Line | Branch | Exec | Source |
---|---|---|---|
1 | /* -*- c++ -*- */ | ||
2 | /* | ||
3 | * Copyright 2012, 2014 Free Software Foundation, Inc. | ||
4 | * | ||
5 | * This file is part of VOLK | ||
6 | * | ||
7 | * SPDX-License-Identifier: LGPL-3.0-or-later | ||
8 | */ | ||
9 | |||
10 | /*! | ||
11 | * \page volk_32f_s32f_convert_32i | ||
12 | * | ||
13 | * \b Overview | ||
14 | * | ||
15 | * Converts a floating point number to a 32-bit integer after applying a | ||
16 | * scaling factor. | ||
17 | * | ||
18 | * <b>Dispatcher Prototype</b> | ||
19 | * \code | ||
20 | * void volk_32f_s32f_convert_32i(int32_t* outputVector, const float* inputVector, const | ||
21 | * float scalar, unsigned int num_points) \endcode | ||
22 | * | ||
23 | * \b Inputs | ||
24 | * \li inputVector: the input vector of floats. | ||
25 | * \li scalar: The value multiplied against each point in the input buffer. | ||
26 | * \li num_points: The number of data points. | ||
27 | * | ||
28 | * \b Outputs | ||
29 | * \li outputVector: The output vector. | ||
30 | * | ||
31 | * \b Example | ||
32 | * Convert floats from [-1,1] to integers with a scale of 5 to maintain smallest delta | ||
33 | * \code | ||
34 | * int N = 10; | ||
35 | * unsigned int alignment = volk_get_alignment(); | ||
36 | * float* increasing = (float*)volk_malloc(sizeof(float)*N, alignment); | ||
37 | * int32_t* out = (int32_t*)volk_malloc(sizeof(int32_t)*N, alignment); | ||
38 | * | ||
39 | * for(unsigned int ii = 0; ii < N; ++ii){ | ||
40 | * increasing[ii] = 2.f * ((float)ii / (float)N) - 1.f; | ||
41 | * } | ||
42 | * | ||
43 | * // Normalize by the smallest delta (0.2 in this example) | ||
44 | * float scale = 5.f; | ||
45 | * | ||
46 | * volk_32f_s32f_convert_32i(out, increasing, scale, N); | ||
47 | * | ||
48 | * for(unsigned int ii = 0; ii < N; ++ii){ | ||
49 | * printf("out[%u] = %i\n", ii, out[ii]); | ||
50 | * } | ||
51 | * | ||
52 | * volk_free(increasing); | ||
53 | * volk_free(out); | ||
54 | * \endcode | ||
55 | */ | ||
56 | |||
57 | #ifndef INCLUDED_volk_32f_s32f_convert_32i_u_H | ||
58 | #define INCLUDED_volk_32f_s32f_convert_32i_u_H | ||
59 | |||
60 | #include <inttypes.h> | ||
61 | #include <limits.h> | ||
62 | #include <stdio.h> | ||
63 | |||
64 | #ifdef LV_HAVE_AVX | ||
65 | #include <immintrin.h> | ||
66 | |||
67 | 2 | static inline void volk_32f_s32f_convert_32i_u_avx(int32_t* outputVector, | |
68 | const float* inputVector, | ||
69 | const float scalar, | ||
70 | unsigned int num_points) | ||
71 | { | ||
72 | 2 | unsigned int number = 0; | |
73 | |||
74 | 2 | const unsigned int eighthPoints = num_points / 8; | |
75 | |||
76 | 2 | const float* inputVectorPtr = (const float*)inputVector; | |
77 | 2 | int32_t* outputVectorPtr = outputVector; | |
78 | |||
79 | 2 | float min_val = INT_MIN; | |
80 | 2 | float max_val = (uint32_t)INT_MAX + 1; | |
81 | float r; | ||
82 | |||
83 | 2 | __m256 vScalar = _mm256_set1_ps(scalar); | |
84 | __m256 inputVal1; | ||
85 | __m256i intInputVal1; | ||
86 | 2 | __m256 vmin_val = _mm256_set1_ps(min_val); | |
87 | 2 | __m256 vmax_val = _mm256_set1_ps(max_val); | |
88 | |||
89 |
2/2✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
|
32768 | for (; number < eighthPoints; number++) { |
90 | 32766 | inputVal1 = _mm256_loadu_ps(inputVectorPtr); | |
91 | 32766 | inputVectorPtr += 8; | |
92 | |||
93 | 98298 | inputVal1 = _mm256_max_ps( | |
94 | _mm256_min_ps(_mm256_mul_ps(inputVal1, vScalar), vmax_val), vmin_val); | ||
95 | 32766 | intInputVal1 = _mm256_cvtps_epi32(inputVal1); | |
96 | |||
97 | _mm256_storeu_si256((__m256i*)outputVectorPtr, intInputVal1); | ||
98 | 32766 | outputVectorPtr += 8; | |
99 | } | ||
100 | |||
101 | 2 | number = eighthPoints * 8; | |
102 |
2/2✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
|
16 | for (; number < num_points; number++) { |
103 | 14 | r = inputVector[number] * scalar; | |
104 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 14 times.
