working on GPU code (f04bcc82) · Commits · Elphel / imagej-elphel

.gitignore

+2 −1

Original line number	Diff line number	Diff line
		@@ -7,3 +7,4 @@ NC393I
		attic
		*.log
		FOCUS-PSF*
		src/main/resources/trained_model
		No newline at end of file

src/main/java/GPUTileProcessor.java

+9 −10

Original line number	Diff line number	Diff line
		@@ -117,7 +117,7 @@ public class GPUTileProcessor {
		copyH2D.WidthInBytes = width_in_bytes;
		copyH2D.Height = height; // /4;

		// for copying results back to host
		// for copying results to host
		CUDA_MEMCPY2D copyD2H = new CUDA_MEMCPY2D();
		copyD2H.srcMemoryType = CUmemorytype.CU_MEMORYTYPE_DEVICE;
		copyD2H.srcDevice = dst_dpointer; // ((test & 1) ==0) ? src_dpointer : dst_dpointer; // copy same data
		@@ -130,8 +130,7 @@ public class GPUTileProcessor {
		copyD2H.WidthInBytes = width_in_bytes;
		copyD2H.Height = height; // /2;

		// Set up the kernel parameters: A pointer to an array
		// of pointers which point to the actual values.
		// kernel parameters: pointer to pointers
		Pointer kernelParameters = Pointer.to(
		Pointer.to(dst_dpointer),
		Pointer.to(src_dpointer),
		@@ -152,16 +151,16 @@ public class GPUTileProcessor {
		0, null, // Shared memory size and stream (shared - only dynamic, static is in code)
		kernelParameters, null); // Kernel- and extra parameters

		// Copy the data from the device back to the host
		// Copy the data from the device to the host
		cuMemcpy2D(copyD2H);
		// clean up
		cuMemFree(src_dpointer);
		cuMemFree(dst_dpointer);
		}

		public int setup() throws IOException // String arg, ImagePlus imagePlus)
		public int setup() throws IOException
		{

		// From code by Marco Hutter - http://www.jcuda.org
		// Enable exceptions and omit all subsequent error checks
		JCudaDriver.setExceptionsEnabled(true);
		JNvrtc.setExceptionsEnabled(true);
		@@ -174,7 +173,9 @@ public class GPUTileProcessor {
		cuCtxCreate(context, 0, device);

		// Obtain the CUDA source code from the CUDA file

		// Get absolute path to the file in resource foldder, then read it as a normal file.
		// When using just Eclipse resources - it does not notice that the file
		// was edited (happens frequently during kernel development).
		ClassLoader classLoader = getClass().getClassLoader();
		File file = new File(classLoader.getResource(GPU_KERNEL_FILE).getFile());
		System.out.println(file.getAbsolutePath());
		@@ -196,9 +197,7 @@ public class GPUTileProcessor {
		}

		/**
		* Create the CUDA function object for the kernel function with the
		* given name that is contained in the given source code
		*
		* Create the kernel function by its name in the source code
		* @param sourceCode The source code
		* @param kernelName The kernel function name
		* @return

src/main/java/ImageDtt.java

+786 −1

File changed.

Preview size limit exceeded, changes collapsed.

src/main/java/TwoQuadCLT.java

+115 −5

Original line number	Diff line number	Diff line
		@@ -20,6 +20,7 @@
		-----------------------------------------------------------------------------
		**
		*/
		import java.io.DataOutputStream;
		import java.io.File;
		import java.io.FileNotFoundException;
		import java.io.FileOutputStream;
		@@ -907,7 +908,80 @@ public class TwoQuadCLT {
		return results;
		}

		public void saveFloatKernels(String file_prefix,
		double [][][][][][] clt_kernels,
		double [][][] image_data,
		double [][][] port_xy,
		boolean transpose) throws IOException {
		if (clt_kernels != null) {
		for (int chn = 0; chn < clt_kernels.length; chn++) {
		String kern_path = file_prefix+"_chn"+chn+(transpose?"_transposed":"")+".kernel";
		String offs_path = file_prefix+"_chn"+chn+(transpose?"_transposed":"")+".kernel_offsets";
		FileOutputStream fos = new FileOutputStream(kern_path);
		DataOutputStream dos = new DataOutputStream(fos);
		for (int ty = 0; ty < clt_kernels[chn][0].length; ty++) {
		for (int tx = 0; tx < clt_kernels[chn][0][ty].length; tx++) {
		for (int col = 0; col < clt_kernels[chn].length; col++) {
		for (int p = 0; p < 4; p++) {
		double [] pa = clt_kernels[chn][col][ty][tx][p];
		for (int i0 = 0; i0 < 64; i0++) {
		int i;
		if (transpose) {
		i = ((i0 & 7) << 3) + ((i0 >>3) & 7);
		} else {
		i = i0;
		}
		dos.writeFloat((float)pa[i]);
		}
		}
		}
		}
		}
		dos.close();
		fos = new FileOutputStream(offs_path);
		dos = new DataOutputStream(fos);

		for (int ty = 0; ty < clt_kernels[chn][0].length; ty++) {
		for (int tx = 0; tx < clt_kernels[chn][0][ty].length; tx++) {
		for (int col = 0; col < clt_kernels[chn].length; col++) {
		double [] pa = clt_kernels[chn][col][ty][tx][4];
		for (int i = 0; i < pa.length; i++) {
		dos.writeFloat((float)pa[i]);
		}
		}
		}
		}

