COMPMID-3325: Add support in gemm_tuner for cl_image

Change-Id: I78f815005516ca0e83366bab017884530c1d2e86
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3518
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 94 insertions, 54 deletions

diff --git a/arm_compute/core/CL/gemm/CLGEMMHelpers.h b/arm_compute/core/CL/gemm/CLGEMMHelpers.h
index b2d9f21c3d..a370f9171a 100644
--- a/arm_compute/core/CL/gemm/CLGEMMHelpers.h
+++ b/arm_compute/core/CL/gemm/CLGEMMHelpers.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 Arm Limited.
+ * Copyright (c) 2019-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -24,6 +24,7 @@
 #ifndef ARM_COMPUTE_CLGEMMHELPERS_H
 #define ARM_COMPUTE_CLGEMMHELPERS_H
 
+#include "arm_compute/core/TensorInfo.h"
 #include "arm_compute/core/Types.h"
 
 namespace arm_compute
@@ -48,6 +49,12 @@ namespace cl_gemm
  */
 std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_lhs_rhs_info(unsigned int m, unsigned int n, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0,
                                                                        bool lhs_interleave, bool rhs_interleave, bool lhs_transpose, bool rhs_transpose);
+
+/** Update padding required to export the OpenCL buffer to OpenCL image2d
+ *
+ * @param[in,out] tensor ITensorInfo of the tensor required to be exported to OpenCL image2d
+ */
+void update_padding_for_cl_image(ITensorInfo *tensor);
 } // namespace cl_gemm
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_CLGEMMHELPERS_H */
diff --git a/examples/gemm_tuner/cl_gemm_reshaped.cpp b/examples/gemm_tuner/cl_gemm_reshaped.cpp
index 8000e1ab7b..93125dd4a3 100644
--- a/examples/gemm_tuner/cl_gemm_reshaped.cpp
+++ b/examples/gemm_tuner/cl_gemm_reshaped.cpp
@@ -26,6 +26,7 @@
 #endif /* ARM_COMPUTE_CL */
 
 #include "CommonGemmExampleOptions.h"
+#include "arm_compute/core/CL/gemm/CLGEMMHelpers.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h"
 #include "arm_compute/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h"
 #include "arm_compute/core/Helpers.h"
@@ -52,15 +53,16 @@ namespace
 /** Structure holding all tunable gemm configs specific to this example/strategy */
 struct GemmConfigs
 {
-    size_t m0{ 4 };                /**< Number of rows processed by the matrix multiplication */
-    size_t n0{ 4 };                /**< Number of columns processed by the matrix multiplication */
-    size_t k0{ 4 };                /**< Number of partial accumulations performed by the matrix multiplication */
-    size_t v0{ 1 };                /**< Number of vertical blocks of size (m0xk0) stored on the same output row */
-    size_t h0{ 1 };                /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */
-    bool   interleave_lhs{ true }; /**< Interleave lhs matrix */
-    bool   transpose_lhs{ true };  /**< Transpose lhs matrix. */
-    bool   interleave_rhs{ true }; /**< Interleave rhs matrix */
-    bool   transpose_rhs{ true };  /**< Transpose rhs matrix. */
+    size_t m0{ 4 };                        /**< Number of rows processed by the matrix multiplication */
+    size_t n0{ 4 };                        /**< Number of columns processed by the matrix multiplication */
+    size_t k0{ 4 };                        /**< Number of partial accumulations performed by the matrix multiplication */
+    size_t v0{ 1 };                        /**< Number of vertical blocks of size (m0xk0) stored on the same output row */
+    size_t h0{ 1 };                        /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */
+    bool   interleave_lhs{ true };         /**< Interleave lhs matrix */
+    bool   transpose_lhs{ true };          /**< Transpose lhs matrix. */
+    bool   interleave_rhs{ true };         /**< Interleave rhs matrix */
+    bool   transpose_rhs{ true };          /**< Transpose rhs matrix. */
+    bool   export_to_cl_image_rhs{ true }; /**< Export rhs matrix to cl_image. */
 };
 
