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>

4 files changed, 171 insertions, 90 deletions

diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h
index 360372d192..9a20769ca1 100644
--- a/arm_compute/runtime/CL/CLFunctions.h
+++ b/arm_compute/runtime/CL/CLFunctions.h
@@ -59,7 +59,8 @@
 #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h"
 #include "arm_compute/runtime/CL/functions/CLGEMM.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMInterleave4x4.h"
-#include "arm_compute/runtime/CL/functions/CLGEMMLowp.h"
+#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h"
+#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h"
 #include "arm_compute/runtime/CL/functions/CLGEMMTranspose1xW.h"
 #include "arm_compute/runtime/CL/functions/CLGaussian3x3.h"
 #include "arm_compute/runtime/CL/functions/CLGaussian5x5.h"
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowp.h b/arm_compute/runtime/CL/functions/CLGEMMLowp.h
deleted file mode 100644
index ffd997f6ec..0000000000
--- a/arm_compute/runtime/CL/functions/CLGEMMLowp.h
+++ /dev/null
@@ -1,89 +0,0 @@
-/*
- * Copyright (c) 2016, 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_CLGEMMLOWP_H__
-#define __ARM_COMPUTE_CLGEMMLOWP_H__
-
-#include "arm_compute/core/CL/ICLKernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h"
-#include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h"
-#include "arm_compute/runtime/CL/CLMemoryGroup.h"
-#include "arm_compute/runtime/CL/CLTensor.h"
-#include "arm_compute/runtime/IFunction.h"
-#include "arm_compute/runtime/IMemoryManager.h"
-
-#include <memory>
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Basic function to execute GEMMLowp on OpenCL. This function calls the following OpenCL kernels:
-*
-*  -# @ref CLGEMMInterleave4x4Kernel
-*  -# @ref CLGEMMTranspose1xWKernel
-*  -# @ref CLGEMMLowpMatrixMultiplyKernel
-*
-*/
-class CLGEMMLowp : public IFunction
-{
-public:
-    /** Constructor */
-    CLGEMMLowp(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
-    /** Initialise the kernel's inputs, output
-     *
-     * @note GEMM_LOWP:  low precision matrix multiply kernel
-     *  This kernel performs the following computation:
-     *
-     *  -# Convert a values from uint8 to int32 and add a_offset to each of them.
-     *  -# Convert b values from uint8 to int32 and add b_offset to each of them.
-     *  -# Compute the int32 matrix product of the resulting a * b.
-     *  -# Add output_offset to each entry of the result.
-     *  -# Multiply each entry of the result and round to the nearest integer
-     *  -# Clamp the resulting int32 values to the [0..255] range and cast to uint8.
-     *
-     * @param[in]  a               First input tensor  (Matrix A). Data types supported: U8.
-     * @param[in]  b               Second input tensor (Matrix B). Data types supported: same as @p a.
-     * @param[out] output          Output tensor. Data types supported: same as @p a.
-     * @param[in]  a_offset        Offset to be added to each element of the matrix A.
-     * @param[in]  b_offset        Offset to be added to each element of the matrix B.
-     * @param[in]  output_offset   Offset to be added to each element of the output matrix
-     * @param[in]  output_mult_int Multiplied with each element of the output matrix
-     * @param[in]  shift           Number of bits to shift right the result.
-     */
-    void configure(const ICLTensor *a, const ICLTensor *b, ICLTensor *output, int32_t a_offset, int32_t b_offset, int32_t output_offset, int32_t output_mult_int, int32_t shift);
-
-    // Inherited methods overridden:
-    void run() override;
-
-private:
-    CLMemoryGroup                  _memory_group;
-    CLGEMMInterleave4x4Kernel      _interleave_kernel;
-    CLGEMMTranspose1xWKernel       _transpose_kernel;
-    CLGEMMLowpMatrixMultiplyKernel _mm_kernel;
-    CLTensor                       _tmp_a;
-    CLTensor                       _tmp_b;
-};
-}
-#endif /*__ARM_COMPUTE_CLGEMMLOWP_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
new file mode 100644
index 0000000000..9944afeac7
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h
@@ -0,0 +1,91 @@
+/*
+ * 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_CLGEMMLOWPMATRIXMULTIPLYCORE_H__
+#define __ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYCORE_H__
+
+#include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h"
+#include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h"
+#include "arm_compute/runtime/CL/CLMemoryGroup.h"
+#include "arm_compute/runtime/CL/CLTensor.h"
+#include "arm_compute/runtime/IFunction.h"
+
+namespace arm_compute
+{
+class IMemoryManager;
+class ICLTensor;
+
+/** Basic function to execute GEMMLowpMatrixMultiplyCore on OpenCL. This function calls the following OpenCL kernels:
+ *
+ *  -# @ref CLGEMMInterleave4x4Kernel  (if the output tensor is a matrix)
+ *  -# @ref CLGEMMTranspose1xWKernel  (if the output tensor is a matrix)
+ *  -# @ref CLGEMMLowpMatrixMultiplyKernel
+ *  -# @ref CLGEMMLowpMatrixAReductionKernel (if the offset of matrix B is not 0)
+ *  -# @ref CLGEMMLowpMatrixBReductionKernel (if the offset of matrix A is not 0)
+ *  -# @ref CLGEMMLowpOffsetContributionKernel
+ *
+*/
+class CLGEMMLowpMatrixMultiplyCore : public IFunction
+{
+public:
+    /** Constructor */
+    CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
+    /** Initialise the kernel's inputs, output
+     *
+     * @note GEMM_LOWP:  low precision GEMM kernel
+     *  This kernel performs the following computations:
+     *
+     *  -# Convert a values from QASYMM8 to int32 and add a_offset to each of them.
