COMPMID-3320: Add cl_image support for GEMMReshaped T_NT

COMPMID-3321: Add cl_image support for GEMMReshaped NT_T

- Added support for cl_image in CLGEMMMatrixMultiplyReshapedKernel (both
  NT and T kernels)
- Extended the tests for the validating rhs_info.export_to_cl_image =
  true
- Added utility macros in OpenCL to load data from a OpenCL image object
- Updated doxygen documentation in CLGEMMMatrixMultiplyReshapedKernel.h
- Updated doxygen documentation in CLGEMMReshapeRHSMatrixKernel.h

Change-Id: I953b10e4ef205d1b76dcbc366e5a91fd5a8e1d5c
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3329
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

3 files changed, 89 insertions, 19 deletions

diff --git a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
index ee8e57fa8c..aeedd50e0b 100644
--- a/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h
@@ -55,19 +55,30 @@ public:
      *       Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the
      *       multiplications. i.e. float c = (half)a * (half)b
      *
-     * @param[in]  input0    Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32. The number of dimensions for the LHS matrix must be less or equal than 4
+     * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will fetch the RHS data using the OpenCL read_image built-in function.
+     *       Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer,
+     *       the following conditions are required:
+     *       -# rhs_info.n0 can only be 4, 8 and 16
+     *       -# rhs_info.k0 can only be 4, 8 and 16
+     *       -# Data type can only be F32
+     *       -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
+     *       -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
+     *       -# input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
+     *       -# input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
+     *
+     * @param[in]  input0    Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32 (only F32 if rhs_info.export_to_cl_image = true). The number of dimensions for the LHS matrix must be less or equal than 4
      * @param[in]  input1    Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3
      * @param[in]  input2    Input tensor containing the bias matrix. Data type supported: same as @p input0.
      * @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]  beta      Weight of the matrix bias
-     * @param[in]  lhs_info  LHS matrix information used for reshaping the input0 tensor.  Only the following values are supported:
+     * @param[in]  lhs_info  LHS matrix information used for reshaping the input0 tensor. Only the following values are supported:
      *                       lhs_info.m0: 2,3,4,5,6,7,8
      *                       lhs_info.k0: 2,3,4,8,16
      *                       lhs_info.transpose: false
      * @param[in]  rhs_info  RHS matrix information used for reshaping the input1 tensor.  Only the following values are supported:
-     *                       rhs_info.n0: 2,3,4,8,16
-     *                       rhs_info.k0: 2,3,4,8,16
+     *                       rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
+     *                       rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
      *                       rhs_info.transpose: true
      * @param[in]  gemm_info GEMM information used to retrieve the original dimensions of the input matrices
      *
@@ -82,8 +93,19 @@ public:
      *       Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the
      *       multiplications. i.e. float c = (half)a * (half)b
      *
+     * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will fetch the RHS data using the OpenCL read_image built-in function.
+     *       Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer,
+     *       the following conditions are required:
+     *       -# rhs_info.n0 can only be 4, 8 and 16
+     *       -# rhs_info.k0 can only be 4, 8 and 16
+     *       -# Data type can only be F32
+     *       -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
+     *       -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
+     *       -# input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
+     *       -# input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
+     *
      * @param[in]  compile_context The compile context to be used.
-     * @param[in]  input0          Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32. The number of dimensions for the LHS matrix must be less or equal than 4
+     * @param[in]  input0          Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32  (only F32 if rhs_info.export_to_cl_image = true). The number of dimensions for the LHS matrix must be less or equal than 4
      * @param[in]  input1          Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3
      * @param[in]  input2          Input tensor containing the bias matrix. Data type supported: same as @p input0.
      * @param[out] output          Output tensor to store the result of matrix multiplication. Data type supported: same as @p input0
@@ -94,8 +116,8 @@ public:
      *                             lhs_info.k0: 2,3,4,8,16
      *                             lhs_info.transpose: false
      * @param[in]  rhs_info        RHS matrix information used for reshaping the input1 tensor.  Only the following values are supported:
-     *                             rhs_info.n0: 2,3,4,8,16
-     *                             rhs_info.k0: 2,3,4,8,16
+     *                             rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
