tensor_stream 0.2.0 → 0.3.0
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +5 -5
- data/.circleci/config.yml +2 -1
- data/CHANGELOG.md +5 -0
- data/README.md +28 -1
- data/benchmark/benchmark.rb +129 -0
- data/lib/tensor_stream.rb +7 -4
- data/lib/tensor_stream/evaluator/buffer.rb +10 -0
- data/lib/tensor_stream/evaluator/evaluator.rb +1 -0
- data/lib/tensor_stream/evaluator/kernels/_bool_operand.cl +45 -0
- data/lib/tensor_stream/evaluator/kernels/_operand.cl +45 -0
- data/lib/tensor_stream/evaluator/kernels/abs.cl +16 -0
- data/lib/tensor_stream/evaluator/kernels/add.cl +5 -0
- data/lib/tensor_stream/evaluator/kernels/argmax.cl +15 -0
- data/lib/tensor_stream/evaluator/kernels/argmin.cl +15 -0
- data/lib/tensor_stream/evaluator/kernels/cast.cl +15 -0
- data/lib/tensor_stream/evaluator/kernels/cond.cl.erb +5 -0
- data/lib/tensor_stream/evaluator/kernels/cos.cl +7 -0
- data/lib/tensor_stream/evaluator/kernels/div.cl.erb +5 -0
- data/lib/tensor_stream/evaluator/kernels/exp.cl +7 -0
- data/lib/tensor_stream/evaluator/kernels/gemm.cl +63 -0
- data/lib/tensor_stream/evaluator/kernels/log.cl +7 -0
- data/lib/tensor_stream/evaluator/kernels/log1p.cl +7 -0
- data/lib/tensor_stream/evaluator/kernels/max.cl +91 -0
- data/lib/tensor_stream/evaluator/kernels/mul.cl +5 -0
- data/lib/tensor_stream/evaluator/kernels/negate.cl +15 -0
- data/lib/tensor_stream/evaluator/kernels/pow.cl +130 -0
- data/lib/tensor_stream/evaluator/kernels/reciprocal.cl +15 -0
- data/lib/tensor_stream/evaluator/kernels/round.cl +7 -0
- data/lib/tensor_stream/evaluator/kernels/sigmoid.cl +8 -0
- data/lib/tensor_stream/evaluator/kernels/sigmoid_grad.cl +54 -0
- data/lib/tensor_stream/evaluator/kernels/sign.cl +23 -0
- data/lib/tensor_stream/evaluator/kernels/sin.cl +8 -0
- data/lib/tensor_stream/evaluator/kernels/sqrt.cl +8 -0
- data/lib/tensor_stream/evaluator/kernels/square.cl +15 -0
- data/lib/tensor_stream/evaluator/kernels/sub.cl +5 -0
- data/lib/tensor_stream/evaluator/kernels/tan.cl +7 -0
- data/lib/tensor_stream/evaluator/kernels/tanh.cl +7 -0
- data/lib/tensor_stream/evaluator/kernels/tanh_grad.cl +6 -0
- data/lib/tensor_stream/evaluator/kernels/where.cl +15 -0
- data/lib/tensor_stream/evaluator/opencl_buffer.rb +30 -0
- data/lib/tensor_stream/evaluator/opencl_evaluator.rb +1095 -0
- data/lib/tensor_stream/evaluator/opencl_template_helper.rb +58 -0
- data/lib/tensor_stream/evaluator/operation_helpers/array_ops_helper.rb +27 -0
- data/lib/tensor_stream/evaluator/ruby_evaluator.rb +20 -31
- data/lib/tensor_stream/graph.rb +4 -2
- data/lib/tensor_stream/math_gradients.rb +3 -0
- data/lib/tensor_stream/operation.rb +29 -2
- data/lib/tensor_stream/ops.rb +14 -2
- data/lib/tensor_stream/placeholder.rb +1 -1
- data/lib/tensor_stream/session.rb +10 -3
- data/lib/tensor_stream/tensor_shape.rb +1 -1
- data/lib/tensor_stream/train/saver.rb +1 -1
- data/lib/tensor_stream/variable.rb +7 -1
- data/lib/tensor_stream/version.rb +1 -1
- data/samples/logistic_regression.rb +2 -1
- data/samples/nearest_neighbor.rb +54 -0
- data/tensor_stream.gemspec +3 -1
- metadata +107 -28
@@ -0,0 +1,23 @@
|
|
1
|
+
__kernel void sign_fp(const int M, const int N, __global const float *A, __global float *C) {
|
2
|
+
// Get the index of the current element to be processed
|
3
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
4
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
5
|
+
float value = A[globalRow * N + globalCol];
|
6
|
+
if (isnan(value) || value == 0.0f) {
|
7
|
+
C[globalRow * N + globalCol] = 0.0;
|
8
|
+
} else {
|
9
|
+
C[globalRow * N + globalCol] = value < 0 ? -1.0 : 1.0;
|
10
|
+
}
|
11
|
+
}
|
12
|
+
|
13
|
+
__kernel void sign_int(const int M, const int N, __global const int *A, __global int *C) {
|
14
|
+
// Get the index of the current element to be processed
|
15
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
16
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
17
|
+
float value = A[globalRow * N + globalCol];
|
18
|
+
if (isnan(value) || value == 0) {
|
19
|
+
C[globalRow * N + globalCol] = 0;
|
20
|
+
} else {
|
21
|
+
C[globalRow * N + globalCol] = value < 0 ? -1 : 1;
|
22
|
+
}
|
23
|
+
}
|
@@ -0,0 +1,8 @@
|
|
1
|
+
|
2
|
+
__kernel void sin_fp(const int M, const int N, __global const float *A, __global float *C) {
|
3
|
+
// Get the index of the current element to be processed
|
4
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
5
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
6
|
+
|
7
|
+
C[globalRow * N + globalCol] = sin(A[globalRow * N + globalCol]);
|
8
|
+
}
|
@@ -0,0 +1,8 @@
|
|
1
|
+
|
2
|
+
__kernel void sqrt_fp(const int M, const int N, __global const float *A, __global float *C) {
|
3
|
+
// Get the index of the current element to be processed
|
4
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
5
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
6
|
+
|
7
|
+
C[globalRow * N + globalCol] = sqrt(A[globalRow * N + globalCol]);
|
8
|
+
}
|
@@ -0,0 +1,15 @@
|
|
1
|
+
__kernel void square_fp(const int M, const int N, __global const float *A, __global float *C) {
|
2
|
+
// Get the index of the current element to be processed
|
3
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
4
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
5
|
+
|
6
|
+
C[globalRow * N + globalCol] = A[globalRow * N + globalCol] * A[globalRow * N + globalCol];
|
7
|
+
}
|
8
|
+
|
9
|
+
__kernel void square_int(const int M, const int N, __global const int *A, __global int *C) {
|
10
|
+
// Get the index of the current element to be processed
|
11
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
12
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
13
|
+
|
14
|
+
C[globalRow * N + globalCol] = A[globalRow * N + globalCol] * A[globalRow * N + globalCol];
|
15
|
+
}
|
@@ -0,0 +1,7 @@
|
|
1
|
+
__kernel void tan_fp(const int M, const int N, __global const float *A, __global float *C) {
|
2
|
+
// Get the index of the current element to be processed
|
3
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
4
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
5
|
+
|
6
|
+
C[globalRow * N + globalCol] = tan(A[globalRow * N + globalCol]);
|
7
|
+
}
|
@@ -0,0 +1,7 @@
|
|
1
|
+
__kernel void tanh_fp(const int M, const int N, __global const float *A, __global float *C) {
|
2
|
+
// Get the index of the current element to be processed
|
3
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
4
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
5
|
+
|
6
|
+
C[globalRow * N + globalCol] = tanh(A[globalRow * N + globalCol]);
|
7
|
+
}
|
@@ -0,0 +1,6 @@
|
|
1
|
+
__kernel void tanh_grad_fp(const int M, const int N, __global const float *A, __global float *C) {
|
2
|
+
// Get the index of the current element to be processed
|
3
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
4
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
5
|
+
C[globalRow * N + globalCol] = 1 - tanh(A[globalRow * N + globalCol]) * tanh(A[globalRow * N + globalCol]);
|
6
|
+
}
|
@@ -0,0 +1,15 @@
|
|
1
|
+
__kernel void where_fp(const int M, const int N, __global const int *PRED, __global const float *A, __global const float *B, __global float *C) {
|
2
|
+
// Get the index of the current element to be processed
|
3
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
4
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
5
|
+
|
6
|
+
C[globalRow * N + globalCol] = PRED[globalRow * N + globalCol] ? A[globalRow * N + globalCol] : B[globalRow * N + globalCol];
|
7
|
+
}
|
8
|
+
|
9
|
+
__kernel void where_int(const int M, const int N, __global const int *PRED, __global const int *A, __global const int *B, __global int *C) {
|
10
|
+
// Get the index of the current element to be processed
|
11
|
+
const int globalRow = get_global_id(0); // Row ID of C (0..M)
|
12
|
+
const int globalCol = get_global_id(1); // Col ID of C (0..N)
|
13
|
+
|
14
|
+
C[globalRow * N + globalCol] = PRED[globalRow * N + globalCol] ? A[globalRow * N + globalCol] : B[globalRow * N + globalCol];
|
15
|
+
}
|
@@ -0,0 +1,30 @@
|
|
1
|
+
module TensorStream
|
2
|
+
class OpenCLBuffer < Buffer
|
3
|
+
include ArrayOpsHelper
|
4
|
+
|
5
|
+
attr_accessor :data_type, :shape, :buffer, :cl_buffer, :op
|
6
|
+
|
7
|
+
def initialize(data_type: , shape:, buffer:, cl_buffer:, op: nil, name: nil)
|
8
|
+
@data_type = data_type
|
9
|
+
@shape = shape
|
10
|
+
@buffer = buffer
|
11
|
+
@cl_buffer = cl_buffer
|
12
|
+
@name = name
|
13
|
+
@op = op
|
14
|
+
end
|
15
|
+
|
16
|
+
def to_ruby
|
17
|
+
return [] if buffer.empty?
