GatedSpectrometer.cu 26.3 KB
Newer Older
1
#include "psrdada_cpp/effelsberg/edd/GatedSpectrometer.cuh"
2
#include "psrdada_cpp/effelsberg/edd/Tools.cuh"
3
4
5
#include "psrdada_cpp/common.hpp"
#include "psrdada_cpp/cuda_utils.hpp"
#include "psrdada_cpp/raw_bytes.hpp"
6

7
#include <cuda.h>
8
#include <cuda_profiler_api.h>
9
#include <thrust/system/cuda/execution_policy.h>
10
11

#include <iostream>
12
#include <iomanip>
13
14
#include <cstring>
#include <sstream>
15
16
17
18
19

namespace psrdada_cpp {
namespace effelsberg {
namespace edd {

20
// Reduce thread local vatiable v in shared array x, so that x[0]
21
template<typename T>
22
__device__ void sum_reduce(T *x, const T &v)
23
24
25
26
27
28
29
30
31
{
  x[threadIdx.x] = v;
  __syncthreads();
  for(int s = blockDim.x / 2; s > 0; s = s / 2)
  {
    if (threadIdx.x < s)
      x[threadIdx.x] += x[threadIdx.x + s];
    __syncthreads();
  }
32
}
33
34


35
36
37
38
39
40
41
42
43
44
45
// If one of the side channel items is lsot, then both are considered as lost
// here
__global__ void mergeSideChannels(uint64_t* __restrict__ A, uint64_t* __restrict__ B, size_t N)
{
  for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; (i < N);
       i += blockDim.x * gridDim.x)
  {
    uint64_t v = A[i] || B[i];
    A[i] = v;
    B[i] = v;
  }
46
47
48
}


49
50
51
52
53
__global__ void gating(float* __restrict__ G0,
        float* __restrict__ G1,
        const uint64_t* __restrict__ sideChannelData,
        size_t N, size_t heapSize, size_t bitpos,
        size_t noOfSideChannels, size_t selectedSideChannel,
54
55
        const float*  __restrict__ _baseLineG0,
        const float*  __restrict__ _baseLineG1,
56
57
58
        float* __restrict__ baseLineNG0,
        float* __restrict__ baseLineNG1,
        uint64_cu* stats_G0, uint64_cu* stats_G1) {
59
  // statistics values for samopels to G0, G1
60
61
62
  uint32_t _G0stats = 0;
  uint32_t _G1stats = 0;

63
64
65
  const float baseLineG0 = _baseLineG0[0];
  const float baseLineG1 = _baseLineG1[0];

66
67
68
  float baselineUpdateG0 = 0;
  float baselineUpdateG1 = 0;

Tobias Winchen's avatar
Tobias Winchen committed
69
  for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; (i < N);
70
       i += blockDim.x * gridDim.x) {
71
72
    const float v = G0[i];

73
74
    const uint64_t sideChannelItem = sideChannelData[((i / heapSize) * (noOfSideChannels)) +
                        selectedSideChannel];
75

76
77
    const unsigned int bit_set = TEST_BIT(sideChannelItem, bitpos);
    const unsigned int heap_lost = TEST_BIT(sideChannelItem, 63);
78
79
80
    G1[i] = (v - baseLineG1) * bit_set * (!heap_lost) + baseLineG1;
    G0[i] = (v - baseLineG0) * (!bit_set) *(!heap_lost) + baseLineG0;

81
82
    _G0stats += (!bit_set) *(!heap_lost);
    _G1stats += bit_set * (!heap_lost);
83
84
85

    baselineUpdateG1 += v * bit_set * (!heap_lost);
    baselineUpdateG0 += v * (!bit_set) *(!heap_lost);
86
  }
87

88
89
90
  __shared__ uint32_t x[1024];

  // Reduce G0, G1
91
  sum_reduce<uint32_t>(x, _G0stats);
92
  if(threadIdx.x == 0) {
93
    atomicAdd(stats_G0,  (uint64_cu) x[threadIdx.x]);
94
  }
95
  __syncthreads();
96
97

  sum_reduce<uint32_t>(x, _G1stats);
98
  if(threadIdx.x == 0) {
99
    atomicAdd(stats_G1,  (uint64_cu) x[threadIdx.x]);
100
  }
101
  __syncthreads();
102

