48 using ElementsProcessed = Steps;
50 Status
validate_arguments(
const ITensorInfo *input0,
const ITensorInfo *input1,
const ITensorInfo *input2,
const ITensorInfo *output,
float alpha,
float beta,
const GEMMLHSMatrixInfo &lhs_info,
51 const GEMMRHSMatrixInfo &rhs_info,
52 const GEMMKernelInfo &gemm_info)
67 && (!gemm_info.broadcast_bias),
68 "Bias addition only supported with broadcast mode in case the input or output has to be reinterpreted as 3D");
72 const unsigned int m = gemm_info.m;
73 const unsigned int n = gemm_info.n;
74 const unsigned int k = gemm_info.k;
76 TensorShape tensor_shape1{ input1->tensor_shape() };
77 tensor_shape1.set(0, n);
78 tensor_shape1.set(1, k);
82 const unsigned int input2_dim0 = input2->dimension(0);
83 const unsigned int input2_dim1 = input2->dimension(1);
86 if(gemm_info.broadcast_bias)
96 const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
98 const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(
compute_rhs_reshaped_shape(tensor_info1, rhs_info));
101 if(gemm_info.reinterpret_input_as_3d)
111 if(output->total_size() != 0)
113 const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(
compute_mm_shape(*input0, *input1, gemm_info));
121 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output,
const GEMMLHSMatrixInfo &lhs_info,
122 const GEMMRHSMatrixInfo &rhs_info,
123 const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
125 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
126 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
127 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
128 bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
132 bool window_changed =
false;
137 if((reinterpret_input_as_3d == reinterpret_output_as_3d) && gemm_info.has_pad_y)
139 reinterpret_output_as_3d =
false;
145 TensorInfo tmp_info(*output);
147 if(reinterpret_output_as_3d)
151 TensorShape tmp_shape(output->tensor_shape());
152 tmp_shape.collapse(2U, 1U);
153 tmp_info.set_tensor_shape(tmp_shape);
157 num_elems_processed_per_iteration_x = rhs_info.n0;
158 num_elems_processed_per_iteration_y = lhs_info.m0;
160 win =
calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
161 win_out =
calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
163 if(input2 !=
nullptr)
165 const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
167 AccessWindowStatic input2_access(input2, 0, 0,
169 input2->dimension(1));
176 Window collapsed = win;
177 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
178 collapsed = win.collapse(win, dimension_to_collapse);
181 return std::make_pair(err, collapsed);
186 : _input0(nullptr), _input1(nullptr), _input2(nullptr), _output(nullptr), _slide_matrix_b(true), _reinterpret_input_as_3d(false), _reinterpret_output_as_3d(false), _use_dummy_work_items(false),
187 _add_bias(false), _broadcast_bias(false), _export_to_cl_image(false), _has_pad_y(false)
215 _add_bias = _input2 !=
nullptr;
217 _export_to_cl_image = rhs_info.export_to_cl_image;
224 if((_reinterpret_input_as_3d == _reinterpret_output_as_3d) && _has_pad_y)
226 _reinterpret_input_as_3d =
false;
227 _reinterpret_output_as_3d =
false;
234 ElementsProcessed num_elements_processed{};
237 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);
239 ICLKernel::configure_internal(win_config.second);
244 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : output->
info()->
dimension(1);
252 const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
255 const unsigned int partial_store_m0 = internal_m % internal_m0;
256 const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
264 build_opts.
add_option_if(gemm_info.broadcast_bias,
"-DBROADCAST_BIAS");
266 build_opts.
add_option_if(rhs_info.interleave,
"-DRHS_INTERLEAVE");
267 build_opts.
add_option_if(_use_dummy_work_items,
"-DDUMMY_WORK_ITEMS");
268 build_opts.
add_option_if(rhs_info.export_to_cl_image,
"-DOPENCL_IMAGE_SUPPORT");
284 build_opts.
add_option_if(_reinterpret_input_as_3d,
"-DREINTERPRET_INPUT_AS_3D");
285 build_opts.
add_option_if(_reinterpret_output_as_3d,
"-DREINTERPRET_OUTPUT_AS_3D");
290 std::string
kernel_name(
"gemm_mm_reshaped_only_rhs_");
291 kernel_name += rhs_info.transpose ?
"t" :
"nt";
292 kernel_name += rhs_info.export_to_cl_image ?
"_texture" :
"";
300 _config_id += (_has_pad_y ?
"" :
"no_pad_y_");
301 _config_id += (_add_bias ?
"add_bias_" :
"");
302 _config_id += (_broadcast_bias ?
"broadcast_bias_" :
"");
303 _config_id += (_reinterpret_input_as_3d ?
"3di_" :
"");
304 _config_id += (_reinterpret_output_as_3d ?
