Commit 460eda96 authored by Andrey Filippov's avatar Andrey Filippov

moved LPF to 4-image generation to match correlations and textures

parent 06b62a05
......@@ -106,6 +106,8 @@
#define DBG_TILE_Y 111 // 66
#define DBG_TILE (DBG_TILE_Y * 324 + DBG_TILE_X)
#undef DBG_MARK_DBG_TILE 1
//56494
// struct tp_task
//#define TASK_SIZE 12
......@@ -879,6 +881,18 @@ extern "C" __global__ void textures_accumulate(
size_t texture_stride, // in floats (now 256*4 = 1024)
float * gpu_texture_tiles); // (number of colors +1 + ?)*16*16 rgba texture tiles
extern "C"
__global__ void imclt_rbg(
float * gpu_clt, // [TILESY][TILESX][NUM_COLORS][DTT_SIZE*DTT_SIZE]
float * gpu_rbg, // WIDTH, 3 * HEIGHT
int apply_lpf,
int mono, // defines lpf filter
int color, // defines location of clt data
int v_offset,
int h_offset,
const size_t dstride); // in floats (pixels)
//===========================
extern "C"
__global__ void correlate2D(
float ** gpu_clt, // [NUM_CAMS] ->[TILESY][TILESX][NUM_COLORS][DTT_SIZE*DTT_SIZE]
......@@ -2274,16 +2288,12 @@ __global__ void textures_accumulate(
} // textures_accumulate()
extern "C"
__global__ void imclt_rbg(
float * gpu_clt, // [TILESY][TILESX][NUM_COLORS][DTT_SIZE*DTT_SIZE]
float * gpu_rbg, // WIDTH, 3 * HEIGHT
int apply_lpf,
int mono,
int color,
int v_offset,
int h_offset,
......@@ -2334,13 +2344,27 @@ __global__ void imclt_rbg(
clt_tile += column + thr3; // first 2 rows
gpu_tile += column; // first 2 rows
if (apply_lpf) {
// lpf - covers 2 rows, as there there are 16 threads
float *lpf0 = lpf_data[mono? 3 :color] + threadIdx.x; // lpf_data[3] - mono
#pragma unroll
for (int q = 0; q < 4; q++){
float *lpf = lpf0;
for (int i = 0; i < DTT_SIZE/2; i++){
*clt_tile= *gpu_tile * (*lpf);
clt_tile += (2 * DTT_SIZE1);
gpu_tile += (2 * DTT_SIZE);
lpf += (2 * DTT_SIZE);
}
}
} else {
#pragma unroll
for (int i = 0; i < DTT_SIZE2; i++){
*clt_tile= *gpu_tile;
clt_tile += (2 * DTT_SIZE1);
gpu_tile += (2 * DTT_SIZE);
}
}
float * mclt_top = ((float*) mclt_tiles) + tile_in_block * (DTT_SIZE2 * DTT_SIZE21) + column;
float * rbg_top = color_plane + (tileY * DTT_SIZE)* dstride + (tileX * DTT_SIZE) + column;
float * mclt_tile = mclt_top;
......@@ -2377,7 +2401,7 @@ __global__ void imclt_rbg(
// save result (back)
// save result (back)
float * rbg_p = rbg_top;
mclt_tile = mclt_top;
if ((tileX == 0) && (tileY == 0)){
......@@ -2387,6 +2411,7 @@ __global__ void imclt_rbg(
mclt_tile += DTT_SIZE21;
rbg_p += dstride; // DTT_SIZE2; // FIXME
}
#ifdef DBG_MARK_DBG_TILE
} else if ((tileX == DBG_TILE_X) && (tileY == DBG_TILE_Y)){
#pragma unroll
for (int i = 0; i < DTT_SIZE2; i++){
......@@ -2394,6 +2419,7 @@ __global__ void imclt_rbg(
mclt_tile += DTT_SIZE21;
rbg_p += dstride; // DTT_SIZE2; // FIXME
}
#endif
} else {
#pragma unroll
for (int i = 0; i < DTT_SIZE2; i++){
......
......@@ -651,7 +651,9 @@ int main(int argc, char **argv)
imclt_rbg<<<grid_imclt,threads_imclt>>>(
gpu_clt_h[ncam], // float * gpu_clt, // [TILESY][TILESX][NUM_COLORS][DTT_SIZE*DTT_SIZE]
gpu_corr_images_h[ncam], // float * gpu_rbg, // WIDTH, 3 * HEIGHT
color, // int color,
1, // int apply_lpf,
0, // int mono, // defines lpf filter
color, // int color, // defines location of clt data
v_offs, // int v_offset,
h_offs, // int h_offset,
dstride_rslt/sizeof(float)); //const size_t dstride); // in floats (pixels)
......
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