COMPMID-1043: Rework GCGEMMMatrixMultiplyKernel interface and allow auto initialization of the tensors

This patch also:
- removes support for already reshaped weights in GCConvolutionLayer
- makes GCConvolutionLayer similar to CLGEMMConvolutionLayer
- enables usage of the GCGEMM function in GCConvolution instead of calling the
  GEMM kernels directly

Change-Id: I3e4a64335555e86e18585d38d8fda4bfdb44e265
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127696
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

5 files changed, 136 insertions, 48 deletions

diff --git a/arm_compute/core/CL/CLTypes.h b/arm_compute/core/CL/CLTypes.h
index ca487814a7..4a03cc9637 100644
--- a/arm_compute/core/CL/CLTypes.h
+++ b/arm_compute/core/CL/CLTypes.h
@@ -24,6 +24,8 @@
 #ifndef __ARM_COMPUTE_CL_TYPES_H__
 #define __ARM_COMPUTE_CL_TYPES_H__
 
+#include "arm_compute/core/GPUTarget.h"
+
 #include <string>
 
 namespace arm_compute
@@ -31,26 +33,6 @@ namespace arm_compute
 /** Default string for the CLKernel configuration id */
 static const std::string default_config_id = "no_config_id";
 
-/** Available GPU Targets */
-enum class GPUTarget
-{
-    UNKNOWN       = 0x101,
-    GPU_ARCH_MASK = 0xF00,
-    MIDGARD       = 0x100,
-    BIFROST       = 0x200,
-    T600          = 0x110,
-    T700          = 0x120,
-    T800          = 0x130,
-    G71           = 0x210,
-    G72           = 0x220,
-    G51           = 0x230,
-    G51BIG        = 0x231,
-    G51LIT        = 0x232,
-    TNOX          = 0x240,
-    TTRX          = 0x250,
-    TBOX          = 0x260
-};
-
 /** Available OpenCL Version */
 enum class CLVersion
 {
diff --git a/arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.h b/arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.h
index 3a0b22f148..cea03a9357 100644
--- a/arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.h
+++ b/arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,6 +25,7 @@
 #define __ARM_COMPUTE_GCGEMMMATRIXMULTIPLYKERNEL_H__
 
 #include "arm_compute/core/GLES_COMPUTE/IGCKernel.h"
+#include "arm_compute/core/GPUTarget.h"
 
 namespace arm_compute
 {
@@ -32,9 +33,6 @@ class IGCTensor;
 
 /** GLES Compute kernel to multiply two input matrices "A" and "B" or to multiply a vector "A" by a matrix "B". All elements of the output matrix/vector will be multiplied by alpha
  *
- * @note If the output tensor is a matrix, the implementation assumes that the input tensors @p input0 and @p input1 are both matrices and reshaped respectively with @ref GCGEMMInterleave4x4Kernel" and @ref GCGEMMTranspose1xWKernel
- * @note If the output tensor is a vector and the data type is F32, the implementation assumes that the first input tensor @p input0 is a vector and the second input tensor @p input1 a matrix. The implementation also assumes that both tensors have not been reshaped
- *
  * @attention The second input tensor must have at least 2 dimensions (matrix)
  *
  */
@@ -64,8 +62,23 @@ public:
      * @param[out] output                    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
      * @param[in]  alpha                     Weight of the matrix product
      * @param[in]  is_interleaved_transposed (Optional) True if input0 and input1 have been reshaped respectively using @ref GCGEMMInterleave4x4Kernel and @ref GCGEMMTranspose1xWKernel
+     * @param[in]  reshape_info              (Optional) GEMM reshape info. If is_interleaved_transposed = true, this object must contain the information to understand how the matrix A and matrix B have been reshaped
+     */
+    void configure(const IGCTensor *input0, const IGCTensor *input1, IGCTensor *output, float alpha, bool is_interleaved_transposed = true, const GEMMReshapeInfo &reshape_info = GEMMReshapeInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref GCGEMMMatrixMultiplyKernel
+     *
+     * @param[in] input0                    Input tensor containing the Matrix A. Data types supported: F16/F32
+     * @param[in] input1                    Input tensor containing the Matrix B. Data type supported: same as @p input0
+     * @param[in] output                    Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
+     * @param[in] alpha                     Weight of the matrix product
+     * @param[in] is_interleaved_transposed True if input0 and input1 have been reshaped respectively using @ref GCGEMMInterleave4x4Kernel and @ref GCGEMMTranspose1xWKernel
+     * @param[in] reshape_info              GEMM reshape info. If is_interleaved_transposed = true, this object must contain the information to understand how the matrix A and matrix B have been reshaped
+     * @param[in] gpu_target                GPU Target
+     *
+     * @return a status
      */
-    void configure(const IGCTensor *input0, const IGCTensor *input1, IGCTensor *output, float alpha, bool is_interleaved_transposed = true);
+    static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, float alpha, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info,
+                           GPUTarget gpu_target);
 