|
14 | if (r > max_val) |
105 | ✗ | r = max_val; | |
106 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 14 times.
|
14 | else if (r < min_val) |
107 | ✗ | r = min_val; | |
108 | 14 | outputVector[number] = (int32_t)rintf(r); | |
109 | } | ||
110 | 2 | } | |
111 | |||
112 | #endif /* LV_HAVE_AVX */ | ||
113 | |||
114 | #ifdef LV_HAVE_SSE2 | ||
115 | #include <emmintrin.h> | ||
116 | |||
117 | 2 | static inline void volk_32f_s32f_convert_32i_u_sse2(int32_t* outputVector, | |
118 | const float* inputVector, | ||
119 | const float scalar, | ||
120 | unsigned int num_points) | ||
121 | { | ||
122 | 2 | unsigned int number = 0; | |
123 | |||
124 | 2 | const unsigned int quarterPoints = num_points / 4; | |
125 | |||
126 | 2 | const float* inputVectorPtr = (const float*)inputVector; | |
127 | 2 | int32_t* outputVectorPtr = outputVector; | |
128 | |||
129 | 2 | float min_val = INT_MIN; | |
130 | 2 | float max_val = (uint32_t)INT_MAX + 1; | |
131 | float r; | ||
132 | |||
133 | 2 | __m128 vScalar = _mm_set_ps1(scalar); | |
134 | __m128 inputVal1; | ||
135 | __m128i intInputVal1; | ||
136 | 2 | __m128 vmin_val = _mm_set_ps1(min_val); | |
137 | 2 | __m128 vmax_val = _mm_set_ps1(max_val); | |
138 | |||
139 |
2/2✓ Branch 0 taken 65534 times.
✓ Branch 1 taken 2 times.
|
65536 | for (; number < quarterPoints; number++) { |
140 | 65534 | inputVal1 = _mm_loadu_ps(inputVectorPtr); | |
141 | 65534 | inputVectorPtr += 4; | |
142 | |||
143 | inputVal1 = | ||
144 | 196602 | _mm_max_ps(_mm_min_ps(_mm_mul_ps(inputVal1, vScalar), vmax_val), vmin_val); | |
145 | 65534 | intInputVal1 = _mm_cvtps_epi32(inputVal1); | |
146 | |||
147 | _mm_storeu_si128((__m128i*)outputVectorPtr, intInputVal1); | ||
148 | 65534 | outputVectorPtr += 4; | |
149 | } | ||
150 | |||
151 | 2 | number = quarterPoints * 4; | |
152 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 2 times.
|
8 | for (; number < num_points; number++) { |
153 | 6 | r = inputVector[number] * scalar; | |
154 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
|
6 | if (r > max_val) |
155 | ✗ | r = max_val; | |
156 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
|
6 | else if (r < min_val) |
157 | ✗ | r = min_val; | |
158 | 6 | outputVector[number] = (int32_t)rintf(r); | |
159 | } | ||
160 | 2 | } | |
161 | |||
162 | #endif /* LV_HAVE_SSE2 */ | ||
163 | |||
164 | |||
165 | #ifdef LV_HAVE_SSE | ||
166 | #include <xmmintrin.h> | ||
167 | |||
168 | 2 | static inline void volk_32f_s32f_convert_32i_u_sse(int32_t* outputVector, | |
169 | const float* inputVector, | ||
170 | const float scalar, | ||
171 | unsigned int num_points) | ||
172 | { | ||
173 | 2 | unsigned int number = 0; | |
174 | |||
175 | 2 | const unsigned int quarterPoints = num_points / 4; | |
176 | |||
177 | 2 | const float* inputVectorPtr = (const float*)inputVector; | |
178 | 2 | int32_t* outputVectorPtr = outputVector; | |
179 | |||
180 | 2 | float min_val = INT_MIN; | |
181 | 2 | float max_val = (uint32_t)INT_MAX + 1; | |
182 | float r; | ||
183 | |||
184 | 2 | __m128 vScalar = _mm_set_ps1(scalar); | |
185 | __m128 ret; | ||
186 | 2 | __m128 vmin_val = _mm_set_ps1(min_val); | |
187 | 2 | __m128 vmax_val = _mm_set_ps1(max_val); | |
188 | |||
189 | __VOLK_ATTR_ALIGNED(16) float outputFloatBuffer[4]; | ||
190 | |||
191 |
2/2✓ Branch 0 taken 65534 times.
