Skip to content

Commit 2e802d9

Browse files
BBufBBuf
committed
Add speed_common_functions.cpp
1 parent de399b7 commit 2e802d9

File tree

1 file changed

+142
-1
lines changed

1 file changed

+142
-1
lines changed

‎speed_common_functions.cpp

+142-1
Original file line numberDiff line numberDiff line change
@@ -133,4 +133,145 @@ int sgn(int X){
133133
if (X > 0) return 1;
134134
if (X < 0) return -1;
135135
return 0;
136-
}
136+
}
137+
138+
//函数12: 获取某个整形变量对应的颜色值
139+
//参考: 无
140+
//简介: 无
141+
void GetRGB(int Color, int *R, int *G, int *B){
142+
*R = Color & 255;
143+
*G = (Color & 65280) / 256;
144+
*B = (Color & 16711680) / 65536;
145+
}
146+
147+
//函数13: 牛顿法近似获取指定数字的算法平方根
148+
//参考: https://www.cnblogs.com/qlky/p/7735145.html
149+
//简介: 仍然是近似算法，近似出了指定数字的平方根
150+
float Sqrt(float X)
151+
{
152+
float HalfX = 0.5f * X; // 对double类型的数字无效
153+
int I = *(int*)&X; // get bits for floating VALUE
154+
I = 0x5f375a86 - (I >> 1); // gives initial guess y0
155+
X = *(float*)&I; // convert bits BACK to float
156+
X = X * (1.5f - HalfX * X * X); // Newton step, repeating increases accuracy
157+
X = X * (1.5f - HalfX * X * X); // Newton step, repeating increases accuracy
158+
X = X * (1.5f - HalfX * X * X); // Newton step, repeating increases accuracy
159+
return 1 / X;
160+
}
161+
162+
//函数14: 无符号短整形直方图数据相加，即是Y = X + Y
163+
//参考: 无
164+
//简介: SSE优化
165+
void HistgramAddShort(unsigned short *X, unsigned short *Y)
166+
{
167+
*(__m128i*)(Y + 0) = _mm_add_epi16(*(__m128i*)&Y[0], *(__m128i*)&X[0]); // 不要想着用自己写的汇编超过他的速度了，已经试过了
168+
*(__m128i*)(Y + 8) = _mm_add_epi16(*(__m128i*)&Y[8], *(__m128i*)&X[8]);
169+
*(__m128i*)(Y + 16) = _mm_add_epi16(*(__m128i*)&Y[16], *(__m128i*)&X[16]);
170+
*(__m128i*)(Y + 24) = _mm_add_epi16(*(__m128i*)&Y[24], *(__m128i*)&X[24]);
171+
*(__m128i*)(Y + 32) = _mm_add_epi16(*(__m128i*)&Y[32], *(__m128i*)&X[32]);
172+
*(__m128i*)(Y + 40) = _mm_add_epi16(*(__m128i*)&Y[40], *(__m128i*)&X[40]);
173+
*(__m128i*)(Y + 48) = _mm_add_epi16(*(__m128i*)&Y[48], *(__m128i*)&X[48]);
174+
*(__m128i*)(Y + 56) = _mm_add_epi16(*(__m128i*)&Y[56], *(__m128i*)&X[56]);
175+
*(__m128i*)(Y + 64) = _mm_add_epi16(*(__m128i*)&Y[64], *(__m128i*)&X[64]);
176+
*(__m128i*)(Y + 72) = _mm_add_epi16(*(__m128i*)&Y[72], *(__m128i*)&X[72]);
177+
*(__m128i*)(Y + 80) = _mm_add_epi16(*(__m128i*)&Y[80], *(__m128i*)&X[80]);
178+
*(__m128i*)(Y + 88) = _mm_add_epi16(*(__m128i*)&Y[88], *(__m128i*)&X[88]);
179+
*(__m128i*)(Y + 96) = _mm_add_epi16(*(__m128i*)&Y[96], *(__m128i*)&X[96]);
180+
*(__m128i*)(Y + 104) = _mm_add_epi16(*(__m128i*)&Y[104], *(__m128i*)&X[104]);
181+
*(__m128i*)(Y + 112) = _mm_add_epi16(*(__m128i*)&Y[112], *(__m128i*)&X[112]);
182+
*(__m128i*)(Y + 120) = _mm_add_epi16(*(__m128i*)&Y[120], *(__m128i*)&X[120]);
183+
