s, #0x1f\n" |
| "sshr v19.4s, v19.4s, #0x1f\n" |
| "sqrdmulh v26.4s, v26.4s, v2.4s\n" |
| "sqrdmulh v27.4s, v27.4s, v2.4s\n" |
| "sqrdmulh v28.4s, v28.4s, v2.4s\n" |
| "sqrdmulh v29.4s, v29.4s, v2.4s\n" |
| "sqrdmulh v30.4s, v30.4s, v2.4s\n" |
| "sqrdmulh v31.4s, v31.4s, v2.4s\n" |
| "sqadd v23.4s, v23.4s, v21.4s\n" |
| "sqadd v24.4s, v24.4s, v20.4s\n" |
| "sqadd v25.4s, v25.4s, v19.4s\n" |
| "and v18.16b, v26.16b, v1.16b\n" |
| "and v17.16b, v27.16b, v1.16b\n" |
| "and v16.16b, v28.16b, v1.16b\n" |
| "and v21.16b, v29.16b, v1.16b\n" |
| "and v20.16b, v30.16b, v1.16b\n" |
| "and v19.16b, v31.16b, v1.16b\n" |
| "sshr v18.4s, v18.4s, #0x1f\n" |
| "sshr v17.4s, v17.4s, #0x1f\n" |
| "sshr v16.4s, v16.4s, #0x1f\n" |
| "sshr v21.4s, v21.4s, #0x1f\n" |
| "sshr v20.4s, v20.4s, #0x1f\n" |
| "sshr v19.4s, v19.4s, #0x1f\n" |
| "sqadd v26.4s, v26.4s, v18.4s\n" |
| "sqadd v27.4s, v27.4s, v17.4s\n" |
| "sqadd v28.4s, v28.4s, v16.4s\n" |
| "sqadd v29.4s, v29.4s, v21.4s\n" |
| "sqadd v30.4s, v30.4s, v20.4s\n" |
| "sqadd v31.4s, v31.4s, v19.4s\n" |
| "srshl v23.4s, v23.4s, v1.4s\n" |
| "srshl v24.4s, v24.4s, v1.4s\n" |
| "srshl v25.4s, v25.4s, v1.4s\n" |
| "srshl v26.4s, v26.4s, v1.4s\n" |
| "srshl v27.4s, v27.4s, v1.4s\n" |
| "srshl v28.4s, v28.4s, v1.4s\n" |
| "srshl v29.4s, v29.4s, v1.4s\n" |
| "srshl v30.4s, v30.4s, v1.4s\n" |
| "srshl v31.4s, v31.4s, v1.4s\n" |
| "add v23.4s, v23.4s, v4.4s\n" |
| "add v24.4s, v24.4s, v4.4s\n" |
| "add v25.4s, v25.4s, v4.4s\n" |
| "add v26.4s, v26.4s, v4.4s\n" |
| "add v27.4s, v27.4s, v4.4s\n" |
| "add v28.4s, v28.4s, v4.4s\n" |
| "add v29.4s, v29.4s, v4.4s\n" |
| "add v30.4s, v30.4s, v4.4s\n" |
| "add v31.4s, v31.4s, v4.4s\n" |
| "smax v23.4s, v23.4s, v8.4s\n" |
| "smax v24.4s, v24.4s, v8.4s\n" |
| "smax v25.4s, v25.4s, v8.4s\n" |
| "smax v26.4s, v26.4s, v8.4s\n" |
| "smax v27.4s, v27.4s, v8.4s\n" |
| "smax v28.4s, v28.4s, v8.4s\n" |
| "smax v29.4s, v29.4s, v8.4s\n" |
| "smax v30.4s, v30.4s, v8.4s\n" |
| "smax v31.4s, v31.4s, v8.4s\n" |
| "smin v23.4s, v23.4s, v7.4s\n" |
| "smin v24.4s, v24.4s, v7.4s\n" |
| "smin v25.4s, v25.4s, v7.4s\n" |
| "smin v26.4s, v26.4s, v7.4s\n" |
| "smin v27.4s, v27.4s, v7.4s\n" |
| "smin v28.4s, v28.4s, v7.4s\n" |
| "smin v29.4s, v29.4s, v7.4s\n" |
| "smin v30.4s, v30.4s, v7.4s\n" |
| "smin v31.4s, v31.4s, v7.4s\n" |
