Links

Content Skeleton

This Page

Previous topic

Cuda Threads and Blocks

Next topic

Chroma Physics

Propagate

python side

096     @profile_if_possible
097     def propagate(self, gpu_geometry, rng_states, nthreads_per_block=64,
098                   max_blocks=1024, max_steps=10, use_weights=False,
099                   scatter_first=0):
...
110         nphotons = self.pos.size
111         step = 0
112         input_queue = np.empty(shape=nphotons+1, dtype=np.uint32)
113         input_queue[0] = 0
114         # Order photons initially in the queue to put the clones next to each other
115         for copy in xrange(self.ncopies):
116             input_queue[1+copy::self.ncopies] = np.arange(self.true_nphotons, dtype=np.uint32) + copy * self.true_nphotons
117         input_queue_gpu = ga.to_gpu(input_queue)
118         output_queue = np.zeros(shape=nphotons+1, dtype=np.uint32)
119         output_queue[0] = 1
120         output_queue_gpu = ga.to_gpu(output_queue)
121
122         while step < max_steps:
123             # Just finish the rest of the steps if the # of photons is low
124             if nphotons < nthreads_per_block * 16 * 8 or use_weights:
125                 nsteps = max_steps - step
126             else:
127                 nsteps = 1
128
129             for first_photon, photons_this_round, blocks in \
130                     chunk_iterator(nphotons, nthreads_per_block, max_blocks):
131                 self.gpu_funcs.propagate(
                             np.int32(first_photon),
                             np.int32(photons_this_round),
                             input_queue_gpu[1:],
                             output_queue_gpu,
                             rng_states,
                             self.pos, self.dir, self.wavelengths, self.pol, self.t, self.flags,
                             self.last_hit_triangles,
                             self.weights,
                             np.int32(nsteps),   ## CAUTION thats max_steps on cuda side
                             np.int32(use_weights),
                             np.int32(scatter_first),
                             gpu_geometry.gpudata,

                             block=(nthreads_per_block,1,1), grid=(blocks, 1))
132
133             step += nsteps
134             scatter_first = 0 # Only allow non-zero in first pass
135
136             if step < max_steps:
137                 temp = input_queue_gpu
138                 input_queue_gpu = output_queue_gpu
139                 output_queue_gpu = temp
140                 # Assign with a numpy array of length 1 to silence
141                 # warning from PyCUDA about setting array with different strides/storage orders.
142                 output_queue_gpu[:1].set(np.ones(shape=1, dtype=np.uint32))
143                 nphotons = input_queue_gpu[:1].get()[0] - 1

///         stick the surviving propagated photons in output_queue into input_queue

144
145         if ga.max(self.flags).get() & (1 << 31):
146             print >>sys.stderr, "WARNING: ABORTED PHOTONS"
147         cuda.Context.get_current().synchronize()

cuda entry points

simon:cuda blyth$ grep -l blockIdx *.*
bvh.cu
daq.cu
hybrid_render.cu
mesh.h
pdf.cu
propagate.cu
random.h
render.cu
tools.cu
transform.cu

cuda propagate

Entry point is propagate, communication via numpy arrays curtesy of pycuda.

