From d02d5edfa15ba6c04a9986a8a362a945cb38ac31 Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Fri, 22 Jan 2021 09:47:04 +0000 Subject: Integrate improved CPU depthwise convolution kernels * Replace assembly kernels for depthwise convolution with more optimized ones. * Add int8 assembly kernels. * Fix implicit padding on optimized kernels Resolves: COMPMID-3867, COMPMID-4361 Change-Id: I0b0867e05f61be4f368f62190d55e14d0ab3ebf2 Signed-off-by: Michele Di Giorgio Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5622 Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas --- .../generic.cpp | 454 +++++++++++++++++++++ 1 file changed, 454 insertions(+) create mode 100644 src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp new file mode 100644 index 0000000000..0124370067 --- /dev/null +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst/generic.cpp @@ -0,0 +1,454 @@ +/* + * 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 +#include + +#if defined(__ARM_FEATURE_SVE) + +namespace arm_conv { +namespace depthwise { + +void sve_fp32_packed_to_nhwc_generic_with_multiplier_output2x8_mla_depthfirst_impl( + const float *const *const inptrs, + float *const *const outptrs, + const float *weights, + const float *bias, + const unsigned int kernel_points, + const unsigned int n_output_channels, + const float activation_min, + const float activation_max +) +{ + const float minmax_vals[2] = { activation_min, activation_max }; + + __asm__ __volatile__( + "ptrue p1.b\n" + "ld1rw { z11.s }, p1/Z, [%x[minmax_vals]]\n" + "mov x28, #0x0\n" + "ld1rw { z10.s }, p1/Z, [%x[minmax_vals], #4]\n" + "whilelt p0.s, x28, %x[n_output_channels]\n" + "1:" // Output channel loop + "mov z16.b, #0x0\n" + "cbz %x[bias], 2f\n" + "ld1w { z16.s }, p0/Z, [%x[bias], x28, LSL #2]\n" + "2:" // Output channel loop: Load bias: Done + "mov z9.d, z16.d\n" + "ld1w { z8.s }, p1/Z, [%x[weights]]\n" + "mov x20, %x[inptrs]\n" + "mov z31.d, z16.d\n" + "ldp x24, x27, [x20], #0x10\n" + "lsr x19, %x[kernel_points], #0x1\n" + "mov z30.d, z16.d\n" + "ld1rqw { z7.s }, p1/Z, [x24]\n" + "mov z29.d, z16.d\n" + "addvl %x[weights], %x[weights], #1\n" + "mov z28.d, z16.d\n" + "ld1rqw { z6.s }, p1/Z, [x24, #16]\n" + "mov z27.d, z16.d\n" + "ld1rqw { z5.s }, p1/Z, [x27]\n" + "mov z26.d, z16.d\n" + "ld1rqw { z4.s }, p1/Z, [x27, #16]\n" + "mov z25.d, z16.d\n" + "mov z24.d, z16.d\n" + "mov z23.d, z16.d\n" + "mov z22.d, z16.d\n" + "mov z21.d, z16.d\n" + "mov z20.d, z16.d\n" + "mov z19.d, z16.d\n" + "mov z18.d, z16.d\n" + "mov z17.d, z16.d\n" + "cbz x19, 6f\n" + "ldp x24, x27, [x20], #0x10\n" + "ld1w { z16.s }, p1/Z, [%x[weights]]\n" + "subs x19, x19, #0x1\n" + "addvl %x[weights], %x[weights], #1\n" + "ld1rqw { z3.s }, p1/Z, [x24]\n" + "ld1rqw { z2.s }, p1/Z, [x24, #16]\n" + "ld1rqw { z1.s }, p1/Z, [x27]\n" + "ld1rqw { z0.s }, p1/Z, [x27, #16]\n" + "beq 4f\n" + "3:" // Output channel loop: Kernel loop + "fmla z9.s, z8.s, z7.s[0]\n" + "ldp x24, x27, [x20], #0x10\n" + "subs x19, x19, #0x1\n" + "fmla z31.s, z8.s, z7.s[1]\n" + "fmla z30.s, z8.s, z7.s[2]\n" + "fmla z29.s, z8.s, z7.s[3]\n" + "ld1rqw { z7.s }, p1/Z, [x24]\n" + "fmla z28.s, z8.s, z6.s[0]\n" + "fmla z27.s, z8.s, z6.s[1]\n" + "fmla z26.s, z8.s, z6.s[2]\n" + "fmla z25.s, z8.s, z6.s[3]\n" + "ld1rqw { z6.s }, p1/Z, [x24, #16]\n" + "fmla z24.s, z8.s, z5.s[0]\n" + "fmla z23.s, z8.s, z5.s[1]\n" + "fmla z22.s, z8.s, z5.s[2]\n" + "fmla z21.s, z8.s, z5.s[3]\n" + "ld1rqw { z5.s }, p1/Z, [x27]\n" + "fmla z20.s, z8.s, z4.s[0]\n" + "fmla z19.s, z8.s, z4.s[1]\n" + "fmla z18.s, z8.s, z4.s[2]\n" + "fmla z17.s, z8.s, z4.s[3]\n" + "ld1rqw { z4.s }, p1/Z, [x27, #16]\n" + "fmla z9.s, z16.s, z3.s[0]\n" + "ld1w { z8.s }, p1/Z, [%x[weights]]\n" + "fmla z31.s, z16.s, z3.s[1]\n" + "ldp x24, x27, [x20], #0x10\n" + "fmla z30.s, z16.s, z3.s[2]\n" + "fmla z29.s, z16.s, z3.s[3]\n" + "ld1rqw { z3.s }, p1/Z, [x24]\n" + "fmla z28.s, z16.s, z2.s[0]\n" + "fmla z27.s, z16.s, z2.s[1]\n" + "fmla z26.s, z16.s, z2.s[2]\n" + "fmla z25.s, z16.s, z2.s[3]\n" + "ld1rqw { z2.s }, p1/Z, [x24, #16]\n" + "fmla z24.s, z16.s, z1.s[0]\n" + "fmla z23.s, z16.s, z1.s[1]\n" + "fmla z22.s, z16.s, z1.s[2]\n" + "fmla z21.s, z16.s, z1.s[3]\n" + "ld1rqw { z1.s }, p1/Z, [x27]\n" + "fmla z20.s, z16.s, z0.s[0]\n" + "fmla z19.s, z16.s, z0.s[1]\n" + "fmla z18.s, z16.s, z0.s[2]\n" + "fmla z17.s, z16.s, z0.s[3]\n" + "ld1rqw { z0.s }, p1/Z, [x27, #16]\n" + "ld1w { z16.s }, p1/Z, [%x[weights], #1, MUL VL]\n" + "addvl %x[weights], %x[weights], #2\n" + "bgt 3b\n" + "4:" // Output channel loop: Kernel loop tail + "tbnz %x[kernel_points], #0, 5f\n" + "fmla