*(__m128i*)(Y + 128) = _mm_add_epi16(*(__m128i*)&Y[128], *(__m128i*)&X[128]);
184+
*(__m128i*)(Y + 136) = _mm_add_epi16(*(__m128i*)&Y[136], *(__m128i*)&X[136]);
185+
*(__m128i*)(Y + 144) = _mm_add_epi16(*(__m128i*)&Y[144], *(__m128i*)&X[144]);
186+
*(__m128i*)(Y + 152) = _mm_add_epi16(*(__m128i*)&Y[152], *(__m128i*)&X[152]);
187+
*(__m128i*)(Y + 160) = _mm_add_epi16(*(__m128i*)&Y[160], *(__m128i*)&X[160]);
188+
*(__m128i*)(Y + 168) = _mm_add_epi16(*(__m128i*)&Y[168], *(__m128i*)&X[168]);
189+
*(__m128i*)(Y + 176) = _mm_add_epi16(*(__m128i*)&Y[176], *(__m128i*)&X[176]);
190+
*(__m128i*)(Y + 184) = _mm_add_epi16(*(__m128i*)&Y[184], *(__m128i*)&X[184]);
191+
*(__m128i*)(Y + 192) = _mm_add_epi16(*(__m128i*)&Y[192], *(__m128i*)&X[192]);
192+
*(__m128i*)(Y + 200) = _mm_add_epi16(*(__m128i*)&Y[200], *(__m128i*)&X[200]);
193+
*(__m128i*)(Y + 208) = _mm_add_epi16(*(__m128i*)&Y[208], *(__m128i*)&X[208]);
194+
*(__m128i*)(Y + 216) = _mm_add_epi16(*(__m128i*)&Y[216], *(__m128i*)&X[216]);
195+
*(__m128i*)(Y + 224) = _mm_add_epi16(*(__m128i*)&Y[224], *(__m128i*)&X[224]);
196+
*(__m128i*)(Y + 232) = _mm_add_epi16(*(__m128i*)&Y[232], *(__m128i*)&X[232]);
197+
*(__m128i*)(Y + 240) = _mm_add_epi16(*(__m128i*)&Y[240], *(__m128i*)&X[240]);
198+
*(__m128i*)(Y + 248) = _mm_add_epi16(*(__m128i*)&Y[248], *(__m128i*)&X[248]);
199+
}
200+
201+
//函数15: 无符号短整形直方图数据相减，即是Y = Y - X
202+
//参考: 无
203+
//简介: SSE优化
204+
void HistgramSubShort(unsigned short *X, unsigned short *Y)
205+
{
206+
*(__m128i*)(Y + 0) = _mm_sub_epi16(*(__m128i*)&Y[0], *(__m128i*)&X[0]);
207+
*(__m128i*)(Y + 8) = _mm_sub_epi16(*(__m128i*)&Y[8], *(__m128i*)&X[8]);
208+
*(__m128i*)(Y + 16) = _mm_sub_epi16(*(__m128i*)&Y[16], *(__m128i*)&X[16]);
209+
*(__m128i*)(Y + 24) = _mm_sub_epi16(*(__m128i*)&Y[24], *(__m128i*)&X[24]);
210+
*(__m128i*)(Y + 32) = _mm_sub_epi16(*(__m128i*)&Y[32], *(__m128i*)&X[32]);
211+
*(__m128i*)(Y + 40) = _mm_sub_epi16(*(__m128i*)&Y[40], *(__m128i*)&X[40]);
212+
*(__m128i*)(Y + 48) = _mm_sub_epi16(*(__m128i*)&Y[48], *(__m128i*)&X[48]);
213+
*(__m128i*)(Y + 56) = _mm_sub_epi16(*(__m128i*)&Y[56], *(__m128i*)&X[56]);
214+
*(__m128i*)(Y + 64) = _mm_sub_epi16(*(__m128i*)&Y[64], *(__m128i*)&X[64]);
215+
*(__m128i*)(Y + 72) = _mm_sub_epi16(*(__m128i*)&Y[72], *(__m128i*)&X[72]);
216+
*(__m128i*)(Y + 80) = _mm_sub_epi16(*(__m128i*)&Y[80], *(__m128i*)&X[80]);
217+
*(__m128i*)(Y + 88) = _mm_sub_epi16(*(__m128i*)&Y[88], *(__m128i*)&X[88]);
218+
*(__m128i*)(Y + 96) = _mm_sub_epi16(*(__m128i*)&Y[96], *(__m128i*)&X[96]);
219+
*(__m128i*)(Y + 104) = _mm_sub_epi16(*(__m128i*)&Y[104], *(__m128i*)&X[104]);
220+
*(__m128i*)(Y + 112) = _mm_sub_epi16(*(__m128i*)&Y[112], *(__m128i*)&X[112]);
221+
*(__m128i*)(Y + 120) = _mm_sub_epi16(*(__m128i*)&Y[120], *(__m128i*)&X[120]);
222+
