debugged initially with jcuda (d8e9a454) · Commits · Elphel / tile_processor_gpu

src/TpHostGpu.cu

+25 −15

Original line number	Diff line number	Diff line
		@@ -5,15 +5,21 @@
		* Author: elphel
		*/
		#include <stdexcept>
		#include <helper_cuda.h> // for checkCudaErrors
		//#include <driver_types.h> // was not needed before, only for indexes - needs __DRIVER_TYPES_H__
		#include <cstdlib>
		#include <cstdio>
		#include <cuda_runtime.h> // cudaFree
		#include <helper_cuda.h> // for checkCudaErrors
		#include <helper_functions.h> // timer functions




		//#include "TpParams.h" // TpHostGpu.h has it
		#include "tp_paths.h"
		#include "tp_files.h"
		#include "tp_utils.h" // for copyalloc_kernel_gpu
		#include "GenerateRgbaHost.h"
		//#include "GenerateRgbaHost.h"
		#include "TpHostGpu.h"

		#define MY_EXCEPTION(aMessage) \
		@@ -79,6 +85,7 @@ void TpHostGpu::allTests(
		int image_dy,
		const float target_disparity,
		const float scale,
		const float fat_zero, // 1000.0
		int quad_combine,
		int use_dp,
		int debug){
		@@ -100,10 +107,10 @@ void TpHostGpu::allTests(
		testConvertDirect (num_runs); // 608
		testImcltRbgAll (num_runs); // 701

		testCorrelate2DIntra (num_runs); // 762 - 885
		testCorrelate2DIntraTD (num_runs, quad_combine); // 886 - 1123
		testCorrelate2DIntra (num_runs, fat_zero); // 762 - 885
		testCorrelate2DIntraTD (num_runs, fat_zero, quad_combine); // 886 - 1123
		setImgBuffersShifted(is_bayer, image_dx, image_dy); // 1171-1188
		testCorrelate2DInterSelf(num_runs); // 1136 - 1411
		testCorrelate2DInterSelf(num_runs, fat_zero); // 1136 - 1411
		testTextures (num_runs, use_dp, debug); // 1422-1664
		testTexturesRGBA (num_runs, use_dp, debug); // 1669-1810
		return;
		@@ -696,7 +703,7 @@ void TpHostGpu::testImcltRbgAll (int num_runs){ // 701
		m_gpu_corr_images_h); // float ** gpu_corr_images_h){
		}

		void TpHostGpu::testCorrelate2DIntra(int num_runs){
		void TpHostGpu::testCorrelate2DIntra(int num_runs, float fat_zero){
		int num_corr_indices = m_tpParams.num_pairs * m_tpParams.num_tiles;
		int numIterations = m_tpParams.debug_tile ? 1 : num_runs;
		int i0 = m_tpParams.debug_tile ? 0 : -1;
		@@ -730,7 +737,7 @@ void TpHostGpu::testCorrelate2DIntra(int num_runs){
		m_tpParams.color_weights[0], // 0.25, // float scale0, // scale for R
		m_tpParams.color_weights[1], // 0.25, // float scale1, // scale for B
		m_tpParams.color_weights[2], // 0.5, // float scale2, // scale for G
		m_tpParams.fat_zero * m_tpParams.fat_zero, // float fat_zero2, // here - absolute
		fat_zero * fat_zero, // float fat_zero2, // here - absolute
		m_gpu_ftasks, // float * gpu_ftasks, // flattened tasks, 27 floats for quad EO, 99 floats for LWIR16
		m_tpParams.tp_tasks_size, // int num_tiles) // number of tiles in task
		m_tpParams.tilesx, // int tilesx, // number of tile rows
		@@ -764,7 +771,7 @@ void TpHostGpu::testCorrelate2DIntra(int num_runs){
		16); //int num_sel_sensors) { // only for interscene
		}