 /** Formatted output of the GemmConfigs type
@@ -84,6 +86,7 @@ struct GemmConfigs
     os << "transpose_lhs : " << (configs.transpose_lhs ? true_str : false_str) << std::endl;
     os << "interleave_rhs : " << (configs.interleave_rhs ? true_str : false_str) << std::endl;
     os << "transpose_rhs : " << (configs.transpose_rhs ? true_str : false_str) << std::endl;
+    os << "export_to_cl_image_rhs : " << (configs.export_to_cl_image_rhs ? true_str : false_str) << std::endl;
     return os;
 }
 
@@ -103,7 +106,8 @@ public:
           h0(parser.add_positional_option<SimpleOption<size_t>>("h0", 1)),
           interleave_lhs(parser.add_positional_option<SimpleOption<size_t>>("interleave_lhs", 1)),
           interleave_rhs(parser.add_positional_option<SimpleOption<size_t>>("interleave_rhs", 1)),
-          transpose_rhs(parser.add_positional_option<SimpleOption<size_t>>("transpose_rhs", 1))
+          transpose_rhs(parser.add_positional_option<SimpleOption<size_t>>("transpose_rhs", 1)),
+          export_to_cl_image_rhs(parser.add_positional_option<SimpleOption<size_t>>("export_to_cl_image_rhs", 1))
     {
         m0->set_help("Number of rows processed by the matrix multiplication");
         n0->set_help("Number of columns processed by the matrix multiplication");
@@ -116,6 +120,7 @@ public:
         // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other
         // 2 variants (both transposed and none transposed)
         transpose_rhs->set_help("Transpose rhs matrix but not lhs matrix (1) / Do not transpose rhs matrix but do transpose lhs matrix (0)");
+        export_to_cl_image_rhs->set_help("Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0)");
     }
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     GemmConfigOptions(const GemmConfigOptions &) = delete;
@@ -138,7 +143,8 @@ public:
     // FIXME: Currently we only support 2 variants of the gemm reshaped kernels in which transpose_lhs and
     // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other
     // 2 variants (both transposed and none transposed)
-    SimpleOption<size_t> *transpose_rhs; /**< Transpose rhs matrix option (1 enable; 0 disable). Also set the lhs matrix transpose option to the opposite. */
+    SimpleOption<size_t> *transpose_rhs;          /**< Transpose rhs matrix option (1 enable; 0 disable). Also set the lhs matrix transpose option to the opposite. */
+    SimpleOption<size_t> *export_to_cl_image_rhs; /**< Export rhs matrix to cl_image.*/
 };
 
 /** Consumes the gemm configuration options and creates a structure containing all information
@@ -159,9 +165,10 @@ GemmConfigs consume_gemm_configs(const GemmConfigOptions &options)
     // FIXME: Currently we only support 2 variants of the gemm reshaped kernels in which transpose_lhs and
     // transpose_rhs are the opposites of each other. In the future we may extend the kernels to include the other
     // 2 variants (both transposed and none transposed)
-    configs.transpose_lhs  = options.transpose_rhs->value() == 0;
-    configs.interleave_rhs = options.interleave_rhs->value() != 0;
-    configs.transpose_rhs  = options.transpose_rhs->value() != 0;
+    configs.transpose_lhs          = options.transpose_rhs->value() == 0;
+    configs.interleave_rhs         = options.interleave_rhs->value() != 0;
+    configs.transpose_rhs          = options.transpose_rhs->value() != 0;
+    configs.export_to_cl_image_rhs = options.export_to_cl_image_rhs->value() != 0;
     return configs;
 }
 
@@ -230,11 +237,12 @@ public:
         lhs_info.transpose  = configs.transpose_lhs;
 
