GatedSpectrometer.cu 13 KB
Newer Older
1
2
3
4
5
#include "psrdada_cpp/effelsberg/edd/GatedSpectrometer.cuh"
#include "psrdada_cpp/common.hpp"
#include "psrdada_cpp/cuda_utils.hpp"
#include "psrdada_cpp/raw_bytes.hpp"
#include <cuda.h>
6
#include <cuda_profiler_api.h>
7
8
9
10
11
12
13
14
15

#include <iostream>

namespace psrdada_cpp {
namespace effelsberg {
namespace edd {


__global__ void gating(float *G0, float *G1, const int64_t *sideChannelData,
16
17
                       size_t N, size_t heapSize, size_t bitpos,
                       size_t noOfSideChannels, size_t selectedSideChannel) {
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
  for (int i = blockIdx.x * blockDim.x + threadIdx.x; (i < N);
       i += blockDim.x * gridDim.x) {
    const float w = G0[i];
    const int64_t sideChannelItem =
        sideChannelData[((i / heapSize) * (noOfSideChannels)) +
                        selectedSideChannel]; // Probably not optimal access as
                                              // same data is copied for several
                                              // threads, but maybe efficiently
                                              // handled by cache?

    const int bit_set = TEST_BIT(sideChannelItem, bitpos);
    G1[i] = w * bit_set;
    G0[i] = w * (!bit_set);
  }
}

Tobias Winchen's avatar
Tobias Winchen committed
34

35
36
37
__global__ void countBitSet(const int64_t *sideChannelData, size_t N, size_t
    bitpos, size_t noOfSideChannels, size_t selectedSideChannel, unsigned int
    *nBitsSet)
38
{
Tobias Winchen's avatar
Tobias Winchen committed
39
40
41
42
43
  // really not optimized reduction, but here only trivial array sizes.
  int i = blockIdx.x * blockDim.x + threadIdx.x;
  __shared__ int x[256];

  if (i == 0)
44
    nBitsSet[0] = 0;
Tobias Winchen's avatar
Tobias Winchen committed
45

46
  if (i * noOfSideChannels + selectedSideChannel < N)
Tobias Winchen's avatar
Tobias Winchen committed
47
48
49
50
51
52
53
54
55
56
57
    x[threadIdx.x] = TEST_BIT(sideChannelData[i * noOfSideChannels + selectedSideChannel], bitpos);
  else
    x[threadIdx.x] = 0;
  __syncthreads();

  for(int s = blockDim.x / 2; s > 0; s = s / 2)
  {
    if (threadIdx.x < s)
      x[threadIdx.x] += x[threadIdx.x + s];
    __syncthreads();
  }
58

Tobias Winchen's avatar
Tobias Winchen committed
59
  if(threadIdx.x == 0)
60
   atomicAdd(nBitsSet, x[threadIdx.x]);
Tobias Winchen's avatar
Tobias Winchen committed
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

template <class HandlerType>
GatedSpectrometer<HandlerType>::GatedSpectrometer(
    std::size_t buffer_bytes, std::size_t nSideChannels,
    std::size_t selectedSideChannel, std::size_t selectedBit,
    std::size_t speadHeapSize, std::size_t fft_length, std::size_t naccumulate,
    std::size_t nbits, float input_level, float output_level,
    HandlerType &handler)
    : _buffer_bytes(buffer_bytes), _nSideChannels(nSideChannels),
      _selectedSideChannel(selectedSideChannel), _selectedBit(selectedBit),
      _speadHeapSize(speadHeapSize), _fft_length(fft_length),
      _naccumulate(naccumulate), _nbits(nbits), _handler(handler), _fft_plan(0),
      _call_count(0) {
  assert(((_nbits == 12) || (_nbits == 8)));
  assert(_naccumulate > 0); // Sanity check
  BOOST_LOG_TRIVIAL(debug)
      << "Creating new GatedSpectrometer instance with parameters: \n"
      << "  fft_length = " << _fft_length << "\n"
      << "  naccumulate = " << _naccumulate << "\n"
      << "  nSideChannels = " << _nSideChannels << "\n"
      << "  speadHeapSize = " << _speadHeapSize << " byte\n"
      << "  selectedSideChannel = " << _selectedSideChannel
      << "  selectedBit = " << _selectedBit;

