summaryrefslogtreecommitdiffstats
path: root/cuda/3d/par3d_bp.cu
blob: 6f76d229c85abc1765feee4eb1f0e5b8f3b662ff (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
/*
-----------------------------------------------------------------------
Copyright: 2010-2018, imec Vision Lab, University of Antwerp
           2014-2018, CWI, Amsterdam

Contact: astra@astra-toolbox.com
Website: http://www.astra-toolbox.com/

This file is part of the ASTRA Toolbox.


The ASTRA Toolbox is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

The ASTRA Toolbox is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.

-----------------------------------------------------------------------
*/

#include <cstdio>
#include <cassert>
#include <iostream>
#include <list>

#include <cuda.h>
#include "util3d.h"

#ifdef STANDALONE
#include "par3d_fp.h"
#include "testutil.h"
#endif

#include "dims3d.h"

typedef texture<float, 3, cudaReadModeElementType> texture3D;

static texture3D gT_par3DProjTexture;

namespace astraCUDA3d {

#define ZSIZE 6
static const unsigned int g_volBlockZ = ZSIZE;

static const unsigned int g_anglesPerBlock = 32;
static const unsigned int g_volBlockX = 16;
static const unsigned int g_volBlockY = 32;

static const unsigned g_MaxAngles = 1024;

__constant__ float gC_C[8*g_MaxAngles];


static bool bindProjDataTexture(const cudaArray* array)
{
	cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>();

	gT_par3DProjTexture.addressMode[0] = cudaAddressModeBorder;
	gT_par3DProjTexture.addressMode[1] = cudaAddressModeBorder;
	gT_par3DProjTexture.addressMode[2] = cudaAddressModeBorder;
	gT_par3DProjTexture.filterMode = cudaFilterModeLinear;
	gT_par3DProjTexture.normalized = false;

	cudaBindTextureToArray(gT_par3DProjTexture, array, channelDesc);

	// TODO: error value?

	return true;
}


__global__ void dev_par3D_BP(void* D_volData, unsigned int volPitch, int startAngle, int angleOffset, const SDimensions3D dims, float fOutputScale)
{
	float* volData = (float*)D_volData;

	int endAngle = startAngle + g_anglesPerBlock;
	if (endAngle > dims.iProjAngles - angleOffset)
		endAngle = dims.iProjAngles - angleOffset;

	// threadIdx: x = rel x
	//            y = rel y

	// blockIdx:  x = x + y
	//            y = z


	const int X = blockIdx.x % ((dims.iVolX+g_volBlockX-1)/g_volBlockX) * g_volBlockX + threadIdx.x;
	const int Y = blockIdx.x / ((dims.iVolX+g_volBlockX-1)/g_volBlockX) * g_volBlockY + threadIdx.y;

	if (X >= dims.iVolX)
		return;
	if (Y >= dims.iVolY)
		return;

	const int startZ = blockIdx.y * g_volBlockZ;

	float fX = X - 0.5f*dims.iVolX + 0.5f;
	float fY = Y - 0.5f*dims.iVolY + 0.5f;
	float fZ = startZ - 0.5f*dims.iVolZ + 0.5f;

	float Z[ZSIZE];
	for(int i=0; i < ZSIZE; i++)
		Z[i] = 0.0f;

	{
		float fAngle = startAngle + angleOffset + 0.5f;

		for (int angle = startAngle; angle < endAngle; ++angle, fAngle += 1.0f)
		{

			float4 fCu = make_float4(gC_C[8*angle+0], gC_C[8*angle+1], gC_C[8*angle+2], gC_C[8*angle+3]);
			float4 fCv = make_float4(gC_C[8*angle+4], gC_C[8*angle+5], gC_C[8*angle+6], gC_C[8*angle+7]);

			float fU = fCu.w + fX * fCu.x + fY * fCu.y + fZ * fCu.z;
			float fV = fCv.w + fX * fCv.x + fY * fCv.y + fZ * fCv.z;

