| /* |
| * Copyright (c) 2021 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. |
| */ |
| |
| #include "pooling.hpp" |
| #include <cstdint> |
| |
| #if defined(ARM_COMPUTE_ENABLE_SVE) && defined(ARM_COMPUTE_ENABLE_SVE2) |
| |
| namespace arm_conv { |
| namespace pooling { |
| |
| |
| void sve_u8q_nhwc_max_generic_depthfirst_impl( |
| const uint64_t, |
| const uint64_t n_valid_cells, |
| uint64_t n_channels, |
| const uint8_t *const *const inptrs, |
| uint8_t *outptr, |
| const Requantize32 &qp |
| ) |
| { |
| __asm__ __volatile__( |
| "ptrue p4.b\n" |
| "mov x28, #0x0\n" |
| "cntb x27\n" |
| "cntb x26, ALL, MUL #2\n" |
| "cntb x25, ALL, MUL #3\n" |
| "whilelt p3.b, x28, %x[n_channels]\n" |
| "whilelt p2.b, x27, %x[n_channels]\n" |
| "whilelt p1.b, x26, %x[n_channels]\n" |
| "whilelt p0.b, x25, %x[n_channels]\n" |
| "b.none 7f\n" |
| "1:" // 4-vectors of channels |
| "mov z10.b, #0x0\n" |
| "mov x19, %x[inptrs]\n" |
| "mov z9.b, #0x0\n" |
| "lsr x24, %x[n_valid_cells], #0x2\n" |
| "mov z8.b, #0x0\n" |
| "mov z7.b, #0x0\n" |
| "cbz x24, 4f\n" |
| "ldp x23, x22, [x19, #0x0]\n" |
| "ldp x21, x20, [x19, #0x10]\n" |
| "add x19, x19, #0x20\n" |
| "subs x24, x24, #0x1\n" |
| "ld1b { z3.b }, p3/Z, [x23, x28]\n" |
| "ld1b { z2.b }, p3/Z, [x22, x28]\n" |
| "ld1b { z1.b }, p3/Z, [x21, x28]\n" |
| "ld1b { z0.b }, p3/Z, [x20, x28]\n" |
| "ld1b { z31.b }, p2/Z, [x23, x27]\n" |
| "ld1b { z30.b }, p2/Z, [x22, x27]\n" |
| "ld1b { z22.b }, p2/Z, [x21, x27]\n" |
| "ld1b { z29.b }, p2/Z, [x20, x27]\n" |
| "ld1b { z28.b }, p1/Z, [x23, x26]\n" |
| "ld1b { z27.b }, p1/Z, [x22, x26]\n" |
| "ld1b { z21.b }, p1/Z, [x21, x26]\n" |
| "ld1b { z26.b }, p1/Z, [x20, x26]\n" |
| "ld1b { z16.b }, p0/Z, [x23, x25]\n" |
| "ld1b { z25.b }, p0/Z, [x22, x25]\n" |
| "ld1b { z20.b }, p0/Z, [x21, x25]\n" |
| "ld1b { z24.b }, p0/Z, [x20, x25]\n" |
| "beq 3f\n" |
| "2:" // 4-vectors of channels: 4 inputs loop |
| "movprfx z19, z3\n umax z19.b, p4/M, z19.b, z2.b\n" |
| "ldp x23, x22, [x19, #0x0]\n" |
| "subs x24, x24, #0x1\n" |
| "movprfx z23, z1\n umax z23.b, p4/M, z23.b, z0.b\n" |
| "ldp x21, x20, [x19, #0x10]\n" |
| "add x19, x19, #0x20\n" |
| "movprfx z18, z31\n umax z18.b, p4/M, z18.b, z30.b\n" |
| "ld1b { z3.b }, p3/Z, [x23, x28]\n" |
| "umax z22.b, p4/M, z22.b, z29.b\n" |
| "movprfx z17, z28\n umax z17.b, p4/M, z17.b, z27.b\n" |
| "ld1b { z2.b }, p3/Z, [x22, x28]\n" |