		void TpHostGpu::testCorrelate2DIntraTD (int num_runs, int quad_combine){ // 886 - 1123
		void TpHostGpu::testCorrelate2DIntraTD (int num_runs, float fat_zero, int quad_combine){ // 886 - 1123
		int numIterations = m_tpParams.debug_tile ? 1 : num_runs;
		int i0 = m_tpParams.debug_tile ? 0 : -1;
		// check/replace names
		@@ -817,7 +824,7 @@ void TpHostGpu::testCorrelate2DIntraTD (int num_runs, int quad_combine){ // 886
		m_tpParams.color_weights[0], // 0.25, // float scale0, // scale for R
		m_tpParams.color_weights[1], // 0.25, // float scale1, // scale for B
		m_tpParams.color_weights[2], // 0.5, // float scale2, // scale for G
		m_tpParams.fat_zero * m_tpParams.fat_zero, // float fat_zero2, // here - absolute
		fat_zero * fat_zero, // float fat_zero2, // here - absolute
		m_gpu_ftasks, // float * gpu_ftasks, // flattened tasks, 27 floats for quad EO, 99 floats for LWIR16
		m_tpParams.tp_tasks_size, // int num_tiles) // number of tiles in task
		m_tpParams.tilesx, // int tilesx, // number of tile rows
		@@ -856,7 +863,7 @@ void TpHostGpu::testCorrelate2DIntraTD (int num_runs, int quad_combine){ // 886
		(float ) 0, // float corr_weights, // null or per-tile weight (fat_zero2 will be divided by it)
		dstride_corr_combo/sizeof(float), // const size_t corr_stride, // in floats
		m_gpu_corrs_combo, // float * gpu_corrs, // correlation output data (pixel domain)
		m_tpParams.fat_zero * m_tpParams.fat_zero, // float fat_zero2, // here - absolute
		fat_zero * fat_zero, // float fat_zero2, // here - absolute
		m_tpParams.corr_out_rad); // int corr_radius); // radius of the output correlation (7 for 15x15)
		printf("corr2D_combine pass: %d\n",i);
		}else { // if (quad_combine) {
		@@ -868,7 +875,7 @@ void TpHostGpu::testCorrelate2DIntraTD (int num_runs, int quad_combine){ // 886
		(float ) 0, // corr_weights, // float corr_weights, // null or per-tile weight (fat_zero2 will be divided by it)
		dstride_corr/sizeof(float), // const size_t corr_stride, // in floats
		m_gpu_corrs, // float * gpu_corrs, // correlation output data (pixel domain)
		m_tpParams.fat_zero * m_tpParams.fat_zero, // float fat_zero2, // here - absolute
		fat_zero * fat_zero, // float fat_zero2, // here - absolute
		m_tpParams.corr_out_rad); // int corr_radius); // radius of the output correlation (7 for 15x15)
		} // if (quad_combine) {

		@@ -971,7 +978,7 @@ void TpHostGpu::testCorrelate2DIntraTD (int num_runs, int quad_combine){ // 886
		} // if (quad_combine) {
		}

		void TpHostGpu::testCorrelate2DInterSelf(int num_runs){ // 889
		void TpHostGpu::testCorrelate2DInterSelf(int num_runs, float fat_zero){ // 889
		int numIterations = m_tpParams.debug_tile ? 1 : num_runs;
		int i0 = m_tpParams.debug_tile ? 0 : -1;
		// check/replace names
		@@ -1087,7 +1094,7 @@ void TpHostGpu::testCorrelate2DInterSelf(int num_runs){ // 889
		(float ) 0, // corr_weights, // float corr_weights, // null or per-tile weight (fat_zero2 will be divided by it)
		dstride_corr/sizeof(float), // const size_t corr_stride, // in floats
		m_gpu_corrs, // float * gpu_corrs, // correlation output data (pixel domain)
		m_tpParams.fat_zero * m_tpParams.fat_zero, // float fat_zero2, // here - absolute
		fat_zero * fat_zero, // float fat_zero2, // here - absolute
		m_tpParams.corr_out_rad); // int corr_radius); // radius of the output correlation (7 for 15x15)
		getLastCudaError("Kernel failure:corr2D_normalize");
		checkCudaErrors(cudaDeviceSynchronize());
		@@ -1440,6 +1447,8 @@ void TpHostGpu::testTexturesRGBA (
		printf("\n1. shared_size=%d, num_cams=%d, colors=%d\n",shared_size, m_tpParams.num_cams, m_tpParams.texture_colors);

		cudaFuncSetAttribute(textures_accumulate, cudaFuncAttributeMaxDynamicSharedMemorySize, shared_size); // 60000); // 5536); // for CC 7.5
		// was not here - next line
		cudaFuncSetAttribute(textures_accumulate, cudaFuncAttributePreferredSharedMemoryCarveout,cudaSharedmemCarveoutMaxShared);
		generate_RBGA<<<1,1>>> (
		m_tpParams.num_cams, // int num_cams, // number of cameras used
		// Parameters to generate texture tasks
		@@ -1788,6 +1797,8 @@ float * TpHostGpu::getCorrTdImg(
		return corr_img;
		}