         GEMMRHSMatrixInfo rhs_info;
-        rhs_info.n0         = configs.n0;
-        rhs_info.k0         = configs.k0;
-        rhs_info.h0         = configs.h0;
-        rhs_info.interleave = configs.interleave_rhs;
-        rhs_info.transpose  = configs.transpose_rhs;
+        rhs_info.n0                 = configs.n0;
+        rhs_info.k0                 = configs.k0;
+        rhs_info.h0                 = configs.h0;
+        rhs_info.interleave         = configs.interleave_rhs;
+        rhs_info.transpose          = configs.transpose_rhs;
+        rhs_info.export_to_cl_image = configs.export_to_cl_image_rhs;
 
         GEMMKernelInfo kernel_info;
         kernel_info.m                       = params.M;
@@ -251,6 +259,11 @@ public:
         // Initialise rhs_reshaped tensor info
         auto_init_if_empty(*rhs_reshaped.info(), rhs.info()->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*rhs.info(), rhs_info)));
 
+        if(rhs_info.export_to_cl_image)
+        {
+            arm_compute::cl_gemm::update_padding_for_cl_image(rhs_reshaped.info());
+        }
+
         // Configure reshape lhs function
         reshape_lhs.configure(&lhs, &lhs_reshaped, lhs_info);
 
diff --git a/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp b/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp
index 4e7b93dba8..45a28abca6 100644
--- a/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp
+++ b/examples/gemm_tuner/cl_gemm_reshaped_rhs_only.cpp
@@ -26,6 +26,7 @@
 #endif /* ARM_COMPUTE_CL */
 
 #include "CommonGemmExampleOptions.h"
+#include "arm_compute/core/CL/gemm/CLGEMMHelpers.h"
 #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/KernelDescriptors.h"
@@ -51,12 +52,13 @@ namespace
 /** Structure holding all tunable gemm configs specific to this example/strategy */
 struct GemmConfigs
 {
-    size_t m0{ 4 };                /**< Number of rows processed by the matrix multiplication */
-    size_t n0{ 4 };                /**< Number of columns processed by the matrix multiplication */
-    size_t k0{ 4 };                /**< Number of partial accumulations performed by the matrix multiplication */
-    size_t h0{ 1 };                /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */
-    bool   interleave_rhs{ true }; /**< Interleave rhs matrix */
-    bool   transpose_rhs{ true };  /**< Transpose rhs matrix */
+    size_t m0{ 4 };                        /**< Number of rows processed by the matrix multiplication */
+    size_t n0{ 4 };                        /**< Number of columns processed by the matrix multiplication */
+    size_t k0{ 4 };                        /**< Number of partial accumulations performed by the matrix multiplication */
+    size_t h0{ 1 };                        /**< Number of horizontal blocks of size (k0xn0) stored on the same output row */
+    bool   interleave_rhs{ true };         /**< Interleave rhs matrix */
+    bool   transpose_rhs{ true };          /**< Transpose rhs matrix */
+    bool   export_to_cl_image_rhs{ true }; /**< Export rhs matrix to cl_image.*/
 };
 
 /** Formatted output of the GemmConfigs type
@@ -77,6 +79,7 @@ struct GemmConfigs
     os << "h0 : " << configs.h0 << std::endl;
     os << "interleave_rhs : " << (configs.interleave_rhs ? true_str : false_str) << std::endl;
     os << "transpose_rhs : " << (configs.transpose_rhs ? true_str : false_str) << std::endl;
+    os << "export_to_cl_image_rhs : " << (configs.export_to_cl_image_rhs ? true_str : false_str) << std::endl;
     return os;
 }
 
