Skip to content

Commit 08a4d5e

Browse files
atveitatveit
committed
initial commit with actual code
1 parent 2e5ba37 commit 08a4d5e

File tree

58 files changed

+6114
-1
lines changed

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

58 files changed

+6114
-1
lines changed

‎CommonForAllPlatforms/ConvolutionLayer.swift‎

Lines changed: 237 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,237 @@
1+
//
2+
// ConvolutionLayer.swift
3+
// MemkiteMetal
4+
//
5+
// Created by Torb Morland & Amund Tveit on 12/12/15.
6+
// Copyright © 2015 Memkite. All rights reserved.
7+
//
8+
9+
import Foundation
10+
import Metal
11+
12+
func getDataFromBlob(blob: NSDictionary) -> ([Float], [Float]) {
13+
print(" ==> getDataFromBlob")
14+
15+
let shape = blob["shape"] as! NSDictionary
16+
let data = blob["data"] as! [Float]
17+
var FloatData = createFloatNumbersArray(data.count)
18+
for i in 0 ..< data.count {
19+
FloatData[i] = data[i]
20+
}
21+
return (shape["dim"] as! [Float], FloatData)
22+
}
23+
24+
25+
26+
func createConvolutionLayerCached(layer: NSDictionary,
27+
inputBuffer: MTLBuffer,
28+
inputShape: [Float],
29+
metalCommandQueue: MTLCommandQueue, metalDefaultLibrary:MTLLibrary, metalDevice:MTLDevice,
30+
inout layer_data_caches: [Dictionary<String,MTLBuffer>],
31+
inout blob_cache: [Dictionary<String,([Float],[Float])>],
32+
layer_number: Int,
33+
layer_string: String) -> (MTLBuffer, MTLCommandBuffer, [Float]) {
34+
35+
let start = NSDate()
36+
37+
print("CREATECONVLAYERCACHED")
38+
39+
// let metalCommandBuffer = metalCommandQueue.commandBuffer()
40+
let metalCommandBuffer = metalCommandQueue.commandBufferWithUnretainedReferences()
41+
42+
var convolution_params_dict:NSDictionary = NSDictionary()
43+
var pad:Float = 0.0
44+
var kernel_size:Float = 1.0
45+
var stride:Float = 1.0
46+
var blobs:[NSDictionary] = []
47+
var weights:[Float] = []
48+
var weight_shape:[Float] = []
49+
var bias_data:[Float] = []
50+
var h:Float = 0.0
51+
var w:Float = 0.0
52+
var result_shape:[Float] = []
53+
var outputCount:Int = 0
54+
55+
var input_dimensions:MetalTensorDimensions = MetalTensorDimensions(n: 0, channels: 0, width: 0, height:0)
56+
var weight_dimensions:MetalTensorDimensions = MetalTensorDimensions(n: 0, channels: 0, width: 0, height:0)
57+
var result_dimensions:MetalTensorDimensions = MetalTensorDimensions(n: 0, channels: 0, width: 0, height:0)
58+
var tensor_dimensions:[MetalTensorDimensions] = []
59+
var col_dimensions:MetalTensorDimensions = MetalTensorDimensions(n: 0, channels: 0, width: 0, height:0)
60+
var col_output:[Float] = []
61+
var convolution_params:MetalConvolutionParameters = MetalConvolutionParameters(pad:0, kernel_size: 0, stride: 0)
62+
63+
64+
print("NOTCACHINGMODE")
65+
convolution_params_dict = layer["convolution_param"] as! NSDictionary
66+
pad = 0.0
67+
kernel_size = 1.0
68+
stride = 1.0
69+
if let val = convolution_params_dict["pad"] as? Float {
70+
pad = val
71+
}
72+
if let val = convolution_params_dict["kernel_size"] as? Float {
73+
kernel_size = val
74+
}
75+
76+
let startblob = NSDate()
77+
78+
79+
if let tmpval = blob_cache[layer_number]["0"] {
80+
print("found blob key = 0 in cache")
81+
(weight_shape, weights) = tmpval
82+
} else {
83+
print("didnt find blob key = 0 in cache")
84+
blobs = layer["blobs"] as! [NSDictionary]
85+
(weight_shape, weights) = getDataFromBlob(blobs[0])
86+
blob_cache[layer_number]["0"] = (weight_shape, weights)
87+
}
88+
89+
// this can be optimized
90+
blobs = layer["blobs"] as! [NSDictionary]
91+
(_, bias_data) = getDataFromBlob(blobs[1])
92+
93+
print("### Time to blob: \(NSDate().timeIntervalSinceDate(startblob))")
94+
95+
96+
/*
97+
let startblob = NSDate()
98+