		//void TpHostGpu::generate_RBGA_host(
		// static // https://stackoverflow.com/questions/15725922/static-function-a-storage-class-may-not-be-specified-here
		void TpHostGpu::generate_RBGA_host(
		int num_cams, // number of cameras used
		// Parameters to generate texture tasks
		@@ -2031,7 +2042,6 @@ void TpHostGpu::generate_RBGA_host(
		pntt, // ntt, // int * num_texture_tiles, // number of texture tiles to process
		ti_offset, // gpu_texture_indices_offset,// add to gpu_texture_indices
		gpu_texture_indices, // + ti_offset, // int * gpu_texture_indices,// packed tile + bits (now only (1 << 7)
		// gpu_texture_indices + ti_offset, // int * gpu_texture_indices,// packed tile + bits (now only (1 << 7)
		gpu_geometry_correction, // struct gc * gpu_geometry_correction,
		colors, // int colors, // number of colors (3/1)
		is_lwir, // int is_lwir, // do not perform shot correction

src/TpHostGpu.h

+19 −19

Original line number	Diff line number	Diff line
		@@ -117,6 +117,7 @@ public:
		int image_dy, // 0
		const float target_disparity, // DBG_DISPARITY == 0.0
		const float scale, // 0.0
		const float fat_zero, // 1000.0
		int quad_combine,
		int use_dp,
		int debug);
		@@ -138,27 +139,13 @@ public:
		// void testImclt (int num_runs); // 682 // not implemented
		void testImcltRbgAll (int num_runs); // 701

		void testCorrelate2DIntra (int num_runs); // 762 - 885
		void testCorrelate2DIntraTD (int num_runs, int quad_combine); // 886 - 1123
		void testCorrelate2DIntra (int num_runs, float fat_zero); // 762 - 885
		void testCorrelate2DIntraTD (int num_runs, float fat_zero, int quad_combine); // 886 - 1123
		//void setImgBuffersShifted(int is_bayer, int image_dx, int image_dy); // 1171-1188
		void testCorrelate2DInterSelf(int num_runs); // 1136 - 1411
		void testCorrelate2DInterSelf(int num_runs, float fat_zero); // 1136 - 1411
		void testTextures (int num_runs, int use_dp, int debug); // 1422-1664
		void testTexturesRGBA (int num_runs, int use_dp, int debug); // 1669-1810


		private:
		void saveClt(const char ** paths, const char * prompt, float ** gpu_clt_h);
		void saveRgb(const char ** paths, const char * prompt, float ** gpu_corr_images_h);


		// for both intra and inter!
		void saveIntraCorrFile(const char * path, const char * prompt, int num_corrs, int num_corr_indices, float * gpu_corrs, int * gpu_corr_indices, int num_sel_sensors);
		void saveInterCorrFile(const char * path, const char * prompt, int num_corrs, int num_corr_indices, float * gpu_corrs_td, int * gpu_corr_indices, int num_sel_sensors);
		void saveInterCorrIndicesFile(const char * path, const char * prompt, int num_corr_indices, int * gpu_corr_indices, int num_sel_sensors);

		float * getCorrImg (int corr_img_size, int num_corr_indices, int * cpu_corr_indices, float * cpu_corr, int num_sel_sensors);
		float * getCorrTdImg(int corr_img_size, int num_corr_indices, int * cpu_corr_indices, float * cpu_corr_td, int num_sel_sensors);
		void generate_RBGA_host( // not a member
		static void generate_RBGA_host( // not a member
		int num_cams, // number of cameras used
		// Parameters to generate texture tasks
		float * gpu_ftasks, // flattened tasks, 27 floats for quad EO, 99 floats for LWIR16p// struct tp_task * gpu_tasks,
		@@ -182,6 +169,19 @@ private:
		const int texture_rbga_stride, // in floats
		float * gpu_texture_tiles); // (number of colors +1 + ?)1616 rgba texture tiles


		private:
		void saveClt(const char ** paths, const char * prompt, float ** gpu_clt_h);
		void saveRgb(const char ** paths, const char * prompt, float ** gpu_corr_images_h);


