cmake working (3c5e4a1e) · Commits · Elphel / tile_processor_gpu

CMakeLists.txt

0 → 100644

+41 −0

Original line number	Original line	Diff line number	Diff line
			cmake_minimum_required(VERSION 3.22 FATAL_ERROR)

			project(tile-processor-gpu LANGUAGES CUDA CXX)

			# get submodules
			find_package(Git QUIET)
			if(GIT_FOUND AND EXISTS "${PROJECT_SOURCE_DIR}/.git")
			option(GIT_SUBMODULE "Check submodules during build" ON)
			if(GIT_SUBMODULE)
			message(STATUS "Submodule update")
			execute_process(
			COMMAND ${GIT_EXECUTABLE} submodule update --init --recursive
			WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
			RESULT_VARIABLE GIT_SUBMOD_RESULT)
			if(NOT GIT_SUBMOD_RESULT EQUAL "0")
			message(FATAL_ERROR "git submodule update --init --recursive failed with ${GIT_SUBMOD_RESULT}, please checkout submodules")
			endif()
			endif()
			endif()


			if(NOT EXISTS "${PROJECT_SOURCE_DIR}/external/cuda-samples/README.md")
			message(FATAL_ERROR "The submodule cuda-samples was not downloaded! GIT_SUBMODULE was turned off or failed. Please update submodules and try again.")
			endif()

			# create the executable
			add_executable(${PROJECT_NAME}
			src/test_tp.cu
			src/dtt8x8.cu
			src/dtt8x8.h
			src/geometry_correction.cu
			src/geometry_correction.h
			src/test_tp.cu
			src/TileProcessor.cuh
			src/TileProcessor.h
			src/tp_defines.h)
			include_directories(external/cuda-samples/Common)
			set_target_properties(${PROJECT_NAME}
			PROPERTIES CUDA_SEPARABLE_COMPILATION ON)

src/test_tp.cu

+153 −152

Original line number	Original line	Diff line number	Diff line
	@@ -49,6 +49,7 @@
	#include <iterator>		#include <iterator>
	#include <vector>		#include <vector>

			// src includes
	#include "TileProcessor.cuh"		#include "TileProcessor.cuh"
	#include "dtt8x8.h"		#include "dtt8x8.h"
	#include "geometry_correction.h"		#include "geometry_correction.h"
	@@ -542,131 +543,131 @@ int main(int argc, char **argv) {

	#if TEST_LWIR		#if TEST_LWIR
	const char *kernel_file[] = {		const char *kernel_file[] = {
	"clt/aux_chn0_transposed.kernel",		"../../clt/aux_chn0_transposed.kernel",
	"clt/aux_chn1_transposed.kernel",		"../../clt/aux_chn1_transposed.kernel",
	"clt/aux_chn2_transposed.kernel",		"../../clt/aux_chn2_transposed.kernel",
	"clt/aux_chn3_transposed.kernel",		"../../clt/aux_chn3_transposed.kernel",
	"clt/aux_chn4_transposed.kernel",		"../../clt/aux_chn4_transposed.kernel",
	"clt/aux_chn5_transposed.kernel",		"../../clt/aux_chn5_transposed.kernel",
	"clt/aux_chn6_transposed.kernel",		"../../clt/aux_chn6_transposed.kernel",
	"clt/aux_chn7_transposed.kernel",		"../../clt/aux_chn7_transposed.kernel",
	"clt/aux_chn8_transposed.kernel",		"../../clt/aux_chn8_transposed.kernel",
	"clt/aux_chn9_transposed.kernel",		"../../clt/aux_chn9_transposed.kernel",
	"clt/aux_chn10_transposed.kernel",		"../../clt/aux_chn10_transposed.kernel",
	"clt/aux_chn11_transposed.kernel",		"../../clt/aux_chn11_transposed.kernel",
	"clt/aux_chn12_transposed.kernel",		"../../clt/aux_chn12_transposed.kernel",
	"clt/aux_chn13_transposed.kernel",		"../../clt/aux_chn13_transposed.kernel",
	"clt/aux_chn14_transposed.kernel",		"../../clt/aux_chn14_transposed.kernel",
	"clt/aux_chn15_transposed.kernel"};		"../../clt/aux_chn15_transposed.kernel"};