@@ -94,7 +97,8 @@ public:
           k0(parser.add_positional_option<SimpleOption<size_t>>("k0", 4)),
           h0(parser.add_positional_option<SimpleOption<size_t>>("h0", 1)),
           interleave_rhs(parser.add_positional_option<SimpleOption<size_t>>("interleave_rhs", 1)),
-          transpose_rhs(parser.add_positional_option<SimpleOption<size_t>>("transpose_rhs", 1))
+          transpose_rhs(parser.add_positional_option<SimpleOption<size_t>>("transpose_rhs", 1)),
+          export_to_cl_image_rhs(parser.add_positional_option<SimpleOption<size_t>>("export_to_cl_image_rhs", 1))
     {
         m0->set_help("Number of rows processed by the matrix multiplication");
         n0->set_help("Number of columns processed by the matrix multiplication");
@@ -102,6 +106,7 @@ public:
         h0->set_help("Number of horizontal blocks of size (k0xn0) stored on the same output row");
         interleave_rhs->set_help("Interleave rhs matrix (1) / Do not interleave rhs matrix (0)");
         transpose_rhs->set_help("Transpose rhs matrix (1) / Do not transpose rhs matrix (0)");
+        export_to_cl_image_rhs->set_help("Export rhs matrix to cl_image (1) / Do not export rhs matrix to cl_image (0)");
     }
     /** Prevent instances of this class from being copied (As this class contains pointers) */
     GemmConfigOptions(const GemmConfigOptions &) = delete;
@@ -114,12 +119,13 @@ public:
     /** Default destructor */
     ~GemmConfigOptions() = default;
 
-    SimpleOption<size_t> *m0;             /**< Number of rows processed by the matrix multiplication option */
-    SimpleOption<size_t> *n0;             /**< Number of columns processed by the matrix multiplication option */
-    SimpleOption<size_t> *k0;             /**< Number of partial accumulations performed by the matrix multiplication option */
-    SimpleOption<size_t> *h0;             /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */
-    SimpleOption<size_t> *interleave_rhs; /**< Interleave rhs matrix option (1 enable; 0 disable) */
-    SimpleOption<size_t> *transpose_rhs;  /**< Transpose rhs matrix option (1 enable; 0 disable) */
+    SimpleOption<size_t> *m0;                     /**< Number of rows processed by the matrix multiplication option */
+    SimpleOption<size_t> *n0;                     /**< Number of columns processed by the matrix multiplication option */
+    SimpleOption<size_t> *k0;                     /**< Number of partial accumulations performed by the matrix multiplication option */
+    SimpleOption<size_t> *h0;                     /**< Number of horizontal blocks of size (k0xn0) stored on the same output row option */
+    SimpleOption<size_t> *interleave_rhs;         /**< Interleave rhs matrix option (1 enable; 0 disable) */
+    SimpleOption<size_t> *transpose_rhs;          /**< Transpose rhs matrix option (1 enable; 0 disable) */
+    SimpleOption<size_t> *export_to_cl_image_rhs; /**< Export rhs matrix to cl_image.*/
 };
 
 /** Consumes the gemm configuration options and creates a structure containing all information
@@ -131,12 +137,13 @@ public:
 GemmConfigs consume_gemm_configs(const GemmConfigOptions &options)
 {
     GemmConfigs configs;
-    configs.m0             = options.m0->value();
-    configs.n0             = options.n0->value();
-    configs.k0             = options.k0->value();
-    configs.h0             = options.h0->value();
-    configs.interleave_rhs = options.interleave_rhs->value() != 0;
-    configs.transpose_rhs  = options.transpose_rhs->value() != 0;
+    configs.m0                     = options.m0->value();
+    configs.n0                     = options.n0->value();
+    configs.k0                     = options.k0->value();
+    configs.h0                     = options.h0->value();
+    configs.interleave_rhs         = options.interleave_rhs->value() != 0;
+    configs.transpose_rhs          = options.transpose_rhs->value() != 0;
+    configs.export_to_cl_image_rhs = options.export_to_cl_image_rhs->value() != 0;
     return configs;
 }
 
@@ -200,11 +207,12 @@ public:
         lhs_info.k0 = configs.k0;
 
