Compute Library
 21.02
CLGEMMMatrixMultiplyReshapedKernel Class Reference

OpenCL kernel to multiply matrices when both the input matrices LHS (input0) and RHS (input1) have been reshaped. More...

#include <CLGEMMMatrixMultiplyReshapedKernel.h>

Collaboration diagram for CLGEMMMatrixMultiplyReshapedKernel:
[legend]

Public Member Functions

 CLGEMMMatrixMultiplyReshapedKernel ()
 Default Constructor. More...
 
 CLGEMMMatrixMultiplyReshapedKernel (const CLGEMMMatrixMultiplyReshapedKernel &)=delete
 Prevent instances of this class from being copied (As this class contains pointers) More...
 
CLGEMMMatrixMultiplyReshapedKerneloperator= (const CLGEMMMatrixMultiplyReshapedKernel &)=delete
 Prevent instances of this class from being copied (As this class contains pointers) More...
 
 CLGEMMMatrixMultiplyReshapedKernel (CLGEMMMatrixMultiplyReshapedKernel &&)=default
 Allow instances of this class to be moved. More...
 
CLGEMMMatrixMultiplyReshapedKerneloperator= (CLGEMMMatrixMultiplyReshapedKernel &&)=default
 Allow instances of this class to be moved. More...
 
void configure (const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
 Initialise the kernel's input and output. More...
 
void configure (const CLCompileContext &compile_context, const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
 Initialise the kernel's input and output. More...
 
void run (const Window &window, cl::CommandQueue &queue) override
 Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue. More...
 
- Public Member Functions inherited from ICLKernel
 ICLKernel ()
 Constructor. More...
 
cl::Kernel & kernel ()
 Returns a reference to the OpenCL kernel of this object. More...
 
template<typename T >
void add_1D_array_argument (unsigned int &idx, const ICLArray< T > *array, const Strides &strides, unsigned int num_dimensions, const Window &window)
 Add the passed 1D array's parameters to the object's kernel's arguments starting from the index idx. More...
 
void add_1D_tensor_argument (unsigned int &idx, const ICLTensor *tensor, const Window &window)
 Add the passed 1D tensor's parameters to the object's kernel's arguments starting from the index idx. More...
 
void add_1D_tensor_argument_if (bool cond, unsigned int &idx, const ICLTensor *tensor, const Window &window)
 Add the passed 1D tensor's parameters to the object's kernel's arguments starting from the index idx if the condition is true. More...
 
void add_2D_tensor_argument (unsigned int &idx, const ICLTensor *tensor, const Window &window)
 Add the passed 2D tensor's parameters to the object's kernel's arguments starting from the index idx. More...
 
void add_2D_tensor_argument_if (bool cond, unsigned int &idx, const ICLTensor *tensor, const Window &window)
 Add the passed 2D tensor's parameters to the object's kernel's arguments starting from the index idx if the condition is true. More...
 
void add_3D_tensor_argument (unsigned int &idx, const ICLTensor *tensor, const Window &window)
 Add the passed 3D tensor's parameters to the object's kernel's arguments starting from the index idx. More...
 
void add_4D_tensor_argument (unsigned int &idx, const ICLTensor *tensor, const Window &window)
 Add the passed 4D tensor's parameters to the object's kernel's arguments starting from the index idx. More...
 
virtual void run_op (ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
 Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue. More...
 
template<typename T >
void add_argument (unsigned int &idx, T value)
 Add the passed parameters to the object's kernel's arguments starting from the index idx. More...
 
void set_lws_hint (const cl::NDRange &lws_hint)
 Set the Local-Workgroup-Size hint. More...
 
cl::NDRange lws_hint () const
 Return the Local-Workgroup-Size hint. More...
 
void set_wbsm_hint (const cl_int &wbsm_hint)
 Set the workgroup batch size modifier hint. More...
 
cl_int wbsm_hint () const
 Return the workgroup batch size modifier hint. More...
 
const std::string & config_id () const
 Get the configuration ID. More...
 
void set_target (GPUTarget target)
 Set the targeted GPU architecture. More...
 
void set_target (cl::Device &device)
 Set the targeted GPU architecture according to the CL device. More...
 
