Skip to content

T
tile_processor_gpu
Commit cde525c8 authored by Andrey Filippov's avatar Andrey Filippov
Browse files
Options

debugging

parent 40239aff
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 128 additions and 53 deletions
src/TileProcessor.cuh
View file @ cde525c8
...@@ -1031,7 +1031,11 @@ __global__ void index_direct( ...@@ -1031,7 +1031,11 @@ __global__ void index_direct(
__global__ void index_correlate( __global__ void index_correlate(
int num_cams, int num_cams,
int * sel_pairs, // int * sel_pairs, // unused bits should be 0
int sel_pairs0,
int sel_pairs1,
int sel_pairs2,
int sel_pairs3,
float * gpu_ftasks, // flattened tasks, 27 floats for quad EO, 99 floats for LWIR16 float * gpu_ftasks, // flattened tasks, 27 floats for quad EO, 99 floats for LWIR16
// struct tp_task * gpu_tasks, // struct tp_task * gpu_tasks,
int num_tiles, // number of tiles in task int num_tiles, // number of tiles in task
...@@ -1161,7 +1165,11 @@ __device__ int get_textures_shared_size( // in bytes ...@@ -1161,7 +1165,11 @@ __device__ int get_textures_shared_size( // in bytes
*/ */
extern "C" __global__ void correlate2D( extern "C" __global__ void correlate2D(
int num_cams, int num_cams,
int * sel_pairs, // int * sel_pairs,
int sel_pairs0,
int sel_pairs1,
int sel_pairs2,
int sel_pairs3,
float ** gpu_clt, // [num_cams] ->[TILES-Y][TILES-X][colors][DTT_SIZE*DTT_SIZE] float ** gpu_clt, // [num_cams] ->[TILES-Y][TILES-X][colors][DTT_SIZE*DTT_SIZE]
int colors, // number of colors (3/1) int colors, // number of colors (3/1)
float scale0, // scale for R float scale0, // scale for R
...@@ -1184,29 +1192,34 @@ extern "C" __global__ void correlate2D( ...@@ -1184,29 +1192,34 @@ extern "C" __global__ void correlate2D(
*pnum_corr_tiles = 0; *pnum_corr_tiles = 0;
index_correlate<<<blocks0,threads0>>>( index_correlate<<<blocks0,threads0>>>(
num_cams, // int num_cams, num_cams, // int num_cams,
sel_pairs, // int * sel_pairs, // sel_pairs, // int * sel_pairs,
sel_pairs0, // int sel_pairs0,
sel_pairs1, // int sel_pairs1,
sel_pairs2, // int sel_pairs2,
sel_pairs3, // int sel_pairs3,
gpu_ftasks, // float * gpu_ftasks, // flattened tasks, 27 floats for quad EO, 99 floats for LWIR16 gpu_ftasks, // float * gpu_ftasks, // flattened tasks, 27 floats for quad EO, 99 floats for LWIR16
// gpu_tasks, // struct tp_task * gpu_tasks, // gpu_tasks, // struct tp_task * gpu_tasks,
num_tiles, // int num_tiles, // number of tiles in task num_tiles, // int num_tiles, // number of tiles in task
tilesx, // int width, // number of tiles in a row tilesx, // int width, // number of tiles in a row
gpu_corr_indices, // int * gpu_corr_indices, // array of correlation tasks gpu_corr_indices, // int * gpu_corr_indices, // array of correlation tasks
pnum_corr_tiles); // int * pnum_corr_tiles); // pointer to the length of correlation tasks array pnum_corr_tiles); // int * pnum_corr_tiles); // pointer to the length of correlation tasks array
cudaDeviceSynchronize(); cudaDeviceSynchronize();
dim3 threads_corr(CORR_THREADS_PER_TILE, CORR_TILES_PER_BLOCK, 1); dim3 threads_corr(CORR_THREADS_PER_TILE, CORR_TILES_PER_BLOCK, 1);
dim3 grid_corr((*pnum_corr_tiles + CORR_TILES_PER_BLOCK-1) / CORR_TILES_PER_BLOCK,1,1); dim3 grid_corr((*pnum_corr_tiles + CORR_TILES_PER_BLOCK-1) / CORR_TILES_PER_BLOCK,1,1);
correlate2D_inner<<<grid_corr,threads_corr>>>( correlate2D_inner<<<grid_corr,threads_corr>>>(
num_cams, // int num_cams, num_cams, // int num_cams,
