COMPMID-697 - Rework GEMMLowp interface on OpenCL

Reworked the interface of GemmLowp in order to make easy the integration
in Android NN

- Added support for different output stage
- Added validation for both matrix multiplication and output stage
- Added bounded relu support in the output stage
- Added in32_t bias support
- Added optimized path for vector by matrix case

This rework is required for:
- Convolution quantized
- Fully connected quantized

Change-Id: I512283d406099cf8c614dd89d0a97ed411143afc
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/110625
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>

1 files changed, 99 insertions, 0 deletions

diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h
new file mode 100644
index 0000000000..aa0583fe81
--- /dev/null
+++ b/arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h
@@ -0,0 +1,99 @@
+/*
+ * 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_CLGEMMLOWREDUCTIONKERNEL_H__
+#define __ARM_COMPUTE_CLGEMMLOWREDUCTIONKERNEL_H__
+
+#include "arm_compute/core/CL/ICLKernel.h"
+
+namespace arm_compute
+{
+class ICLTensor;
+
+/** Common interface for all OpenCL reduction kernels */
+class ICLGEMMLowpReductionKernel : public ICLKernel
+{
+public:
+    /** Constructor */
+    ICLGEMMLowpReductionKernel();
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    ICLGEMMLowpReductionKernel(const ICLGEMMLowpReductionKernel &) = delete;
+    /** Prevent instances of this class from being copied (As this class contains pointers)*/
+    ICLGEMMLowpReductionKernel &operator=(const ICLGEMMLowpReductionKernel &) = delete;
+    /** Allow instances of this class to be moved */
+    ICLGEMMLowpReductionKernel(ICLGEMMLowpReductionKernel &&) = default;
+    /** Allow instances of this class to be moved */
+    ICLGEMMLowpReductionKernel &operator=(ICLGEMMLowpReductionKernel &&) = default;
+
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  input  Input tensor. Data type supported: S8
+     * @param[out] output Output row-vector of sums of all the entries in each row/col of input tensor. Data type supported: S32
+     */
+    virtual void configure(const ICLTensor *input, ICLTensor *output) = 0;
+
+protected:
+    const ICLTensor *_input;
+    ICLTensor       *_output;
+};
+
+/** OpenCL kernel used to compute the row-vectors of sums of all the entries in each row of Matrix A.
+ *
+ * @note This stage is needed to handle the offset of matrix product
+ *       https://github.com/google/gemmlowp/blob/master/doc/low-precision.md
+ */
+class CLGEMMLowpMatrixAReductionKernel : public ICLGEMMLowpReductionKernel
+{
+public:
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  mtx_a          Input tensor. Data type supported: QASYMM8
+     * @param[out] vector_sum_row Output row-vector of sums of all the entries in each row of mtx_a. Data type supported: S32
+     */
+    void configure(const ICLTensor *mtx_a, ICLTensor *vector_sum_row) override;
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+};
+
+/** OpenCL kernel used to compute the row-vectors of sums of all the entries in each column of Matrix B.
+ *
+ * @note This stage is needed to handle the offset of matrix product
+ *       https://github.com/google/gemmlowp/blob/master/doc/low-precision.md
+ */
+class CLGEMMLowpMatrixBReductionKernel : public ICLGEMMLowpReductionKernel
+{
+public:
+    /** Initialise the kernel's input and output.
+     *
+     * @param[in]  mtx_b          Input tensor. Data type supported: Data type supported: QASYMM8
+     * @param[out] vector_sum_col Output row-vector of sums of all the entries in each column of mtx_b. Data type supported: S32
+     */
+    void configure(const ICLTensor *mtx_b, ICLTensor *vector_sum_col) override;
+
+    // Inherited methods overridden:
+    void run(const Window &window, cl::CommandQueue &queue) override;
+};
+} // namespace arm_compute
+
+#endif /* __ARM_COMPUTE_CLGEMMLOWREDUCTIONKERNEL_H__ */