| "uzp1 v23.16b, v23.16b, v23.16b\n" |
| "uzp1 v24.16b, v24.16b, v24.16b\n" |
| "uzp1 v25.16b, v25.16b, v25.16b\n" |
| "uzp1 v26.16b, v26.16b, v26.16b\n" |
| "uzp1 v27.16b, v27.16b, v27.16b\n" |
| "uzp1 v28.16b, v28.16b, v28.16b\n" |
| "uzp1 v29.16b, v29.16b, v29.16b\n" |
| "uzp1 v30.16b, v30.16b, v30.16b\n" |
| "uzp1 v31.16b, v31.16b, v31.16b\n" |
| "uzp1 v23.16b, v23.16b, v23.16b\n" |
| "uzp1 v24.16b, v24.16b, v24.16b\n" |
| "str s23, [x28, x11]\n" |
| "uzp1 v25.16b, v25.16b, v25.16b\n" |
| "uzp1 v26.16b, v26.16b, v26.16b\n" |
| "str s24, [x27, x11]\n" |
| "uzp1 v27.16b, v27.16b, v27.16b\n" |
| "uzp1 v28.16b, v28.16b, v28.16b\n" |
| "str s25, [x26, x11]\n" |
| "uzp1 v29.16b, v29.16b, v29.16b\n" |
| "uzp1 v30.16b, v30.16b, v30.16b\n" |
| "str s26, [x25, x11]\n" |
| "uzp1 v31.16b, v31.16b, v31.16b\n" |
| "str s27, [x24, x11]\n" |
| "str s28, [x23, x11]\n" |
| "str s29, [x22, x11]\n" |
| "str s30, [x21, x11]\n" |
| "str s31, [x20, x11]\n" |
| "add x11, x11, #0x4\n" |
| "cmp x11, x12, LSL #2\n" |
| "blt 1b\n" |
| "6:" // Oddments |
| "tst %x[n_channels], #0x3\n" |
| "beq 24f\n" |
| "movi v23.4s, #0x0\n" |
| "cbz %x[bias], 9f\n" |
| "add x20, %x[bias], x11, LSL #2\n" |
| "tbz %x[n_channels], #1, 7f\n" |
| "ld1 { v23.d }[0], [x20], #0x8\n" |
| "tbz %x[n_channels], #0, 8f\n" |
| "ld1 { v23.s }[2], [x20], #0x4\n" |
| "b 8f\n" |
| "7:" // Oddments: Load bias: Bit 1: Unset |
| "ld1 { v23.s }[0], [x20], #0x4\n" |
| "8:" // Oddments: Load bias: Bit 1: End |
| "9:" // Oddments: Load bias: Done |
| "ldr s0, [%x[params]], #0x4\n" |
| "mov x21, %x[inptrs]\n" |
| "ldp x10, x9, [x21], #0x10\n" |
| "mov v24.16b, v23.16b\n" |
| "ldp x28, x27, [x21], #0x10\n" |
| "ldp x26, x25, [x21], #0x10\n" |
| "mov v25.16b, v23.16b\n" |
| "mov v26.16b, v23.16b\n" |
| "ldp x24, x23, [x21], #0x10\n" |
| "ldr x22, [x21], #0x8\n" |
| "mov v27.16b, v23.16b\n" |
| "mov v28.16b, v23.16b\n" |
| "mov v29.16b, v23.16b\n" |
| "mov v30.16b, v23.16b\n" |
| "add x10, x10, x11\n" |
| "add x9, x9, x11\n" |
| "mov v31.16b, v23.16b\n" |
| "ssubl v0.8h, v0.8b, v5.8b\n" |
| "add x28, x28, x11\n" |
| "add x27, x27, x11\n" |
| "add x26, x26, x11\n" |
| "add x25, x25, x11\n" |
| "add x24, x24, x11\n" |
| "add x23, x23, x11\n" |
| "add x22, x22, x11\n" |
| "tbz %x[n_channels], #1, 10f\n" |
| "ldr h14, [x10], #0x2\n" |
| "ldr h15, [x9], #0x2\n" |