gpu_clt, // float ** gpu_clt, // [num_cams] ->[TILES-Y][TILES-X][colors][DTT_SIZE*DTT_SIZE] gpu_clt, // float ** gpu_clt, // [num_cams] ->[TILES-Y][TILES-X][colors][DTT_SIZE*DTT_SIZE]
colors, // int colors, // number of colors (3/1) colors, // int colors, // number of colors (3/1)
scale0, // float scale0, // scale for R scale0, // float scale0, // scale for R
scale1, // float scale1, // scale for B scale1, // float scale1, // scale for B
scale2, // float scale2, // scale for G scale2, // float scale2, // scale for G
fat_zero, // float fat_zero, // here - absolute fat_zero, // float fat_zero, // here - absolute
*pnum_corr_tiles, // size_t num_corr_tiles, // number of correlation tiles to process *pnum_corr_tiles, // size_t num_corr_tiles, // number of correlation tiles to process
gpu_corr_indices, // int * gpu_corr_indices, // packed tile+pair gpu_corr_indices, // int * gpu_corr_indices, // packed tile+pair
corr_stride, // const size_t corr_stride, // in floats corr_stride, // const size_t corr_stride, // in floats
corr_radius, // int corr_radius, // radius of the output correlation (7 for 15x15) corr_radius, // int corr_radius, // radius of the output correlation (7 for 15x15)
gpu_corrs); // float * gpu_corrs); // correlation output data gpu_corrs); // float * gpu_corrs); // correlation output data
} }
} }
...@@ -2493,7 +2506,11 @@ __global__ void create_nonoverlap_list( ...@@ -2493,7 +2506,11 @@ __global__ void create_nonoverlap_list(
*/ */
__global__ void index_correlate( __global__ void index_correlate(
int num_cams, int num_cams,
int * sel_pairs, // unused bits should be 0 // int * sel_pairs, // unused bits should be 0
int sel_pairs0,
int sel_pairs1,
int sel_pairs2,
int sel_pairs3,
float * gpu_ftasks, // flattened tasks, 27 floats for quad EO, 99 floats for LWIR16 float * gpu_ftasks, // flattened tasks, 27 floats for quad EO, 99 floats for LWIR16
// struct tp_task * gpu_tasks, // struct tp_task * gpu_tasks,
int num_tiles, // number of tiles in task int num_tiles, // number of tiles in task
...@@ -2505,6 +2522,7 @@ __global__ void index_correlate( ...@@ -2505,6 +2522,7 @@ __global__ void index_correlate(
if (num_tile >= num_tiles){ if (num_tile >= num_tiles){
return; return;
} }
int sel_pairs[] = {sel_pairs0, sel_pairs1, sel_pairs2, sel_pairs3};
// int task_size = get_task_size(num_cams); // int task_size = get_task_size(num_cams);
int task_task =get_task_task(num_tile, gpu_ftasks, num_cams); int task_task =get_task_task(num_tile, gpu_ftasks, num_cams);
if (((task_task >> TASK_CORR_BITS) & 1) == 0){ // needs correlation. Maybe just check task_task != 0? if (((task_task >> TASK_CORR_BITS) & 1) == 0){ // needs correlation. Maybe just check task_task != 0?
......
src/TileProcessor.h
View file @ cde525c8
...@@ -65,7 +65,11 @@ extern "C" __global__ void convert_direct( // called with a single block, single ...@@ -65,7 +65,11 @@ extern "C" __global__ void convert_direct( // called with a single block, single
extern "C" __global__ void correlate2D( extern "C" __global__ void correlate2D(
int num_cams, int num_cams,
int * sel_pairs, // int * sel_pairs,
int sel_pairs0,
int sel_pairs1,
int sel_pairs2,
int sel_pairs3,
float ** gpu_clt, // [NUM_CAMS] ->[TILES-Y][TILES-X][NUM_COLORS][DTT_SIZE*DTT_SIZE] float ** gpu_clt, // [NUM_CAMS] ->[TILES-Y][TILES-X][NUM_COLORS][DTT_SIZE*DTT_SIZE]
int colors, // number of colors (3/1) int colors, // number of colors (3/1)
float scale0, // scale for R float scale0, // scale for R
......
