gpu intra to batch (1b16c1e5) · Commits · Elphel / imagej-elphel

src/main/java/com/elphel/imagej/gpu/GPUTileProcessor.java

+23 −13

Original line number	Diff line number	Diff line
		@@ -896,7 +896,7 @@ public class GPUTileProcessor {
		}
		cuMemcpyHtoD(gpu_tasks, Pointer.to(ftasks), TPTASK_SIZE * num_task_tiles * Sizeof.FLOAT);
		}

		/*
		public void setCorrIndices(int [] corr_indices)
		{
		num_corr_tiles = corr_indices.length;
		@@ -906,7 +906,6 @@ public class GPUTileProcessor {
		}
		cuMemcpyHtoD(gpu_corr_indices, Pointer.to(fcorr_indices), num_corr_tiles * Sizeof.FLOAT);
		}

		public void setTextureIndices(int [] texture_indices) // never used
		{
		num_texture_tiles = texture_indices.length;
		@@ -916,6 +915,7 @@ public class GPUTileProcessor {
		}
		cuMemcpyHtoD(gpu_texture_indices, Pointer.to(ftexture_indices), num_texture_tiles * Sizeof.FLOAT);
		}
		*/

		public int [] getTextureIndices()
		{
		@@ -1733,6 +1733,7 @@ public class GPUTileProcessor {
		*/

		public void execCorr2D_normalize(
		boolean combo, // normalize combo correlations (false - per-pair ones)
		double fat_zero,
		int corr_radius) {
		if (GPU_CORR2D_NORMALIZE_kernel == null)
		@@ -1740,13 +1741,12 @@ public class GPUTileProcessor {
		IJ.showMessage("Error", "No GPU kernel: GPU_CORR2D_NORMALIZE_kernel");
		return;
		}
		// float [] fnum_corrs = new float[1];
		// cuMemcpyDtoH(Pointer.to(fnum_corrs), gpu_num_corr_tiles, 1 * Sizeof.FLOAT);
		// int num_tiles = Float.floatToIntBits(fnum_corrs[0])/num_pairs; // number of correlation tiles calculated

		int [] GridFullWarps = {1, 1, 1};
		int [] ThreadsFullWarps = {1, 1, 1};
		Pointer kernelParameters = Pointer.to(
		Pointer kernelParameters;
		if (combo) {
		kernelParameters = Pointer.to(
		Pointer.to(new int[] { num_corr_combo_tiles }), // num_task_tiles }), // int num_corr_tiles, // number of correlation tiles to process
		Pointer.to(new int[] { corr_stride_combo_td }),// const size_t corr_stride_td, // in floats
		Pointer.to(gpu_corrs_combo_td), // float * gpu_corrs_combo); // combined correlation output (one per tile)
		@@ -1754,6 +1754,16 @@ public class GPUTileProcessor {
		Pointer.to(gpu_corrs_combo), // float * gpu_corrs, // correlation output data (pixel domain)
		Pointer.to(new float[] {(float) fat_zero }), // float fat_zero, // here - absolute
		Pointer.to(new int[] { corr_radius })); // int corr_radius, // radius of the output correlation (7 for 15x15)
		} else {
		kernelParameters = Pointer.to(
		Pointer.to(new int[] { num_corr_tiles }), // num_task_tiles }), // int num_corr_tiles, // number of correlation tiles to process
		Pointer.to(new int[] { corr_stride_td }),// const size_t corr_stride_td, // in floats
		Pointer.to(gpu_corrs_td), // float * gpu_corrs_combo); // combined correlation output (one per tile)
		Pointer.to(new int[] { corr_stride }), // const size_t corr_stride, // in floats
		Pointer.to(gpu_corrs), // float * gpu_corrs, // correlation output data (pixel domain)
		Pointer.to(new float[] {(float) fat_zero }), // float fat_zero, // here - absolute
		Pointer.to(new int[] { corr_radius })); // int corr_radius, // radius of the output correlation (7 for 15x15)
		}

		cuCtxSynchronize();
		// Call the kernel function

src/main/java/com/elphel/imagej/tileprocessor/ImageDtt.java

+65 −45

Original line number	Diff line number	Diff line
		@@ -258,39 +258,6 @@ public class ImageDtt extends ImageDttCPU {
		" debug_tileX="+debug_tileX+" debug_tileY="+debug_tileY+" globalDebugLevel="+globalDebugLevel);
		}