		dos.close();
		}
		}

		if (image_data != null) {
		for (int chn = 0; chn < image_data.length; chn++) {
		String img_path = file_prefix+"_chn"+chn+".bayer";
		FileOutputStream fos = new FileOutputStream(img_path);
		DataOutputStream dos = new DataOutputStream(fos);
		for (int i = 0; i < image_data[chn][0].length; i++) {
		dos.writeFloat((float) (image_data[chn][0][i] + image_data[chn][1][i] + image_data[chn][2][i]));
		}
		dos.close();
		}
		}
		if (port_xy != null) {
		for (int chn = 0; chn < port_xy[0].length; chn++) {
		String img_path = file_prefix+"_chn"+chn+".portsxy";
		FileOutputStream fos = new FileOutputStream(img_path);
		DataOutputStream dos = new DataOutputStream(fos);
		for (int i = 0; i < port_xy.length; i++) {
		dos.writeFloat((float) (port_xy[i][chn][0])); // x-offset
		dos.writeFloat((float) (port_xy[i][chn][1])); // y-offset
		}
		dos.close();
		}

		}

		}

		public ImagePlus [] processCLTQuadCorrPairGpu(
		GPUTileProcessor gPUTileProcessor,
		@@ -1003,6 +1077,8 @@ public class TwoQuadCLT {
		double [][][][][][] clt_kernels_main = quadCLT_main.getCLTKernels(); // [4][3][123][164]{[64],[64],[64],[64],[8]}
		double [][][][][][] clt_kernels_aux = quadCLT_aux.getCLTKernels();


		//[4][3][123][164][5][]
		double [][] dbg_kern = clt_kernels_main[0][0][0][0];
		// here all data is ready (images, kernels) to try GPU code

		@@ -1033,13 +1109,12 @@ public class TwoQuadCLT {
		"converted",
		dbg_titles);

		if (debugLevel < 1000) {
		return null;
		}

		double [][][] port_xy_main_dbg = new double [tilesX*tilesY][][];
		double [][][] port_xy_aux_dbg = new double [tilesX*tilesY][][];

		final double [][][][][][][] clt_bidata = // new double[2][quad][nChn][tilesY][tilesX][][]; // first index - main/aux
		image_dtt.clt_bi_quad (
		image_dtt.clt_bi_quad_dbg (
		clt_parameters, // final EyesisCorrectionParameters.CLTParameters clt_parameters,
		clt_parameters.fat_zero, // final double fatzero, // May use correlation fat zero from 2 different parameters - fat_zero and rig.ml_fatzero
		notch_mode, // final boolean notch_mode, // use notch filter for inter-camera correlation to detect poles
		@@ -1070,9 +1145,44 @@ public class TwoQuadCLT {
		// woi_tops, // final int [][] woi_tops,
		ers_delay, // final double [][][] ers_delay, // if not null - fill with tile center acquisition delay
		threadsMax, // final int threadsMax, // maximal number of threads to launch
		debugLevel); // final int globalDebugLevel);
		debugLevel, // final int globalDebugLevel);
		port_xy_main_dbg, // final double [][][] port_xy_main_dbg, // for each tile/port save x,y pixel coordinates (gpu code development)
		port_xy_aux_dbg); // final double [][][] port_xy_aux_dbg) // for each tile/port save x,y pixel coordinates (gpu code development)

		String kernel_dir = "/home/eyesis/workspace-python3/nvidia_dct8x8/clt/";
		boolean [][] what_to_save = {{false,false,true}, {false,false,true}};
		try {
		saveFloatKernels(
		kernel_dir +"main", // String file_prefix,
		(what_to_save[0][0]?clt_kernels_main:null), // double [][][][][][] clt_kernels, // null
		(what_to_save[0][1]?quadCLT_main.image_data:null),
		(what_to_save[0][2]?port_xy_main_dbg:null), // double [][][] port_xy,
		true);
		} catch (IOException e) {
		System.out.println("Failed to save flattened kernels tp "+kernel_dir);
		// TODO Auto-generated catch block
		e.printStackTrace();
		} // boolean transpose);

		try {
		saveFloatKernels(
		kernel_dir +"aux", // String file_prefix,
		(what_to_save[1][0]?clt_kernels_aux:null), // double [][][][][][] clt_kernels, // null
		(what_to_save[1][1]?quadCLT_aux.image_data:null),
		(what_to_save[1][2]?port_xy_aux_dbg:null), // double [][][] port_xy,
		true);
		} catch (IOException e) {
		System.out.println("Failed to save flattened kernels tp "+kernel_dir);
		// TODO Auto-generated catch block
		e.printStackTrace();
		} // boolean transpose);



		if (debugLevel < 1000) {
		return null;
		}

		if (ers_delay !=null) {
		showERSDelay(ers_delay);
		}

src/main/resources/dtt8x8.cuh

+1 −0

Original line number	Diff line number	Diff line
		@@ -43,6 +43,7 @@
		#define DTTTEST_BLK_STRIDE (DTTTEST_BLOCK_WIDTH+1)
		#define DTT_SIZE 8

		//#define CUDART_INF_F __int_as_float(0x7f800000)
		/*
		Python code to generate constant coefficients:
		def dct_constants():