COMPMID-661: Add QASYMM8 support (and basic tests) to CLDepthwiseConvolution3x3 kernel (#28)

Change-Id: I51bebe74e3814c1245812ad575fe7854d460674f
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/109864
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>

5 files changed, 303 insertions, 13 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 94cc02a705..9a2bb81708 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -187,6 +187,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "copy_planes_3p", "channel_combine.cl" },
     { "copy_to_keypoint", "fast_corners.cl" },
     { "depthwise_convolution_3x3", "depthwise_convolution.cl" },
+    { "depthwise_convolution_3x3_quantized", "depthwise_convolution_quantized.cl" },
     { "depthwise_im2col", "depthwise_convolution.cl" },
     { "depthwise_vector_to_tensor", "depthwise_convolution.cl" },
     { "depthwise_weights_reshape", "depthwise_convolution.cl" },
@@ -419,6 +420,10 @@ const std::map<std::string, std::string> CLKernelLibrary::_program_source_map =
 #include "./cl_kernels/depthwise_convolution.clembed"
     },
     {
+        "depthwise_convolution_quantized.cl",
+#include "./cl_kernels/depthwise_convolution_quantized.clembed"
+    },
+    {
         "dequantization_layer.cl",
 #include "./cl_kernels/dequantization_layer.clembed"
     },
diff --git a/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
new file mode 100644
index 0000000000..19a509bd0a
--- /dev/null
+++ b/src/core/CL/cl_kernels/depthwise_convolution_quantized.cl
@@ -0,0 +1,258 @@
+/*
+ * 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.
+ */
+
+#include "helpers_asymm.h"
+
+#if defined(CONV_STRIDE_X)
+
+#if CONV_STRIDE_X == 1
+#define convolution1x3 convolution1x3_stride_1
+#elif CONV_STRIDE_X == 2
+#define convolution1x3 convolution1x3_stride_2
+#elif CONV_STRIDE_X == 3
+#define convolution1x3 convolution1x3_stride_3
+#else /* CONV_STRIDE_X */
+#error "Stride not supported"
+#endif /* CONV_STRIDE_X */
+
+/** Compute a 1D horizontal convolution of size 3 and stride 1 for uchar type.
+ *
+ * @param[in] left_pixel    Pointer to the left pixel.
+ * @param[in] left_coeff    Weight of the left pixel
+ * @param[in] middle_coeff  Weight of the middle pixel
+ * @param[in] right_coeff   Weight of the right pixel
+ * @param[in] input_offset  Quantized offset of zero point of the input tensor data range
+ * @param[in] weight_offset Quantized offset of zero point of the weights tensor data range
+ *
+ * @return a int2 containing 2 convoluted values.
+ */
+inline int2 convolution1x3_stride_1(__global const uchar *left_pixel,
+                                    const int             left_coeff,
+                                    const int             middle_coeff,
+                                    const int             right_coeff,
+                                    const int             input_offset,
+                                    const int             weight_offset)
+{
+    int4 temp = CONVERT(vload4(0, left_pixel), int4);
+
+    int2 left   = CONVERT(temp.s01, int2);
+    int2 middle = CONVERT(temp.s12, int2);
+    int2 right  = CONVERT(temp.s23, int2);
+
+    return (left + input_offset) * (int2)(left_coeff + weight_offset) + (middle + input_offset) * (int2)(middle_coeff + weight_offset) + (right + input_offset) * (int2)(right_coeff + weight_offset);
+}
+
+/** Compute a 1D horizontal convolution of size 3 and stride 2 for uchar type.
+ *
+ * @param[in] left_pixel    Pointer to the left pixel.
+ * @param[in] left_coeff    Weight of the left pixel
+ * @param[in] middle_coeff  Weight of the middle pixel
+ * @param[in] right_coeff   Weight of the right pixel
+ * @param[in] input_offset  Quantized offset of zero point of the input tensor data range
+ * @param[in] weight_offset Quantized offset of zero point of the weights tensor data range
+ *
+ * @return a int2 containing 2 convoluted values.
+ */
+inline int2 convolution1x3_stride_2(__global const uchar *left_pixel,
+                                    const int             left_coeff,
+                                    const int             middle_coeff,
+                                    const int             right_coeff,
+                                    const int             input_offset,
+                                    const int             weight_offset)
+{
+    int4 temp0 = CONVERT(vload4(0, left_pixel), int4);
+    int  temp1 = CONVERT(*(left_pixel + 4 * sizeof(uchar)), int);
+
+    int2 left   = CONVERT(temp0.s02, int2);
+    int2 middle = CONVERT(temp0.s13, int2);
+    int2 right  = CONVERT((int2)(temp0.s2, temp1), int2);
+
+    return (left + input_offset) * (int2)(left_coeff + weight_offset) + (middle + input_offset) * (int2)(middle_coeff + weight_offset) + (right + input_offset) * (int2)(right_coeff + weight_offset);
+}
+
+/** Compute a 1D horizontal convolution of size 3 and stride 3 for uchar type.
+ *
+ * @param[in] left_pixel    Pointer to the left pixel.
