Module: TensorStream::OpenCLHelpers::NNOps

Included in:
Evaluator::OpenclEvaluator
Defined in:
lib/tensor_stream/opencl/nn_ops.rb

Overview

Collection of math functions for interfacing with OpenCL kernels

Class Method Summary collapse

Class Method Details

.included(klass) ⇒ Object 



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
 
# File 'lib/tensor_stream/opencl/nn_ops.rb', line 6

def NNOps.included(klass)
  klass.class_eval do

    # Fast in place multiply subtract assign
    register_op :apply_gradient_descent do |_context, tensor, inputs|
      _target_var, learning_rate, delta = inputs

      assign = tensor.inputs[0] || tensor

      assign.buffer.dirty = true # force buffer copy when variable is read externally
      output_buffer = assign.buffer

      m, n = output_buffer.shape
      work_group = [m || 1, n || 1]
      cl_m = OpenCL::Int1.new(m || 1)
      cl_n = OpenCL::Int1.new(n || 1)

      event_wait_list = build_event_wait_list([assign.buffer, learning_rate, delta])
      method_call = :"apply_gradient_#{output_buffer.data_type}"
      event = _cl_program("apply_gradient", dtype: output_buffer.data_type).send(method_call, _opencl_queue, work_group, cl_m, cl_n, delta.cl_buffer, learning_rate.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
      output_buffer.op = event
      output_buffer
    end

    # updates for gradient descent with momentum
    register_op :apply_momentum do |_context, tensor, inputs|
      target_var, momentum_var, learning_rate, grad, momentum = inputs

      assign = tensor.inputs[0] || tensor
      assign_acc = tensor.inputs[1]
      assign.buffer.dirty = true # force buffer copy when variable is read externally
      assign_acc.buffer.dirty = true # force buffer copy when variable is read externally

      output_buffer = assign.buffer

      m, n = output_buffer.shape
      work_group = [m || 1, n || 1]
      cl_m = OpenCL::Int1.new(m || 1)
      cl_n = OpenCL::Int1.new(n || 1)

      event_wait_list = build_event_wait_list([assign.buffer, assign_acc.buffer, learning_rate, grad, momentum])
      method_call = :"apply_momentum_#{output_buffer.data_type}"
      event = _cl_program("apply_momentum", nesterov: tensor.options[:use_nesterov], dtype: output_buffer.data_type).
                  send(method_call, _opencl_queue, work_group, cl_m, cl_n, grad.cl_buffer,
                      learning_rate.cl_buffer, momentum.cl_buffer, output_buffer.cl_buffer,
                      assign_acc.buffer.cl_buffer, event_wait_list: event_wait_list)
      output_buffer.op = event
      assign_acc.buffer.op = event
      output_buffer
    end

    register_op :apply_adadelta do |context, tensor, inputs|
      _target_var, _accum, _accum_update, lr, rho, epsilon, grad = inputs
      assign = tensor.inputs[0] || tensor
      assign_acc = tensor.inputs[1]
      assign_acc_update = tensor.inputs[2]

      # mark variable buffers as dirty
      assign.buffer.dirty = true # force buffer copy when variable is read externally
      assign_acc.buffer.dirty = true # force buffer copy when variable is read externally
      assign_acc_update.buffer.dirty = true # force buffer copy when variable is read externally

      output_buffer = assign.buffer

      m, n = output_buffer.shape
      work_group = [m || 1, n || 1]
      cl_m = OpenCL::Int1.new(m || 1)
      cl_n = OpenCL::Int1.new(n || 1)

      event_wait_list = build_event_wait_list(inputs)
      method_call = :"apply_adadelta_#{output_buffer.data_type}"
      event = _cl_program('apply_adadelta', dtype: output_buffer.data_type)
                          .send(method_call, _opencl_queue, work_group, cl_m, cl_n,
                                lr.cl_buffer,
                                rho.cl_buffer,
                                epsilon.cl_buffer,
                                grad.cl_buffer,
                                assign.buffer.cl_buffer,
                                assign_acc.buffer.cl_buffer,
                                assign_acc_update.buffer.cl_buffer,
                                event_wait_list: event_wait_list)
      output_buffer.op = event
      assign_acc.buffer.op = event
      assign_acc_update.buffer.op = event
      output_buffer
    end

    # Adam optimization algorithm
    register_op :apply_adam do |_context, tensor, inputs|
      _target_var, _m, _v, beta1_power, beta2_power, lr_t, beta1_t, beta2_t, epsilon_t, grad = inputs

      assign = tensor.inputs[0] || tensor
      assign_m = tensor.inputs[1]
      assign_v = tensor.inputs[2]