     // Inherited methods overridden:
     void run(const Window &window) override;
diff --git a/arm_compute/core/GPUTarget.h b/arm_compute/core/GPUTarget.h
new file mode 100644
index 0000000000..8a5ca80f49
--- /dev/null
+++ b/arm_compute/core/GPUTarget.h
@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2018 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_GPUTARGET_H__
+#define __ARM_COMPUTE_GPUTARGET_H__
+
+namespace arm_compute
+{
+/** Available GPU Targets */
+enum class GPUTarget
+{
+    UNKNOWN       = 0x101,
+    GPU_ARCH_MASK = 0xF00,
+    MIDGARD       = 0x100,
+    BIFROST       = 0x200,
+    T600          = 0x110,
+    T700          = 0x120,
+    T800          = 0x130,
+    G71           = 0x210,
+    G72           = 0x220,
+    G51           = 0x230,
+    G51BIG        = 0x231,
+    G51LIT        = 0x232,
+    TNOX          = 0x240,
+    TTRX          = 0x250,
+    TBOX          = 0x260
+};
+} // namespace arm_compute
+#endif /* __ARM_COMPUTE_GPUTARGET_H__ */
diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.h
index 54b17b40bb..fa29f447c8 100644
--- a/arm_compute/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.h
+++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCConvolutionLayer.h
@@ -27,15 +27,13 @@
 
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCCol2ImKernel.h"
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCFillBorderKernel.h"
-#include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMInterleave4x4Kernel.h"
-#include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.h"
-#include "arm_compute/core/GLES_COMPUTE/kernels/GCGEMMTranspose1xWKernel.h"
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCIm2ColKernel.h"
 #include "arm_compute/core/GLES_COMPUTE/kernels/GCWeightsReshapeKernel.h"
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/GLES_COMPUTE/GCMemoryGroup.h"
 #include "arm_compute/runtime/GLES_COMPUTE/GCTensor.h"
 #include "arm_compute/runtime/GLES_COMPUTE/functions/GCActivationLayer.h"
+#include "arm_compute/runtime/GLES_COMPUTE/functions/GCGEMM.h"
 #include "arm_compute/runtime/IFunction.h"
 
 #include <memory>
@@ -46,7 +44,6 @@ class IGCTensor;
 
 /** Function to reshape and transpose the weights. This function calls the following kernels:
  * -# @ref GCWeightsReshapeKernel
- * -# @ref GCGEMMTranspose1xWKernel
  */
 class GCConvolutionLayerReshapeWeights : public IFunction
 {
@@ -55,22 +52,18 @@ public:
     GCConvolutionLayerReshapeWeights();
     /** Set the input and output tensors.
      *
-     * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
-     *                          Data type supported: F16/F32.
-     * @param[in]  biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
-     * @param[out] output       Destination tensor. Data types supported: Same as @p weights.
-     * @param[in]  transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise.
-     *                          Data types supported: Same as @p weights.
+     * @param[in]  weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM].
+     *                     Data type supported: F16/F32.
+     * @param[in]  biases  Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights.
+     * @param[out] output  Destination tensor. Data types supported: Same as @p weights.
      */
-    void configure(const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, bool transpose1xW);
+    void configure(const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output);
     // Inherited methods overridden:
     void run() override;
 
 private:
-    GCWeightsReshapeKernel   _weights_reshape_kernel;
-    GCGEMMTranspose1xWKernel _weights_transposed_kernel;
-    GCTensor                 _weights_reshaped;
-    bool                     _transpose1xW;
+    GCWeightsReshapeKernel _weights_reshape_kernel;
+    GCTensor               _weights_reshaped;
 };
 
 /** Basic function to compute the convolution layer. This function calls the following GLES kernels:
@@ -86,7 +79,14 @@ class GCConvolutionLayer : public IFunction
 public:
     /** Default constructor */
     GCConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr);
-
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    GCConvolutionLayer(const GCConvolutionLayer &) = delete;
+    /** Default move constructor */
+    GCConvolutionLayer(GCConvolutionLayer &&) = default;
+    /** Prevent instances of this class from being copied (As this class contains pointers) */
+    GCConvolutionLayer &operator=(const GCConvolutionLayer &) = delete;
+    /** Default move assignment operator */
+    GCConvolutionLayer &operator=(GCConvolutionLayer &&) = default;
     /** Set the input and output tensors.
      *
      * @param[in]  input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
@@ -105,6 +105,26 @@ public:
      */
     void configure(const IGCTensor *input, const IGCTensor *weights, const IGCTensor *biases, IGCTensor *output, const PadStrideInfo &conv_info,
                    const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref GCConvolutionLayer.
+     *
+     * @param[in]  input        Source tensor. 3 lower dimensions represent a single input [width, height, IFM],
+     *                          while every optional dimension from 4 and above represent a batch of inputs.
+     *                          Data types supported: QS8/QASYMM8/QS16/F16/F32.
+     * @param[in]  weights      Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input.
+     * @param[in]  biases       Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM].
+     *                          Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type.
+     * @param[out] output       Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs.
+     *                          Data types supported: Same as @p input.
+     * @param[in]  conv_info    Contains padding and stride information described in @ref PadStrideInfo.
+     * @param[in]  weights_info Specifies if the weights tensor has been reshaped with GCWeightsReshapeKernel. If this is not part of the fully connected layer the weights
+     *                          tensor has also been transposed with GCGEMMTranspose1xWKernel. Data type supported: Same as @p input.
+     * @param[in]  dilation     (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
+     * @param[in]  act_info     (Optional) Activation layer information in case of a fused activation.
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
+                           const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo());
 
