File size: 8,399 Bytes
61c2d32
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#include <ATen/ATen.h>

#include <cuda_fp16.h>

#include <vector>

#include "utils/checks.h"
#include "utils/cuda.cuh"
#include "inplace_abn.h"

#include <ATen/cuda/CUDAContext.h>

// Operations for reduce
struct SumOpH {
  __device__ SumOpH(const half *t, int c, int s)
      : tensor(t), chn(c), sp(s) {}
  __device__ __forceinline__ float operator()(int batch, int plane, int n) {
    return __half2float(tensor[(batch * chn + plane) * sp + n]);
  }
  const half *tensor;
  const int chn;
  const int sp;
};

struct VarOpH {
  __device__ VarOpH(float m, const half *t, int c, int s)
      : mean(m), tensor(t), chn(c), sp(s) {}
  __device__ __forceinline__ float operator()(int batch, int plane, int n) {
    const auto t = __half2float(tensor[(batch * chn + plane) * sp + n]);
    return (t - mean) * (t - mean);
  }
  const float mean;
  const half *tensor;
  const int chn;
  const int sp;
};

struct GradOpH {
  __device__ GradOpH(float _weight, float _bias, const half *_z, const half *_dz, int c, int s)
      : weight(_weight), bias(_bias), z(_z), dz(_dz), chn(c), sp(s) {}
  __device__ __forceinline__ Pair<float> operator()(int batch, int plane, int n) {
    float _y = (__half2float(z[(batch * chn + plane) * sp + n]) - bias) / weight;
    float _dz = __half2float(dz[(batch * chn + plane) * sp + n]);
    return Pair<float>(_dz, _y * _dz);
  }
  const float weight;
  const float bias;
  const half *z;
  const half *dz;
  const int chn;
  const int sp;
};

/***********
 * mean_var
 ***********/

__global__ void mean_var_kernel_h(const half *x, float *mean, float *var, int num, int chn, int sp) {
  int plane = blockIdx.x;
  float norm = 1.f / static_cast<float>(num * sp);

  float _mean = reduce<float, SumOpH>(SumOpH(x, chn, sp), plane, num, sp) * norm;
  __syncthreads();
  float _var = reduce<float, VarOpH>(VarOpH(_mean, x, chn, sp), plane, num, sp) * norm;

  if (threadIdx.x == 0) {
    mean[plane] = _mean;
    var[plane] = _var;
  }
}

std::vector<at::Tensor> mean_var_cuda_h(at::Tensor x) {
  CHECK_CUDA_INPUT(x);

  // Extract dimensions
  int64_t num, chn, sp;
  get_dims(x, num, chn, sp);

  // Prepare output tensors
  auto mean = at::empty({chn},x.options().dtype(at::kFloat));
  auto var = at::empty({chn},x.options().dtype(at::kFloat));

  // Run kernel
  dim3 blocks(chn);
  dim3 threads(getNumThreads(sp));
  auto stream = at::cuda::getCurrentCUDAStream();
  mean_var_kernel_h<<<blocks, threads, 0, stream>>>(
      reinterpret_cast<half*>(x.data<at::Half>()),
      mean.data<float>(),
      var.data<float>(),
      num, chn, sp);

  return {mean, var};
}

/**********
 * forward
 **********/

__global__ void forward_kernel_h(half *x, const float *mean, const float *var, const float *weight, const float *bias,
                                 bool affine, float eps, int num, int chn, int sp) {
  int plane = blockIdx.x;

  const float _mean = mean[plane];
  const float _var = var[plane];
  const float _weight = affine ? abs(weight[plane]) + eps : 1.f;
  const float _bias = affine ? bias[plane] : 0.f;

  const float mul = rsqrt(_var + eps) * _weight;

  for (int batch = 0; batch < num; ++batch) {
    for (int n = threadIdx.x; n < sp; n += blockDim.x) {
      half *x_ptr = x + (batch * chn + plane) * sp + n;
      float _x = __half2float(*x_ptr);
      float _y = (_x - _mean) * mul + _bias;

      *x_ptr = __float2half(_y);
    }
  }
}

at::Tensor forward_cuda_h(at::Tensor x, at::Tensor mean, at::Tensor var, at::Tensor weight, at::Tensor bias,
                        bool affine, float eps) {
  CHECK_CUDA_INPUT(x);
  CHECK_CUDA_INPUT(mean);
  CHECK_CUDA_INPUT(var);
  CHECK_CUDA_INPUT(weight);
  CHECK_CUDA_INPUT(bias);

  // Extract dimensions
  int64_t num, chn, sp;
  get_dims(x, num, chn, sp);

  // Run kernel
  dim3 blocks(chn);
  dim3 threads(getNumThreads(sp));
  auto stream = at::cuda::getCurrentCUDAStream();
  forward_kernel_h<<<blocks, threads, 0, stream>>>(
      reinterpret_cast<half*>(x.data<at::Half>()),
      mean.data<float>(),
      var.data<float>(),
      weight.data<float>(),
      bias.data<float>(),
      affine, eps, num, chn, sp);

  return x;
}

__global__ void edz_eydz_kernel_h(const half *z, const half *dz, const float *weight, const float *bias,
                                float *edz, float *eydz, bool affine, float eps, int num, int chn, int sp) {
  int plane = blockIdx.x;

  float _weight = affine ? abs(weight[plane]) + eps : 1.f;
  float _bias = affine ? bias[plane] : 0.f;