| "umax z21.b, p4/M, z21.b, z26.b\n" |
| "ld1b { z1.b }, p3/Z, [x21, x28]\n" |
| "umax z16.b, p4/M, z16.b, z25.b\n" |
| "ld1b { z0.b }, p3/Z, [x20, x28]\n" |
| "umax z20.b, p4/M, z20.b, z24.b\n" |
| "ld1b { z31.b }, p2/Z, [x23, x27]\n" |
| "umax z19.b, p4/M, z19.b, z23.b\n" |
| "ld1b { z30.b }, p2/Z, [x22, x27]\n" |
| "umax z18.b, p4/M, z18.b, z22.b\n" |
| "ld1b { z22.b }, p2/Z, [x21, x27]\n" |
| "umax z17.b, p4/M, z17.b, z21.b\n" |
| "ld1b { z29.b }, p2/Z, [x20, x27]\n" |
| "umax z16.b, p4/M, z16.b, z20.b\n" |
| "ld1b { z28.b }, p1/Z, [x23, x26]\n" |
| "umax z10.b, p4/M, z10.b, z19.b\n" |
| "ld1b { z27.b }, p1/Z, [x22, x26]\n" |
| "umax z9.b, p4/M, z9.b, z18.b\n" |
| "ld1b { z21.b }, p1/Z, [x21, x26]\n" |
| "umax z8.b, p4/M, z8.b, z17.b\n" |
| "ld1b { z26.b }, p1/Z, [x20, x26]\n" |
| "umax z7.b, p4/M, z7.b, z16.b\n" |
| "ld1b { z16.b }, p0/Z, [x23, x25]\n" |
| "ld1b { z25.b }, p0/Z, [x22, x25]\n" |
| "ld1b { z20.b }, p0/Z, [x21, x25]\n" |
| "ld1b { z24.b }, p0/Z, [x20, x25]\n" |
| "bgt 2b\n" |
| "3:" // 4-vectors of channels: 4 inputs tail |
| "movprfx z19, z3\n umax z19.b, p4/M, z19.b, z2.b\n" |
| "movprfx z23, z1\n umax z23.b, p4/M, z23.b, z0.b\n" |
| "movprfx z18, z31\n umax z18.b, p4/M, z18.b, z30.b\n" |
| "umax z22.b, p4/M, z22.b, z29.b\n" |
| "movprfx z17, z28\n umax z17.b, p4/M, z17.b, z27.b\n" |
| "umax z21.b, p4/M, z21.b, z26.b\n" |
| "umax z16.b, p4/M, z16.b, z25.b\n" |
| "umax z20.b, p4/M, z20.b, z24.b\n" |
| "umax z19.b, p4/M, z19.b, z23.b\n" |
| "umax z18.b, p4/M, z18.b, z22.b\n" |
| "umax z17.b, p4/M, z17.b, z21.b\n" |
| "umax z16.b, p4/M, z16.b, z20.b\n" |
| "umax z10.b, p4/M, z10.b, z19.b\n" |
| "umax z9.b, p4/M, z9.b, z18.b\n" |
| "umax z8.b, p4/M, z8.b, z17.b\n" |
| "umax z7.b, p4/M, z7.b, z16.b\n" |
| "4:" // 4-vectors of channels: After loop |
| "ands x20, %x[n_valid_cells], #0x3\n" |
| "beq 6f\n" |
| "5:" // 4-vectors of channels: Single input loop |
| "ldr x23, [x19], #0x8\n" |
| "subs x20, x20, #0x1\n" |
| "ld1b { z3.b }, p3/Z, [x23, x28]\n" |
| "umax z10.b, p4/M, z10.b, z3.b\n" |
| "ld1b { z31.b }, p2/Z, [x23, x27]\n" |
| "ld1b { z28.b }, p1/Z, [x23, x26]\n" |
| "umax z9.b, p4/M, z9.b, z31.b\n" |
| "ld1b { z16.b }, p0/Z, [x23, x25]\n" |
| "umax z8.b, p4/M, z8.b, z28.b\n" |
| "umax z7.b, p4/M, z7.b, z16.b\n" |
| "bgt 5b\n" |
| "6:" // 4-vectors of channels: Single input loop: End |
| "mov z6.s, #0x0\n" |
