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
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
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
|
#include "highpassfilter.h"
#include "../../util/rng.h"
#include <cmath>
#include <stdexcept>
#ifdef __SSE__
#define USE_INTRINSICS
#endif
#ifdef USE_INTRINSICS
#include <xmmintrin.h>
#endif
HighPassFilter::~HighPassFilter() {
delete[] _hKernel;
delete[] _vKernel;
}
void HighPassFilter::applyLowPassSimple(const Image2DPtr &image) {
// Guassian convolution can be separated in two 1D convolution
// because of properties of the 2D Gaussian function.
Image2DPtr temp =
Image2D::CreateZeroImagePtr(image->Width(), image->Height());
size_t hKernelMid = _hWindowSize / 2;
for (size_t i = 0; i < _hWindowSize; ++i) {
const num_t kernelValue = _hKernel[i];
const size_t xStart = (i >= hKernelMid) ? 0 : (hKernelMid - i),
xEnd = (i <= hKernelMid)
? image->Width()
: (image->Width() + hKernelMid > i
? (image->Width() - i + hKernelMid)
: 0);
for (unsigned y = 0; y < image->Height(); ++y) {
for (unsigned x = xStart; x < xEnd; ++x)
temp->AddValue(x, y, image->Value(x + i - hKernelMid, y) * kernelValue);
}
}
image->SetAll(0.0);
size_t vKernelMid = _vWindowSize / 2;
for (size_t i = 0; i < _vWindowSize; ++i) {
const num_t kernelValue = _vKernel[i];
const size_t yStart = (i >= vKernelMid) ? 0 : (vKernelMid - i),
yEnd = (i <= vKernelMid)
? image->Height()
: ((image->Height() + vKernelMid > i)
? (image->Height() - i + vKernelMid)
: 0);
for (unsigned y = yStart; y < yEnd; ++y) {
for (unsigned x = 0; x < image->Width(); ++x)
image->AddValue(x, y, temp->Value(x, y + i - vKernelMid) * kernelValue);
}
}
}
void HighPassFilter::applyLowPassSSE(const Image2DPtr &image) {
#ifdef USE_INTRINSICS
Image2DPtr temp =
Image2D::CreateZeroImagePtr(image->Width(), image->Height());
unsigned hKernelMid = _hWindowSize / 2;
for (unsigned i = 0; i < _hWindowSize; ++i) {
const num_t k = _hKernel[i];
const __m128 k4 = _mm_set_ps(k, k, k, k);
unsigned
/* xStart is the first column to start writing to. Note that it might be
* larger than the width. */
xStart = (i >= hKernelMid) ? 0 : (hKernelMid - i),
xEnd = (i <= hKernelMid) ? image->Width()
: (image->Width() + hKernelMid > i
? (image->Width() - i + hKernelMid)
: 0);
for (unsigned y = 0; y < image->Height(); ++y) {
float *tempPtr = temp->ValuePtr(xStart, y);
const float *imagePtr = image->ValuePtr(xStart + i - hKernelMid, y);
unsigned x = xStart;
for (; x + 4 < xEnd; x += 4) {
const __m128 imageVal = _mm_loadu_ps(imagePtr),
tempVal = _mm_loadu_ps(tempPtr);
// *tempPtr += k * (*imagePtr);
_mm_storeu_ps(tempPtr, _mm_add_ps(tempVal, _mm_mul_ps(imageVal, k4)));
tempPtr += 4;
imagePtr += 4;
}
for (; x < xEnd; ++x) {
*tempPtr += k * (*imagePtr);
++tempPtr;
++imagePtr;
}
}
}
image->SetAll(0.0);
unsigned vKernelMid = _vWindowSize / 2;
for (unsigned i = 0; i < _vWindowSize; ++i) {
const num_t k = _vKernel[i];
const __m128 k4 = _mm_set_ps(k, k, k, k);
const unsigned yStart = (i >= vKernelMid) ? 0 : (vKernelMid - i),
yEnd = (i <= vKernelMid)
? image->Height()
: ((image->Height() + vKernelMid > i)
? (image->Height() - i + vKernelMid)
: 0);
for (unsigned y = yStart; y < yEnd; ++y) {
const float *tempPtr = temp->ValuePtr(0, y + i - vKernelMid);
float *imagePtr = image->ValuePtr(0, y);
unsigned x = 0;
for (; x + 4 < image->Width(); x += 4) {
const __m128 imageVal = _mm_load_ps(imagePtr),
tempVal = _mm_load_ps(tempPtr);
// *imagePtr += k * (*tempPtr);
_mm_store_ps(imagePtr, _mm_add_ps(imageVal, _mm_mul_ps(tempVal, k4)));
tempPtr += 4;
imagePtr += 4;
}
for (; x < image->Width(); ++x) {
*imagePtr += k * (*tempPtr);
++tempPtr;
++imagePtr;
}
}
}
#else
throw std::runtime_error("SSE function called without SSE available");
#endif
}
Image2DPtr HighPassFilter::ApplyHighPass(const Image2DCPtr &image,
const Mask2DCPtr &mask) {
Image2DPtr outputImage = ApplyLowPass(image, mask);
outputImage->SubtractAsRHS(image);
return outputImage;
}
Image2DPtr HighPassFilter::ApplyLowPass(const Image2DCPtr &image,
const Mask2DCPtr &mask) {
