GatedSpectrometer.cu 16.1 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

#include <iostream>
9
10
#include <cstring>
#include <sstream>
11
12
13
14
15

namespace psrdada_cpp {
namespace effelsberg {
namespace edd {

Tobias Winchen's avatar
Tobias Winchen committed
16
__global__ void gating(float* __restrict__ G0, float* __restrict__ G1, const int64_t* __restrict__ sideChannelData,
17
                       size_t N, size_t heapSize, size_t bitpos,
Tobias Winchen's avatar
Tobias Winchen committed
18
                       size_t noOfSideChannels, size_t selectedSideChannel, const float* __restrict__ _baseLineN) {
19
  float baseLine = (*_baseLineN) / N;
Tobias Winchen's avatar
Tobias Winchen committed
20
  for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; (i < N);
21
       i += blockDim.x * gridDim.x) {
Tobias Winchen's avatar
Tobias Winchen committed
22
    const float w = G0[i] - baseLine;
23
24
25
26
27
28
29
30
    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);
Tobias Winchen's avatar
Tobias Winchen committed
31
32
    G1[i] = w * bit_set + baseLine;
    G0[i] = w * (!bit_set) + baseLine;
33
34
35
  }
}

Tobias Winchen's avatar
Tobias Winchen committed
36

37
__global__ void countBitSet(const int64_t *sideChannelData, size_t N, size_t
38
    bitpos, size_t noOfSideChannels, size_t selectedSideChannel, size_t
39
    *nBitsSet)
40
{
Tobias Winchen's avatar
Tobias Winchen committed
41
42
  // really not optimized reduction, but here only trivial array sizes.
  int i = blockIdx.x * blockDim.x + threadIdx.x;
43
  __shared__ unsigned int x[256];
Tobias Winchen's avatar
Tobias Winchen committed
44
45

  if (i == 0)
46
    nBitsSet[0] = 0;
Tobias Winchen's avatar
Tobias Winchen committed
47

48
  if (i * noOfSideChannels + selectedSideChannel < N)
Tobias Winchen's avatar
Tobias Winchen committed
49
50
51
52
53
54
55
56
57
58
59
    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();
  }
60

Tobias Winchen's avatar
Tobias Winchen committed
61
  if(threadIdx.x == 0)
62
   nBitsSet[0] += x[threadIdx.x];
Tobias Winchen's avatar
Tobias Winchen committed
63
}
64

65

Tobias Winchen's avatar
Tobias Winchen committed
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
// blocksize for the array sum kernel
#define array_sum_Nthreads 1024

__global__ void array_sum(float *in, size_t N, float *out) {
  extern __shared__ float data[];

  size_t tid = threadIdx.x;

  float ls = 0;

  for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; (i < N);
       i += blockDim.x * gridDim.x) {
    ls += in[i]; // + in[i + blockDim.x];   // loading two elements increase the used bandwidth by ~10% but requires matching blocksize and size of input array
  }

  data[tid] = ls;
  __syncthreads();

  for (size_t i = blockDim.x / 2; i > 0; i /= 2) {
    if (tid < i) {
      data[tid] += data[tid + i];
    }
    __syncthreads();
  }

  // unroll last warp
  // if (tid < 32)
  //{
  //  warpReduce(data, tid);
  //}

  if (tid == 0) {
    out[blockIdx.x] = data[0];
  }
}


103
104
template <class HandlerType, typename IntegratedPowerType>
GatedSpectrometer<HandlerType, IntegratedPowerType>::GatedSpectrometer(
105
106
107
108
109
110
111
112
113
114
    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) {
115
116

  // Sanity checks
117
  assert(((_nbits == 12) || (_nbits == 8)));
118
119
120
121
122
  assert(_naccumulate > 0);

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

123
  BOOST_LOG_TRIVIAL(info)
124
      << "Creating new GatedSpectrometer instance with parameters: \n"
125
126
127
128
129
130
131
      << "  fft_length           " << _fft_length << "\n"
      << "  naccumulate          " << _naccumulate << "\n"
      << "  nSideChannels        " << _nSideChannels << "\n"
      << "  speadHeapSize        " << _speadHeapSize << " byte\n"
      << "  selectedSideChannel  " << _selectedSideChannel << "\n"
      << "  selectedBit          " << _selectedBit << "\n"
      << "  output bit depth     " << sizeof(IntegratedPowerType) * 8;
132
133
134
135
136
137
138
139
140
141

  _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
142
  BOOST_LOG_TRIVIAL(info) << "Resulting memory configuration: \n"
143
144
                           << "  totalSizeOfHeap: " << _totalHeapSize << " byte\n"
                           << "  number of heaps per buffer: " << _nHeaps << "\n"
145
                           << "  resulting gap: " << _gapSize << " byte\n"
146
                           << "  datablock size in buffer: " << _dataBlockBytes << " byte\n";
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