     // Inherited methods overridden:
     void run() override;
@@ -115,20 +135,30 @@ private:
      * @param input                     Input tensor. Data types supported: F16/F32.
      * @param weights                   Weights tensor. Data type supported: Same as @p input.
      * @param output                    Output tensor. Data types supported: Same as @p input,
-     * @param is_interleaved_transposed Flag that signals if matrix is interleaved transposed
      */
-    void configure_mm(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output, bool is_interleaved_transposed = true);
+    void configure_mm(const IGCTensor *input, const IGCTensor *weights, IGCTensor *output);
+    /** Static function to check if given info will lead to a valid configuration of @ref GCGEMMConvolutionLayer matrix multiply routines
+     *
+     * @param[in] input   Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32.
+     * @param[in] weights Weights tensor. Data type supported: Same as @p input.
+     * @param[in] output  Output tensor. Data types supported: Same as @p input,
+     *                                      except for input of QASYMM8 type where output should be of S32 type.
+     *
+     * @return a status
+     */
+    static Status validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output);
 
 private:
     GCMemoryGroup                    _memory_group;
     GCConvolutionLayerReshapeWeights _reshape_weights;
     GCIm2ColKernel                   _input_im2col_kernel;
-    GCGEMMInterleave4x4Kernel        _input_interleave_kernel;
-    GCGEMMMatrixMultiplyKernel       _mm_kernel;
+    GCGEMM                           _mm_gemm;
     GCCol2ImKernel                   _output_col2im_kernel;
     GCFillBorderKernel               _fill_border;
     GCActivationLayer                _activationlayer_function;
 
+    const IGCTensor *_original_weights;
+
     GCTensor _input_im2col_reshaped;
     GCTensor _input_interleaved_reshaped;
     GCTensor _weights_reshaped;
@@ -136,9 +166,7 @@ private:
     GCTensor _gemm_output;
     GCTensor _tmp_output;
 
-    bool _append_bias;
-    bool _is_fully_connected_convolution;
-    bool _are_weights_reshaped;
+    bool _is_first_run;
     bool _is_activationlayer_enabled;
 };
 }
diff --git a/arm_compute/runtime/GLES_COMPUTE/functions/GCGEMM.h b/arm_compute/runtime/GLES_COMPUTE/functions/GCGEMM.h
index 31ad0abaa0..a1d6c8a438 100644
--- a/arm_compute/runtime/GLES_COMPUTE/functions/GCGEMM.h
+++ b/arm_compute/runtime/GLES_COMPUTE/functions/GCGEMM.h
@@ -69,6 +69,20 @@ public:
      *                       if the reshape of matrix B should happen only for the first run
      */
     void configure(const IGCTensor *a, const IGCTensor *b, const IGCTensor *c, IGCTensor *output, float alpha, float beta, const GEMMInfo &gemm_info = GEMMInfo());
+    /** Static function to check if given info will lead to a valid configuration of @ref GCGEMM.
+     *
+     * @param[in]  a         First input tensor  (Matrix or Vector A). Data types supported: F16/F32
+     * @param[in]  b         Second input tensor (Matrix B). Data type supported: same as @p a.
+     * @param[in]  c         Third input tensor  (Matrix C). It can be a nullptr if just the multiplication between @p a and @p b is needed. Data type supported: same as @p a.
+     * @param[out] output    Output tensor. Data type supported: same as @p a
+     * @param[in]  alpha     Weight of the matrix product
+     * @param[in]  beta      Weight of matrix C
+     * @param[in]  gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and
+     *                       if the reshape of matrix B should happen only for the first run
+     *
+     * @return a status
+     */
+    static Status validate(const ITensorInfo *a, const ITensorInfo *b, const IGCTensor *c, const ITensorInfo *output, const float alpha, const float beta, const GEMMInfo &gemm_info = GEMMInfo());
 
     // Inherited methods overridden:
     void run() override;
@@ -83,6 +97,8 @@ private:
     GCTensor                   _tmp_b;
     bool                       _is_interleaved_transposed;
     bool                       _run_addition;
+    bool                       _is_first_run;
+    bool                       _reshape_b_only_on_first_run;
 };
 }