diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp | 365 |
1 files changed, 183 insertions, 182 deletions
diff --git a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp index 51a69a42be..f3f4950a1f 100644 --- a/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/pooling/kernels/sve_u8_nhwc_avg_generic_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,11 +23,12 @@ */ #include <cstdint> +#include <cstddef> #include <cstring> #include <cmath> -#if defined(__ARM_FEATURE_SVE) && defined(SVE2) +#if defined(ARM_COMPUTE_ENABLE_SVE) namespace arm_conv { namespace pooling { @@ -84,30 +85,31 @@ void sve_u8_nhwc_avg_generic_depthfirst_impl( f_rescale_value *= 2.0f; } - rescale_value = static_cast<int32_t>(round(f_rescale_value * static_cast<float>(1ll << 31))); - if (static_cast<int64_t>(rescale_value) == (1ll << 31)) + int64_t long_rescale_value = round(f_rescale_value * static_cast<float>(1ll << 31)); + if (long_rescale_value == (1ll << 31)) { shift_value++; - rescale_value >>= 1; + long_rescale_value >>= 1; } + rescale_value = static_cast<int32_t>(long_rescale_value); } __asm__ __volatile__( - "ptrue p4.b\n" - "mov x26, #0x0\n" - "cntb x25\n" - "cntb x24, ALL, MUL #2\n" - "cntb x23, ALL, MUL #3\n" + "mov x27, #0x0\n" + "cntb x26\n" + "cntb x25, ALL, MUL #2\n" + "cntb x24, ALL, MUL #3\n" + "whilelt p4.b, x27, %x[n_channels]\n" "whilelt p3.b, x26, %x[n_channels]\n" "whilelt p2.b, x25, %x[n_channels]\n" "whilelt p1.b, x24, %x[n_channels]\n" - "whilelt p0.b, x23, %x[n_channels]\n" + "ptrue p0.b\n" "b.none 7f\n" "1:" // 4-vectors of channels + "lsr x23, %x[n_valid_cells], #0x1\n" "mov z15.s, #0x0\n" - "mov x19, %x[inptrs]\n" "mov z14.s, #0x0\n" - "lsr x22, %x[n_valid_cells], #0x1\n" + "mov x22, %x[inptrs]\n" "mov z13.s, #0x0\n" "mov z12.s, #0x0\n" "mov z11.s, #0x0\n" @@ -122,43 +124,43 @@ void sve_u8_nhwc_avg_generic_depthfirst_impl( "mov z2.s, #0x0\n" "mov z1.s, #0x0\n" "mov z0.s, #0x0\n" - "cbz x22, 4f\n" - "ldp x21, x20, [x19, #0x0]\n" - "ld1b { z31.b }, p3/Z, [x21, x26]\n" - "add x19, x19, #0x10\n" - "ld1b { z30.b }, p3/Z, [x20, x26]\n" - "subs x22, x22, #0x1\n" - "ld1b { z29.b }, p2/Z, [x21, x25]\n" - "ld1b { z28.b }, p2/Z, [x20, x25]\n" - "ld1b { z27.b }, p1/Z, [x21, x24]\n" - "ld1b { z26.b }, p1/Z, [x20, x24]\n" - "ld1b { z25.b }, p0/Z, [x21, x23]\n" - "ld1b { z24.b }, p0/Z, [x20, x23]\n" + "cbz x23, 4f\n" + "ldp x21, x20, [x22, #0x0]\n" + "subs x23, x23, #0x1\n" + "add