51 using ElementsProcessed = Steps;
53 Status
validate_arguments(
const ITensorInfo *input0,
const ITensorInfo *input1,
const ITensorInfo *input2,
const ITensorInfo *output,
float alpha,
float beta,
const GEMMLHSMatrixInfo &lhs_info,
54 const GEMMRHSMatrixInfo &rhs_info,
55 const GEMMKernelInfo &gemm_info)
68 && (!gemm_info.broadcast_bias),
69 "Bias addition only supported with broadcast mode in case the input or output has to be reinterpreted as 3D");
73 const unsigned int m = gemm_info.m;
74 const unsigned int n = gemm_info.n;
75 const unsigned int k = gemm_info.k;
84 if(gemm_info.reinterpret_input_as_3d)
95 const unsigned int input2_dim0 = input2->dimension(0);
96 const unsigned int input2_dim1 = input2->dimension(1);
99 if(gemm_info.broadcast_bias)
109 if(output->total_size() != 0)
111 const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(
compute_mm_shape(*input0, *input1, gemm_info));
119 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output,
const GEMMLHSMatrixInfo &lhs_info,
120 const GEMMRHSMatrixInfo &rhs_info,
121 const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
123 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
124 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
125 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
126 bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
130 bool window_changed =
false;
134 if(reinterpret_input_as_3d == reinterpret_output_as_3d)
136 reinterpret_output_as_3d =
false;
142 TensorInfo tmp_info(*output);
144 if(reinterpret_output_as_3d)
148 TensorShape tmp_shape(output->tensor_shape());
149 tmp_shape.collapse(2U, 1U);
150 tmp_info.set_tensor_shape(tmp_shape);
154 num_elems_processed_per_iteration_x = rhs_info.n0;
155 num_elems_processed_per_iteration_y = lhs_info.m0;
157 win =
calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
158 win_out =
calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
160 AccessWindowStatic input0_access(input0, 0, 0,
161 input0->dimension(0),
162 input0->dimension(1));
163 AccessWindowStatic input1_access(input1, 0, 0,
164 ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x),
165 input1->dimension(1));
166 AccessWindowStatic output_access(output, 0, 0,
167 output->dimension(0),
168 output->dimension(1));
170 if(input2 !=
nullptr)
172 const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
174 AccessWindowStatic input2_access(input2, 0, 0,
176 input2->dimension(1));
187 output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape()));
191 Window collapsed = win;
192 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
193 collapsed = win.collapse(win, dimension_to_collapse);
196 return std::make_pair(err, collapsed);
201 : _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),
202 _add_bias(false), _broadcast_bias(false)
230 _add_bias = _input2 !=
nullptr;
235 if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
237 _reinterpret_input_as_3d =
false;
238 _reinterpret_output_as_3d =
false;
242 const unsigned int num_dimensions_input0 = _input0->info()->num_dimensions();
245 ElementsProcessed num_elements_processed{};
248 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);
250 ICLKernel::configure_internal(win_config.second);
255 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.m : output->
info()->
dimension(1);
261 const unsigned int partial_store_m0 = internal_m % lhs_info.m0;
262 const unsigned int partial_store_n0 = gemm_info.n % rhs_info.n0;
266 const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0);
274 build_opts.
add_option_if(gemm_info.broadcast_bias,
"-DBROADCAST_BIAS");
275 build_opts.
add_option_if(_reinterpret_input_as_3d,
"-DREINTERPRET_INPUT_AS_3D");
276 build_opts.
add_option_if(_reinterpret_output_as_3d,
"-DREINTERPRET_OUTPUT_AS_3D");
280 build_opts.
add_option_if(_use_dummy_work_items,
"-DDUMMY_WORK_ITEMS");
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_" :
"");
329 ElementsProcessed num_elements_processed{};
332 input1->
clone().get(),
333 input2 !=
nullptr ? input2->
clone().get() :
nullptr,
334 output->
clone().get(),
338 num_elements_processed)
361 if(_reinterpret_input_as_3d)
374 _kernel.setArg<cl_uint>(idx0,
static_cast<unsigned int>(total_cross_plane_pad));
377 if(_reinterpret_output_as_3d)
390 _kernel.setArg<cl_uint>(idx0,
static_cast<unsigned int>(total_cross_plane_pad));
400 slice_b = slice_matrix_b;
403 unsigned int idx = 0;
411 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(_input0->
info()->
strides_in_bytes()[2]));
412 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(_input1->
info()->
strides_in_bytes()[2]));
415 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(_input2->
info()->
strides_in_bytes()[2]));
417 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(_output->
info()->
strides_in_bytes()[2]));
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.
const Window & window() const
The maximum window the kernel can be executed on.
CLGEMMMatrixMultiplyNativeKernel()
Default Constructor.
Descriptor used by the GEMM kernels.
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.
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.
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.
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.
#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 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 CLGEMMMatrixMultiplyNati...
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.
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.
static constexpr unsigned int num_arguments_per_2D_tensor()
Returns the number of arguments enqueued per 2D tensor object.
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.
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.
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.
Manages all the OpenCL kernels compilation and caching, provides accessors for the OpenCL Context...
#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)
Wrapper to configure the Khronos OpenCL C++ header.
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.
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
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)
void run(const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue...
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)