103
104
105
  //reuse shared array
  float *y = (float*) x;
  //update the baseline array
106
  sum_reduce<float>(y, baselineUpdateG0);
107
  if(threadIdx.x == 0) {
108
    atomicAdd(baseLineNG0, y[threadIdx.x]);
109
  }
Tobias Winchen's avatar
Tobias Winchen committed
110
  __syncthreads();
111
112

  sum_reduce<float>(y, baselineUpdateG1);
113
  if(threadIdx.x == 0) {
114
    atomicAdd(baseLineNG1, y[threadIdx.x]);
115
  }
116
  __syncthreads();
Tobias Winchen's avatar
Tobias Winchen committed
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
// Updates the baselines of the gates for the polarization set for the next
// block
// only few output blocks per input block thus execution on only one thread.
// Important is that the execution is async on the GPU.
__global__ void update_baselines(float*  __restrict__ baseLineG0,
        float*  __restrict__ baseLineG1,
        float* __restrict__ baseLineNG0,
        float* __restrict__ baseLineNG1,
        uint64_cu* stats_G0, uint64_cu* stats_G1,
        size_t N)
{
    size_t NG0 = 0;
    size_t NG1 = 0;

    for (size_t i =0; i < N; i++)
    {
       NG0 += stats_G0[i];
       NG1 += stats_G1[i];
    }

    baseLineG0[0] = baseLineNG0[0] / NG0;
    baseLineG1[0] = baseLineNG1[0] / NG1;
    baseLineNG0[0] = 0;
    baseLineNG1[0] = 0;
}





151
152
template <class HandlerType>
GatedSpectrometer<HandlerType>::GatedSpectrometer(
153
154
    const DadaBufferLayout &dadaBufferLayout,
    std::size_t selectedSideChannel, std::size_t selectedBit, std::size_t fft_length, std::size_t naccumulate,
155
    std::size_t nbits, float input_level, float output_level,
156
    HandlerType &handler) : _dadaBufferLayout(dadaBufferLayout),
157
      _selectedSideChannel(selectedSideChannel), _selectedBit(selectedBit),
158
      _fft_length(fft_length),
159
      _naccumulate(naccumulate), _nbits(nbits), _handler(handler), _fft_plan(0),
160
      _call_count(0), _nsamps_per_heap(4096), _processing_efficiency(0.){
161
162

  // Sanity checks
163
  assert(((_nbits == 12) || (_nbits == 8)));
164
165
166
167
168
  assert(_naccumulate > 0);

  // check for any device errors
  CUDA_ERROR_CHECK(cudaDeviceSynchronize());

169
  BOOST_LOG_TRIVIAL(info)
170
      << "Creating new GatedSpectrometer instance with parameters: \n"
171
172
      << "  fft_length           " << _fft_length << "\n"
      << "  naccumulate          " << _naccumulate << "\n"
173
174
      << "  nSideChannels        " << _dadaBufferLayout.getNSideChannels() << "\n"
      << "  speadHeapSize        " << _dadaBufferLayout.getHeapSize() << " byte\n"
175
176
177
      << "  selectedSideChannel  " << _selectedSideChannel << "\n"
      << "  selectedBit          " << _selectedBit << "\n"
      << "  output bit depth     " << sizeof(IntegratedPowerType) * 8;
178

179
180
  assert((_dadaBufferLayout.getNSideChannels() == 0) ||
         (selectedSideChannel < _dadaBufferLayout.getNSideChannels()));  // Sanity check of side channel value
181
182
  assert(selectedBit < 64); // Sanity check of selected bit

183
   _nsamps_per_buffer = _dadaBufferLayout.sizeOfData() * 8 / nbits;
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202

  _nsamps_per_output_spectra = fft_length * naccumulate;
  int nBlocks;
  if (_nsamps_per_output_spectra <= _nsamps_per_buffer)
  { // one buffer block is used for one or multiple output spectra
    size_t N = _nsamps_per_buffer / _nsamps_per_output_spectra;
    // All data in one block has to be used
    assert(N * _nsamps_per_output_spectra == _nsamps_per_buffer);
    nBlocks = 1;
  }
  else
  { // multiple blocks are integrated intoone output
    size_t N =  _nsamps_per_output_spectra /  _nsamps_per_buffer;
    // All data in multiple blocks has to be used
    assert(N * _nsamps_per_buffer == _nsamps_per_output_spectra);
    nBlocks = N;
  }
  BOOST_LOG_TRIVIAL(debug) << "Integrating  " << _nsamps_per_output_spectra << " samples from " << nBlocks << " into one spectra.";

