diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_max_generic_depthfirst/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_max_generic_depthfirst/generic.cpp | 418 |
1 files changed, 418 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_max_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_max_generic_depthfirst/generic.cpp new file mode 100644 index 0000000000..c8e8e7d399 --- /dev/null +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sme_u8q_nhwc_max_generic_depthfirst/generic.cpp @@ -0,0 +1,418 @@ +/* + * Copyright (c) 2022-2023 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. + */ + +#include "pooling.hpp" +#include <cstdint> +#include <cstddef> + +#if defined(ARM_COMPUTE_ENABLE_SME) + +namespace arm_conv { +namespace pooling { + + +void sme_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__( + ".inst 0xd503477f // SMSTART ZA\n" + "mov x9, #0x0\n" + "cntb x28\n" + "cntb x27, ALL, MUL #2\n" + "cntb x26, ALL, MUL #3\n" + "whilelt p4.b, x9, %x[n_channels]\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" + "ptrue p0.b\n" + "b.none 7f\n" + "1:" // 4-vectors of channels + "lsr x25, %x[n_valid_cells], #0x2\n" + "mov z5.b, #0x0\n" + "mov z3.b, #0x0\n" + "mov x24, %x[inptrs]\n" + "mov z2.b, #0x0\n" + "mov z1.b, #0x0\n" + "cbz x25, 4f\n" + "ldp x23, x22, [x24, #0x0]\n" + "subs x25, x25, #0x1\n" + "ld1b { z0.b }, p4/Z, [x23, x9]\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" + "ld1b { z31.b }, p4/Z, [x22, x9]\n" + "ld1b { z23.b }, p4/Z, [x21, x9]\n" + "ld1b { z30.b }, p4/Z, [x20, x9]\n" + "ld1b { z18.b }, p3/Z, [x23, x28]\n" + "ld1b { z29.b }, p3/Z, [x22, x28]\n" + "ld1b { z22.b }, p3/Z, [x21, x28]\n" + "ld1b { z28.b }, p3/Z, [x20, x28]\n" + "ld1b { z17.b }, p2/Z, [x23, x27]\n" + "ld1b { z27.b }, p2/Z, [x22, x27]\n" + "ld1b { z21.b }, p2/Z, [x21, x27]\n" + "ld1b { z26.b }, p2/Z, [x20, x27]\n" + "ld1b { z16.b }, p1/Z, [x23, x26]\n" + "ld1b { z25.b }, p1/Z, [x22, x26]\n" + "ld1b { z20.b }, p1/Z, [x21, x26]\n" + "ld1b { z24.b }, p1/Z, [x20, x26]\n" + "beq 3f\n" + "2:" // 4-vectors of channels: 4 inputs loop + "movprfx z19, z0\n umax z19.b, p0/M, z19.b, z31.b\n" + "umax z23.b, p0/M, z23.b, z30.b\n" + "ldp x23, x22, [x24, #0x0]\n" + "subs x25, x25, #0x1\n" + "umax z18.b, p0/M, z18.b, z29.b\n" + "umax z22.b, p0/M, z22.b, z28.b\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" + "umax z17.b, p0/M, z17.b, z27.b\n" + "umax z21.b, p0/M, z21.b, z26.b\n" + "ld1b { z0.b }, p4/Z, [x23, x9]\n" + "umax z16.b, p0/M, z16.b, z25.b\n" + "umax z20.b, p0/M, z20.b, z24.b\n" + "ld1b { z31.b }, p4/Z, [x22, x9]\n" + "umax