			for (int idx = 0; idx < ZSIZE; ++idx) {

				float fVal = tex3D(gT_par3DProjTexture, fU, fAngle, fV);
				Z[idx] += fVal;

				fU += fCu.z;
				fV += fCv.z;
			}

		}
	}

	int endZ = ZSIZE;
	if (endZ > dims.iVolZ - startZ)
		endZ = dims.iVolZ - startZ;

	for(int i=0; i < endZ; i++)
		volData[((startZ+i)*dims.iVolY+Y)*volPitch+X] += Z[i] * fOutputScale;
}

// supersampling version
__global__ void dev_par3D_BP_SS(void* D_volData, unsigned int volPitch, int startAngle, int angleOffset, const SDimensions3D dims, int iRaysPerVoxelDim, float fOutputScale)
{
	float* volData = (float*)D_volData;

	int endAngle = startAngle + g_anglesPerBlock;
	if (endAngle > dims.iProjAngles - angleOffset)
		endAngle = dims.iProjAngles - angleOffset;

	// threadIdx: x = rel x
	//            y = rel y

	// blockIdx:  x = x + y
    //            y = z


	// TO TRY: precompute part of detector intersection formulas in shared mem?
	// TO TRY: inner loop over z, gather ray values in shared mem

	const int X = blockIdx.x % ((dims.iVolX+g_volBlockX-1)/g_volBlockX) * g_volBlockX + threadIdx.x;
	const int Y = blockIdx.x / ((dims.iVolX+g_volBlockX-1)/g_volBlockX) * g_volBlockY + threadIdx.y;

	if (X >= dims.iVolX)
		return;
	if (Y >= dims.iVolY)
		return;

	const int startZ = blockIdx.y * g_volBlockZ;
	int endZ = startZ + g_volBlockZ;
	if (endZ > dims.iVolZ)
		endZ = dims.iVolZ;

	float fX = X - 0.5f*dims.iVolX + 0.5f - 0.5f + 0.5f/iRaysPerVoxelDim;
	float fY = Y - 0.5f*dims.iVolY + 0.5f - 0.5f + 0.5f/iRaysPerVoxelDim;
	float fZ = startZ - 0.5f*dims.iVolZ + 0.5f - 0.5f + 0.5f/iRaysPerVoxelDim;

	const float fSubStep = 1.0f/iRaysPerVoxelDim;

	fOutputScale /= (iRaysPerVoxelDim*iRaysPerVoxelDim*iRaysPerVoxelDim);


	for (int Z = startZ; Z < endZ; ++Z, fZ += 1.0f)
	{

		float fVal = 0.0f;
		float fAngle = startAngle + angleOffset + 0.5f;

		for (int angle = startAngle; angle < endAngle; ++angle, fAngle += 1.0f)
		{
			const float fCux = gC_C[8*angle+0];
			const float fCuy = gC_C[8*angle+1];
			const float fCuz = gC_C[8*angle+2];
			const float fCuc = gC_C[8*angle+3];
			const float fCvx = gC_C[8*angle+4];
			const float fCvy = gC_C[8*angle+5];
			const float fCvz = gC_C[8*angle+6];
			const float fCvc = gC_C[8*angle+7];

			float fXs = fX;
			for (int iSubX = 0; iSubX < iRaysPerVoxelDim; ++iSubX) {
			float fYs = fY;
			for (int iSubY = 0; iSubY < iRaysPerVoxelDim; ++iSubY) {
			float fZs = fZ;
			for (int iSubZ = 0; iSubZ < iRaysPerVoxelDim; ++iSubZ) {

				const float fU = fCuc + fXs * fCux + fYs * fCuy + fZs * fCuz;
				const float fV = fCvc + fXs * fCvx + fYs * fCvy + fZs * fCvz;

				fVal += tex3D(gT_par3DProjTexture, fU, fAngle, fV);
				fZs += fSubStep;
			}
			fYs += fSubStep;
			}
			fXs += fSubStep;
			}