x22, x22, #0x10\n" + "ld1b { z31.b }, p4/Z, [x21, x27]\n" + "ld1b { z30.b }, p4/Z, [x20, x27]\n" + "ld1b { z29.b }, p3/Z, [x21, x26]\n" + "ld1b { z28.b }, p3/Z, [x20, x26]\n" + "ld1b { z27.b }, p2/Z, [x21, x25]\n" + "ld1b { z26.b }, p2/Z, [x20, x25]\n" + "ld1b { z25.b }, p1/Z, [x21, x24]\n" + "ld1b { z24.b }, p1/Z, [x20, x24]\n" "beq 3f\n" "2:" // 4-vectors of channels: 2 inputs loop ".inst 0x455e0bf7 // uaddlb z23.h, z31.b, z30.b\n" - "ldp x21, x20, [x19, #0x0]\n" - "add x19, x19, #0x10\n" ".inst 0x455e0ff6 // uaddlt z22.h, z31.b, z30.b\n" - "ld1b { z31.b }, p3/Z, [x21, x26]\n" + "ldp x21, x20, [x22, #0x0]\n" + "subs x23, x23, #0x1\n" ".inst 0x455c0bb5 // uaddlb z21.h, z29.b, z28.b\n" - "subs x22, x22, #0x1\n" ".inst 0x455c0fb4 // uaddlt z20.h, z29.b, z28.b\n" - "ld1b { z30.b }, p3/Z, [x20, x26]\n" + "add x22, x22, #0x10\n" + "ld1b { z31.b }, p4/Z, [x21, x27]\n" ".inst 0x455a0b73 // uaddlb z19.h, z27.b, z26.b\n" - "ld1b { z29.b }, p2/Z, [x21, x25]\n" ".inst 0x455a0f72 // uaddlt z18.h, z27.b, z26.b\n" - "ld1b { z28.b }, p2/Z, [x20, x25]\n" + "ld1b { z30.b }, p4/Z, [x20, x27]\n" + "ld1b { z29.b }, p3/Z, [x21, x26]\n" ".inst 0x45580b31 // uaddlb z17.h, z25.b, z24.b\n" - "ld1b { z27.b }, p1/Z, [x21, x24]\n" ".inst 0x45580f30 // uaddlt z16.h, z25.b, z24.b\n" - "ld1b { z26.b }, p1/Z, [x20, x24]\n" + "ld1b { z28.b }, p3/Z, [x20, x26]\n" + "ld1b { z27.b }, p2/Z, [x21, x25]\n" ".inst 0x459749ef // uaddwb z15.s, z15.s, z23.h\n" - "ld1b { z25.b }, p0/Z, [x21, x23]\n" ".inst 0x45974dce // uaddwt z14.s, z14.s, z23.h\n" - "ld1b { z24.b }, p0/Z, [x20, x23]\n" + "ld1b { z26.b }, p2/Z, [x20, x25]\n" + "ld1b { z25.b }, p1/Z, [x21, x24]\n" ".inst 0x459649ad // uaddwb z13.s, z13.s, z22.h\n" ".inst 0x45964d8c // uaddwt z12.s, z12.s, z22.h\n" + "ld1b { z24.b }, p1/Z, [x20, x24]\n" ".inst 0x4595496b // uaddwb z11.s, z11.s, z21.h\n" ".inst 0x45954d4a // uaddwt z10.s, z10.s, z21.h\n" ".inst 0x45944929 // uaddwb z9.s, z9.s, z20.h\n" @@ -198,219 +200,218 @@ void sve_u8_nhwc_avg_generic_depthfirst_impl( ".inst 0x45904821 // uaddwb z1.s, z1.s, z16.h\n" ".inst 0x45904c00 // uaddwt z0.s, z0.s, z16.h\n" "4:" // 4-vectors of channels: After loop - "ands x20, %x[n_valid_cells], #0x1\n" + "ands x21, %x[n_valid_cells], #0x1\n" "beq 6f\n" "5:" // 4-vectors of channels: Single input loop - "ldr x21, [x19], #0x8\n" - "subs x20, x20, #0x1\n" - "ld1b { z31.b }, p3/Z, [x21, x26]\n" - ".inst 0x4508abf1 // ushllb z17.h, z31.b, #0x0\n" - "ld1b { z29.b }, p2/Z, [x21, x25]\n" - ".inst 0x4508aff0 // ushllt z16.h, z31.b, #0x0\n" - "ld1b { z27.b }, p1/Z, [x21, x24]\n" - ".inst 0x459149ef // uaddwb z15.s, z15.s, z17.h\n" - "ld1b { z25.b }, p0/Z, [x21, x23]\n" - ".inst 0x45914dce // uaddwt z14.s, z14.s, z17.h\n" - ".inst 0x459049ad // uaddwb z13.s, z13.s, z16.h\n" - ".inst 0x45904d8c // uaddwt z12.s, z12.s, z16.h\n" - ".inst 0x4508abb0 // ushllb z16.h, z29.b, #0x0\n" - ".inst 0x4590496b // uaddwb z11.s, z11.s, z16.h\n" - ".inst 0x45904d4a // uaddwt z10.s, z10.s, z16.h\n" - ".inst 0x4508afb0 // ushllt z16.h, z29.b, #0x0\n" - ".inst 0x45904929 // uaddwb z9.s, z9.s, z16.h\n" - ".inst 0x45904d08 // uaddwt z8.s, z8.s, z16.h\n" - ".inst 0x4508ab70 // ushllb z16.h, z27.b, #0x0\n" - ".inst 0x459048e7 // uaddwb z7.s, z7.s, z16.h\n" - ".inst 0x45904cc6 // uaddwt z6.s, z6.s, z16.h\n" - ".inst 0x4508af70 // ushllt z16.h, z27.b, #0x0\n" - ".inst 0x459048a5 // uaddwb z5.s, z5.s, z16.h\n" - ".inst 0x45904c84 // uaddwt z4.s, z4.s, z16.h\n" - ".inst 0x4508ab30 // ushllb z16.h, z25.b, #0x0\n" - ".inst 0x45904863 // uaddwb z3.s, z3.s, z16.h\n" - ".inst 0x45904c42 // uaddwt z2.s, z2.s, z16.h\n" - ".inst 0x4508af30 // ushllt z16.h, z25.b, #0x0\n" + "ldr x20, [x22], #0x8\n" + "ld1b { z16.b }, p4/Z, [x20, x27]\n" + ".inst 0x4508aa17 // ushllb z23.h, z16.b, #0x0\n" + ".inst 0x4508ae16 // ushllt z22.h, z16.b, #0x0\n" + "ld1b { z16.b }, p3/Z, [x20, x26]\n" + "ld1b { z17.b }, p2/Z, [x20, x25]\n" + ".inst 0x4508aa15 // ushllb z21.h, z16.b, #0x0\n" + ".inst 0x4508ae14 // ushllt z20.h, z16.b, #0x0\n" + "ld1b { z16.b }, p1/Z, [x20, x24]\n" + ".inst 0x4508aa33 // ushllb z19.h, z17.b, #0x0\n" + ".inst 0x4508ae32 // ushllt z18.h, z17.b, #0x0\n" + "subs x21, x21, #0x1\n" + ".inst 0x4508aa11 // ushllb z17.h, z16.b, #0x0\n" + ".inst 0x4508ae10 // ushllt z16.h, z16.b, #0x0\n" + ".inst 0x459749ef // uaddwb z15.s, z15.s, z23.h\n" + ".inst 0x45974dce // uaddwt z14.s, z14.s, z23.h\n" + ".inst 0x459649ad // uaddwb z13.s, z13.s, z22.h\n" + ".inst 0x45964d8c // uaddwt z12.s, z12.s, z22.h\n" + ".inst 0x4595496b // uaddwb z11.s, z11.s, z21.h\n" + ".inst 0x45954d4a // uaddwt z10.s, z10.s, z21.h\n" + ".inst 0x45944929 // uaddwb z9.s, z9.s, z20.h\n" + ".inst 0x45944d08 // uaddwt z8.s, z8.s, z20.h\n" + ".inst 0x459348e7 // uaddwb z7.s, z7.s, z19.h\n" + ".inst 0x45934cc6 // uaddwt z6.s, z6.s, z19.h\n" + ".inst 0x459248a5 // uaddwb z5.s, z5.s, z18.h\n" + ".inst 0x45924c84 // uaddwt z4.s, z4.s, z18.h\n" + ".inst 0x45914863 // uaddwb z3.s, z3.s, z17.h\n" + ".inst 0x45914c42 // uaddwt z2.s, z2.s, z17.h\n" ".inst 0x45904821 // uaddwb z1.s, z1.s, z16.h\n" ".inst 0x45904c00 // uaddwt z0.s, z0.s, z16.h\n" "bgt 5b\n" "6:" // 4-vectors of channels: Single input loop: End - "mov z20.s, #0x0\n" - "ld1rw { z17.s }, p4/Z, [%x[rescale_ptr]]\n" - "mov z19.s, #0xff\n" - "ld1rw { z16.s }, p4/Z, [%x[shift_ptr]]\n" + "ld1rw { z17.s }, p0/Z, [%x[rescale_ptr]]\n" + "ld1rw { z16.s }, p0/Z, [%x[shift_ptr]]\n" ".inst 0x04b175ef // sqdmulh z15.s, z15.s, z17.s\n" ".inst 0x04b175ce // sqdmulh z14.s, z14.s, z17.s\n" ".inst 0x04b175ad // sqdmulh z13.s, z13.s, z17.s\n" ".inst 0x04b1758c // sqdmulh z12.s, z12.s, z17.s\n" + ".inst 0x4482820f // srshl z15.s, p0/M, z15.s, z16.s\n" + ".inst 0x4482820e // srshl z14.s, p0/M, z14.s, z16.s\n" ".inst 0x04b1756b // sqdmulh z11.s, z11.s, z17.s\n" ".inst 0x04b1754a // sqdmulh z10.s, z10.s, z17.s\n" + ".inst 0x4482820d // srshl z13.s, p0/M, z13.s, z16.s\n" + ".inst 0x4482820c // srshl z12.s, p0/M, z12.s, z16.s\n" ".inst 0x04b17529 // sqdmulh z9.s, z9.s, z17.s\n" ".inst 0x04b17508 // sqdmulh z8.s, z8.s, z17.s\n" + ".inst 0x4482820b // srshl z11.s, p0/M, z11.s, z16.s\n" + ".inst 0x4482820a // srshl z10.s, p0/M, z10.s, z16.s\n" ".inst 0x04b174e7 // sqdmulh z7.s, z7.s, z17.s\n" ".inst 0x04b174c6 // sqdmulh z6.s, z6.s, z17.s\n" + ".inst 0x44828209 // srshl z9.s, p0/M, z9.s, z16.s\n" + ".inst 0x44828208 // srshl z8.s, p0/M, z8.s, z16.s\n" ".inst 0x04b174a5 // sqdmulh z5.s, z5.s, z17.s\n" ".inst 0x04b17484 // sqdmulh z4.s, z4.s, z17.s\n" + ".inst 0x44828207 // srshl z7.s, p0/M, z7.s, z16.s\n" + ".inst 0x44828206 // srshl z6.s, p0/M, z6.s, z16.s\n" ".inst 0x04b17463 // sqdmulh z3.s, z3.s, z17.s\n" ".inst 0x04b17442 // sqdmulh z2.s, z2.s, z17.s\n" + ".inst 0x44828205 // srshl z5.s, p0/M, z5.s, z16.s\n" + ".inst 0x44828204 // srshl z4.s, p0/M, z4.s, z16.s\n" ".inst 0x04b17421 // sqdmulh z1.s, z1.s, z17.s\n" ".inst 0x04b17400 // sqdmulh z0.s, z0.s, z17.s\n" - ".inst 0x4482920f // srshl z15.s, p4/M, z15.s, z16.s\n" - ".inst 0x4482920e // srshl z14.s, p4/M, z14.s, z16.s\n" - ".inst 0x4482920d // srshl z13.s, p4/M, z13.s, z16.s\n" - ".inst 0x4482920c // srshl z12.s, p4/M, z12.s, z16.s\n" - ".inst 0x4482920b // srshl z11.s, p4/M, z11.s, z16.s\n" - ".inst 0x4482920a // srshl z10.s, p4/M, z10.s, z16.s\n" - ".inst 0x44829209 // srshl z9.s, p4/M, z9.s, z16.s\n" - ".inst 