203
  _nchans = _fft_length / 2 + 1;
204
  int batch = _nsamps_per_buffer / _fft_length;
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
  float dof = 2 * _naccumulate;
  float scale =
      std::pow(input_level * std::sqrt(static_cast<float>(_nchans)), 2);
  float offset = scale * dof;
  float scaling = scale * std::sqrt(2 * dof) / output_level;
  BOOST_LOG_TRIVIAL(debug)
      << "Correction factors for 8-bit conversion: offset = " << offset
      << ", scaling = " << scaling;

  BOOST_LOG_TRIVIAL(debug) << "Generating FFT plan";
  int n[] = {static_cast<int>(_fft_length)};
  CUFFT_ERROR_CHECK(cufftPlanMany(&_fft_plan, 1, n, NULL, 1, _fft_length, NULL,
                                  1, _nchans, CUFFT_R2C, batch));
  cufftSetStream(_fft_plan, _proc_stream);

  BOOST_LOG_TRIVIAL(debug) << "Allocating memory";
221
222
223
224
  polarization0._raw_voltage.resize(_dadaBufferLayout.sizeOfData() / sizeof(uint64_t));
  polarization1._raw_voltage.resize(_dadaBufferLayout.sizeOfData() / sizeof(uint64_t));
  polarization0._sideChannelData.resize(_dadaBufferLayout.getNSideChannels() * _dadaBufferLayout.getNHeaps());
  polarization1._sideChannelData.resize(_dadaBufferLayout.getNSideChannels() * _dadaBufferLayout.getNHeaps());
225
  BOOST_LOG_TRIVIAL(debug) << "  Input voltages size (in 64-bit words): "
226
                           << polarization0._raw_voltage.size();
227
228
  _unpacked_voltage_G0.resize(_nsamps_per_buffer);
  _unpacked_voltage_G1.resize(_nsamps_per_buffer);
Tobias Winchen's avatar
Tobias Winchen committed
229

230
231
232
233
234
235
236
237
    polarization0._baseLineG0.resize(1);
    polarization0._baseLineG0_update.resize(1);
    polarization0._baseLineG1.resize(1);
    polarization0._baseLineG1_update.resize(1);
    polarization1._baseLineG0.resize(1);
    polarization1._baseLineG0_update.resize(1);
    polarization1._baseLineG1.resize(1);
    polarization1._baseLineG1_update.resize(1);
238

239
240
  BOOST_LOG_TRIVIAL(debug) << "  Unpacked voltages size (in samples): "
                           << _unpacked_voltage_G0.size();
241
242
243
244
  polarization0._channelised_voltage_G0.resize(_nchans * batch);
  polarization0._channelised_voltage_G1.resize(_nchans * batch);
  polarization1._channelised_voltage_G0.resize(_nchans * batch);
  polarization1._channelised_voltage_G1.resize(_nchans * batch);
245
  BOOST_LOG_TRIVIAL(debug) << "  Channelised voltages size: "
246
247
248
249
250
251
252
                           << polarization0._channelised_voltage_G0.size();

   stokes_G0.resize(_nchans, batch / (_naccumulate / nBlocks));
   stokes_G1.resize(_nchans, batch / (_naccumulate / nBlocks));

  // on the host full output is stored together with sci data in one buffer
  _host_power_db.resize( 8 * (_nchans * sizeof(IntegratedPowerType) + sizeof(size_t)) * batch / (_naccumulate / nBlocks));
253
254
255
256
257
258
259
260
261
262

  CUDA_ERROR_CHECK(cudaStreamCreate(&_h2d_stream));
  CUDA_ERROR_CHECK(cudaStreamCreate(&_proc_stream));
  CUDA_ERROR_CHECK(cudaStreamCreate(&_d2h_stream));
  CUFFT_ERROR_CHECK(cufftSetStream(_fft_plan, _proc_stream));

