-
Notifications
You must be signed in to change notification settings - Fork 0
/
minMax.cu
307 lines (282 loc) · 8.35 KB
/
minMax.cu
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
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
#ifndef __MINMAX.CU__
#define __MINMAX.CU__
#define LENA_WIDTH 512
#define LENA_HEIGHT 512
#define LENA_SIZE 262144 //512*512
//two different min and max implementations in order to test divergence
#define MIN y ^ ((x ^ y) & -(x < y))
#define MAX x ^ ((x ^ y) & -(x < y))
//#define MIN (x < y) ? x : y
//#define MAX (x < y) ? y : x
/*
This kernel finds mins and maxes for each block and stores them in global memory side by side.
Inspired by parallel reduction v7
Keeps thread_count*2 elements in shared memory for maxes and mins
Each thread takes multiple elements by traversing diffrent parts of data
Loop unrolling contributes to the performance
When tid is less than 32 there is no need to use __syncthreads() because there is one warp left.
When tid==0 it takes the min and max to the global memory and exits.
*/
template<unsigned int blockSize>
__global__ void findMinMax(float *g_idata, float *res) {
extern __shared__ float data[];
// Dynamically taking shared memory
// Dividing shared memory in the following way
float *minData = data;
float *maxData = &data[blockSize];
unsigned int tid = threadIdx.x;
unsigned int i = blockIdx.x * (blockSize * 2) + tid;
unsigned int gridSize = blockSize * 2 * gridDim.x;
minData[tid] = INT32_MAX;
maxData[tid] = INT32_MIN;
int x, y, min, max;
//Traversing all the data by jumping as much as gridSize elements
while (i < LENA_SIZE) {
x = g_idata[i];
y = g_idata[i + blockSize];
min = MIN;
max = MAX;
x = minData[tid];
y = min;
minData[tid] = MIN;
x = maxData[tid];
y = max;
maxData[tid] = MAX;
i += gridSize;
}
__syncthreads();
if (blockSize >= 1024) {
if (tid < 512) {
x = minData[tid];
y = minData[tid + 512];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 512];
maxData[tid] = MAX;
}
__syncthreads();
}
if (blockSize >= 512) {
if (tid < 256) {
x = minData[tid];
y = minData[tid + 256];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 256];
maxData[tid] = MAX;
}
__syncthreads();
}
if (blockSize >= 256) {
if (tid < 128) {
x = minData[tid];
y = minData[tid + 128];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 128];
maxData[tid] = MAX;
}
__syncthreads();
}
if (blockSize >= 128) {
if (tid < 64) {
x = minData[tid];
y = minData[tid + 64];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 64];
maxData[tid] = MAX;
}
__syncthreads();
}
if (tid < 32) {
if (blockSize >= 64) {
x = minData[tid];
y = minData[tid + 32];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 32];
maxData[tid] = MAX;
}
if (blockSize >= 32) {
x = minData[tid];
y = minData[tid + 16];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 16];
maxData[tid] = MAX;
}
if (blockSize >= 16) {
x = minData[tid];
y = minData[tid + 8];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 8];
maxData[tid] = MAX;
}
if (blockSize >= 8) {
x = minData[tid];
y = minData[tid + 4];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 4];
maxData[tid] = MAX;
}
if (blockSize >= 4) {
x = minData[tid];
y = minData[tid + 2];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 2];
maxData[tid] = MAX;
}
if (blockSize >= 2) {
x = minData[tid];
y = minData[tid + 1];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 1];
maxData[tid] = MAX;
}
}
// store min&max in res array side by side
if (tid == 0) {
res[blockIdx.x] = minData[0];
res[gridDim.x + blockIdx.x] = maxData[0];
}
}
/*
This traverses all the image and recalculate the values according to the given min and constant values found before
*/
__global__ void calcPixelVal(float *g_idata, float* constant, float* min)
{
unsigned int i = blockIdx.x * blockDim.x + threadIdx.x;
if(i<LENA_SIZE)g_idata[i]=(g_idata[i]-(*min))*(*constant);
}
/*This is called after min&max value is called
Only one thread executes this
*/
__device__ void calcConstant(float* constant, float *min, float *max) {
*constant = 255.0f / (*max - *min);
}
/* This is nearly same with the findMinMax() kernel.
After gathering data from first kernel only a few mins and maxes are left
With one block this kernel finds the min&max and calls the calcConstant()
The idea is same with the findMinMax() kernel but this just loads 2 elements for each thread
*/
template<unsigned int blockSize>
__global__ void finalMinMax(float *g_idata, float *constant, float *min, float *max) {
extern __shared__ float data[];
float *minData = data;
float *maxData = &data[blockSize];
unsigned int tid = threadIdx.x;
int x, y;
//load 2 elemens for each
x = g_idata[tid];
y = g_idata[tid + blockSize];
minData[tid] = MIN;
x = g_idata[tid + blockSize * 2];
y = g_idata[tid + blockSize * 2 + blockSize];
maxData[tid] = MAX;
__syncthreads();
if (blockSize >= 1024) {
if (tid < 512) {
x = minData[tid];
y = minData[tid + 512];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 512];
maxData[tid] = MAX;
}
__syncthreads();
}
if (blockSize >= 512) {
if (tid < 256) {
x = minData[tid];
y = minData[tid + 256];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 256];
maxData[tid] = MAX;
}
__syncthreads();
}
if (blockSize >= 256) {
if (tid < 128) {
x = minData[tid];
y = minData[tid + 128];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 128];
maxData[tid] = MAX;
}
__syncthreads();
}
if (blockSize >= 128) {
if (tid < 64) {
x = minData[tid];
y = minData[tid + 64];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 64];
maxData[tid] = MAX;
}
__syncthreads();
}
if (tid < 32) {
if (blockSize >= 64) {
x = minData[tid];
y = minData[tid + 32];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 32];
maxData[tid] = MAX;
}
if (blockSize >= 32) {
x = minData[tid];
y = minData[tid + 16];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 16];
maxData[tid] = MAX;
}
if (blockSize >= 16) {
x = minData[tid];
y = minData[tid + 8];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 8];
maxData[tid] = MAX;
}
if (blockSize >= 8) {
x = minData[tid];
y = minData[tid + 4];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 4];
maxData[tid] = MAX;
}
if (blockSize >= 4) {
x = minData[tid];
y = minData[tid + 2];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 2];
maxData[tid] = MAX;
}
if (blockSize >= 2) {
x = minData[tid];
y = minData[tid + 1];
minData[tid] = MIN;
x = maxData[tid];
y = maxData[tid + 1];
maxData[tid] = MAX;
}
}
if (tid == 0) {
*min = minData[0];
*max = maxData[0];
calcConstant(constant, min, max);
}
}
#endif