51 using namespace misc::shape_calculator;
55 using ElementsProcessed = Steps;
57 Status validate_arguments(
const ITensorInfo *src0,
const ITensorInfo *src1,
const ITensorInfo *
dst,
const GEMMKernelInfo &gemm_info,
58 const ITensorInfo *vector_sum_col,
const ITensorInfo *vector_sum_row,
const ITensorInfo *bias,
59 const ITensorInfo *output_multipliers,
const ITensorInfo *output_shifts)
74 const GEMMRHSMatrixInfo rhs_info = gemm_info.rhs_info;
75 const GEMMLHSMatrixInfo lhs_info = gemm_info.lhs_info;
76 const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
78 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");
83 const int m = gemm_info.m;
84 const int n = gemm_info.n;
85 const int k = gemm_info.k;
87 TensorShape tensor_shape1{ src1->tensor_shape() };
88 tensor_shape1.set(0, n);
89 tensor_shape1.set(1, k);
91 const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1);
95 if(gemm_info.reinterpret_input_as_3d)
105 const TensorShape expected_dst_shape =
compute_mm_shape(*src0, *src1, gemm_info);
106 if(dst->total_size() != 0)
108 const TensorInfo tensor_info_dst = dst->clone()->set_tensor_shape(expected_dst_shape);
128 "Only GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT is supported");
134 if(gemm_info.a_offset != 0)
141 if(gemm_info.b_offset != 0)
146 const bool reinterpret_as_3d = expected_dst_shape.num_dimensions() > 1 && expected_dst_shape.y() != vector_sum_row->tensor_shape().x();
152 if(expected_dst_shape.num_dimensions() > 1)
154 const unsigned int dst_batch_idx = reinterpret_as_3d ? 3 : 2;
156 TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape();
157 vector_sum_row_shape.collapse_from(1);
158 TensorShape collapsed_dst_shape(expected_dst_shape);
159 collapsed_dst_shape.collapse_from(dst_batch_idx);
162 "vector_sum_row must have the same number of batches of dst tensor");
164 if(gemm_info.a_offset != 0)
166 TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape();
167 vector_sum_col_shape.collapse_from(1);
170 "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");
175 if(dst->total_size() != 0)
181 if(output_multipliers !=
nullptr && output_shifts !=
nullptr)
187 if(output_stage.is_quantized_per_channel)
197 std::pair<Status, Window> validate_and_configure_window(
const ITensorInfo *src0,
const ITensorInfo *src1, ITensorInfo *dst,
const GEMMKernelInfo &gemm_info,
198 ITensorInfo *vector_sum_col,
const ITensorInfo *vector_sum_row, ITensorInfo *bias,
199 ITensorInfo *output_multipliers, ITensorInfo *output_shifts, ElementsProcessed &num_elements_processed)
201 const GEMMLowpOutputStageInfo output_stage = gemm_info.output_stage;
203 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
204 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
205 bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d;
206 bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d != 0);
210 bool window_changed =
false;
214 if(reinterpret_input_as_3d == reinterpret_output_as_3d)
216 reinterpret_output_as_3d =
false;
220 const TensorShape expected_dst_shape =
compute_mm_shape(*src0, *src1, gemm_info);
223 auto_init_if_empty(*dst, src0->clone()->set_tensor_shape(expected_dst_shape).set_data_type(output_stage.output_data_type));
230 TensorInfo tmp_info(*dst);
232 if(reinterpret_output_as_3d)
236 TensorShape tmp_shape(dst->tensor_shape());
237 tmp_shape.collapse(2U, 1U);
238 tmp_info.set_tensor_shape(tmp_shape);
242 num_elems_processed_per_iteration_x = gemm_info.rhs_info.n0;
243 num_elems_processed_per_iteration_y = gemm_info.lhs_info.m0;
245 win =
calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
246 win_out =
calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
250 if(gemm_info.a_offset != 0)
252 AccessWindowHorizontal vector_sum_col_access(vector_sum_col, 0, num_elems_processed_per_iteration_x);
260 AccessWindowHorizontal bias_access(bias, 0, num_elems_processed_per_iteration_x);
264 if(output_multipliers !=
nullptr && output_stage.is_quantized_per_channel)
266 AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_processed_per_iteration_x);
267 AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_processed_per_iteration_x);
274 Window collapsed = win;
275 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
276 collapsed = win.collapse(win, dimension_to_collapse);
279 return std::make_pair(err, collapsed);
294 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, dst, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts));
300 const int32_t a_offset = gemm_info.
a_offset;
301 const int32_t b_offset = gemm_info.
b_offset;
310 if(_reinterpret_input_as_3d == _reinterpret_output_as_3d)
312 _reinterpret_input_as_3d =
false;
313 _reinterpret_output_as_3d =
false;
318 _slide_matrix_b = (src1->
num_dimensions() >= num_dimensions_src0);
320 ElementsProcessed num_elements_processed{};
323 auto win_config = validate_and_configure_window(src0, src1, dst, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts, num_elements_processed);
325 ICLKernel::configure_internal(win_config.second);
330 const unsigned int internal_m = _reinterpret_output_as_3d ? gemm_info.