		}

		volData[(Z*dims.iVolY+Y)*volPitch+X] += fVal * fOutputScale;
	}

}

bool Par3DBP_Array(cudaPitchedPtr D_volumeData,
                   cudaArray *D_projArray,
                   const SDimensions3D& dims, const SPar3DProjection* angles,
                   const SProjectorParams3D& params)
{
	bindProjDataTexture(D_projArray);

	float fOutputScale = params.fOutputScale * params.fVolScaleX * params.fVolScaleY * params.fVolScaleZ;

	for (unsigned int th = 0; th < dims.iProjAngles; th += g_MaxAngles) {
		unsigned int angleCount = g_MaxAngles;
		if (th + angleCount > dims.iProjAngles)
			angleCount = dims.iProjAngles - th;

		// transfer angles to constant memory
		float* tmp = new float[8*dims.iProjAngles];

		// NB: We increment angles at the end of the loop body.


		// TODO: Use functions from dims3d.cu for this:

#define TRANSFER_TO_CONSTANT(expr,name) do { for (unsigned int i = 0; i < angleCount; ++i) tmp[8*i + name] = (expr) ; } while (0)

#define DENOM (angles[i].fRayX*angles[i].fDetUY*angles[i].fDetVZ - angles[i].fRayX*angles[i].fDetUZ*angles[i].fDetVY - angles[i].fRayY*angles[i].fDetUX*angles[i].fDetVZ + angles[i].fRayY*angles[i].fDetUZ*angles[i].fDetVX + angles[i].fRayZ*angles[i].fDetUX*angles[i].fDetVY - angles[i].fRayZ*angles[i].fDetUY*angles[i].fDetVX)

		TRANSFER_TO_CONSTANT( ( - (angles[i].fRayY*angles[i].fDetVZ - angles[i].fRayZ*angles[i].fDetVY)) / DENOM , 0 );
		TRANSFER_TO_CONSTANT( ( (angles[i].fRayX*angles[i].fDetVZ - angles[i].fRayZ*angles[i].fDetVX)) / DENOM , 1 );
		TRANSFER_TO_CONSTANT( (- (angles[i].fRayX*angles[i].fDetVY - angles[i].fRayY*angles[i].fDetVX) ) / DENOM , 2 );
		TRANSFER_TO_CONSTANT( (-(angles[i].fDetSY*angles[i].fDetVZ - angles[i].fDetSZ*angles[i].fDetVY)*angles[i].fRayX + (angles[i].fRayY*angles[i].fDetVZ - angles[i].fRayZ*angles[i].fDetVY)*angles[i].fDetSX - (angles[i].fRayY*angles[i].fDetSZ - angles[i].fRayZ*angles[i].fDetSY)*angles[i].fDetVX) / DENOM , 3 );

		TRANSFER_TO_CONSTANT( ((angles[i].fRayY*angles[i].fDetUZ - angles[i].fRayZ*angles[i].fDetUY) ) / DENOM , 4 );
		TRANSFER_TO_CONSTANT( (- (angles[i].fRayX*angles[i].fDetUZ - angles[i].fRayZ*angles[i].fDetUX) ) / DENOM , 5 );
		TRANSFER_TO_CONSTANT( ((angles[i].fRayX*angles[i].fDetUY - angles[i].fRayY*angles[i].fDetUX) ) / DENOM , 6 );
		TRANSFER_TO_CONSTANT( ((angles[i].fDetSY*angles[i].fDetUZ - angles[i].fDetSZ*angles[i].fDetUY)*angles[i].fRayX - (angles[i].fRayY*angles[i].fDetUZ - angles[i].fRayZ*angles[i].fDetUY)*angles[i].fDetSX + (angles[i].fRayY*angles[i].fDetSZ - angles[i].fRayZ*angles[i].fDetSY)*angles[i].fDetUX ) / DENOM , 7 );