  _unpacker.reset(new Unpacker(_proc_stream));
} // constructor


263
264
265

template <class HandlerType>
GatedSpectrometer<HandlerType>::~GatedSpectrometer() {
266
267
268
269
270
271
272
273
274
  BOOST_LOG_TRIVIAL(debug) << "Destroying GatedSpectrometer";
  if (!_fft_plan)
    cufftDestroy(_fft_plan);
  cudaStreamDestroy(_h2d_stream);
  cudaStreamDestroy(_proc_stream);
  cudaStreamDestroy(_d2h_stream);
}


275
276
277

template <class HandlerType>
void GatedSpectrometer<HandlerType>::init(RawBytes &block) {
278
  BOOST_LOG_TRIVIAL(debug) << "GatedSpectrometer init called";
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
  std::stringstream headerInfo;
  headerInfo << "\n"
      << "# Gated spectrometer parameters: \n"
      << "fft_length               " << _fft_length << "\n"
      << "nchannels                " << _fft_length << "\n"
      << "naccumulate              " << _naccumulate << "\n"
      << "selected_side_channel    " << _selectedSideChannel << "\n"
      << "selected_bit             " << _selectedBit << "\n"
      << "output_bit_depth         " << sizeof(IntegratedPowerType) * 8;

  size_t bEnd = std::strlen(block.ptr());
  if (bEnd + headerInfo.str().size() < block.total_bytes())
  {
    std::strcpy(block.ptr() + bEnd, headerInfo.str().c_str());
  }
  else
  {
    BOOST_LOG_TRIVIAL(warning) << "Header of size " << block.total_bytes()
      << " bytes already contains " << bEnd
      << "bytes. Cannot add gated spectrometer info of size "
      << headerInfo.str().size() << " bytes.";
  }

302
303
304
305
  _handler.init(block);
}


306
307
308

template <class HandlerType>
void GatedSpectrometer<HandlerType>::gated_fft(
309
        PolarizationData &data,
310
311
312
313
  thrust::device_vector<uint64_cu> &_noOfBitSetsIn_G0,
  thrust::device_vector<uint64_cu> &_noOfBitSetsIn_G1
        )
{
314
315
316
  BOOST_LOG_TRIVIAL(debug) << "Unpacking raw voltages";
  switch (_nbits) {
  case 8:
317
    _unpacker->unpack<8>(data._raw_voltage.b(), _unpacked_voltage_G0);
318
319
    break;
  case 12:
320
    _unpacker->unpack<12>(data._raw_voltage.b(), _unpacked_voltage_G0);
321
322
323
324
    break;
  default:
    throw std::runtime_error("Unsupported number of bits");
  }
325

326
  // Loop over outputblocks, for case of multiple output blocks per input block
327
  int step = data._sideChannelData.b().size() / _noOfBitSetsIn_G0.size();
328

329
  for (size_t i = 0; i < _noOfBitSetsIn_G0.size(); i++)
330
  { // ToDo: Should be in one kernel call
331
  gating<<<1024, 1024, 0, _proc_stream>>>(
332
333
334
335
336
337
338
      thrust::raw_pointer_cast(_unpacked_voltage_G0.data() + i * step * _nsamps_per_heap),
      thrust::raw_pointer_cast(_unpacked_voltage_G1.data() + i * step * _nsamps_per_heap),
      thrust::raw_pointer_cast(data._sideChannelData.b().data() + i * step),
      _unpacked_voltage_G0.size() / _noOfBitSetsIn_G0.size(),
      _dadaBufferLayout.getHeapSize(),
      _selectedBit,
      _dadaBufferLayout.getNSideChannels(),
339
      _selectedSideChannel,
340
341
      thrust::raw_pointer_cast(data._baseLineG0.data()),
      thrust::raw_pointer_cast(data._baseLineG1.data()),
342
343
      thrust::raw_pointer_cast(data._baseLineG0_update.data()),
      thrust::raw_pointer_cast(data._baseLineG1_update.data()),
344
345
      thrust::raw_pointer_cast(_noOfBitSetsIn_G0.data() + i),
      thrust::raw_pointer_cast(_noOfBitSetsIn_G1.data() + i)
346
      );
347
  }
348