  _sideChannelSize = nSideChannels * sizeof(int64_t);
  _totalHeapSize = _speadHeapSize + _sideChannelSize;
  _nHeaps = buffer_bytes / _totalHeapSize;
  _gapSize = (buffer_bytes - _nHeaps * _totalHeapSize);
  _dataBlockBytes = _nHeaps * _speadHeapSize;
  assert((nSideChannels == 0) ||
         (selectedSideChannel <
          nSideChannels));  // Sanity check of side channel value
  assert(selectedBit < 64); // Sanity check of selected bit
  BOOST_LOG_TRIVIAL(debug) << "Resulting memory configuration: \n"
                           << "  totalSizeOfHeap: " << _totalHeapSize
                           << " byte\n"
                           << "  number of heaps per buffer: " << _nHeaps
                           << "\n"
                           << "  resulting gap: " << _gapSize << " byte\n"
                           << "  datablock size in buffer: " << _dataBlockBytes
                           << " byte\n";

  std::size_t nsamps_per_buffer = _dataBlockBytes * 8 / nbits;
  std::size_t n64bit_words = _dataBlockBytes / sizeof(uint64_t);
  _nchans = _fft_length / 2 + 1;
  int batch = nsamps_per_buffer / _fft_length;
  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";
  _raw_voltage_db.resize(n64bit_words);
  _sideChannelData_db.resize(_sideChannelSize * _nHeaps);
  BOOST_LOG_TRIVIAL(debug) << "  Input voltages size (in 64-bit words): "
                           << _raw_voltage_db.size();
  _unpacked_voltage_G0.resize(nsamps_per_buffer);
  _unpacked_voltage_G1.resize(nsamps_per_buffer);
  BOOST_LOG_TRIVIAL(debug) << "  Unpacked voltages size (in samples): "
                           << _unpacked_voltage_G0.size();
133
  _channelised_voltage.resize(_nchans * batch);
134
  BOOST_LOG_TRIVIAL(debug) << "  Channelised voltages size: "
135
                           << _channelised_voltage.size();
136
137
138
139
140
141
  _power_db_G0.resize(_nchans * batch / _naccumulate);
  _power_db_G1.resize(_nchans * batch / _naccumulate);
  BOOST_LOG_TRIVIAL(debug) << "  Powers size: " << _power_db_G0.size() << ", "
                           << _power_db_G1.size();
  // on the host both power are stored in the same data buffer
  _host_power_db.resize( _power_db_G0.size() + _power_db_G1 .size());
Tobias Winchen's avatar
Tobias Winchen committed
142
  _noOfBitSetsInSideChannel.resize(1);
143
144
145
146
147
148

  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));

Tobias Winchen's avatar
Tobias Winchen committed
149
  // Create and record process status events to signal that processing chain is clear
Tobias Winchen's avatar
Tobias Winchen committed
150
151
152
153
154
  CUDA_ERROR_CHECK(cudaEventCreateWithFlags(&_procA, cudaEventDisableTiming));
  CUDA_ERROR_CHECK(cudaEventRecord(_procA, _proc_stream));
  CUDA_ERROR_CHECK(cudaEventCreateWithFlags(&_procB, cudaEventDisableTiming));
  CUDA_ERROR_CHECK(cudaEventRecord(_procB, _proc_stream));

155
156
157
158
159
160
161
162
163
164
165
166
167
168
  _unpacker.reset(new Unpacker(_proc_stream));
  _detector.reset(new DetectorAccumulator(_nchans, _naccumulate, scaling,
                                          offset, _proc_stream));
} // constructor


template <class HandlerType>
GatedSpectrometer<HandlerType>::~GatedSpectrometer() {
  BOOST_LOG_TRIVIAL(debug) << "Destroying GatedSpectrometer";
  if (!_fft_plan)
    cufftDestroy(_fft_plan);
  cudaStreamDestroy(_h2d_stream);
  cudaStreamDestroy(_proc_stream);
  cudaStreamDestroy(_d2h_stream);
Tobias Winchen's avatar
Tobias Winchen committed
169
170
  cudaEventDestroy(_procA);
  cudaEventDestroy(_procB);
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
}