✓ Branch 1 taken 2 times.
|
65536 | for (; number < quarterPoints; number++) { |
192 | 65534 | ret = _mm_loadu_ps(inputVectorPtr); | |
193 | 65534 | inputVectorPtr += 4; | |
194 | |||
195 | 196602 | ret = _mm_max_ps(_mm_min_ps(_mm_mul_ps(ret, vScalar), vmax_val), vmin_val); | |
196 | |||
197 | _mm_store_ps(outputFloatBuffer, ret); | ||
198 | 65534 | *outputVectorPtr++ = (int32_t)rintf(outputFloatBuffer[0]); | |
199 | 65534 | *outputVectorPtr++ = (int32_t)rintf(outputFloatBuffer[1]); | |
200 | 65534 | *outputVectorPtr++ = (int32_t)rintf(outputFloatBuffer[2]); | |
201 | 65534 | *outputVectorPtr++ = (int32_t)rintf(outputFloatBuffer[3]); | |
202 | } | ||
203 | |||
204 | 2 | number = quarterPoints * 4; | |
205 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 2 times.
|
8 | for (; number < num_points; number++) { |
206 | 6 | r = inputVector[number] * scalar; | |
207 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
|
6 | if (r > max_val) |
208 | ✗ | r = max_val; | |
209 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
|
6 | else if (r < min_val) |
210 | ✗ | r = min_val; | |
211 | 6 | outputVector[number] = (int32_t)rintf(r); | |
212 | } | ||
213 | 2 | } | |
214 | |||
215 | #endif /* LV_HAVE_SSE */ | ||
216 | |||
217 | |||
218 | #ifdef LV_HAVE_GENERIC | ||
219 | |||
220 | 4 | static inline void volk_32f_s32f_convert_32i_generic(int32_t* outputVector, | |
221 | const float* inputVector, | ||
222 | const float scalar, | ||
223 | unsigned int num_points) | ||
224 | { | ||
225 | 4 | int32_t* outputVectorPtr = outputVector; | |
226 | 4 | const float* inputVectorPtr = inputVector; | |
227 | 4 | const float min_val = (float)INT_MIN; | |
228 | 4 | const float max_val = (float)((uint32_t)INT_MAX + 1); | |
229 | |||
230 |
2/2✓ Branch 0 taken 524284 times.
✓ Branch 1 taken 4 times.
|
524288 | for (unsigned int number = 0; number < num_points; number++) { |
231 | 524284 | const float r = *inputVectorPtr++ * scalar; | |
232 | int s; | ||
233 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 524284 times.
|
524284 | if (r >= max_val) |
234 | ✗ | s = INT_MAX; | |
235 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 524284 times.
