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_direct.cpp | 364 +++++++++++++++++++++ 1 file changed, 364 insertions(+) create mode 100644 src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp new file mode 100644 index 0000000000..d899255e84 --- /dev/null +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/sve_fp32_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp @@ -0,0 +1,364 @@ +/* + * 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_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( + const unsigned int n_tile_rows, + const unsigned int n_tile_cols, + const float *inptr, + int64_t ld_input_row, + int64_t ld_input_col, + float *outptr, + int64_t ld_output_row, + int64_t ld_output_col, + const void *params, + unsigned int n_channels, + const float activation_min, + const float activation_max +) +{ + struct Args + { + const uint64_t n_tile_rows, n_tile_cols; + const float *inptr; + const uint64_t ld_input_row; + const uint64_t ld_input_col; + float *outptr; + const uint64_t ld_output_row; + const uint64_t ld_output_col; + const void *params; + const float min, max; + + uint64_t tile_i = 0, tile_j = 0; + + Args( + const unsigned int n_tile_rows, + const unsigned int n_tile_cols, + const float *inptr, + int64_t ld_input_row, + int64_t ld_input_col, + float *outptr, + int64_t ld_output_row, + int64_t ld_output_col, + const void *params, + const float activation_min, + const float activation_max + ) : n_tile_rows(n_tile_rows), n_tile_cols(n_tile_cols), inptr(inptr), + ld_input_row(ld_input_row), ld_input_col(ld_input_col), outptr(outptr), + ld_output_row(ld_output_row), ld_output_col(ld_output_col), + params(params), min(activation_min), max(activation_max) + { + } + }; + + Args params_struct( + n_tile_rows, n_tile_cols, + inptr, ld_input_row, ld_input_col, + outptr, ld_output_row, ld_output_col, + params, activation_min, activation_max + ); + + __asm__ __volatile__( + "ptrue p3.b\n" + "mov x8, #0x0\n" + "mov x17, #0x0\n" + "1:" // Tile loop + "str x8, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "mov x21, #0x2\n" + "str x17, [%x[params_struct], %[offsetof_args_tile_j]]\n" + "cntb x16\n" + "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n" + "add x16, x16, XZR, LSL #4\n" + "ldr x20, [%x[params_struct], %[offsetof_args_ld_input_row]]\n" + "cntb x14\n" + "ldr x13, [%x[params_struct], %[offsetof_args_ld_input_col]]\n" + "cntb x12\n" + "ldr x11, [%x[params_struct], %[offsetof_args_inptr]]\n" + "mul x19, x8, x20\n" // offset = tile_i * ld_input_row + "ldr x22, [%x[params_struct], %[offsetof_args_ld_output_row]]\n" + "madd x19, x17, x13, x19\n" // offset += tile_j * ld_input_col + "ldr x10, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" + "mul x19, x19, x21\n" // offset *= kernel_stride * output_size + "ldr x9, [%x[params_struct], %[offsetof_args_outptr]]\n" + "add x11, x11, x19, LSL #2\n" // inptr[0] += offset * sizeof(float) + "ld1rw { z18.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n" + "add x28, x11, x20, LSL #2\n" + "ld1rw { z17.