m : dst->
dimension(1);
334 const unsigned int internal_m0 = std::min(internal_m, lhs_info.
m0);
337 const unsigned int partial_store_m0 = internal_m % internal_m0;
338 const unsigned int partial_store_n0 = gemm_info.
n % rhs_info.n0;
342 build_opts.
add_option_if(_reinterpret_input_as_3d,
"-DREINTERPRET_INPUT_AS_3D");
343 build_opts.
add_option_if(_reinterpret_output_as_3d,
"-DREINTERPRET_OUTPUT_AS_3D");
347 build_opts.
add_option_if(rhs_info.interleave,
"-DRHS_INTERLEAVE");
348 build_opts.
add_option_if(_use_dummy_work_items,
"-DDUMMY_WORK_ITEMS");
361 std::string
kernel_name(
"gemmlowp_mm_reshaped_only_rhs_");
362 kernel_name += rhs_info.transpose ?
"t" :
"nt";
366 kernel_name +=
"_fused_output_stage_fixedpoint";
367 _fuse_output_stage =
true;
369 if(a_offset != 0 && vector_sum_col !=
nullptr)
381 build_opts.
add_option_if(_is_quantized_per_channel,
"-DPER_CHANNEL_QUANTIZATION");
401 _config_id += (_reinterpret_input_as_3d ?
"3di_" :
"");
402 _config_id += (_reinterpret_output_as_3d ?
"3do_" :
"");
427 ElementsProcessed num_elements_processed{};
428 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, dst, gemm_info, vector_sum_col, vector_sum_row, bias, output_multipliers, output_shifts));
433 vector_sum_col !=
nullptr ? vector_sum_col->
clone().get() :
nullptr,
434 vector_sum_row !=
nullptr ? vector_sum_row->
clone().get() :
nullptr,
435 bias !=
nullptr ? bias->
clone().get() :
nullptr,
436 output_multipliers !=
nullptr ? output_multipliers->
clone().get() :
nullptr,
437 output_shifts !=
nullptr ? output_shifts->
clone().get() :
nullptr,
438 num_elements_processed)
458 if(src1->info()->num_dimensions() < 3)
470 if(_reinterpret_input_as_3d)
473 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3;
474 const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->info()->padding().bottom;
475 _kernel.setArg<cl_uint>(idx0,
static_cast<unsigned int>(total_cross_plane_pad));
478 if(_reinterpret_output_as_3d)
481 const unsigned int idx0 = 3 * num_arguments_per_2D_tensor() + 3 + (_reinterpret_input_as_3d ? 1 : 0);
482 const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom;
483 _kernel.setArg<cl_uint>(idx0,
static_cast<unsigned int>(total_cross_plane_pad));
508 slice_b = slice_matrix_b;
511 unsigned int idx = 0;
512 add_2D_tensor_argument(idx, src0, slice);
513 add_2D_tensor_argument(idx, src1, slice_b);
514 add_2D_tensor_argument(idx, dst, slice);
515 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
516 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
517 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
518 if(_reinterpret_input_as_3d)
524 if(_reinterpret_output_as_3d)
530 if(_fuse_output_stage)
532 add_2D_tensor_argument_if((vector_sum_col !=
nullptr), idx, vector_sum_col, win_vector_sum_col);
533 add_2D_tensor_argument_if((vector_sum_row !=
nullptr), idx, vector_sum_row, win_vector_sum_row);
534 add_1D_tensor_argument_if((bias !=
nullptr), idx, bias, biases_slice);
535 add_1D_tensor_argument_if(_is_quantized_per_channel, idx, output_multipliers, biases_slice);
536 add_1D_tensor_argument_if(_is_quantized_per_channel, idx, output_shifts, biases_slice);
538 enqueue(queue, *
this, slice, lws_hint(), _use_dummy_work_items);
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.
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.
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...
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.
void configure(const CLCompileContext &compile_context, const ITensorInfo *src0, const ITensorInfo *src1, ITensorInfo *dst, const GEMMKernelInfo &gemm_info, ITensorInfo *vector_sum_col=nullptr, const ITensorInfo *vector_sum_row=nullptr, ITensorInfo *bias=nullptr, ITensorInfo *output_multipliers=nullptr, ITensorInfo *output_shifts=nullptr)
Initialise the kernel's source and destination.
#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.
Manages all the OpenCL kernels compilation and caching, provides accessors for the OpenCL Context...
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...
static Status validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *dst, 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.
#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.
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
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 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.
ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel()
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.
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.
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.
ITensor * get_tensor(int id)
Get tensor of a given id from the pac.
GEMMRHSMatrixInfo rhs_info
RHS matrix information used for reshaping the RHS matrix.
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...
#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.
unsigned int num_dimensions() const
Returns the effective dimensionality of the tensor.
#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.
#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
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.
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)