*(__m128i*)(Y + 128) = _mm_sub_epi16(*(__m128i*)&Y[128], *(__m128i*)&X[128]);
223+
*(__m128i*)(Y + 136) = _mm_sub_epi16(*(__m128i*)&Y[136], *(__m128i*)&X[136]);
224+
*(__m128i*)(Y + 144) = _mm_sub_epi16(*(__m128i*)&Y[144], *(__m128i*)&X[144]);
225+
*(__m128i*)(Y + 152) = _mm_sub_epi16(*(__m128i*)&Y[152], *(__m128i*)&X[152]);
226+
*(__m128i*)(Y + 160) = _mm_sub_epi16(*(__m128i*)&Y[160], *(__m128i*)&X[160]);
227+
*(__m128i*)(Y + 168) = _mm_sub_epi16(*(__m128i*)&Y[168], *(__m128i*)&X[168]);
228+
*(__m128i*)(Y + 176) = _mm_sub_epi16(*(__m128i*)&Y[176], *(__m128i*)&X[176]);
229+
*(__m128i*)(Y + 184) = _mm_sub_epi16(*(__m128i*)&Y[184], *(__m128i*)&X[184]);
230+
*(__m128i*)(Y + 192) = _mm_sub_epi16(*(__m128i*)&Y[192], *(__m128i*)&X[192]);
231+
*(__m128i*)(Y + 200) = _mm_sub_epi16(*(__m128i*)&Y[200], *(__m128i*)&X[200]);
232+
*(__m128i*)(Y + 208) = _mm_sub_epi16(*(__m128i*)&Y[208], *(__m128i*)&X[208]);
233+
*(__m128i*)(Y + 216) = _mm_sub_epi16(*(__m128i*)&Y[216], *(__m128i*)&X[216]);
234+
*(__m128i*)(Y + 224) = _mm_sub_epi16(*(__m128i*)&Y[224], *(__m128i*)&X[224]);
235+
*(__m128i*)(Y + 232) = _mm_sub_epi16(*(__m128i*)&Y[232], *(__m128i*)&X[232]);
236+
*(__m128i*)(Y + 240) = _mm_sub_epi16(*(__m128i*)&Y[240], *(__m128i*)&X[240]);
237+
*(__m128i*)(Y + 248) = _mm_sub_epi16(*(__m128i*)&Y[248], *(__m128i*)&X[248]);
238+
}
239+
240+
//函数16: 无符号短整形直方图数据相加减，即是Z = Z + Y - X
241+
//参考: 无
242+
//简介: SSE优化
243+
void HistgramSubAddShort(unsigned short *X, unsigned short *Y, unsigned short *Z)
244+
{
245+
*(__m128i*)(Z + 0) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[0], *(__m128i*)&Z[0]), *(__m128i*)&X[0]); // 不要想着用自己写的汇编超过他的速度了，已经试过了
246+
*(__m128i*)(Z + 8) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[8], *(__m128i*)&Z[8]), *(__m128i*)&X[8]);
247+
*(__m128i*)(Z + 16) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[16], *(__m128i*)&Z[16]), *(__m128i*)&X[16]);
248+
*(__m128i*)(Z + 24) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[24], *(__m128i*)&Z[24]), *(__m128i*)&X[24]);
249+
*(__m128i*)(Z + 32) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[32], *(__m128i*)&Z[32]), *(__m128i*)&X[32]);
250+
*(__m128i*)(Z + 40) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[40], *(__m128i*)&Z[40]), *(__m128i*)&X[40]);
251+
*(__m128i*)(Z + 48) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[48], *(__m128i*)&Z[48]), *(__m128i*)&X[48]);
252+
*(__m128i*)(Z + 56) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[56], *(__m128i*)&Z[56]), *(__m128i*)&X[56]);
253+
*(__m128i*)(Z + 64) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[64], *(__m128i*)&Z[64]), *(__m128i*)&X[64]);
254+
*(__m128i*)(Z + 72) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[72], *(__m128i*)&Z[72]), *(__m128i*)&X[72]);
255+
