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 | 353 +++++++++++++++++++++ 1 file changed, 353 insertions(+) create mode 100644 src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp new file mode 100644 index 0000000000..bb9931c20f --- /dev/null +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp @@ -0,0 +1,353 @@ +/* + * 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 "arm_gemm.hpp" +#include +#include + +#if defined(__ARM_FEATURE_SVE) + +namespace arm_conv { +namespace depthwise { + +void sve_u8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst_impl( + const uint8_t *const *const inptrs, + uint8_t *const *const outptrs, + const void *params, + unsigned int n_output_channels, + const arm_gemm::Requantize32& qp +) +{ + __asm__ __volatile__( + "mov z31.s, #0x0\n" + "ldr x24, [%x[inptrs], #0x0]\n" + "ptrue p2.b\n" + "mov z18.s, #0x0\n" + "ldr x23, [%x[inptrs], #0x8]\n" + "lsl x9, %x[n_channels], #0x2\n" + "mov z29.s, #0x0\n" + "ldr x22, [%x[inptrs], #0x10]\n" + "addvl SP, SP, #-8\n" + "mov z28.s, #0x0\n" + "ldr x21, [%x[inptrs], #0x18]\n" + "mov x19, #0x9\n" + "mov z13.s, #0x0\n" + "ldr x20, [%x[inptrs], #0x20]\n" + "whilelt p1.b, XZR, x19\n" + "mov z14.s, #0x0\n" + "ld1b { z7.b }, p1/Z, [x24]\n" + "mov x19, #0x3\n" + "mov z15.s, #0x0\n" + "ld1b { z3.b }, p1/Z, [x23]\n" + "whilelt p0.b, XZR, x19\n" + "mov z11.b, p0/z, #0x1\n" + "ld1b { z4.b }, p1/Z, [x22]\n" + "mov x28, #0x0\n" + "mov z10.d, z7.d\n" + "ld1b { z6.b }, p1/Z, [x21]\n" + "mov x27, #0x0\n" + "ext z10.b, z10.b, z10.b, #0x2\n" + "ld1b { z5.b }, p1/Z, [x20]\n" + "whilelt p1.b, x28, x9\n" + "mov z17.d, z7.d\n" + "ld1rw { z30.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_b_offset]]\n" + "mov z26.d, z7.d\n" + "ldp x26, x25, [%x[outptrs], #0x0]\n" + "ext z17.b, z17.b, z17.b, #0x4\n" + "ldp x24, x23, [%x[outptrs], #0x10]\n" + "ext z26.b, z26.b, z26.b, #0x6\n" + "ldp x22, x21, [%x[outptrs], #0x20]\n" + "mov z19.d, z3.d\n" + "ldp x20, x19, [%x[outptrs], #0x30]\n" + "ext z19.b, z19.b, z19.b, #0x2\n" + "ld1rw { z12.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_c_offset]]\n" + "zip1 z7.s, z7.s, z17.s\n" + "ld1rw { z16.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_minval]]\n" + "zip1 z10.s, z10.s, z26.s\n" + "ld1rw { z0.s }, p2/Z, [%x[qp], %[offsetof_Requantize32_maxval]]\n" + "zip1 z7.s, z7.s, z10.s\n" + "ld1w { z1.s }, p1/Z, [%x[params]]\n" + "mov z7.q, z7.q[0]\n" + "ld1b { z8.b }, p1/Z, [%x[params], #1, MUL VL]\n" + "mov z17.d, z3.d\n" + "ld1b { z9.b }, p1/Z, [%x[params], #2, MUL VL]\n" + "ext z17.b, z17.b, z17.b, #0x4\n" + "ld1b { z10.b }, p1/Z, [%x[params], #3, MUL VL]\n" + "addvl %x[params], %x[params], #4\n" + "mov z2.d, z3.d\n" + "mov z20.d, z4.d\n" + "ext z2.b, z2.b, z2.b, #0x6\n" + "zip1 z3.s, z3.s, z17.s\n" + "ext z20.b, z20.b, z20.b, #0x2\n" + "mov