  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);
Tobias Winchen's avatar
Tobias Winchen committed
174
175

  _baseLineN.resize(array_sum_Nthreads);
176
177
  BOOST_LOG_TRIVIAL(debug) << "  Unpacked voltages size (in samples): "
                           << _unpacked_voltage_G0.size();
178
  _channelised_voltage.resize(_nchans * batch);
179
  BOOST_LOG_TRIVIAL(debug) << "  Channelised voltages size: "
180
                           << _channelised_voltage.size();
181
182
  _power_db.resize(_nchans * batch / _naccumulate * 2);  // hold on and off spectra to simplify output
  BOOST_LOG_TRIVIAL(debug) << "  Powers size: " << _power_db.size() / 2; 
183
184
185
186
  // on the host both power are stored in the same data buffer together with
  // the number of bit sets
  _noOfBitSetsInSideChannel.resize( batch / _naccumulate);
  _host_power_db.resize( _power_db.size() * sizeof(IntegratedPowerType) + 2 * sizeof(size_t) * _noOfBitSetsInSideChannel.size());
187
188
189
190
191
192
193

  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));
Tobias Winchen's avatar
Tobias Winchen committed
194
  _detector.reset(new DetectorAccumulator<IntegratedPowerType>(_nchans, _naccumulate, scaling,
195
196
197
198
                                          offset, _proc_stream));
} // constructor


199
200
template <class HandlerType, typename IntegratedPowerType>
GatedSpectrometer<HandlerType, IntegratedPowerType>::~GatedSpectrometer() {
201
202
203
204
205
206
207
208
209
  BOOST_LOG_TRIVIAL(debug) << "Destroying GatedSpectrometer";
  if (!_fft_plan)
    cufftDestroy(_fft_plan);
  cudaStreamDestroy(_h2d_stream);
  cudaStreamDestroy(_proc_stream);
  cudaStreamDestroy(_d2h_stream);
}


210
211
template <class HandlerType, typename IntegratedPowerType>
void GatedSpectrometer<HandlerType, IntegratedPowerType>::init(RawBytes &block) {
212
  BOOST_LOG_TRIVIAL(debug) << "GatedSpectrometer init called";
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
  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.";
  }

236
237
238
239
  _handler.init(block);
}


240
241
template <class HandlerType, typename IntegratedPowerType>
void GatedSpectrometer<HandlerType, IntegratedPowerType>::process(
242
    thrust::device_vector<RawVoltageType> const &digitiser_raw,
243
    thrust::device_vector<int64_t> const &sideChannelData,
244
    thrust::device_vector<IntegratedPowerType> &detected, thrust::device_vector<size_t> &noOfBitSet) {
245
246
247
248
249
250
251
252
253
254
255
  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
256
257
258
  BOOST_LOG_TRIVIAL(debug) << "Calculate baseline";
  psrdada_cpp::effelsberg::edd::array_sum<<<64, array_sum_Nthreads, array_sum_Nthreads * sizeof(float), _proc_stream>>>(thrust::raw_pointer_cast(_unpacked_voltage_G0.data()), _unpacked_voltage_G0.size(), thrust::raw_pointer_cast(_baseLineN.data()));
  psrdada_cpp::effelsberg::edd::array_sum<<<1, array_sum_Nthreads, array_sum_Nthreads * sizeof(float), _proc_stream>>>(thrust::raw_pointer_cast(_baseLineN.data()), _baseLineN.size(), thrust::raw_pointer_cast(_baseLineN.data()));
259

260
  BOOST_LOG_TRIVIAL(debug) << "Perform gating";
261
262
  gating<<<1024, 1024, 0, _proc_stream>>>(
      thrust::raw_pointer_cast(_unpacked_voltage_G0.data()),
263
264
      thrust::raw_pointer_cast(_unpacked_voltage_G1.data()),
      thrust::raw_pointer_cast(sideChannelData.data()),
265
      _unpacked_voltage_G0.size(), _speadHeapSize, _selectedBit, _nSideChannels,
Tobias Winchen's avatar
Tobias Winchen committed
266
      _selectedSideChannel, thrust::raw_pointer_cast(_baseLineN.data()));
267

268
269
270
271
272
273
274
275
  for (size_t i = 0; i < _noOfBitSetsInSideChannel.size(); i++)
  { // ToDo: Should be in one kernel call
    countBitSet<<<(sideChannelData.size()+255)/256, 256, 0,
      _proc_stream>>>(thrust::raw_pointer_cast(sideChannelData.data() + i * sideChannelData.size() / _noOfBitSetsInSideChannel.size() ),
          sideChannelData.size() / _noOfBitSetsInSideChannel.size(), _selectedBit,
          _nSideChannels, _selectedBit,
          thrust::raw_pointer_cast(noOfBitSet.data() + i));
  }
276
277
278
279
280