| "ldr h16, [x28], #0x2\n" |
| "ldr h17, [x27], #0x2\n" |
| "ldr h18, [x26], #0x2\n" |
| "ldr h19, [x25], #0x2\n" |
| "ldr h20, [x24], #0x2\n" |
| "ldr h21, [x23], #0x2\n" |
| "ldr h22, [x22], #0x2\n" |
| "tbz %x[n_channels], #0, 11f\n" |
| "ld1 { v14.b }[2], [x10], #0x1\n" |
| "ld1 { v15.b }[2], [x9], #0x1\n" |
| "ld1 { v16.b }[2], [x28], #0x1\n" |
| "ld1 { v17.b }[2], [x27], #0x1\n" |
| "ld1 { v18.b }[2], [x26], #0x1\n" |
| "ld1 { v19.b }[2], [x25], #0x1\n" |
| "ld1 { v20.b }[2], [x24], #0x1\n" |
| "ld1 { v21.b }[2], [x23], #0x1\n" |
| "ld1 { v22.b }[2], [x22], #0x1\n" |
| "b 11f\n" |
| "10:" // Oddments: Load: Bit 1: Unset |
| "ldr b14, [x10], #0x1\n" |
| "ldr b15, [x9], #0x1\n" |
| "ldr b16, [x28], #0x1\n" |
| "ldr b17, [x27], #0x1\n" |
| "ldr b18, [x26], #0x1\n" |
| "ldr b19, [x25], #0x1\n" |
| "ldr b20, [x24], #0x1\n" |
| "ldr b21, [x23], #0x1\n" |
| "ldr b22, [x22], #0x1\n" |
| "11:" // Oddments: Load: Bit 1: End |
| "subs x20, %x[n_points], #0x1\n" |
| "ssubl v14.8h, v14.8b, v6.8b\n" |
| "ssubl v15.8h, v15.8b, v6.8b\n" |
| "ssubl v16.8h, v16.8b, v6.8b\n" |
| "ssubl v17.8h, v17.8b, v6.8b\n" |
| "ssubl v18.8h, v18.8b, v6.8b\n" |
| "ssubl v19.8h, v19.8b, v6.8b\n" |
| "ssubl v20.8h, v20.8b, v6.8b\n" |
| "ssubl v21.8h, v21.8b, v6.8b\n" |
| "ssubl v22.8h, v22.8b, v6.8b\n" |
| "ble 15f\n" |
| "12:" // Oddments: Planar loop |
| "ldp x10, x9, [x21], #0x10\n" |
| "ldp x28, x27, [x21], #0x10\n" |
| "smlal v23.4s, v14.4h, v0.4h\n" |
| "smlal v24.4s, v15.4h, v0.4h\n" |
| "ldp x26, x25, [x21], #0x10\n" |
| "ldp x24, x23, [x21], #0x10\n" |
| "smlal v25.4s, v16.4h, v0.4h\n" |
| "smlal v26.4s, v17.4h, v0.4h\n" |
| "smlal v27.4s, v18.4h, v0.4h\n" |
| "smlal v28.4s, v19.4h, v0.4h\n" |
| "ldr x22, [x21], #0x8\n" |
| "add x10, x10, x11\n" |
| "smlal v29.4s, v20.4h, v0.4h\n" |
| "smlal v30.4s, v21.4h, v0.4h\n" |
| "add x9, x9, x11\n" |
| "add x28, x28, x11\n" |
| "smlal v31.4s, v22.4h, v0.4h\n" |
| "ldr s0, [%x[params]], #0x4\n" |
| "ssubl v0.8h, v0.8b, v5.8b\n" |
| "add x27, x27, x11\n" |
| "add x26, x26, x11\n" |
| "add x25, x25, x11\n" |
| "add x24, x24, x11\n" |
| "add x23, x23, x11\n" |
| "add x22, x22, x11\n" |
| "tbz %x[n_channels], #1, 13f\n" |
| "ldr h14, [x10], #0x2\n" |
| "ldr h15, [x9], #0x2\n" |
| "ldr h16, [x28], #0x2\n" |
| "ldr h17, [x27], #0x2\n" |
| "ldr h18, [x26], #0x2\n" |