|
524284 | else if (r < min_val) |
236 | ✗ | s = INT_MIN; | |
237 | else | ||
238 | 524284 | s = (int32_t)rintf(r); | |
239 | 524284 | *outputVectorPtr++ = s; | |
240 | } | ||
241 | 4 | } | |
242 | |||
243 | #endif /* LV_HAVE_GENERIC */ | ||
244 | |||
245 | |||
246 | #endif /* INCLUDED_volk_32f_s32f_convert_32i_u_H */ | ||
247 | #ifndef INCLUDED_volk_32f_s32f_convert_32i_a_H | ||
248 | #define INCLUDED_volk_32f_s32f_convert_32i_a_H | ||
249 | |||
250 | #include <inttypes.h> | ||
251 | #include <stdio.h> | ||
252 | #include <volk/volk_common.h> | ||
253 | |||
254 | #ifdef LV_HAVE_AVX | ||
255 | #include <immintrin.h> | ||
256 | |||
257 | 2 | static inline void volk_32f_s32f_convert_32i_a_avx(int32_t* outputVector, | |
258 | const float* inputVector, | ||
259 | const float scalar, | ||
260 | unsigned int num_points) | ||
261 | { | ||
262 | 2 | unsigned int number = 0; | |
263 | |||
264 | 2 | const unsigned int eighthPoints = num_points / 8; | |
265 | |||
266 | 2 | const float* inputVectorPtr = (const float*)inputVector; | |
267 | 2 | int32_t* outputVectorPtr = outputVector; | |
268 | |||
269 | 2 | float min_val = INT_MIN; | |
270 | 2 | float max_val = (uint32_t)INT_MAX + 1; | |
271 | float r; | ||
272 | |||
273 | 2 | __m256 vScalar = _mm256_set1_ps(scalar); | |
274 | __m256 inputVal1; | ||
275 | __m256i intInputVal1; | ||
276 | 2 | __m256 vmin_val = _mm256_set1_ps(min_val); | |
277 | 2 | __m256 vmax_val = _mm256_set1_ps(max_val); | |
278 | |||
279 |
2/2✓ Branch 0 taken 32766 times.
✓ Branch 1 taken 2 times.
|
32768 | for (; number < eighthPoints; number++) { |
280 | 32766 | inputVal1 = _mm256_load_ps(inputVectorPtr); | |
281 | 32766 | inputVectorPtr += 8; | |
282 | |||
283 | 98298 | inputVal1 = _mm256_max_ps( | |
284 | _mm256_min_ps(_mm256_mul_ps(inputVal1, vScalar), vmax_val), vmin_val); | ||
285 | 32766 | intInputVal1 = _mm256_cvtps_epi32(inputVal1); | |
286 | |||
287 | _mm256_store_si256((__m256i*)outputVectorPtr, intInputVal1); | ||
288 | 32766 | outputVectorPtr += 8; | |
289 | } | ||
290 | |||
291 | 2 | number = eighthPoints * 8; | |
292 |
2/2✓ Branch 0 taken 14 times.
✓ Branch 1 taken 2 times.
|
16 | for (; number < num_points; number++) { |
293 | 14 | r = inputVector[number] * scalar; | |
294 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 14 times.
|
14 | if (r > max_val) |
295 | ✗ | r = max_val; | |
296 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 14 times.
|
14 | else if (r < min_val) |
297 | ✗ | r = min_val; | |
298 | 14 | outputVector[number] = (int32_t)rintf(r); | |
299 | } | ||
300 | 2 | } | |
301 | |||
302 | #endif /* LV_HAVE_AVX */ | ||
303 | |||
304 | |||
305 | #ifdef LV_HAVE_SSE2 | ||
306 | #include <emmintrin.h> | ||
307 | |||
308 | 2 | static inline void volk_32f_s32f_convert_32i_a_sse2(int32_t* outputVector, | |
309 | const float* inputVector, | ||
310 | const float scalar, | ||
311 | unsigned int num_points) | ||
312 | { | ||
313 | 2 | unsigned int number = 0; | |
314 | |||
315 | 2 | const unsigned int quarterPoints = num_points / 4; | |
316 | |||
317 | 2 | const float* inputVectorPtr = (const float*)inputVector; | |
318 | 2 | int32_t* outputVectorPtr = outputVector; | |
319 | |||
320 | 2 | float min_val = INT_MIN; | |
321 | 2 | float max_val = (uint32_t)INT_MAX + 1; | |
322 | float r; | ||
323 | |||
324 | 2 | __m128 vScalar = _mm_set_ps1(scalar); | |
325 | __m128 inputVal1; | ||
326 | __m128i intInputVal1; | ||
327 | 2 | __m128 vmin_val = _mm_set_ps1(min_val); | |
328 | 2 | __m128 vmax_val = _mm_set_ps1(max_val); | |
329 | |||
330 |
2/2✓ Branch 0 taken 65534 times.