z19.b, p0/M, z19.b, z23.b\n" + "umax z18.b, p0/M, z18.b, z22.b\n" + "ld1b { z23.b }, p4/Z, [x21, x9]\n" + "umax z17.b, p0/M, z17.b, z21.b\n" + "umax z16.b, p0/M, z16.b, z20.b\n" + "ld1b { z30.b }, p4/Z, [x20, x9]\n" + "umax z5.b, p0/M, z5.b, z19.b\n" + "umax z3.b, p0/M, z3.b, z18.b\n" + "ld1b { z18.b }, p3/Z, [x23, x28]\n" + "umax z2.b, p0/M, z2.b, z17.b\n" + "umax z1.b, p0/M, z1.b, z16.b\n" + "ld1b { z29.b }, p3/Z, [x22, x28]\n" + "ld1b { z22.b }, p3/Z, [x21, x28]\n" + "ld1b { z28.b }, p3/Z, [x20, x28]\n" + "ld1b { z17.b }, p2/Z, [x23, x27]\n" + "ld1b { z27.b }, p2/Z, [x22, x27]\n" + "ld1b { z21.b }, p2/Z, [x21, x27]\n" + "ld1b { z26.b }, p2/Z, [x20, x27]\n" + "ld1b { z16.b }, p1/Z, [x23, x26]\n" + "ld1b { z25.b }, p1/Z, [x22, x26]\n" + "ld1b { z20.b }, p1/Z, [x21, x26]\n" + "ld1b { z24.b }, p1/Z, [x20, x26]\n" + "bgt 2b\n" + "3:" // 4-vectors of channels: 4 inputs tail + "movprfx z19, z0\n umax z19.b, p0/M, z19.b, z31.b\n" + "umax z23.b, p0/M, z23.b, z30.b\n" + "umax z18.b, p0/M, z18.b, z29.b\n" + "umax z22.b, p0/M, z22.b, z28.b\n" + "umax z17.b, p0/M, z17.b, z27.b\n" + "umax z21.b, p0/M, z21.b, z26.b\n" + "umax z16.b, p0/M, z16.b, z25.b\n" + "umax z20.b, p0/M, z20.b, z24.b\n" + "umax z19.b, p0/M, z19.b, z23.b\n" + "umax z18.b, p0/M, z18.b, z22.b\n" + "umax z17.b, p0/M, z17.b, z21.b\n" + "umax z16.b, p0/M, z16.b, z20.b\n" + "umax z5.b, p0/M, z5.b, z19.b\n" + "umax z3.b, p0/M, z3.b, z18.b\n" + "umax z2.b, p0/M, z2.b, z17.b\n" + "umax z1.b, p0/M, z1.b, z16.b\n" + "4:" // 4-vectors of channels: After loop + "ands x21, %x[n_valid_cells], #0x3\n" + "beq 6f\n" + "5:" // 4-vectors of channels: Single input loop + "ldr x20, [x24], #0x8\n" + "ld1b { z16.b }, p4/Z, [x20, x9]\n" + "subs x21, x21, #0x1\n" + "umax z5.b, p0/M, z5.b, z16.b\n" + "ld1b { z16.b }, p3/Z, [x20, x28]\n" + "umax z3.b, p0/M, z3.b, z16.b\n" + "ld1b { z16.b }, p2/Z, [x20, x27]\n" + "umax z2.b, p0/M, z2.b, z16.b\n" + "ld1b { z16.b }, p1/Z, [x20, x26]\n" + "umax z1.b, p0/M, z1.b, z16.b\n" + "bgt 5b\n" + "6:" // 4-vectors of channels: Single input loop: End + "add x20, %x[quant_params], %[offsetof_qp_input_offset]\n" + "ld1rw { z4.s }, p0/Z, [x20]\n" + ".inst 0x4508a8b7 // ushllb z23.h, z5.b, #0x0\n" + ".inst 0x4508acb9 // ushllt z25.h, z5.b, #0x0\n" + ".inst 0x4508a876 // ushllb z22.h, z3.b, #0x0\n" + ".inst 0x4508ac72 // ushllt z18.h, z3.b, #0x0\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z3.s }, p0/Z, [x20]\n" + ".inst 0x4508a855 // ushllb z21.h, z2.b, #0x0\n" + ".inst 0x4508ac51 // ushllt z17.h, z2.b, #0x0\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "ld1rw { z2.s }, p0/Z, [x20]\n" + ".inst 0x4508a834 // ushllb z20.h, z1.b, #0x0\n" + ".inst 0x4508ac38 // ushllt z24.h, z1.b, #0x0\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "ld1rw { z19.s }, p0/Z, [x20]\n" + "neg z4.s, p0/M, z4.s\n" + ".inst 0x45974081 // saddwb z1.s, z4.s, z23.h\n" + "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" + "ld1rw { z16.s }, p0/Z, [x20]\n" + ".inst 0x45974497 // saddwt z23.s, z4.s, z23.h\n" + ".inst 0x45994080 // saddwb z0.s, z4.s, z25.h\n" + ".inst 0x4599449f // saddwt z31.s, z4.s, z25.h\n" + ".inst 0x4596409e // saddwb z30.s, z4.s, z22.h\n" + ".inst 0x45964496 // saddwt z22.s, z4.s, z22.h\n" + ".inst 0x4592409d // saddwb z29.s, z4.s, z18.h\n" + ".inst 0x45924492 // saddwt z18.s, z4.s, z18.h\n" + ".inst 0x4595409c // saddwb z28.s, z4.s, z21.h\n" + ".inst 0x45954495 // saddwt z21.s, z4.s, z21.h\n" + ".inst 0x4591409b // saddwb z27.s, z4.s, z17.h\n" + ".inst 0x45914491 // saddwt z17.s, z4.s, z17.h\n" + ".inst 0x4594409a // saddwb z26.s, z4.s, z20.h\n" + ".inst 0x45944494 // saddwt z20.s, z4.s, z20.h\n" + ".inst 0x45984099 // saddwb z25.s, z4.s, z24.h\n" + ".inst 0x45984498 // saddwt z24.s, z4.s, z24.h\n" + ".inst 0x44828061 // srshl z1.s, p0/M, z1.s, z3.s\n" + ".inst 0x44828077 // srshl z23.s, p0/M, z23.s, z3.s\n" + ".inst 0x44828060 // srshl z0.s, p0/M, z0.s, z3.s\n" + ".inst 0x4482807f // srshl z31.s, p0/M, z31.s, z3.s\n" + ".inst 0x4482807e // srshl z30.s, p0/M, z30.s, z3.s\n" + ".inst 0x44828076 // srshl z22.s, p0/M, z22.s, z3.s\n" + ".inst 0x4482807d // srshl z29.s, p0/M, z29.s, z3.s\n" + ".inst 0x44828072 // srshl z18.s, p0/M, z18.s, z3.s\n" + ".inst 0x4482807c // srshl z28.s, p0/M, z28.s, z3.s\n" + ".inst 0x44828075 // srshl z21.s, p0/M, z21.s, z3.s\n" + ".inst 0x4482807b // srshl z27.s, p0/M, z27.s, z3.s\n" + ".inst 0x44828071 // srshl z17.s, p0/M, z17.s, z3.s\n" + ".inst 0x4482807a // srshl z26.s, p0/M, z26.s, z3.s\n" + ".inst 0x44828074 // srshl z20.s, p0/M, z20.s, z3.s\n" + ".inst 0x44828079 // srshl z25.s, p0/M, z25.s, z3.s\n" + ".inst 0x44828078 // srshl z24.s, p0/M, z24.s, z3.s\n" + ".inst 0x04a27421 // sqrdmulh z1.s, z1.s, z2.s\n" + ".inst 0x04a276f7 // sqrdmulh z23.s, z23.s, z2.s\n" + ".inst 0x04a27400 // sqrdmulh z0.s, z0.s, z2.s\n" + ".inst 0x04a277ff // sqrdmulh z31.s, z31.s, z2.s\n" + ".inst 0x04a277de // sqrdmulh z30.s, z30.s, z2.s\n" + ".inst 0x04a276d6 // sqrdmulh z22.s, z22.s, z2.s\n" + ".inst 0x04a277bd // sqrdmulh z29.s, z29.s, z2.s\n" + ".inst 0x04a27652 // sqrdmulh z18.s, z18.s, z2.s\n" + ".inst 0x04a2779c // sqrdmulh z28.s, z28.s, z2.s\n" + ".inst 0x04a276b5 // sqrdmulh