| "ldr h19, [x25], #0x2\n" |
| "ldr h20, [x24], #0x2\n" |
| "ldr h21, [x23], #0x2\n" |
| "ldr h22, [x22], #0x2\n" |
| "tbz %x[n_channels], #0, 14f\n" |
| "ld1 { v14.b }[2], [x10], #0x1\n" |
| "ld1 { v15.b }[2], [x9], #0x1\n" |
| "ld1 { v16.b }[2], [x28], #0x1\n" |
| "ld1 { v17.b }[2], [x27], #0x1\n" |
| "ld1 { v18.b }[2], [x26], #0x1\n" |
| "ld1 { v19.b }[2], [x25], #0x1\n" |
| "ld1 { v20.b }[2], [x24], #0x1\n" |
| "ld1 { v21.b }[2], [x23], #0x1\n" |
| "ld1 { v22.b }[2], [x22], #0x1\n" |
| "b 14f\n" |
| "13:" // Oddments: Planar loop: Load: Bit 1: Unset |
| "ldr b14, [x10], #0x1\n" |
| "ldr b15, [x9], #0x1\n" |
| "ldr b16, [x28], #0x1\n" |
| "ldr b17, [x27], #0x1\n" |
| "ldr b18, [x26], #0x1\n" |
| "ldr b19, [x25], #0x1\n" |
| "ldr b20, [x24], #0x1\n" |
| "ldr b21, [x23], #0x1\n" |
| "ldr b22, [x22], #0x1\n" |
| "14:" // Oddments: Planar loop: Load: Bit 1: End |
| "subs x20, x20, #0x1\n" |
| "ssubl v14.8h, v14.8b, v6.8b\n" |
| "ssubl v15.8h, v15.8b, v6.8b\n" |
| "ssubl v16.8h, v16.8b, v6.8b\n" |
| "ssubl v17.8h, v17.8b, v6.8b\n" |
| "ssubl v18.8h, v18.8b, v6.8b\n" |
| "ssubl v19.8h, v19.8b, v6.8b\n" |
| "ssubl v20.8h, v20.8b, v6.8b\n" |
| "ssubl v21.8h, v21.8b, v6.8b\n" |
| "ssubl v22.8h, v22.8b, v6.8b\n" |
| "bgt 12b\n" |
| "15:" // Oddments: Planar tail |
| "smlal v23.4s, v14.4h, v0.4h\n" |
| "smlal v24.4s, v15.4h, v0.4h\n" |
| "smlal v25.4s, v16.4h, v0.4h\n" |
| "smlal v26.4s, v17.4h, v0.4h\n" |
| "smlal v27.4s, v18.4h, v0.4h\n" |
| "smlal v28.4s, v19.4h, v0.4h\n" |
| "smlal v29.4s, v20.4h, v0.4h\n" |
| "smlal v30.4s, v21.4h, v0.4h\n" |
| "smlal v31.4s, v22.4h, v0.4h\n" |
| "cbz %x[rq_mul_ptr], 21f\n" |
| "add x22, %x[rq_mul_ptr], x11, LSL #2\n" |
| "add x21, %x[rq_right_shift_ptr], x11, LSL #2\n" |
| "add x20, %x[rq_left_shift_ptr], x11, LSL #2\n" |
| "tbz %x[n_channels], #1, 18f\n" |
| "ld1 { v2.d }[0], [x22], #0x8\n" |
| "ld1 { v1.d }[0], [x21], #0x8\n" |
| "cbz %x[rq_left_shift_ptr], 16f\n" |
| "ld1 { v3.d }[0], [x20], #0x8\n" |
| "16:" // Oddments: Load quantisation parameters: Bit 1: Load left shift: Done |
| "tbz %x[n_channels], #0, 20f\n" |
| "ld1 { v2.s }[2], [x22], #0x4\n" |
| "ld1 { v1.s }[2], [x21], #0x4\n" |
| "cbz %x[rq_left_shift_ptr], 17f\n" |
| "ld1 { v3.s }[2], [x20], #0x4\n" |
| "17:" // Oddments: Load quantisation parameters: Bit 1: Bit 0: Load left shift: Done |
