Openvx讀書筆記

OpenVX讀書筆記

summary

	high level	low level
ovx	strong typed eg VX_API_ENTRY vx_node VX_API_CALL vxColorConvertNode(vx_graph graph, vx_image input, vx_image output);	weak typed, eg OpenVX.dll!agoCreateNode(_vx_graph * graph, int kernel_id)
tbb	strong typed make_edge(tbb::flow::output_port<1>(gpu_slm_split_n), tbb::flow::input_port<1>(gpu_slm_mat_mult_n)) tbb::flow::function_node< validation_args_type > mat_validation_n(g, tbb::flow::unlimited, { // Get references to matrixes const tbb::flow::gfx_buffer& GPU_SLM_MAT = std::get<0>(result); const tbb::flow::gfx_buffer& CPU_SLM_MAT = std::get<1>(result); const tbb::flow::gfx_buffer& CPU_NAIVE_MAT = std::get<2>(result); // Verify results // Check that slm algorithm produces correct results on CPU: validate_mat("matrix multiply: 'SLM' CPU vs. CPU", SIZE_Y, SIZE_X, CPU_SLM_MAT.data(), CPU_NAIVE_MAT.data()); // Verify Gen results: validate_mat("matrix multiply: SLM Gen vs. CPU", SIZE_Y, SIZE_X, GPU_SLM_MAT.data(), CPU_NAIVE_MAT.data()); });	Not sure
G-API	strong typed	TODO

// ovx: \vis_bep_12\C\Users\test\Downloads\amdovx-core-0.9-beta2\amdovx-core-0.9-beta2 // tbb: C:\Users\test\Downloads\tbb2017_20170604oss_win\tbb2017_20170604oss

How to register Kernel

Define a enum

VX_KERNEL_COLOR_CONVERT = VX_KERNEL_BASE(VX_ID_KHRONOS, VX_LIBRARY_KHR_BASE) + 0x1,

Registrtion

OVX_KERNEL_ENTRY( VX_KERNEL_COLOR_CONVERT         , ColorConvert, "color_convert",             AIN_AOUT,             ATYPE_II           , false ), 

the parameters meaning

#define OVX_KERNEL_ENTRY(kernel_id,name,kname,argCfg,argType,validRectReset) \

#define ATYPE_II                               { VX_TYPE_IMAGE, VX_TYPE_IMAGE }

AIN_AOUT: 1 in, 1 out
ATYPE_II: 2 image types

Implement “DramaDivideNode” operation, it is used to select the best suited for this PC architecture

int agoDramaDivideNode(AgoNodeList * nodeList, AgoNode * anode)
{
	// save parameter list
	vx_uint32 paramCount = anode->paramCount;
	AgoData * paramList[AGO_MAX_PARAMS]; memcpy(paramList, anode->paramList, sizeof(paramList));
	// divide the node depending on the type
	int status = -1;
	switch (anode->akernel->id)
	{
		case VX_KERNEL_COLOR_CONVERT:
			status = agoDramaDivideColorConvertNode(nodeList, anode);
			break;

the function is called by optimize function

>	OpenVX.dll!agoCreateNode(_vx_graph * graph, int kernel_id) Line 2699	C++
 	OpenVX.dll!agoDramaDivideAppend(AgoNodeList * nodeList, _vx_node * anode, int new_kernel_id, _vx_reference * * paramList, unsigned int paramCount) Line 37	C++
 	OpenVX.dll!agoDramaDivideAppend(AgoNodeList * nodeList, _vx_node * anode, int new_kernel_id) Line 56	C++
 	OpenVX.dll!agoDramaDivideColorConvertNode(AgoNodeList * nodeList, _vx_node * anode) Line 244	C++
 	OpenVX.dll!agoDramaDivideNode(AgoNodeList * nodeList, _vx_node * anode) Line 1818	C++
 	OpenVX.dll!agoOptimizeDramaDivide(_vx_graph * agraph) Line 1962	C++
 	OpenVX.dll!agoOptimizeDrama(_vx_graph * agraph) Line 522	C++
 	OpenVX.dll!agoOptimizeGraph(_vx_graph * agraph) Line 209	C++
 	OpenVX.dll!vxVerifyGraph(_vx_graph * graph) Line 2450	C++
 	runvx.exe!CVxEngine::ProcessGraph(std::vector<char const *,std::allocator<char const *> > * graphNameList, unsigned __int64 beginIndex) Line 285	C++

How to schedule graph?

What optimization is done in optimize()?

Choose the best

Published on Dec 14, 2019 in categories vision

上一篇 G Api notes
下一篇 Visionworks openvx