349
350
351
352
353
354
355
356
357
358
359
    // only few output blocks per input block thus execution on only one thread.
    // Important is that the execution is async on the GPU.
    update_baselines<<<1,1,0, _proc_stream>>>(
        thrust::raw_pointer_cast(data._baseLineG0.data()),
        thrust::raw_pointer_cast(data._baseLineG1.data()),
        thrust::raw_pointer_cast(data._baseLineG0_update.data()),
        thrust::raw_pointer_cast(data._baseLineG1_update.data()),
        thrust::raw_pointer_cast(_noOfBitSetsIn_G0.data()),
        thrust::raw_pointer_cast(_noOfBitSetsIn_G1.data()),
        _noOfBitSetsIn_G0.size()
            );
360
361
362
363
364

  BOOST_LOG_TRIVIAL(debug) << "Performing FFT 1";
  UnpackedVoltageType *_unpacked_voltage_ptr =
      thrust::raw_pointer_cast(_unpacked_voltage_G0.data());
  ChannelisedVoltageType *_channelised_voltage_ptr =
365
      thrust::raw_pointer_cast(data._channelised_voltage_G0.data());
366
367
368
369
370
  CUFFT_ERROR_CHECK(cufftExecR2C(_fft_plan, (cufftReal *)_unpacked_voltage_ptr,
                                 (cufftComplex *)_channelised_voltage_ptr));

  BOOST_LOG_TRIVIAL(debug) << "Performing FFT 2";
  _unpacked_voltage_ptr = thrust::raw_pointer_cast(_unpacked_voltage_G1.data());
371
  _channelised_voltage_ptr = thrust::raw_pointer_cast(data._channelised_voltage_G1.data());
372
373
374
  CUFFT_ERROR_CHECK(cufftExecR2C(_fft_plan, (cufftReal *)_unpacked_voltage_ptr,
                                 (cufftComplex *)_channelised_voltage_ptr));

Tobias Winchen's avatar
Tobias Winchen committed
375
  CUDA_ERROR_CHECK(cudaStreamSynchronize(_proc_stream));
376
  BOOST_LOG_TRIVIAL(debug) << "Exit processing";
377
378
379
} // process


380
381
template <class HandlerType>
bool GatedSpectrometer<HandlerType>::operator()(RawBytes &block) {
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
    ++_call_count;
    BOOST_LOG_TRIVIAL(debug) << "GatedSpectrometer operator() called (count = "
                             << _call_count << ")";
    if (block.used_bytes() != _dadaBufferLayout.getBufferSize()) {
      // Stop on unexpected buffer size
      BOOST_LOG_TRIVIAL(error) << "Unexpected Buffer Size - Got "
                               << block.used_bytes() << " byte, expected "
                               << _dadaBufferLayout.getBufferSize() << " byte)";
      CUDA_ERROR_CHECK(cudaDeviceSynchronize());
      cudaProfilerStop();
      return true;
    }

    // Copy data to device
    CUDA_ERROR_CHECK(cudaStreamSynchronize(_h2d_stream));
    polarization0.swap();
    polarization1.swap();

    BOOST_LOG_TRIVIAL(debug) << "   block.used_bytes() = " <<
        block.used_bytes() << ", dataBlockBytes = " <<
        _dadaBufferLayout.sizeOfData() << "\n";

    // Copy the data with stride to the GPU:
    // CPU: P1P2P1P2P1P2 ...
    // GPU: P1P1P1 ... P2P2P2 ...
    // If this is a bottleneck the gating kernel could sort the layout out
    // during copy
    int heapsize_bytes = _nsamps_per_heap * _nbits / 8;
    CUDA_ERROR_CHECK(cudaMemcpy2DAsync(
      static_cast<void *>(polarization0._raw_voltage.a_ptr()),
        heapsize_bytes,
        static_cast<void *>(block.ptr()),
        2 * heapsize_bytes,
        heapsize_bytes, _dadaBufferLayout.sizeOfData() / heapsize_bytes/ 2,
        cudaMemcpyHostToDevice, _h2d_stream));

    CUDA_ERROR_CHECK(cudaMemcpy2DAsync(
      static_cast<void *>(polarization1._raw_voltage.a_ptr()),
        heapsize_bytes,
        static_cast<void *>(block.ptr()) + heapsize_bytes,
        2 * heapsize_bytes,
        heapsize_bytes, _dadaBufferLayout.sizeOfData() / heapsize_bytes/ 2,
        cudaMemcpyHostToDevice, _h2d_stream));

