50 using ElementsProcessed = Steps;
52 Status
validate_arguments(
const ITensorInfo *input0,
const ITensorInfo *input1,
const ITensorInfo *output,
const GEMMKernelInfo &gemm_info,
53 const ITensorInfo *vector_sum_col,
const ITensorInfo *vector_sum_row,
const ITensorInfo *bias,
54 const ITensorInfo *output_multipliers,
const ITensorInfo *output_shifts)
69 const GEMMRHSMatrixInfo rhs_info = gemm_info.rhs_info;
70 const GEMMLHSMatrixInfo lhs_info = gemm_info.lhs_info;
71 const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
73 ARM_COMPUTE_RETURN_ERROR_ON_MSG((((rhs_info.k0 & (rhs_info.k0 - 1)) && rhs_info.k0 != 3) || (rhs_info.k0 > 16)),
"Only 2,3,4,8,16 are supported for k0");
78 const int m = gemm_info.m;
79 const int n = gemm_info.n;
80 const int k = gemm_info.k;
82 TensorShape tensor_shape1{ input1->tensor_shape() };
83 tensor_shape1.set(0, n);
84 tensor_shape1.set(1, k);
86 const TensorInfo tensor_info1 = input1->clone()->set_tensor_shape(tensor_shape1);
87 const TensorInfo tensor_info_reshaped1 = input1->clone()->set_tensor_shape(
compute_rhs_reshaped_shape(tensor_info1, rhs_info));
90 if(gemm_info.reinterpret_input_as_3d)
100 const TensorShape expected_output_shape =
compute_mm_shape(*input0, *input1, gemm_info);
101 if(output->total_size() != 0)
103 const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(expected_output_shape);
123 "Only GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT is supported");
129 if(gemm_info.a_offset != 0)
136 if(gemm_info.b_offset != 0)
141 const bool reinterpret_as_3d = expected_output_shape.num_dimensions() > 1 && expected_output_shape.y() != vector_sum_row->tensor_shape().x();
144 ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (expected_output_shape[1] * expected_output_shape[2]));
147 if(expected_output_shape.num_dimensions() > 1)
149 const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2;
151 TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
152 vector_sum_row_shape.collapse_from(1);
153 TensorShape collapsed_output_shape(expected_output_shape);
154 collapsed_output_shape.collapse_from(output_batch_idx);
157 "vector_sum_row must have the same number of batches of output tensor");
159 if(gemm_info.a_offset != 0)
161 TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape();
162 vector_sum_col_shape.collapse_from(1);
165 "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1");
170 if(output->total_size() != 0)
176 if(output_multipliers !=
nullptr && output_shifts !=
nullptr)
182 if(output_stage.is_quantized_per_channel)
192 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input0, ITensorInfo *input1, ITensorInfo *output,
const GEMMKernelInfo &gemm_info,
193 ITensorInfo *vector_sum_col, ITensorInfo *vector_sum_row, ITensorInfo *bias,
194 ITensorInfo *output_multipliers, ITensorInfo *output_shifts, ElementsProcessed &num_elements_processed)
196 const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
198 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
199 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
200 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
201 bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d != 0);
205 bool window_changed =
false;
209 if(reinterpret_input_as_3d == reinterpret_output_as_3d)
211 reinterpret_output_as_3d =
false;
215 const TensorShape expected_output_shape =
compute_mm_shape(*input0, *input1, gemm_info);
218 auto_init_if_empty(*output, input0->clone()->set_tensor_shape(expected_output_shape).set_data_type(output_stage.output_data_type));
225 TensorInfo tmp_info(*output);
227 if(reinterpret_output_as_3d)
231 TensorShape tmp_shape(output->tensor_shape());
232 tmp_shape.collapse(2U, 1U);
233 tmp_info.set_tensor_shape(tmp_shape);
237 num_elems_processed_per_iteration_x = gemm_info.rhs_info.n0;
238 num_elems_processed_per_iteration_y = gemm_info.lhs_info.m0;
240 win =
calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
241 win_out =
calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
245 if(gemm_info.a_offset != 0)
247 AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration_x);
255 AccessWindowHorizontal bias_access(bias, 0, num_elems_processed_per_iteration_x);
259 if(output_multipliers !=
nullptr && output_multipliers->dimension(0) > 1)
261 AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_processed_per_iteration_x);
262 AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_processed_per_iteration_x);
269 Window collapsed = win;
270 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(output->num_dimensions()), 2u);
271 collapsed = win.collapse(win, dimension_to_collapse);
274 return std::make_pair(err, collapsed);
282 _vector_sum_col(nullptr),
283 _vector_sum_row(nullptr),
285 _output_multipliers(nullptr),
286 _output_shifts(nullptr),
287 _slide_matrix_b(true),
288 _reinterpret_input_as_3d(false),
289 _reinterpret_output_as_3d(false),
290 _use_dummy_work_items(false),
291 _is_quantized_per_channel(false),
292 _fuse_output_stage(false)
300 configure(
CLKernelLibrary::get().get_compile_context(), input0, input1, output, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts);
313 vector_sum_col !=
nullptr ? vector_sum_col->
info() :
nullptr,
314 vector_sum_row !=
nullptr ? vector_sum_row->
info() :
nullptr,
315 bias !=
nullptr ? bias->
info() :
nullptr,
316 output_multipliers !=
nullptr ? output_multipliers->
info() :
nullptr,
317 output_shifts !=
nullptr ? output_shifts->
info() :
nullptr));
319 auto padding_info =
get_padding_info({ input0, input1, output, vector_sum_row });
323 const int32_t a_offset = gemm_info.
a_offset;
324 const int32_t b_offset = gemm_info.
b_offset;
329 _vector_sum_col = vector_sum_col;
330 _vector_sum_row = vector_sum_row;
332 _output_multipliers = output_multipliers;
333 _output_shifts = output_shifts;
341 if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
343 _reinterpret_input_as_3d =
false;
344 _reinterpret_output_as_3d =
false;
351 ElementsProcessed num_elements_processed{};
354 auto win_config = validate_and_configure_window(input0->
info(),
358 vector_sum_col !=
nullptr ? vector_sum_col->
info() :
nullptr,
359 vector_sum_row !=
nullptr ? vector_sum_row->
info() :
nullptr,
360 bias !=
nullptr ? bias->
info() :
nullptr,
361 output_multipliers !=
nullptr ? output_multipliers->
info() :
nullptr,
362 output_shifts !=
nullptr ? output_shifts->
info() :
nullptr,
363 num_elements_processed);
365 ICLKernel::configure_internal(win_config.second);
370 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.
m : output->
info()->
dimension(1);
374 const unsigned int internal_m0 = std::min(internal_m, lhs_info.
m0);
377 const unsigned int partial_store_m0 = internal_m % internal_m0;
378 const unsigned int partial_store_n0 = gemm_info.
n % rhs_info.n0;
382 build_opts.
add_option_if(_reinterpret_input_as_3d,
"-DREINTERPRET_INPUT_AS_3D");
383 build_opts.
add_option_if(_reinterpret_output_as_3d,
"-DREINTERPRET_OUTPUT_AS_3D");
387 build_opts.
add_option_if(rhs_info.interleave,
"-DRHS_INTERLEAVE");
388 build_opts.
add_option_if(_use_dummy_work_items,
"-DDUMMY_WORK_ITEMS");
401 std::string
kernel_name(
"gemmlowp_mm_reshaped_only_rhs_");
407 _fuse_output_stage =
true;
409 if(a_offset != 0 && vector_sum_col !=
nullptr)
421 build_opts.
add_option_if(_is_quantized_per_channel,
"-DPER_CHANNEL_QUANTIZATION");
441 _config_id += (_reinterpret_input_as_3d ?
"3di_" :
"");
442 _config_id += (_reinterpret_output_as_3d ?
"3do_" :
"");
467 ElementsProcessed num_elements_processed{};
470 input1->
clone().get(),
471 output->
clone().get(),
473 vector_sum_col !=
nullptr ? vector_sum_col->
clone().get() :
nullptr,
474 vector_sum_row !=
nullptr ? vector_sum_row->
clone().get() :
nullptr,
475 bias !=
nullptr ? bias->
clone().get() :
nullptr,
476 output_multipliers !=
nullptr ? output_multipliers->
clone().get() :
nullptr,
477 output_shifts !=
nullptr ? output_shifts->
clone().get() :
nullptr,
478 num_elements_processed)
501 if(_reinterpret_input_as_3d)
506 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
509 if(_reinterpret_output_as_3d)
514 _kernel.setArg<cl_uint>(idx0, static_cast<unsigned int>(total_cross_plane_pad));
539 slice_b = slice_matrix_b;
542 unsigned int idx = 0;
546 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input0->
info()->
strides_in_bytes()[2]));
547 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_input1->
info()->
strides_in_bytes()[2]));
548 _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(_output->
info()->
strides_in_bytes()[2]));
549 if(_reinterpret_input_as_3d)
555 if(_reinterpret_output_as_3d)
561 if(_fuse_output_stage)
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 ...
unsigned int top
top of the border
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
Class describing the value of a pixel for any image format.
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
const Window & window() const
The maximum window the kernel can be executed on.
Quantize using a fixed point multiplication.
bool dot8_supported(const cl::Device &device)
Helper function to check whether the cl_arm_integer_dot_product_int8 extension is supported.
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 ...
std::unordered_map< const ITensorInfo *, PaddingSize > get_padding_info(std::initializer_list< const ITensorInfo * > infos)
Stores padding information before configuring a kernel.
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...
CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel()
Default Constructor.
std::string get_cl_dot8_acc_type_from_data_type(const DataType &dt)
Translates a tensor data type to the appropriate OpenCL dot8 accumulator type.
#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.
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
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
int32_t gemmlowp_offset
GEMMLowp output stage offset used for quantizing to QASYMM8.
int32_t gemmlowp_max_bound
GEMMLowp max value used to saturate down the output result before converting back to QASYMM8.
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
GEMMLowpOutputStageType type
GEMMLowp output stage type.
GEMMLHSMatrixInfo lhs_info
LHS matrix information used to retrieve the number of rows processed by each thread.
Copyright (c) 2017-2021 Arm Limited.
bool is_quantized_per_channel
GEMMLowp quantized per-channel flag.
std::vector< int32_t > gemmlowp_shifts
GEMMLowp output stage multiplier used for quantizing to QASYMM8.
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
1 channel, 1 S32 per channel
void add_option(std::string option)
Adds option to the existing build option list.
unsigned int m
Number of LHS rows.
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.
static Status validate(const ITensorInfo *input0, const ITensorInfo *input1, const ITensorInfo *output, const GEMMKernelInfo &gemm_info, const ITensorInfo *vector_sum_col=nullptr, const ITensorInfo *vector_sum_row=nullptr, const ITensorInfo *bias=nullptr, const ITensorInfo *output_multipliers=nullptr, const ITensorInfo *output_shifts=nullptr)
Static function to check if given info will lead to a valid configuration of CLGEMMLowpMatrixMultiply...
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.
int32_t b_offset
Offset to be added to each element of the matrix B.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
quantized, asymmetric fixed-point 8-bit number unsigned
std::vector< int32_t > gemmlowp_multipliers
GEMMLowp output stage multiplier used for quantizing to QASYMM8.
std::string get_cl_type_from_data_type(const DataType &dt)
Translates a tensor data type to the appropriate OpenCL type.
GEMMLowpOutputStageInfo output_stage
GEMMLowp output stage information.
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.
GEMMLowp output stage info.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
Quantize using a floating point multiplication.
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.
Quantize using an integer multiplication.
#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.
quantized, symmetric fixed-point 8-bit number
quantized, symmetric per channel fixed-point 8-bit number
int32_t a_offset
Offset to be added to each element of the matrix A.
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.
GEMMRHSMatrixInfo rhs_info
RHS matrix information used for reshaping the RHS matrix.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
#define ARM_COMPUTE_CREATE_ERROR(error_code, msg)
Creates an error with a given message.
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Manages all the OpenCL kernels compilation and caching, provides accessors for the OpenCL Context.
unsigned int num_dimensions() const
Returns the effective dimensionality of the tensor.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
void run(const Window &window, cl::CommandQueue &queue) override
Enqueue the OpenCL kernel to process the given window on the passed OpenCL command queue.
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
Wrapper to configure the Khronos OpenCL C++ header.
unsigned int k
Number of LHS columns or RHS rows.
#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.
quantized, asymmetric fixed-point 8-bit number signed
virtual const Strides & strides_in_bytes() const =0
The strides in bytes for accessing each dimension of the tensor.
int32_t gemmlowp_min_bound
GEMMLowp min value used to saturate down the output result before converting back to QASYMM8.
Window first_slice_window_3D() const
First 3D slice of the window.
void configure(const ICLTensor *input0, const ICLTensor *input1, ICLTensor *output, const GEMMKernelInfo &gemm_info, const ICLTensor *vector_sum_col=nullptr, const ICLTensor *vector_sum_row=nullptr, const ICLTensor *bias=nullptr, const ICLTensor *output_multipliers=nullptr, const ICLTensor *output_shifts=nullptr)
Initialise the kernel's input and output.
std::tuple< PixelValue, PixelValue > get_min_max(DataType dt)
Compute the mininum and maximum values a data type can take.
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)