| "add x19, %x[quant_params], %[offsetof_qp_input_offset]\n" |
| "ld1rw { z5.s }, p4/Z, [x19]\n" |
| "mov z4.s, #0xff\n" |
| "add x19, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" |
| ".inst 0x4508a951 // ushllb z17.h, z10.b, #0x0\n" |
| "ld1rw { z3.s }, p4/Z, [x19]\n" |
| "add x19, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" |
| ".inst 0x4508ad50 // ushllt z16.h, z10.b, #0x0\n" |
| "ld1rw { z2.s }, p4/Z, [x19]\n" |
| "add x19, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" |
| ".inst 0x4508a937 // ushllb z23.h, z9.b, #0x0\n" |
| "ld1rw { z1.s }, p4/Z, [x19]\n" |
| "add x19, %x[quant_params], %[offsetof_qp_output_offset]\n" |
| ".inst 0x4508ad36 // ushllt z22.h, z9.b, #0x0\n" |
| "ld1rw { z0.s }, p4/Z, [x19]\n" |
| ".inst 0x4508a912 // ushllb z18.h, z8.b, #0x0\n" |
| ".inst 0x4508ad15 // ushllt z21.h, z8.b, #0x0\n" |
| ".inst 0x4508a8f4 // ushllb z20.h, z7.b, #0x0\n" |
| ".inst 0x4508acf3 // ushllt z19.h, z7.b, #0x0\n" |
| "neg z5.s, p4/M, z5.s\n" |
| ".inst 0x459140bf // saddwb z31.s, z5.s, z17.h\n" |
| ".inst 0x459144b1 // saddwt z17.s, z5.s, z17.h\n" |
| ".inst 0x459040be // saddwb z30.s, z5.s, z16.h\n" |
| ".inst 0x459044b0 // saddwt z16.s, z5.s, z16.h\n" |
| ".inst 0x459740bd // saddwb z29.s, z5.s, z23.h\n" |
| ".inst 0x459744bc // saddwt z28.s, z5.s, z23.h\n" |
| ".inst 0x459640bb // saddwb z27.s, z5.s, z22.h\n" |
| ".inst 0x459644ba // saddwt z26.s, z5.s, z22.h\n" |
| ".inst 0x459240b9 // saddwb z25.s, z5.s, z18.h\n" |
| ".inst 0x459244b2 // saddwt z18.s, z5.s, z18.h\n" |
| ".inst 0x459540b8 // saddwb z24.s, z5.s, z21.h\n" |
| ".inst 0x459544b7 // saddwt z23.s, z5.s, z21.h\n" |
| ".inst 0x459440b6 // saddwb z22.s, z5.s, z20.h\n" |
| ".inst 0x459444b5 // saddwt z21.s, z5.s, z20.h\n" |
| ".inst 0x459340b4 // saddwb z20.s, z5.s, z19.h\n" |
| ".inst 0x459344b3 // saddwt z19.s, z5.s, z19.h\n" |
| ".inst 0x4482905f // srshl z31.s, p4/M, z31.s, z2.s\n" |
| ".inst 0x44829051 // srshl z17.s, p4/M, z17.s, z2.s\n" |
| ".inst 0x4482905e // srshl z30.s, p4/M, z30.s, z2.s\n" |
| ".inst 0x44829050 // srshl z16.s, p4/M, z16.s, z2.s\n" |
| ".inst 0x04a377ff // sqrdmulh z31.s, z31.s, z3.s\n" |
| ".inst 0x04a37631 // sqrdmulh z17.s, z17.s, z3.s\n" |
| ".inst 0x04a377de // sqrdmulh z30.s, z30.s, z3.s\n" |
| ".inst 0x04a37610 // sqrdmulh z16.s, z16.s, z3.s\n" |
| ".inst 0x4482903f // srshl z31.s, p4/M, z31.s, z1.s\n" |