|
18
|
+
if shape.empty?
|
19
|
+
return buffer[0] != 0 if data_type == :boolean
|
20
|
+
return buffer[0]
|
21
|
+
end
|
22
|
+
|
23
|
+
result = buffer.reshape(*shape.reverse).to_a
|
24
|
+
if data_type == :boolean
|
25
|
+
result = process_function_op(result, ->(a, _b) { a != 0 })
|
26
|
+
end
|
27
|
+
result
|
28
|
+
end
|
29
|
+
end
|
30
|
+
end
|
@@ -0,0 +1,1095 @@
|
|
1
|
+
require 'tensor_stream/evaluator/operation_helpers/random_gaussian'
|
2
|
+
require 'tensor_stream/evaluator/operation_helpers/array_ops_helper'
|
3
|
+
require 'tensor_stream/evaluator/operation_helpers/math_helper'
|
4
|
+
require 'tensor_stream/evaluator/opencl_buffer'
|
5
|
+
require 'tensor_stream/evaluator/opencl_template_helper'
|
6
|
+
require 'distribution'
|
7
|
+
require 'opencl_ruby_ffi'
|
8
|
+
require 'narray_ffi'
|
9
|
+
|
10
|
+
module TensorStream
|
11
|
+
module Evaluator
|
12
|
+
class FullEvalNotPossible < RuntimeError
|
13
|
+
end
|
14
|
+
|
15
|
+
# Errors during graph evaluation
|
16
|
+
class EvaluatorExcecutionException < RuntimeError
|
17
|
+
attr_reader :tensor
|
18
|
+
|
19
|
+
def initialize(exception, tensor)
|
20
|
+
@exception = exception
|
21
|
+
@tensor = tensor
|
22
|
+
end
|
23
|
+
|
24
|
+
def wrapped_exception
|
25
|
+
@exception
|
26
|
+
end
|
27
|
+
end
|
28
|
+
|
29
|
+
## PURE ruby evaluator used for testing and development
|
30
|
+
class OpenclEvaluator
|
31
|
+
attr_accessor :retain
|
32
|
+
|
33
|
+
include TensorStream::OpHelper
|
34
|
+
include TensorStream::ArrayOpsHelper
|
35
|
+
include TensorStream::MathHelper
|
36
|
+
|
37
|
+
def initialize(session, context, thread_pool: nil, log_intermediates: false, preferred_device: nil)
|
38
|
+
@session = session
|
39
|
+
@context = context
|
40
|
+
@log_intermediates = log_intermediates
|
41
|
+
@preferred_device = preferred_device
|
42
|
+
@retain = context[:retain] || []
|
43
|
+
@thread_pool = thread_pool || Concurrent::ImmediateExecutor.new
|
44
|
+
|
45
|
+
@context[:compute_history] = [] if log_intermediates
|
46
|
+
end
|
47
|
+
|
48
|
+
# opencl evaluator main entrypoint
|
49
|
+
def run(tensor, execution_context)
|
50
|
+
_create_opencl_context
|
51
|
+
# _prepare_kernels
|
52
|
+
|
53
|
+
read_final_result(complete_eval(tensor, execution_context))
|
54
|
+
end
|
55
|
+
|
56
|
+
def complete_eval(tensor, context)
|
57
|
+
create_command_queue
|
58
|
+
buffer = _run(tensor, context)
|
59
|
+
if buffer.is_a?(Array)
|
60
|
+
buffer = buffer.collect do |b|
|
61
|
+
next b if b.buffer.size.zero?
|
62
|
+
_opencl_queue.enqueue_read_buffer(b.cl_buffer, b.buffer, event_wait_list: [b.op].compact)
|
63
|
+
b
|
64
|
+
end
|
65
|
+
else
|
66
|
+
return buffer if buffer.nil? || buffer.buffer.size.zero?
|
67
|
+
_opencl_queue.enqueue_read_buffer(buffer.cl_buffer, buffer.buffer, event_wait_list: [buffer.op].compact)
|
68
|
+
end
|
69
|
+
|
70
|
+
_opencl_queue.finish
|
71
|
+
buffer
|
72
|
+
end
|
73
|
+
|
74
|
+
def opencl_device
|
75
|
+
@context[:_cache][:_opencl_device]
|
76
|
+
end
|
77
|
+
|
78
|
+
protected
|
79
|
+
|
80
|
+
# read result from opencl and convert to ruby
|
81
|
+
def read_final_result(buffer)
|
82
|
+
return buffer.map { |b| read_final_result(b) } if buffer.is_a?(Array)
|
83
|
+
return nil if buffer.nil?
|
84
|
+
|
85
|
+
buffer.to_ruby
|
86
|
+
end
|
87
|
+
|
88
|
+
def _create_opencl_context
|
89
|
+
@context[:_cache][:_opencl_device] ||= begin
|
90
|
+
if @preferred_device
|
91
|
+
@preferred_device
|
92
|
+
else
|
93
|
+
device, _score, _platform, _index = choose_best_device
|
94
|
+
device
|
95
|
+
end
|
96
|
+
end
|
97
|
+
@context[:_cache][:_opencl_context] ||= OpenCL.create_context(opencl_device)
|
98
|
+
end
|
99
|
+
|
100
|
+
def choose_best_device
|
101
|
+
@best_device ||= begin
|
102
|
+
devices = OpenCL.platforms.flat_map do |p|
|
103
|
+
p.devices.select { |d| d.available > 0 }.each_with_index.collect do |d, index|
|
104
|
+
score = 0
|
105
|
+
if d.type.to_s == 'CPU'
|
106
|
+
score += 1
|
107
|
+
elsif d.type.to_s == 'GPU'
|
108
|
+
score += 4
|
109
|
+
end
|
110
|
+
|
111
|
+
score += d.max_compute_units
|
112
|
+
|
113
|
+
[d, score, p.name, index]
|
114
|
+
end
|
115
|
+
end
|
116
|
+
end
|
117
|
+
devices.max { |a| a[1] }
|
118
|
+
end
|
119
|
+
|
120
|
+
def create_command_queue
|
121
|
+
supported_proprties = opencl_device.queue_properties.names
|
122
|
+
properties = []
|
123
|
+
properties << OpenCL::CommandQueue::PROFILING_ENABLE if supported_proprties.include?('PROFILING_ENABLE')
|
124
|
+
properties << OpenCL::CommandQueue::OUT_OF_ORDER_EXEC_MODE_ENABLE if supported_proprties.include?('OUT_OF_ORDER_EXEC_MODE_ENABLE')
|
125
|
+
@context[:_cache][:_opencl_queue] ||= _opencl_context.create_command_queue(opencl_device, properties: properties)
|
126
|
+
end
|
127
|
+
|
128
|
+
def _opencl_context
|
129
|
+
@context[:_cache][:_opencl_context]
|
130
|
+
end
|
131
|
+
|
132
|
+
def _opencl_queue
|
133
|
+
@context[:_cache][:_opencl_queue]
|
134
|
+
end
|
135
|
+
|
136
|
+
def cl_template_path(kernel, extension)
|
137
|
+
File.join(File.dirname(__FILE__), 'kernels', "#{kernel}.#{extension}")
|
138
|
+
end
|
139
|
+
|
140
|
+
def _cl_program(kernel)
|
141
|
+
@context[:_cache]["_opencl_kernel_#{kernel}"] ||= begin
|
142
|
+
filename = %w[cl.erb cl].map { |ext| cl_template_path(kernel, ext) }.find { |n| File.exist?(n) }
|
143
|
+
source = File.read(filename)
|
144
|
+
source = OpenclTemplateHelper.new(source).generate
|
145
|
+
program = _opencl_context.create_program_with_source(source)
|
146
|
+
program.build
|
147
|
+
rescue OpenCL::Error::BUILD_PROGRAM_FAILURE => e
|
148
|
+
puts "OpenCL Compile error: #{program.build_log}"
|
149
|
+
raise e
|
150
|
+
end
|
151
|
+
end
|
152
|
+
|
153
|
+
def _run(tensor, execution_context)
|
154
|
+
return tensor if tensor.is_a?(OpenCLBuffer)
|
155
|
+
return tensor.map { |t| _run(t, execution_context) } if tensor.is_a?(Array)
|
156
|
+
|
157
|
+
return tensor if retain.include?(tensor) # if var is in retain don't eval to value
|
158
|
+
|
159
|
+
tensor = tensor.call if tensor.is_a?(Proc)
|
160
|
+
|
161
|
+
child_context = execution_context.dup
|
162
|
+
res = if tensor.is_a?(Operation)
|
163
|
+
eval_operation(tensor, child_context)
|
164
|
+
elsif tensor.is_a?(Variable)
|
165
|
+
eval_variable(tensor, child_context)
|
166
|
+
elsif tensor.is_a?(Placeholder)
|
167
|
+
resolve_placeholder(tensor, child_context)
|
168
|
+
else
|
169
|
+
eval_tensor(tensor, child_context)
|
170
|
+
end
|
171
|
+
execution_context.deep_merge!(returns: child_context[:returns])
|
172
|
+
res
|
173
|
+
end
|
174
|
+
|
175
|
+
def eval_variable(tensor, child_context)
|
176
|
+
raise "variable #{tensor.name} not initalized" if tensor.value.nil? && (tensor.buffer.nil? || !tensor.buffer.dirty)
|
177
|
+
tensor.buffer = wrap_opencl(tensor, name: tensor.name) if tensor.buffer.nil?
|
178
|
+
tensor.buffer
|
179
|
+
end
|
180
|
+
|
181
|
+
def eval_operation(tensor, child_context)
|
182
|
+
return @context[tensor.name] if @context.key?(tensor.name)
|
183
|
+
|
184
|
+
a = resolve_placeholder(tensor.items[0], child_context) if tensor.items && tensor.items[0]
|
185
|
+
b = resolve_placeholder(tensor.items[1], child_context) if tensor.items && tensor.items[1]
|
186
|
+
|
187
|
+
case tensor.operation
|
188
|
+
when :concat
|
189
|
+
input_a = read_final_result(complete_eval(a, child_context))
|
190
|
+
arr = concat_array(input_a, tensor.options[:axis])
|
191
|
+
convert_to_opencl(arr.flatten, shape_eval(arr), data_type: tensor.data_type, name: tensor.name)
|
192
|
+
when :cond
|
193
|
+
pred = complete_eval(tensor.options[:pred], child_context)
|
194
|
+
a = _run(a, child_context)
|
195
|
+
b = _run(b, child_context)
|
196
|
+
|
197
|
+
if all_true?(pred.buffer)
|
198
|
+
a
|
199
|
+
else
|
200
|
+
b
|
201
|
+
end
|
202
|
+
when :identity
|
203
|
+
_run(a, child_context)
|
204
|
+
when :eye
|
205
|
+
rows = complete_eval(a, child_context)
|
206
|
+
columns = complete_eval(b, child_context)
|
207
|
+
shape = [rows.buffer[0], columns.buffer[0]]
|
208
|
+
eye_arr = Array.new(rows.buffer[0]) do |i|
|
209
|
+
Array.new(columns.buffer[0]) do |col|
|
210
|
+
if fp_type?(tensor.data_type)
|
211
|
+
i == col ? 1.0 : 0.0
|
212
|
+
else
|
213
|
+
i == col ? 1 : 0
|
214
|
+
end
|
215
|
+
end
|
216
|
+
end
|
217
|
+
|
218
|
+
convert_to_opencl(eye_arr.flatten, shape, data_type: tensor.data_type, name: tensor.name)
|
219
|
+
when :pad
|
220
|
+
a = read_final_result(complete_eval(a, child_context))
|
221
|
+
p = read_final_result(complete_eval(tensor.options[:paddings], child_context))
|
222
|
+
|
223
|
+
padding = arr_pad(a, p, tensor.data_type)
|
224
|
+
convert_to_opencl(padding.flatten, shape_eval(padding), data_type: tensor.data_type, name: tensor.name)
|
225
|
+
when :tile
|
226
|
+
input = read_final_result(complete_eval(a, child_context))
|
227
|
+
multiples = read_final_result(complete_eval(b, child_context))
|
228
|
+
|
229
|
+
rank = get_rank(input)
|
230
|
+
raise '1D or higher tensor required' if rank.zero?
|
231
|
+
raise "invalid multiple size passed #{rank} != #{multiples.size}" if rank != multiples.size
|
232
|
+
|
233
|
+
tile = tile_arr(input, 0, multiples)
|
234
|
+
arr = tile.nil? ? [] : tile
|
235
|
+
convert_to_opencl(arr.flatten, shape_eval(arr), data_type: tensor.data_type, name: tensor.name)
|
236
|
+
when :assign
|
237
|
+
assign_var(tensor, b, child_context)
|
238
|
+
when :assign_add
|
239
|
+
a = _run(a, child_context)
|
240
|
+
b = _run(b, child_context)
|
241
|
+
|
242
|
+
value = execute_2_operand_func('add', tensor, a, b, child_context)
|
243
|
+
assign_var(tensor, value, child_context)
|
244
|
+
when :assign_sub
|
245
|
+
a = _run(a, child_context)
|
246
|
+
b = _run(b, child_context)
|
247
|
+
|
248
|
+
value = execute_2_operand_func('sub', tensor, a, b, child_context)
|
249
|
+
assign_var(tensor, value, child_context)
|
250
|
+
when :less
|
251
|
+
execute_2_operand_func('less', tensor, a, b, child_context, 'cond')
|
252
|
+
when :less_equal
|
253
|
+
execute_2_operand_func('less_equal', tensor, a, b, child_context, 'cond')
|
254
|
+
when :greater
|
255
|
+
execute_2_operand_func('greater', tensor, a, b, child_context, 'cond')
|
256
|
+
when :greater_equal
|
257
|
+
execute_2_operand_func('greater_equal', tensor, a, b, child_context, 'cond')
|
258
|
+
when :equal
|
259
|
+
execute_2_operand_func('equal', tensor, a, b, child_context, 'cond')
|
260
|
+
when :not_equal
|
261
|
+
execute_2_operand_func('not_equal', tensor, a, b, child_context, 'cond')
|
262
|
+
when :logical_and
|
263
|
+
execute_2_operand_func('logical_and', tensor, a, b, child_context, 'cond')
|
264
|
+
when :where
|
265
|
+
pred = tensor.options[:pred]
|
266
|
+
execute_cond_func('where', tensor, pred, a, b, child_context)
|
267
|
+
when :max
|
268
|
+
execute_2_operand_func('max', tensor, a, b, child_context)
|
269
|
+
when :add
|
270
|
+
execute_2_operand_func('add', tensor, a, b, child_context)
|
271
|
+
when :div
|
272
|
+
execute_2_operand_func('div', tensor, a, b, child_context)
|
273
|
+
when :sub
|
274
|
+
execute_2_operand_func('sub', tensor, a, b, child_context)
|
275
|
+
when :matmul
|
276
|
+
a = _run(a, child_context)
|
277
|
+
b = _run(b, child_context)
|
278
|
+
|
279
|
+
m = a.shape[0]
|
280
|
+
n = b.shape[1]
|
281
|
+
v = b.shape[0]
|
282
|
+
k = a.shape[1]
|
283
|
+
|
284
|
+
m, k = [a.shape[1], a.shape[0]] if tensor.options[:transpose_a]
|
285
|
+
n, v = [b.shape[0], b.shape[1]] if tensor.options[:transpose_b]
|
286
|
+
|
287
|
+
result_shape = [m, n]
|
288
|
+
|
289
|
+
raise "#{tensor.items[0].name} rank must be greater than 1" if a.shape.size < 2
|
290
|
+
raise "#{tensor.items[1].name} rank must be greater than 1" if b.shape.size < 2
|
291
|
+
raise "incompatible shape sizes for matrix multiplication (#{a.shape[1]} != #{b.shape[0]}) #{a.shape} vs #{b.shape}" if k != v
|
292
|
+
|
293
|
+
dtype = TensorStream::Ops::FLOATING_POINT_TYPES.include?(tensor.data_type) ? 'fp' : 'int'
|
294
|
+
a, b = type_cast(a, b)
|
295
|
+
output_buffer = _create_result_buffer(a.data_type, result_shape, tensor.name)
|
296
|
+
|
297
|
+
cl_m = OpenCL::Int1.new(m)
|
298
|
+
cl_n = OpenCL::Int1.new(n)
|
299
|
+
cl_k = OpenCL::Int1.new(k)
|
300
|
+
|
301
|
+
transpose_a = OpenCL::Int1.new(tensor.options[:transpose_a] ? 1 : 0)
|
302
|
+
transpose_b = OpenCL::Int1.new(tensor.options[:transpose_b] ? 1 : 0)
|
303
|
+
|
304
|
+
output_buffer.op = _cl_program('gemm').send(:"gemm_#{dtype}", _opencl_queue, result_shape, cl_m, cl_n, cl_k, transpose_a, transpose_b, a.cl_buffer, b.cl_buffer, output_buffer.cl_buffer)
|
305
|
+
output_buffer
|
306
|
+
when :mul
|
307
|
+
execute_2_operand_func('mul', tensor, a, b, child_context)
|
308
|
+
when :pow
|
309
|
+
execute_2_operand_func('pow', tensor, a, b, child_context)
|
310
|
+
when :cast
|
311
|
+
a = _run(a, child_context)
|
312
|
+
if a.data_type != tensor.data_type
|
313
|
+
buffer = _create_result_buffer(tensor.data_type, a.shape, tensor.name)
|
314
|
+
s_dtype = TensorStream::Ops::FLOATING_POINT_TYPES.include?(a.data_type) ? 'fp' : 'int'
|
315
|
+
t_dtype = TensorStream::Ops::FLOATING_POINT_TYPES.include?(tensor.data_type) ? 'fp' : 'int'
|
316
|
+
m, n = a.shape
|
317
|
+
cl_m = OpenCL::Int1.new(m || 1)
|
318
|
+
cl_n = OpenCL::Int1.new(n || 1)
|
319
|
+
work_group = [m || 1, n || 1]
|
320
|
+
|
321
|
+
buffer.op = _cl_program("cast").send(:"cast_#{s_dtype}_#{t_dtype}",_opencl_queue, work_group, cl_m, cl_n, a.cl_buffer, buffer.cl_buffer)
|
322
|
+
buffer
|
323
|
+
else
|
324
|
+
a
|
325
|
+
end
|
326
|
+
when :sign
|
327
|
+
execute_func('sign', tensor, a, child_context)