"3do_" :
"");
305 _config_id += (gemm_info.activation_info.enabled() ?
"fused_activation_" :
"");
333 ElementsProcessed num_elements_processed{};
336 input1->
clone().get(),
337 input2 !=
nullptr ? input2->
clone().get() :
nullptr,
338 output->
clone().get(),
342 num_elements_processed)
359 const size_t lhs_idx_batch_size = _reinterpret_input_as_3d && !_has_pad_y ? 3u : 2u;
360 const size_t rhs_idx_batch_size = 2u;
361 const size_t bia_idx_batch_size = 2u;
362 const size_t out_idx_batch_size = _reinterpret_output_as_3d && !_has_pad_y ? 3u : 2u;
375 ARM_COMPUTE_ERROR_ON(!_has_pad_y && ((total_cross_plane_pad_lhs != 0) || (total_cross_plane_pad_out != 0)));
377 cl::Image2D input1_image2d;
379 if(_export_to_cl_image)
394 slice_b = slice_matrix_b;
397 unsigned int idx = 0;
403 if(_export_to_cl_image)
405 _kernel.setArg(idx++, input1_image2d);
419 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(_input0->
info()->
strides_in_bytes()[lhs_idx_batch_size]));
422 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(_input1->
info()->
strides_in_bytes()[rhs_idx_batch_size]));
427 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(_input2->
info()->
strides_in_bytes()[bia_idx_batch_size]));
431 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(_output->
info()->
strides_in_bytes()[out_idx_batch_size]));
434 if(_reinterpret_input_as_3d && _has_pad_y)
436 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(total_cross_plane_pad_lhs));
440 if(_reinterpret_output_as_3d && _has_pad_y)
442 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(total_cross_plane_pad_out));
unsigned int top
top of the border
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...
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
bool broadcast_bias
Flag used to broadcast the bias addition.
#define ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(tensor)
const Window & window() const
The maximum window the kernel can be executed on.
Descriptor used by the GEMM kernels.
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 ...
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.
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.
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 CLGEMMMatrixMultiplyResh...
unsigned int depth_output_gemm3d
Depth of the output tensor in case is reinterpreted as 3D.
cl::NDRange lws_hint() const
Return the Local-Workgroup-Size hint.
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...
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
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.
TensorShape compute_mm_shape(const ITensorInfo &input0, const ITensorInfo &input1, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info)
Calculate the matrix multiplication output shape of two tensors.
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
CLGEMMMatrixMultiplyReshapedOnlyRHSKernel()
Default Constructor.
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
const std::string & string_from_activation_func(ActivationLayerInfo::ActivationFunction act)
Translates a given activation function to a string.
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
GEMM LHS (Left Hand Side) matrix information.
Store the tensor's metadata.
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Describe one of the image's dimensions with a start, end and step.
unsigned int bottom
bottom of the border
std::string lower_string(const std::string &val)
Lower a given string.
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Copyright (c) 2017-2021 Arm Limited.
1 channel, 1 F16 per channel
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
void add_option(std::string option)
Adds option to the existing build option list.
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.
const std::string & string_from_data_type(DataType dt)
Convert a data type identity into a string.
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
bool update_window_and_padding(Window &win, Ts &&... patterns)
Update window and padding size for each of the access patterns.
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
GEMM RHS (Right Hand Side) matrix information.
std::string float_to_string_with_full_precision(float val)
Create a string with the float in full precision.
auto ceil_to_multiple(S value, T divisor) -> decltype(((value+divisor - 1)/divisor) *divisor)
Computes the smallest number larger or equal to value that is a multiple of divisor.
std::string get_cl_type_from_data_type(const DataType &dt)
Translates a tensor data type to the appropriate OpenCL type.
TensorShape compute_rhs_reshaped_shape(const ITensorInfo &a, const GEMMRHSMatrixInfo &rhs_info)
Calculate the Right Hand Side matrix reshaped shape.
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
bool reinterpret_input_as_3d
Flag used to reinterpret the input as 3D.
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
void add_option_if(bool cond, std::string option)
Adds option if a given condition is true;.
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
virtual PaddingSize padding() const =0
Padding of tensor.
bool slide_window_slice_3D(Window &slice) const
Slide the passed 3D window slice.
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
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.
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.
bool has_pad_y
Flag used to indicate if the input/output tensors have internal pad on the y direction.
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
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...
Interface for OpenCL tensor.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
#define ARM_COMPUTE_CREATE_ERROR(error_code, msg)
Creates an error with a given message.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
std::unordered_map< const ITensorInfo *, PaddingSize > get_padding_info(std::initializer_list< const ITensorInfo *> infos)
Stores padding information before configuring a kernel.
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
void run(const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue...
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...
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
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 ...
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
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.
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.
Describe a multidimensional execution window.
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)