z17.d, z4.d\n" + "zip1 z19.s, z19.s, z2.s\n" + "zip1 z3.s, z3.s, z19.s\n" + "mov z3.q, z3.q[0]\n" + "ext z17.b, z17.b, z17.b, #0x4\n" + "mov z26.d, z4.d\n" + "ext z26.b, z26.b, z26.b, #0x6\n" + "mov z21.d, z6.d\n" + "zip1 z4.s, z4.s, z17.s\n" + "ext z21.b, z21.b, z21.b, #0x2\n" + "zip1 z20.s, z20.s, z26.s\n" + "zip1 z4.s, z4.s, z20.s\n" + "mov z4.q, z4.q[0]\n" + "mov z17.d, z6.d\n" + "ext z17.b, z17.b, z17.b, #0x4\n" + "mov z20.d, z6.d\n" + "ext z20.b, z20.b, z20.b, #0x6\n" + "mov z19.d, z5.d\n" + "zip1 z6.s, z6.s, z17.s\n" + "ext z19.b, z19.b, z19.b, #0x2\n" + "zip1 z21.s, z21.s, z20.s\n" + "zip1 z6.s, z6.s, z21.s\n" + "mov z6.q, z6.q[0]\n" + "mov z17.d, z5.d\n" + "ext z17.b, z17.b, z17.b, #0x4\n" + "mov z20.d, z5.d\n" + "ext z20.b, z20.b, z20.b, #0x6\n" + "mov z11.s, z11.s[0]\n" + "zip1 z5.s, z5.s, z17.s\n" + "mov z25.s, #0x0\n" + "zip1 z19.s, z19.s, z20.s\n" + "zip1 z5.s, z5.s, z19.s\n" + "mov z5.q, z5.q[0]\n" + "mov z26.s, #0x0\n" + "mov z27.s, #0x0\n" + "mov z24.s, #0x0\n" + "mov z23.s, #0x0\n" + "mov z22.s, #0x0\n" + "mov z21.s, #0x0\n" + "mov z17.s, #0x0\n" + "mov z20.s, #0x0\n" + "mov z2.s, #0x0\n" + "mov z19.s, #0x0\n" + "udot z31.s, z11.b, z7.b[0]\n" + "udot z18.s, z11.b, z7.b[1]\n" + "udot z29.s, z11.b, z7.b[2]\n" + "udot z28.s, z11.b, z7.b[3]\n" + "udot z13.s, z11.b, z3.b[0]\n" + "udot z14.s, z11.b, z3.b[1]\n" + "udot z15.s, z11.b, z3.b[2]\n" + "udot z25.s, z11.b, z3.b[3]\n" + "udot z26.s, z11.b, z4.b[0]\n" + "udot z27.s, z11.b, z4.b[1]\n" + "udot z24.s, z11.b, z4.b[2]\n" + "udot z23.s, z11.b, z4.b[3]\n" + "udot z22.s, z11.b, z6.b[0]\n" + "udot z21.s, z11.b, z6.b[1]\n" + "udot z17.s, z11.b, z6.b[2]\n" + "udot z20.s, z11.b, z6.b[3]\n" + "udot z2.s, z11.b, z5.b[0]\n" + "udot z19.s, z11.b, z5.b[1]\n" + "mov z31.d, z31.d\n" + "mov z18.d, z18.d\n" + "mov z29.d, z29.d\n" + "mov z28.d, z28.d\n" + "add z31.s, z31.s, z13.s\n" + "mov z13.s, #0x0\n" + "udot z13.s, z11.b, z5.b[2]\n" + "add z18.s, z18.s, z14.s\n" + "mov z14.s, #0x0\n" + "udot z14.s, z11.b, z5.b[3]\n" + "add z29.s, z29.s, z15.s\n" + "add z28.s, z28.s, z25.s\n" + "add z31.s, z31.s, z26.s\n" + "add z18.s, z18.s, z27.s\n" + "add z29.s, z29.s, z24.s\n" + "add z28.s, z28.s, z23.s\n" + "mov z26.d, z26.d\n" + "mov z25.d, z27.d\n" + "mov z24.d, z24.d\n" + "mov z23.d, z23.d\n" + "add z26.s, z26.s, z22.s\n" + "add z25.s, z25.s, z21.s\n" + "add z24.s, z24.s, z17.s\n" + "add z23.s, z23.s, z20.s\n" + "add z26.s, z26.s, z2.s\n" + "add z25.s, z25.s, z19.s\n" + "add z24.s, z24.s, z13.s\n" + "add z23.s, z23.s, z14.s\n" + "neg z30.s, p2/M, z30.s\n" + "mul z31.s, p2/M, z31.s, z30.s\n" + "st1w { z31.s }, p2, [SP]\n" + "add z31.s, z31.s, z1.s\n" + "mul z18.s, p2/M, z18.s, z30.s\n" + "st1w { z18.s }, p2, [SP, #1, MUL VL]\n" + "add z18.s, z18.s, z1.s\n" + "mul z29.s, p2/M, z29.s, z30.s\n" + "st1w { z29.s }, p2, [SP, #2, MUL VL]\n" + "add z29.s, z29.s, z1.s\n" + "mul z28.s, p2/M, z28.s, z30.s\n" + "st1w { z28.s }, p2, [SP, #3, MUL VL]\n" + "add z28.s, z28.s, z1.s\n" + "mul z26.s, p2/M, z26.s, z30.s\n" + "st1w { z26.s }, p2, [SP, #4, MUL VL]\n" + "add z26.s, z26.s, z1.s\n" + "mul z25.s, p2/M, z25.s, z30.s\n" + "st1w { z25.s }, p2, [SP, #5, MUL VL]\n" + "add z25.s, z25.s, z1.s\n" + "mul z24.s, p2/M, z24.s, z30.s\n" + "st1w { z24.s }, p2, [SP, #6, MUL VL]\n" + "add z24.s, z24.s, z1.s\n" + "mul z23.s, p2/M, z23.s, z30.s\n" + "st1w { z23.s }, p2, [SP, #7, MUL VL]\n" + "add z23.s, z23.s, z1.s\n" + "1:" // Loop + "udot z31.s, z8.b, z7.b[0]\n" + "ld1w { z22.s }, p2/Z, [%x[params]]\n" + "incb x28\n" + "udot z18.s, z8.b, z7.b[1]\n" + "ld1w { z21.s }, p2/Z, [%x[params], #1, MUL VL]\n" + "whilelt p0.s, x27, %x[n_channels]\n" + "udot z29.s, z8.b, z7.b[2]\n" + "whilelt p1.b, x28, x9\n" + "ld1w { z1.s }, p1/Z, [%x[params], #2, MUL VL]\n" + "udot z28.s, z8.b, z7.b[3]\n" + "udot z26.s, z8.b, z4.b[0]\n" + "udot z25.s, z8.b, z4.b[1]\n" + "udot z24.s, z8.b, z4.b[2]\n" + "udot z23.s, z8.b, z4.b[3]\n" + "ld1b { z8.b }, p1/Z, [%x[params], #3, MUL VL]\n" + "udot z31.s, z9.b, z3.b[0]\n" + "udot z18.s, z9.b, z3.b[1]\n" + "udot z29.s, z9.b, z3.b[2]\n" + "udot z28.s, z9.b, z3.b[3]\n" + "udot z26.s, z9.b, z6.b[0]\n" + "udot z25.s, z9.b, z6.b[1]\n" + "udot z24.s, z9.b, z6.b[2]\n" + "udot z23.s, z9.b, z6.b[3]\n" + "ld1b { z9.b }, p1/Z, [%x[params], #4, MUL VL]\n" + "udot z31.s, z10.b, z4.b[0]\n" + "udot z18.s, z10.b, z4.b[1]\n" + "udot z29.s, z10.b, z4.b[2]\n" + "udot z28.s, z10.b, z4.b[3]\n" + "udot z26.s, z10.b, z5.b[0]\n" + "udot z25.s, z10.b, z5.b[1]\n" + "udot z24.s, z10.b, z5.b[2]\n" + "udot z23.s, z10.b, z5.b[3]\n" + "ld1b { z10.b }, p1/Z, [%x[params], #5, MUL VL]\n" + "addvl %x[params], %x[params], #6\n" + ".inst 0x04b677ff // sqrdmulh z31.s, z31.s, z22.s\n" + ".inst 0x04b67652 // sqrdmulh z18.s, z18.s, z22.s\n" + ".inst 0x04b677bd // sqrdmulh z29.s, z29.s, z22.s\n" + ".inst 0x04b6779c // sqrdmulh z28.s, z28.s, z22.s\n" + ".inst 0x04b6775a // sqrdmulh z26.s, z26.s, z22.s\n" + "and z20.d, z31.d, z21.d\n" + "asr z20.s, z20.s, #0x1f\n" + "and z19.d, z18.d, z21.d\n" + "and z14.d, z29.d, z21.d\n" + "asr z19.s, z19.s, #0x1f\n" + "and z17.d, z28.d, z21.d\n" + "and z2.d, z26.d, z21.d\n" + "asr z14.s, z14.s, #0x1f\n" + ".inst 0x04b67739 // sqrdmulh z25.s, z25.s, z22.s\n" + "asr z17.s, z17.s, #0x1f\n" + "sqadd z31.s, z31.s, z20.s\n" + ".inst 0x04b67718 // sqrdmulh z24.s, z24.s, z22.s\n" + "asr z2.s, z2.s, #0x1f\n" + ".inst 0x04b676f7 // sqrdmulh z23.s, z23.s, z22.s\n" + "sqadd z18.s, z18.s, z19.s\n" + "sqadd z29.s, z29.s, z14.s\n" + "and z27.d, z25.d, z21.d\n" + "asr z27.s, z27.s, #0x1f\n" + "sqadd z28.s, z28.s, z17.s\n" + "sqadd z26.s, z26.s, z2.s\n" + "and