    CUDA_ERROR_CHECK(cudaMemcpy2DAsync(
        static_cast<void *>(polarization0._sideChannelData.a_ptr()),
        sizeof(uint64_t),
        static_cast<void *>(block.ptr() + _dadaBufferLayout.sizeOfData() + _dadaBufferLayout.sizeOfGap()),
        2 * sizeof(uint64_t),
        sizeof(uint64_t),
        _dadaBufferLayout.sizeOfSideChannelData() / 2 / sizeof(uint64_t),
        cudaMemcpyHostToDevice, _h2d_stream));

    CUDA_ERROR_CHECK(cudaMemcpy2DAsync(
        static_cast<void *>(polarization1._sideChannelData.a_ptr()),
        sizeof(uint64_t),
        static_cast<void *>(block.ptr() + _dadaBufferLayout.sizeOfData() + _dadaBufferLayout.sizeOfGap() + sizeof(uint64_t)),
        2 * sizeof(uint64_t),
        sizeof(uint64_t),
        _dadaBufferLayout.sizeOfSideChannelData() / 2 / sizeof(uint64_t), cudaMemcpyHostToDevice, _h2d_stream));

    BOOST_LOG_TRIVIAL(debug) << "First side channel item: 0x" <<   std::setw(16)
        << std::setfill('0') << std::hex <<
        (reinterpret_cast<uint64_t*>(block.ptr() + _dadaBufferLayout.sizeOfData()
                                     + _dadaBufferLayout.sizeOfGap()))[0] << std::dec;
447

448

449
450
451
  if (_call_count == 1) {
    return false;
  }
452

453
  // process data
454
455
  // check if new outblock is started:  _call_count -1 because this is the block number on the device
  bool newBlock = (((_call_count-1) * _nsamps_per_buffer) % _nsamps_per_output_spectra == 0);
456

457
458
  // only if  a newblock is started the output buffer is swapped. Otherwise the
  // new data is added to it
459
  if (newBlock)
460
  {
461
462
463
464
465
      BOOST_LOG_TRIVIAL(debug) << "Starting new output block.";
      stokes_G0.swap();
      stokes_G1.swap();
      stokes_G0.reset(_proc_stream);
      stokes_G1.reset(_proc_stream);
466
  }
467

468
469
  mergeSideChannels<<<1024, 1024, 0, _proc_stream>>>(thrust::raw_pointer_cast(polarization0._sideChannelData.a().data()),
          thrust::raw_pointer_cast(polarization1._sideChannelData.a().data()), polarization1._sideChannelData.a().size());
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494

  gated_fft(polarization0, stokes_G0._noOfBitSets.a(), stokes_G1._noOfBitSets.a());
  gated_fft(polarization1, stokes_G0._noOfBitSets.a(), stokes_G1._noOfBitSets.a());

  stokes_accumulate<<<1024, 1024, 0, _proc_stream>>>(
          thrust::raw_pointer_cast(polarization0._channelised_voltage_G0.data()),
          thrust::raw_pointer_cast(polarization1._channelised_voltage_G0.data()),
          thrust::raw_pointer_cast(stokes_G0.I.a().data()),
          thrust::raw_pointer_cast(stokes_G0.Q.a().data()),
          thrust::raw_pointer_cast(stokes_G0.U.a().data()),
          thrust::raw_pointer_cast(stokes_G0.V.a().data()),
          _nchans, _naccumulate
          );

  stokes_accumulate<<<1024, 1024, 0, _proc_stream>>>(
          thrust::raw_pointer_cast(polarization0._channelised_voltage_G1.data()),
          thrust::raw_pointer_cast(polarization1._channelised_voltage_G1.data()),
          thrust::raw_pointer_cast(stokes_G1.I.a().data()),
          thrust::raw_pointer_cast(stokes_G1.Q.a().data()),
          thrust::raw_pointer_cast(stokes_G1.U.a().data()),
          thrust::raw_pointer_cast(stokes_G1.V.a().data()),
          _nchans, _naccumulate
          );


495
  CUDA_ERROR_CHECK(cudaStreamSynchronize(_proc_stream));
Tobias Winchen's avatar
Tobias Winchen committed
496

497
  if ((_call_count == 2) || (!newBlock)) {
498
499
500
    return false;
  }

501
  // copy data to host if block is finished
502
  CUDA_ERROR_CHECK(cudaStreamSynchronize(_d2h_stream));
503
  _host_power_db.swap();
504
505
  // OUTPUT MEMORY LAYOUT:
  // I G0, IG1,Q G0, QG1, U G0,UG1,V G0,VG1, 8xSCI, ...
506