z21.s, z21.s, z2.s\n" + ".inst 0x04a2777b // sqrdmulh z27.s, z27.s, z2.s\n" + ".inst 0x04a27631 // sqrdmulh z17.s, z17.s, z2.s\n" + ".inst 0x04a2775a // sqrdmulh z26.s, z26.s, z2.s\n" + ".inst 0x04a27694 // sqrdmulh z20.s, z20.s, z2.s\n" + ".inst 0x04a27739 // sqrdmulh z25.s, z25.s, z2.s\n" + ".inst 0x04a27718 // sqrdmulh z24.s, z24.s, z2.s\n" + ".inst 0x44828261 // srshl z1.s, p0/M, z1.s, z19.s\n" + ".inst 0x44828277 // srshl z23.s, p0/M, z23.s, z19.s\n" + ".inst 0x44828260 // srshl z0.s, p0/M, z0.s, z19.s\n" + ".inst 0x4482827f // srshl z31.s, p0/M, z31.s, z19.s\n" + ".inst 0x4482827e // srshl z30.s, p0/M, z30.s, z19.s\n" + ".inst 0x44828276 // srshl z22.s, p0/M, z22.s, z19.s\n" + ".inst 0x4482827d // srshl z29.s, p0/M, z29.s, z19.s\n" + ".inst 0x44828272 // srshl z18.s, p0/M, z18.s, z19.s\n" + ".inst 0x4482827c // srshl z28.s, p0/M, z28.s, z19.s\n" + ".inst 0x44828275 // srshl z21.s, p0/M, z21.s, z19.s\n" + ".inst 0x4482827b // srshl z27.s, p0/M, z27.s, z19.s\n" + ".inst 0x44828271 // srshl z17.s, p0/M, z17.s, z19.s\n" + ".inst 0x4482827a // srshl z26.s, p0/M, z26.s, z19.s\n" + ".inst 0x44828274 // srshl z20.s, p0/M, z20.s, z19.s\n" + ".inst 0x44828279 // srshl z25.s, p0/M, z25.s, z19.s\n" + ".inst 0x44828278 // srshl z24.s, p0/M, z24.s, z19.s\n" + "add z1.s, z1.s, z16.s\n" + "add z23.s, z23.s, z16.s\n" + "add z0.s, z0.s, z16.s\n" + "add z31.s, z31.s, z16.s\n" + "add z30.s, z30.s, z16.s\n" + "add z22.s, z22.s, z16.s\n" + "add z29.s, z29.s, z16.s\n" + "add z18.s, z18.s, z16.s\n" + "add z28.s, z28.s, z16.s\n" + "add z21.s, z21.s, z16.s\n" + "add z27.s, z27.s, z16.s\n" + "add z17.s, z17.s, z16.s\n" + "add z26.s, z26.s, z16.s\n" + "add z20.s, z20.s, z16.s\n" + "add z25.s, z25.s, z16.s\n" + "add z24.s, z24.s, z16.s\n" + "mov z16.s, #0x0\n" + "mov z19.s, #0xff\n" + "smax z1.s, p0/M, z1.s, z16.s\n" + "smax z23.s, p0/M, z23.s, z16.s\n" + "smax z0.s, p0/M, z0.s, z16.s\n" + "smax z31.s, p0/M, z31.s, z16.s\n" + "smax z30.s, p0/M, z30.s, z16.s\n" + "smax z22.s, p0/M, z22.s, z16.s\n" + "smax z29.s, p0/M, z29.s, z16.s\n" + "smax z18.s, p0/M, z18.s, z16.s\n" + "smax z28.s, p0/M, z28.s, z16.s\n" + "smax z21.s, p0/M, z21.s, z16.s\n" + "smax z27.s, p0/M, z27.s, z16.s\n" + "smax z17.s, p0/M, z17.s, z16.s\n" + "smax z26.s, p0/M, z26.s, z16.s\n" + "smax z20.s, p0/M, z20.s, z16.s\n" + "smax z25.s, p0/M, z25.s, z16.s\n" + "smax z24.s, p0/M, z24.s, z16.s\n" + "smin z1.s, p0/M, z1.s, z19.s\n" + "smin z23.s, p0/M, z23.s, z19.s\n" + "smin z0.s, p0/M, z0.s, z19.s\n" + "trn1 z23.h, z1.h, z23.h\n" + "smin z31.s, p0/M, z31.s, z19.s\n" + "smin z30.s, p0/M, z30.s, z19.s\n" + "trn1 