z17.d, z24.d, z21.d\n" + "asr z17.s, z17.s, #0x1f\n" + "and z15.d, z23.d, z21.d\n" + ".inst 0x44828abf // srshl z31.s, p2/M, z31.s, z21.s\n" + "asr z15.s, z15.s, #0x1f\n" + "sqadd z25.s, z25.s, z27.s\n" + ".inst 0x44828ab2 // srshl z18.s, p2/M, z18.s, z21.s\n" + "add z31.s, z31.s, z12.s\n" + "sqadd z24.s, z24.s, z17.s\n" + ".inst 0x44828abd // srshl z29.s, p2/M, z29.s, z21.s\n" + "add z18.s, z18.s, z12.s\n" + "sqadd z23.s, z23.s, z15.s\n" + "smin z31.s, p2/M, z31.s, z0.s\n" + "add z29.s, z29.s, z12.s\n" + "smin z18.s, p2/M, z18.s, z0.s\n" + ".inst 0x44828abc // srshl z28.s, p2/M, z28.s, z21.s\n" + "smax z31.s, p2/M, z31.s, z16.s\n" + "st1b { z31.s }, p0, [x26, x27]\n" + "add z28.s, z28.s, z12.s\n" + "smax z18.s, p2/M, z18.s, z16.s\n" + "ld1w { z31.s }, p2/Z, [SP]\n" + "smin z29.s, p2/M, z29.s, z0.s\n" + "st1b { z18.s }, p0, [x25, x27]\n" + "add z31.s, z31.s, z1.s\n" + "smin z28.s, p2/M, z28.s, z0.s\n" + "ld1w { z18.s }, p2/Z, [SP, #1, MUL VL]\n" + "smax z29.s, p2/M, z29.s, z16.s\n" + "st1b { z29.s }, p0, [x24, x27]\n" + "add z18.s, z18.s, z1.s\n" + "smax z28.s, p2/M, z28.s, z16.s\n" + "ld1w { z29.s }, p2/Z, [SP, #2, MUL VL]\n" + ".inst 0x44828aba // srshl z26.s, p2/M, z26.s, z21.s\n" + "st1b { z28.s }, p0, [x23, x27]\n" + "add z29.s, z29.s, z1.s\n" + ".inst 0x44828ab9 // srshl z25.s, p2/M, z25.s, z21.s\n" + "ld1w { z28.s }, p2/Z, [SP, #3, MUL VL]\n" + "add z26.s, z26.s, z12.s\n" + ".inst 0x44828ab8 // srshl z24.s, p2/M, z24.s, z21.s\n" + ".inst 0x44828ab7 // srshl z23.s, p2/M, z23.s, z21.s\n" + "add z25.s, z25.s, z12.s\n" + "add z28.s, z28.s, z1.s\n" + "add z24.s, z24.s, z12.s\n" + "add z23.s, z23.s, z12.s\n" + "smin z26.s, p2/M, z26.s, z0.s\n" + "smin z25.s, p2/M, z25.s, z0.s\n" + "smin z24.s, p2/M, z24.s, z0.s\n" + "smin z23.s, p2/M, z23.s, z0.s\n" + "smax z26.s, p2/M, z26.s, z16.s\n" + "st1b { z26.s }, p0, [x22, x27]\n" + "smax z25.s, p2/M, z25.s, z16.s\n" + "smax z24.s, p2/M, z24.s, z16.s\n" + "ld1w { z26.s }, p2/Z, [SP, #4, MUL VL]\n" + "smax z23.s, p2/M, z23.s, z16.s\n" + "st1b { z25.s }, p0, [x21, x27]\n" + "add z26.s, z26.s, z1.s\n" + "st1b { z24.s }, p0, [x20, x27]\n" + "st1b { z23.s }, p0, [x19, x27]\n" + "incw x27\n" + "ld1w { z25.s }, p2/Z, [SP, #5, MUL VL]\n" + "add z25.s, z25.s, z1.s\n" + "ld1w { z24.s }, p2/Z, [SP, #6, MUL VL]\n" + "ld1w { z23.s }, p2/Z, [SP, #7, MUL VL]\n" + "add z24.s, z24.s, z1.s\n" + "add z23.s, z23.s, z1.s\n" + "b.any 1b\n" + "addvl SP, SP, #8\n" + : [params] "+&r" (params) + : [inptrs] "r" (inptrs), [n_channels] "r" (n_output_channels), [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)), [outptrs] "r" (outptrs), [qp] "r" (&qp) + : "cc", "memory", "p0", "p1", "p2", "x9", "x19", "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 depthwise +} // namespace arm_conv + +#endif // defined(__ARM_FEATURE_SVE) -- cgit v1.2.1