0x44829208 // srshl z8.s, p4/M, z8.s, z16.s\n" - ".inst 0x44829207 // srshl z7.s, p4/M, z7.s, z16.s\n" - ".inst 0x44829206 // srshl z6.s, p4/M, z6.s, z16.s\n" - ".inst 0x44829205 // srshl z5.s, p4/M, z5.s, z16.s\n" - ".inst 0x44829204 // srshl z4.s, p4/M, z4.s, z16.s\n" - ".inst 0x44829203 // srshl z3.s, p4/M, z3.s, z16.s\n" - ".inst 0x44829202 // srshl z2.s, p4/M, z2.s, z16.s\n" - ".inst 0x44829201 // srshl z1.s, p4/M, z1.s, z16.s\n" - ".inst 0x44829200 // srshl z0.s, p4/M, z0.s, z16.s\n" - "smax z15.s, p4/M, z15.s, z20.s\n" - "smax z14.s, p4/M, z14.s, z20.s\n" - "smax z13.s, p4/M, z13.s, z20.s\n" - "smax z12.s, p4/M, z12.s, z20.s\n" - "smin z15.s, p4/M, z15.s, z19.s\n" - "smin z14.s, p4/M, z14.s, z19.s\n" - "smin z13.s, p4/M, z13.s, z19.s\n" - "smin z12.s, p4/M, z12.s, z19.s\n" - "smax z11.s, p4/M, z11.s, z20.s\n" + ".inst 0x44828203 // srshl z3.s, p0/M, z3.s, z16.s\n" + ".inst 0x44828202 // srshl z2.s, p0/M, z2.s, z16.s\n" + ".inst 0x44828201 // srshl z1.s, p0/M, z1.s, z16.s\n" + ".inst 0x44828200 // srshl z0.s, p0/M, z0.s, z16.s\n" + "mov z16.s, #0x0\n" + "mov z18.s, #0xff\n" + "smax z15.s, p0/M, z15.s, z16.s\n" + "smax z14.s, p0/M, z14.s, z16.s\n" + "smax z13.s, p0/M, z13.s, z16.s\n" + "smax z12.s, p0/M, z12.s, z16.s\n" + "smax z11.s, p0/M, z11.s, z16.s\n" + "smax z10.s, p0/M, z10.s, z16.s\n" + "smax z9.s, p0/M, z9.s, z16.s\n" + "smax z8.s, p0/M, z8.s, z16.s\n" + "smax z7.s, p0/M, z7.s, z16.s\n" + "smax z6.s, p0/M, z6.s, z16.s\n" + "smax z5.s, p0/M, z5.s, z16.s\n" + "smax z4.s, p0/M, z4.s, z16.s\n" + "smax z3.s, p0/M, z3.s, z16.s\n" + "smax z2.s, p0/M, z2.s, z16.s\n" + "smax z1.s, p0/M, z1.s, z16.s\n" + "smax z0.s, p0/M, z0.s, z16.s\n" + "smin z15.s, p0/M, z15.s, z18.s\n" + "smin z14.s, p0/M, z14.s, z18.s\n" "trn1 z17.h, z15.h, z14.h\n" - "smax z10.s, p4/M, z10.s, z20.s\n" + "smin z13.s, p0/M, z13.s, z18.s\n" + "smin z12.s, p0/M, z12.s, z18.s\n" "trn1 z16.h, z13.h, z12.h\n" - "smin z11.s, p4/M, z11.s, z19.s\n" "trn1 z16.b, z17.b, z16.b\n" - "st1b { z16.b }, p3, [%x[outptr], x26]\n" - "smin z10.s, p4/M, z10.s, z19.s\n" - "incb x26, ALL, MUL #4\n" - "smax z9.s, p4/M, z9.s, z20.s\n" - "smax z8.s, p4/M, z8.s, z20.s\n" - "smax z7.s, p4/M, z7.s, z20.s\n" - "smax z6.s, p4/M, z6.s, z20.s\n" - "trn1 z18.h, z11.h, z10.h\n" - "smin z9.s, p4/M, z9.s, z19.s\n" - "smin z8.s, p4/M, z8.s, z19.s\n" - "smin z7.s, p4/M, z7.s, z19.s\n" - "smin z6.s, p4/M, z6.s, z19.s\n" - "smax z5.s, p4/M, z5.s, z20.s\n" + "smin