GPUTarget get_target () const
 Get the targeted GPU architecture. More...
 
size_t get_max_workgroup_size ()
 Get the maximum workgroup size for the device the CLKernelLibrary uses. More...
 
template<unsigned int dimension_size>
void add_tensor_argument (unsigned &idx, const ICLTensor *tensor, const Window &window)
 
template<typename T , unsigned int dimension_size>
void add_array_argument (unsigned &idx, const ICLArray< T > *array, const Strides &strides, unsigned int num_dimensions, const Window &window)
 Add the passed array's parameters to the object's kernel's arguments starting from the index idx. More...
 
- Public Member Functions inherited from IKernel
 IKernel ()
 Constructor. More...
 
virtual ~IKernel ()=default
 Destructor. More...
 
virtual bool is_parallelisable () const
 Indicates whether or not the kernel is parallelisable. More...
 
virtual BorderSize border_size () const
 The size of the border for that kernel. More...
 
const Windowwindow () const
 The maximum window the kernel can be executed on. More...
 

Static Public Member Functions

static Status validate (const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
 Static function to check if given info will lead to a valid configuration of CLGEMMMatrixMultiplyReshapedKernel. More...
 
- Static Public Member Functions inherited from ICLKernel
static constexpr unsigned int num_arguments_per_1D_array ()
 Returns the number of arguments enqueued per 1D array object. More...
 
static constexpr unsigned int num_arguments_per_1D_tensor ()
 Returns the number of arguments enqueued per 1D tensor object. More...
 
static constexpr unsigned int num_arguments_per_2D_tensor ()
 Returns the number of arguments enqueued per 2D tensor object. More...
 
static constexpr unsigned int num_arguments_per_3D_tensor ()
 Returns the number of arguments enqueued per 3D tensor object. More...
 
static constexpr unsigned int num_arguments_per_4D_tensor ()
 Returns the number of arguments enqueued per 4D tensor object. More...
 
static cl::NDRange gws_from_window (const Window &window)
 Get the global work size given an execution window. More...
 

Detailed Description

OpenCL kernel to multiply matrices when both the input matrices LHS (input0) and RHS (input1) have been reshaped.

Note
The input matrices input0 and input1 must be reshaped through CLGEMMReshapeLHSMatrixKernel and CLGEMMReshapeRHSMatrixKernel

Definition at line 39 of file CLGEMMMatrixMultiplyReshapedKernel.h.

Constructor & Destructor Documentation

◆ CLGEMMMatrixMultiplyReshapedKernel() [1/3]

Default Constructor.

Definition at line 206 of file CLGEMMMatrixMultiplyReshapedKernel.cpp.

207  : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_output_as_3d(false), _use_dummy_work_items(false), _add_bias(false),
208  _broadcast_bias(false), _export_to_cl_image(false), _k(1)
209 {
210 }

◆ CLGEMMMatrixMultiplyReshapedKernel() [2/3]

Prevent instances of this class from being copied (As this class contains pointers)

◆ CLGEMMMatrixMultiplyReshapedKernel() [3/3]

Allow instances of this class to be moved.

Member Function Documentation

◆ configure() [1/2]

void configure ( const ICLTensor input0,
const ICLTensor input1,
const ICLTensor input2,
ICLTensor output,
float  alpha,
float  beta,
const GEMMLHSMatrixInfo lhs_info,
const GEMMRHSMatrixInfo rhs_info,
const GEMMKernelInfo gemm_info 
)

Initialise the kernel's input and output.

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
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:
  1. rhs_info.n0 can only be 4, 8 and 16
  2. rhs_info.k0 can only be 4, 8 and 16
  3. Data type can only be F32
  4. The platform should support the OpenCL cl_khr_image2d_from_buffer extension
  5. The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
  6. input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
  7. input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
Parameters
[in]input0Input 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
[in]input1Input tensor containing the RHS reshaped matrix. Data type supported: same as input0. The number of dimensions for the RHS matrix must be less or equal than 3
[in]input2Input tensor containing the bias matrix. Data type supported: same as input0.
[out]outputOutput tensor to store the result of matrix multiplication. Data type supported: same as input0
[in]alphaWeight of the matrix product
[in]betaWeight of the matrix bias
[in]lhs_infoLHS 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
[in]rhs_infoRHS matrix information used for reshaping the input1 tensor. Only the following values are supported: 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
[in]gemm_infoGEMM information used to retrieve the original dimensions of the input matrices
Note
lhs_info.k0 must be equal to rhs_info.k0

Definition at line 212 of file CLGEMMMatrixMultiplyReshapedKernel.cpp.