✓ Branch 1 taken 2 times.
|
65536 | for (; number < quarterPoints; number++) { |
331 | 65534 | inputVal1 = _mm_load_ps(inputVectorPtr); | |
332 | 65534 | inputVectorPtr += 4; | |
333 | |||
334 | inputVal1 = | ||
335 | 196602 | _mm_max_ps(_mm_min_ps(_mm_mul_ps(inputVal1, vScalar), vmax_val), vmin_val); | |
336 | 65534 | intInputVal1 = _mm_cvtps_epi32(inputVal1); | |
337 | |||
338 | _mm_store_si128((__m128i*)outputVectorPtr, intInputVal1); | ||
339 | 65534 | outputVectorPtr += 4; | |
340 | } | ||
341 | |||
342 | 2 | number = quarterPoints * 4; | |
343 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 2 times.
|
8 | for (; number < num_points; number++) { |
344 | 6 | r = inputVector[number] * scalar; | |
345 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
|
6 | if (r > max_val) |
346 | ✗ | r = max_val; | |
347 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
|
6 | else if (r < min_val) |
348 | ✗ | r = min_val; | |
349 | 6 | outputVector[number] = (int32_t)rintf(r); | |
350 | } | ||
351 | 2 | } | |
352 | |||
353 | #endif /* LV_HAVE_SSE2 */ | ||
354 | |||
355 | |||
356 | #ifdef LV_HAVE_SSE | ||
357 | #include <xmmintrin.h> | ||
358 | |||
359 | 2 | static inline void volk_32f_s32f_convert_32i_a_sse(int32_t* outputVector, | |
360 | const float* inputVector, | ||
361 | const float scalar, | ||
362 | unsigned int num_points) | ||
363 | { | ||
364 | 2 | unsigned int number = 0; | |
365 | |||
366 | 2 | const unsigned int quarterPoints = num_points / 4; | |
367 | |||
368 | 2 | const float* inputVectorPtr = (const float*)inputVector; | |
369 | 2 | int32_t* outputVectorPtr = outputVector; | |
370 | |||
371 | 2 | float min_val = INT_MIN; | |
372 | 2 | float max_val = (uint32_t)INT_MAX + 1; | |
373 | float r; | ||
374 | |||
375 | 2 | __m128 vScalar = _mm_set_ps1(scalar); | |
376 | __m128 ret; | ||
377 | 2 | __m128 vmin_val = _mm_set_ps1(min_val); | |
378 | 2 | __m128 vmax_val = _mm_set_ps1(max_val); | |
379 | |||
380 | __VOLK_ATTR_ALIGNED(16) float outputFloatBuffer[4]; | ||
381 | |||
382 |
2/2✓ Branch 0 taken 65534 times.
✓ Branch 1 taken 2 times.
|
65536 | for (; number < quarterPoints; number++) { |
383 | 65534 | ret = _mm_load_ps(inputVectorPtr); | |
384 | 65534 | inputVectorPtr += 4; | |
385 | |||
386 | 196602 | ret = _mm_max_ps(_mm_min_ps(_mm_mul_ps(ret, vScalar), vmax_val), vmin_val); | |
387 | |||
388 | _mm_store_ps(outputFloatBuffer, ret); | ||
389 | 65534 | *outputVectorPtr++ = (int32_t)rintf(outputFloatBuffer[0]); | |
390 | 65534 | *outputVectorPtr++ = (int32_t)rintf(outputFloatBuffer[1]); | |
391 | 65534 | *outputVectorPtr++ = (int32_t)rintf(outputFloatBuffer[2]); | |
392 | 65534 | *outputVectorPtr++ = (int32_t)rintf(outputFloatBuffer[3]); | |
393 | } | ||
394 | |||
395 | 2 | number = quarterPoints * 4; | |
396 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 2 times.
|
8 | for (; number < num_points; number++) { |
397 | 6 | r = inputVector[number] * scalar; | |
398 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
|
6 | if (r > max_val) |
399 | ✗ | r = max_val; | |
400 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6 times.
|
6 | else if (r < min_val) |
401 | ✗ | r = min_val; | |
402 | 6 | outputVector[number] = (int32_t)rintf(r); | |
403 | } | ||
404 | 2 | } | |
405 | |||
406 | #endif /* LV_HAVE_SSE */ | ||
407 | |||
408 | |||
409 | #ifdef LV_HAVE_GENERIC | ||
410 | |||
411 | 2 | static inline void volk_32f_s32f_convert_32i_a_generic(int32_t* outputVector, | |
412 | const float* inputVector, | ||
413 | const float scalar, | ||
414 | unsigned int num_points) | ||
415 | { | ||
416 | 2 | volk_32f_s32f_convert_32i_generic(outputVector, inputVector, scalar, num_points); | |
417 | 2 | } | |
418 | |||
419 | #endif /* LV_HAVE_GENERIC */ | ||
420 | |||
421 | #endif /* INCLUDED_volk_32f_s32f_convert_32i_a_H */ | ||
422 |