initializeKernel();
Image2DPtr outputImage =
Image2D::CreateUnsetImagePtr(image->Width(), image->Height()),
weights =
Image2D::CreateUnsetImagePtr(image->Width(), image->Height());
setFlaggedValuesToZeroAndMakeWeights(image, outputImage, mask, weights);
applyLowPass(outputImage);
applyLowPass(weights);
elementWiseDivide(outputImage, weights);
weights.reset();
return outputImage;
}
void HighPassFilter::initializeKernel() {
if (_hKernel == 0) {
_hKernel = new num_t[_hWindowSize];
const int midPointX = _hWindowSize / 2;
for (int x = 0; x < (int)_hWindowSize; ++x)
_hKernel[x] =
RNG::EvaluateUnnormalizedGaussian(x - midPointX, _hKernelSigmaSq);
}
if (_vKernel == 0) {
_vKernel = new num_t[_vWindowSize];
const int midPointY = _vWindowSize / 2;
for (int y = 0; y < (int)_vWindowSize; ++y)
_vKernel[y] =
RNG::EvaluateUnnormalizedGaussian(y - midPointY, _vKernelSigmaSq);
}
}
void HighPassFilter::setFlaggedValuesToZeroAndMakeWeightsSimple(
const Image2DCPtr &inputImage, const Image2DPtr &outputImage,
const Mask2DCPtr &inputMask, const Image2DPtr &weightsOutput) {
const size_t width = inputImage->Width();
for (size_t y = 0; y < inputImage->Height(); ++y) {
for (size_t x = 0; x < width; ++x) {
if (inputMask->Value(x, y) || !std::isfinite(inputImage->Value(x, y))) {
outputImage->SetValue(x, y, 0.0);
weightsOutput->SetValue(x, y, 0.0);
} else {
outputImage->SetValue(x, y, inputImage->Value(x, y));
weightsOutput->SetValue(x, y, 1.0);
}
}
}
}
void HighPassFilter::setFlaggedValuesToZeroAndMakeWeightsSSE(
const Image2DCPtr &inputImage, const Image2DPtr &outputImage,
const Mask2DCPtr &inputMask, const Image2DPtr &weightsOutput) {
#ifdef USE_INTRINSICS
const size_t width = inputImage->Width();
const __m128i zero4i = _mm_set_epi32(0, 0, 0, 0);
const __m128 zero4 = _mm_set_ps(0.0, 0.0, 0.0, 0.0);
const __m128 one4 = _mm_set_ps(1.0, 1.0, 1.0, 1.0);
for (size_t y = 0; y < inputImage->Height(); ++y) {
const bool *rowPtr = inputMask->ValuePtr(0, y);
const float *inputPtr = inputImage->ValuePtr(0, y);
float *outputPtr = outputImage->ValuePtr(0, y);
float *weightsPtr = weightsOutput->ValuePtr(0, y);
const float *end = inputPtr + width;
while (inputPtr < end) {
// Assign each integer to one bool in the mask
// Convert false to 0xFFFFFFFF and true to 0
__m128 conditionMask = _mm_castsi128_ps(_mm_cmpeq_epi32(
_mm_set_epi32(rowPtr[3] || !std::isfinite(inputPtr[3]),
rowPtr[2] || !std::isfinite(inputPtr[2]),
rowPtr[1] || !std::isfinite(inputPtr[1]),
rowPtr[0] || !std::isfinite(inputPtr[0])),
zero4i));
_mm_store_ps(weightsPtr, _mm_or_ps(_mm_and_ps(conditionMask, one4),
_mm_andnot_ps(conditionMask, zero4)));
_mm_store_ps(outputPtr,
_mm_or_ps(_mm_and_ps(conditionMask, _mm_load_ps(inputPtr)),
_mm_andnot_ps(conditionMask, zero4)));
rowPtr += 4;
outputPtr += 4;
inputPtr += 4;
weightsPtr += 4;
}
}
#else
throw std::runtime_error("SSE function called without SSE available");
#endif
}
void HighPassFilter::elementWiseDivideSimple(const Image2DPtr &leftHand,
const Image2DCPtr &rightHand) {
for (unsigned y = 0; y < leftHand->Height(); ++y) {
for (unsigned x = 0; x < leftHand->Width(); ++x) {
if (rightHand->Value(x, y) == 0.0)
leftHand->SetValue(x, y, 0.0);
else
leftHand->SetValue(x, y,
leftHand->Value(x, y) / rightHand->Value(x, y));
}
}
}
void HighPassFilter::elementWiseDivideSSE(const Image2DPtr &leftHand,
const Image2DCPtr &rightHand) {
#ifdef USE_INTRINSICS
const __m128 zero4 = _mm_set_ps(0.0, 0.0, 0.0, 0.0);
for (unsigned y = 0; y < leftHand->Height(); ++y) {
float *leftHandPtr = leftHand->ValuePtr(0, y);
const float *rightHandPtr = rightHand->ValuePtr(0, y);
float *end = leftHandPtr + leftHand->Width();
while (leftHandPtr < end) {
__m128 l = _mm_load_ps(leftHandPtr), r = _mm_load_ps(rightHandPtr);
__m128 conditionMask = _mm_cmpeq_ps(r, zero4);
_mm_store_ps(leftHandPtr,
_mm_or_ps(_mm_and_ps(conditionMask, zero4),
_mm_andnot_ps(conditionMask, _mm_div_ps(l, r))));
leftHandPtr += 4;
rightHandPtr += 4;
}
}
#else
throw std::runtime_error("SSE function called without SSE available");
#endif
}
|
