Optimize Transposed Convolution for CL backend (FP32/16)

This patch optimizes transposed convolution for CL backend by rewriting it in a single kernel instead of three (flip_kernel + upsample + conv). The new kernel skips the upsampling step which reduces the input space of convolution by stride_x * stride_y, resulting in significant performance improvement. It also skips the kernel flipping by traversing the weights accordingly, thus reduces the memory footprint.

Resolves: COMPMID-5676
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: I8a333212dc7c5f7f0597aa58b0d56d44814baa14
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8588
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 90 insertions, 0 deletions

diff --git a/src/gpu/cl/operators/ClTransposedConvolution.h b/src/gpu/cl/operators/ClTransposedConvolution.h
new file mode 100644
index 0000000000..bc04387df5
--- /dev/null
+++ b/src/gpu/cl/operators/ClTransposedConvolution.h
@@ -0,0 +1,90 @@
+/*
+ * Copyright (c) 2022 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_CL_TRANSPOSED_CONVOLUTION_H
+#define ARM_COMPUTE_CL_TRANSPOSED_CONVOLUTION_H
+
+#include "src/gpu/cl/ClCompileContext.h"
+#include "src/gpu/cl/IClKernel.h"
+#include "src/gpu/cl/IClOperator.h"
+namespace arm_compute
+{
+namespace opencl
+{
+/** Basic function to simulate a directly convolution layer. This function calls the following OpenCL kernels:
+ *
+ * -# @ref opencl::ClTransposedConvolution
+ */
+class ClTransposedConvolution : public IClOperator
+{
+public:
+    /** Default constructor */
+    ClTransposedConvolution() = default;
+    /** Default Destructor */
+    ~ClTransposedConvolution() = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    ClTransposedConvolution(const ClTransposedConvolution &) = delete;
+    /** Default move constructor */
+    ClTransposedConvolution(ClTransposedConvolution &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    ClTransposedConvolution &operator=(const ClTransposedConvolution &) = delete;
+    /** Default move assignment operator */
+    ClTransposedConvolution &operator=(ClTransposedConvolution &&) = default;
+
+    /** Set the input, weights, biases and output tensors.
+     *
+     * @note: Only NHWC data layout is supported
+     *
+     * @param[in]  compile_context The compile context to be used.
+     * @param[in]  input           Input tensor info with dimensions [IFM, width, height, batch]
+     *                             Data types supported: F16/F32.
+     * @param[in]  weights         Weight tensor info with dimensions [IFM, width, height, OFM].
+     *                             Data type supported: Same as @p input
+     * @param[in]  biases          (Optional) Biases tensor info. Biases are 1D tensor with dimension [OFM].
+     *                             Data type supported: Should match @p input data type
+     * @param[out] output          Output tensor info with dimensions [OFM, width, height, batch]
+     *                             The 1st dimension must be equal to the 4th dimension of the @p weights tensor.
+     *                             Data types supported: Same as @p input.
+     * @param[in]  deconv_info     Contains padding and stride information described in @ref PadStrideInfo.
+     *
+     */
+    void configure(const CLCompileContext &compile_context, const ITensorInfo *input, const ITensorInfo *weights,
+                   const ITensorInfo *biases, ITensorInfo *output, const PadStrideInfo &deconv_info);
+    /** Static function to check if given info will lead to a valid configuration
+     *
+     * Similar to ClTransposedConvolution::configure()
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases,
+                           const ITensorInfo *output, const PadStrideInfo &deconv_info);
+
+    // Inherited method overridden
+    void run(ITensorPack &tensors) override;
+
+private:
+    std::unique_ptr<IClKernel> _transposed_conv_kernel{ nullptr };
+};
+} // namespace opencl
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_CL_TRANSPOSED_CONVOLUTION_H */
\ No newline at end of file