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n" + "add x27, x28, x20, LSL #2\n" + "ld1w { z16.s }, p3/Z, [x15]\n" + "mov z31.d, z16.d\n" + "ld1w { z0.s }, p3/Z, [x15, #1, MUL VL]\n" + "add x26, x27, x20, LSL #2\n" + "mov z30.d, z16.d\n" + "ld1w { z1.s }, p3/Z, [x15, #2, MUL VL]\n" + "add x25, x13, x13\n" + "mov z29.d, z16.d\n" + "ld1w { z2.s }, p3/Z, [x15, #3, MUL VL]\n" + "add x24, x25, x13\n" + "mov z28.d, z16.d\n" + "ld1w { z3.s }, p3/Z, [x15, #4, MUL VL]\n" + "add x14, x14, x13, LSL #4\n" + "ld1w { z4.s }, p3/Z, [x15, #5, MUL VL]\n" + "add x12, x12, x25, LSL #4\n" + "ld1w { z5.s }, p3/Z, [x15, #6, MUL VL]\n" + "cntb x23\n" + "ld1w { z6.s }, p3/Z, [x15, #7, MUL VL]\n" + "add x23, x23, x24, LSL #4\n" + "prfm pldl1keep, [x28, x14]\n" + "mov x20, #0x2\n" + "prfm pldl1keep, [x11, x16]\n" + "mul x19, x8, x22\n" // offset = tile_i * ld_output_row + "prfm pldl1keep, [x11, x23]\n" + "madd x19, x17, x10, x19\n" // offset += tile_j * ld_output_col + "prfm pldl1keep, [x28, x12]\n" + "mul x19, x19, x20\n" // offset *= output_tile_size + "prfm pldl1keep, [x27, x14]\n" + "add x9, x9, x19, LSL #2\n" // outptrs[0] += offset * sizeof(float) + "mov x21, #0x0\n" + "add x22, x9, x22, LSL #2\n" + "cntw x20\n" + "sub x19, XZR, x20\n" + "whilelt p2.s, XZR, %x[n_channels]\n" + "ld1w { z9.s }, p2/Z, [x28, x13, LSL #2]\n" + "ld1w { z10.s }, p2/Z, [x11]\n" + "addvl x15, x15, #16\n" + "ld1w { z11.s }, p2/Z, [x11, x24, LSL #2]\n" + "cmp x20, %x[n_channels]\n" + "ld1w { z7.s }, p3/Z, [x15, #-8, MUL VL]\n" + "ld1w { z8.s }, p3/Z, [x15, #-7, MUL VL]\n" + "addvl x15, x15, #-6\n" + "ld1w { z12.s }, p2/Z, [x28, x25, LSL #2]\n" + "ld1w { z13.s }, p2/Z, [x27, x13, LSL #2]\n" + "bge 3f\n" + "2:" // Tile loop: Channel loop + "fmla z31.s, p3/M, z4.s, z9.s\n" + "prfm pldl1keep, [x26, x16]\n" + "whilelt p1.s, x20, %x[n_channels]\n" + "fmla z30.s, p3/M, z3.s, z9.s\n" + "prfm pldl1keep, [x26, x23]\n" + "incw x19\n" + "fmla z29.s, p3/M, z1.s, z9.s\n" + "prfm pldl1keep, [x11, x14]\n" + "mov p0.b, p2.b\n" + "fmla z28.s, p3/M, z0.s, z9.s\n" + "ld1w { z9.s }, p2/Z, [x26]\n" + "incw x21\n" + "fmla z31.s, p3/M, z0.s, z10.s\n" + "prfm pldl1keep, [x11, x12]\n" + "incw x20\n" + "fmla z30.s, p3/M, z2.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x26, x24, LSL #2]\n" + "ld1w { z10.s }, p2/Z, [x27, x25, LSL #2]\n" + "fmla z29.s, p3/M, z2.s, z12.s\n" + "prfm pldl1keep, [x27, x12]\n" + "fmla z28.s, p3/M, z1.s, z12.s\n" + "prfm pldl1keep, [x28, x16]\n" + "fmla z31.s, p3/M, z5.s, z12.s\n" + "prfm pldl1keep, [x28, x23]\n" + "fmla z30.s, p3/M, z4.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x11, x13, LSL #2]\n" + "fmla z29.s, p3/M, z6.s, z9.s\n" + "ld1w { z9.s }, p2/Z, [x11, x25, LSL #2]\n" + "fmla z28.s, p3/M, z3.s, z13.s\n" + "addvl x11, x11, #1\n" + "fmla z31.s, p3/M, z7.s, z13.s\n" + "prfm pldl1keep, [x27, x16]\n" + "prfm pldl1keep, [x27, x23]\n" + "fmla z30.s, p3/M, z6.s, z13.s\n" + "prfm pldl1keep, [x26, x14]\n" + "fmla z29.s, p3/M, z4.s, z13.s\n" + "prfm pldl1keep, [x26, x12]\n" + "fmla z28.s, p3/M, z8.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x28]\n" + "fmla z31.s, p3/M, z1.s, z12.s\n" + "ld1w { z16.s }, p3/Z, [x15]\n" + "fmla z30.s, p3/M, z0.