z11.s, p0/M, z11.s, z18.s\n" + "smin z10.s, p0/M, z10.s, z18.s\n" + "trn1 z17.h, z11.h, z10.h\n" + "st1b { z16.b }, p4, [%x[outptr], x27]\n" + "smin z9.s, p0/M, z9.s, z18.s\n" + "smin z8.s, p0/M, z8.s, z18.s\n" "trn1 z16.h, z9.h, z8.h\n" - "smax z4.s, p4/M, z4.s, z20.s\n" + "trn1 z16.b, z17.b, z16.b\n" + "smin z7.s, p0/M, z7.s, z18.s\n" + "smin z6.s, p0/M, z6.s, z18.s\n" "trn1 z17.h, z7.h, z6.h\n" - "trn1 z16.b, z18.b, z16.b\n" - "st1b { z16.b }, p2, [%x[outptr], x25]\n" - "smin z5.s, p4/M, z5.s, z19.s\n" - "incb x25, ALL, MUL #4\n" - "smin z4.s, p4/M, z4.s, z19.s\n" - "smax z3.s, p4/M, z3.s, z20.s\n" - "smax z2.s, p4/M, z2.s, z20.s\n" - "smax z1.s, p4/M, z1.s, z20.s\n" - "smax z0.s, p4/M, z0.s, z20.s\n" + "st1b { z16.b }, p3, [%x[outptr], x26]\n" + "smin z5.s, p0/M, z5.s, z18.s\n" + "smin z4.s, p0/M, z4.s, z18.s\n" "trn1 z16.h, z5.h, z4.h\n" - "smin z3.s, p4/M, z3.s, z19.s\n" "trn1 z16.b, z17.b, z16.b\n" - "st1b { z16.b }, p1, [%x[outptr], x24]\n" - "smin z2.s, p4/M, z2.s, z19.s\n" - "incb x24, ALL, MUL #4\n" - "smin z1.s, p4/M, z1.s, z19.s\n" - "smin z0.s, p4/M, z0.s, z19.s\n" + "smin z3.s, p0/M, z3.s, z18.s\n" + "smin z2.s, p0/M, z2.s, z18.s\n" "trn1 z17.h, z3.h, z2.h\n" + "st1b { z16.b }, p2, [%x[outptr], x25]\n" + "smin z1.s, p0/M, z1.s, z18.s\n" + "smin z0.s, p0/M, z0.s, z18.s\n" "trn1 z16.h, z1.h, z0.h\n" "trn1 z16.b, z17.b, z16.b\n" - "st1b { z16.b }, p0, [%x[outptr], x23]\n" - "incb x23, ALL, MUL #4\n" - "whilelt p0.b, x23, %x[n_channels]\n" + "st1b { z16.b }, p1, [%x[outptr], x24]\n" + "incb x24, ALL, MUL #4\n" + "whilelt p1.b, x24, %x[n_channels]\n" + "incb x27, ALL, MUL #4\n" + "incb x26, ALL, MUL #4\n" + "incb x25, ALL, MUL #4\n" "b.any 1b\n" "7:" // Single vector of channels - "whilelt p3.b, x26, %x[n_channels]\n" + "whilelt p4.b, x27, %x[n_channels]\n" "b.none 14f\n" "8:" // Single vector of channels: Loop + "lsr x23, %x[n_valid_cells], #0x1\n" "mov z15.s, #0x0\n" - "mov x19, %x[inptrs]\n" "mov z14.s, #0x0\n" - "lsr x22, %x[n_valid_cells], #0x1\n" + "mov x22, %x[inptrs]\n" "mov z13.s, #0x0\n" "mov z12.s, #0x0\n" - "cbz x22, 11f\n" - "ldp x21, x20, [x19, #0x0]\n" - "ld1b { z31.b }, p3/Z, [x21, x26]\n" - "add x19, x19, #0x10\n" - "ld1b { z30.b }, p3/Z, [x20, x26]\n" - "subs x22, x22, #0x1\n" + "cbz x23, 11f\n" + "ldp x21, x20, [x22, #0x0]\n" + "subs x23, x23, #0x1\n" + "add x22, x22, #0x10\n" + "ld1b { z31.b }, p4/Z, [x21, x27]\n" + "ld1b { z30.b }, p4/Z, [x20, x27]\n" "beq 10f\n" "9:" // Single vector of channels: Loop: 2 inputs loop - ".inst 0x455e0bf7 // uaddlb z23.h, z31.b, z30.b\n" - "ldp x21, x20, [x19, #0x0]\n" - "add x19, x19, #0x10\n" - ".inst 0x455e0ff6 // uaddlt z22.h, z31.b, z30.b\n" - "ld1b { z31.b }, p3/Z, [x21, x26]\n" - ".inst 0x459749ef // uaddwb z15.s, z15.s, z23.h\n" - "subs x22, x22, #0x1\n" - ".inst 0x45974dce // uaddwt z14.s, z14.s, z23.h\n" - "ld1b { z30.b }, p3/Z, [x20, x26]\n" - ".inst 0x459649ad // uaddwb z13.s, z13.s, z22.h\n" - ".inst 0x45964d8c // uaddwt z12.s, z12.s, z22.h\n" + ".inst 0x455e0bf1 // uaddlb z17.h, z31.b, z30.b\n" + ".inst 0x455e0ff0 // uaddlt z16.h, z31.b, z30.b\n" + "ldp x21, x20, [x22, #0x0]\n" + "subs x23, x23, #0x1\n" + ".inst 0x459149ef // uaddwb z15.s, z15.s, z17.h\n" + ".inst 0x45914dce // uaddwt z14.s, z14.s, z17.h\n" + "add x22, x22, #0x10\n" + "ld1b { z31.b }, p4/Z, [x21, x27]\n" + ".inst 0x459049ad // uaddwb z13.s, z13.s, z16.h\n" + ".inst 0x45904d8c // uaddwt z12.s, z12.s, z16.h\n" + "ld1b { z30.b }, p4/Z, [x20, x27]\n" "bgt 9b\n" "10:" // Single vector of channels: Loop: 2 inputs tail - ".inst 0x455e0bf7 // uaddlb z23.h, z31.b, z30.b\n" - ".inst 0x455e0ff6 // uaddlt z22.h, z31.b, z30.b\n" - ".inst 0x459749ef // uaddwb z15.s, z15.s, z23.h\n" - ".inst 0x45974dce // uaddwt z14.s, z14.s, z23.h\n" - ".inst 0x459649ad // uaddwb z13.s, z13.s, z22.h\n" - ".inst 0x45964d8c // uaddwt z12.s, z12.s, z22.h\n" + ".inst 0x455e0bf1 // uaddlb z17.h, z31.b, z30.b\n" + ".inst 0x455e0ff0 // uaddlt z16.h, z31.b, z30.b\n" + ".inst 0x459149ef // uaddwb z15.s, z15.s, z17.h\n" + ".inst 0x45914dce // uaddwt z14.s, z14.s, z17.h\n" + ".inst 0x459049ad // uaddwb z13.s, z13.s, z16.h\n" + ".inst 0x45904d8c // uaddwt z12.s, z12.s, z16.h\n" "11:" // Single vector of channels: Loop: After loop - "ands x20, %x[n_valid_cells], #0x1\n" + "ands x21, %x[n_valid_cells], #0x1\n" "beq 13f\n" "12:" // Single vector of channels: Loop: Single input loop - "ldr x21, [x19], #0x8\n" - "subs x20, x20, #0x1\n" - "ld1b { z31.b }, p3/Z, [x21, x26]\n" - ".inst 0x4508abf1 // ushllb z17.h, z31.b, #0x0\n" - ".inst 0x4508aff0 // ushllt z16.h, z31.b, #0x0\n" + "ldr x20, [x22], #0x8\n" + "ld1b { z16.b }, p4/Z, [x20, x27]\n" + ".inst 0x4508aa11 // ushllb z17.h, z16.b, #0x0\n" + ".inst 0x4508ae10 // ushllt z16.h, z16.b, #0x0\n" + "subs x21, x21, #0x1\n" ".inst 0x459149ef // uaddwb z15.s, z15.s, z17.h\n" ".inst 0x45914dce // uaddwt z14.s, z14.s, z17.h\n" ".inst 