|
328
|
+
when :exp
|
329
|
+
execute_func('exp', tensor, a, child_context)
|
330
|
+
when :log
|
331
|
+
execute_func('log', tensor, a, child_context)
|
332
|
+
when :sin
|
333
|
+
execute_func('sin', tensor, a, child_context)
|
334
|
+
when :tan
|
335
|
+
execute_func('tan', tensor, a, child_context)
|
336
|
+
when :cos
|
337
|
+
execute_func('cos', tensor, a, child_context)
|
338
|
+
when :abs
|
339
|
+
execute_func('abs', tensor, a, child_context)
|
340
|
+
when :sqrt
|
341
|
+
execute_func('sqrt', tensor, a, child_context)
|
342
|
+
when :negate
|
343
|
+
execute_func('negate', tensor, a, child_context)
|
344
|
+
when :square
|
345
|
+
execute_func('square', tensor, a, child_context)
|
346
|
+
when :reciprocal
|
347
|
+
execute_func('reciprocal', tensor, a, child_context)
|
348
|
+
when :tanh
|
349
|
+
execute_func('tanh', tensor, a, child_context)
|
350
|
+
when :tanh_grad
|
351
|
+
execute_func('tanh_grad', tensor, a, child_context)
|
352
|
+
when :sigmoid
|
353
|
+
execute_func('sigmoid', tensor, a, child_context)
|
354
|
+
when :log1p
|
355
|
+
execute_func('log1p', tensor, a, child_context)
|
356
|
+
when :round
|
357
|
+
execute_func('round', tensor, a, child_context)
|
358
|
+
when :sigmoid_grad
|
359
|
+
execute_2_operand_func('sigmoid_grad', tensor, a, b, child_context)
|
360
|
+
when :truncate
|
361
|
+
a = _run(a, child_context)
|
362
|
+
b = _run(b, child_context)
|
363
|
+
|
364
|
+
if a.shape.size.zero?
|
365
|
+
a
|
366
|
+
else
|
367
|
+
input_b = read_final_result(b)
|
368
|
+
if a.shape == input_b
|
369
|
+
a
|
370
|
+
else
|
371
|
+
input_a = read_final_result(a)
|
372
|
+
if input_b == []
|
373
|
+
if a.buffer.size == 1
|
374
|
+
a.shape = input_b
|
375
|
+
a
|
376
|
+
else
|
377
|
+
wrap_opencl(a.buffer[0], data_type: a.data_type, name: tensor.name)
|
378
|
+
end
|
379
|
+
else
|
380
|
+
wrap_opencl(truncate(input_a, input_b), data_type: a.data_type, name: tensor.name)
|
381
|
+
end
|
382
|
+
end
|
383
|
+
end
|
384
|
+
when :zeros, :ones, :zeros_like, :ones_like
|
385
|
+
shape = if %i[zeros_like ones_like].include?(tensor.operation)
|
386
|
+
_run(a, child_context).shape
|
387
|
+
else
|
388
|
+
read_final_result(complete_eval(a, child_context)) || tensor.shape.shape
|
389
|
+
end
|
390
|
+
|
391
|
+
func = if %i[zeros zeros_like].include?(tensor.operation)
|
392
|
+
-> { tensor.data_type == :int32 ? 0 : 0.0 }
|
393
|
+
else
|
394
|
+
-> { tensor.data_type == :int32 ? 1 : 1.0 }
|
395
|
+
end
|
396
|
+
|
397
|
+
size = shape.empty? ? 1 : shape.reduce(:*)
|
398
|
+
|
399
|
+
buffer = if TensorStream::Ops::FLOATING_POINT_TYPES.include?(tensor.data_type)
|
400
|
+
NArray.sfloat(size)
|
401
|
+
elsif TensorStream::Ops::INTEGER_TYPES.include?(tensor.data_type)
|
402
|
+
NArray.int(size)
|
403
|
+
else
|
404
|
+
raise "unsupported type #{tensor.data_type}"
|
405
|
+
end
|
406
|
+
|
407
|
+
data = if !shape.empty?
|
408
|
+
Array.new(size) do |index|
|
409
|
+
func.call
|
410
|
+
end
|
411
|
+
else
|
412
|
+
func.call
|
413
|
+
end
|
414
|
+
|
415
|
+
convert_to_opencl(data, shape, data_type: tensor.data_type, name: tensor.name)
|
416
|
+
when :broadcast_transform
|
417
|
+
a = _run(a, child_context)
|
418
|
+
b = _run(b, child_context)
|
419
|
+
|
420
|
+
if a.shape == b.shape
|
421
|
+
[a, b]
|
422
|
+
else
|
423
|
+
input_a = read_final_result(complete_eval(a, child_context))
|
424
|
+
input_b = read_final_result(complete_eval(b, child_context))
|
425
|
+
b_a, b_b = broadcast(input_a, input_b)
|
426
|
+
[ wrap_opencl(b_a, data_type: a.data_type, name: "#{tensor.name}_a"),
|
427
|
+
wrap_opencl(b_b, data_type: a.data_type, name: "#{tensor.name}_b")]
|
428
|
+
end
|
429
|
+
when :print
|
430
|
+
a = _run(a, child_context)
|
431
|
+
b = _run(b, child_context)
|
432
|
+
input_b = complete_eval(b, child_context)
|
433
|
+
input_b = read_final_result(input_b)
|
434
|
+
puts "#{tensor.options.fetch(:message, '')} #{input_b}"
|
435
|
+
a
|
436
|
+
when :rank
|
437
|
+
a = _run(a, child_context)
|
438
|
+
wrap_opencl(a.shape.size, data_type: tensor.data_type, name: tensor.name)
|
439
|
+
when :stop_gradient
|
440
|
+
_run(a, child_context)
|
441
|
+
when :slice
|
442
|
+
input_a = complete_eval(a, child_context)
|
443
|
+
input_b = read_final_result(complete_eval(b, child_context))
|
444
|
+
size = tensor.options[:size]
|
445
|
+
|
446
|
+
slice_param = input_b.zip(size).collect { |x, y| x..x + y - 1 }.reverse
|
447
|
+
|
448
|
+
new_buf = input_a.buffer.reshape(*input_a.shape.reverse)
|
449
|
+
sliced = new_buf.slice[*slice_param]
|
450
|
+
convert_to_opencl(sliced.flatten, sliced.shape.reverse, data_type: a.data_type, name: tensor.name)
|
451
|
+
when :transpose
|
452
|
+
input_a = complete_eval(a, child_context)
|
453
|
+
t_param = Array.new(input_a.shape.size) { |index| index }.reverse
|
454
|
+
transposed = input_a.buffer.reshape(*input_a.shape.reverse).transpose(*t_param)
|
455
|
+
convert_to_opencl(transposed.flatten, transposed.shape.reverse, data_type: a.data_type, name: tensor.name)
|
456
|
+
when :index
|
457
|
+
a = complete_eval(a, child_context)
|
458
|
+
input_a = read_final_result(a)
|
459
|
+
index = read_final_result(complete_eval(b, child_context))
|
460
|
+
|
461
|
+
if a.is_a?(Array)
|
462
|
+
a[index]
|
463
|
+
else
|
464
|
+
new_shape = a.shape.dup
|
465
|
+
new_shape.shift
|
466
|
+
convert_to_opencl(input_a[index], new_shape, data_type: a.data_type, name: tensor.name)
|
467
|
+
end
|
468
|
+
when :broadcast_gradient_args
|
469
|
+
a = complete_eval(a, child_context)
|
470
|
+
b = complete_eval(b, child_context)
|
471
|
+
|
472
|
+
wrap_opencl(get_broadcast_gradient_args(a.buffer.to_a, b.buffer.to_a), data_type: a.data_type, name: tensor.name)
|
473
|
+
when :shape
|
474
|
+
a = _run(a, child_context)
|
475
|
+
|
476
|
+
wrap_opencl(a.shape, name: tensor.name, data_type: tensor.options[:out_type] || :float32)
|
477
|
+
when :reshape
|
478
|
+
arr = complete_eval(a, child_context)
|
479
|
+
new_shape = read_final_result(complete_eval(b, child_context))
|
480
|
+
|
481
|
+
if new_shape.size.zero? && arr.buffer.size == 1
|
482
|
+
arr.shape = new_shape
|
483
|
+
arr
|
484
|
+
else
|
485
|
+
new_shape = TensorShape.fix_inferred_elements(new_shape, arr.buffer.size)
|
486
|
+
arr.shape = new_shape
|
487
|
+
arr
|
488
|
+
end
|
489
|
+
when :random_uniform
|
490
|
+
maxval = tensor.options.fetch(:maxval, 1)
|
491
|
+
minval = tensor.options.fetch(:minval, 0)
|
492
|
+
seed = tensor.options[:seed]
|
493
|
+
|
494
|
+
random = _get_randomizer(tensor, seed)
|
495
|
+
generator = -> { random.rand * (maxval - minval) + minval }
|
496
|
+
shape = tensor.options[:shape] || tensor.shape.shape
|
497
|
+
|
498
|
+
convert_to_opencl(generate_vector(shape, generator: generator), shape, data_type: tensor.data_type, name: tensor.name)