template <class HandlerType>
void GatedSpectrometer<HandlerType>::init(RawBytes &block) {
  BOOST_LOG_TRIVIAL(debug) << "GatedSpectrometer init called";
  _handler.init(block);
}


template <class HandlerType>
void GatedSpectrometer<HandlerType>::process(
    thrust::device_vector<RawVoltageType> const &digitiser_raw,
    thrust::device_vector<RawVoltageType> const &sideChannelData,
    thrust::device_vector<IntegratedPowerType> &detected_G0,
186
    thrust::device_vector<IntegratedPowerType> &detected_G1, thrust::device_vector<unsigned int> &noOfBitSet) {
187
188
189
190
191
192
193
194
195
196
197
  BOOST_LOG_TRIVIAL(debug) << "Unpacking raw voltages";
  switch (_nbits) {
  case 8:
    _unpacker->unpack<8>(digitiser_raw, _unpacked_voltage_G0);
    break;
  case 12:
    _unpacker->unpack<12>(digitiser_raw, _unpacked_voltage_G0);
    break;
  default:
    throw std::runtime_error("Unsupported number of bits");
  }
Tobias Winchen's avatar
Tobias Winchen committed
198
199
  // raw voltage buffer is free again
  CUDA_ERROR_CHECK(cudaEventRecord(_procB, _proc_stream));
200

201
  BOOST_LOG_TRIVIAL(debug) << "Perform gating";
202
203
204
205
206
207
208
  const int64_t *sideCD =
      (int64_t *)(thrust::raw_pointer_cast(sideChannelData.data()));
  gating<<<1024, 1024, 0, _proc_stream>>>(
      thrust::raw_pointer_cast(_unpacked_voltage_G0.data()),
      thrust::raw_pointer_cast(_unpacked_voltage_G1.data()), sideCD,
      _unpacked_voltage_G0.size(), _speadHeapSize, _selectedBit, _nSideChannels,
      _selectedSideChannel);
209

Tobias Winchen's avatar
Tobias Winchen committed
210
211
212
213
  countBitSet<<<(sideChannelData.size()+255)/256, 256, 0,
    _proc_stream>>>(sideCD, sideChannelData.size(), _selectedBit,
        _nSideChannels, _selectedBit,
        thrust::raw_pointer_cast(noOfBitSet.data()));
214
215
216
217
218

  BOOST_LOG_TRIVIAL(debug) << "Performing FFT 1";
  UnpackedVoltageType *_unpacked_voltage_ptr =
      thrust::raw_pointer_cast(_unpacked_voltage_G0.data());
  ChannelisedVoltageType *_channelised_voltage_ptr =
219
      thrust::raw_pointer_cast(_channelised_voltage.data());
220
221
  CUFFT_ERROR_CHECK(cufftExecR2C(_fft_plan, (cufftReal *)_unpacked_voltage_ptr,
                                 (cufftComplex *)_channelised_voltage_ptr));
222
  _detector->detect(_channelised_voltage, detected_G0);
223
224
225
226
227
228
229

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

//  CUDA_ERROR_CHECK(cudaStreamSynchronize(_proc_stream));
230
  _detector->detect(_channelised_voltage, detected_G1);
231
  BOOST_LOG_TRIVIAL(debug) << "Exit processing";
232
233
234
235
236
237
238
239
240
241
242
243
} // process


template <class HandlerType>
bool GatedSpectrometer<HandlerType>::operator()(RawBytes &block) {
  ++_call_count;
  BOOST_LOG_TRIVIAL(debug) << "GatedSpectrometer operator() called (count = "
                           << _call_count << ")";
  if (block.used_bytes() != _buffer_bytes) { /* Unexpected buffer size */
    BOOST_LOG_TRIVIAL(error) << "Unexpected Buffer Size - Got "
                             << block.used_bytes() << " byte, expected "
                             << _buffer_bytes << " byte)";
Tobias Winchen's avatar
Tobias Winchen committed
244
245
246
    cudaDeviceSynchronize();
    cudaProfilerStop();
    return true;
247
248
  }