0x459049ad // uaddwb z13.s, z13.s, z16.h\n" ".inst 0x45904d8c // uaddwt z12.s, z12.s, z16.h\n" "bgt 12b\n" "13:" // Single vector of channels: Loop: Single input loop: End - "mov z20.s, #0x0\n" - "ld1rw { z17.s }, p4/Z, [%x[rescale_ptr]]\n" - "mov z19.s, #0xff\n" - "ld1rw { z16.s }, p4/Z, [%x[shift_ptr]]\n" + "ld1rw { z17.s }, p0/Z, [%x[rescale_ptr]]\n" + "ld1rw { z16.s }, p0/Z, [%x[shift_ptr]]\n" ".inst 0x04b175ef // sqdmulh z15.s, z15.s, z17.s\n" ".inst 0x04b175ce // sqdmulh z14.s, z14.s, z17.s\n" ".inst 0x04b175ad // sqdmulh z13.s, z13.s, z17.s\n" ".inst 0x04b1758c // sqdmulh z12.s, z12.s, z17.s\n" - ".inst 0x4482920f // srshl z15.s, p4/M, z15.s, z16.s\n" - ".inst 0x4482920e // srshl z14.s, p4/M, z14.s, z16.s\n" - ".inst 0x4482920d // srshl z13.s, p4/M, z13.s, z16.s\n" - ".inst 0x4482920c // srshl z12.s, p4/M, z12.s, z16.s\n" - "smax z15.s, p4/M, z15.s, z20.s\n" - "smax z14.s, p4/M, z14.s, z20.s\n" - "smax z13.s, p4/M, z13.s, z20.s\n" - "smax z12.s, p4/M, z12.s, z20.s\n" - "smin z15.s, p4/M, z15.s, z19.s\n" - "smin z14.s, p4/M, z14.s, z19.s\n" - "smin z13.s, p4/M, z13.s, z19.s\n" - "smin z12.s, p4/M, z12.s, z19.s\n" + ".inst 0x4482820f // srshl z15.s, p0/M, z15.s, z16.s\n" + ".inst 0x4482820e // srshl z14.s, p0/M, z14.s, z16.s\n" + ".inst 0x4482820d // srshl z13.s, p0/M, z13.s, z16.s\n" + ".inst 0x4482820c // srshl z12.s, p0/M, z12.s, z16.s\n" + "mov z17.s, #0x0\n" + "mov z16.s, #0xff\n" + "smax z15.s, p0/M, z15.s, z17.s\n" + "smax z14.s, p0/M, z14.s, z17.s\n" + "smax z13.s, p0/M, z13.s, z17.s\n" + "smax z12.s, p0/M, z12.s, z17.s\n" + "smin z15.s, p0/M, z15.s, z16.s\n" + "smin z14.s, p0/M, z14.s, z16.s\n" "trn1 z17.h, z15.h, z14.h\n" + "smin z13.s, p0/M, z13.s, z16.s\n" + "smin z12.s, p0/M, z12.s, z16.s\n" "trn1 z16.h, z13.h, z12.h\n" "trn1 z16.b, z17.b, z16.b\n" - "st1b { z16.b }, p3, [%x[outptr], x26]\n" - "incb x26\n" - "whilelt p3.b, x26, %x[n_channels]\n" + "st1b { z16.b }, p4, [%x[outptr], x27]\n" + "incb x27\n" + "whilelt p4.b, x27, %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), [outptr] "r" (outptr), [rescale_ptr] "r" (&rescale_value), [shift_ptr] "r" (&shift_value) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "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" + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "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_FEATURE_SVE) && defined(SVE2) +#endif // defined(ARM_COMPUTE_ENABLE_SVE) |