arm_compute::opencl::kernels::gemm Namespace Reference

Data Structures
class	CLGEMMConfigArray
	Basic container for the OpenCL GEMM configuration functions. More...
class	ClGemmDefaultConfigNativeBifrost
	Bifrost based OpenCL GEMMNative configuration. More...
class	ClGemmDefaultConfigNativeMidgard
	Midgard based OpenCL GEMMNative configuration. More...
class	ClGemmDefaultConfigNativeValhall
	Valhall based OpenCL GEMMNative configuration. More...
class	ClGemmDefaultConfigReshapedBifrost
	Bifrost based OpenCL GEMMReshaped configuration. More...
class	ClGemmDefaultConfigReshapedRhsOnlyBifrost
	Bifrost based OpenCL GEMMReshapedOnlyRHS configuration. More...
class	ClGemmDefaultConfigReshapedRhsOnlyValhall
	Valhall based OpenCL GEMMReshapedOnlyRHS configuration. More...
class	ClGemmDefaultConfigReshapedValhall
	Valhall based OpenCL GEMMReshaped configuration. More...
class	ClGemmNativeKernelConfigurationFactory
	CLGEMMNative factory class. More...
class	ClGemmReshapedKernelConfigurationFactory
	CLGEMMReshaped factory class. More...
class	ClGemmReshapedOnlyRhsKernelConfigurationFactory
	CLGEMMReshapedOnlyRHS factory class. More...
class	IClGemmKernelConfig
	Basic interface for the GEMM kernel configuration. More...

Functions
std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo >	configure_lhs_rhs_info (unsigned int m, unsigned int n, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool lhs_interleave, bool rhs_interleave, bool lhs_transpose, bool rhs_transpose, bool export_to_cl_image=false)
	Configure GEMMLHSMatrixInfo and GEMMRHSMatrixInfo. More...
std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo >	select_lhs_rhs_info (std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo > info_img, std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo > info_buf, unsigned int n, unsigned int k, unsigned int b, DataType data_type)
	Select GEMMLHSMatrixInfo and GEMMRHSMatrixInfo. More...
void	update_padding_for_cl_image (ITensorInfo *tensor)
	Update padding required to export the OpenCL buffer to OpenCL image2d. More...
Status	validate_image2d_support_on_rhs (const ITensorInfo &tensor_reshaped_info, const GEMMRHSMatrixInfo &rhs_info)
	Utility function to validate the image2d OpenCL object support on the RHS reshaped matrix. More...

Function Documentation

configure_lhs_rhs_info()

std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo > configure_lhs_rhs_info	(	unsigned int	m,
		unsigned int	n,
		unsigned int	m0,
		unsigned int	n0,
		unsigned int	k0,
		unsigned int	v0,
		unsigned int	h0,
		bool	lhs_interleave,
		bool	rhs_interleave,
		bool	lhs_transpose,
		bool	rhs_transpose,
		bool	export_to_cl_image = false
	)

Configure GEMMLHSMatrixInfo and GEMMRHSMatrixInfo.

Parameters

[in]	m	Number of rows (M) in the LHS matrix not reshaped
[in]	n	Number of columns (N) in the RHS matrix not reshaped
[in]	m0	Number of rows processed by each thread/work-item
[in]	n0	Number of columns processed by each thread/work-item
[in]	k0	Number of inner accumulation performed by each thread/work-item
[in]	v0	Number of vertical blocks of size (m0xk0) stored on the same output row
[in]	h0	Number of horizontal blocks of size (k0xn0) stored on the same output row
[in]	lhs_interleave	True if the v0 (m0xk0) blocks have to be interleaved in the output row
[in]	rhs_interleave	True if the h0 (k0xn0) blocks have to be interleaved in the output row
[in]	lhs_transpose	True if the (m0xk0) block has to be transposed before been stored
[in]	rhs_transpose	True if the (k0xn0) block has to be transposed before been stored
[in]	export_to_cl_image	(Optional) True if the RHS reshaped matrix has to be exported to cl_image

Returns

GEMMLHSMatrixInfo and GEMMRHSMatrixInfo

Definition at line 41 of file ClGemmHelpers.cpp.