	const char *kernel_offs_file[] = {		const char *kernel_offs_file[] = {
	"clt/aux_chn0_transposed.kernel_offsets",		"../../clt/aux_chn0_transposed.kernel_offsets",
	"clt/aux_chn1_transposed.kernel_offsets",		"../../clt/aux_chn1_transposed.kernel_offsets",
	"clt/aux_chn2_transposed.kernel_offsets",		"../../clt/aux_chn2_transposed.kernel_offsets",
	"clt/aux_chn3_transposed.kernel_offsets",		"../../clt/aux_chn3_transposed.kernel_offsets",
	"clt/aux_chn4_transposed.kernel_offsets",		"../../clt/aux_chn4_transposed.kernel_offsets",
	"clt/aux_chn5_transposed.kernel_offsets",		"../../clt/aux_chn5_transposed.kernel_offsets",
	"clt/aux_chn6_transposed.kernel_offsets",		"../../clt/aux_chn6_transposed.kernel_offsets",
	"clt/aux_chn7_transposed.kernel_offsets",		"../../clt/aux_chn7_transposed.kernel_offsets",
	"clt/aux_chn8_transposed.kernel_offsets",		"../../clt/aux_chn8_transposed.kernel_offsets",
	"clt/aux_chn9_transposed.kernel_offsets",		"../../clt/aux_chn9_transposed.kernel_offsets",
	"clt/aux_chn10_transposed.kernel_offsets",		"../../clt/aux_chn10_transposed.kernel_offsets",
	"clt/aux_chn11_transposed.kernel_offsets",		"../../clt/aux_chn11_transposed.kernel_offsets",
	"clt/aux_chn12_transposed.kernel_offsets",		"../../clt/aux_chn12_transposed.kernel_offsets",
	"clt/aux_chn13_transposed.kernel_offsets",		"../../clt/aux_chn13_transposed.kernel_offsets",
	"clt/aux_chn14_transposed.kernel_offsets",		"../../clt/aux_chn14_transposed.kernel_offsets",
	"clt/aux_chn15_transposed.kernel_offsets"};		"../../clt/aux_chn15_transposed.kernel_offsets"};

	const char *image_files[] = {		const char *image_files[] = {
	"clt/aux_chn0.bayer",		"../../clt/aux_chn0.bayer",
	"clt/aux_chn1.bayer",		"../../clt/aux_chn1.bayer",
	"clt/aux_chn2.bayer",		"../../clt/aux_chn2.bayer",
	"clt/aux_chn3.bayer",		"../../clt/aux_chn3.bayer",
	"clt/aux_chn4.bayer",		"../../clt/aux_chn4.bayer",
	"clt/aux_chn5.bayer",		"../../clt/aux_chn5.bayer",
	"clt/aux_chn6.bayer",		"../../clt/aux_chn6.bayer",
	"clt/aux_chn7.bayer",		"../../clt/aux_chn7.bayer",
	"clt/aux_chn8.bayer",		"../../clt/aux_chn8.bayer",
	"clt/aux_chn9.bayer",		"../../clt/aux_chn9.bayer",
	"clt/aux_chn10.bayer",		"../../clt/aux_chn10.bayer",
	"clt/aux_chn11.bayer",		"../../clt/aux_chn11.bayer",
	"clt/aux_chn12.bayer",		"../../clt/aux_chn12.bayer",
	"clt/aux_chn13.bayer",		"../../clt/aux_chn13.bayer",
	"clt/aux_chn14.bayer",		"../../clt/aux_chn14.bayer",
	"clt/aux_chn15.bayer"};		"../../clt/aux_chn15.bayer"};

