50 using ElementsProcessed = Steps;
52 inline Status validate_arguments(
const ITensorInfo *src0,
const ITensorInfo *src1,
const ITensorInfo *src2,
const ITensorInfo *
dst,
float beta,
53 bool is_interleaved_transposed,
const GEMMReshapeInfo &reshape_info,
bool fp_mixed_precision)
62 ARM_COMPUTE_RETURN_ERROR_ON_MSG(is_interleaved_transposed && reshape_info.reinterpret_input_as_3d(),
"The input tensor cannot be reinterpreted as 3D if is_interleaved_transposed is true");
63 ARM_COMPUTE_RETURN_ERROR_ON_MSG(src1->num_dimensions() > 2 && reshape_info.reinterpret_input_as_3d(),
"The src1 tensor cannot have more than 2 dimensions if src0 has to be reinterpreted as 3D");
65 && (!reshape_info.broadcast_bias()),
66 "Bias addition only supported with broadcast mode in case the input or dst has to be reinterpreted as 3D");
68 if(!is_interleaved_transposed)
74 const unsigned int m = reshape_info.reinterpret_input_as_3d() ? src0->dimension(1) * src0->dimension(2) : src0->dimension(1);
75 const unsigned int n = src1->dimension(0);
76 const unsigned int src2_dim0 = src2->dimension(0);
77 const unsigned int src2_dim1 = src2->dimension(1);
80 if(reshape_info.broadcast_bias())
92 GEMMRHSMatrixInfo rhs_info;
93 GEMMLHSMatrixInfo lhs_info;
94 const auto m =
static_cast<unsigned int>(reshape_info.m());
95 const auto n =
static_cast<unsigned int>(reshape_info.n());
96 const int k = reshape_info.k();
97 const int mult_transpose1xW_width = reshape_info.mult_transpose1xW_width();
98 const int mult_interleave4x4_height = reshape_info.mult_interleave4x4_height();
99 rhs_info.n0 = max_cl_vector_width / src1->element_size();
101 rhs_info.h0 = mult_transpose1xW_width;
102 rhs_info.interleave =
false;
103 rhs_info.transpose =
false;
106 lhs_info.v0 = mult_interleave4x4_height;
107 lhs_info.interleave =
true;
108 lhs_info.transpose =
true;
110 TensorShape tensor_shape0{ src0->tensor_shape() };
111 tensor_shape0.set(0, k);
112 tensor_shape0.set(1, m);
114 TensorShape tensor_shape1{ src1->tensor_shape() };
115 tensor_shape1.set(0, n);
116 tensor_shape1.set(1, k);
118 const TensorInfo tensor_info0 = src0->clone()->set_tensor_shape(tensor_shape0);
119 const TensorInfo tensor_info1 = src1->clone()->set_tensor_shape(tensor_shape1);
129 const unsigned int src2_dim0 = src2->dimension(0);
130 const unsigned int src2_dim1 = src2->dimension(1);
133 if(reshape_info.broadcast_bias())
144 if(dst->total_size() != 0)
154 inline std::pair<Status, Window> validate_and_configure_window(ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst,
155 float beta,
bool is_interleaved_transposed,
const GEMMReshapeInfo &reshape_info,
GPUTarget gpu_target,
156 ElementsProcessed &num_elements_processed)
159 bool window_changed =
false;
164 unsigned int &num_elems_processed_per_iteration_x = num_elements_processed[0];
165 unsigned int &num_elems_processed_per_iteration_y = num_elements_processed[1];
166 bool reinterpret_input_as_3d = reshape_info.reinterpret_input_as_3d();
167 bool reinterpret_output_as_3d = (reshape_info.depth_output_gemm3d() != 0);
171 if(reinterpret_input_as_3d == reinterpret_output_as_3d)
173 reinterpret_input_as_3d =
false;
174 reinterpret_output_as_3d =
false;
180 TensorInfo tmp_info(*dst);
182 if(reinterpret_output_as_3d)
186 TensorShape tmp_shape(dst->tensor_shape());
187 tmp_shape.collapse(2U, 1U);
188 tmp_info.set_tensor_shape(tmp_shape);
191 if(is_interleaved_transposed)
197 num_elems_processed_per_iteration_x = max_cl_vector_width /
data_size_from_type(data_type);
198 num_elems_processed_per_iteration_y = 4;
200 win =
calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
203 const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
205 const int bias_processed_per_iteration_y = reshape_info.broadcast_bias() ? 1 : num_elems_processed_per_iteration_y;
207 AccessWindowStatic src2_access(src2, 0, 0,
217 num_elems_processed_per_iteration_x = max_cl_vector_width /
data_size_from_type(data_type);
218 num_elems_processed_per_iteration_y = std::min(static_cast<int>(dst->dimension(1)), 4);
224 num_elems_processed_per_iteration_x = (src1->dimension(0) <= 1000 && src0->num_dimensions() == 1) ? 2 : 4;
228 win =
calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
229 win_out =
calculate_max_window(*dst, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y));
230 AccessWindowStatic src0_access(src0, 0, 0, src0->dimension(0), src0->dimension(1));
231 AccessWindowStatic src1_access(src1, 0, 0,
ceil_to_multiple(src1->dimension(0), num_elems_processed_per_iteration_x), src1->dimension(1));
232 AccessWindowStatic dst_access(dst, 0, 0,
238 const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x;
240 AccessWindowStatic src2_access(src2, 0, 0,
256 Window collapsed = win;
257 const unsigned int dimension_to_collapse = std::min(static_cast<unsigned int>(dst->num_dimensions()), 2u);
258 collapsed = win.collapse(win, dimension_to_collapse);
261 return std::make_pair(err, collapsed);
278 is_interleaved_transposed, reshape_info, fp_mixed_precision));
290 _reinterpret_input_as_3d =
false;
304 ElementsProcessed num_elements_processed{};
307 auto win_config = validate_and_configure_window(src0, src1, src2, dst, beta, is_interleaved_transposed, reshape_info,
308 gpu_target, num_elements_processed);
310 ICLKernel::configure_internal(win_config.second);
316 const unsigned int n = dst->
dimension(0);
321 const unsigned int m0 = num_elements_processed.y();
322 const unsigned int n0 = num_elements_processed.x();
325 const unsigned int partial_store_m0 = internal_m % m0;
326 const unsigned int partial_store_n0 = n % n0;
335 build_opts.