blobs = layer["blobs"] as! [NSDictionary]
99+
(weight_shape, weights) = getDataFromBlob(blobs[0])
100+
(_, bias_data) = getDataFromBlob(blobs[1])
101+
print("### Time to blob: \(NSDate().timeIntervalSinceDate(startblob))")
102+
*/
103+
104+
105+
h = (inputShape[2] + 2 * pad - kernel_size) / stride + 1
106+
w = (inputShape[3] + 2 * pad - kernel_size) / stride + 1
107+
result_shape = [inputShape[0], weight_shape[0], h, w]
108+
outputCount = Int(result_shape.reduce(1, combine: *))
109+
110+
// Create input and output vectors, and corresponding metal buffer
111+
input_dimensions = MetalTensorDimensions(n: inputShape[0], channels: inputShape[1], width: inputShape[2], height: inputShape[3])
112+
weight_dimensions = MetalTensorDimensions(n: weight_shape[0], channels: weight_shape[1], width: weight_shape[2], height: weight_shape[3])
113+
col_dimensions = MetalTensorDimensions(n: inputShape[0], channels: inputShape[1] * weight_shape[2] * weight_shape[3], width: inputShape[2], height: inputShape[3])
114+
result_dimensions = MetalTensorDimensions(n: result_shape[0], channels: result_shape[1], width: result_shape[2], height: result_shape[3])
115+
tensor_dimensions = [input_dimensions, weight_dimensions, col_dimensions, result_dimensions]
116+
117+
118+
col_output = createFloatNumbersArray(Int(col_dimensions.n * col_dimensions.channels * col_dimensions.height * col_dimensions.width))
119+
120+
121+
convolution_params = MetalConvolutionParameters(pad: pad, kernel_size: kernel_size, stride: stride)
122+
print("AFTER NOTCACHINGMODE")
123+
124+
125+
print("BEFORE THE BIG CALL")
126+
127+
let resultBuffer = addConvolutionCommandToCommandBufferCached(metalCommandBuffer, inputBuffer: inputBuffer, im2ColCount: col_output.count, weights: weights, outputCount: outputCount, convolution_params: convolution_params, tensor_dimensions: tensor_dimensions, bias: bias_data, metalDefaultLibrary: metalDefaultLibrary, metalDevice:metalDevice, layer_data_caches: &layer_data_caches, layer_number: layer_number,layer_string: layer_string)
128+
//metalCommandBuffer.commit()
129+
130+
print("AFTER BIG CALL")
131+
132+
print("### Time to setup convolution layer: \(NSDate().timeIntervalSinceDate(start))")
133+
134+
135+
return (resultBuffer, metalCommandBuffer, result_shape)
136+
137+
}
138+
139+
func addConvolutionCommandToCommandBufferCached(commandBuffer: MTLCommandBuffer,
140+
inputBuffer: MTLBuffer,
141+
im2ColCount: Int,
142+
weights: [Float],
143+
outputCount: Int,
144+
convolution_params: MetalConvolutionParameters,
145+
tensor_dimensions: [MetalTensorDimensions],
146+
bias: [Float],
147+
metalDefaultLibrary:MTLLibrary, metalDevice:MTLDevice,
148+
inout layer_data_caches: [Dictionary<String,MTLBuffer>],
149+
layer_number: Int,
150+
layer_string: String) -> MTLBuffer {
151+
152+
let start = NSDate()
153+
154+
print("before output and col_output")
155+
156+
var output:[Float] = []
157+
var col_output:[Float] = []
158+
159+
output = createFloatNumbersArray(outputCount)
160+
col_output = createFloatNumbersArray(im2ColCount)
161+
162+
print("before setupshaderinpipeline")
163+
164+
let (_, im2colComputePipelineState, _) = setupShaderInMetalPipeline("im2col", metalDefaultLibrary: metalDefaultLibrary, metalDevice: metalDevice)
165+
166+
let resultMetalBuffer = createOrReuseFloatMetalBuffer("resultMetalBuffer", data: output, cache: &layer_data_caches, layer_number: layer_number, metalDevice: metalDevice)
167+
168+
print("after resultmetalbuffer")
169+
170+
let weightMetalBuffer = createOrReuseFloatMetalBuffer("weightMetalBuffer", data: weights, cache: &layer_data_caches, layer_number:layer_number, metalDevice: metalDevice)
171+
172+
173+
let convolutionParamsMetalBuffer = createOrReuseConvolutionParametersMetalBuffer("convolutionParamsMetalBuffer", data: convolution_params, cache: &layer_data_caches, layer_number: layer_number, metalDevice: metalDevice)
174+
let tensorDimensionsMetalBuffer = createOrReuseTensorDimensionsVectorMetalBuffer("tensorDimensionsMetalBuffer", data: tensor_dimensions, cache: &layer_data_caches, layer_number: layer_number, metalDevice: metalDevice)
175+
176+
let colOutputMetalBuffer = createOrReuseFloatMetalBuffer("colOutputMetalBuffer", data: col_output, cache: &layer_data_caches, layer_number: layer_number, metalDevice: metalDevice)
177+
let biasMetalBuffer = createOrReuseFloatMetalBuffer("bias", data: bias, cache: &layer_data_caches, layer_number:layer_number, metalDevice: metalDevice)
178+
179+
180+
// Create Metal compute command encoder for im2col
181+
var metalComputeCommandEncoder = commandBuffer.computeCommandEncoder()
182+
metalComputeCommandEncoder.setBuffer(inputBuffer, offset: 0, atIndex: 0)
183+
metalComputeCommandEncoder.setBuffer(tensorDimensionsMetalBuffer, offset: 0, atIndex: 1)
184+
metalComputeCommandEncoder.setBuffer(convolutionParamsMetalBuffer, offset: 0, atIndex: 2)
185+
metalComputeCommandEncoder.setBuffer(colOutputMetalBuffer, offset: 0, atIndex: 3)
186+
//metalComputeCommandEncoder.setComputePipelineState(im2colComputePipelineState)
187+
188+
189+
// Set the shader function that Metal will use
190+
metalComputeCommandEncoder.setComputePipelineState(im2colComputePipelineState)
191+
192+
// Set up thread groups on GPU
193+
var threadsPerGroup = MTLSize(width:im2colComputePipelineState.threadExecutionWidth,height:1,depth:1)
194+
// ensure at least 1 threadgroup
195+
print("before mtlsize 2")
196+
var numThreadgroups = MTLSize(width:(col_output.count-1)/im2colComputePipelineState.threadExecutionWidth + 1, height:1, depth:1)
197+
metalComputeCommandEncoder.dispatchThreadgroups(numThreadgroups, threadsPerThreadgroup: threadsPerGroup)
198+
199+
print("after dispatch")
200+
201+
// Finalize configuration
202+
metalComputeCommandEncoder.endEncoding()
203+
204+
205+
206+
207+
let (_, convolutionComputePipelineState, _) = setupShaderInMetalPipeline("convolution_layer", metalDefaultLibrary: metalDefaultLibrary, metalDevice: metalDevice)
208+
metalComputeCommandEncoder = commandBuffer.computeCommandEncoder()
209+
210+
// Create Metal Compute Command Encoder and add input and output buffers to it
211+
metalComputeCommandEncoder.setBuffer(resultMetalBuffer, offset: 0, atIndex: 0)
212+
metalComputeCommandEncoder.setBuffer(weightMetalBuffer, offset: 0, atIndex: 1)
213+
metalComputeCommandEncoder.setBuffer(tensorDimensionsMetalBuffer, offset: 0, atIndex: 2)
214+
metalComputeCommandEncoder.setBuffer(colOutputMetalBuffer, offset: 0, atIndex: 3)
215+
metalComputeCommandEncoder.setBuffer(biasMetalBuffer, offset: 0, atIndex: 4)
216+
217+
// Set the shader function that Metal will use
218+
metalComputeCommandEncoder.setComputePipelineState(convolutionComputePipelineState)
219+
220+
// Set up thread groups on GPU
221+
threadsPerGroup = MTLSize(width:convolutionComputePipelineState.threadExecutionWidth,height:1,depth:1)
222+
// ensure at least 1 threadgroup
223+
numThreadgroups = MTLSize(width:(outputCount-1)/convolutionComputePipelineState.threadExecutionWidth + 1, height:1, depth:1)
224+
metalComputeCommandEncoder.dispatchThreadgroups(numThreadgroups, threadsPerThreadgroup: threadsPerGroup)
225+
226+
// Finalize configuration
227+
metalComputeCommandEncoder.endEncoding()
228+
229+
print("after endencoding")
230+
231+
print("#### Time to add convolution layer: \(NSDate().timeIntervalSinceDate(start))")
232+
233+
234+
return resultMetalBuffer
235+
236+
}
237+