507
  for (size_t i = 0; i < stokes_G0._noOfBitSets.size(); i++)
508
  {
509
510
511
    size_t memslicesize = (_nchans * sizeof(IntegratedPowerType));
    size_t memOffset = 8 * i * (memslicesize +  + sizeof(size_t));
    // Copy  II QQ UU VV
512
513
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr() + memOffset) ,
514
515
516
517
518
519
520
521
                        static_cast<void *>(stokes_G0.I.b_ptr() + i * memslicesize),
                        _nchans * sizeof(IntegratedPowerType),
                        cudaMemcpyDeviceToHost, _d2h_stream));

    CUDA_ERROR_CHECK(
        cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr() + memOffset + 1 * memslicesize) ,
                        static_cast<void *>(stokes_G1.I.b_ptr() + i * memslicesize),
                        _nchans * sizeof(IntegratedPowerType),
522
                        cudaMemcpyDeviceToHost, _d2h_stream));
523

524
    CUDA_ERROR_CHECK(
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
        cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr() + memOffset + 2 * memslicesize) ,
                        static_cast<void *>(stokes_G0.Q.b_ptr() + i * memslicesize),
                        _nchans * sizeof(IntegratedPowerType),
                        cudaMemcpyDeviceToHost, _d2h_stream));

    CUDA_ERROR_CHECK(
        cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr() + memOffset + 3 * memslicesize) ,
                        static_cast<void *>(stokes_G1.Q.b_ptr() + i * memslicesize),
                        _nchans * sizeof(IntegratedPowerType),
                        cudaMemcpyDeviceToHost, _d2h_stream));

    CUDA_ERROR_CHECK(
        cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr() + memOffset + 4 * memslicesize) ,
                        static_cast<void *>(stokes_G0.U.b_ptr() + i * memslicesize),
                        _nchans * sizeof(IntegratedPowerType),
                        cudaMemcpyDeviceToHost, _d2h_stream));

    CUDA_ERROR_CHECK(
        cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr() + memOffset + 5 * memslicesize) ,
                        static_cast<void *>(stokes_G1.U.b_ptr() + i * memslicesize),
                        _nchans * sizeof(IntegratedPowerType),
                        cudaMemcpyDeviceToHost, _d2h_stream));

    CUDA_ERROR_CHECK(
        cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr() + memOffset + 6 * memslicesize) ,
                        static_cast<void *>(stokes_G0.V.b_ptr() + i * memslicesize),
                        _nchans * sizeof(IntegratedPowerType),
                        cudaMemcpyDeviceToHost, _d2h_stream));

    CUDA_ERROR_CHECK(
        cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr() + memOffset + 7 * memslicesize) ,
                        static_cast<void *>(stokes_G1.V.b_ptr() + i * memslicesize),
                        _nchans * sizeof(IntegratedPowerType),
                        cudaMemcpyDeviceToHost, _d2h_stream));