	const char *ports_offs_xy_file[] = {		const char *ports_offs_xy_file[] = {
	"clt/aux_chn0.portsxy",		"../../clt/aux_chn0.portsxy",
	"clt/aux_chn1.portsxy",		"../../clt/aux_chn1.portsxy",
	"clt/aux_chn2.portsxy",		"../../clt/aux_chn2.portsxy",
	"clt/aux_chn3.portsxy",		"../../clt/aux_chn3.portsxy",
	"clt/aux_chn4.portsxy",		"../../clt/aux_chn4.portsxy",
	"clt/aux_chn5.portsxy",		"../../clt/aux_chn5.portsxy",
	"clt/aux_chn6.portsxy",		"../../clt/aux_chn6.portsxy",
	"clt/aux_chn7.portsxy",		"../../clt/aux_chn7.portsxy",
	"clt/aux_chn8.portsxy",		"../../clt/aux_chn8.portsxy",
	"clt/aux_chn9.portsxy",		"../../clt/aux_chn9.portsxy",
	"clt/aux_chn10.portsxy",		"../../clt/aux_chn10.portsxy",
	"clt/aux_chn11.portsxy",		"../../clt/aux_chn11.portsxy",
	"clt/aux_chn12.portsxy",		"../../clt/aux_chn12.portsxy",
	"clt/aux_chn13.portsxy",		"../../clt/aux_chn13.portsxy",
	"clt/aux_chn14.portsxy",		"../../clt/aux_chn14.portsxy",
	"clt/aux_chn15.portsxy"};		"../../clt/aux_chn15.portsxy"};

	//#ifndef DBG_TILE		//#ifndef DBG_TILE
	#ifdef SAVE_CLT		#ifdef SAVE_CLT
	const char *ports_clt_file[] = {// never referenced		const char *ports_clt_file[] = {// never referenced
	"clt/aux_chn0.clt",		"../../clt/aux_chn0.clt",
	"clt/aux_chn1.clt",		"../../clt/aux_chn1.clt",
	"clt/aux_chn2.clt",		"../../clt/aux_chn2.clt",
	"clt/aux_chn3.clt",		"../../clt/aux_chn3.clt",
	"clt/aux_chn4.clt",		"../../clt/aux_chn4.clt",
	"clt/aux_chn5.clt",		"../../clt/aux_chn5.clt",
	"clt/aux_chn6.clt",		"../../clt/aux_chn6.clt",
	"clt/aux_chn7.clt",		"../../clt/aux_chn7.clt",
	"clt/aux_chn8.clt",		"../../clt/aux_chn8.clt",
	"clt/aux_chn9.clt",		"../../clt/aux_chn9.clt",
	"clt/aux_chn10.clt",		"../../clt/aux_chn10.clt",
	"clt/aux_chn11.clt",		"../../clt/aux_chn11.clt",
	"clt/aux_chn12.clt",		"../../clt/aux_chn12.clt",
	"clt/aux_chn13.clt",		"../../clt/aux_chn13.clt",
	"clt/aux_chn14.clt",		"../../clt/aux_chn14.clt",
	"clt/aux_chn15.clt"};		"../../clt/aux_chn15.clt"};

	#endif		#endif
	const char *result_rbg_file[] = {		const char *result_rbg_file[] = {
	"clt/aux_chn0.rbg",		"../../clt/aux_chn0.rbg",
	"clt/aux_chn1.rbg",		"../../clt/aux_chn1.rbg",
	"clt/aux_chn2.rbg",		"../../clt/aux_chn2.rbg",
	"clt/aux_chn3.rbg",		"../../clt/aux_chn3.rbg",
	"clt/aux_chn4.rbg",		"../../clt/aux_chn4.rbg",
	"clt/aux_chn5.rbg",		"../../clt/aux_chn5.rbg",
	"clt/aux_chn6.rbg",		"../../clt/aux_chn6.rbg",
	"clt/aux_chn7.rbg",		"../../clt/aux_chn7.rbg",
	"clt/aux_chn8.rbg",		"../../clt/aux_chn8.rbg",
	"clt/aux_chn9.rbg",		"../../clt/aux_chn9.rbg",
	"clt/aux_chn10.rbg",		"../../clt/aux_chn10.rbg",
	"clt/aux_chn11.rbg",		"../../clt/aux_chn11.rbg",
	"clt/aux_chn12.rbg",		"../../clt/aux_chn12.rbg",
	"clt/aux_chn13.rbg",		"../../clt/aux_chn13.rbg",
	"clt/aux_chn14.rbg",		"../../clt/aux_chn14.rbg",
	"clt/aux_chn15.rbg"};		"../../clt/aux_chn15.rbg"};
	//#endif		//#endif
	const char *result_corr_file = "clt/aux_corr.corr";		const char *result_corr_file = "../../clt/aux_corr.corr";
	const char *result_corr_quad_file = "clt/aux_corr-quad.corr";		const char *result_corr_quad_file = "../../clt/aux_corr-quad.corr";
	const char *result_corr_td_norm_file = "clt/aux_corr-td-norm.corr";		const char *result_corr_td_norm_file = "../../clt/aux_corr-td-norm.corr";
	/// const char* result_corr_cross_file = "clt/aux_corr-cross.corr";		/// const char* result_corr_cross_file = "../../clt/aux_corr-cross.corr";
	const char *result_textures_file = "clt/aux_texture_nodp.rgba";		const char *result_textures_file = "../../clt/aux_texture_nodp.rgba";
	const char *result_diff_rgb_combo_file = "clt/aux_diff_rgb_combo_nodp.drbg";		const char *result_diff_rgb_combo_file = "../../clt/aux_diff_rgb_combo_nodp.drbg";
	const char *result_textures_rgba_file = "clt/aux_texture_rgba_nodp.rgba";		const char *result_textures_rgba_file = "../../clt/aux_texture_rgba_nodp.rgba";