| "b 20f\n" |
| "18:" // Oddments: Load quantisation parameters: Bit 1: Unset |
| "ld1 { v2.s }[0], [x22], #0x4\n" |
| "ld1 { v1.s }[0], [x21], #0x4\n" |
| "cbz %x[rq_left_shift_ptr], 19f\n" |
| "ld1 { v3.s }[0], [x20], #0x4\n" |
| "19:" // Oddments: Load quantisation parameters: Bit 1: Unset: Bit 0: Load left shift: Done |
| |
| "20:" // Oddments: Load quantisation parameters: Bit 1: End |
| |
| "21:" // Oddments: Load quantisation parameters: Done |
| "sshl v23.4s, v23.4s, v3.4s\n" |
| "sshl v24.4s, v24.4s, v3.4s\n" |
| "ldp x28, x27, [%x[outptrs], #0x0]\n" |
| "ldp x26, x25, [%x[outptrs], #0x10]\n" |
| "sshl v25.4s, v25.4s, v3.4s\n" |
| "sqrdmulh v23.4s, v23.4s, v2.4s\n" |
| "ldp x24, x23, [%x[outptrs], #0x20]\n" |
| "ldp x22, x21, [%x[outptrs], #0x30]\n" |
| "sqrdmulh v24.4s, v24.4s, v2.4s\n" |
| "sqrdmulh v25.4s, v25.4s, v2.4s\n" |
| "ldr x20, [%x[outptrs], #0x40]\n" |
| "add x28, x28, x11\n" |
| "and v21.16b, v23.16b, v1.16b\n" |
| "and v20.16b, v24.16b, v1.16b\n" |
| "add x27, x27, x11\n" |
| "add x26, x26, x11\n" |
| "and v19.16b, v25.16b, v1.16b\n" |
| "sshl v26.4s, v26.4s, v3.4s\n" |
| "add x25, x25, x11\n" |
| "add x24, x24, x11\n" |
| "sshl v27.4s, v27.4s, v3.4s\n" |
| "sshl v28.4s, v28.4s, v3.4s\n" |
| "add x23, x23, x11\n" |
| "add x22, x22, x11\n" |
| "sshl v29.4s, v29.4s, v3.4s\n" |
| "sshl v30.4s, v30.4s, v3.4s\n" |
| "add x21, x21, x11\n" |
| "add x20, x20, x11\n" |
| "sshl v31.4s, v31.4s, v3.4s\n" |
| "sshr v21.4s, v21.4s, #0x1f\n" |
| "sshr v20.4s, v20.4s, #0x1f\n" |
| "sshr v19.4s, v19.4s, #0x1f\n" |
| "sqrdmulh v26.4s, v26.4s, v2.4s\n" |
| "sqrdmulh v27.4s, v27.4s, v2.4s\n" |
| "sqrdmulh v28.4s, v28.4s, v2.4s\n" |
| "sqrdmulh v29.4s, v29.4s, v2.4s\n" |
| "sqrdmulh v30.4s, v30.4s, v2.4s\n" |
| "sqrdmulh v31.4s, v31.4s, v2.4s\n" |
| "sqadd v23.4s, v23.4s, v21.4s\n" |
| "sqadd v24.4s, v24.4s, v20.4s\n" |
| "sqadd v25.4s, v25.4s, v19.4s\n" |
| "and v18.16b, v26.16b, v1.16b\n" |
| "and v17.16b, v27.16b, v1.16b\n" |
| "and v16.16b, v28.16b, v1.16b\n" |
| "and v21.16b, v29.16b, v1.16b\n" |
| "and v20.16b, v30.16b, v1.16b\n" |
| "and v19.16b, v31.16b, v1.16b\n" |
| "sshr v18.4s, v18.4s, #0x1f\n" |
| "sshr v17.4s, v17.4s, #0x1f\n" |
| "sshr v16.4s, v16.4s, #0x1f\n" |
| "sshr v21.4s, v21.4s, #0x1f\n" |
| "sshr v20.4s, v20.4s, #0x1f\n" |