References CLKernelLibrary::get().

215 {
216  configure(CLKernelLibrary::get().get_compile_context(), input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info);
217 }
void configure(const ICLTensor *input0, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
Initialise the kernel&#39;s input and output.
static CLKernelLibrary & get()
Access the KernelLibrary singleton.

◆ configure() [2/2]

void configure ( const CLCompileContext compile_context,
const ICLTensor input0,
const ICLTensor input1,
const ICLTensor input2,
ICLTensor output,
float  alpha,
float  beta,
const GEMMLHSMatrixInfo lhs_info,
const GEMMRHSMatrixInfo rhs_info,
const GEMMKernelInfo gemm_info 
)

Initialise the kernel's input and output.

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
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:
  1. rhs_info.n0 can only be 4, 8 and 16
  2. rhs_info.k0 can only be 4, 8 and 16
  3. Data type can only be F32
  4. The platform should support the OpenCL cl_khr_image2d_from_buffer extension
  5. The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
  6. input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
  7. input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
Parameters
[in]compile_contextThe compile context to be used.
[in]input0Input 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
[in]input1Input tensor containing the RHS reshaped matrix. Data type supported: same as input0. The number of dimensions for the RHS matrix must be less or equal than 3
[in]input2Input tensor containing the bias matrix. Data type supported: same as input0.
[out]outputOutput tensor to store the result of matrix multiplication. Data type supported: same as input0
[in]alphaWeight of the matrix product
[in]betaWeight of the matrix bias
[in]lhs_infoLHS 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
[in]rhs_infoRHS matrix information used for reshaping the input1 tensor. Only the following values are supported: 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
[in]gemm_infoGEMM information used to retrieve the original dimensions of the input matrices
Note
lhs_info.k0 must be equal to rhs_info.k0

Definition at line 219 of file CLGEMMMatrixMultiplyReshapedKernel.cpp.

References ARM_COMPUTE_ERROR_ON_NULLPTR, ARM_COMPUTE_ERROR_THROW_ON, arm_compute::get_padding_info(), ITensor::info(), arm_compute::helpers::float_ops::is_zero(), and arm_compute::validate_arguments().