	const char *result_textures_file_dp = "clt/aux_texture_dp.rgba";		const char *result_textures_file_dp = "../../clt/aux_texture_dp.rgba";
	const char *result_diff_rgb_combo_file_dp = "clt/aux_diff_rgb_combo_dp.drbg";		const char *result_diff_rgb_combo_file_dp = "../../clt/aux_diff_rgb_combo_dp.drbg";
	const char *result_textures_rgba_file_dp = "clt/aux_texture_rgba_dp.rgba";		const char *result_textures_rgba_file_dp = "../../clt/aux_texture_rgba_dp.rgba";

	const char *rByRDist_file = "clt/aux.rbyrdist";		const char *rByRDist_file = "../../clt/aux.rbyrdist";
	const char *correction_vector_file = "clt/aux.correction_vector";		const char *correction_vector_file = "../../clt/aux.correction_vector";
	const char *geometry_correction_file = "clt/aux.geometry_correction";		const char *geometry_correction_file = "../../clt/aux.geometry_correction";

	float color_weights[] = {		float color_weights[] = {
	1.0, // float weight0, // scale for R 0.5 / (1.0 + 0.5 +0.2)		1.0, // float weight0, // scale for R 0.5 / (1.0 + 0.5 +0.2)
	@@ -682,56 +683,56 @@ int main(int argc, char **argv) {

	#else		#else
	const char *kernel_file[] = {		const char *kernel_file[] = {
	"clt/main_chn0_transposed.kernel",		"../../clt/main_chn0_transposed.kernel",
	"clt/main_chn1_transposed.kernel",		"../../clt/main_chn1_transposed.kernel",
	"clt/main_chn2_transposed.kernel",		"../../clt/main_chn2_transposed.kernel",
	"clt/main_chn3_transposed.kernel"};		"../../clt/main_chn3_transposed.kernel"};

	const char *kernel_offs_file[] = {		const char *kernel_offs_file[] = {
	"clt/main_chn0_transposed.kernel_offsets",		"../../clt/main_chn0_transposed.kernel_offsets",
	"clt/main_chn1_transposed.kernel_offsets",		"../../clt/main_chn1_transposed.kernel_offsets",
	"clt/main_chn2_transposed.kernel_offsets",		"../../clt/main_chn2_transposed.kernel_offsets",
	"clt/main_chn3_transposed.kernel_offsets"};		"../../clt/main_chn3_transposed.kernel_offsets"};

	const char *image_files[] = {		const char *image_files[] = {
	"clt/main_chn0.bayer",		"../../clt/main_chn0.bayer",
	"clt/main_chn1.bayer",		"../../clt/main_chn1.bayer",
	"clt/main_chn2.bayer",		"../../clt/main_chn2.bayer",
	"clt/main_chn3.bayer"};		"../../clt/main_chn3.bayer"};

