From d621bca4e963555a99be4328c8d49d1813789649 Mon Sep 17 00:00:00 2001 From: Chunosov Date: Fri, 3 Nov 2017 17:33:15 +0700 Subject: COMPMID-661: directconv-uint8 (#20) Change-Id: I84f7a1ce3658be0d3c91e65096467258af48f0b6 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/94341 Tested-by: Kaizen Reviewed-by: Anthony Barbier --- tests/validation/CPP/ConvolutionLayer.cpp | 118 ++++++++++++++++++++++++----- tests/validation/CPP/UtilsQuantizedAsymm.h | 57 ++++++++++++++ 2 files changed, 157 insertions(+), 18 deletions(-) create mode 100644 tests/validation/CPP/UtilsQuantizedAsymm.h (limited to 'tests/validation/CPP') diff --git a/tests/validation/CPP/ConvolutionLayer.cpp b/tests/validation/CPP/ConvolutionLayer.cpp index ab3690a493..aa73869a0e 100644 --- a/tests/validation/CPP/ConvolutionLayer.cpp +++ b/tests/validation/CPP/ConvolutionLayer.cpp @@ -23,11 +23,15 @@ */ #include "ConvolutionLayer.h" +#include "tests/validation/CPP/Utils.h" +#include "tests/validation/CPP/UtilsQuantizedAsymm.h" #include "tests/validation/FixedPoint.h" #include "tests/validation/Helpers.h" #include "tests/framework/Asserts.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" + namespace arm_compute { namespace test @@ -45,9 +49,14 @@ inline bool is_valid_pixel(int i, int min, int max) // 3D convolution for floating point type template ::value, int>::type = 0> -void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights, int fixed_point_position) +void convolution3d(const SimpleTensor &in, const SimpleTensor &weights, const SimpleTensor &bias, SimpleTensor &out, + int i_offset, int w_offset, int b_offset, int o_offset, + int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights) { - ARM_COMPUTE_UNUSED(fixed_point_position); + const T *in_ptr = in.data() + i_offset; + const T *w_ptr = weights.data() + w_offset; + const T *b_ptr = bias.data() + b_offset; + T *out_ptr = out.data() + o_offset; const int half_width_weights = width_weights / 2; const int half_height_weights = height_weights / 2; @@ -72,8 +81,8 @@ void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, const int idx = xk + half_width_weights; const int idy = yk + half_height_weights; - const T i_value = in[offset_slice_in + xk + yk * width_in]; - const T w_value = weights[idx + idy * width_weights + ifm * width_weights * height_weights]; + const T i_value = in_ptr[offset_slice_in + xk + yk * width_in]; + const T w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights]; acc += i_value * w_value; } @@ -82,14 +91,21 @@ void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, } // Accumulate the bias and store the result - *out = acc + (*bias); + *out_ptr = acc + (*b_ptr); } // 3D convolution for fixed point type template ::value, int>::type = 0> -void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights, - int fixed_point_position) +void convolution3d(const SimpleTensor &in, const SimpleTensor &weights, const SimpleTensor &bias, SimpleTensor &out, + int i_offset, int w_offset, int b_offset, int o_offset, + int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights) { + const T *in_ptr = in.data() + i_offset; + const T *w_ptr = weights.data() + w_offset; + const T *b_ptr = bias.data() + b_offset; + T *out_ptr = out.data() + o_offset; + int fixed_point_position = in.fixed_point_position(); + const int half_width_weights = width_weights / 2; const int half_height_weights = height_weights / 2; @@ -116,8 +132,8 @@ void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, const int idx = xk + half_width_weights; const int idy = yk + half_height_weights; - const fixed_point i_value(in[offset_slice_in + xk + yk * width_in], fixed_point_position, true); - const fixed_point w_value(weights[idx + idy * width_weights + ifm * width_weights * height_weights], fixed_point_position, true); + const fixed_point i_value(in_ptr[offset_slice_in + xk + yk * width_in], fixed_point_position, true); + const fixed_point w_value(w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights], fixed_point_position, true); const fixed_point iw = i_value * w_value; acc = iw + acc; } @@ -126,12 +142,79 @@ void convolution3d(const T *in, const T *weights, const T *bias, T *out, int xi, } // Get the bias - const fixed_point b(*bias, fixed_point_position, true); + const fixed_point b(*b_ptr, fixed_point_position, true); // Accumulate the bias and covert back acc = acc + b; fixed_point res(acc); - *out = res.raw(); + *out_ptr = res.raw(); +} + +// 3D convolution for QASYMM8 type +template <> +void convolution3d(const SimpleTensor &in, const SimpleTensor &weights, const SimpleTensor &bias, SimpleTensor &out, + int i_offset, int w_offset, int b_offset, int o_offset, + int xi, int yi, int width_in, int height_in, int depth_in, int width_weights, int height_weights) +{ + const uint8_t *in_ptr = in.data() + i_offset; + const uint8_t *w_ptr = weights.data() + w_offset; + const