‎CommonForAllPlatforms/DeepNetwork+JSON.swift‎

Lines changed: 26 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,26 @@
1+
//
2+
// DeepNetwork+JSON.swift
3+
// MemkiteMetal
4+
//
5+
// Created by Amund Tveit & Torb Morland on 12/12/15.
6+
// Copyright © 2015 Memkite. All rights reserved.
7+
//
8+
9+
import Foundation
10+
11+
public extension DeepNetwork {
12+
13+
func loadJSONFile(filename: String) -> NSDictionary? {
14+
print(" ==> loadJSONFile(filename=\(filename)")
15+
16+
do {
17+
let bundle = NSBundle.mainBundle()
18+
let path = bundle.pathForResource(filename, ofType: "json")!
19+
let jsonData = NSData(contentsOfFile: path)
20+
print(" <== loadJSONFile")
21+
return try NSJSONSerialization.JSONObjectWithData(jsonData!, options: .AllowFragments) as? NSDictionary
22+
} catch _ {
23+
return nil
24+
}
25+
}
26+
}

‎CommonForAllPlatforms/DeepNetwork+SetupNetworkFromDict.swift‎

Lines changed: 86 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,86 @@
1+
//
2+
// DeepNetwork+SetupNetworkFromDict.swift
3+
// MemkiteMetal
4+
//
5+
// Created by Amund Tveit & Torb Morland on 12/12/15.
6+
// Copyright © 2015 Memkite. All rights reserved.
7+
//
8+
9+
import Foundation
10+
import Metal
11+
12+
public extension DeepNetwork {
13+
14+
func setupNetworkFromDict(deepNetworkAsDict: NSDictionary, inputimage: MTLBuffer, inputshape: [Float]) {
15+
16+
let start = NSDate()
17+
18+
gpuCommandLayers = []
19+
20+
print(" ==> setupNetworkFromDict()")
21+
// Add input image
22+
var layer_number = 0
23+
layer_data_caches.append(Dictionary<String, MTLBuffer>()) // for input
24+
pool_type_caches.append(Dictionary<String,String>())
25+
blob_cache.append(Dictionary<String,([Float],[Float])>())
26+
namedDataLayers.append(("input", inputimage))
27+
++layer_number
28+
29+
30+
// Add remaining network
31+
var previousBuffer:MTLBuffer = inputimage
32+
var previousShape:[Float] = inputshape
33+
34+
self.deepNetworkAsDict = deepNetworkAsDict
35+
36+
// create new command buffer for next layer
37+
var currentCommandBuffer: MTLCommandBuffer = metalCommandQueue.commandBufferWithUnretainedReferences()
38+
39+
var t = NSDate()
40+
for layer in deepNetworkAsDict["layer"] as! [NSDictionary] {
41+
if let type = layer["type"] as? String {
42+
let layer_string = layer["name"] as! String
43+
44+
layer_data_caches.append(Dictionary<String, MTLBuffer>())
45+
pool_type_caches.append(Dictionary<String,String>())