+     *                             rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
      *                             rhs_info.transpose: true
      * @param[in]  gemm_info       GEMM information used to retrieve the original dimensions of the input matrices
      *
@@ -107,7 +129,22 @@ public:
                    const GEMMKernelInfo    &gemm_info);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMMatrixMultiplyReshapedKernel
      *
-     * @param[in] input0    Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32. The number of dimensions for the LHS matrix must be less or equal than 4
+     * @note The F16 computation also supports mixed precision through the gemm_info.fp_mixed_precision flag.
+     *       Mixed precision combines different floating precisions during the computation, in particular, F32 for the accumulations and F16 for the
+     *       multiplications. i.e. float c = (half)a * (half)b
+     *
+     * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will fetch the RHS data using the OpenCL read_image built-in function.
+     *       Reading from the OpenCL image object can increase the performance. However, since the OpenCL image object is created importing the OpenCL buffer,
+     *       the following conditions are required:
+     *       -# rhs_info.n0 can only be 4, 8 and 16
+     *       -# rhs_info.k0 can only be 4, 8 and 16
+     *       -# Data type can only be F32
+     *       -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
+     *       -# The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
+     *       -# input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
+     *       -# input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
+     *
+     * @param[in] input0    Input tensor containing the LHS reshaped matrix. Data type supported: F16/F32  (only F32 if rhs_info.export_to_cl_image = true). The number of dimensions for the LHS matrix must be less or equal than 4
      * @param[in] input1    Input tensor containing the RHS reshaped matrix. Data type supported: same as @p input0. The number of dimensions for the RHS matrix must be less or equal than 3
      * @param[in] input2    Input tensor info containing the bias matrix. 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
@@ -118,8 +155,8 @@ public:
      *                      lhs_info.k0: 2,3,4,8,16
      *                      lhs_info.transpose: false
      * @param[in] rhs_info  RHS matrix information used for reshaping the input1 tensor.  Only the following values are supported:
-     *                      rhs_info.n0: 2,3,4,8,16
-     *                      rhs_info.k0: 2,3,4,8,16
+     *                      rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
+     *                      rhs_info.k0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image = true)
      *                      rhs_info.transpose: true
      * @param[in] gemm_info GEMM information used to retrieve the original dimensions of the input matrices
      *
@@ -141,10 +178,10 @@ private:
     ICLTensor       *_output;
     bool             _slide_matrix_b;
     bool             _reinterpret_output_as_3d;
-    unsigned int     _k;
     bool             _use_dummy_work_items;
     bool             _add_bias;
     bool             _broadcast_bias;
+    bool             _export_to_cl_image;
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_CLGEMMMATRIXMULTIPLYRESHAPEDKERNEL_H*/
\ No newline at end of file
diff --git a/arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h b/arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h
index 0e6352bdbb..59a4aaa912 100644
--- a/arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h
+++ b/arm_compute/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h
@@ -48,12 +48,23 @@ public:
     CLGEMMReshapeRHSMatrixKernel &operator=(CLGEMMReshapeRHSMatrixKernel &&) = default;
     /** Initialise the kernel's input and output.
      *
+     * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will guarantee the OpenCL pitch alignment for the output tensor,
+     *       required to create a OpenCL image object from buffer in @ref CLGEMMMatrixMultiplyReshapedKernel and in @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
+     *       Since the OpenCL image object is created importing the OpenCL buffer, the following conditions are required:
+     *       -# rhs_info.n0 can only be 4, 8 and 16
+     *       -# rhs_info.k0 can only be 4, 8 and 16
+     *       -# Data type can only be F32
+     *       -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
+     *       -# output width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
+     *       -# output (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
+     *       -# The output tensor should be only consumed by @ref CLGEMMMatrixMultiplyReshapedKernel or @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
+     *
      * @param[in]  input    Input tensor. Data types supported: All
      * @param[out] output   Output tensor. Data type supported: same as @p input
      * @param[in]  rhs_info RHS matrix information to be used for reshaping. This object contains all the necessary
      *                      information to reshape the input tensor. Only the following values are supported:
-     *                      rhs_info.n0: 2,3,4,8,16
-     *                      rhs_info.k0: 1,2,3,4,8,16 (k0 = 1 only if rhs_info.transpose = false)
+     *                      rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image == true)