src/geometry_correction.cu
View file @ cde525c8
...@@ -37,11 +37,16 @@ ...@@ -37,11 +37,16 @@
*/ */
#ifndef JCUDA #ifndef JCUDA
#include "tp_defines.h" #include "tp_defines.h"
#include "dtt8x8.h" #include "dtt8x8.h"
#include "geometry_correction.h" #include "geometry_correction.h"
#endif // #ifndef JCUDA #endif // #ifndef JCUDA
#ifndef get_task_size
#define get_task_size(x) (sizeof(struct tp_task)/sizeof(float) - 6 * (NUM_CAMS - x))
#endif
// Using NUM_CAMS threads per tile // Using NUM_CAMS threads per tile
#define THREADS_PER_BLOCK_GEOM (TILES_PER_BLOCK_GEOM * NUM_CAMS) #define THREADS_PER_BLOCK_GEOM (TILES_PER_BLOCK_GEOM * NUM_CAMS)
///#define CYCLES_COPY_GC ((sizeof(struct gc)/sizeof(float) + THREADS_PER_BLOCK_GEOM - 1) / THREADS_PER_BLOCK_GEOM) ///#define CYCLES_COPY_GC ((sizeof(struct gc)/sizeof(float) + THREADS_PER_BLOCK_GEOM - 1) / THREADS_PER_BLOCK_GEOM)
......
src/test_tp.cu
View file @ cde525c8
This diff is collapsed.
src/tp_defines.h
View file @ cde525c8
...@@ -41,14 +41,22 @@ ...@@ -41,14 +41,22 @@
#ifndef JCUDA #ifndef JCUDA
#include <stdio.h> #include <stdio.h>
#define THREADSX (DTT_SIZE) #define THREADSX (DTT_SIZE)
#define TEST_LWIR 1 #define TEST_LWIR 0
#define NUM_CAMS 16 // now maximal number of cameras #define NUM_CAMS 16 // now maximal number of cameras
#define NUM_PAIRS 6 //#define NUM_PAIRS 6
#define NUM_COLORS 1 //3 #define NUM_COLORS 1 //3
#define IMG_WIDTH 2592 // kernels [num_cams][num_colors][KERNELS_HOR][KERNELS_VERT][4][64]
#define IMG_HEIGHT 1936 #if TEST_LWIR
#define KERNELS_HOR 164 #define IMG_WIDTH 640
#define KERNELS_VERT 123 #define IMG_HEIGHT 512
#define KERNELS_HOR 82 // 80+2
#define KERNELS_VERT 66 // 64+2
#else
#define IMG_WIDTH 2592
#define IMG_HEIGHT 1936
#define KERNELS_HOR 164 // 2592 / 16 + 2
#define KERNELS_VERT 123 // 1936 / 16 + 2
#endif
#define KERNELS_LSTEP 4 #define KERNELS_LSTEP 4
#define THREADS_PER_TILE 8 #define THREADS_PER_TILE 8
#define TILES_PER_BLOCK 4 #define TILES_PER_BLOCK 4
...@@ -73,7 +81,7 @@ ...@@ -73,7 +81,7 @@
#define TASK_TEXTURE_W_BIT 3 // Texture with West neighbor #define TASK_TEXTURE_W_BIT 3 // Texture with West neighbor
#define TASK_TEXTURE_BIT 3 // bit to request texture calculation int task field of struct tp_task #define TASK_TEXTURE_BIT 3 // bit to request texture calculation int task field of struct tp_task
#define LIST_TEXTURE_BIT 7 // bit to request texture calculation #define LIST_TEXTURE_BIT 7 // bit to request texture calculation
#define CORR_OUT_RAD 4 #define CORR_OUT_RAD 7 // full tile (15x15), was 4 (9x9)
#define FAT_ZERO_WEIGHT 0.0001 // add to port weights to avoid nan #define FAT_ZERO_WEIGHT 0.0001 // add to port weights to avoid nan
#define THREADS_DYNAMIC_BITS 5 // treads in block for CDP creation of the texture list #define THREADS_DYNAMIC_BITS 5 // treads in block for CDP creation of the texture list
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment