Module: TensorStream::OpenCLHelpers::MathOps

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

Overview

Collection of math functions for interfacing with OpenCL kernels

Class Method Summary collapse

Class Method Details

.included(klass) ⇒ Object 



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
 
# File 'lib/tensor_stream/opencl/math_ops.rb', line 5

def MathOps.included(klass)
  klass.class_eval do
    %i[max min add real_div div sub floor_mod mod mul pow sigmoid_grad squared_difference].each do |op|
      register_op op do |_context, tensor, inputs|
        execute_2_operand_func(op.to_s, tensor, inputs[0], inputs[1])
      end
    end

    register_op :add_n do |_context, tensor, inputs|
      if inputs.size == 1
        inputs[0]
      else
        work_group = if inputs[0].shape.size > 2
                       [ inputs[0].shape.reduce(:*) / inputs[0].shape.last, inputs[0].shape.last]
                     else
                       m, n = inputs[0].shape
                       [m || 1, n || 1]
                     end

        cl_m = OpenCL::Int1.new(work_group[0])
        cl_n = OpenCL::Int1.new(work_group[1])
        cl_switch = OpenCL::Int1.new(0)
        dtype = tensor.data_type

        output_buffer = _create_result_buffer(tensor.data_type, inputs[0].shape, "out_#{tensor.name}")
        inputs_queue = inputs.dup
        a = inputs_queue.pop
        until inputs_queue.empty?
          b = inputs_queue.pop
          event_wait_list = build_event_wait_list([a, b])
          method_call = :"add_#{a.data_type}_#{b.data_type}"
          event = _cl_program('add', a: a.data_type, b: b.data_type, dtype: dtype).send(method_call, _opencl_queue, work_group, cl_m, cl_n, cl_switch, a.cl_buffer, b.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
          a = output_buffer
          a.op = event
        end

        output_buffer.op = a.op
        output_buffer
      end
    end

    register_op :floor_div do |context, tensor, inputs|
      if fp_type?(tensor.data_type)
        execute_2_operand_func('floor_div', tensor, inputs[0], inputs[1])
      else
        execute_2_operand_func('div', tensor, inputs[0], inputs[1])
      end
    end

    register_op :mat_mul do |_context, tensor, inputs|
      a, b = inputs

      a_matrix_shape = a.shape.dup
      b_matrix_shape = b.shape.dup

      k = a_matrix_shape.pop
      m = a_matrix_shape.pop
      n = b_matrix_shape.pop
      v = b_matrix_shape.pop

      if tensor.options[:transpose_a]
        m, k = k, m
      end

      if tensor.options[:transpose_b]
        n, v = v, n
      end

      result_shape = [a_matrix_shape.first, m, n].compact
      work_group = [a_matrix_shape.first || 1, m, n]

      raise "#{tensor.inputs[0].name} rank must be greater than 1" if a.shape.size < 2
      raise "#{tensor.inputs[1].name} rank must be greater than 1" if b.shape.size < 2
      raise "#{tensor.inputs[0].name} unsupported rank" if b.shape.size > 3 || a.shape.size > 3
      raise "incompatible shape sizes for matrix multiplication (#{a.shape[1]} != #{b.shape[0]}) #{a.shape} vs #{b.shape}" if k != v

      dtype = tensor.data_type
      a, b = auto_type_cast(a, b, name: "#{tensor.name}/cast_#{a.name}_#{b.data_type}")
      output_buffer = _create_result_buffer(a.data_type, result_shape, tensor.name)

      cl_m = OpenCL::Int1.new(m)
      cl_n = OpenCL::Int1.new(n)
      cl_k = OpenCL::Int1.new(k)

      event_wait_list = build_event_wait_list([a, b])
      output_buffer.op = _cl_program('gemm', ta: !!tensor.options[:transpose_a], tb: !!tensor.options[:transpose_b], n: m * n, n_a: m * k, n_b: n * v, dtype: dtype).send(:"gemm_#{dtype}", _opencl_queue, work_group, cl_m, cl_n, cl_k, a.cl_buffer, b.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)

      output_buffer
    end

    register_op :bias_add do |context, tensor, inputs|
      value, bias = inputs
      output_buffer = _create_result_buffer(value.data_type, value.shape, tensor.name)
      result_shape = value.shape.dup
      bias_length = result_shape.pop
      work_group = [result_shape.reduce(:*)]
      event_wait_list = build_event_wait_list([value, bias])
      dtype = tensor.data_type
      output_buffer.op = _cl_program('bias_add', n: bias_length, dtype: dtype)
        .send(:"bias_add_#{dtype}", _opencl_queue, work_group, value.cl_buffer,
              bias.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
      output_buffer
    end

    register_op :bias_add_grad do |context, tensor, inputs|
      received_grad = inputs[0]
      bias_size = received_grad.shape.last
      output_buffer = _create_result_buffer(received_grad.data_type, [bias_size], tensor.name)
      work_group = [bias_size]

      received_grad_shape = received_grad.shape.dup
      received_grad_shape.pop
      item_rows = received_grad_shape.reduce(:*)
      dtype = tensor.data_type
      output_buffer.op = _cl_program('bias_add_grad', n: bias_size, rows: item_rows, dtype: dtype)
        .send(:"bias_add_grad_#{dtype}", _opencl_queue, work_group, received_grad.cl_buffer,
              output_buffer.cl_buffer, event_wait_list: build_event_wait_list([received_grad]))
      output_buffer
    end

    %i[sign exp tan acos asin sin cos abs sqrt negate square reciprocal tanh tanh_grad sigmoid log1p round floor ceil log].each do |op|
      register_op op, noop: true do |context, tensor, inputs|
        execute_func(op.to_s, tensor, inputs[0], context)
      end
    end

    %i[sum mean].each do |op|
      register_op op do |context, tensor, inputs|
        reduction(context, tensor, inputs[0], inputs[1], op.to_sym)
      end
    end

    register_op :prod do |context, tensor, inputs|
      if inputs[0].shape == [0]
        convert_to_opencl([1.0], [], data_type: inputs[0].data_type, name: tensor.name)
      else
        reduction(context, tensor, inputs[0], inputs[1], :prod)
      end
    end

    %i[argmin argmax].each do |op|
      register_op op do |context, tensor, inputs|
        value, axis = inputs
        rank = value.shape.size
        axis = 0 if axis.nil?

        axis = axis.is_a?(OpenCLBuffer) ? read_final_result(axis) : axis
        raise TensorStream::InvalidArgumentError, "Expected dimension in the range [#{-rank},#{rank}) but got #{axis}" if axis < -rank || axis >= rank

        reduce_multi_axis(context, tensor, value, axis, 'arg', op.to_sym)
       end
    end

    def reduction(child_context, tensor, value, axis, func)
      if axis.nil?
        value = _run(value, child_context)
        size = value.shape.reduce(:*) || 1
        if value.shape.empty? # for scalars, just return as is
          value
        else
          reduction_threads = 32
          items_per_thread_threshold = 4

          output_buffer = _create_result_buffer(value.data_type, [], tensor.name)
          event_wait_list = build_event_wait_list([value])

          if (size > reduction_threads) && ((size / reduction_threads) > items_per_thread_threshold)
            items_per_thread = size / reduction_threads
            extra_items = size % reduction_threads
            intermediate_output_buffer = _create_result_buffer(value.data_type, [reduction_threads], tensor.name)

            temp_values = if extra_items.zero?
                            _cl_program(func, dtype: value.data_type, index: 0, n: items_per_thread, w: items_per_thread).
                              send(:"#{func}_#{value.data_type}", _opencl_queue, [reduction_threads], value.cl_buffer, intermediate_output_buffer.cl_buffer, event_wait_list: event_wait_list)
                          else
                            [_cl_program(func, dtype: value.data_type, index: 0, n: items_per_thread, w: items_per_thread).
                              send(:"#{func}_#{value.data_type}", _opencl_queue, [reduction_threads - 1], value.cl_buffer, intermediate_output_buffer.cl_buffer, event_wait_list: event_wait_list),
                            _cl_program(func, dtype: value.data_type, index: reduction_threads - 1, n: items_per_thread + extra_items,  w: items_per_thread).send(:"#{func}_#{value.data_type}", _opencl_queue, [1], value.cl_buffer, intermediate_output_buffer.cl_buffer, event_wait_list: event_wait_list)]
                          end
            output_buffer.op = _cl_program(func, dtype: value.data_type, n: reduction_threads, index: 0, w: 0).send(:"#{func}_#{value.data_type}", _opencl_queue, [1], value.cl_buffer, output_buffer.cl_buffer, event_wait_list: temp_values)
            output_buffer
          else
            output_buffer.op = _cl_program(func, dtype: value.data_type, n: size, index: 0, w: 0).send(:"#{func}_#{value.data_type}", _opencl_queue, [1], value.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
            output_buffer
          end
         end
      else
        reduce_multi_axis(child_context, tensor, value, axis, 'reduce', func)
      end
    end

    def reduce_multi_axis(child_context, tensor, value, axis, prog, func)
      return value if value.shape.empty?

      rank = value.shape.size

      axis = axis.is_a?(OpenCLBuffer) ? read_final_result(axis) : axis
      axis = [axis] unless axis.is_a?(Array)
      return value if axis.empty?
      # remap negative values
      axis.map! { |axis| axis < 0 ? rank - axis.abs : axis }

      new_shape = value.shape.collect.with_index { |v, index| axis.include?(index) ? nil : v }.compact

      buffer_shape = tensor.options[:keepdims] ? _reduced_shape(value.shape.dup, axis) : new_shape
      output_buffer = _create_result_buffer(tensor.options[:output_type] || tensor.data_type, buffer_shape, tensor.name)

      work_group = new_shape.empty? ? [1] : new_shape
      dtype = value.data_type

      output_buffer.op = _cl_program("#{prog}_axis", f: func, axis: axis, shape: value.shape, o_shape: new_shape, dtype: dtype, out_dtype: tensor.options[:output_type])
          .send("#{prog}_axis_#{dtype}", _opencl_queue, work_group, value.cl_buffer,
                output_buffer.cl_buffer, event_wait_list: build_event_wait_list([value]))

      output_buffer
    end
  end
end