| "sshr v19.4s, v19.4s, #0x1f\n" |
| "sqadd v26.4s, v26.4s, v18.4s\n" |
| "sqadd v27.4s, v27.4s, v17.4s\n" |
| "sqadd v28.4s, v28.4s, v16.4s\n" |
| "sqadd v29.4s, v29.4s, v21.4s\n" |
| "sqadd v30.4s, v30.4s, v20.4s\n" |
| "sqadd v31.4s, v31.4s, v19.4s\n" |
| "srshl v23.4s, v23.4s, v1.4s\n" |
| "srshl v24.4s, v24.4s, v1.4s\n" |
| "srshl v25.4s, v25.4s, v1.4s\n" |
| "srshl v26.4s, v26.4s, v1.4s\n" |
| "srshl v27.4s, v27.4s, v1.4s\n" |
| "srshl v28.4s, v28.4s, v1.4s\n" |
| "srshl v29.4s, v29.4s, v1.4s\n" |
| "srshl v30.4s, v30.4s, v1.4s\n" |
| "srshl v31.4s, v31.4s, v1.4s\n" |
| "add v23.4s, v23.4s, v4.4s\n" |
| "add v24.4s, v24.4s, v4.4s\n" |
| "add v25.4s, v25.4s, v4.4s\n" |
| "add v26.4s, v26.4s, v4.4s\n" |
| "add v27.4s, v27.4s, v4.4s\n" |
| "add v28.4s, v28.4s, v4.4s\n" |
| "add v29.4s, v29.4s, v4.4s\n" |
| "add v30.4s, v30.4s, v4.4s\n" |
| "add v31.4s, v31.4s, v4.4s\n" |
| "smax v23.4s, v23.4s, v8.4s\n" |
| "smax v24.4s, v24.4s, v8.4s\n" |
| "smax v25.4s, v25.4s, v8.4s\n" |
| "smax v26.4s, v26.4s, v8.4s\n" |
| "smax v27.4s, v27.4s, v8.4s\n" |
| "smax v28.4s, v28.4s, v8.4s\n" |
| "smax v29.4s, v29.4s, v8.4s\n" |
| "smax v30.4s, v30.4s, v8.4s\n" |
| "smax v31.4s, v31.4s, v8.4s\n" |
| "smin v23.4s, v23.4s, v7.4s\n" |
| "smin v24.4s, v24.4s, v7.4s\n" |
| "smin v25.4s, v25.4s, v7.4s\n" |
| "smin v26.4s, v26.4s, v7.4s\n" |
| "smin v27.4s, v27.4s, v7.4s\n" |
| "smin v28.4s, v28.4s, v7.4s\n" |
| "smin v29.4s, v29.4s, v7.4s\n" |
| "smin v30.4s, v30.4s, v7.4s\n" |
| "smin v31.4s, v31.4s, v7.4s\n" |
| "uzp1 v23.16b, v23.16b, v23.16b\n" |
| "uzp1 v24.16b, v24.16b, v24.16b\n" |
| "uzp1 v25.16b, v25.16b, v25.16b\n" |
| "uzp1 v26.16b, v26.16b, v26.16b\n" |
| "uzp1 v27.16b, v27.16b, v27.16b\n" |
| "uzp1 v28.16b, v28.16b, v28.16b\n" |
| "uzp1 v29.16b, v29.16b, v29.16b\n" |
| "uzp1 v30.16b, v30.16b, v30.16b\n" |
| "uzp1 v31.16b, v31.16b, v31.16b\n" |
| "uzp1 v23.16b, v23.16b, v23.16b\n" |
| "uzp1 v24.16b, v24.16b, v24.16b\n" |
| "uzp1 v25.16b, v25.16b, v25.16b\n" |
| "uzp1 v26.16b, v26.16b, v26.16b\n" |
| "uzp1 v27.16b, v27.16b, v27.16b\n" |
| "uzp1 v28.16b, v28.16b, v28.16b\n" |
| "uzp1 v29.16b, v29.16b, v29.16b\n" |
| "uzp1 v30.16b, v30.16b, v30.16b\n" |
| "uzp1 v31.16b, v31.16b, v31.16b\n" |
| "tbz %x[n_channels], #1, 22f\n" |
| "st1 { v23.h }[0], [x28], #0x2\n" |
| "st1 { v24.h }[0], [x27], #0x2\n" |