		// TODO: Remove unused
		/**
		final int [][] zi =
		{{ 0, 1, 2, 3},
		{-1, 0, -3, 2},
		{-2, -3, 0, 1},
		{ 3, -2, -1, 0}};
		final int [][] corr_pairs ={ // {first, second, rot} rot: 0 - as is, 1 - swap y,x // not used in lwir
		{0,1,0},
		{2,3,0},
		{0,2,1},
		{1,3,1}};

		final double[][] port_offsets = { // lwir: used only in textures to scale differences
		{-0.5, -0.5},
		{ 0.5, -0.5},
		{-0.5, 0.5},
		{ 0.5, 0.5}};
		final int transform_len = transform_size * transform_size;
		final double [] filter = doubleGetCltLpfFd(corr_sigma);
		*/
		// prepare disparity maps and weights
		//// final int max_search_radius = (int) Math.abs(max_corr_radius); // use negative max_corr_radius for squares instead of circles?
		//// final int max_search_radius_poly = 1;
		/**
		if (globalDebugLevel > 0){
		System.out.println("max_corr_radius= "+max_corr_radius);
		System.out.println("max_search_radius= "+max_search_radius);
		System.out.println("max_search_radius_poly="+max_search_radius_poly);
		System.out.println("gpu_fat_zero= "+gpu_fat_zero);
		System.out.println("disparity_array[0][0]= "+disparity_array[0][0]);
		}
		*/

		// add optional initialization of debug layers here
		boolean need_macro = false;
		@@ -445,17 +412,52 @@ public class ImageDtt extends ImageDttCPU {
		// does it need correlations?
		if (fneed_corr) {
		//Generate 2D phase correlations from the CLT representation
		/*
		gpuQuad.execCorr2D(
		col_weights, // scales,// double [] scales,
		gpu_fat_zero, // double fat_zero);
		gpu_corr_rad); // int corr_radius
		//Show 2D correlations
		// int [] wh = new int[2];
		final int [] corr_indices = gpuQuad.getCorrIndices();
		final float [][] fcorr2D = gpuQuad.getCorr2D(
		gpu_corr_rad); // int corr_rad);
		final float [][] fcorr2D = gpuQuad.getCorr2D(gpu_corr_rad); // int corr_rad);
		*/
		gpuQuad.execCorr2D_TD(col_weights); // Get TD version of correlations (may be read out and saved)
		final int [] corr_indices = gpuQuad.getCorrIndices();
		gpuQuad.execCorr2D_normalize(
		false, // boolean combo, // normalize combo correlations (false - per-pair ones)
		gpu_fat_zero, // double fat_zero);
		gpu_corr_rad); // int corr_radius
		final float [][] fcorr2D = gpuQuad.getCorr2D(gpu_corr_rad); // int corr_rad);
		// calculate combine quad correlation
		gpuQuad.execCorr2D_combine( // calculate cross pairs
		true, // boolean init_corr, // initialize output tiles (false - add to current)
		GPUTileProcessor.NUM_PAIRS, // int num_pairs_in, // typically 6 - number of pairs per tile (tile task should have same number per each tile
		0x0f); // int pairs_mask // selected pairs (0x3 - horizontal, 0xc - vertical, 0xf - quad, 0x30 - cross)
		// normalize and convert to pixel domain
		gpuQuad.execCorr2D_normalize(
		true, // boolean combo, // normalize combo correlations (false - per-pair ones)
		gpu_fat_zero, // double fat_zero);
		gpu_corr_rad); // int corr_radius

		final int [] corr_quad_indices = gpuQuad.getCorrComboIndices(); // get quad
		final float [][] fcorr2D_quad = gpuQuad.getCorr2DCombo(gpu_corr_rad);
		// calculate and get cross here
		gpuQuad.execCorr2D_combine( // calculate cross pairs
		true, // boolean init_corr, // initialize output tiles (false - add to current)
		GPUTileProcessor.NUM_PAIRS, // int num_pairs_in, // typically 6 - number of pairs per tile (tile task should have same number per each tile
		0x30); // int pairs_mask // selected pairs (0x3 - horizontal, 0xc - vertical, 0xf - quad, 0x30 - cross)

		gpuQuad.execCorr2D_normalize(
		true, // boolean combo, // normalize combo correlations (false - per-pair ones)
		gpu_fat_zero, // double fat_zero);
		gpu_corr_rad); // int corr_radius

		// final int [] corr_cross_indices = gpuQuad.getCorrComboIndices(); // cross indices are the quad
		final float [][] fcorr2D_cross = gpuQuad.getCorr2DCombo(gpu_corr_rad);


		if (corr_indices.length > 0) {
		if (true) {
		/*
		if (true) { // debugging only
		int [] wh = new int[2];
		double [][] dbg_corr = GPUTileProcessor.getCorr2DView(
		tilesX,
		@@ -471,7 +473,7 @@ public class ImageDtt extends ImageDttCPU {
		"dbg-corr2D", // name+"-CORR2D-D"+clt_parameters.disparity,
		GPUTileProcessor.getCorrTitles());
		}