|
499
|
+
when :random_normal
|
500
|
+
random = _get_randomizer(tensor, seed)
|
501
|
+
r = RandomGaussian.new(tensor.options.fetch(:mean), tensor.options.fetch(:stddev), -> { random.rand })
|
502
|
+
random = _get_randomizer(tensor, seed)
|
503
|
+
generator = -> { r.rand }
|
504
|
+
shape = tensor.options[:shape] || tensor.shape.shape
|
505
|
+
|
506
|
+
convert_to_opencl(generate_vector(shape, generator: generator), shape, data_type: tensor.data_type, name: tensor.name)
|
507
|
+
when :glorot_uniform
|
508
|
+
random = _get_randomizer(tensor, seed)
|
509
|
+
|
510
|
+
shape = tensor.options[:shape] || tensor.shape.shape
|
511
|
+
fan_in, fan_out = if shape.size.zero?
|
512
|
+
[1, 1]
|
513
|
+
elsif shape.size == 1
|
514
|
+
[1, shape[0]]
|
515
|
+
else
|
516
|
+
[shape[0], shape.last]
|
517
|
+
end
|
518
|
+
|
519
|
+
limit = Math.sqrt(6.0 / (fan_in + fan_out))
|
520
|
+
|
521
|
+
minval = -limit
|
522
|
+
maxval = limit
|
523
|
+
|
524
|
+
generator = -> { random.rand * (maxval - minval) + minval }
|
525
|
+
convert_to_opencl(generate_vector(shape, generator: generator), shape, data_type: tensor.data_type, name: tensor.name)
|
526
|
+
when :flow_group
|
527
|
+
tensor.items.collect { |item| _run(item, child_context) }
|
528
|
+
when :sum
|
529
|
+
reduction(child_context, tensor, a, b, :sum)
|
530
|
+
when :mean
|
531
|
+
reduction(child_context, tensor, a, b, :mean)
|
532
|
+
when :prod
|
533
|
+
input_a = complete_eval(a, child_context)
|
534
|
+
if input_a.buffer.empty?
|
535
|
+
convert_to_opencl([1.0], [], data_type: a.data_type, name: tensor.name)
|
536
|
+
else
|
537
|
+
reduction(child_context, tensor, a, b, :prod)
|
538
|
+
end
|
539
|
+
when :argmin
|
540
|
+
a = complete_eval(a, child_context)
|
541
|
+
axis = tensor.options[:axis] || 0
|
542
|
+
arr = a.buffer.reshape(*a.shape.reverse).to_a
|
543
|
+
op = get_op_with_axis(arr, axis, 0, a.data_type, ->(a, b) { a < b })
|
544
|
+
convert_to_opencl(op, shape_eval(op), data_type: tensor.data_type, name: tensor.name)
|
545
|
+
when :argmax
|
546
|
+
a = complete_eval(a, child_context)
|
547
|
+
axis = tensor.options[:axis] || 0
|
548
|
+
arr = a.buffer.reshape(*a.shape.reverse).to_a
|
549
|
+
op = get_op_with_axis(arr, axis, 0, a.data_type, ->(a, b) { a > b })
|
550
|
+
convert_to_opencl(op, shape_eval(op), data_type: tensor.data_type, name: tensor.name)
|
551
|
+
else
|
552
|
+
raise "unknown op #{tensor.operation}"
|
553
|
+
end.tap do |result|
|
554
|
+
if tensor.breakpoint
|
555
|
+
a = read_final_result(complete_eval(a, child_context))
|
556
|
+
b = read_final_result(complete_eval(b, child_context))
|
557
|
+
result = read_final_result(complete_eval(result, child_context))
|
558
|
+
|
559
|
+
tensor.breakpoint.call(tensor, a, b, result)
|
560
|
+
end
|
561
|
+
if @log_intermediates
|
562
|
+
@context[:compute_history] << {
|
563
|
+
name: tensor.name,
|
564
|
+
type: tensor.data_type,
|
565
|
+
shape: shape_eval(result),
|
566
|
+
source: tensor.source,
|
567
|
+
description: tensor.to_math(true, 1),
|
568
|
+
value: result
|
569
|
+
}
|
570
|
+
end
|
571
|
+
@context[tensor.name] = result
|
572
|
+
end
|
573
|
+
rescue EvaluatorExcecutionException => e
|
574
|
+
raise e
|
575
|
+
rescue StandardError => e
|
576
|
+
puts e.message
|
577
|
+
puts e.backtrace.join("\n")
|
578
|
+
|
579
|
+
# shape_a = a.shape.shape if a
|
580
|
+
# shape_b = b.shape.shape if b
|
581
|
+
# dtype_a = a.data_type if a
|
582
|
+
# dtype_b = b.data_type if b
|
583
|
+
# a = complete_eval(a, child_context)
|
584
|
+
# b = complete_eval(b, child_context)
|
585
|
+
# puts "name: #{tensor.given_name}"
|
586
|
+
# # puts "op: #{tensor.to_math(true, 1)}"
|
587
|
+
# puts "A #{shape_a} #{dtype_a}: #{a}" if a
|
588
|
+
# puts "B #{shape_b} #{dtype_b}: #{b}" if b
|
589
|
+
# dump_intermediates if @log_intermediates
|
590
|
+
# File.write('/home/jedld/workspace/tensor_stream/samples/error.graphml', TensorStream::Graphml.new.get_string(tensor, @session))
|
591
|
+
|
592
|
+
# File.write('/Users/josephemmanueldayo/workspace/gradients.graphml', TensorStream::Graphml.new.get_string(tensor, @session))
|
593
|
+
raise EvaluatorExcecutionException.new(e, tensor), "error #{e.message} while evaluating #{tensor.name} : #{tensor.to_math(true,1)} defined at #{tensor.source}"
|
594
|
+
end
|
595
|
+
|
596
|
+
def eval_tensor(tensor, child_context)
|
597
|
+
return tensor unless tensor.is_a?(Tensor)
|
598
|
+
|
599
|
+
cache_key = "#{tensor.graph.object_id}_opencl_#{tensor.name}"
|
600
|
+
return @context[cache_key] if @context.key?(cache_key)
|
601
|
+
return @context[:_cache][cache_key] if tensor.is_const && @context[:_cache][cache_key]
|
602
|
+
@context[cache_key] = if tensor.value.is_a?(Tensor)
|
603
|
+
_run(tensor.value, child_context)
|
604
|
+
else
|
605
|
+
wrap_opencl(tensor, name: tensor.name)
|
606
|
+
end
|
607
|
+
@context[:_cache][cache_key] = @context[cache_key] if tensor.is_const
|
608
|
+
end
|
609
|
+
|
610
|
+
private
|
611
|
+
|
612
|
+
def assign_var(tensor, b, child_context)
|
613
|
+
assign = tensor.items[0] || tensor
|
614
|
+
buffer = complete_eval(b, child_context)
|
615
|
+
if assign.buffer
|
616
|
+
assign.buffer.op = _opencl_queue.enqueue_write_buffer(assign.buffer.cl_buffer, buffer.buffer)
|
617
|
+
else
|
618
|
+
assign.buffer = convert_to_opencl(read_final_result(buffer), buffer.shape, data_type: tensor.data_type, name: tensor.name)
|
619
|
+
end
|
620
|
+
assign.buffer.dirty = true
|
621
|
+
assign.buffer
|
622
|
+
end
|
623
|
+
|
624
|
+
def execute_2_operand_func(op_name, tensor, input_a, input_b, child_context, prog_name = nil)
|
625
|
+
a = _run(input_a, child_context)
|
626
|
+
b = _run(input_b, child_context)
|
627
|
+
a, b = type_cast(a, b)
|
628
|
+
dtype = TensorStream::Ops::FLOATING_POINT_TYPES.include?(tensor.data_type) ? 'fp' : 'int'
|
629
|
+
result_shape = TensorShape.infer_shape(a.shape, b.shape)
|
630
|
+
|
631
|
+
output_buffer = _create_result_buffer(tensor.data_type, result_shape, tensor.name)
|
632
|
+
a, b, prog, switch_operands = select_program(a, b, op_name)
|
633
|
+
m, n = result_shape
|
634
|
+
work_group = [m || 1, n || 1]
|
635
|
+
cl_m = OpenCL::Int1.new(m || 1)
|
636
|
+
cl_n = OpenCL::Int1.new(n || 1)
|
637
|
+
cl_switch = OpenCL::Int1.new(switch_operands) # no need to switch for addition
|
638
|
+
|
639
|
+
event_wait_list = [a.op, b.op].compact # add dependency wait list
|
640
|
+
|
641
|
+
event = if prog == "#{op_name}_b"
|
642
|
+
cl_m_b, cl_n_b = if b.shape.size == 2
|
643
|
+
[ OpenCL::Int1.new(b.shape[0]), OpenCL::Int1.new(b.shape[1]) ]
|
644
|
+
elsif b.shape.size == 1
|
645
|
+
[ OpenCL::Int1.new(1), OpenCL::Int1.new(b.shape[0]) ]
|
646
|
+
else
|
647
|
+
raise "rank > 2 not supported!"