*(__m128i*)(Z + 80) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[80], *(__m128i*)&Z[80]), *(__m128i*)&X[80]);
256+
*(__m128i*)(Z + 88) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[88], *(__m128i*)&Z[88]), *(__m128i*)&X[88]);
257+
*(__m128i*)(Z + 96) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[96], *(__m128i*)&Z[96]), *(__m128i*)&X[96]);
258+
*(__m128i*)(Z + 104) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[104], *(__m128i*)&Z[104]), *(__m128i*)&X[104]);
259+
*(__m128i*)(Z + 112) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[112], *(__m128i*)&Z[112]), *(__m128i*)&X[112]);
260+
*(__m128i*)(Z + 120) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[120], *(__m128i*)&Z[120]), *(__m128i*)&X[120]);
261+
*(__m128i*)(Z + 128) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[128], *(__m128i*)&Z[128]), *(__m128i*)&X[128]);
262+
*(__m128i*)(Z + 136) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[136], *(__m128i*)&Z[136]), *(__m128i*)&X[136]);
263+
*(__m128i*)(Z + 144) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[144], *(__m128i*)&Z[144]), *(__m128i*)&X[144]);
264+
*(__m128i*)(Z + 152) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[152], *(__m128i*)&Z[152]), *(__m128i*)&X[152]);
265+
*(__m128i*)(Z + 160) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[160], *(__m128i*)&Z[160]), *(__m128i*)&X[160]);
266+
*(__m128i*)(Z + 168) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[168], *(__m128i*)&Z[168]), *(__m128i*)&X[168]);
267+
*(__m128i*)(Z + 176) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[176], *(__m128i*)&Z[176]), *(__m128i*)&X[176]);
268+
*(__m128i*)(Z + 184) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[184], *(__m128i*)&Z[184]), *(__m128i*)&X[184]);
269+
*(__m128i*)(Z + 192) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[192], *(__m128i*)&Z[192]), *(__m128i*)&X[192]);
270+
*(__m128i*)(Z + 200) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[200], *(__m128i*)&Z[200]), *(__m128i*)&X[200]);
271+
*(__m128i*)(Z + 208) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[208], *(__m128i*)&Z[208]), *(__m128i*)&X[208]);
272+
*(__m128i*)(Z + 216) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[216], *(__m128i*)&Z[216]), *(__m128i*)&X[216]);
273+
*(__m128i*)(Z + 224) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[224], *(__m128i*)&Z[224]), *(__m128i*)&X[224]);
274+
*(__m128i*)(Z + 232) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[232], *(__m128i*)&Z[232]), *(__m128i*)&X[232]);
275+
*(__m128i*)(Z + 240) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[240], *(__m128i*)&Z[240]), *(__m128i*)&X[240]);
276+
*(__m128i*)(Z + 248) = _mm_sub_epi16(_mm_add_epi16(*(__m128i*)&Y[248], *(__m128i*)&Z[248]), *(__m128i*)&X[248]);
277+
}

0 commit comments

Comments
 (0)