249
  //CUDA_ERROR_CHECK(cudaStreamSynchronize(_h2d_stream));
250
251
  _raw_voltage_db.swap();
  _sideChannelData_db.swap();
Tobias Winchen's avatar
Tobias Winchen committed
252
  std::swap(_procA, _procB);
253
254
255

  BOOST_LOG_TRIVIAL(debug) << "   block.used_bytes() = " << block.used_bytes()
                           << ", dataBlockBytes = " << _dataBlockBytes << "\n";
Tobias Winchen's avatar
Tobias Winchen committed
256

Tobias Winchen's avatar
Tobias Winchen committed
257
258
  // If necessary wait until the raw data has been processed
  CUDA_ERROR_CHECK(cudaEventSynchronize(_procA));
Tobias Winchen's avatar
Tobias Winchen committed
259

260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
  CUDA_ERROR_CHECK(cudaMemcpyAsync(static_cast<void *>(_raw_voltage_db.a_ptr()),
                                   static_cast<void *>(block.ptr()),
                                   _dataBlockBytes, cudaMemcpyHostToDevice,
                                   _h2d_stream));
  CUDA_ERROR_CHECK(cudaMemcpyAsync(
      static_cast<void *>(_sideChannelData_db.a_ptr()),
      static_cast<void *>(block.ptr() + _dataBlockBytes + _gapSize),
      _sideChannelSize * _nHeaps, cudaMemcpyHostToDevice, _h2d_stream));

  if (_call_count == 1) {
    return false;
  }

  // Synchronize all streams
  _power_db_G0.swap();
  _power_db_G1.swap();
Tobias Winchen's avatar
Tobias Winchen committed
276
  _noOfBitSetsInSideChannel.swap();
277
278

  process(_raw_voltage_db.b(), _sideChannelData_db.b(), _power_db_G0.a(),
279
          _power_db_G1.a(), _noOfBitSetsInSideChannel.a());
280

Tobias Winchen's avatar
Tobias Winchen committed
281
  // signal that data block has been processed
Tobias Winchen's avatar
Tobias Winchen committed
282
283
  //CUDA_ERROR_CHECK(cudaStreamSynchronize(_proc_stream));

284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
  if (_call_count == 2) {
    return false;
  }

  //CUDA_ERROR_CHECK(cudaStreamSynchronize(_d2h_stream));
  _host_power_db.swap();
  CUDA_ERROR_CHECK(
      cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr()),
                      static_cast<void *>(_power_db_G0.b_ptr()),
                      _power_db_G0.size() * sizeof(IntegratedPowerType),
                      cudaMemcpyDeviceToHost, _d2h_stream));
  CUDA_ERROR_CHECK(cudaMemcpyAsync(
      static_cast<void *>(_host_power_db.a_ptr() +
                          (_power_db_G0.size() * sizeof(IntegratedPowerType))),
      static_cast<void *>(_power_db_G1.b_ptr()),
      _power_db_G1.size() * sizeof(IntegratedPowerType), cudaMemcpyDeviceToHost,
      _d2h_stream));

Tobias Winchen's avatar
Tobias Winchen committed
302
303
304
  int R[1];
  CUDA_ERROR_CHECK(cudaMemcpyAsync(static_cast<void *>(R),
        static_cast<void *>(_noOfBitSetsInSideChannel.b_ptr()),
305
          1 * sizeof(unsigned int),cudaMemcpyDeviceToHost, _d2h_stream));
306

Tobias Winchen's avatar
Tobias Winchen committed
307
  BOOST_LOG_TRIVIAL(info) << "NOOF BIT SET IN SIDE CHANNEL: " << R[0] << std::endl;
308

309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
  if (_call_count == 3) {
    return false;
  }

  // Wrap _detected_host_previous in a RawBytes object here;
  RawBytes bytes(reinterpret_cast<char *>(_host_power_db.b_ptr()),
                 _host_power_db.size() * sizeof(IntegratedPowerType),
                 _host_power_db.size() * sizeof(IntegratedPowerType));
  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).
  return _handler(bytes);
} // operator ()

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