      # mark variable buffers as dirty
      assign.buffer.dirty = true # force buffer copy when variable is read externally
      assign_m.buffer.dirty = true # force buffer copy when variable is read externally
      assign_v.buffer.dirty = true # force buffer copy when variable is read externally

      output_buffer = assign.buffer

      m, n = output_buffer.shape
      work_group = [m || 1, n || 1]
      cl_m = OpenCL::Int1.new(m || 1)
      cl_n = OpenCL::Int1.new(n || 1)

      event_wait_list = build_event_wait_list(inputs)
      method_call = :"apply_adam_#{output_buffer.data_type}"
      event = _cl_program("apply_adam", dtype: output_buffer.data_type)
                          .send(method_call, _opencl_queue, work_group, cl_m, cl_n,
                                grad.cl_buffer,
                                lr_t.cl_buffer,
                                beta1_power.cl_buffer,
                                beta2_power.cl_buffer,
                                beta1_t.cl_buffer,
                                beta2_t.cl_buffer,
                                epsilon_t.cl_buffer,
                                assign_m.buffer.cl_buffer,
                                assign.buffer.cl_buffer,
                                assign_v.buffer.cl_buffer,
                                event_wait_list: event_wait_list)
      output_buffer.op = event
      assign_m.buffer.op = event
      assign_v.buffer.op = event
      output_buffer
    end

    register_op :softmax do |_context, tensor, inputs|
      a = inputs[0]
      event_wait_list = build_event_wait_list(inputs)
      dtype = tensor.data_type
      output_buffer = _create_result_buffer(tensor.data_type, a.shape, tensor.name)

      m, n = a.shape
      work_group = [m]
      n = m if n.nil?
      cl_n = OpenCL::Int1.new(n || 1)

      event = _cl_program("softmax", dtype: dtype).send(:"softmax_#{dtype}", _opencl_queue, work_group, cl_n, a.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
      output_buffer.op = event
      output_buffer
    end

    register_op :log_softmax do |_context, tensor, inputs|
      a = inputs[0] # logits
      event_wait_list = build_event_wait_list(inputs)
      dtype = tensor.data_type
      output_buffer = _create_result_buffer(tensor.data_type, a.shape, tensor.name)

      m, n = a.shape
      work_group = [m]
      n = m if n.nil?
      cl_n = OpenCL::Int1.new(n || 1)

      event = _cl_program("log_softmax", dtype: dtype).send(:"log_softmax_#{dtype}", _opencl_queue, work_group, cl_n, a.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
      output_buffer.op = event
      output_buffer
    end

    register_op :softmax_cross_entropy_with_logits_v2 do |context, tensor, inputs|
      a = inputs[0] # logits
      b = inputs[1] # labels
      event_wait_list = build_event_wait_list(inputs)
      dtype = tensor.data_type
      output_buffer = _create_result_buffer(tensor.data_type, a.shape, tensor.name)
      output_buffer_backprop = _create_result_buffer(tensor.data_type, a.shape, "#{tensor.name}_2")
      rank = a.shape.size - 1
      m, n = a.shape
      work_group = [m]
      n = m if n.nil?
      cl_n = OpenCL::Int1.new(n || 1)

      event = _cl_program("softmax_cross", dtype: dtype).send(:"softmax_cross_#{dtype}", _opencl_queue, work_group, cl_n, a.cl_buffer, b.cl_buffer,
                           output_buffer.cl_buffer, output_buffer_backprop.cl_buffer, event_wait_list: event_wait_list)
      output_buffer.op = event
      output_buffer_backprop.op = event

      loss = reduction(context, tensor, output_buffer, rank, :sum)
      TensorStream::Evaluator::OutputGroup.new([loss, output_buffer_backprop],  [tensor.inputs[0].data_type, tensor.inputs[0].data_type])
    end

    register_op :softmax_cross_entropy_with_logits_v2_grad do |_context, tensor, inputs|
      a = inputs[0] # logits
      b = inputs[1] # labels
      c = inputs[2] # grads
      event_wait_list = build_event_wait_list(inputs)
      dtype = tensor.data_type
      output_buffer = _create_result_buffer(tensor.data_type, a.shape, tensor.name)

      m, n = a.shape
      work_group = [m]
      n = m if n.nil?
      cl_n = OpenCL::Int1.new(n || 1)

      event = _cl_program("softmax_cross_grad", dtype: dtype).send(:"softmax_cross_grad_#{dtype}", _opencl_queue, work_group, cl_n, a.cl_buffer, b.cl_buffer, c.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
      output_buffer.op = event
      output_buffer
    end

    register_op :softmax_grad do |_context, tensor, inputs|
      a, grad = inputs

      event_wait_list = build_event_wait_list(inputs)
      dtype = tensor.data_type
      output_buffer = _create_result_buffer(tensor.data_type, a.shape, tensor.name)

      m, n = a.shape
      work_group = [m]
      n = m if n.nil?
      cl_n = OpenCL::Int1.new(n || 1)
      event = _cl_program('softmax_grad', dtype: dtype, size: n).send(:"softmax_grad_#{dtype}", _opencl_queue, work_group, cl_n, a.cl_buffer, grad.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
      output_buffer.op = event
      output_buffer
    end
  end
end