uint8_t *b_ptr = bias.data() + b_offset; + uint8_t *out_ptr = out.data() + o_offset; + + const int input_offset = -in.quantization_info().offset; + const float input_scale = in.quantization_info().scale; + const int weights_offset = -weights.quantization_info().offset; + const float weights_scale = weights.quantization_info().scale; + const int output_offset = out.quantization_info().offset; + const float output_scale = out.quantization_info().scale; + + int output_multiplier = 0; + int output_shift = 0; + const float multiplier = input_scale * weights_scale / output_scale; + arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); + + const int half_width_weights = width_weights / 2; + const int half_height_weights = height_weights / 2; + + // Reset accumulator + int32_t acc(0); + + // Compute a 2D convolution for each IFM and accumulate the result + for(int ifm = 0; ifm < depth_in; ++ifm) + { + // Compute the offset for the input slice + const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in; + + // Compute 2D convolution + for(int yk = -half_height_weights; yk <= half_height_weights; ++yk) + { + for(int xk = -half_width_weights; xk <= half_width_weights; ++xk) + { + // Check if the pixel is out-of-bound + if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in)) + { + const int idx = xk + half_width_weights; + const int idy = yk + half_height_weights; + + const uint8_t i_value = in_ptr[offset_slice_in + xk + yk * width_in]; + const uint8_t w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights]; + + acc += (i_value + input_offset) * (w_value + weights_offset); + } + } + } + } + + // Accumulate the bias + acc += (*b_ptr); + + acc = asymm_rounding_divide_by_pow2(asymm_int_mult(acc, output_multiplier), output_shift); + acc += output_offset; + acc = std::max(acc, 0); + acc = std::min(acc, 255); + + // Store the result + *out_ptr = acc; } } // namespace @@ -139,7 +222,7 @@ template SimpleTensor convolution_layer(const SimpleTensor &src, const SimpleTensor &weights, const SimpleTensor &bias, const TensorShape &output_shape, const PadStrideInfo &info) { // Create reference - SimpleTensor dst{ output_shape, src.data_type(), 1, src.fixed_point_position() }; + SimpleTensor dst{ output_shape, src.data_type(), 1, src.fixed_point_position(), src.quantization_info() }; // Compute reference const int width_in = src.shape().x(); @@ -182,14 +265,11 @@ SimpleTensor convolution_layer(const SimpleTensor &src, const SimpleTensor ARM_COMPUTE_ASSERT(yo < height_out); // Compute 3D convolution - convolution3d(src.data() + offset_in, - weights.data() + ofm * width_weights * height_weights * depth_weights, - bias.data() + ofm, - dst.data() + offset_out, + convolution3d(src, weights, bias, dst, + offset_in, ofm * width_weights * height_weights * depth_weights, ofm, offset_out, xi, yi, width_in, height_in, depth_in, - width_weights, height_weights, - src.fixed_point_position()); + width_weights, height_weights); } } } @@ -206,6 +286,8 @@ template SimpleTensor convolution_layer(const SimpleTensor &sr const PadStrideInfo &info); template SimpleTensor convolution_layer(const SimpleTensor &src, const SimpleTensor &weights, const SimpleTensor &bias, const TensorShape &output_shape, const PadStrideInfo &info); +template SimpleTensor convolution_layer(const SimpleTensor &src, const SimpleTensor &weights, const SimpleTensor &bias, const TensorShape &output_shape, + const PadStrideInfo &info); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/CPP/UtilsQuantizedAsymm.h b/tests/validation/CPP/UtilsQuantizedAsymm.h new file mode 100644 index 0000000000..b7b69d588a --- /dev/null +++ b/tests/validation/CPP/UtilsQuantizedAsymm.h @@ -0,0 +1,57 @@ +/* + * Copyright (c) 2017 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. + */ +#ifndef __ARM_COMPUTE_TEST_VALIDATION_UTILS_QUANTIZED_ASYMM_H__ +#define __ARM_COMPUTE_TEST_VALIDATION_UTILS_QUANTIZED_ASYMM_H__ + +#include + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/** Rounded to nearest division by a power-of-two. */ +inline int32_t asymm_rounding_divide_by_pow2(int32_t x, int exponent) +{ + const int32_t mask = (1 << exponent) - 1; + const int32_t threshold = (mask >> 1) + (x < 0 ? 1 : 0); + return (x >> exponent) + ((x & mask) > threshold ? 1 : 0); +} + +/** Multiplication of two integers. The same as ARMv7 NEON VQRDMULH instruction. */ +inline int32_t asymm_int_mult(int32_t a, int32_t b) +{ + bool overflow = a == b && a == std::numeric_limits::min(); + int64_t a_64(a); + int64_t b_64(b); + int64_t ab_64 = a_64 * b_64; + int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30)); + int32_t ab_x2_high32 = static_cast((ab_64 + nudge) / (1ll << 31)); + return overflow ? std::numeric_limits::max() : ab_x2_high32; +} +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* __ARM_COMPUTE_TEST_VALIDATION_UTILS_QUANTIZED_ASYMM_H__ */ -- cgit v1.2.1