| ".inst 0x44829031 // srshl z17.s, p4/M, z17.s, z1.s\n" |
| ".inst 0x4482903e // srshl z30.s, p4/M, z30.s, z1.s\n" |
| ".inst 0x44829030 // srshl z16.s, p4/M, z16.s, z1.s\n" |
| "add z31.s, z31.s, z0.s\n" |
| "add z17.s, z17.s, z0.s\n" |
| "add z30.s, z30.s, z0.s\n" |
| "add z16.s, z16.s, z0.s\n" |
| ".inst 0x4482905d // srshl z29.s, p4/M, z29.s, z2.s\n" |
| ".inst 0x4482905c // srshl z28.s, p4/M, z28.s, z2.s\n" |
| ".inst 0x4482905b // srshl z27.s, p4/M, z27.s, z2.s\n" |
| ".inst 0x4482905a // srshl z26.s, p4/M, z26.s, z2.s\n" |
| ".inst 0x04a377bd // sqrdmulh z29.s, z29.s, z3.s\n" |
| ".inst 0x04a3779c // sqrdmulh z28.s, z28.s, z3.s\n" |
| ".inst 0x04a3777b // sqrdmulh z27.s, z27.s, z3.s\n" |
| ".inst 0x04a3775a // sqrdmulh z26.s, z26.s, z3.s\n" |
| ".inst 0x4482903d // srshl z29.s, p4/M, z29.s, z1.s\n" |
| ".inst 0x4482903c // srshl z28.s, p4/M, z28.s, z1.s\n" |
| ".inst 0x4482903b // srshl z27.s, p4/M, z27.s, z1.s\n" |
| ".inst 0x4482903a // srshl z26.s, p4/M, z26.s, z1.s\n" |
| "add z29.s, z29.s, z0.s\n" |
| "add z28.s, z28.s, z0.s\n" |
| "add z27.s, z27.s, z0.s\n" |
| "add z26.s, z26.s, z0.s\n" |
| ".inst 0x44829059 // srshl z25.s, p4/M, z25.s, z2.s\n" |
| ".inst 0x44829052 // srshl z18.s, p4/M, z18.s, z2.s\n" |
| "smax z31.s, p4/M, z31.s, z6.s\n" |
| "smax z17.s, p4/M, z17.s, z6.s\n" |
| ".inst 0x04a37739 // sqrdmulh z25.s, z25.s, z3.s\n" |
| ".inst 0x04a37652 // sqrdmulh z18.s, z18.s, z3.s\n" |
| "smin z31.s, p4/M, z31.s, z4.s\n" |
| "smin z17.s, p4/M, z17.s, z4.s\n" |
| ".inst 0x44829039 // srshl z25.s, p4/M, z25.s, z1.s\n" |
| ".inst 0x44829032 // srshl z18.s, p4/M, z18.s, z1.s\n" |
| "smax z30.s, p4/M, z30.s, z6.s\n" |
| "trn1 z17.h, z31.h, z17.h\n" |
| "add z25.s, z25.s, z0.s\n" |
| "add z18.s, z18.s, z0.s\n" |
| ".inst 0x44829058 // srshl z24.s, p4/M, z24.s, z2.s\n" |
| ".inst 0x44829057 // srshl z23.s, p4/M, z23.s, z2.s\n" |
| "smin z30.s, p4/M, z30.s, z4.s\n" |
| "smax z16.s, p4/M, z16.s, z6.s\n" |
| ".inst 0x04a37718 // sqrdmulh z24.s, z24.s, z3.s\n" |
| ".inst 0x04a376f7 // sqrdmulh z23.s, z23.s, z3.s\n" |
| "smax z29.s, p4/M, z29.s, z6.s\n" |
| "smin z16.s, p4/M, z16.s, z4.s\n" |
| ".inst 0x44829038 // srshl z24.s, p4/M, z24.s, z1.s\n" |
| ".inst 0x44829037 // srshl z23.s, p4/M, z23.s, z1.s\n" |
| "smin z29.s, p4/M, z29.s, z4.s\n" |
| "trn1 z16.h, z30.h, z16.h\n" |
| "add z24.s, z24.s, z0.s\n" |
| "add z23.s, z23.s, z0.s\n" |
