Compute Library
 21.02
CLGEMMHelpers.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
25 
31 
32 #include <utility>
33 
34 namespace arm_compute
35 {
36 namespace cl_gemm
37 {
39 
40 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,
41  bool lhs_interleave, bool rhs_interleave, bool lhs_transpose, bool rhs_transpose, bool export_to_cl_image)
42 {
43  ARM_COMPUTE_ERROR_ON(m0 == 0 || n0 == 0);
44  v0 = std::max(std::min(static_cast<int>(m / m0), static_cast<int>(v0)), static_cast<int>(1));
45  h0 = std::max(std::min(static_cast<int>(n / n0), static_cast<int>(h0)), static_cast<int>(1));
46 
47  const GEMMLHSMatrixInfo lhs_info(m0, k0, v0, lhs_transpose, lhs_interleave);
48  const GEMMRHSMatrixInfo rhs_info(n0, k0, h0, rhs_transpose, rhs_interleave, export_to_cl_image);
49 
50  return std::make_pair(lhs_info, rhs_info);
51 }
52 
53 std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> select_lhs_rhs_info(std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> info_img,
54  std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> info_buf,
55  unsigned int n, unsigned int k, unsigned int b, DataType data_type)
56 {
57  const TensorInfo tensor_rhs_info(TensorShape(n, k, b), 1, data_type);
58  const TensorShape shape = compute_rhs_reshaped_shape(tensor_rhs_info, info_img.second);
59  const TensorInfo tensor_reshaped_info(shape, 1, data_type);
60 
61  if(bool(validate_image2d_support_on_rhs(tensor_reshaped_info, info_img.second)))
62  {
63  return info_img;
64  }
65  else
66  {
67  return info_buf;
68  }
69 }
70 
72 {
73  constexpr unsigned int num_floats_per_pixel = 4;
74 
75  const unsigned int stride_y_in_elements = tensor->strides_in_bytes()[1] / tensor->element_size();
76  const unsigned int pixel_aligment = get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device());
77  const unsigned int row_pitch_alignment = pixel_aligment * num_floats_per_pixel;
78  const unsigned int round_up_width = ((stride_y_in_elements + row_pitch_alignment - 1) / row_pitch_alignment) * row_pitch_alignment;
79  const unsigned int padding = round_up_width - stride_y_in_elements;
80 
81  tensor->extend_padding(PaddingSize(0, padding, 0, 0));
82 }
83 
84 Status validate_image2d_support_on_rhs(const ITensorInfo &tensor_reshaped_info, const GEMMRHSMatrixInfo &rhs_info)
85 {
86  if(rhs_info.export_to_cl_image)
87  {
88  ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.n0 == 2) || (rhs_info.n0 == 3), "Export to cl_image only supported with n0 = 4, 8 or 16");
89  ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.k0 == 2) || (rhs_info.k0 == 3), "Export to cl_image only supported with k0 = 4, 8 or 16");
91  ARM_COMPUTE_RETURN_ERROR_ON_MSG(!image2d_from_buffer_supported(CLKernelLibrary::get().get_device()), "The extension cl_khr_image2d_from_buffer is not supported on the target platform");
92  ARM_COMPUTE_RETURN_ERROR_ON_MSG(get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device()) == 0, "Impossible to retrieve the cl_image pitch alignment");
93 
94  // Check the width and height of the output tensor.
95  // Since we cannot create a 3d image from a buffer, the third dimension is collapsed on the second dimension
96  const size_t max_image_w = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>();
97  const size_t max_image_h = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>();
98 
99  ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.tensor_shape()[0] > max_image_w * 4, "Not supported width for cl_image");
100  ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.tensor_shape()[1] * tensor_reshaped_info.tensor_shape()[2] > max_image_h, "Not supported height for cl_image");
101  }
102 
103  return Status{};
104 }
105 } // namespace cl_gemm
106 } // namespace arm_compute
bool image2d_from_buffer_supported(const cl::Device &device)
Helper function to check whether the cl_khr_image2d_from_buffer extension is supported.
Definition: CLHelpers.cpp:368
Shape of a tensor.
Definition: TensorShape.h:39
SimpleTensor< float > b
Definition: DFT.cpp:157
Status validate_image2d_support_on_rhs(const ITensorInfo &tensor_reshaped_info, const GEMMRHSMatrixInfo &rhs_info)
Utility function to validate the image2d OpenCL object support on the RHS reshaped matrix...
1 channel, 1 F32 per channel
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
GEMM LHS (Left Hand Side) matrix information.
Definition: Types.h:1968
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
Status class.
Definition: Error.h:52
bool export_to_cl_image
True if the reshaped rhs has to be exported to cl_image.
Definition: Types.h:1995
Copyright (c) 2017-2021 Arm Limited.
1 channel, 1 F16 per channel
const DataType data_type
Definition: Im2Col.cpp:150
unsigned int k0
Number of partial accumulations performed by the matrix multiplication.
Definition: Types.h:1991
void update_padding_for_cl_image(ITensorInfo *tensor)
Update padding required to export the OpenCL buffer to OpenCL image2d.
GEMM RHS (Right Hand Side) matrix information.
Definition: Types.h:1983
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
unsigned int n0
Number of columns processed by the matrix multiplication.
Definition: Types.h:1990
TensorShape compute_rhs_reshaped_shape(const ITensorInfo &a, const GEMMRHSMatrixInfo &rhs_info)
Calculate the Right Hand Side matrix reshaped shape.
virtual size_t element_size() const =0
Element size in bytes calculated as data_size() * num_channels()
BorderSize PaddingSize
Container for 2D padding size.
Definition: Types.h:382
size_t get_cl_image_pitch_alignment(const cl::Device &device)
Helper function to get the cl_image pitch alignment in pixels.
Definition: CLHelpers.cpp:373
Manages all the OpenCL kernels compilation and caching, provides accessors for the OpenCL Context...
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, bool export_to_cl_image)
Configure GEMMLHSMatrixInfo and GEMMRHSMatrixInfo.
Wrapper to configure the Khronos OpenCL C++ header.
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
Definition: Error.h:244
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:45
virtual const Strides & strides_in_bytes() const =0
The strides in bytes for accessing each dimension of the tensor.
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(t,...)
Definition: Validate.h:694
DataType
Available data types.
Definition: Types.h:77
virtual bool extend_padding(const PaddingSize &padding)=0
Update the offset to the first element, the strides and the total size.
std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo > select_lhs_rhs_info(std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo > info_img, std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo > info_buf, unsigned int n, unsigned int k, unsigned int b, DataType data_type)
Select GEMMLHSMatrixInfo and GEMMRHSMatrixInfo.
const cl::Device & get_device()
Gets the CL device for which the programs are created.