  Pair<float> res = reduce<Pair<float>, GradOpH>(GradOpH(_weight, _bias, z, dz, chn, sp), plane, num, sp);
  __syncthreads();

  if (threadIdx.x == 0) {
    edz[plane] = res.v1;
    eydz[plane] = res.v2;
  }
}

std::vector<at::Tensor> edz_eydz_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor weight, at::Tensor bias,
                                      bool affine, float eps) {
  CHECK_CUDA_INPUT(z);
  CHECK_CUDA_INPUT(dz);
  CHECK_CUDA_INPUT(weight);
  CHECK_CUDA_INPUT(bias);

  // Extract dimensions
  int64_t num, chn, sp;
  get_dims(z, num, chn, sp);

  auto edz = at::empty({chn},z.options().dtype(at::kFloat));
  auto eydz = at::empty({chn},z.options().dtype(at::kFloat));

  // Run kernel
  dim3 blocks(chn);
  dim3 threads(getNumThreads(sp));
  auto stream = at::cuda::getCurrentCUDAStream();
  edz_eydz_kernel_h<<<blocks, threads, 0, stream>>>(
        reinterpret_cast<half*>(z.data<at::Half>()),
        reinterpret_cast<half*>(dz.data<at::Half>()),
        weight.data<float>(),
        bias.data<float>(),
        edz.data<float>(),
        eydz.data<float>(),
        affine, eps, num, chn, sp);
 
  return {edz, eydz};
}

__global__ void backward_kernel_h(const half *z, const half *dz, const float *var, const float *weight, const float *bias, const float *edz,
                                  const float *eydz, half *dx, bool affine, float eps, int num, int chn, int sp) {
  int plane = blockIdx.x;

  float _weight = affine ? abs(weight[plane]) + eps : 1.f;
  float _bias = affine ? bias[plane] : 0.f;
  float _var = var[plane];
  float _edz = edz[plane];
  float _eydz = eydz[plane];

  float _mul = _weight * rsqrt(_var + eps);
  float count = float(num * sp);

  for (int batch = 0; batch < num; ++batch) {
    for (int n = threadIdx.x; n < sp; n += blockDim.x) {
      float _dz = __half2float(dz[(batch * chn + plane) * sp + n]);
      float _y = (__half2float(z[(batch * chn + plane) * sp + n]) - _bias) / _weight;

      dx[(batch * chn + plane) * sp + n] = __float2half((_dz - _edz / count - _y * _eydz / count) * _mul);
    }
  }
}

at::Tensor backward_cuda_h(at::Tensor z, at::Tensor dz, at::Tensor var, at::Tensor weight, at::Tensor bias,
                                      at::Tensor edz, at::Tensor eydz, bool affine, float eps) {
  CHECK_CUDA_INPUT(z);
  CHECK_CUDA_INPUT(dz);
  CHECK_CUDA_INPUT(var);
  CHECK_CUDA_INPUT(weight);
  CHECK_CUDA_INPUT(bias);
  CHECK_CUDA_INPUT(edz);
  CHECK_CUDA_INPUT(eydz);

  // Extract dimensions
  int64_t num, chn, sp;
  get_dims(z, num, chn, sp);

  auto dx = at::zeros_like(z);

  // Run kernel
  dim3 blocks(chn);
  dim3 threads(getNumThreads(sp));
  auto stream = at::cuda::getCurrentCUDAStream();
  backward_kernel_h<<<blocks, threads, 0, stream>>>(
        reinterpret_cast<half*>(z.data<at::Half>()),
        reinterpret_cast<half*>(dz.data<at::Half>()),
        var.data<float>(),
        weight.data<float>(),
        bias.data<float>(),
        edz.data<float>(),
        eydz.data<float>(),
        reinterpret_cast<half*>(dx.data<at::Half>()),
        affine, eps, num, chn, sp);

  return dx;
}

__global__ void leaky_relu_backward_impl_h(half *z, half *dz, float slope, int64_t count) {
  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < count;  i += blockDim.x * gridDim.x){
    float _z = __half2float(z[i]);
    if (_z < 0) {
      dz[i] = __float2half(__half2float(dz[i]) * slope);
      z[i] = __float2half(_z / slope);
    }
  }
}

void leaky_relu_backward_cuda_h(at::Tensor z, at::Tensor dz, float slope) {
  CHECK_CUDA_INPUT(z);
  CHECK_CUDA_INPUT(dz);

  int64_t count = z.numel();
  dim3 threads(getNumThreads(count));
  dim3 blocks = (count + threads.x - 1) / threads.x;
  auto stream = at::cuda::getCurrentCUDAStream();
  leaky_relu_backward_impl_h<<<blocks, threads, 0, stream>>>(
      reinterpret_cast<half*>(z.data<at::Half>()),
      reinterpret_cast<half*>(dz.data<at::Half>()),
      slope, count);
}