		*/

		final int corr_length = fcorr2D[0].length;// all correlation tiles have the same size
		// assuming that the correlation pairs sets are the same for each tile that has correlations
		@@ -481,7 +483,7 @@ public class ImageDtt extends ImageDttCPU {
		for (int i = 1; (i < corr_indices.length) && ((corr_indices[i] >> GPUTileProcessor.CORR_NTILE_SHIFT) == nt0) ; i++) {
		nc0++;
		}
		final int num_tile_corr = nc0;
		final int num_tile_corr = nc0; // normally 6
		final int num_tiles = corr_indices.length / num_tile_corr;


		@@ -504,7 +506,9 @@ public class ImageDtt extends ImageDttCPU {
		(imgdtt_params.lma_debug_level > 1)); // boolean debug);

		for (int indx_tile = ai.getAndIncrement(); indx_tile < num_tiles; indx_tile = ai.getAndIncrement()) {
		double [][] corrs = new double [GPUTileProcessor.NUM_PAIRS][corr_length]; // 225-long (15x15)
		// double [][] corrs = new double [GPUTileProcessor.NUM_PAIRS][corr_length]; // 225-long (15x15)
		// added quad and cross combos
		double [][] corrs = new double [GPUTileProcessor.NUM_PAIRS + 2][corr_length]; // 225-long (15x15)
		int indx_corr = indx_tile * num_tile_corr;
		int nt = (corr_indices[indx_corr] >> GPUTileProcessor.CORR_NTILE_SHIFT);
		int tileX = nt % tilesX;
		@@ -520,7 +524,21 @@ public class ImageDtt extends ImageDttCPU {
		}
		indx_corr++;
		}
		// add 2 combo layers
		int pair = GPUTileProcessor.NUM_PAIRS; // 6
		nt = (corr_quad_indices[indx_tile] >> GPUTileProcessor.CORR_NTILE_SHIFT); // corr_quad_indices - different sequence
		for (int i = 0; i < corr_length; i++) {
		corrs[pair][i] = gpu_corr_scale * fcorr2D_quad[indx_tile][i]; // from float to double
		}
		// indices for cross are the same as for quad
		pair++;
		for (int i = 0; i < corr_length; i++) {
		corrs[pair][i] = gpu_corr_scale * fcorr2D_cross[indx_tile][i]; // from float to double
		}

		// does not include combo
		int used_pairs = pair_mask; // imgdtt_params.dbg_pair_mask; //TODO: use tile tasks

		int tile_lma_debug_level = ((tileX == debug_tileX) && (tileY == debug_tileY))? (imgdtt_params.lma_debug_level-1) : -2;
		boolean debugTile =(tileX == debug_tileX) && (tileY == debug_tileY) && (globalDebugLevel > -1);

		@@ -611,8 +629,10 @@ public class ImageDtt extends ImageDttCPU {
		clt_corr_partial[tileY][tileX][0][3] = corrs[3]; // 4
		clt_corr_partial[tileY][tileX][1][0] = corrs[4]; // 5
		clt_corr_partial[tileY][tileX][1][1] = corrs[5]; // 6
		clt_corr_partial[tileY][tileX][1][2] = corr2d.debugStrip(strip_hor); // 7
		clt_corr_partial[tileY][tileX][1][3] = corr2d.debugStrip(strip_vert); // 8
		clt_corr_partial[tileY][tileX][1][2] = corrs[6]; // 5
		clt_corr_partial[tileY][tileX][1][3] = corrs[7]; // 6
		// clt_corr_partial[tileY][tileX][1][2] = corr2d.debugStrip(strip_hor); // 7
		// clt_corr_partial[tileY][tileX][1][3] = corr2d.debugStrip(strip_vert); // 8
		clt_corr_partial[tileY][tileX][2][0] = corr2d.debugStrip(strips[4]); // 9
		clt_corr_partial[tileY][tileX][2][1] = corr2d.debugStrip(strips[5]); // 10
		clt_corr_partial[tileY][tileX][2][2] = corr2d.debugStrip2(strip_hor); // 11

src/main/java/com/elphel/imagej/tileprocessor/QuadCLT.java

+2 −0

Original line number	Diff line number	Diff line
		@@ -677,6 +677,7 @@ public class QuadCLT extends QuadCLTCPU {
		0x0f); // int pairs_mask // selected pairs (0x3 - horizontal, 0xc - vertical, 0xf - quad, 0x30 - cross)

		quadCLT_main.getGPU().execCorr2D_normalize(
		true, // boolean combo, // normalize combo correlations (false - per-pair ones)
		fat_zero, // double fat_zero);
		clt_parameters.gpu_corr_rad); // int corr_radius

		@@ -803,6 +804,7 @@ public class QuadCLT extends QuadCLTCPU {
		0x30); // int pairs_mask // selected pairs (0x3 - horizontal, 0xc - vertical, 0xf - quad, 0x30 - cross)

		quadCLT_main.getGPU().execCorr2D_normalize(
		true, // boolean combo, // normalize combo correlations (false - per-pair ones)
		fat_zero, // double fat_zero);
		clt_parameters.gpu_corr_rad); // int corr_radius

Admin message