223 {
224  ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
225 
226  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), (input2 != nullptr ? input2->info() : nullptr), output->info(), alpha, beta, lhs_info, rhs_info, gemm_info));
227 
228  auto padding_info = get_padding_info({ input0, output });
229  _input0 = input0;
230  _input1 = input1;
231  _input2 = helpers::float_ops::is_zero(beta) ? nullptr : input2;
232  _output = output;
233  _reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
234  _use_dummy_work_items = preferred_dummy_work_items_support(CLKernelLibrary::get().get_device());
235  _add_bias = _input2 != nullptr;
236  _broadcast_bias = gemm_info.broadcast_bias;
237  _export_to_cl_image = rhs_info.export_to_cl_image;
238  _k = gemm_info.k;
239 
240  // Check if we need to slide the matrix B
241  const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
242  _slide_matrix_b = (_input1->info()->num_dimensions() >= num_dimensions_input0);
243 
244  ElementsProcessed num_elements_processed{};
245 
246  // Configure kernel window
247  auto win_config = validate_and_configure_window(input0->info(), input1->info(), input2 != nullptr ? input2->info() : nullptr, output->info(), lhs_info, rhs_info, gemm_info, num_elements_processed);
248  ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
249  ICLKernel::configure_internal(win_config.second);
250 
251  const bool enable_mixed_precision = gemm_info.fp_mixed_precision;
252  const DataType data_type = input0->info()->data_type();
253 
254  // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding.
255  const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : output->info()->dimension(1);
256 
257  const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
258  const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
259 
260  // Create build options
261  CLBuildOptions build_opts;
262  build_opts.add_option_if(!(helpers::float_ops::is_one(alpha)), "-DALPHA=" + float_to_string_with_full_precision(alpha));
263  build_opts.add_option_if(_input2 != nullptr, "-DBETA=" + float_to_string_with_full_precision(beta));
264  build_opts.add_option_if(helpers::float_ops::is_one(beta), "-DUNIT_BETA");
265  build_opts.add_option_if(_reinterpret_output_as_3d, "-DREINTERPRET_OUTPUT_AS_3D");
266  build_opts.add_option_if(_reinterpret_output_as_3d, "-DHEIGHT_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(1)));
267  build_opts.add_option_if(_reinterpret_output_as_3d, "-DDEPTH_GEMM3D=" + support::cpp11::to_string(output->info()->dimension(2)));
268  build_opts.add_option_if(gemm_info.broadcast_bias, "-DBROADCAST_BIAS");
269  build_opts.add_option_if(!_slide_matrix_b, "-DMATRIX_B_DEPTH=" + support::cpp11::to_string(input1->info()->dimension(2)));
270  build_opts.add_option_if(lhs_info.interleave, "-DLHS_INTERLEAVE");
271  build_opts.add_option_if(rhs_info.interleave, "-DRHS_INTERLEAVE");
272  build_opts.add_option_if(lhs_info.transpose, "-DLHS_TRANSPOSE");
273  build_opts.add_option_if(_use_dummy_work_items, "-DDUMMY_WORK_ITEMS");
274  build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation())));
275  build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a()));
276  build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b()));
277  build_opts.add_option_if(enable_mixed_precision, "-DMIXED_PRECISION");
278  build_opts.add_option_if(rhs_info.export_to_cl_image, "-DOPENCL_IMAGE_SUPPORT");
279  build_opts.add_option("-DRHS_HEIGHT=" + support::cpp11::to_string(input1->info()->dimension(1)));
280  build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
281  build_opts.add_option("-DDATA_TYPE_ACCUMULATOR=" + (enable_mixed_precision ? get_cl_type_from_data_type(DataType::F32) : get_cl_type_from_data_type(data_type)));
282  build_opts.add_option("-DM=" + support::cpp11::to_string(gemm_info.m));
283  build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n));
284  build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k));
285  build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0));
286  build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0));
287  build_opts.add_option("-DK0=" + support::cpp11::to_string(lhs_info.k0));
288  build_opts.add_option("-DV0=" + support::cpp11::to_string(lhs_info.v0));
289  build_opts.add_option("-DH0=" + support::cpp11::to_string(rhs_info.h0));
290  build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0));
291  build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0));
292 
293  std::string kernel_name("gemm_mm_reshaped_");
294  kernel_name += lhs_info.transpose ? "lhs_t_" : "lhs_nt_";
295  kernel_name += rhs_info.transpose ? "rhs_t" : "rhs_nt";
296  kernel_name += rhs_info.export_to_cl_image ? "_texture" : "";
297 
298  // Create kernel
299  _kernel = create_kernel(compile_context, kernel_name, build_opts.options());
300 
301  // Set config_id for enabling LWS tuning
302  _config_id = kernel_name;
303  _config_id += "_";
304  _config_id += (_add_bias ? "add_bias_" : "");