  • self.flags on corresponds to histories array
  • id identifies the CUDA thread, corresponding to a single photon
  • photon parameters indexed into the arrays with photon_id
112 __global__ void
113 propagate(int first_photon, int nthreads, unsigned int *input_queue,
114       unsigned int *output_queue, curandState *rng_states,
115       float3 *positions, float3 *directions,
116       float *wavelengths, float3 *polarizations,
117       float *times, unsigned int *histories,
118       int *last_hit_triangles, float *weights,
119       int max_steps, int use_weights, int scatter_first,
120       Geometry *g)
121 {
122     __shared__ Geometry sg;
123
124     if (threadIdx.x == 0)
125     sg = *g;
//
// shared geometry between threads
//
126
127     __syncthreads();
128
129     int id = blockIdx.x*blockDim.x + threadIdx.x;
//
//  id points at the single photon to propagate in this parallel thread
//
130
131     if (id >= nthreads)
132     return;
133
134     g = &sg;
135
136     curandState rng = rng_states[id];
137
138     int photon_id = input_queue[first_photon + id];
139
140     Photon p;
141     p.position = positions[photon_id];
142     p.direction = directions[photon_id];
143     p.direction /= norm(p.direction);
144     p.polarization = polarizations[photon_id];
145     p.polarization /= norm(p.polarization);
146     p.wavelength = wavelengths[photon_id];
147     p.time = times[photon_id];
148     p.last_hit_triangle = last_hit_triangles[photon_id];
149     p.history = histories[photon_id];
150     p.weight = weights[photon_id];
151
152     if (p.history & (NO_HIT | BULK_ABSORB | SURFACE_DETECT | SURFACE_ABSORB | NAN_ABORT))
153     return;
154
155     State s;
156
157     int steps = 0;
158     while (steps < max_steps) {
159     steps++;
160
161     int command;
162
163     // check for NaN and fail
164     if (isnan(p.direction.x*p.direction.y*p.direction.z*p.position.x*p.position.y*p.position.z)) {
165         p.history |= NO_HIT | NAN_ABORT;
166         break;
167     }
168
169     fill_state(s, p, g);
170
171     if (p.last_hit_triangle == -1)
172         break;
173
174     command = propagate_to_boundary(p, s, rng, use_weights, scatter_first);
//
//      propagate_* only changes p (?) refering to state s
//
175     scatter_first = 0; // Only use the scatter_first value once
176
177     if (command == BREAK)
178         break;
179
180     if (command == CONTINUE)
181         continue;
182
183     if (s.surface_index != -1) {
184       command = propagate_at_surface(p, s, rng, g, use_weights);
185
186         if (command == BREAK)
187         break;
188
189         if (command == CONTINUE)
190         continue;
191     }
192
193     propagate_at_boundary(p, s, rng);
194
195     } // while (steps < max_steps)
196
197     rng_states[id] = rng;
198     positions[photon_id] = p.position;
199     directions[photon_id] = p.direction;
200     polarizations[photon_id] = p.polarization;
201     wavelengths[photon_id] = p.wavelength;
202     times[photon_id] = p.time;
203     histories[photon_id] = p.history;
204     last_hit_triangles[photon_id] = p.last_hit_triangle;
205     weights[photon_id] = p.weight;
206
207     // Not done, put photon in output queue
208     if ((p.history & (NO_HIT | BULK_ABSORB | SURFACE_DETECT | SURFACE_ABSORB | NAN_ABORT)) == 0) {
//
//       the photon lives on thanks to
//            RAYLEIGH_SCATTER REFLECT_DIFFUSE REFLECT_SPECULAR SURFACE_REEMIT SURFACE_TRANSMIT BULK_REEMIT
//
//
209     int out_idx = atomicAdd(output_queue, 1);
210     output_queue[out_idx] = photon_id;
//
//     http://supercomputingblog.com/cuda/cuda-tutorial-4-atomic-operations/
//
//         This atomicAdd function can be called within a kernel. When a thread executes this operation, a memory address is read,
//         has the value of val added to it, and the result is written back to memory.
//         The original value of the memory at location ?address? is returned to the thread.
//
211     }
212 } // propagate

chroma/cuda/photon.h

584 __device__ int
585 propagate_at_surface(Photon &p, State &s, curandState &rng, Geometry *geometry,
586                      bool use_weights=false)
587 {
588     Surface *surface = geometry->surfaces[s.surface_index];
589
590     if (surface->model == SURFACE_COMPLEX)
591         return propagate_complex(p, s, rng, surface, use_weights);
592     else if (surface->model == SURFACE_WLS)
593         return propagate_at_wls(p, s, rng, surface, use_weights);
594     else {
595         // use default surface model: do a combination of specular and
596         // diffuse reflection, detection, and absorption based on relative
597         // probabilties
  • chroma/doc/source/surface.rst