	const char *ports_offs_xy_file[] = {		const char *ports_offs_xy_file[] = {
	"clt/main_chn0.portsxy",		"../../clt/main_chn0.portsxy",
	"clt/main_chn1.portsxy",		"../../clt/main_chn1.portsxy",
	"clt/main_chn2.portsxy",		"../../clt/main_chn2.portsxy",
	"clt/main_chn3.portsxy"};		"../../clt/main_chn3.portsxy"};
	#ifdef SAVE_CLT		#ifdef SAVE_CLT
	const char *ports_clt_file[] = {// never referenced		const char *ports_clt_file[] = {// never referenced
	"clt/main_chn0.clt",		"../../clt/main_chn0.clt",
	"clt/main_chn1.clt",		"../../clt/main_chn1.clt",
	"clt/main_chn2.clt",		"../../clt/main_chn2.clt",
	"clt/main_chn3.clt"};		"../../clt/main_chn3.clt"};
	#endif		#endif
	const char *result_rbg_file[] = {		const char *result_rbg_file[] = {
	"clt/main_chn0.rbg",		"../../clt/main_chn0.rbg",
	"clt/main_chn1.rbg",		"../../clt/main_chn1.rbg",
	"clt/main_chn2.rbg",		"../../clt/main_chn2.rbg",
	"clt/main_chn3.rbg"};		"../../clt/main_chn3.rbg"};
	//#endif		//#endif
	const char *result_corr_file = "clt/main_corr.corr";		const char *result_corr_file = "../../clt/main_corr.corr";
	const char *result_corr_quad_file = "clt/main_corr-quad.corr";		const char *result_corr_quad_file = "../../clt/main_corr-quad.corr";
	const char *result_corr_td_norm_file = "clt/aux_corr-td-norm.corr";		const char *result_corr_td_norm_file = "../../clt/aux_corr-td-norm.corr";
	/// const char* result_corr_cross_file = "clt/main_corr-cross.corr";		/// const char* result_corr_cross_file = "../../clt/main_corr-cross.corr";
	const char *result_textures_file = "clt/main_texture_nodp.rgba";		const char *result_textures_file = "../../clt/main_texture_nodp.rgba";
	const char *result_diff_rgb_combo_file = "clt/main_diff_rgb_combo_nodp.drbg";		const char *result_diff_rgb_combo_file = "../../clt/main_diff_rgb_combo_nodp.drbg";
	const char *result_textures_rgba_file = "clt/main_texture_rgba_nodp.rgba";		const char *result_textures_rgba_file = "../../clt/main_texture_rgba_nodp.rgba";

	const char *result_textures_file_dp = "clt/main_texture_dp.rgba";		const char *result_textures_file_dp = "../../clt/main_texture_dp.rgba";
	const char *result_diff_rgb_combo_file_dp = "clt/main_diff_rgb_combo_dp.drbg";		const char *result_diff_rgb_combo_file_dp = "../../clt/main_diff_rgb_combo_dp.drbg";
	const char *result_textures_rgba_file_dp = "clt/main_texture_rgba_dp.rgba";		const char *result_textures_rgba_file_dp = "../../clt/main_texture_rgba_dp.rgba";

	const char *rByRDist_file = "clt/main.rbyrdist";		const char *rByRDist_file = "../../clt/main.rbyrdist";
	const char *correction_vector_file = "clt/main.correction_vector";		const char *correction_vector_file = "../../clt/main.correction_vector";
	const char *geometry_correction_file = "clt/main.geometry_correction";		const char *geometry_correction_file = "../../clt/main.geometry_correction";

	float color_weights[] = {		float color_weights[] = {
	0.294118, // float weight0, // scale for R 0.5 / (1.0 + 0.5 +0.2)		0.294118, // float weight0, // scale for R 0.5 / (1.0 + 0.5 +0.2)
	@@ -1626,8 +1627,8 @@ int main(int argc, char **argv) {
	(corr_size_combo * corr_size_combo) * sizeof(float),		(corr_size_combo * corr_size_combo) * sizeof(float),
	num_corr_combo,		num_corr_combo,
	cudaMemcpyDeviceToHost));		cudaMemcpyDeviceToHost));
	// const char* result_corr_quad_file = "clt/main_corr-quad.corr";		// const char* result_corr_quad_file = "../../clt/main_corr-quad.corr";
	// const char* result_corr_cross_file = "clt/main_corr-cross.corr";		// const char* result_corr_cross_file = "../../clt/main_corr-cross.corr";

	#ifndef NSAVE_CORR		#ifndef NSAVE_CORR
	printf("Writing phase correlation data to %s\n", result_corr_quad_file);		printf("Writing phase correlation data to %s\n", result_corr_quad_file);