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x28, x24, LSL #2]\n" + "fmla z29.s, p3/M, z5.s, z10.s\n" + "addvl x28, x28, #1\n" + "fmla z28.s, p3/M, z4.s, z10.s\n" + "prfm pldl1keep, [x28, x14]\n" + "prfm pldl1keep, [x11, x16]\n" + "fmla z31.s, p3/M, z2.s, z9.s\n" + "prfm pldl1keep, [x11, x23]\n" + "fmla z30.s, p3/M, z1.s, z9.s\n" + "ld1w { z9.s }, p2/Z, [x27]\n" + "fmla z29.s, p3/M, z0.s, z11.s\n" + "prfm pldl1keep, [x28, x12]\n" + "fmla z28.s, p3/M, z2.s, z12.s\n" + "ld1w { z0.s }, p3/Z, [x15, #1, MUL VL]\n" + "fmla z31.s, p3/M, z8.s, z10.s\n" + "ld1w { z1.s }, p3/Z, [x15, #2, MUL VL]\n" + "fmla z30.s, p3/M, z7.s, z10.s\n" + "ld1w { z10.s }, p2/Z, [x27, x24, LSL #2]\n" + "fmla z29.s, p3/M, z3.s, z9.s\n" + "addvl x27, x27, #1\n" + "fmla z31.s, p3/M, z3.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x26, x13, LSL #2]\n" + "fmla z28.s, p3/M, z5.s, z10.s\n" + "ld1w { z13.s }, p1/Z, [x27, x13, LSL #2]\n" + "fmla z30.s, p3/M, z5.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x26, x25, LSL #2]\n" + "whilelt p2.s, x21, %x[n_channels]\n" + "fmla z29.s, p3/M, z7.s, z11.s\n" + "prfm pldl1keep, [x27, x14]\n" + "addvl x26, x26, #1\n" + "fmla z31.s, p3/M, z6.s, z9.s\n" + "ld1w { z9.s }, p1/Z, [x28, x13, LSL #2]\n" + "cmp x20, %x[n_channels]\n" + "fmla z30.s, p3/M, z8.s, z10.s\n" + "ld1w { z10.s }, p1/Z, [x11]\n" + "fmla z28.s, p3/M, z6.s, z11.s\n" + "ld1w { z11.s }, p1/Z, [x11, x24, LSL #2]\n" + "ld1w { z2.s }, p3/Z, [x15, #3, MUL VL]\n" + "fmla z29.s, p3/M, z8.s, z12.s\n" + "ld1w { z3.s }, p3/Z, [x15, #4, MUL VL]\n" + "fmax z31.s, p3/M, z31.s, z18.s\n" + "ld1w { z4.s }, p3/Z, [x15, #5, MUL VL]\n" + "fmla z28.s, p3/M, z7.s, z12.s\n" + "fmax z30.s, p3/M, z30.s, z18.s\n" + "ld1w { z12.s }, p1/Z, [x28, x25, LSL #2]\n" + "ld1w { z5.s }, p3/Z, [x15, #6, MUL VL]\n" + "fmax z29.s, p3/M, z29.s, z18.s\n" + "ld1w { z6.s }, p3/Z, [x15, #7, MUL VL]\n" + "fmin z31.s, p3/M, z31.s, z17.s\n" + "addvl x15, x15, #16\n" + "fmin z30.s, p3/M, z30.s, z17.s\n" + "ld1w { z7.s }, p3/Z, [x15, #-8, MUL VL]\n" + "fmax z28.s, p3/M, z28.s, z18.s\n" + "ld1w { z8.s }, p3/Z, [x15, #-7, MUL VL]\n" + "addvl x15, x15, #-6\n" + "fmin z29.s, p3/M, z29.s, z17.s\n" + "st1w { z31.s }, p0, [x9]\n" + "mov z31.d, z16.d\n" + "fmin z28.s, p3/M, z28.s, z17.s\n" + "st1w { z30.s }, p0, [x9, x10, LSL #2]\n" + "addvl x9, x9, #1\n" + "mov z30.d, z16.d\n" + "st1w { z29.s }, p0, [x22]\n" + "mov z29.d, z16.d\n" + "st1w { z28.s }, p0, [x22, x10, LSL #2]\n" + "mov z28.d, z16.d\n" + "addvl x22, x22, #1\n" + "blt 2b\n" + "3:" // Tile loop: Channel tail + "fmla z31.s, p3/M, z4.s, z9.s\n" + "prfm pldl1keep, [x26, x16]\n" + "mov p0.b, p2.b\n" + "fmla z30.s, p3/M, z3.s, z9.s\n" + "prfm pldl1keep, [x26, x23]\n" + "fmla z29.s, p3/M, z1.s, z9.s\n" + "prfm pldl1keep, [x11, x14]\n" + "fmla z28.s, p3/M, z0.s, z9.s\n" + "ld1w { z9.s }, p2/Z, [x26]\n" + "prfm pldl1keep, [x11, x12]\n" + "fmla z31.s, p3/M, z0.s, z10.s\n" + "ld1w { z10.s }, p2/Z, [x27, x25, LSL #2]\n" + "fmla z30.s, p3/M, z2.s, z11.s\n" + "fmla z29.s, p3/M, z2.s, z12.s\n" + "ld1w { z11.s }, p2/Z, [x26, x24, LSL #2]\n" + "fmla z28.s, p3/M, z1.s, z12.s\n" + "prfm pldl1keep, [x27, x12]\n" + "prfm pldl1keep, [x28, x16]\n" + "fmla z31.s, p3/M, z5.s, z12.s\n" + "prfm pldl1keep, [x28, x23]\n" + "fmla z30.s, p3/M, z4.s, z12.s\n" + "fmla z29.s, p3/M, z6.s, z9.s\n" + "ld1w { z12.s }, p2/Z, [x11, x13, LSL #2]\n" + "fmla z28.s, p3/M, z3.s, z13.s\n" + "ld1w { z9.s }, p2/Z, [x11, x25, LSL #2]\n" + "prfm pldl1keep, [x27, x16]\n" + "fmla z31.s, p3/M, z7.s, z13.s\n" + "prfm pldl1keep, [x27, x23]\n" + "fmla z30.s, p3/M, z6.s, z13.s\n" + "fmla z29.s, p3/M, z4.s, z13.s\n" + "prfm pldl1keep, [x26, x14]\n" + "fmla z28.s, p3/M, z8.