52 using ElementsProcessed = Steps;
54 const auto post_op_utils = experimental::PostOpCLKernelUtils(
73 Status
validate_arguments(
const ITensorInfo *src0,
const ITensorInfo *src1,
const ITensorInfo *src2,
const ITensorInfo *
dst,
float alpha,
float beta,
const GEMMLHSMatrixInfo &
lhs_info,
88 && (!gemm_info.broadcast_bias),
89 "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D");
94 const unsigned int m = gemm_info.m;
95 const unsigned int n = gemm_info.n;
96 const unsigned int k = gemm_info.k;
105 if(gemm_info.reinterpret_input_as_3d)
116 const unsigned int src2_dim0 = src2->dimension(0);
117 const unsigned int src2_dim1 = src2->dimension(1);
120 if(gemm_info.broadcast_bias)
130 if(dst->total_size() != 0)
141 std::pair<Status, Window>
validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst,
const GEMMLHSMatrixInfo &lhs_info,
142 const GEMMRHSMatrixInfo &rhs_info,
143 const GEMMKernelInfo &gemm_info, ElementsProcessed &num_elements_processed)
145 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
146 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
147 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
148 bool reinterpret_output_as_3d = gemm_info.depth_output_gemm3d != 0;
152 bool window_changed =
false;
156 if(reinterpret_input_as_3d == reinterpret_output_as_3d)
158 reinterpret_output_as_3d =
false;
164 TensorInfo tmp_info(*dst);
166 if(reinterpret_output_as_3d)
170 TensorShape tmp_shape(dst->tensor_shape());
171 tmp_shape.collapse(2U, 1U);
172 tmp_info.set_tensor_shape(tmp_shape);
176 num_elems_processed_per_iteration_x = rhs_info.n0;
177 num_elems_processed_per_iteration_y = lhs_info.m0;
179 win =
calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
180 win_out =
calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
182 AccessWindowStatic src0_access(src0, 0, 0,
185 AccessWindowStatic src1_access(src1, 0, 0,
188 AccessWindowStatic dst_access(dst, 0, 0,
194 const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
196 AccessWindowStatic src2_access(src2, 0, 0,
211 Window collapsed = win;
212 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
213 collapsed = win.collapse(win, dimension_to_collapse);
216 return std::make_pair(err, collapsed);
241 _add_bias = src2 !=
nullptr;
242 _num_post_op_args = gemm_info.
post_ops.total_num_arguments();
246 if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
248 _reinterpret_input_as_3d =
false;
249 _reinterpret_output_as_3d =
false;
254 _slide_matrix_b = (src1->
num_dimensions() >= num_dimensions_src0);
256 ElementsProcessed num_elements_processed{};
259 auto win_config =
validate_and_configure_window(src0, src1, src2 !=
nullptr ? src2 :
nullptr, dst, lhs_info, rhs_info, gemm_info, num_elements_processed);
261 IClKernel::configure_internal(win_config.second);
266 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.
m : dst->
dimension(1);
268 const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? dst->
dimension(1) : src0->
dimension(1);
269 const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? dst->
dimension(2) : src0->
dimension(2);
272 const unsigned int partial_store_m0 = internal_m % lhs_info.
m0;
273 const unsigned int partial_store_n0 = gemm_info.
n % rhs_info.
n0;
277 const unsigned int internal_m0 = std::min(internal_m, lhs_info.
m0);
289 build_opts.
add_option_if(_reinterpret_input_as_3d,
"-DREINTERPRET_INPUT_AS_3D");
290 build_opts.
add_option_if(_reinterpret_output_as_3d,
"-DREINTERPRET_OUTPUT_AS_3D");
294 build_opts.
add_option_if(_use_dummy_work_items,
"-DDUMMY_WORK_ITEMS");
303 post_op_utils.set_post_ops_cl_build_options(build_opts, gemm_info.
post_ops);
313 post_op_utils.set_post_ops_cl_kernel_name(kernel_name, gemm_info.
post_ops);
324 _config_id += (_add_bias ?