		// for both intra and inter!
		void saveIntraCorrFile(const char * path, const char * prompt, int num_corrs, int num_corr_indices, float * gpu_corrs, int * gpu_corr_indices, int num_sel_sensors);
		void saveInterCorrFile(const char * path, const char * prompt, int num_corrs, int num_corr_indices, float * gpu_corrs_td, int * gpu_corr_indices, int num_sel_sensors);
		void saveInterCorrIndicesFile(const char * path, const char * prompt, int num_corr_indices, int * gpu_corr_indices, int num_sel_sensors);

		float * getCorrImg (int corr_img_size, int num_corr_indices, int * cpu_corr_indices, float * cpu_corr, int num_sel_sensors);
		float * getCorrTdImg(int corr_img_size, int num_corr_indices, int * cpu_corr_indices, float * cpu_corr_td, int num_sel_sensors);
		void hfree(float *& p); // {if (p) free (p);}
		void hfree(struct CltExtra *& p);
		void gfree(float *& p);
		@@ -193,9 +193,9 @@ private:
		void gfree(struct trot_deriv *& p);
		void gfree(float **& p);
		void gfree(struct CltExtra **& p);

		};




		#endif /* SRC_TPHOSTGPU_H_ */

src/TpParams.h

+1 −1

Original line number	Diff line number	Diff line
		@@ -49,7 +49,7 @@ public:
		static constexpr int tp_task_centerxy_offset = TP_TASK_CENTERXY_OFFSET;// 3
		static constexpr int tp_task_scale_offset = TP_TASK_SCALE_OFFSET;// 5
		static constexpr int tp_task_xy_offset = TP_TASK_XY_OFFSET;// 6
		static constexpr float fat_zero = 1000.0f; // 300.0f; // 30.0;
		// static constexpr float fat_zero = 1000.0f; // 300.0f; // 30.0;

		static constexpr int convert_direct_indexing_threads = CONVERT_DIRECT_INDEXING_THREADS; //
		static constexpr int convert_direct_indexing_threads_log2 = CONVERT_DIRECT_INDEXING_THREADS_LOG2; //

src/geometry_correction.cu

+16 −5

Original line number	Diff line number	Diff line
		@@ -112,12 +112,23 @@ __constant__ float ROTS_TEMPLATE[7][3][3][3] = {// ...{cos,sin,const}...
		{{ 0, 0,0},{0, 0,0},{ 0, 0,0}},
		}
		};

		// TODO: Make offsets calculate in compile time, to avoid NVRTC(in java): " error: dynamic initialization is not supported for a __constant__ variable"
		__constant__ int angles_offsets [4] {15,0,30,30};
		/*
		__constant__ int angles_offsets [4] {
		(int) (offsetof4(corr_vector, azimuth)),
		(int) (offsetof4(corr_vector, tilt)),
		(int) (offsetof4(corr_vector, roll)),
		(int) (offsetof4(corr_vector, roll))};
		*/
		/*
		__constant__ int angles_offsets [4] = {
		offsetof(corr_vector, azimuth)/sizeof(float),
		offsetof(corr_vector, tilt) /sizeof(float),
		offsetof(corr_vector, roll) /sizeof(float),
		offsetof(corr_vector, roll) /sizeof(float)};
		(int) (offsetof(corr_vector, azimuth)/sizeof(float)),
		(int) (offsetof(corr_vector, tilt) /sizeof(float)),
		(int) (offsetof(corr_vector, roll) /sizeof(float)),
		(int) (offsetof(corr_vector, roll) /sizeof(float))};

		*/
		__constant__ int mm_seq [3][3][3]={
		{
		{6,5,12}, // a_t * a_z -> tmp0

src/geometry_correction.h

+5 −0

Original line number	Diff line number	Diff line
		@@ -51,6 +51,11 @@
		((size_t)&(((st *)0)->m))
		//#define offsetof(TYPE, MEMBER) __builtin_offsetof (TYPE, MEMBER)
		#endif
		#ifndef offsetof4
		#define offsetof4(st, m) \
		(((size_t)&(((st *)0)->m))>>2)
		//#define offsetof(TYPE, MEMBER) __builtin_offsetof (TYPE, MEMBER)
		#endif


		#define SCENE_UNITS_SCALE 0.001 // meters from mm

Admin message