|
648
|
+
end
|
649
|
+
_cl_program("#{prog_name || op_name}").send(:"#{prog}_#{dtype}", _opencl_queue, work_group, cl_m, cl_n, cl_m_b, cl_n_b, cl_switch, a.cl_buffer, b.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
|
650
|
+
else
|
651
|
+
_cl_program("#{prog_name || op_name}").send(:"#{prog}_#{dtype}", _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)
|
652
|
+
end
|
653
|
+
|
654
|
+
output_buffer.op = event
|
655
|
+
output_buffer
|
656
|
+
end
|
657
|
+
|
658
|
+
def execute_cond_func(op_name, tensor, pred, input_a, input_b, child_context)
|
659
|
+
p = _run(pred, child_context)
|
660
|
+
a = _run(input_a, child_context)
|
661
|
+
b = _run(input_b, child_context)
|
662
|
+
|
663
|
+
a, b = type_cast(a, b)
|
664
|
+
dtype = TensorStream::Ops::FLOATING_POINT_TYPES.include?(tensor.data_type) ? 'fp' : 'int'
|
665
|
+
|
666
|
+
output_buffer = _create_result_buffer(tensor.data_type, p.shape, tensor.name)
|
667
|
+
|
668
|
+
m, n = p.shape
|
669
|
+
work_group = [m || 1, n || 1]
|
670
|
+
cl_m = OpenCL::Int1.new(m || 1)
|
671
|
+
cl_n = OpenCL::Int1.new(n || 1)
|
672
|
+
|
673
|
+
event_wait_list = [a.op, b.op, p.op].compact # add dependency wait list
|
674
|
+
output_buffer.op = _cl_program("#{op_name}").send(:"#{op_name}_#{dtype}", _opencl_queue, work_group, cl_m, cl_n, p.cl_buffer, a.cl_buffer, b.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
|
675
|
+
output_buffer
|
676
|
+
end
|
677
|
+
|
678
|
+
def execute_func(op_name, tensor, a, child_context)
|
679
|
+
a = _run(a, child_context)
|
680
|
+
event_wait_list = [a.op].compact
|
681
|
+
dtype = TensorStream::Ops::FLOATING_POINT_TYPES.include?(tensor.data_type) ? 'fp' : 'int'
|
682
|
+
output_buffer = _create_result_buffer(tensor.data_type, a.shape, tensor.name)
|
683
|
+
|
684
|
+
m, n = a.shape
|
685
|
+
work_group = [m || 1, n || 1]
|
686
|
+
cl_m = OpenCL::Int1.new(m || 1)
|
687
|
+
cl_n = OpenCL::Int1.new(n || 1)
|
688
|
+
|
689
|
+
event = _cl_program("#{op_name}").send(:"#{op_name}_#{dtype}", _opencl_queue, work_group, cl_m, cl_n, a.cl_buffer, output_buffer.cl_buffer, event_wait_list: event_wait_list)
|
690
|
+
output_buffer.op = event
|
691
|
+
output_buffer
|
692
|
+
end
|
693
|
+
|
694
|
+
def type_cast(a, b)
|
695
|
+
return [a, b] if a.data_type == b.data_type
|
696
|
+
m, n = b.shape
|
697
|
+
work_group = [m || 1, n || 1]
|
698
|
+
buffer = buffer_for(b.shape, b.data_type)
|
699
|
+
if (TensorStream::Ops::FLOATING_POINT_TYPES.include?(a.data_type.to_sym))
|
700
|
+
if TensorStream::Ops::INTEGER_TYPES.include?(b.data_type.to_sym)
|
701
|
+
cl_m = OpenCL::Int1.new(m || 1)
|
702
|
+
cl_n = OpenCL::Int1.new(n || 1)
|
703
|
+
|
704
|
+
_cl_program("cast").cast_int_fp(_opencl_queue, work_group, cl_m, cl_n, b.cl_buffer, buffer.cl_buffer)
|
705
|
+
return [a, buffer]
|
706
|
+
end
|
707
|
+
elsif TensorStream::Ops::INTEGER_TYPES.include?(a.data_type.to_sym)
|
708
|
+
if TensorStream::Ops::FLOATING_POINT_TYPES.include?(b.data_type.to_sym)
|
709
|
+
cl_m = OpenCL::Int1.new(m || 1)
|
710
|
+
cl_n = OpenCL::Int1.new(n || 1)
|
711
|
+
_cl_program("cast").cast_fp_int(_opencl_queue, work_group, cl_m, cl_n, b.cl_buffer, buffer.cl_buffer)
|
712
|
+
return [a, buffer]
|
713
|
+
end
|
714
|
+
end
|
715
|
+
|
716
|
+
[a, b]
|
717
|
+
end
|
718
|
+
|
719
|
+
def buffer_for(shape, data_type)
|
720
|
+
size = shape.empty? ? 1 : shape.reduce(:*)
|
721
|
+
|
722
|
+
buffer = allocate_narray_for_type(data_type, size)
|
723
|
+
|
724
|
+
cl_buffer = _opencl_context.create_buffer(buffer.size * buffer.element_size)
|
725
|
+
OpenCLBuffer.new(data_type: data_type, shape: shape, buffer: buffer, cl_buffer: cl_buffer)
|
726
|
+
end
|
727
|
+
|
728
|
+
def wrap_opencl(tensor, data_type: nil, name: nil)
|
729
|
+
value, shape = if tensor.is_a?(Tensor)
|
730
|
+
[tensor.value, tensor.shape.shape]
|
731
|
+
else
|
732
|
+
[tensor , shape_eval(tensor)]
|
733
|
+
end
|
734
|
+
|
735
|
+
convert_to_opencl(value, shape, data_type: data_type || tensor.data_type, name: name)
|
736
|
+
end
|
737
|
+
|
738
|
+
def convert_to_opencl(value, shape, data_type: nil, name: nil)
|
739
|
+
if !value.is_a?(Array) && !value.is_a?(NArray)
|
740
|
+
value = [value]
|
741
|
+
end
|
742
|
+
|
743
|
+
cache_key = "_cl_object_#{name}_#{shape.join('_')}"
|
744
|
+
cl_object = if name && @context[:_cache][cache_key]
|
745
|
+
@context[:_cache][cache_key]
|
746
|
+
else
|
747
|
+
narray_size = shape.reduce(:*) || 1
|
748
|
+
|
749
|
+
buffer = if value.is_a?(NArray)
|
750
|
+
value
|
751
|
+
else
|
752
|
+
allocate_narray_for_type(data_type, narray_size)
|
753
|
+
end
|
754
|
+
|
755
|
+
cl_buffer_size = shape.empty? ? 1 : shape.reduce(:*)
|
756
|
+
|
757
|
+
cl_buffer = if !value.flatten.empty?