References ARM_COMPUTE_ERROR_ON, arm_compute::test::validation::lhs_info, and arm_compute::test::validation::rhs_info.

Referenced by ClGemmDefaultConfigNativeBifrost::configure(), ClGemmDefaultConfigNativeMidgard::configure(), ClGemmDefaultConfigReshapedBifrost::configure(), ClGemmDefaultConfigReshapedValhall::configure(), ClGemmDefaultConfigReshapedRhsOnlyValhall::configure(), ClGemmDefaultConfigReshapedRhsOnlyBifrost::configure(), ClGemmDefaultConfigNativeValhall::configure(), arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_config_native(), arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_config_reshaped(), and arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs().

{
    ARM_COMPUTE_ERROR_ON(m0 == 0 || n0 == 0);
    v0 = std::max(std::min(static_cast<int>(m / m0), static_cast<int>(v0)), static_cast<int>(1));
    h0 = std::max(std::min(static_cast<int>(n / n0), static_cast<int>(h0)), static_cast<int>(1));

    const GEMMLHSMatrixInfo lhs_info(m0, k0, v0, lhs_transpose, lhs_interleave);
    const GEMMRHSMatrixInfo rhs_info(n0, k0, h0, rhs_transpose, rhs_interleave, export_to_cl_image);

    return std::make_pair(lhs_info, rhs_info);
}

select_lhs_rhs_info()

std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo > select_lhs_rhs_info	(	std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo >	info_img,
		std::pair< GEMMLHSMatrixInfo, GEMMRHSMatrixInfo >	info_buf,
		unsigned int	n,
		unsigned int	k,
		unsigned int	b,
		DataType	data_type
	)

Select GEMMLHSMatrixInfo and GEMMRHSMatrixInfo.

This function accepts two pairs of GEMMLHSMatrixInfo/GEMMRHSMatrixInfo where only the first is with cl_image2d support, and selects the valid one validating the GEMMRHSMatrixInfo. If the validation passes, the functions will return the first GEMMLHSMatrixInfo/GEMMRHSMatrixInfo pair with cl_image2d support.

Parameters

[in]	info_img	GEMMLHSMatrixInfo/GEMMRHSMatrixInfo with cl_image2d support
[in]	info_buf	GEMMLHSMatrixInfo/GEMMRHSMatrixInfo to fall-back if cl_image2d cannot be used
[in]	n	Number of columns (N) in the RHS matrix not reshaped
[in]	k	Number of rows (K) in the RHS matrix not reshaped
[in]	b	Batch size
[in]	data_type	Data type

Returns

GEMMLHSMatrixInfo and GEMMRHSMatrixInfo

Definition at line 54 of file ClGemmHelpers.cpp.

References arm_compute::misc::shape_calculator::compute_rhs_reshaped_shape(), arm_compute::test::validation::shape, and validate_image2d_support_on_rhs().

Referenced by ClGemmDefaultConfigReshapedBifrost::configure(), ClGemmDefaultConfigReshapedRhsOnlyValhall::configure(), ClGemmDefaultConfigReshapedRhsOnlyBifrost::configure(), and ClGemmDefaultConfigReshapedValhall::configure().

{
    const TensorInfo  tensor_rhs_info(TensorShape(n, k, b), 1, data_type);
    const TensorShape shape = misc::shape_calculator::compute_rhs_reshaped_shape(tensor_rhs_info, info_img.second);
    const TensorInfo  tensor_reshaped_info(shape, 1, data_type);

    if(bool(validate_image2d_support_on_rhs(tensor_reshaped_info, info_img.second)))
    {
        return info_img;
    }
    else
    {
        return info_buf;
    }
}

update_padding_for_cl_image()

void update_padding_for_cl_image

(

ITensorInfo *

tensor

)

Update padding required to export the OpenCL buffer to OpenCL image2d.