+     *                      rhs_info.k0: 1,2,3,4,8,16 (k0 = 1 only if rhs_info.transpose = false), (only 4, 8 and 16 if rhs_info.export_to_cl_image == true)
      *                      rhs_info.h0: greater than 0
      *                      rhs_info.transpose: true, false
      *                      rhs_info.interleave: true, false
@@ -61,13 +72,24 @@ public:
     void configure(const ICLTensor *input, ICLTensor *output, const GEMMRHSMatrixInfo &rhs_info);
     /** Initialise the kernel's input and output.
      *
+     * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will guarantee the OpenCL pitch alignment for the output tensor,
+     *       required to create a OpenCL image object from buffer in @ref CLGEMMMatrixMultiplyReshapedKernel and in @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
+     *       Since the OpenCL image object is created importing the OpenCL buffer, the following conditions are required:
+     *       -# rhs_info.n0 can only be 4, 8 and 16
+     *       -# rhs_info.k0 can only be 4, 8 and 16
+     *       -# Data type can only be F32
+     *       -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
+     *       -# output width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
+     *       -# output (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
+     *       -# The output tensor should be only consumed by @ref CLGEMMMatrixMultiplyReshapedKernel or @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
+     *
      * @param[in]  compile_context The compile context to be used.
      * @param[in]  input           Input tensor. Data types supported: All
      * @param[out] output          Output tensor. Data type supported: same as @p input
      * @param[in]  rhs_info        RHS matrix information to be used for reshaping. This object contains all the necessary
      *                             information to reshape the input tensor. Only the following values are supported:
-     *                             rhs_info.n0: 2,3,4,8,16
-     *                             rhs_info.k0: 1,2,3,4,8,16 (k0 = 1 only if rhs_info.transpose = false)
+     *                             rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image == true)
+     *                             rhs_info.k0: 1,2,3,4,8,16 (k0 = 1 only if rhs_info.transpose = false), (only 4, 8 and 16 if rhs_info.export_to_cl_image == true)
      *                             rhs_info.h0: greater than 0
      *                             rhs_info.transpose: true, false
      *                             rhs_info.interleave: true, false
@@ -75,12 +97,23 @@ public:
     void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const GEMMRHSMatrixInfo &rhs_info);
     /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMReshapeRHSMatrixKernel
      *
+     * @note If rhs_info.export_to_cl_image = true, this OpenCL kernel will guarantee the OpenCL pitch alignment for the output tensor,
+     *       required to create a OpenCL image object from buffer in @ref CLGEMMMatrixMultiplyReshapedKernel and in @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
+     *       Since the OpenCL image object is created importing the OpenCL buffer, the following conditions are required:
+     *       -# rhs_info.n0 can only be 4, 8 and 16
+     *       -# rhs_info.k0 can only be 4, 8 and 16
+     *       -# Data type can only be F32
+     *       -# The platform should support the OpenCL cl_khr_image2d_from_buffer extension
+     *       -# output width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
+     *       -# output (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
+     *       -# The output tensor should be only consumed by @ref CLGEMMMatrixMultiplyReshapedKernel or @ref CLGEMMMatrixMultiplyReshapedOnlyRHSKernel
+     *
      * @param[in] input    Input tensor info. Data types supported: All
      * @param[in] output   Output tensor info which stores the interleaved matrix. Data type supported: same as @p input.
      * @param[in] rhs_info RHS matrix information to be used for reshaping. This object contains all the necessary
      *                     information to reshape the input tensor. Only the following values are supported:
-     *                     rhs_info.n0: 2,3,4,8,16
-     *                     rhs_info.k0: 1,2,3,4,8,16 (k0 = 1 only if rhs_info.transpose = false)
+     *                     rhs_info.n0: 2,3,4,8,16 (only 4, 8 and 16 if rhs_info.export_to_cl_image == true)
+     *                     rhs_info.k0: 1,2,3,4,8,16 (k0 = 1 only if rhs_info.transpose = false),(only 4, 8 and 16 if rhs_info.export_to_cl_image == true)
      *                     rhs_info.h0: greater than 0
      *                     rhs_info.transpose: true, false
      *                     rhs_info.interleave: true, false
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h
index 4e73edba4b..d151496537 100644
--- a/arm_compute/core/Types.h
+++ b/arm_compute/core/Types.h
@@ -1912,8 +1912,8 @@ struct GEMMLHSMatrixInfo
 struct GEMMRHSMatrixInfo
 {
     GEMMRHSMatrixInfo() = default;
-    GEMMRHSMatrixInfo(unsigned int n, unsigned int k, unsigned int h, bool trans, bool inter)
-        : n0(n), k0(k), h0(h), transpose(trans), interleave(inter)
+    GEMMRHSMatrixInfo(unsigned int n, unsigned int k, unsigned int h, bool trans, bool inter, bool export_to_cl_img)
+        : n0(n), k0(k), h0(h), transpose(trans), interleave(inter), export_to_cl_image(export_to_cl_img)
     {
     }
     unsigned int n0{ 1 };                     /**< Number of columns processed by the matrix multiplication */