+     *  -# Convert b values from QASYMM8 to int32 add b_offset to each of them.
+     *  -# Compute the matrix product of the resulting a * b in int32.
+     *
+     * @param[in]  a      First input tensor  (Matrix A). Data type supported: QASYMM8.
+     * @param[in]  b      Second input tensor (Matrix B). Data type supported: same as @p a
+     * @param[out] output Output tensor. Data type supported: Data type supported: S32
+     */
+    void configure(const ICLTensor *a, const ICLTensor *b, ICLTensor *output);
+
+    // Inherited methods overridden:
+    void run() override;
+
+private:
+    CLMemoryGroup                      _memory_group;
+    CLGEMMLowpMatrixMultiplyKernel     _mm_kernel;
+    CLGEMMInterleave4x4Kernel          _mtx_a_reshape_kernel;
+    CLGEMMTranspose1xWKernel           _mtx_b_reshape_kernel;
+    CLGEMMLowpMatrixAReductionKernel   _mtx_a_reduction_kernel;
+    CLGEMMLowpMatrixBReductionKernel   _mtx_b_reduction_kernel;
+    CLGEMMLowpOffsetContributionKernel _offset_contribution_kernel;
+    CLTensor                           _vector_sum_col;
+    CLTensor                           _vector_sum_row;
+    CLTensor                           _tmp_a;
+    CLTensor                           _tmp_b;
+    int32_t                            _a_offset;
+    int32_t                            _b_offset;
+    bool                               _is_interleaved_transposed;
+};
+}
+#endif /*__ARM_COMPUTE_CLGEMMLOWPMATRIXMULTIPLYCORE_H__ */
diff --git a/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
new file mode 100644
index 0000000000..5c176a960b
--- /dev/null
+++ b/arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h
@@ -0,0 +1,78 @@
+/*
+ * 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_CLGEMMLOWPOUTPUTSTAGE_H__
+#define __ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H__
+
+#include "arm_compute/runtime/CL/ICLSimpleFunction.h"
+
+/** This file contains all available output stages for GEMMLowp on OpenCL.
+ *
+ *  In gemmlowp, the "output stage" is the process that takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyCore),
+ *  and processes it to obtain the final ASYMM8 value.
+ *
+ *  More information about the GEMMLowp output stage can be found at https://github.com/google/gemmlowp/blob/master/doc/output.md
+ */
+
+namespace arm_compute
+{
+class ITensor;
+
+/** Basic function to execute CLGEMMLowpQuantizeDownInt32ToUint8Scale on OpenCL.
+ *
+ *  CLGEMMLowpQuantizeDownInt32ToUint8Scale depends on 3 parameters: result_offset, result_mult_int, result_shift
+ *  The final result is:
+ *
+ *  ((input[i][k] + result_offset) * result_mult_int) >> result_shift
+ *
+ * In case the bias tensor is provided, the final result is:
+ *
+ *  ((input[i][k] + result_offset) * result_mult_int + bias[k]) >> result_shift
+ *
+ *  This function calls the following OpenCL kernels:
+ *
+ * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel
+ *
+ * @note The function accepts also 2 optional input arguments (min and max) which can be used to implement "rectified linear unit" activation functions
+ *       before the result is shifted right by result_shift
+*/
+class CLGEMMLowpQuantizeDownInt32ToUint8Scale : public ICLSimpleFunction
+{
+public:
+    /** Initialise the kernel's inputs, output
+     *
+     * @param[in]  input           Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32
+     * @param[in]  bias            Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required.
+     *                             Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input.
+     * @param[out] output          Output tensor. Data type supported: Data type supported: QASYMM8
+     * @param[in]  result_offset   Offset to be added to each element of the input matrix
+     * @param[in]  result_mult_int Value to be multiplied to each element of the input matrix when once the result_offset has been add
+     * @param[in]  result_shift    Number of bits to shift right the result before converting back to QASYMM8
+     * @param[in]  min             (Optional) Min value used to saturate down the output result before converting back to QASYMM8
+     * @param[in]  max             (Optional) Max value used to saturate up the output result before converting back to QASYMM8,
+     *                             Along with @p min, this value can be used to implement "rectified linear unit" activation functions
+     */
+    void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, int result_offset, int result_mult_int, int result_shift, int min = 0, int max = 0);
+};
+}
+#endif /*__ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H__ */
\ No newline at end of file