Parameters

[in,out]

tensor

ITensorInfo of the tensor required to be exported to OpenCL image2d

Definition at line 72 of file ClGemmHelpers.cpp.

References ARM_COMPUTE_ERROR_ON_MSG, ITensorInfo::element_size(), ITensorInfo::extend_padding(), CLKernelLibrary::get(), arm_compute::get_cl_image_pitch_alignment(), ITensorInfo::padding(), BorderSize::right, and ITensorInfo::strides_in_bytes().

Referenced by ClDirectConv2dKernel::configure(), CLDepthwiseConvolutionLayerNativeKernel::configure(), and ClDirectConvolutionKernelComponent::generate_build_options().

{
    constexpr unsigned int num_floats_per_pixel = 4;

    const unsigned int stride_y_in_elements = tensor->strides_in_bytes()[1] / tensor->element_size();
    const unsigned int pixel_alignment      = get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device());

    ARM_COMPUTE_ERROR_ON_MSG(pixel_alignment == 0, "Cannot retrieve cl_image pitch alignment");
    if(pixel_alignment == 0)
    {
        return;
    }

    const unsigned int row_pitch_alignment = pixel_alignment * num_floats_per_pixel;
    const unsigned int round_up_width      = ((stride_y_in_elements + row_pitch_alignment - 1) / row_pitch_alignment) * row_pitch_alignment;
    const unsigned int padding             = round_up_width - stride_y_in_elements;

    tensor->extend_padding(PaddingSize(0, tensor->padding().right + padding, 0, 0));
}

validate_image2d_support_on_rhs()

Status validate_image2d_support_on_rhs	(	const ITensorInfo &	tensor_reshaped_info,
		const GEMMRHSMatrixInfo &	rhs_info
	)

Utility function to validate the image2d OpenCL object support on the RHS reshaped matrix.

Parameters

[in]	tensor_reshaped_info	TensorInfo for the RHS reshaped matrix
[in]	rhs_info	GEMMRHSMatrixInfo

Returns

Status reporting if we can use the image2d OpenCL object on the RHS reshaped matrix

Definition at line 92 of file ClGemmHelpers.cpp.

References ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN, ARM_COMPUTE_RETURN_ERROR_ON_MSG, GEMMRHSMatrixInfo::export_to_cl_image, arm_compute::F16, arm_compute::F32, CLKernelLibrary::get(), arm_compute::get_cl_image_pitch_alignment(), CLKernelLibrary::get_device(), arm_compute::image2d_from_buffer_supported(), GEMMRHSMatrixInfo::k0, GEMMRHSMatrixInfo::n0, and ITensorInfo::tensor_shape().

Referenced by ClGemmDefaultConfigReshapedBifrost::configure(), ClGemmDefaultConfigReshapedRhsOnlyBifrost::configure(), and select_lhs_rhs_info().

{
    if(rhs_info.export_to_cl_image)
    {
        ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.n0 == 2) || (rhs_info.n0 == 3), "Export to cl_image only supported with n0 = 4, 8 or 16");
        ARM_COMPUTE_RETURN_ERROR_ON_MSG((rhs_info.k0 == 2) || (rhs_info.k0 == 3), "Export to cl_image only supported with k0 = 4, 8 or 16");
        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(&tensor_reshaped_info, DataType::F32, DataType::F16);
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(!image2d_from_buffer_supported(CLKernelLibrary::get().get_device()), "The extension cl_khr_image2d_from_buffer is not supported on the target platform");
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device()) == 0, "Impossible to retrieve the cl_image pitch alignment");

        // Check the width and height of the output tensor.
        // Since we cannot create a 3d image from a buffer, the third dimension is collapsed on the second dimension
        const size_t max_image_w = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>();
        const size_t max_image_h = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>();

        ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.tensor_shape()[0] > max_image_w * 4, "Not supported width for cl_image");
        ARM_COMPUTE_RETURN_ERROR_ON_MSG(tensor_reshaped_info.tensor_shape()[1] * tensor_reshaped_info.tensor_shape()[2] > max_image_h, "Not supported height for cl_image");
    }

    return Status{};
}