  BOOST_LOG_TRIVIAL(debug) << "Performing FFT 1";
  UnpackedVoltageType *_unpacked_voltage_ptr =
      thrust::raw_pointer_cast(_unpacked_voltage_G0.data());
  ChannelisedVoltageType *_channelised_voltage_ptr =
281
      thrust::raw_pointer_cast(_channelised_voltage.data());
282
283
  CUFFT_ERROR_CHECK(cufftExecR2C(_fft_plan, (cufftReal *)_unpacked_voltage_ptr,
                                 (cufftComplex *)_channelised_voltage_ptr));
284
  _detector->detect(_channelised_voltage, detected, 2, 0);
285
286
287
288
289
290

  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));

291
  CUDA_ERROR_CHECK(cudaStreamSynchronize(_proc_stream));
292
  _detector->detect(_channelised_voltage, detected, 2, 1);
Tobias Winchen's avatar
Tobias Winchen committed
293
  CUDA_ERROR_CHECK(cudaStreamSynchronize(_proc_stream));
294
  BOOST_LOG_TRIVIAL(debug) << "Exit processing";
295
296
297
} // process


298
299
template <class HandlerType, typename IntegratedPowerType>
bool GatedSpectrometer<HandlerType, IntegratedPowerType>::operator()(RawBytes &block) {
300
301
302
303
304
305
306
  ++_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
307
308
309
    cudaDeviceSynchronize();
    cudaProfilerStop();
    return true;
310
311
  }

312
  // Copy data to device
313
  CUDA_ERROR_CHECK(cudaStreamSynchronize(_h2d_stream));
314
315
316
317
318
  _raw_voltage_db.swap();
  _sideChannelData_db.swap();

  BOOST_LOG_TRIVIAL(debug) << "   block.used_bytes() = " << block.used_bytes()
                           << ", dataBlockBytes = " << _dataBlockBytes << "\n";
Tobias Winchen's avatar
Tobias Winchen committed
319

320
321
322
323
324
325
326
327
328
329
330
331
332
  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;
  }

333
  //process data
334
  _power_db.swap();
Tobias Winchen's avatar
Tobias Winchen committed
335
  _noOfBitSetsInSideChannel.swap();
336

337
  process(_raw_voltage_db.b(), _sideChannelData_db.b(), _power_db.a(), _noOfBitSetsInSideChannel.a());
338
  CUDA_ERROR_CHECK(cudaStreamSynchronize(_proc_stream));
Tobias Winchen's avatar
Tobias Winchen committed
339

340
341
342
343
  if (_call_count == 2) {
    return false;
  }

344
  //copy data to host
345
  CUDA_ERROR_CHECK(cudaStreamSynchronize(_d2h_stream));
346
  _host_power_db.swap();
347
  std::swap(_noOfBitSetsInSideChannel_host[0], _noOfBitSetsInSideChannel_host[1]);
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366

  for (size_t i = 0; i < _noOfBitSetsInSideChannel.size(); i++)
  {
    size_t memOffset = 2 * i * (_nchans * sizeof(IntegratedPowerType) + sizeof(size_t));
    // copy 2x channel data
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync(static_cast<void *>(_host_power_db.a_ptr() + memOffset) ,
                        static_cast<void *>(_power_db.b_ptr() + 2 * i * _nchans),
                        2 * _nchans * sizeof(IntegratedPowerType),
                        cudaMemcpyDeviceToHost, _d2h_stream));
    // copy noOf bit set data
    CUDA_ERROR_CHECK(
        cudaMemcpyAsync( static_cast<void *>(_host_power_db.a_ptr() + memOffset + 2 * _nchans ),
          static_cast<void *>(_noOfBitSetsInSideChannel.b_ptr() + i ),
            1 * sizeof(size_t),
            cudaMemcpyDeviceToHost, _d2h_stream));
}

  BOOST_LOG_TRIVIAL(debug) << "Copy Data back to host";
367

368
369
370
371
  if (_call_count == 3) {
    return false;
  }

372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
  // calculate off value
  BOOST_LOG_TRIVIAL(info) << "Buffer block: " << _call_count << " with " << _noOfBitSetsInSideChannel.size() << " output heaps:"; 
  for (size_t i = 0; i < _noOfBitSetsInSideChannel.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));
    *off_values =  _nHeaps - (*on_values);

    BOOST_LOG_TRIVIAL(info) << "    " << i << ": No of bit set in side channel: " << *on_values << " / " << *off_values << std::endl;
  }


  // call handler


  // Wrap in a RawBytes object here;
390
  RawBytes bytes(reinterpret_cast<char *>(_host_power_db.b_ptr()),
391
392
                 _host_power_db.size(),
                 _host_power_db.size());
393
394
395
396
  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
397
398
399

  _handler(bytes);
  return false; //
400
401
402
403
404
405
} // operator ()

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