|
758
|
+
cl_buffer_size = 1 if cl_buffer_size.zero?
|
759
|
+
_opencl_context.create_buffer(cl_buffer_size * buffer.element_size)
|
760
|
+
else
|
761
|
+
nil
|
762
|
+
end
|
763
|
+
|
764
|
+
@context[:_cache][cache_key] = OpenCLBuffer.new(name: name, data_type: data_type, shape: shape, buffer: buffer, cl_buffer: cl_buffer)
|
765
|
+
end
|
766
|
+
|
767
|
+
if value.is_a?(Array)
|
768
|
+
value.flatten.each_with_index do |element, index|
|
769
|
+
if element.is_a?(Tensor)
|
770
|
+
cl_object.buffer[index] = read_final_result(complete_eval(element, {}))
|
771
|
+
else
|
772
|
+
cl_object.buffer[index] = Tensor.cast_dtype(element, data_type)
|
773
|
+
end
|
774
|
+
end
|
775
|
+
elsif value.is_a?(NArray)
|
776
|
+
cl_object.buffer = value
|
777
|
+
else
|
778
|
+
cl_object.buffer[0] = Tensor.cast_dtype(value, data_type)
|
779
|
+
end
|
780
|
+
|
781
|
+
write_op = if cl_object.cl_buffer && !value.nil? && (!value.is_a?(Array) || !value.empty?)
|
782
|
+
_opencl_queue.enqueue_write_buffer(cl_object.cl_buffer, cl_object.buffer)
|
783
|
+
end
|
784
|
+
cl_object.op = write_op
|
785
|
+
cl_object
|
786
|
+
end
|
787
|
+
|
788
|
+
def allocate_narray_for_type(data_type, narray_size)
|
789
|
+
if TensorStream::Ops::FLOATING_POINT_TYPES.include?(data_type.to_sym) || TensorStream::Ops::FLOATING_POINT_TYPES.include?(data_type.to_sym)
|
790
|
+
NArray.sfloat(narray_size)
|
791
|
+
elsif TensorStream::Ops::INTEGER_TYPES.include?(data_type.to_sym) || TensorStream::Ops::INTEGER_TYPES.include?(data_type.to_sym)
|
792
|
+
NArray.int(narray_size)
|
793
|
+
elsif data_type.to_sym == :boolean
|
794
|
+
NArray.int(narray_size)
|
795
|
+
else
|
796
|
+
raise "unsupported type #{data_type}"
|
797
|
+
end
|
798
|
+
end
|
799
|
+
|
800
|
+
def _create_result_buffer(data_type, shape, name)
|
801
|
+
@context[:_cache]["_result_#{name}_#{shape.join('_')}"] ||= begin
|
802
|
+
size = shape.empty? ? 1 : shape.reduce(:*)
|
803
|
+
buffer = allocate_narray_for_type(data_type, size)
|
804
|
+
cl_buffer = _opencl_context.create_buffer(buffer.size * buffer.element_size)
|
805
|
+
OpenCLBuffer.new(data_type: data_type, shape: shape, buffer: buffer, cl_buffer: cl_buffer)
|
806
|
+
end
|
807
|
+
end
|
808
|
+
|
809
|
+
def get_op_with_axis(a, target_axis, current_axis, output_type, op = ->(t, u) { t > u })
|
810
|
+
if target_axis == current_axis
|
811
|
+
if a[0].is_a?(Array)
|
812
|
+
(0...a[0].size).each.collect do |column_index|
|
813
|
+
max = nil
|
814
|
+
max_index = 0
|
815
|
+
a.each_with_index do |row, row_index|
|
816
|
+
if max.nil? || op.call(row[column_index], max)
|
817
|
+
max = row[column_index]
|
818
|
+
max_index = row_index
|
819
|
+
end
|
820
|
+
end
|
821
|
+
|
822
|
+
Tensor.cast_dtype(max_index, output_type)
|
823
|
+
end
|
824
|
+
else
|
825
|
+
max = nil
|
826
|
+
max_index = 0
|
827
|
+
a.each_with_index do |x, index|
|
828
|
+
if max.nil? || op.call(x, max)
|
829
|
+
max = x
|
830
|
+
max_index = index
|
831
|
+
end
|
832
|
+
end
|
833
|
+
Tensor.cast_dtype(max_index, output_type)
|
834
|
+
end
|
835
|
+
else
|
836
|
+
a.collect do |row|
|
837
|
+
get_op_with_axis(row, target_axis, current_axis + 1, output_type, op)
|
838
|
+
end
|
839
|
+
end
|
840
|
+
end
|
841
|
+
|
842
|
+
def reduction(child_context, tensor, a, b, func)
|
843
|
+
input = complete_eval(a, child_context)
|
844
|
+
axis = read_final_result(complete_eval(b, child_context))
|
845
|
+
if axis.nil?
|
846
|
+
convert_to_opencl(input.buffer.send(func), [], data_type: tensor.data_type, name: tensor.name)
|
847
|
+
else
|
848
|
+
return input if input.shape.empty?
|
849
|
+
value = input.buffer.reshape(*input.shape.reverse)
|
850
|
+
rank = input.shape.size - 1
|
851
|
+
|
852
|
+
if axis.is_a?(Array)
|
853
|
+
axis.map{ |x| rank - x.abs }.sort.reverse.each do |x|
|
854
|
+
value = value.send(func, x)
|
855
|
+
end
|
856
|
+
else
|
857
|
+
value = value.send(func, rank - axis.abs)
|
858
|
+
end
|
859
|
+
|
860
|
+
new_shape = if value.is_a?(NArray)
|
861
|
+
value.shape.reverse
|
862
|
+
else
|
863
|
+
value = [value]
|
864
|
+
[]
|
865
|
+
end
|
866
|
+
|
867
|
+
if tensor.options[:keepdims]
|
868
|
+
new_shape = reduced_shape(input.shape.dup, axis)
|
869
|
+
end
|
870
|
+
|
871
|
+
convert_to_opencl(value.flatten, new_shape, data_type: tensor.data_type, name: tensor.name)
|
872
|
+
end
|
873
|
+
end
|
874
|
+
|
875
|
+
def arr_pad(arr, paddings, data_type = :float32, rank = 0)
|
876
|
+
raise "padding #{paddings[rank]} needs to have to elements [before, after]" if paddings[rank].size != 2
|
877
|
+
|
878
|
+
before = paddings[rank][0]
|
879
|
+
after = paddings[rank][1]
|
880
|
+
pad_value = fp_type?(data_type) ? 0.0 : 0
|
881
|
+
if arr[0].is_a?(Array)
|
882
|
+
next_dim_elem = arr.collect { |a| arr_pad(a, paddings, data_type, rank + 1) }
|
883
|
+
padding = deep_dup_array(next_dim_elem[0], pad_value)
|
884
|
+
Array.new(before) { padding } + next_dim_elem + Array.new(after) { padding }
|
885
|
+
else
|
886
|
+
Array.new(before) { pad_value } + arr + Array.new(after) { pad_value }
|
887
|
+
end
|
888
|
+
end
|
889
|
+
|
890
|
+
def deep_dup_array(arr, value = nil)
|
891
|
+
if arr.is_a?(Array)
|
892
|
+
arr.dup.collect do |a|
|
893
|
+
deep_dup_array(a, value)
|
894
|
+
end
|
895
|
+
else
|
896
|
+
value.nil? ? arr : value
|
897
|
+
end
|
898
|
+
end
|
899
|
+
|
900
|
+
def matmul_const_transform(mat, mat_b, tensor)
|
901
|
+
if !mat.is_a?(Array)
|
902
|
+
compat_shape = shape_eval(mat_b).reverse
|
903
|
+
func = -> { tensor.data_type == :int32 ? mat.to_i : mat.to_f }
|
904
|
+
|
905
|
+
generate_vector(compat_shape, generator: func)
|
906
|
+
else
|
907
|
+
mat
|
908
|
+
end
|
909
|
+
end
|
910
|
+
|
911
|
+
# determine possible reduction axis to be used
|
912
|
+
def _broadcast_gradient_op(vector_shape1, vector_shape2, level)
|
913
|
+
va_rank = _rank_from_shape(vector_shape1)
|
914
|
+
vb_rank = _rank_from_shape(vector_shape2)
|
915
|
+
return [] if vector_shape1 == vector_shape2 # same shape so no reductions
|
916
|
+
|
917
|
+
shape2_r = vector_shape2.reverse
|
918
|
+
|
919
|
+
vector_shape1.reverse.each_with_index.collect do |s, index|
|
920
|
+
next va_rank - index - 1 if index >= shape2_r.size
|
921
|
+
next nil if shape2_r[index] == s
|
922
|
+
next nil if shape2_r[index] > s
|
923
|
+
va_rank - index - 1
|
924
|
+
end.compact
|
925
|
+
end
|
926
|
+
|
927
|
+
# selects variants of cl programs depending on input
|
928
|
+
def select_program(input_a, input_b, op)
|
929
|
+
return [input_a, input_b, "#{op}", 0] if input_a.shape == input_b.shape
|
930
|
+
|
931
|
+
return [input_b, input_a, "#{op}_c", 1] if input_a.shape.empty? || input_a.shape.reduce(:*) == 1 # A is scalar?