+ * @param[in] left_coeff    Weight of the left pixel
+ * @param[in] middle_coeff  Weight of the middle pixel
+ * @param[in] right_coeff   Weight of the right pixel
+ * @param[in] input_offset  Quantized offset of zero point of the input tensor data range
+ * @param[in] weight_offset Quantized offset of zero point of the weights tensor data range
+ *
+ * @return a int2 containing 2 convoluted values.
+ */
+inline int2 convolution1x3_stride_3(__global const uchar *left_pixel,
+                                    const int             left_coeff,
+                                    const int             middle_coeff,
+                                    const int             right_coeff,
+                                    const int             input_offset,
+                                    const int             weight_offset)
+{
+    int4 temp0 = CONVERT(vload4(0, left_pixel), int4);
+    int2 temp1 = CONVERT(vload2(0, (left_pixel + 4 * sizeof(uchar))), int2);
+
+    int2 left   = CONVERT(temp0.s03, int2);
+    int2 middle = CONVERT((int2)(temp0.s1, temp1.s0), int2);
+    int2 right  = CONVERT((int2)(temp0.s2, temp1.s1), int2);
+
+    return (left + input_offset) * (int2)(left_coeff + weight_offset) + (middle + input_offset) * (int2)(middle_coeff + weight_offset) + (right + input_offset) * (int2)(right_coeff + weight_offset);
+}
+
+/** Apply a 3x3 convolution matrix to a single channel QASYMM8 input image and return the result.
+ *
+ * Convolution matrix layout:
+ *
+ * [ mat0, mat1, mat2 ]\n
+ * [ mat3, mat4, mat5 ]\n
+ * [ mat6, mat7, mat8 ]\n
+ *
+ * @param[in] src               A pointer to source Image structure
+ * @param[in] mat0              Coefficient from the convolution matrix
+ * @param[in] mat1              Coefficient from the convolution matrix
+ * @param[in] mat2              Coefficient from the convolution matrix
+ * @param[in] mat3              Coefficient from the convolution matrix
+ * @param[in] mat4              Coefficient from the convolution matrix
+ * @param[in] mat5              Coefficient from the convolution matrix
+ * @param[in] mat6              Coefficient from the convolution matrix
+ * @param[in] mat7              Coefficient from the convolution matrix
+ * @param[in] mat8              Coefficient from the convolution matrix
+ * @param[in] input_offset      Quantized offset of zero point of the input tensor data range
+ * @param[in] weight_offset     Quantized offset of zero point of the weights tensor data range
+ * @param[in] output_offset     Quantized offset of zero point of the output tensor data range
+ * @param[in] output_multiplier Output scale multiplier
+ * @param[in] output_shift      Output scale divisor exponent
+ * @param[in] bias              (Optional) Bias value
+ *
+ * @return a uchar2 containing 2 convoluted values.
+ */
+inline uchar2 convolution3x3(
+    Image      *src,
+    const uchar mat0, const uchar mat1, const uchar mat2,
+    const uchar mat3, const uchar mat4, const uchar mat5,
+    const uchar mat6, const uchar mat7, const uchar mat8,
+    const int input_offset, const int weight_offset, const int output_offset,
+    const int output_multiplier, const int output_shift
+#if defined(HAS_BIAS)
+    ,
+    const int bias
+#endif //defined(HAS_BIAS)
+)
+{
+    int2 pixels;
+
+    pixels = convolution1x3(offset(src, 0, 0), mat0, mat1, mat2, input_offset, weight_offset);
+    pixels += convolution1x3(offset(src, 0, 1), mat3, mat4, mat5, input_offset, weight_offset);
+    pixels += convolution1x3(offset(src, 0, 2), mat6, mat7, mat8, input_offset, weight_offset);
+#if defined(HAS_BIAS)
+    pixels += (int2)(bias);
+#endif //defined(HAS_BIAS)
+
+    pixels = ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(pixels, output_multiplier, output_shift, 2);
+    pixels = pixels + output_offset;
+    pixels = clamp(pixels, 0, 255);
+
+    return CONVERT(pixels, uchar2);
+}
+
+/** This function computes the horizontal integral of the image.
+  *
+  * @param[in] src_ptr                               Pointer to the source image. Supported data types: QASYMM8
+  * @param[in] src_stride_x                          Stride of the source image in X dimension (in bytes)
+  * @param[in] src_step_x                            src_stride_x * number of elements along X processed per workitem(in bytes)
+  * @param[in] src_stride_y                          Stride of the source image in Y dimension (in bytes)
+  * @param[in] src_step_y                            src_stride_y * number of elements along Y processed per workitem(in bytes)
+  * @param[in] src_offset_first_element_in_bytes     The offset of the first element in the source image
+  * @param[in] src_stride_z                          Stride of the source tensor in Z dimension (in bytes)
+  * @param[in] src_step_z                            src_stride_z * number of elements along Y processed per workitem(in bytes)
+  * @param[in] dst_ptr                               Pointer to the destination tensor. Supported data types: QASYMM8