| "trn1 z16.b, z17.b, z16.b\n" |
| "st1b { z16.b }, p3, [%x[outptr], x28]\n" |
| ".inst 0x44829056 // srshl z22.s, p4/M, z22.s, z2.s\n" |
| "incb x28, ALL, MUL #4\n" |
| ".inst 0x44829055 // srshl z21.s, p4/M, z21.s, z2.s\n" |
| ".inst 0x44829054 // srshl z20.s, p4/M, z20.s, z2.s\n" |
| ".inst 0x44829053 // srshl z19.s, p4/M, z19.s, z2.s\n" |
| "smax z28.s, p4/M, z28.s, z6.s\n" |
| ".inst 0x04a376d6 // sqrdmulh z22.s, z22.s, z3.s\n" |
| ".inst 0x04a376b5 // sqrdmulh z21.s, z21.s, z3.s\n" |
| ".inst 0x04a37694 // sqrdmulh z20.s, z20.s, z3.s\n" |
| ".inst 0x04a37673 // sqrdmulh z19.s, z19.s, z3.s\n" |
| ".inst 0x44829036 // srshl z22.s, p4/M, z22.s, z1.s\n" |
| ".inst 0x44829035 // srshl z21.s, p4/M, z21.s, z1.s\n" |
| ".inst 0x44829034 // srshl z20.s, p4/M, z20.s, z1.s\n" |
| ".inst 0x44829033 // srshl z19.s, p4/M, z19.s, z1.s\n" |
| "add z22.s, z22.s, z0.s\n" |
| "add z21.s, z21.s, z0.s\n" |
| "add z20.s, z20.s, z0.s\n" |
| "add z19.s, z19.s, z0.s\n" |
| "smax z27.s, p4/M, z27.s, z6.s\n" |
| "smax z26.s, p4/M, z26.s, z6.s\n" |
| "smax z25.s, p4/M, z25.s, z6.s\n" |
| "smin z28.s, p4/M, z28.s, z4.s\n" |
| "smin z27.s, p4/M, z27.s, z4.s\n" |
| "smin z26.s, p4/M, z26.s, z4.s\n" |
| "smin z25.s, p4/M, z25.s, z4.s\n" |
| "trn1 z17.h, z29.h, z28.h\n" |
| "smax z18.s, p4/M, z18.s, z6.s\n" |
| "trn1 z16.h, z27.h, z26.h\n" |
| "smax z24.s, p4/M, z24.s, z6.s\n" |
| "trn1 z16.b, z17.b, z16.b\n" |
| "st1b { z16.b }, p2, [%x[outptr], x27]\n" |
| "smin z18.s, p4/M, z18.s, z4.s\n" |
| "incb x27, ALL, MUL #4\n" |
| "smin z24.s, p4/M, z24.s, z4.s\n" |
| "smax z23.s, p4/M, z23.s, z6.s\n" |
| "smax z22.s, p4/M, z22.s, z6.s\n" |
| "smax z21.s, p4/M, z21.s, z6.s\n" |
| "trn1 z18.h, z25.h, z18.h\n" |
| "smin z23.s, p4/M, z23.s, z4.s\n" |
| "smin z22.s, p4/M, z22.s, z4.s\n" |
| "smin z21.s, p4/M, z21.s, z4.s\n" |
| "smax z20.s, p4/M, z20.s, z6.s\n" |
| "trn1 z16.h, z24.h, z23.h\n" |
| "smax z19.s, p4/M, z19.s, z6.s\n" |
| "trn1 z17.h, z22.h, z21.h\n" |
| "trn1 z16.b, z18.b, z16.b\n" |
| "st1b { z16.b }, p1, [%x[outptr], x26]\n" |
| "smin z20.s, p4/M, z20.s, z4.s\n" |
| "incb x26, ALL, MUL #4\n" |
| "smin z19.s, p4/M, z19.s, z4.s\n" |
| "trn1 z16.h, z20.h, z19.h\n" |
| "trn1 z16.b, z17.b, z16.b\n" |
| "st1b { z16.b }, p0, [%x[outptr], x25]\n" |
| "incb x25, ALL, MUL #4\n" |
| "whilelt p0.b, x25, %x[n_channels]\n" |
| "b.any 1b\n" |
| "7:" // Single vector of channels |
| "whilelt p3.b, x28, %x[n_channels]\n" |
| "b.none 14f\n" |
| "8:" // Single vector of channels: Loop |
| "mov z10.b, #0x0\n" |
| "mov x19, %x[inptrs]\n" |
| "lsr x24, %x[n_valid_cells], #0x2\n" |
| "cbz x24, 11f\n" |
| "ldp x23, x22, [x19, #0x0]\n" |
| "ldp x21, x20, [x19, #0x10]\n" |
| "add x19, x19, #0x20\n" |
| "subs x24, x24, #0x1\n" |
| "ld1b { z3.b }, p3/Z, [x23, x28]\n" |
| "ld1b { z2.b }, p3/Z, [x22, x28]\n" |
| "ld1b { z1.b }, p3/Z, [x21, x28]\n" |
| "ld1b { z0.b }, p3/Z, [x20, x28]\n" |
| "beq 10f\n" |
| "9:" // Single vector of channels: Loop: 4 inputs loop |
| "movprfx z19, z3\n umax z19.b, p4/M, z19.b, z2.b\n" |
| "ldp x23, x22, [x19, #0x0]\n" |
| "subs x24, x24, #0x1\n" |
| "movprfx z23, z1\n umax z23.b, p4/M, z23.b, z0.b\n" |
| "ldp x21, x20, [x19, #0x10]\n" |
| "add x19, x19, #0x20\n" |
| "umax z19.b, p4/M, z19.b, z23.b\n" |
| "ld1b { z3.b }, p3/Z, [x23, x28]\n" |
| "ld1b { z2.b }, p3/Z, [x22, x28]\n" |
| "umax z10.b, p4/M, z10.b, z19.b\n" |
| "ld1b { z1.b }, p3/Z, [x21, x28]\n" |
| "ld1b { z0.b }, p3/Z, [x20, x28]\n" |
| "bgt 9b\n" |
| "10:" // Single vector of channels: Loop: 4 inputs tail |
| "movprfx z19, z3\n umax z19.b, p4/M, z19.b, z2.b\n" |
| "movprfx z23, z1\n umax z23.b, p4/M, z23.b, z0.b\n" |
| "umax z19.b, p4/M, z19.b, z23.b\n" |
| "umax z10.b, p4/M, z10.b, z19.b\n" |
| "11:" // Single vector of channels: Loop: After loop |
| "ands x20, %x[n_valid_cells], #0x3\n" |
| "beq 13f\n" |
| "12:" // Single vector of channels: Loop: Single input loop |
| "ldr x23, [x19], #0x8\n" |
| "subs x20, x20, #0x1\n" |
| "ld1b { z3.b }, p3/Z, [x23, x28]\n" |
| "umax z10.b, p4/M, z10.b, z3.b\n" |
| "bgt 12b\n" |
| "13:" // Single vector of channels: Loop: Single input loop: End |
| "mov z6.s, #0x0\n" |
| "add x19, %x[quant_params], %[offsetof_qp_input_offset]\n" |
| "ld1rw { z5.s }, p4/Z, [x19]\n" |
| "mov z4.s, #0xff\n" |
| "add x19, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" |
| ".inst 0x4508a951 // ushllb z17.h, z10.b, #0x0\n" |
| "ld1rw { z3.s }, p4/Z, [x19]\n" |
| "add x19, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" |
| ".inst 0x4508ad50 // ushllt z16.h, z10.b, #0x0\n" |
| "ld1rw { z2.s }, p4/Z, [x19]\n" |
| "add x19, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" |
| "neg z5.s, p4/M, z5.s\n" |
| "ld1rw { z1.s }, p4/Z, [x19]\n" |
| "add x19, %x[quant_params], %[offsetof_qp_output_offset]\n" |