    // Copy SCI
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync( static_cast<void *>(_host_power_db.a_ptr() + memOffset + 8 * memslicesize),
          static_cast<void *>(stokes_G0._noOfBitSets.b_ptr() + i ),
            1 * sizeof(size_t),
            cudaMemcpyDeviceToHost, _d2h_stream));
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync( static_cast<void *>(_host_power_db.a_ptr() + memOffset + 8 * memslicesize + 1 * sizeof(size_t)),
          static_cast<void *>(stokes_G1._noOfBitSets.b_ptr() + i ),
            1 * sizeof(size_t),
            cudaMemcpyDeviceToHost, _d2h_stream));
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync( static_cast<void *>(_host_power_db.a_ptr() + memOffset + 8 * memslicesize + 2 * sizeof(size_t)),
          static_cast<void *>(stokes_G0._noOfBitSets.b_ptr() + i ),
574
575
576
            1 * sizeof(size_t),
            cudaMemcpyDeviceToHost, _d2h_stream));
    CUDA_ERROR_CHECK(
577
578
        cudaMemcpyAsync( static_cast<void *>(_host_power_db.a_ptr() + memOffset + 8 * memslicesize + 3 * sizeof(size_t)),
          static_cast<void *>(stokes_G1._noOfBitSets.b_ptr() + i ),
579
580
            1 * sizeof(size_t),
            cudaMemcpyDeviceToHost, _d2h_stream));
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync( static_cast<void *>(_host_power_db.a_ptr() + memOffset + 8 * memslicesize + 4 * sizeof(size_t)),
          static_cast<void *>(stokes_G0._noOfBitSets.b_ptr() + i ),
            1 * sizeof(size_t),
            cudaMemcpyDeviceToHost, _d2h_stream));
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync( static_cast<void *>(_host_power_db.a_ptr() + memOffset + 8 * memslicesize + 5 * sizeof(size_t)),
          static_cast<void *>(stokes_G1._noOfBitSets.b_ptr() + i ),
            1 * sizeof(size_t),
            cudaMemcpyDeviceToHost, _d2h_stream));
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync( static_cast<void *>(_host_power_db.a_ptr() + memOffset + 8 * memslicesize + 6 * sizeof(size_t)),
          static_cast<void *>(stokes_G0._noOfBitSets.b_ptr() + i ),
            1 * sizeof(size_t),
            cudaMemcpyDeviceToHost, _d2h_stream));
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync( static_cast<void *>(_host_power_db.a_ptr() + memOffset + 8 * memslicesize + 7 * sizeof(size_t)),
          static_cast<void *>(stokes_G1._noOfBitSets.b_ptr() + i ),
            1 * sizeof(size_t),
            cudaMemcpyDeviceToHost, _d2h_stream));

602
  }
603
604

  BOOST_LOG_TRIVIAL(debug) << "Copy Data back to host";
605

606
607
608
609
  if (_call_count == 3) {
    return false;
  }

610
  // calculate off value
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
  //BOOST_LOG_TRIVIAL(info) << "Buffer block: " << _call_count-3 << " with " << _noOfBitSetsIn_G0.size() << "x2 output heaps:";
  //size_t total_samples_lost = 0;
  //for (size_t i = 0; i < _noOfBitSetsIn_G0.size(); i++)
  //{
  //  size_t memOffset = 2 * i * (_nchans * sizeof(IntegratedPowerType) + sizeof(size_t));

  //  size_t* on_values = reinterpret_cast<size_t*> (_host_power_db.b_ptr() + memOffset + 2 * _nchans * sizeof(IntegratedPowerType));
  //  size_t* off_values = reinterpret_cast<size_t*> (_host_power_db.b_ptr() + memOffset + 2 * _nchans * sizeof(IntegratedPowerType) + sizeof(size_t));

  //  size_t samples_lost = _nsamps_per_output_spectra - (*on_values) - (*off_values);
  //  total_samples_lost += samples_lost;

  //  BOOST_LOG_TRIVIAL(info) << "    Heap " << i << ":\n"
  //    <<"                            Samples with  bit set  : " << *on_values << std::endl
  //    <<"                            Samples without bit set: " << *off_values << std::endl
  //    <<"                            Samples lost           : " << samples_lost << " out of " << _nsamps_per_output_spectra << std::endl;
  //}
  //double efficiency = 1. - double(total_samples_lost) / (_nsamps_per_output_spectra * _noOfBitSetsIn_G0.size());
  //double prev_average = _processing_efficiency / (_call_count- 3 - 1);
  //_processing_efficiency += efficiency;
  //double average = _processing_efficiency / (_call_count-3);
  //BOOST_LOG_TRIVIAL(info) << "Total processing efficiency of this buffer block:" << std::setprecision(6) << efficiency << ". Run average: " << average << " (Trend: " << std::showpos << (average - prev_average) << ")";
633
634

  // Wrap in a RawBytes object here;
635
  RawBytes bytes(reinterpret_cast<char *>(_host_power_db.b_ptr()),
636
637
                 _host_power_db.size(),
                 _host_power_db.size());
638
639
640
641
  BOOST_LOG_TRIVIAL(debug) << "Calling handler";
  // The handler can't do anything asynchronously without a copy here
  // as it would be unsafe (given that it does not own the memory it
  // is being passed).
Tobias Winchen's avatar
Tobias Winchen committed
642
643
644

  _handler(bytes);
  return false; //
645
646
647
648
649
650
} // operator ()

} // edd
} // effelsberg
} // psrdada_cpp