add_option_if(_reinterpret_input_as_3d,
"-DREINTERPRET_INPUT_AS_3D");
348 if(is_interleaved_transposed)
371 kernel_name +=
"_acc32";
388 kernel_name =
"gemm_mm_floating_point";
396 kernel_name +=
"_acc32";
413 kernel_name =
"gemm_mm_floating_point";
420 _config_id =
"gemm_";
421 _config_id += (is_interleaved_transposed ?
"reshaped_" :
"");
422 _config_id += (
_add_bias ?
"add_bias_" :
"");
423 _config_id += (reshape_info.
broadcast_bias() ?
"broadcast_bias_" :
"");
424 _config_id += (fp_mixed_precision ?
"fp_mixed_" :
"");
425 _config_id += (_reinterpret_input_as_3d ?
"3di_" :
"");
446 ElementsProcessed num_elements_processed{};
449 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src0, src1, src2, dst, beta, is_interleaved_transposed, reshape_info, fp_mixed_precision));
452 (src2 !=
nullptr) ? src2->
clone().get() :
nullptr,
455 is_interleaved_transposed,
458 num_elements_processed)
477 if(src1->info()->num_dimensions() < 3)
495 const unsigned int total_cross_plane_pad = src0->info()->padding().top + src0->info()->padding().bottom;
496 _kernel.setArg<cl_uint>(idx0,
static_cast<unsigned int>(total_cross_plane_pad));
503 const unsigned int total_cross_plane_pad = dst->info()->padding().top + dst->info()->padding().bottom;
504 _kernel.setArg<cl_uint>(idx0,
static_cast<unsigned int>(total_cross_plane_pad));
514 slice_b = slice_matrix_b;
517 unsigned int idx = 0;
525 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(src0->info()->strides_in_bytes()[2]));
526 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(src1->info()->strides_in_bytes()[2]));
529 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(src2->info()->strides_in_bytes()[2]));
531 _kernel.setArg<cl_uint>(idx++,
static_cast<unsigned int>(dst->info()->strides_in_bytes()[2]));
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)
int mult_interleave4x4_height() const
Multiplication factor for the height of the 4x4 interleaved block.
#define ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(tensor)
const Window & window() const
The maximum window the kernel can be executed on.
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.
cl::NDRange lws_hint() const
Return the Local-Workgroup-Size hint.
void set_lws_hint(const cl::NDRange &lws_hint)
Set the Local-Workgroup-Size hint.
int mult_transpose1xW_width() const
Multiplication factor for the width of the 1xW transposed block.
float a() const
Get the alpha value.
GEMM reshape information class.
#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.
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...
GPUTarget get_arch_from_target(GPUTarget target)
Helper function to get the GPU arch.
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.
Activation Layer Information class.
Copyright (c) 2017-2021 Arm Limited.
1 channel, 1 F16 per channel
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
ClGemmMatrixMultiplyKernel()
void add_option(std::string option)
Adds option to the existing build option list.
bool _reinterpret_input_as_3d
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
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.
void configure(const ClCompileContext &compile_context, ITensorInfo *src0, ITensorInfo *src1, ITensorInfo *src2, ITensorInfo *dst, float alpha, float beta=0.f, bool is_interleaved_transposed=true, const GEMMReshapeInfo &reshape_info=GEMMReshapeInfo(), bool fp_mixed_precision=false, const ActivationLayerInfo &activation_info=ActivationLayerInfo())
Initialise the kernel's input, output and alpha.
TensorShape compute_lhs_reshaped_shape(const ITensorInfo &a, const GEMMLHSMatrixInfo &lhs_info, bool reinterpret_input_as_3d=false)
Calculate the Left Hand Side matrix reshaped shape.
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.
static Status validate(const ITensorInfo *src0, const ITensorInfo *src1, const ITensorInfo *src2, const ITensorInfo *dst, float alpha, float beta, bool is_interleaved_transposed, const GEMMReshapeInfo &reshape_info, GPUTarget gpu_target, bool fp_mixed_precision=false, const ActivationLayerInfo &activation_info=ActivationLayerInfo())
Static function to check if given info will lead to a valid configuration.
GPUTarget get_target() const
Get the targeted GPU architecture.
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...
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
size_t data_size_from_type(DataType data_type)
The size in bytes of the data type.
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.
#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.
bool broadcast_bias() const
Flag which specifies whether to broadcast the shape of the bias tensor.
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...
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.
GPUTarget
Available GPU Targets.
#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.
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...
int depth_output_gemm3d() const
Depth (third dimension) of the output tensor to be used with the GEMM3D kernel.
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(...)
bool _reinterpret_output_as_3d
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.
DataType
Available data types.
bool reinterpret_input_as_3d() const
Flag which specifies if the input tensor has to be reinterpreted as 3D.
Describe a multidimensional execution window.
bool is_data_type_float(DataType dt)
Check if a given data type is of floating point type.
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
SimpleTensor< T > slice(const SimpleTensor< T > &src, Coordinates starts, Coordinates ends)