Hugging Face's logo

bds2714
/
jukebox

Model card Files
xet
Community
jukebox
File size: 4,845 Bytes 
c508d7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
#include <ATen/ATen.h>
#include <ATen/AccumulateType.h>
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/Exceptions.h>
// Another possibility:
// #include <torch/all.h>

#include <assert.h>

#include "type_shim.h"
#include "multi_tensor_apply.cuh"

#define BLOCK_SIZE 512
#define ILP 4

template<typename x_t>
struct L2NormFunctor
{
  __device__ __forceinline__ void operator()(
    int chunk_size,
    volatile int* noop_gmem,
    TensorListMetadata<1>& tl,
    float* output,
    float* output_per_tensor,
    bool per_tensor,
    int max_chunks_per_tensor)
  {
    // I'd like this kernel to propagate infs/nans.
    // if(*noop_gmem == 1)
    //   return;

    int tensor_loc = tl.block_to_tensor[blockIdx.x];
    int chunk_idx = tl.block_to_chunk[blockIdx.x];
    int n = tl.sizes[tensor_loc];

    x_t* x = (x_t*)tl.addresses[0][tensor_loc];
    x += chunk_idx*chunk_size;

    n -= chunk_idx*chunk_size;

    __shared__ float s_vals[512];

    float vals[ILP]; // = {0}; // this probably works too but I want to be sure...
    for(int i = 0; i < ILP; i++)
      vals[i] = 0.f;

    for(int i_start = 0; i_start < n && i_start < chunk_size; i_start += blockDim.x*ILP)
    {
      #pragma unroll
      for(int ii = 0; ii < ILP; ii++)
      {
        int i = i_start + threadIdx.x + ii*blockDim.x;
        if(i < n && i < chunk_size)
        {
          float next = static_cast<float>(x[i]);
          vals[ii] += next*next;
        }
      }
    }

    float val = 0.f;
    for(int i = 0; i < ILP; i++)
        val += vals[i];

    float final = reduce_block_into_lanes(s_vals, val);

    if(threadIdx.x == 0)
    {
      if(!isfinite(final))
        *noop_gmem = 1; // Blindly fire off a write.  These will race but that's ok.
      output[blockIdx.x] += final;
      if(per_tensor)
        output_per_tensor[(tl.start_tensor_this_launch + tensor_loc)*max_chunks_per_tensor + chunk_idx] = final;
    }
  }
};


__global__ void cleanup(
  float* output,
  float* output_per_tensor,
  float* ret,
  float* ret_per_tensor,
  bool per_tensor,
  int max_chunks_per_tensor)
{
  __shared__ float vals[512];

  if(blockIdx.x == 0)
  {
    float val = 0;
    if(threadIdx.x < 320)
      val = output[threadIdx.x];

    float final = reduce_block_into_lanes(vals, val);

    if(threadIdx.x == 0)
      *ret = sqrt(final);
  }

  if(per_tensor)
  {
    float* output_this_tensor = output_per_tensor + blockIdx.x*max_chunks_per_tensor;

    float val = 0;
    for(int i = threadIdx.x; i < max_chunks_per_tensor; i += blockDim.x)
      val += output_this_tensor[i];

    float final = reduce_block_into_lanes(vals, val);

    if(threadIdx.x == 0)
      ret_per_tensor[blockIdx.x] = sqrt(final);
  }
}


std::tuple<at::Tensor, at::Tensor> multi_tensor_l2norm_cuda(
  int chunk_size,
  at::Tensor noop_flag,
  std::vector<std::vector<at::Tensor>> tensor_lists,
  at::optional<bool> per_tensor_python)
{
  bool per_tensor = per_tensor_python.has_value() ? per_tensor_python.value() : false;

  auto float_options = tensor_lists[0][0].options().dtype(at::kFloat);
  auto output = at::zeros({320}, float_options);

  at::Tensor output_per_tensor;
  at::Tensor ret_per_tensor;

  int ntensors = tensor_lists[0].size();
  int max_chunks_per_tensor = -1;

  if(per_tensor)
  {
    for(int t = 0; t < ntensors; t++)
    {
      int max_chunks_this_tensor = (tensor_lists[0][t].numel() + chunk_size - 1)/chunk_size;
      if(max_chunks_this_tensor > max_chunks_per_tensor)
        max_chunks_per_tensor = max_chunks_this_tensor;
    }
    output_per_tensor = at::zeros({ntensors*max_chunks_per_tensor}, float_options);
    ret_per_tensor = at::empty({ntensors}, float_options);
  }
  else
  {
    ret_per_tensor = at::empty({0}, float_options);
  }

  DISPATCH_FLOAT_AND_HALF(tensor_lists[0][0].scalar_type(), 0, "multi_tensor_l2norm_cuda",
    multi_tensor_apply<1>(
      BLOCK_SIZE,
      chunk_size,
      noop_flag,
      tensor_lists,
      L2NormFunctor<scalar_t_0>(),
      output.data<float>(),
      per_tensor ? output_per_tensor.data<float>() : nullptr,
      per_tensor,
      max_chunks_per_tensor);)

  AT_CUDA_CHECK(cudaGetLastError());

  // AT_CUDA_CHECK(cudaDeviceSynchronize());

  // This involves one more small kernel launches, but will be negligible end to end.
  // I could get rid of these by hacking the functor + multi tensor harness with persistence
  // logic, but keeping it simple for now
  auto ret = at::empty({1}, output.options());
  auto stream = at::cuda::getCurrentCUDAStream();
  cleanup<<<per_tensor ? ntensors : 1, 512, 0, stream>>>(
    output.data<float>(),
    per_tensor ? output_per_tensor.data<float>() : nullptr,
    ret.data<float>(),
    per_tensor ? ret_per_tensor.data<float>() : nullptr,
    per_tensor,
    max_chunks_per_tensor);

  return std::tuple<at::Tensor, at::Tensor>(ret, ret_per_tensor);
}