+  * @param[in] dst_stride_x                          Stride of the destination tensor in X dimension (in bytes)
+  * @param[in] dst_step_x                            dst_stride_x * number of elements along X processed per workitem(in bytes)
+  * @param[in] dst_stride_y                          Stride of the destination tensor in Y dimension (in bytes)
+  * @param[in] dst_step_y                            dst_stride_y * number of elements along Y processed per workitem(in bytes)
+  * @param[in] dst_stride_z                          Stride of the destination tensor in Z dimension (in bytes)
+  * @param[in] dst_step_z                            dst_stride_z * number of elements along Y processed per workitem(in bytes)
+  * @param[in] dst_offset_first_element_in_bytes     The offset of the first element in the destination tensor
+  * @param[in] weights_ptr                           Pointer to the weights tensor. Supported data types: QASYMM8
+  * @param[in] weights_stride_x                      Stride of the weights tensor in X dimension (in bytes)
+  * @param[in] weights_step_x                        weights_stride_x * number of elements along X processed per workitem(in bytes)
+  * @param[in] weights_stride_y                      Stride of the weights tensor in Y dimension (in bytes)
+  * @param[in] weights_step_y                        weights_stride_y * number of elements along Y processed per workitem(in bytes)
+  * @param[in] weights_stride_z                      Stride of the weights tensor in Z dimension (in bytes)
+  * @param[in] weights_step_z                        weights_stride_z * number of elements along Y processed per workitem(in bytes)
+  * @param[in] weights_offset_first_element_in_bytes The offset of the first element in the weights tensor
+  * @param[in] biases_ptr                            (Optional) Pointer to the biases vector. Supported data types: QASYMM8
+  * @param[in] biases_stride_x                       (Optional) Stride of the biases vector in X dimension (in bytes)
+  * @param[in] biases_step_x                         (Optional) biases_stride_x * number of elements along X processed per workitem(in bytes)
+  * @param[in] biases_offset_first_element_in_bytes  (Optional) The offset of the first element in the biases vector
+  * @param[in] input_offset                          Quantized offset of zero point of the input tensor data range
+  * @param[in] weight_offset                         Quantized offset of zero point of the weights tensor data range
+  * @param[in] output_offset                         Quantized offset of zero point of the output tensor data range
+  * @param[in] output_multiplier                     Output scale multiplier
+  * @param[in] output_shift                          Output scale divisor exponent
+  */
+
+__kernel void depthwise_convolution_3x3_quantized(
+    TENSOR3D_DECLARATION(src),
+    TENSOR3D_DECLARATION(dst),
+    TENSOR3D_DECLARATION(weights),
+#if defined(HAS_BIAS)
+    VECTOR_DECLARATION(biases),
+#endif //defined(HAS_BIAS)
+    int input_offset,
+    int weight_offset,
+    int output_offset,
+    int output_multiplier,
+    int output_shift)
+{
+    Image    src     = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(src);
+    Image    dst     = CONVERT_TENSOR3D_TO_IMAGE_STRUCT(dst);
+    Tensor3D weights = CONVERT_TO_TENSOR3D_STRUCT(weights);
+#if defined(HAS_BIAS)
+    Vector biases = CONVERT_TO_VECTOR_STRUCT_NO_STEP(biases);
+#endif //defined(HAS_BIAS)
+
+    uchar3 offset          = (uchar3)(0, 1, 2) * (uchar3)weights_stride_y;
+    uchar3 weights_values0 = vload3(0, weights.ptr + offset.s0);
+    uchar3 weights_values1 = vload3(0, weights.ptr + offset.s1);
+    uchar3 weights_values2 = vload3(0, weights.ptr + offset.s2);
+
+#if defined(HAS_BIAS)
+    int bias_value = *((__global int *)(vector_offset(&biases, get_global_id(2))));
+#endif //defined(HAS_BIAS)
+
+    uchar2 pixels = convolution3x3(&src, weights_values0.s0, weights_values0.s1, weights_values0.s2,
+                                   weights_values1.s0, weights_values1.s1, weights_values1.s2,
+                                   weights_values2.s0, weights_values2.s1, weights_values2.s2,
+                                   input_offset, weight_offset, output_offset,
+                                   output_multiplier, output_shift
+#if defined(HAS_BIAS)
+                                   ,
+                                   bias_value
+#endif //defined(HAS_BIAS)
+                                  );
+
+    vstore2(pixels, 0, dst.ptr);
+}
+
+#endif //defined(CONV_STRIDE_X)
diff --git a/src/core/CL/cl_kernels/helpers_asymm.h b/src/core/CL/cl_kernels/helpers_asymm.h
index 3c1d58bda1..b44d0f1fd2 100644
--- a/src/core/CL/cl_kernels/helpers_asymm.h
+++ b/src/core/CL/cl_kernels/helpers_asymm.h
@@ -44,6 +44,7 @@
         return (x >> exponent) + select(zero, one, (x & mask) > threshold);                                        \
     }
 
+ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2)
 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8)
 ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16)
 
@@ -80,6 +81,7 @@ ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16)
         return select(ab_x2_high32, INT_MAX, overflow);                                                      \
     }
 
+ASYMM_MULT_IMP(2)
 ASYMM_MULT_IMP(8)
 ASYMM_MULT_IMP(16)
 
diff --git a/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
index 5f42450b9f..208d06d7cd 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolution3x3Kernel.cpp
@@ -33,6 +33,7 @@
 #include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/core/Utils.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
 
 using namespace arm_compute;
 
@@ -48,14 +49,22 @@ BorderSize CLDepthwiseConvolution3x3Kernel::border_size() const
 
 void CLDepthwiseConvolution3x3Kernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights);
     ARM_COMPUTE_ERROR_ON(weights->info()->dimension(0) != 3 || weights->info()->dimension(1) != 3);
 
     if(biases != nullptr)
     {
-        ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+        if(is_data_type_quantized_asymmetric(weights->info()->data_type()))
+        {
+            ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+        }
+        else
+        {
+            ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+        }
         ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(2));
         ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1);
     }
@@ -80,13 +89,12 @@ void CLDepthwiseConvolution3x3Kernel::configure(const ICLTensor *input, const IC
 
     // Set build options
     ARM_COMPUTE_ERROR_ON(_conv_stride_x < 1 || _conv_stride_x > 3);
-    std::set<std::string> options{ "-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x) };
-    if(_biases != nullptr)
-    {
-        options.emplace("-DHAS_BIAS");
-    }
+    CLBuildOptions build_opts;
+    build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(_conv_stride_x));
+    build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
 
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_convolution_3x3", options));
+    std::string kernel_name = is_data_type_quantized_asymmetric(_input->info()->data_type()) ? "depthwise_convolution_3x3_quantized" : "depthwise_convolution_3x3";
+    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Configure kernel window
     const unsigned int num_elems_processed_per_iteration = 2;
@@ -105,6 +113,23 @@ void CLDepthwiseConvolution3x3Kernel::configure(const ICLTensor *input, const IC
     output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape()));
 
     ICLKernel::configure(win);
+
+    // Set static arguments
+    if(is_data_type_quantized_asymmetric(_input->info()->data_type()))
+    {
+        float multiplier        = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale;
+        int   output_multiplier = 0;
+        int   output_shift      = 0;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+        unsigned int idx = 3 * num_arguments_per_3D_tensor() + ((_biases != nullptr) ? num_arguments_per_1D_tensor() : 0);
+
+        _kernel.setArg(idx++, -_input->info()->quantization_info().offset);
+        _kernel.setArg(idx++, -_weights->info()->quantization_info().offset);
+        _kernel.setArg(idx++, _output->info()->quantization_info().offset);
+        _kernel.setArg(idx++, output_multiplier);
+        _kernel.setArg(idx++, output_shift);
+    }
 }
 
 void CLDepthwiseConvolution3x3Kernel::run(const Window &window, cl::CommandQueue &queue)
diff --git a/src/runtime/CL/functions/CLDepthwiseConvolution.cpp b/src/runtime/CL/functions/CLDepthwiseConvolution.cpp
index 156565950a..23a20a3011 100644
--- a/src/runtime/CL/functions/CLDepthwiseConvolution.cpp
+++ b/src/runtime/CL/functions/CLDepthwiseConvolution.cpp
@@ -37,9 +37,9 @@ CLDepthwiseConvolution3x3::CLDepthwiseConvolution3x3()
 
 void CLDepthwiseConvolution3x3::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32);
-    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, weights);
 
     _kernel.configure(input, weights, biases, output, conv_info);
     _border_handler.configure(input, _kernel.border_size(), BorderMode::CONSTANT, PixelValue(0));