| ".inst 0x459140bf // saddwb z31.s, z5.s, z17.h\n" |
| "ld1rw { z0.s }, p4/Z, [x19]\n" |
| ".inst 0x459144b1 // saddwt z17.s, z5.s, z17.h\n" |
| ".inst 0x459040be // saddwb z30.s, z5.s, z16.h\n" |
| ".inst 0x459044b0 // saddwt z16.s, z5.s, z16.h\n" |
| ".inst 0x4482905f // srshl z31.s, p4/M, z31.s, z2.s\n" |
| ".inst 0x44829051 // srshl z17.s, p4/M, z17.s, z2.s\n" |
| ".inst 0x4482905e // srshl z30.s, p4/M, z30.s, z2.s\n" |
| ".inst 0x44829050 // srshl z16.s, p4/M, z16.s, z2.s\n" |
| ".inst 0x04a377ff // sqrdmulh z31.s, z31.s, z3.s\n" |
| ".inst 0x04a37631 // sqrdmulh z17.s, z17.s, z3.s\n" |
| ".inst 0x04a377de // sqrdmulh z30.s, z30.s, z3.s\n" |
| ".inst 0x04a37610 // sqrdmulh z16.s, z16.s, z3.s\n" |
| ".inst 0x4482903f // srshl z31.s, p4/M, z31.s, z1.s\n" |
| ".inst 0x44829031 // srshl z17.s, p4/M, z17.s, z1.s\n" |
| ".inst 0x4482903e // srshl z30.s, p4/M, z30.s, z1.s\n" |
| ".inst 0x44829030 // srshl z16.s, p4/M, z16.s, z1.s\n" |
| "add z31.s, z31.s, z0.s\n" |
| "add z17.s, z17.s, z0.s\n" |
| "add z30.s, z30.s, z0.s\n" |
| "add z16.s, z16.s, z0.s\n" |
| "smax z31.s, p4/M, z31.s, z6.s\n" |
| "smax z17.s, p4/M, z17.s, z6.s\n" |
| "smax z30.s, p4/M, z30.s, z6.s\n" |
| "smax z16.s, p4/M, z16.s, z6.s\n" |
| "smin z31.s, p4/M, z31.s, z4.s\n" |
| "smin z17.s, p4/M, z17.s, z4.s\n" |
| "smin z30.s, p4/M, z30.s, z4.s\n" |
| "smin z16.s, p4/M, z16.s, z4.s\n" |
| "trn1 z17.h, z31.h, z17.h\n" |
| "trn1 z16.h, z30.h, z16.h\n" |
| "trn1 z16.b, z17.b, z16.b\n" |
| "st1b { z16.b }, p3, [%x[outptr], x28]\n" |
| "incb x28\n" |
| "whilelt p3.b, x28, %x[n_channels]\n" |
| "b.any 8b\n" |
| "14:" // End |
| |
| : |
| : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [n_valid_cells] "r" (n_valid_cells), [offsetof_qp_input_offset] "I" (offsetof(Requantize32, input_offset)), [offsetof_qp_output_offset] "I" (offsetof(Requantize32, output_offset)), [offsetof_qp_per_layer_left_shift] "I" (offsetof(Requantize32, per_layer_left_shift)), [offsetof_qp_per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [offsetof_qp_per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [outptr] "r" (outptr), [quant_params] "r" (&qp) |
| : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" |
| ); |
| } |
| |
| } // namespace pooling |
| } // namespace arm_conv |
| |
| #endif // defined(ARM_COMPUTE_ENABLE_SVE) && defined(ARM_COMPUTE_ENABLE_SVE2) |