46+
blob_cache.append(Dictionary<String,([Float],[Float])>())
47+
48+
49+
if type == "ReLU" {
50+
self.gpuCommandLayers.append(currentCommandBuffer)
51+
//(previousBuffer, currentCommandBuffer) = createRectifierLayer(previousBuffer)
52+
(previousBuffer, currentCommandBuffer) = createRectifierLayer(previousBuffer, metalCommandQueue:metalCommandQueue, metalDefaultLibrary:metalDefaultLibrary, metalDevice:metalDevice)
53+
self.namedDataLayers.append((layer["name"]! as! String, previousBuffer))
54+
} else if type == "Pooling" {
55+
self.gpuCommandLayers.append(currentCommandBuffer)
56+
// (previousBuffer, currentCommandBuffer, previousShape) = createPoolingLayer(layer, inputBuffer: previousBuffer, inputShape: previousShape)
57+
(previousBuffer, currentCommandBuffer, previousShape) = createPoolingLayerCached(layer, inputBuffer: previousBuffer, inputShape: previousShape, metalCommandQueue: metalCommandQueue, metalDefaultLibrary: metalDefaultLibrary, metalDevice: metalDevice, pool_type_caches: &pool_type_caches, layer_data_caches: &layer_data_caches, layer_number: layer_number, layer_string: layer_string)
58+
self.namedDataLayers.append((layer["name"]! as! String, previousBuffer))
59+
} else if type == "Convolution" {
60+
self.gpuCommandLayers.append(currentCommandBuffer)
61+
// (previousBuffer, currentCommandBuffer, previousShape) = createConvolutionLayer(layer, inputBuffer: previousBuffer, inputShape: previousShape)
62+
(previousBuffer, currentCommandBuffer, previousShape) = createConvolutionLayerCached(layer, inputBuffer: previousBuffer, inputShape: previousShape, metalCommandQueue: metalCommandQueue, metalDefaultLibrary:metalDefaultLibrary, metalDevice:metalDevice, layer_data_caches: &layer_data_caches, blob_cache: &blob_cache, layer_number: layer_number, layer_string: layer_string)
63+
64+
65+
self.namedDataLayers.append((layer["name"]! as! String, previousBuffer))
66+
}
67+
let name = layer["name"] as! String
68+
print("\(name): \(NSDate().timeIntervalSinceDate(t))")
69+
t = NSDate()
70+
++layer_number
71+
72+
}
73+
}
74+
75+
self.gpuCommandLayers.append(currentCommandBuffer)
76+
77+
print("bar")
78+
79+
print("AFTER LAYER DATA CHACES = \(layer_data_caches[0])")
80+
81+
print("POOL TYPE CACHES = \(pool_type_caches)")
82+
83+
print("Time to set up network: \(NSDate().timeIntervalSinceDate(start))")
84+
85+
}
86+
}

0 commit comments

Comments
 (0)