305  _config_id += (_broadcast_bias ? "broadcast_bias_" : "");
306  _config_id += (_reinterpret_output_as_3d ? "3do_" : "");
307  _config_id += (gemm_info.activation_info.enabled() ? "fused_activation_" : "");
308  _config_id += lower_string(string_from_data_type(input0->info()->data_type()));
309  _config_id += "_";
310  _config_id += (enable_mixed_precision ? "mixed_precision_" : "");
311  _config_id += support::cpp11::to_string(output->info()->dimension(1));
312  _config_id += "_";
313  _config_id += support::cpp11::to_string(output->info()->dimension(0));
314  _config_id += "_";
315  _config_id += support::cpp11::to_string(gemm_info.k);
316  _config_id += "_";
317  _config_id += support::cpp11::to_string(output->info()->dimension(2));
318  _config_id += "_";
319  _config_id += support::cpp11::to_string(lhs_info.m0);
320  _config_id += "_";
321  _config_id += support::cpp11::to_string(rhs_info.n0);
322  _config_id += "_";
323  _config_id += support::cpp11::to_string(lhs_info.k0);
324  _config_id += "_";
325  _config_id += support::cpp11::to_string(lhs_info.v0);
326  _config_id += "_";
327  _config_id += support::cpp11::to_string(rhs_info.h0);
328  _config_id += "_";
329  _config_id += support::cpp11::to_string(lhs_info.interleave);
330  _config_id += "_";
331  _config_id += support::cpp11::to_string(rhs_info.interleave);
332 
334 }
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
bool is_one(float a, float epsilon=0.00001f)
Checks if the input floating point number is 1.0f checking if the difference is within a range define...
Definition: float_ops.h:97
bool broadcast_bias
Flag used to broadcast the bias addition.
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
const StringSet & options() const
Gets the current options list set.
unsigned int v0
Number of vertical blocks of size (m0xk0) stored on the same output row.
Definition: Types.h:1977
unsigned int depth_output_gemm3d
Depth of the output tensor in case is reinterpreted as 3D.
bool preferred_dummy_work_items_support(const cl::Device &device)
Helper function to check if "dummy work-items" are preferred to have a power of two NDRange In case d...
Definition: CLHelpers.cpp:361
std::string to_string(T &&value)
Convert integer and float values to string.
virtual DataType data_type() const =0
Data type used for each element of the tensor.
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
const std::string & string_from_activation_func(ActivationLayerInfo::ActivationFunction act)
Translates a given activation function to a string.
Definition: Utils.cpp:163
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
std::string lower_string(const std::string &val)
Lower a given string.
Definition: Utils.cpp:350
bool interleave
True if the v0 (m0xk0) blocks have to be interleaved in the output row.
Definition: Types.h:1979
bool export_to_cl_image
True if the reshaped rhs has to be exported to cl_image.
Definition: Types.h:1995
void add_option(std::string option)
Adds option to the existing build option list.
bool transpose
True if the (m0xk0) block has to be transposed before been stored.
Definition: Types.h:1978
const DataType data_type
Definition: Im2Col.cpp:150
cl::Kernel create_kernel(const CLCompileContext &ctx, const std::string &kernel_name, const std::set< std::string > &build_opts=std::set< std::string >())
Creates an opencl kernel using a compile context.
Definition: CLHelpers.cpp:403
const std::string & string_from_data_type(DataType dt)
Convert a data type identity into a string.
Definition: Utils.cpp:135
std::string float_to_string_with_full_precision(float val)
Create a string with the float in full precision.
Definition: Utils.h:1262
std::string kernel_name
std::string get_cl_type_from_data_type(const DataType &dt)
Translates a tensor data type to the appropriate OpenCL type.
Definition: CLHelpers.cpp:37
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
void add_option_if(bool cond, std::string option)
Adds option if a given condition is true;.
bool has_padding_changed(const std::unordered_map< const ITensorInfo *, PaddingSize > &padding_map)
Check if the previously stored padding info has changed after configuring a kernel.
Definition: Utils.cpp:528
std::unordered_map< const ITensorInfo *, PaddingSize > get_padding_info(std::initializer_list< const ITensorInfo *> infos)
Stores padding information before configuring a kernel.
Definition: Utils.cpp:513
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
unsigned int k
Number of LHS columns or RHS rows.
bool is_zero(float a, float epsilon=0.00001f)
Checks if the input floating point number is 0.0f checking if the difference is within a range define...
Definition: float_ops.h:109
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
unsigned int k0
Number of partial accumulations performed by the matrix multiplication.
Definition: Types.h:1976
unsigned int m0
Number of rows processed by the matrix multiplication.
Definition: Types.h:1975
DataType
Available data types.
Definition: Types.h:77

◆ operator=() [1/2]

Prevent instances of this class from being copied (As this class contains pointers)

◆ operator=() [2/2]

Allow instances of this class to be moved.

◆ run()

void run ( const Window window,
cl::CommandQueue &  queue 
)
overridevirtual

Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue.

Note
The queue is not flushed by this method, and therefore the kernel will not have been executed by the time this method returns.
Parameters
[in]windowRegion on which to execute the kernel. (Must be a valid region of the window returned by window()).
[in,out]queueCommand queue on which to enqueue the kernel.

Reimplemented from ICLKernel.

Definition at line 355 of file CLGEMMMatrixMultiplyReshapedKernel.cpp.

References ARM_COMPUTE_ERROR_ON, ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW, ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL, Window::DimX, Window::DimY, Window::first_slice_window_3D(), ITensor::info(), ITensorInfo::num_dimensions(), Window::set(), arm_compute::test::validation::reference::slice(), ITensorInfo::strides_in_bytes(), and IKernel::window().

356 {
359 
360  if(_input1->info()->num_dimensions() < 3)
361  {
362  // The stride_z for matrix B must be zero if we do not slice
363  ARM_COMPUTE_ERROR_ON(_input1->info()->strides_in_bytes()[3] != 0);
364  }
365 
367  Window slice_matrix_b = slice;
368 
369  slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1));
370  slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1));
371 
372  const unsigned int total_cross_plane_pad = _output->info()->padding().top + _output->info()->padding().bottom;
373 
374  cl::Image2D input1_image2d;
375 
376  if(_export_to_cl_image)
377  {
378  const TensorShape shape2d(_input1->info()->dimension(0) / 4, _input1->info()->dimension(1) * _input1->info()->dimension(2));
379  const size_t image_row_pitch = _input1->info()->strides_in_bytes()[1];
380 
381  input1_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), _input1->cl_buffer(), shape2d, _input1->info()->data_type(), image_row_pitch);
382  }
383 
384  do
385  {
386  Window slice_b = slice;
387  // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2
388  // This scenario can happen when the matrix multiplication is used to perform a convolution operation
389  if(!_slide_matrix_b)
390  {
391  slice_b = slice_matrix_b;
392  }
393 
394  unsigned int idx = 0;
395 
396  // LHS buffer
397  add_2D_tensor_argument(idx, _input0, slice);
398 
399  // RHS buffer or RHS OpenCL image (_export_to_cl_image == true)
400  if(_export_to_cl_image)
401  {
402  _kernel.setArg(idx++, input1_image2d);
403  }
404  else
405  {
406  add_2D_tensor_argument(idx, _input1, slice_b);
407  }
408 
409  // Bias buffer (_add_bias == true)
410  add_2D_tensor_argument_if(_add_bias, idx, _input2, slice);
411 
412  // Output buffer
413  add_2D_tensor_argument(idx, _output, slice);
414 
415  // K dimension (not used if _export_to_cl_image == true)
416  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_k));
417 
418  // LHS stride_z
419  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->info()->strides_in_bytes()[2]));
420 
421  // RHS stride_z (not used if _export_to_cl_image == true)
422  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->info()->strides_in_bytes()[2]));
423 
424  // Bias stride_z (if _add_bias == true)
425  if(_add_bias)
426  {
427  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input2->info()->strides_in_bytes()[2]));
428  }
429 
430  // Output stride_z
431  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->info()->strides_in_bytes()[2]));
432 
433  // Cross-plan padding (if _reinterpret_output_as_3d = true)
434  if(_reinterpret_output_as_3d)
435  {
436  _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(total_cross_plane_pad));
437  }
438 
439  // Dispatch kernel
440  enqueue(queue, *this, slice, lws_hint(), _use_dummy_work_items);
441  }
442  while(window.slide_window_slice_3D(slice));
443 }
unsigned int top
top of the border
Definition: Types.h:375
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
Shape of a tensor.
Definition: TensorShape.h:39
void add_2D_tensor_argument_if(bool cond, unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 2D tensor&#39;s parameters to the object&#39;s kernel&#39;s arguments starting from the index idx ...
Definition: ICLKernel.h:159
void enqueue(IGCKernel &kernel, const Window &window, const gles::NDRange &lws=gles::NDRange(1U, 1U, 1U))
Add the kernel to the command queue with the given window.
Definition: IGCKernel.cpp:41
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
cl::NDRange lws_hint() const
Return the Local-Workgroup-Size hint.
Definition: ICLKernel.h:276
virtual DataType data_type() const =0
Data type used for each element of the tensor.
#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.
Describe one of the image&#39;s dimensions with a start, end and step.
Definition: Window.h:77
unsigned int bottom
bottom of the border
Definition: Types.h:377
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
virtual PaddingSize padding() const =0
Padding of tensor.
bool slide_window_slice_3D(Window &slice) const
Slide the passed 3D window slice.
Definition: Window.h:335
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:941
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
void add_2D_tensor_argument(unsigned int &idx, const ICLTensor *tensor, const Window &window)
Add the passed 2D tensor&#39;s parameters to the object&#39;s kernel&#39;s arguments starting from the index idx...
Definition: ICLKernel.h:148
virtual const cl::Buffer & cl_buffer() const =0
Interface to be implemented by the child class to return a reference to the OpenCL buffer containing ...
cl::Image2D create_image2d_from_buffer(const cl::Context &ctx, const cl::Buffer &buffer, const TensorShape &shape2d, DataType data_type, size_t image_row_pitch)
Create a cl::Image2D object from an OpenCL buffer.
Definition: CLUtils.cpp:29
virtual const Strides & strides_in_bytes() const =0
The strides in bytes for accessing each dimension of the tensor.
Window first_slice_window_3D() const
First 3D slice of the window.
Definition: Window.h:291
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:205
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)