|
932
|
+
return [input_a, input_b, "#{op}_c", 0] if input_b.shape.empty? || input_a.shape.reduce(:*) == 1 # B is scalar?
|
933
|
+
|
934
|
+
return [input_b, input_a, "#{op}_b", 1] if input_a.shape.size < input_b.shape.size
|
935
|
+
|
936
|
+
if input_a.shape.size == input_b.shape.size
|
937
|
+
input_a.shape.zip(input_b.shape).each do |s1, s2|
|
938
|
+
return [input_b, input_a, "#{op}_b", 1] if s1 < s2
|
939
|
+
end
|
940
|
+
end
|
941
|
+
|
942
|
+
[input_a, input_b, "#{op}_b", 0]
|
943
|
+
end
|
944
|
+
|
945
|
+
def _rank_from_shape(shape)
|
946
|
+
shape.is_a?(Array) ? shape.size : 0
|
947
|
+
end
|
948
|
+
|
949
|
+
def get_broadcast_gradient_args(input_a, input_b)
|
950
|
+
return [] if get_rank(input_b).zero? && get_rank(input_a).zero?
|
951
|
+
return nil if get_rank(input_b).zero?
|
952
|
+
# ruby scalar
|
953
|
+
if get_rank(input_a).zero?
|
954
|
+
_broadcast_gradient_op(input_b, input_a, 0, true)
|
955
|
+
elsif get_rank(input_a) > 0
|
956
|
+
_broadcast_gradient_op(input_a, input_b, 0)
|
957
|
+
end
|
958
|
+
end
|
959
|
+
|
960
|
+
def concat_array(values, axis)
|
961
|
+
combined_array = values.shift
|
962
|
+
axis = get_rank(combined_array) - 1 if axis == -1
|
963
|
+
|
964
|
+
values.each do |v|
|
965
|
+
combined_array = concat(combined_array, v, axis)
|
966
|
+
end
|
967
|
+
combined_array
|
968
|
+
end
|
969
|
+
|
970
|
+
def concat(a, b, axis)
|
971
|
+
if axis.zero?
|
972
|
+
a + b
|
973
|
+
else
|
974
|
+
a.each_with_index.collect do |i, index|
|
975
|
+
concat(i, b[index], axis - 1)
|
976
|
+
end
|
977
|
+
end
|
978
|
+
end
|
979
|
+
|
980
|
+
def resolve_placeholder(placeholder, _execution_context = {})
|
981
|
+
return nil if placeholder.nil?
|
982
|
+
return placeholder if retain.include?(placeholder)
|
983
|
+
|
984
|
+
var = if placeholder.is_a?(Placeholder)
|
985
|
+
@context[placeholder.name.to_sym].tap do |c|
|
986
|
+
raise "missing placeholder #{placeholder.name}" if c.nil?
|
987
|
+
end
|
988
|
+
else
|
989
|
+
placeholder
|
990
|
+
end
|
991
|
+
|
992
|
+
return convert_to_opencl(var, shape_eval(var), data_type: placeholder.data_type, name: placeholder.name) unless var.is_a?(Tensor)
|
993
|
+
Tensor.cast_dtype(var, placeholder.data_type)
|
994
|
+
end
|
995
|
+
|
996
|
+
def reduce_axis(current_axis, axis, val, keep_dims, f = ->(a, b) { a + b })
|
997
|
+
return val unless val.is_a?(Array)
|
998
|
+
|
999
|
+
r = val.collect do |v|
|
1000
|
+
reduce_axis(current_axis + 1, axis, v, keep_dims, f)
|
1001
|
+
end
|
1002
|
+
|
1003
|
+
should_reduce_axis = axis.nil? || (axis.is_a?(Array) && axis.include?(current_axis)) || (current_axis == axis)
|
1004
|
+
|
1005
|
+
if should_reduce_axis
|
1006
|
+
reduced_val = r[0]
|
1007
|
+
if r.size > 1
|
1008
|
+
reduced_val = f.call(r[0..val.size])
|
1009
|
+
elsif r.size == 0
|
1010
|
+
reduced_val = f.call(nil)
|
1011
|
+
end
|
1012
|
+
keep_dims ? [ reduced_val ] : reduced_val
|
1013
|
+
else
|
1014
|
+
r
|
1015
|
+
end
|
1016
|
+
end
|
1017
|
+
|
1018
|
+
# handle 3 tensor math operations
|
1019
|
+
def call_3way_vector_op(v_a, v_b, v_c, child_context, op = ->(a, b, c) { a + b + c })
|
1020
|
+
return op.call(v_a, v_b, v_c) unless v_a.is_a?(Array)
|
1021
|
+
|
1022
|
+
v_a.each_with_index.collect do |v1, index|
|
1023
|
+
v2 = v_b[index]
|
1024
|
+
v3 = v_c[index]
|
1025
|
+
if v1.is_a?(Array)
|
1026
|
+
call_3way_vector_op(v1, v2, v3, child_context, op)
|
1027
|
+
else
|
1028
|
+
op.call(v1, v2, v3)
|
1029
|
+
end
|
1030
|
+
end
|
1031
|
+
end
|
1032
|
+
|
1033
|
+
def all_true?(arr)
|
1034
|
+
if arr.is_a?(Array) || arr.is_a?(NArray)
|
1035
|
+
arr.each do |a|
|
1036
|
+
return false unless all_true?(a)
|
1037
|
+
end
|
1038
|
+
return true
|
1039
|
+
end
|
1040
|
+
|
1041
|
+
arr != 0
|
1042
|
+
end
|
1043
|
+
|
1044
|
+
def generate_vector(shape, dtype: :float32, generator:)
|
1045
|
+
if shape.is_a?(Integer)
|
1046
|
+
Array.new(shape) do
|
1047
|
+
generator.call
|
1048
|
+
end
|
1049
|
+
elsif shape.size > 1
|
1050
|
+
Array.new(shape[0]) do
|
1051
|
+
generate_vector(shape[1..shape.size], generator: generator, dtype: dtype)
|
1052
|
+
end
|
1053
|
+
elsif shape.size == 1
|
1054
|
+
Array.new(shape[0]) do
|
1055
|
+
generator.call
|
1056
|
+
end
|
1057
|
+
elsif shape.size.zero?
|
1058
|
+
generator.call
|
1059
|
+
end
|
1060
|
+
end
|
1061
|
+
|
1062
|
+
def _get_randomizer(tensor, seed)
|
1063
|
+
if tensor.graph.random_seed && seed
|
1064
|
+
Random.new(tensor.graph.random_seed ^ seed)
|
1065
|
+
elsif tensor.graph.random_seed
|
1066
|
+
@session.randomizer[tensor.graph.object_id] ||= Random.new(tensor.graph.random_seed)
|
1067
|
+
@session.randomizer[tensor.graph.object_id]
|
1068
|
+
elsif seed
|
1069
|
+
@session.randomizer[tensor.operation] ||= Random.new(seed)
|
1070
|
+
@session.randomizer[tensor.operation]
|
1071
|
+
else
|
1072
|
+
Random.new
|
1073
|
+
end
|
1074
|
+
end
|
1075
|
+
|
1076
|
+
def dump_intermediates
|
1077
|
+
arr = []
|
1078
|
+
arr << "============== start ==================="
|
1079
|
+
@context[:compute_history].each_with_index do |history, index|
|
1080
|
+
arr << "------------------------------------"
|
1081
|
+
arr << history[:name]
|
1082
|
+
arr << "#{history[:type]} #{history[:shape]}"
|
1083
|
+
arr << history[:source]
|
1084
|
+
arr << history[:description]
|
1085
|
+
arr << ""
|
1086
|
+
arr << history[:value].to_json
|
1087
|
+
arr << "------------------------------------"
|
1088
|
+
end
|
1089
|
+
arr << "============== end ====================="
|
1090
|
+
str = arr.join("\n")
|
1091
|
+
File.write("/tmp/intermediates.txt", str)
|
1092
|
+
end
|
1093
|
+
end
|
1094
|
+
end
|
1095
|
+
end
|