z9.s, z8.s, z7.s[0]\n" + "ldr x19, [%x[outptrs], #0x0]\n" + "fmla z31.s, z8.s, z7.s[1]\n" + "ldr x20, [%x[outptrs], #0x8]\n" + "fmla z30.s, z8.s, z7.s[2]\n" + "ldr x21, [%x[outptrs], #0x10]\n" + "fmla z29.s, z8.s, z7.s[3]\n" + "ldr x22, [%x[outptrs], #0x18]\n" + "fmla z28.s, z8.s, z6.s[0]\n" + "ldr x23, [%x[outptrs], #0x20]\n" + "fmla z27.s, z8.s, z6.s[1]\n" + "ldr x24, [%x[outptrs], #0x28]\n" + "fmla z26.s, z8.s, z6.s[2]\n" + "ldr x25, [%x[outptrs], #0x30]\n" + "fmla z25.s, z8.s, z6.s[3]\n" + "ldr x26, [%x[outptrs], #0x38]\n" + "fmla z24.s, z8.s, z5.s[0]\n" + "fmla z23.s, z8.s, z5.s[1]\n" + "fmla z22.s, z8.s, z5.s[2]\n" + "fmla z21.s, z8.s, z5.s[3]\n" + "fmla z20.s, z8.s, z4.s[0]\n" + "fmla z19.s, z8.s, z4.s[1]\n" + "fmla z18.s, z8.s, z4.s[2]\n" + "fmla z17.s, z8.s, z4.s[3]\n" + "fmla z9.s, z16.s, z3.s[0]\n" + "fmla z31.s, z16.s, z3.s[1]\n" + "fmla z30.s, z16.s, z3.s[2]\n" + "fmla z29.s, z16.s, z3.s[3]\n" + "fmla z28.s, z16.s, z2.s[0]\n" + "fmla z27.s, z16.s, z2.s[1]\n" + "fmla z26.s, z16.s, z2.s[2]\n" + "fmla z25.s, z16.s, z2.s[3]\n" + "fmla z24.s, z16.s, z1.s[0]\n" + "fmla z23.s, z16.s, z1.s[1]\n" + "fmla z22.s, z16.s, z1.s[2]\n" + "fmla z21.s, z16.s, z1.s[3]\n" + "fmla z20.s, z16.s, z0.s[0]\n" + "fmla z19.s, z16.s, z0.s[1]\n" + "fmla z18.s, z16.s, z0.s[2]\n" + "fmla z17.s, z16.s, z0.s[3]\n" + "fmin z9.s, p1/M, z9.s, z10.s\n" + "fmin z31.s, p1/M, z31.s, z10.s\n" + "fmin z30.s, p1/M, z30.s, z10.s\n" + "fmin z29.s, p1/M, z29.s, z10.s\n" + "fmax z9.s, p1/M, z9.s, z11.s\n" + "st1w { z9.s }, p0, [x19, x28, LSL #2]\n" + "fmax z31.s, p1/M, z31.s, z11.s\n" + "fmax z30.s, p1/M, z30.s, z11.s\n" + "ldr x19, [%x[outptrs], #0x40]\n" + "fmax z29.s, p1/M, z29.s, z11.s\n" + "st1w { z31.s }, p0, [x20, x28, LSL #2]\n" + "fmin z28.s, p1/M, z28.s, z10.s\n" + "fmin z27.s, p1/M, z27.s, z10.s\n" + "st1w { z30.s }, p0, [x21, x28, LSL #2]\n" + "fmin z26.s, p1/M, z26.s, z10.s\n" + "st1w { z29.s }, p0, [x22, x28, LSL #2]\n" + "fmin z25.s, p1/M, z25.s, z10.s\n" + "ldr x20, [%x[outptrs], #0x48]\n" + "fmin z24.s, p1/M, z24.s, z10.s\n" + "ldr x21, [%x[outptrs], #0x50]\n" + "fmax z28.s, p1/M, z28.s, z11.s\n" + "ldr x22, [%x[outptrs], #0x58]\n" + "fmax z27.s, p1/M, z27.s, z11.s\n" + "st1w { z28.s }, p0, [x23, x28, LSL #2]\n" + "fmax z26.s, p1/M, z26.s, z11.s\n" + "fmax z25.s, p1/M, z25.s, z11.s\n" + "st1w { z27.s }, p0, [x24, x28, LSL #2]\n" + "fmax z24.s, p1/M, z24.s, z11.s\n" + "ldr x23, [%x[outptrs], #0x60]\n" + "fmin z23.s, p1/M, z23.s, z10.s\n" + "ldr x24, [%x[outptrs], #0x68]\n" + "fmin z22.s, p1/M, z22.s, z10.s\n" + "st1w { z26.s }, p0, [x25, x28, LSL #2]\n" + "fmin z21.s, p1/M, z21.s, z10.s\n" + "st1w { z25.s }, p0, [x26, x28, LSL #2]\n" + "fmin z20.s, p1/M, z20.s, z10.s\n" + "st1w { z24.s }, p0, [x19, x28, LSL #2]\n" + "fmax z23.s, p1/M, z23.s, z11.s\n" + "ldr x25, [%x[outptrs], #0x70]\n" + "fmax z22.s, p1/M, z22.s, z11.s\n" + "ldr x26, [%x[outptrs], #0x78]\n" + "fmax z21.s, p1/M, z21.s, z11.s\n" + "st1w { z23.s }, p0, [x20, x28, LSL #2]\n" + "fmax z20.s, p1/M, z20.s, z11.s\n" + "fmin z19.s, p1/M, z19.s, z10.s\n" + "st1w { z22.s }, p0, [x21, x28, LSL #2]\n" + "fmin z18.s, p1/M, z18.s, z10.s\n" + "st1w { z21.s }, p0, [x22, x28, LSL #2]\n" + "fmin z17.s, p1/M, z17.s, z10.s\n" + "st1w { z20.s }, p0, [x23, x28, LSL #2]\n" + "fmax z19.s, p1/M, z19.s, z11.s\n" + "fmax z18.s, p1/M, z18.s, z11.s\n" + "st1w { z19.s }, p0, [x24, x28, LSL #2]\n" + "fmax z17.s, p1/M, z17.s, z11.s\n" + "st1w { z18.s }, p0, [x25, x28, LSL #2]\n" + "st1w { z17.s }, p0, [x26, x28, LSL #2]\n" + "b 7f\n" + "5:" // Output channel loop: Odd tail + "fmla z9.s, z8.s, z7.s[0]\n" + "ldp x24, x27, [x20], #0x10\n" + "fmla z31.s, z8.s, z7.s[1]\n" + "ldr x19, [%x[outptrs], #0x0]\n" + "fmla z30.s, z8.s, z7.s[2]\n" + "ldr x20, [%x[outptrs], #0x8]\n" + "fmla z29.s, z8.s, z7.s[3]\n" + "ld1rqw { z7.s }, p1/Z, [x24]\n" + "fmla z28.s, z8.s, z6.s[0]\n" + "ldr x21, [%x[outptrs], #0x10]\n" + "fmla z27.s, z8.s, z6.s[1]\n" + "ldr x22, [%x[outptrs], #0x18]\n" + "fmla z26.s, z8.s, z6.s[2]\n" + "ldr x23, [%x[outptrs], #0x20]\n" + "fmla z25.s, z8.s, z6.s[3]\n" + "ld1rqw { z6.s }, p1/Z, [x24, #16]\n" + "fmla z24.s, z8.s, z5.s[0]\n" + "ldr x24, [%x[outptrs], #0x28]\n" + "fmla z23.s, z8.s, z5.s[1]\n" + "ldr x25, [%x[outptrs], #0x30]\n" + "fmla z22.s, z8.s, z5.s[2]\n" + "ldr x26, [%x[outptrs], #0x38]\n" + "fmla z21.s, z8.s, z5.s[3]\n" + "ld1rqw { z5.s }, p1/Z, [x27]\n" + "fmla z20.s, z8.s, z4.s[0]\n" + "fmla z19.s, z8.s, z4.s[1]\n" + "fmla z18.s, z8.s, z4.s[2]\n" + "fmla z17.s, z8.s, z4.s[3]\n" + "ld1rqw { z4.s }, p1/Z, [x27, #16]\n" + "fmla z9.s, z16.s, z3.s[0]\n" + "ld1w { z8.s }, p1/Z, [%x[weights]]\n" + "addvl %x[weights], %x[weights], #1\n" + "fmla z31.s, z16.s, z3.s[1]\n" + "fmla z30.s, z16.s, z3.s[2]\n" + "fmla z29.s, z16.s, z3.s[3]\n" + "fmla z28.s, z16.s, z2.s[0]\n" + "fmla z27.s, z16.s, z2.s[1]\n" + "fmla z26.s, z16.s, z2.s[2]\n" + "fmla z25.s, z16.s, z2.s[3]\n" + "fmla z24.s, z16.s, z1.s[0]\n" + "fmla z23.s, z16.s, z1.s[1]\n" + "fmla z22.s, z16.s, z1.s[2]\n" + "fmla z21.s, z16.s, z1.s[3]\n" + "fmla z20.s, z16.s, z0.s[0]\n" + "fmla z19.s, z16.s, z0.s[1]\n" + "fmla z18.s, z16.s, z0.s[2]\n" + "fmla z17.s, z16.s, z0.s[3]\n" + "fmla z9.s, z8.s, z7.s[0]\n" + "fmla z31.s, z8.s, z7.s[1]\n" + "fmla z30.s, z8.s, z7.s[2]\n" + "fmla z29.s, z8.s, z7.s[3]\n" + "fmla z28.s, z8.s, z6.s[0]\n" + "fmla z27.s, z8.s, z6.s[1]\n" + "fmla z26.s, z8.s, z6.s[2]\n" + "fmla z25.s, z8.s, z6.s[3]\n" + "fmla z24.s, z8.s, z5.s[0]\n" + "fmla z23.s, z8.s, z5.s[1]\n" + "fmla z22.s, z8.s, z5.s[2]\n" + "fmla z21.s, z8.s, z5.s[3]\n" + "fmla z20.s, z8.s, z4.s[0]\n" + "fmla z19.s, z8.s, z4.s[1]\n" + "fmla z18.s, z8.s, z4.s[2]\n" + "fmla z17.s, z8.s, z4.s[3]\n" + "fmin z9.s, p1/M, z9.s, z10.s\n" + "fmin z31.s, p1/M, z31.s, z10.s\n" + "fmin z30.s, p1/M, z30.s, z10.s\n" + "fmin z29.s, p1/M, z29.s, z10.s\n" + "fmax z9.s, p1/M, z9.s, z11.s\n" + "st1w { z9.s }, p0, [x19, x28, LSL #2]\n" + "fmax z31.s, p1/M, z31.s, z11.s\n" + "fmax z30.s, p1/M, z30.s, z11.s\n" + "ldr x19, [%x[outptrs], #0x40]\n" + "fmax z29.s, p1/M, z29.s, z11.s\n" + "st1w { z31.s }, p0, [x20, x28, LSL #2]\n" + "fmin z28.s, p1/M, z28.s, z10.s\n" + "fmin z27.s, p1/M, z27.s, z10.s\n" + "st1w { z30.s }, p0, [x21, x28, LSL #2]\n" + "fmin z26.s, p1/M, z26.s, z10.s\n" + "st1w { z29.s }, p0, [x22, x28, LSL #2]\n" + "fmin z25.s, p1/M, z25.s, z10.s\n" + "ldr x20, [%x[outptrs], #0x48]\n" + "fmin z24.s, p1/M, z24.s, z10.s\n" + "ldr x21, [%x[outptrs], #0x50]\n" + "fmax z28.s, p1/M, z28.s, z11.s\n" + "ldr x22, [%x[outptrs], #0x58]\n" + "fmax z27.s, p1/M, z27.s, z11.s\n" + "st1w { z28.s }, p0, [x23, x28, LSL #2]\n" + "fmax z26.s, p1/M, z26.s, z11.s\n" + "fmax z25.s, p1/M, z25.s, z11.s\n" + "st1w { z27.s }, p0, [x24, x28, LSL #2]\n" + "fmax z24.s, p1/M, z24.s, z11.s\n" + "ldr x23, [%x[outptrs], #0x60]\n" + "fmin z23.s, p1/M, z23.s, z10.s\n" + "ldr x24, [%x[outptrs], #0x68]\n" + "fmin z22.s, p1/M, z22.s, z10.s\n" + "st1w { z26.s }, p0, [x25, x28, LSL #2]\n" + "fmin z21.s, p1/M, z21.s, z10.s\n" + "st1w { z25.s }, p0, [x26, x28, LSL #2]\n" + "fmin z20.s, p1/M, z20.s, z10.s\n" + "st1w { z24.s }, p0, [x19, x28, LSL #2]\n" + "fmax z23.s, p1/M, z23.s, z11.s\n" + "ldr x25, [%x[outptrs], #0x70]\n" + "fmax z22.s, p1/M, z22.s, z11.s\n" + "ldr x26, [%x[outptrs], #0x78]\n" + "fmax z21.s, p1/M, z21.s, z11.s\n" + "st1w { z23.s }, p0, [x20, x28, LSL #2]\n" + "fmax z20.s, p1/M, z20.s, z11.s\n" + "fmin z19.s, p1/M, z19.s, z10.s\n" + "st1w { z22.s }, p0, [x21, x28, LSL #2]\n" + "fmin z18.s, p1/M, z18.s, z10.s\n" + "st1w { z21.s }, p0, [x22, x28, LSL #2]\n" + "fmin z17.s, p1/M, z17.s, z10.s\n" + "st1w { z20.s }, p0, [x23, x28, LSL #2]\n" + "fmax z19.s, p1/M, z19.s, z11.s\n" + "fmax z18.s, p1/M, z18.s, z11.s\n" + "st1w { z19.s }, p0, [x24, x28, LSL #2]\n" + "fmax z17.s, p1/M, z17.s, z11.s\n" + "st1w { z18.s }, p0, [x25, x28, LSL #2]\n" + "st1w { z17.s }, p0, [x26, x28, LSL #2]\n" + "b 7f\n" + "6:" // Output channel loop: Single kernel point + "fmla z9.s, z8.s, z7.s[0]\n" + "ldr x19, [%x[outptrs], #0x0]\n" + "fmla z31.s, z8.s, z7.s[1]\n" + "ldr x20, [%x[outptrs], #0x8]\n" + "fmla z30.s, z8.s, z7.s[2]\n" + "ldr x21, [%x[outptrs], #0x10]\n" + "fmla z29.s, z8.s, z7.s[3]\n" + "ldr x22, [%x[outptrs], #0x18]\n" + "fmla z28.s, z8.s, z6.s[0]\n" + "ldr x23, [%x[outptrs], #0x20]\n" + "fmla z27.s, z8.s, z6.s[1]\n" + "ldr x24, [%x[outptrs], #0x28]\n" + "fmla z26.s, z8.s, z6.s[2]\n" + "ldr x25, [%x[outptrs], #0x30]\n" + "fmla z25.s, z8.s, z6.s[3]\n" + "ldr x26, [%x[outptrs], #0x38]\n" + "fmla z24.s, z8.s, z5.s[0]\n" + "fmla z23.s, z8.s, z5.s[1]\n" + "fmla z22.s, z8.s, z5.s[2]\n" + "fmla z21.s, z8.s, z5.s[3]\n" + "fmla z20.s, z8.s, z4.s[0]\n" + "fmla z19.s, z8.s, z4.s[1]\n" + "fmla z18.s, z8.s, z4.s[2]\n" + "fmla z17.s, z8.s, z4.s[3]\n" + "fmin z9.s, p1/M, z9.s, z10.s\n" + "fmin z31.s, p1/M, z31.s, z10.s\n" + "fmin z30.s, p1/M, z30.s, z10.s\n" + "fmin z29.s, p1/M, z29.s, z10.s\n" + "fmax z9.s, p1/M, z9.s, z11.s\n" + "st1w { z9.s }, p0, [x19, x28, LSL #2]\n" + "fmax z31.s, p1/M, z31.s, z11.s\n" + "fmax z30.s, p1/M, z30.s, z11.s\n" + "ldr x19, [%x[outptrs], #0x40]\n" + "fmax z29.s, p1/M, z29.s, z11.s\n" + "st1w { z31.s }, p0, [x20, x28, LSL #2]\n" + "fmin z28.s, p1/M, z28.s, z10.s\n" + "fmin z27.s, p1/M, z27.s, z10.s\n" + "st1w { z30.s }, p0, [x21, x28, LSL #2]\n" + "fmin z26.s, p1/M, z26.s, z10.s\n" + "st1w { z29.s }, p0, [x22, x28, LSL #2]\n" + "fmin z25.s, p1/M, z25.s, z10.s\n" + "ldr x20, [%x[outptrs], #0x48]\n" + "fmin z24.s, p1/M, z24.s, z10.s\n" + "ldr x21, [%x[outptrs], #0x50]\n" + "fmax z28.s, p1/M, z28.s, z11.s\n" + "ldr x22, [%x[outptrs], #0x58]\n" + "fmax z27.s, p1/M, z27.s, z11.s\n" + "st1w { z28.s }, p0, [x23, x28, LSL #2]\n" + "fmax z26.s, p1/M, z26.s, z11.s\n" + "fmax z25.s, p1/M, z25.s, z11.s\n" + "st1w { z27.s }, p0, [x24, x28, LSL #2]\n" + "fmax z24.s, p1/M, z24.s, z11.s\n" + "ldr x23, [%x[outptrs], #0x60]\n" + "fmin z23.s, p1/M, z23.s, z10.s\n" + "ldr x24, [%x[outptrs], #0x68]\n" + "fmin z22.s, p1/M, z22.s, z10.s\n" + "st1w { z26.s }, p0, [x25, x28, LSL #2]\n" + "fmin z21.s, p1/M, z21.s, z10.s\n" + "st1w { z25.s }, p0, [x26, x28, LSL #2]\n" + "fmin z20.s, p1/M, z20.s, z10.s\n" + "st1w { z24.s }, p0, [x19, x28, LSL #2]\n" + "fmax z23.s, p1/M, z23.s, z11.s\n" + "ldr x25, [%x[outptrs], #0x70]\n" + "fmax z22.s, p1/M, z22.s, z11.s\n" + "ldr x26, [%x[outptrs], #0x78]\n" + "fmax z21.s, p1/M, z21.s, z11.s\n" + "st1w { z23.s }, p0, [x20, x28, LSL #2]\n" + "fmax z20.s, p1/M, z20.s, z11.s\n" + "fmin z19.s, p1/M, z19.s, z10.s\n" + "st1w { z22.s }, p0, [x21, x28, LSL #2]\n" + "fmin z18.s, p1/M, z18.s, z10.s\n" + "st1w { z21.s }, p0, [x22, x28, LSL #2]\n" + "fmin z17.s, p1/M, z17.s, z10.s\n" + "st1w { z20.s }, p0, [x23, x28, LSL #2]\n" + "fmax z19.s, p1/M, z19.s, z11.s\n" + "fmax z18.s, p1/M, z18.s, z11.s\n" + "st1w { z19.s }, p0, [x24, x28, LSL #2]\n" + "fmax z17.s, p1/M, z17.s, z11.s\n" + "st1w { z18.s }, p0, [x25, x28, LSL #2]\n" + "st1w { z17.s }, p0, [x26, x28, LSL #2]\n" + "7:" // Output channel loop: Done + "incw x28\n" + "whilelt p0.s, x28, %x[n_output_channels]\n" + "b.any 1b\n" + : [weights] "+&r" (weights) + : [bias] "r" (bias), [inptrs] "r" (inptrs), [kernel_points] "r" ((uint64_t) kernel_points), [minmax_vals] "r" (minmax_vals), [n_output_channels] "r" ((uint64_t) n_output_channels), [outptrs] "r" (outptrs) + : "cc", "memory", "p0", "p1", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + ); +} + +} // namespace depthwise +} // namespace arm_conv + +#endif // defined(__ARM_FEATURE_SVE) -- cgit v1.2.1