| "st1 { v25.h }[0], [x26], #0x2\n" |
| "st1 { v26.h }[0], [x25], #0x2\n" |
| "st1 { v27.h }[0], [x24], #0x2\n" |
| "st1 { v28.h }[0], [x23], #0x2\n" |
| "st1 { v29.h }[0], [x22], #0x2\n" |
| "st1 { v30.h }[0], [x21], #0x2\n" |
| "st1 { v31.h }[0], [x20], #0x2\n" |
| "tbz %x[n_channels], #0, 23f\n" |
| "st1 { v23.b }[2], [x28], #0x1\n" |
| "st1 { v24.b }[2], [x27], #0x1\n" |
| "st1 { v25.b }[2], [x26], #0x1\n" |
| "st1 { v26.b }[2], [x25], #0x1\n" |
| "st1 { v27.b }[2], [x24], #0x1\n" |
| "st1 { v28.b }[2], [x23], #0x1\n" |
| "st1 { v29.b }[2], [x22], #0x1\n" |
| "st1 { v30.b }[2], [x21], #0x1\n" |
| "st1 { v31.b }[2], [x20], #0x1\n" |
| "b 23f\n" |
| "22:" // Oddments: Store: Bit 1: Unset |
| "st1 { v23.b }[0], [x28], #0x1\n" |
| "st1 { v24.b }[0], [x27], #0x1\n" |
| "st1 { v25.b }[0], [x26], #0x1\n" |
| "st1 { v26.b }[0], [x25], #0x1\n" |
| "st1 { v27.b }[0], [x24], #0x1\n" |
| "st1 { v28.b }[0], [x23], #0x1\n" |
| "st1 { v29.b }[0], [x22], #0x1\n" |
| "st1 { v30.b }[0], [x21], #0x1\n" |
| "st1 { v31.b }[0], [x20], #0x1\n" |
| "23:" // Oddments: Store: Bit 1: End |
| |
| "24:" // End |
| |
| : [params] "+&r" (params) |
| : [bias] "r" (qp.bias), [inptrs] "r" (inptrs), [n_channels] "r" ((uint64_t) n_channels), [n_points] "r" ((uint64_t) n_points), [offsetof_Requantize32_a_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [offsetof_Requantize32_b_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [offsetof_Requantize32_c_offset] "I" (offsetof(arm_gemm::Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(arm_gemm::Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(arm_gemm::Requantize32, minval)), [offsetof_Requantize32_per_layer_left_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_left_shift)), [offsetof_Requantize32_per_layer_mul] "I" (offsetof(arm_gemm::Requantize32, per_layer_mul)), [offsetof_Requantize32_per_layer_right_shift] "I" (offsetof(arm_gemm::Requantize32, per_layer_right_shift)), [outptrs] "r" (outptrs), [qp] "r" (&qp), [rq_left_shift_ptr] "r" (qp.per_channel_left_shifts), [rq_mul_ptr] "r" (qp.per_channel_muls), [rq_right_shift_ptr] "r" (qp.per_channel_right_shifts) |
| : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" |
| ); |
| } |
| |
| } // namespace depthwise |
| } // namespace arm_conv |
| #endif // defined(__aarch64__) |