"add_bias_" :
"");
325 _config_id += (gemm_info.
broadcast_bias ?
"broadcast_bias_" :
"");
326 _config_id += (_reinterpret_input_as_3d ?
"3di_" :
"");
327 _config_id += (_reinterpret_output_as_3d ?
"3do_" :
"");
352 ElementsProcessed num_elements_processed{};
356 src2 !=
nullptr ? src2->
clone().get() :
nullptr,
361 num_elements_processed)
380 if(src1->info()->num_dimensions() < 3)
392 if(_reinterpret_input_as_3d)
404 const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->info()->padding().bottom;
405 _kernel.setArg<cl_uint>(idx0,
static_cast<unsigned int>(total_cross_plane_pad));
408 if(_reinterpret_output_as_3d)
420 const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom;
421 _kernel.setArg<cl_uint>(idx0,
static_cast<unsigned int>(total_cross_plane_pad));
431 slice_b = slice_matrix_b;
434 unsigned int idx = 0;
443 for(
size_t i = 0; i < _num_post_op_args; ++i)
448 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
449 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
452 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(src2->info()->strides_in_bytes()[2]));
454 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
456 for(
size_t i = 0; i < _num_post_op_args; ++i)
459 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(post_op_arg->info()->strides_in_bytes()[2]));
463 _kernel.setArg<cl_int>(idx++, _m);
464 _kernel.setArg<cl_int>(idx++, _n);
465 _kernel.setArg<cl_int>(idx++, _k);
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...
static Status validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, 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.
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.
Descriptor used by the GEMM kernels.
bool enabled() const
Check if initialised.
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
void enqueue(cl::CommandQueue &queue, ICLKernel &kernel, const Window &window, const cl::NDRange &lws_hint=CLKernelLibrary::get().default_ndrange(), bool use_dummy_work_items=false)
Add the kernel to the command queue with the given window.
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...
float a() const
Get the alpha value.
#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.
Manages all the OpenCL kernels compilation and caching, provides accessors for the OpenCL Context...
ActivationLayerInfo activation_info
Activation function to perform after the matrix multiplication.
std::string lower_string(const std::string &val)
Lower a given string.
Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const PadStrideInfo &conv_info)
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Copyright (c) 2017-2023 Arm Limited.
1 channel, 1 F16 per channel
TensorType get_post_op_arg_type(size_t index)
Get post op argument TensorType from post op argument index in a flattened, ordered post op argument ...
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
void add_option(std::string option)
Adds option to the existing build option list.
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
unsigned int k0
Number of partial accumulations performed by the matrix multiplication.
unsigned int m
Number of LHS rows.
std::string upper_string(const std::string &val)
Raise a given string to upper case.
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.
unsigned int n
Number of RHS columns.
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.
void configure(const ClCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const GEMMKernelInfo &gemm_info)
Initialise the kernel's input and dst.
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.
unsigned int n0
Number of columns processed by the matrix multiplication.
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.
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.
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.
ClGemmMatrixMultiplyNativeKernel()
#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...
std::pair< Status, Window > validate_and_configure_window(ITensorInfo *src, ITensorInfo *dst)
ITensor * get_tensor(int id)
Get tensor of a given id from the pac.
#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.
Wrapper to configure the Khronos OpenCL C++ header.
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...
experimental::PostOpList< ITensorInfo * > post_ops
(EXPERIMENTAL_POST_OPS) Specifies a list of post ops to be fused after the main op.
#define ARM_COMPUTE_RETURN_ERROR_ON_MSG(cond, msg)
If the condition is true, an error is returned.
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
unsigned int m0
Number of rows processed by the matrix multiplication.
ActivationFunction activation() const
Get the type of activation function.
float b() const
Get the beta value.
Window first_slice_window_3D() const
First 3D slice of the window.
void run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue...
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)