z16.h, z0.h, z31.h\n" + "smin z22.s, p0/M, z22.s, z19.s\n" + "smin z29.s, p0/M, z29.s, z19.s\n" + "trn1 z22.h, z30.h, z22.h\n" + "smin z18.s, p0/M, z18.s, z19.s\n" + "smin z28.s, p0/M, z28.s, z19.s\n" + "trn1 z18.h, z29.h, z18.h\n" + "smin z21.s, p0/M, z21.s, z19.s\n" + "smin z27.s, p0/M, z27.s, z19.s\n" + "trn1 z21.h, z28.h, z21.h\n" + "smin z17.s, p0/M, z17.s, z19.s\n" + "smin z26.s, p0/M, z26.s, z19.s\n" + "trn1 z17.h, z27.h, z17.h\n" + "smin z20.s, p0/M, z20.s, z19.s\n" + "smin z25.s, p0/M, z25.s, z19.s\n" + "trn1 z20.h, z26.h, z20.h\n" + "smin z24.s, p0/M, z24.s, z19.s\n" + "trn1 z19.h, z25.h, z24.h\n" + "trn1 z16.b, z23.b, z16.b\n" + "trn1 z18.b, z22.b, z18.b\n" + "trn1 z17.b, z21.b, z17.b\n" + "st1b { z16.b }, p4, [%x[outptr], x9]\n" + "incb x9, ALL, MUL #4\n" + "trn1 z16.b, z20.b, z19.b\n" + "st1b { z18.b }, p3, [%x[outptr], x28]\n" + "incb x28, ALL, MUL #4\n" + "st1b { z17.b }, p2, [%x[outptr], x27]\n" + "incb x27, ALL, MUL #4\n" + "st1b { z16.b }, p1, [%x[outptr], x26]\n" + "incb x26, ALL, MUL #4\n" + "whilelt p1.b, x26, %x[n_channels]\n" + "b.any 1b\n" + "7:" // Single vector of channels + "whilelt p4.b, x9, %x[n_channels]\n" + "b.none 14f\n" + "8:" // Single vector of channels: Loop + "lsr x25, %x[n_valid_cells], #0x2\n" + "mov z5.b, #0x0\n" + "mov x24, %x[inptrs]\n" + "cbz x25, 11f\n" + "ldp x20, x22, [x24, #0x0]\n" + "subs x25, x25, #0x1\n" + "ld1b { z0.b }, p4/Z, [x20, x9]\n" + "ldp x21, x20, [x24, #0x10]\n" + "add x24, x24, #0x20\n" + "ld1b { z31.b }, p4/Z, [x22, x9]\n" + "ld1b { z23.b }, p4/Z, [x21, x9]\n" + "ld1b { z30.b }, p4/Z, [x20, x9]\n" + "beq 10f\n" + "9:" // Single vector of channels: Loop: 4 inputs loop + "movprfx z16, z0\n umax z16.b, p0/M, z16.b, z31.b\n" + "movprfx z17, z23\n umax z17.b, p0/M, z17.b, z30.b\n" + "ldp x23, x22, [x24, #0x0]\n" + "subs x25, x25, #0x1\n" + "umax z16.b, p0/M, z16.b, z17.b\n" + "ldp x21, x20, [x24, #0x10]\n" + "umax z5.b, p0/M, z5.b, z16.b\n" + "add x24, x24, #0x20\n" + "ld1b { z0.b }, p4/Z, [x23, x9]\n" + "ld1b { z31.b }, p4/Z, [x22, x9]\n" + "ld1b { z23.b }, p4/Z, [x21, x9]\n" + "ld1b { z30.b }, p4/Z, [x20, x9]\n" + "bgt 9b\n" + "10:" // Single vector of channels: Loop: 4 inputs tail + "movprfx z16, z0\n umax z16.b, p0/M, z16.b, z31.b\n" + "movprfx z17, z23\n umax z17.b, p0/M, z17.b, z30.b\n" + "umax z16.b, p0/M, z16.b, z17.b\n" + "umax z5.b, p0/M, z5.b, z16.b\n" + "11:" // Single vector of channels: Loop: After loop + "ands x21, %x[n_valid_cells], #0x3\n" + "beq 13f\n" + "12:" // Single vector of channels: Loop: Single input loop + "ldr x20, [x24], #0x8\n" + "ld1b { z16.b }, p4/Z, [x20, x9]\n" + "subs x21, x21, #0x1\n" + "umax z5.b, p0/M, z5.b, z16.b\n" + "bgt 12b\n" + "13:" // Single vector of channels: Loop: Single input loop: End + "add x20, %x[quant_params], %[offsetof_qp_input_offset]\n" + "ld1rw { z18.s }, p0/Z, [x20]\n" + ".inst 0x4508a8b1 // ushllb z17.h, z5.b, #0x0\n" + ".inst 0x4508acb0 // ushllt z16.h, z5.b, #0x0\n" + "neg z18.s, p0/M, z18.s\n" + ".inst 0x45914257 // saddwb z23.s, z18.s, z17.h\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_left_shift]\n" + "ld1rw { z22.s }, p0/Z, [x20]\n" + ".inst 0x45914655 // saddwt z21.s, z18.s, z17.h\n" + ".inst 0x45904254 // saddwb z20.s, z18.s, z16.h\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_mul]\n" + "ld1rw { z19.s }, p0/Z, [x20]\n" + ".inst 0x45904652 // saddwt z18.s, z18.s, z16.h\n" + ".inst 0x448282d7 // srshl z23.s, p0/M, z23.s, z22.s\n" + "add x20, %x[quant_params], %[offsetof_qp_per_layer_right_shift]\n" + "ld1rw { z17.s }, p0/Z, [x20]\n" + ".inst 0x448282d5 // srshl z21.s, p0/M, z21.s, z22.s\n" + ".inst 0x448282d4 // srshl z20.s, p0/M, z20.s, z22.s\n" + "add x20, %x[quant_params], %[offsetof_qp_output_offset]\n" + "ld1rw { z16.s }, p0/Z, [x20]\n" + ".inst 0x448282d2 // srshl z18.s, p0/M, z18.s, z22.s\n" + ".inst 0x04b376f7 // sqrdmulh z23.s, z23.s, z19.s\n" + ".inst 0x04b376b5 // sqrdmulh z21.s, z21.s, z19.s\n" + ".inst 0x04b37694 // sqrdmulh z20.s, z20.s, z19.s\n" + ".inst 0x04b37652 // sqrdmulh z18.s, z18.s, z19.s\n" + ".inst 0x44828237 // srshl z23.s, p0/M, z23.s, z17.s\n" + ".inst 0x44828235 // srshl z21.s, p0/M, z21.s, z17.s\n" + ".inst 0x44828234 // srshl z20.s, p0/M, z20.s, z17.s\n" + ".inst 0x44828232 // srshl z18.s, p0/M, z18.s, z17.s\n" + "add z23.s, z23.s, z16.s\n" + "add z21.s, z21.s, z16.s\n" + "add z20.s, z20.s, z16.s\n" + "add z18.s, z18.s, z16.s\n" + "mov z17.s, #0x0\n" + "mov z16.s, #0xff\n" + "smax z23.s, p0/M, z23.s, z17.s\n" + "smax z21.s, p0/M, z21.s, z17.s\n" + "smax z20.s, p0/M, z20.s, z17.s\n" + "smax z18.s, p0/M, z18.s, z17.s\n" + "smin z23.s, p0/M, z23.s, z16.s\n" + "smin z21.s, p0/M, z21.s, z16.s\n" + "smin z20.s, p0/M, z20.s, z16.s\n" + "trn1 z17.h, z23.h, z21.h\n" + "smin z18.s, p0/M, z18.s, z16.s\n" + "trn1 z16.h, z20.h, z18.h\n" + "trn1 z16.b, z17.b, z16.b\n" + "st1b { z16.b }, p4, [%x[outptr], x9]\n" + "incb x9\n" + "whilelt p4.b, x9, %x[n_channels]\n" + "b.any 8b\n" + "14:" // End + ".inst 0xd503467f // SMSTOP\n" + : + : [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", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "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_SME) |