#undef TRANSFER_TO_CONSTANT
#undef DENOM
		cudaMemcpyToSymbol(gC_C, tmp, angleCount*8*sizeof(float), 0, cudaMemcpyHostToDevice); 

		delete[] tmp;

		dim3 dimBlock(g_volBlockX, g_volBlockY);

		dim3 dimGrid(((dims.iVolX+g_volBlockX-1)/g_volBlockX)*((dims.iVolY+g_volBlockY-1)/g_volBlockY), (dims.iVolZ+g_volBlockZ-1)/g_volBlockZ);

		// timeval t;
		// tic(t);

		for (unsigned int i = 0; i < angleCount; i += g_anglesPerBlock) {
			// printf("Calling BP: %d, %dx%d, %dx%d to %p\n", i, dimBlock.x, dimBlock.y, dimGrid.x, dimGrid.y, (void*)D_volumeData.ptr); 
			if (params.iRaysPerVoxelDim == 1)
				dev_par3D_BP<<<dimGrid, dimBlock>>>(D_volumeData.ptr, D_volumeData.pitch/sizeof(float), i, th, dims, fOutputScale);
			else
				dev_par3D_BP_SS<<<dimGrid, dimBlock>>>(D_volumeData.ptr, D_volumeData.pitch/sizeof(float), i, th, dims, params.iRaysPerVoxelDim, fOutputScale);
		}

		cudaTextForceKernelsCompletion();

		angles = angles + angleCount;
		// printf("%f\n", toc(t));

	}

	return true;
}

bool Par3DBP(cudaPitchedPtr D_volumeData,
            cudaPitchedPtr D_projData,
            const SDimensions3D& dims, const SPar3DProjection* angles,
            const SProjectorParams3D& params)
{
	// transfer projections to array

	cudaArray* cuArray = allocateProjectionArray(dims);
	transferProjectionsToArray(D_projData, cuArray, dims);

	bool ret = Par3DBP_Array(D_volumeData, cuArray, dims, angles, params);

	cudaFreeArray(cuArray);

	return ret;
}


}

#ifdef STANDALONE
int main()
{
	SDimensions3D dims;
	dims.iVolX = 256;
	dims.iVolY = 256;
	dims.iVolZ = 256;
	dims.iProjAngles = 180;
	dims.iProjU = 512;
	dims.iProjV = 512;
	dims.iRaysPerDet = 1;

	cudaExtent extentV;
	extentV.width = dims.iVolX*sizeof(float);
	extentV.height = dims.iVolY;
	extentV.depth = dims.iVolZ;

	cudaPitchedPtr volData; // pitch, ptr, xsize, ysize

	cudaMalloc3D(&volData, extentV);

	cudaExtent extentP;
	extentP.width = dims.iProjU*sizeof(float);
	extentP.height = dims.iProjAngles;
	extentP.depth = dims.iProjV;

	cudaPitchedPtr projData; // pitch, ptr, xsize, ysize

	cudaMalloc3D(&projData, extentP);
	cudaMemset3D(projData, 0, extentP);

	float* slice = new float[256*256];
	cudaPitchedPtr ptr;
	ptr.ptr = slice;
	ptr.pitch = 256*sizeof(float);
	ptr.xsize = 256*sizeof(float);
	ptr.ysize = 256;

	for (unsigned int i = 0; i < 256*256; ++i)
		slice[i] = 1.0f;
	for (unsigned int i = 0; i < 256; ++i) {
		cudaExtent extentS;
		extentS.width = dims.iVolX*sizeof(float);
		extentS.height = dims.iVolY;
		extentS.depth = 1;
		cudaPos sp = { 0, 0, 0 };
		cudaPos dp = { 0, 0, i };
		cudaMemcpy3DParms p;
		p.srcArray = 0;
		p.srcPos = sp;
		p.srcPtr = ptr;
		p.dstArray = 0;
		p.dstPos = dp;
		p.dstPtr = volData;
		p.extent = extentS;
		p.kind = cudaMemcpyHostToDevice;
		cudaMemcpy3D(&p);
#if 0
		if (i == 128) {
			for (unsigned int j = 0; j < 256*256; ++j)
				slice[j] = 0.0f;
		}
#endif 
	}