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x28]\n" + "prfm pldl1keep, [x26, x12]\n" + "fmla z31.s, p3/M, z1.s, z12.s\n" + "ldr x8, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "fmla z30.s, p3/M, z0.s, z12.s\n" + "add x21, x8, #0x1\n" + "fmla z29.s, p3/M, z5.s, z10.s\n" + "ld1w { z12.s }, p2/Z, [x28, x24, LSL #2]\n" + "fmla z31.s, p3/M, z2.s, z9.s\n" + "ldr x17, [%x[params_struct], %[offsetof_args_tile_j]]\n" + "add x17, x17, #0x1\n" + "fmla z30.s, p3/M, z1.s, z9.s\n" + "ld1w { z9.s }, p2/Z, [x27]\n" + "fmla z31.s, p3/M, z8.s, z10.s\n" + "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" + "fmla z28.s, p3/M, z4.s, z10.s\n" + "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" + "cmp x17, x19\n" + "fmla z30.s, p3/M, z7.s, z10.s\n" + "ld1w { z10.s }, p2/Z, [x27, x24, LSL #2]\n" + "fmla z31.s, p3/M, z3.s, z11.s\n" + "csel x17, x17, XZR, LT\n" + "fmla z29.s, p3/M, z0.s, z11.s\n" + "ld1w { z11.s }, p2/Z, [x26, x13, LSL #2]\n" + "csel x8, x8, x21, LT\n" + "fmla z28.s, p3/M, z2.s, z12.s\n" + "cmp x8, x20\n" + "fmla z30.s, p3/M, z5.s, z12.s\n" + "ld1w { z12.s }, p2/Z, [x26, x25, LSL #2]\n" + "fmla z31.s, p3/M, z6.s, z9.s\n" + "fmla z29.s, p3/M, z3.s, z9.s\n" + "fmla z28.s, p3/M, z5.s, z10.s\n" + "fmla z30.s, p3/M, z8.s, z10.s\n" + "fmla z29.s, p3/M, z7.s, z11.s\n" + "fmax z31.s, p3/M, z31.s, z18.s\n" + "fmla z28.s, p3/M, z6.s, z11.s\n" + "fmax z30.s, p3/M, z30.s, z18.s\n" + "fmla z29.s, p3/M, z8.s, z12.s\n" + "fmin z31.s, p3/M, z31.s, z17.s\n" + "st1w { z31.s }, p0, [x9]\n" + "fmla z28.s, p3/M, z7.s, z12.s\n" + "fmin z30.s, p3/M, z30.s, z17.s\n" + "st1w { z30.s }, p0, [x9, x10, LSL #2]\n" + "fmax z29.s, p3/M, z29.s, z18.s\n" + "fmax z28.s, p3/M, z28.s, z18.s\n" + "fmin z29.s, p3/M, z29.s, z17.s\n" + "st1w { z29.s }, p0, [x22]\n" + "fmin z28.s, p3/M, z28.s, z17.s\n" + "st1w { z28.s }, p0, [x22, x10, LSL #2]\n" + "blt 1b\n" + : + : [n_channels] "r" ((unsigned long) n_channels), [offsetof_args_inptr] "I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col] "I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row] "I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col] "I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row] "I" (offsetof(Args, ld_output_row)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_n_tile_cols] "I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows] "I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr] "I" (offsetof(Args, outptr)), [offsetof_args_params] "I" (offsetof(Args, params)), [offsetof_args_tile_i] "I" (offsetof(Args, tile_i)), [offsetof_args_tile_j] "I" (offsetof(Args, tile_j)), [params_struct] "r" (¶ms_struct) + : "cc", "memory", "p0", "p1", "p2", "p3", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z16", "z17", "z18", "z28", "z29", "z30", "z31" + ); +} + +} // namespace depthwise +} // namespace arm_conv + +#endif // defined(__ARM_FEATURE_SVE) -- cgit v1.2.1