         GEMMRHSMatrixInfo rhs_info;
-        rhs_info.n0         = configs.n0;
-        rhs_info.k0         = configs.k0;
-        rhs_info.h0         = configs.h0;
-        rhs_info.interleave = configs.interleave_rhs;
-        rhs_info.transpose  = configs.transpose_rhs;
+        rhs_info.n0                 = configs.n0;
+        rhs_info.k0                 = configs.k0;
+        rhs_info.h0                 = configs.h0;
+        rhs_info.interleave         = configs.interleave_rhs;
+        rhs_info.transpose          = configs.transpose_rhs;
+        rhs_info.export_to_cl_image = configs.export_to_cl_image_rhs;
 
         GEMMKernelInfo kernel_info;
         kernel_info.m                       = params.M;
@@ -218,6 +226,10 @@ public:
         // Initialise rhs_reshaped tensor info
         auto_init_if_empty(*rhs_reshaped.info(), rhs.info()->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*rhs.info(), rhs_info)));
 
+        if(rhs_info.export_to_cl_image)
+        {
+            arm_compute::cl_gemm::update_padding_for_cl_image(rhs_reshaped.info());
+        }
         // Configure function
         gemm.configure(&lhs, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info);
 
diff --git a/src/core/CL/gemm/CLGEMMHelpers.cpp b/src/core/CL/gemm/CLGEMMHelpers.cpp
index 8f54d3f670..1165e70273 100644
--- a/src/core/CL/gemm/CLGEMMHelpers.cpp
+++ b/src/core/CL/gemm/CLGEMMHelpers.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019 Arm Limited.
+ * Copyright (c) 2019-2020 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -23,6 +23,8 @@
  */
 #include "arm_compute/core/CL/gemm/CLGEMMHelpers.h"
 
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+
 #include <utility>
 
 namespace arm_compute
@@ -51,5 +53,18 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_lhs_rhs_info(unsigned
 
     return std::make_pair(lhs_info, rhs_info);
 }
+
+void update_padding_for_cl_image(ITensorInfo *tensor)
+{
+    constexpr unsigned int num_floats_per_pixel = 4;
+
+    const unsigned int stride_y_in_elements = tensor->strides_in_bytes()[1] / tensor->element_size();
+    const unsigned int pixel_aligment       = get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device());
+    const unsigned int row_pitch_alignment  = pixel_aligment * num_floats_per_pixel;
+    const unsigned int round_up_width       = ((stride_y_in_elements + row_pitch_alignment - 1) / row_pitch_alignment) * row_pitch_alignment;
+    const unsigned int padding              = round_up_width - stride_y_in_elements;
+
+    tensor->extend_padding(PaddingSize(0, padding, 0, 0));
+}
 } // namespace cl_gemm
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp b/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp
index 5e09899ec8..c57066ae03 100644
--- a/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp
+++ b/src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.cpp
@@ -29,6 +29,7 @@
 #include "arm_compute/core/CL/CLValidate.h"
 #include "arm_compute/core/CL/ICLTensor.h"
 #include "arm_compute/core/CL/OpenCL.h"
+#include "arm_compute/core/CL/gemm/CLGEMMHelpers.h"
 #include "arm_compute/core/Error.h"
 #include "arm_compute/core/Helpers.h"
 #include "arm_compute/core/TensorInfo.h"
@@ -107,15 +108,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 
     if(rhs_info.export_to_cl_image)
     {
-        constexpr unsigned int num_floats_per_pixel = 4;
-
-        const unsigned int stride_y_in_elements = output->strides_in_bytes()[1] / output->element_size();
-        const unsigned int pixel_aligment       = get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device());
-        const unsigned int row_pitch_alignment  = pixel_aligment * num_floats_per_pixel;
-        const unsigned int round_up_width       = ((stride_y_in_elements + row_pitch_alignment - 1) / row_pitch_alignment) * row_pitch_alignment;
-        const unsigned int padding              = round_up_width - stride_y_in_elements;
-
-        output->extend_padding(PaddingSize(0, padding, 0, 0));
+        arm_compute::cl_gemm::update_padding_for_cl_image(output);
     }
 
     // Collapse along the Z direction