	SPar3DProjection angle[180];
	angle[0].fRayX = 1;
	angle[0].fRayY = 0;
	angle[0].fRayZ = 0;

	angle[0].fDetSX = 512;
	angle[0].fDetSY = -256;
	angle[0].fDetSZ = -256;

	angle[0].fDetUX = 0;
	angle[0].fDetUY = 1;
	angle[0].fDetUZ = 0;

	angle[0].fDetVX = 0;
	angle[0].fDetVY = 0;
	angle[0].fDetVZ = 1;

#define ROTATE0(name,i,alpha) do { angle[i].f##name##X = angle[0].f##name##X * cos(alpha) - angle[0].f##name##Y * sin(alpha); angle[i].f##name##Y = angle[0].f##name##X * sin(alpha) + angle[0].f##name##Y * cos(alpha); } while(0)
	for (int i = 1; i < 180; ++i) {
		angle[i] = angle[0];
		ROTATE0(Ray, i, i*2*M_PI/180);
		ROTATE0(DetS, i, i*2*M_PI/180);
		ROTATE0(DetU, i, i*2*M_PI/180);
		ROTATE0(DetV, i, i*2*M_PI/180);
	}
#undef ROTATE0

	astraCUDA3d::Par3DFP(volData, projData, dims, angle, 1.0f);
#if 1
	float* bufs = new float[180*512];

	for (int i = 0; i < 512; ++i) {
		cudaMemcpy(bufs, ((float*)projData.ptr)+180*512*i, 180*512*sizeof(float), cudaMemcpyDeviceToHost);

		printf("%d %d %d\n", projData.pitch, projData.xsize, projData.ysize);

		char fname[20];
		sprintf(fname, "sino%03d.png", i);
		saveImage(fname, 180, 512, bufs, 0, 512);
	}

	float* bufp = new float[512*512];

	for (int i = 0; i < 180; ++i) {
		for (int j = 0; j < 512; ++j) {
			cudaMemcpy(bufp+512*j, ((float*)projData.ptr)+180*512*j+512*i, 512*sizeof(float), cudaMemcpyDeviceToHost);
		}

		char fname[20];
		sprintf(fname, "proj%03d.png", i);
		saveImage(fname, 512, 512, bufp, 0, 512);
	}
#endif		
	for (unsigned int i = 0; i < 256*256; ++i)
		slice[i] = 0.0f;
	for (unsigned int i = 0; i < 256; ++i) {
		cudaExtent extentS;
		extentS.width = dims.iVolX*sizeof(float);
		extentS.height = dims.iVolY;
		extentS.depth = 1;
		cudaPos sp = { 0, 0, 0 };
		cudaPos dp = { 0, 0, i };
		cudaMemcpy3DParms p;
		p.srcArray = 0;
		p.srcPos = sp;
		p.srcPtr = ptr;
		p.dstArray = 0;
		p.dstPos = dp;
		p.dstPtr = volData;
		p.extent = extentS;
		p.kind = cudaMemcpyHostToDevice;
		cudaMemcpy3D(&p);
	}

	astraCUDA3d::Par3DBP(volData, projData, dims, angle, 1.0f);
#if 1
	float* buf = new float[256*256];

	for (int i = 0; i < 256; ++i) {
		cudaMemcpy(buf, ((float*)volData.ptr)+256*256*i, 256*256*sizeof(float), cudaMemcpyDeviceToHost);

		printf("%d %d %d\n", volData.pitch, volData.xsize, volData.ysize);

		char fname[20];
		sprintf(fname, "vol%03d.png", i);
		saveImage(fname, 256, 256, buf, 0, 60000);
	}
#endif

}
#endif