◆ validate()

Status validate ( const ITensorInfo input0,
const ITensorInfo input1,
const ITensorInfo input2,
const ITensorInfo output,
float  alpha,
float  beta,
const GEMMLHSMatrixInfo lhs_info,
const GEMMRHSMatrixInfo rhs_info,
const GEMMKernelInfo gemm_info 
)
static

Static function to check if given info will lead to a valid configuration of CLGEMMMatrixMultiplyReshapedKernel.

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
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:
  1. rhs_info.n0 can only be 4, 8 and 16
  2. rhs_info.k0 can only be 4, 8 and 16
  3. Data type can only be F32
  4. The platform should support the OpenCL cl_khr_image2d_from_buffer extension
  5. The stride Y for the input1 should satisfy the OpenCL pitch alignment requirement
  6. input1 width should be less or equal to (CL_DEVICE_IMAGE2D_MAX_WIDTH * 4)
  7. input1 (height * depth) should be less or equal to CL_DEVICE_IMAGE2D_MAX_HEIGHT
Parameters
[in]input0Input 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
[in]input1Input tensor containing the RHS reshaped matrix. Data type supported: same as input0. The number of dimensions for the RHS matrix must be less or equal than 3
[in]input2Input tensor info containing the bias matrix. Data type supported: same as input0.
[in]outputOutput tensor to store the result of matrix multiplication. Data type supported: same as input0
[in]alphaWeight of the matrix product
[in]betaWeight of the matrix bias
[in]lhs_infoLHS 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
[in]rhs_infoRHS matrix information used for reshaping the input1 tensor. Only the following values are supported: 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
[in]gemm_infoGEMM information used to retrieve the original dimensions of the input matrices
Note
lhs_info.k0 must be equal to rhs_info.k0
Returns
a status

Definition at line 336 of file CLGEMMMatrixMultiplyReshapedKernel.cpp.

References ARM_COMPUTE_RETURN_ON_ERROR, ICloneable< T >::clone(), and arm_compute::validate_arguments().

Referenced by CLGEMMReshapeRHSMatrixKernelManaged::configure(), and arm_compute::test::validation::DATA_TEST_CASE().

339 {
340  ElementsProcessed num_elements_processed{};
341  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input0, input1, input2, output, alpha, beta, lhs_info, rhs_info, gemm_info));
342  ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input0->clone().get(),
343  input1->clone().get(),
344  input2 != nullptr ? input2->clone().get() : nullptr,
345  output->clone().get(),
346  lhs_info,
347  rhs_info,
348  gemm_info,
349  num_elements_processed)
350  .first);
351 
352  return Status{};
353 }
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
Status class.
Definition: Error.h:52
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)

The documentation for this class was generated from the following files: