GCC Code Coverage Report

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_16i_x4_quad_max_star_16i
12			*
13			* \b Overview
14			*
15			* <FIXME>
16			*
17			* <b>Dispatcher Prototype</b>
18			* \code
19			* void volk_16i_x4_quad_max_star_16i(short* target, short* src0, short* src1, short*
20			* src2, short* src3, unsigned int num_points) \endcode
21			*
22			* \b Inputs
23			* \li src0: The input vector 0.
24			* \li src1: The input vector 1.
25			* \li src2: The input vector 2.
26			* \li src3: The input vector 3.
27			* \li num_points: The number of data points.
28			*
29			* \b Outputs
30			* \li target: The output value.
31			*
32			* \b Example
33			* \code
34			* int N = 10000;
35			*
36			* volk_16i_x4_quad_max_star_16i();
37			*
38			* volk_free(x);
39			* \endcode
40			*/
41
42			#ifndef INCLUDED_volk_16i_x4_quad_max_star_16i_a_H
43			#define INCLUDED_volk_16i_x4_quad_max_star_16i_a_H
44
45			#include <inttypes.h>
46			#include <stdio.h>
47
48			#ifdef LV_HAVE_SSE2
49
50			#include <emmintrin.h>
51
52		✗	static inline void volk_16i_x4_quad_max_star_16i_a_sse2(short* target,
53			short* src0,
54			short* src1,
55			short* src2,
56			short* src3,
57			unsigned int num_points)
58			{
59		✗	const unsigned int num_bytes = num_points * 2;
60
61		✗	int i = 0;
62
63		✗	int bound = (num_bytes >> 4);
64		✗	int bound_copy = bound;
65		✗	int leftovers = (num_bytes >> 1) & 7;
66
67			__m128i *p_target, *p_src0, *p_src1, *p_src2, *p_src3;
68		✗	p_target = (__m128i*)target;
69		✗	p_src0 = (__m128i*)src0;
70		✗	p_src1 = (__m128i*)src1;
71		✗	p_src2 = (__m128i*)src2;
72		✗	p_src3 = (__m128i*)src3;
73
74			__m128i xmm1, xmm2, xmm3, xmm4, xmm5, xmm6, xmm7, xmm8;
75
76		✗	while (bound_copy > 0) {
77		✗	xmm1 = _mm_load_si128(p_src0);
78		✗	xmm2 = _mm_load_si128(p_src1);
79		✗	xmm3 = _mm_load_si128(p_src2);
80		✗	xmm4 = _mm_load_si128(p_src3);
81
82		✗	xmm5 = _mm_setzero_si128();
83		✗	xmm6 = _mm_setzero_si128();
84		✗	xmm7 = xmm1;
85		✗	xmm8 = xmm3;
86
87		✗	xmm1 = _mm_sub_epi16(xmm2, xmm1);
88
89		✗	xmm3 = _mm_sub_epi16(xmm4, xmm3);
90
91		✗	xmm5 = _mm_cmpgt_epi16(xmm1, xmm5);
92		✗	xmm6 = _mm_cmpgt_epi16(xmm3, xmm6);
93
94		✗	xmm2 = _mm_and_si128(xmm5, xmm2);
95		✗	xmm4 = _mm_and_si128(xmm6, xmm4);
96		✗	xmm5 = _mm_andnot_si128(xmm5, xmm7);
97		✗	xmm6 = _mm_andnot_si128(xmm6, xmm8);
98
99		✗	xmm5 = _mm_add_epi16(xmm2, xmm5);
100		✗	xmm6 = _mm_add_epi16(xmm4, xmm6);
101
102		✗	xmm1 = _mm_xor_si128(xmm1, xmm1);
103		✗	xmm2 = xmm5;
104		✗	xmm5 = _mm_sub_epi16(xmm6, xmm5);
105		✗	p_src0 += 1;
106		✗	bound_copy -= 1;
107
108		✗	xmm1 = _mm_cmpgt_epi16(xmm5, xmm1);
109		✗	p_src1 += 1;
110
111		✗	xmm6 = _mm_and_si128(xmm1, xmm6);
112
113		✗	xmm1 = _mm_andnot_si128(xmm1, xmm2);
114		✗	p_src2 += 1;
115
116		✗	xmm1 = _mm_add_epi16(xmm6, xmm1);
117		✗	p_src3 += 1;
118
119			_mm_store_si128(p_target, xmm1);
120		✗	p_target += 1;
121			}
122
123
124			/*__VOLK_ASM __VOLK_VOLATILE
125			(
126			"volk_16i_x4_quad_max_star_16i_a_sse2_L1:\n\t"
127			"cmp $0, %[bound]\n\t"
128			"je volk_16i_x4_quad_max_star_16i_a_sse2_END\n\t"
129
130			"movaps (%[src0]), %%xmm1\n\t"
131			"movaps (%[src1]), %%xmm2\n\t"
132			"movaps (%[src2]), %%xmm3\n\t"
133			"movaps (%[src3]), %%xmm4\n\t"
134
135			"pxor %%xmm5, %%xmm5\n\t"
136			"pxor %%xmm6, %%xmm6\n\t"
137			"movaps %%xmm1, %%xmm7\n\t"
138			"movaps %%xmm3, %%xmm8\n\t"
139			"psubw %%xmm2, %%xmm1\n\t"
140			"psubw %%xmm4, %%xmm3\n\t"
141
142			"pcmpgtw %%xmm1, %%xmm5\n\t"
143			"pcmpgtw %%xmm3, %%xmm6\n\t"
144
145			"pand %%xmm5, %%xmm2\n\t"
146			"pand %%xmm6, %%xmm4\n\t"
147			"pandn %%xmm7, %%xmm5\n\t"
148			"pandn %%xmm8, %%xmm6\n\t"
149
150			"paddw %%xmm2, %%xmm5\n\t"
151			"paddw %%xmm4, %%xmm6\n\t"
152
153			"pxor %%xmm1, %%xmm1\n\t"
154			"movaps %%xmm5, %%xmm2\n\t"
155
156			"psubw %%xmm6, %%xmm5\n\t"
157			"add $16, %[src0]\n\t"
158			"add $-1, %[bound]\n\t"
159
160			"pcmpgtw %%xmm5, %%xmm1\n\t"
161			"add $16, %[src1]\n\t"
162
163			"pand %%xmm1, %%xmm6\n\t"
164
165			"pandn %%xmm2, %%xmm1\n\t"
166			"add $16, %[src2]\n\t"
167
168			"paddw %%xmm6, %%xmm1\n\t"
169			"add $16, %[src3]\n\t"
170
171			"movaps %%xmm1, (%[target])\n\t"
172			"addw $16, %[target]\n\t"
173			"jmp volk_16i_x4_quad_max_star_16i_a_sse2_L1\n\t"
174
175			"volk_16i_x4_quad_max_star_16i_a_sse2_END:\n\t"
176			:
177			:[bound]"r"(bound), [src0]"r"(src0), [src1]"r"(src1), [src2]"r"(src2),
178			[src3]"r"(src3), [target]"r"(target)
179			:
180			);
181			*/
182
183		✗	short temp0 = 0;
184		✗	short temp1 = 0;
185		✗	for (i = bound * 8; i < (bound * 8) + leftovers; ++i) {
186		✗	temp0 = ((short)(src0[i] - src1[i]) > 0) ? src0[i] : src1[i];
187		✗	temp1 = ((short)(src2[i] - src3[i]) > 0) ? src2[i] : src3[i];
188		✗	target[i] = ((short)(temp0 - temp1) > 0) ? temp0 : temp1;
189			}
190		✗	return;
191			}
192
193			#endif /*LV_HAVE_SSE2*/
194
195			#ifdef LV_HAVE_NEON
196
197			#include <arm_neon.h>
198
199			static inline void volk_16i_x4_quad_max_star_16i_neon(short* target,
200			short* src0,
201			short* src1,
202			short* src2,
203			short* src3,
204			unsigned int num_points)
205			{
206			const unsigned int eighth_points = num_points / 8;
207			unsigned i;
208
209			int16x8_t src0_vec, src1_vec, src2_vec, src3_vec;
210			int16x8_t diff12, diff34;
211			int16x8_t comp0, comp1, comp2, comp3;
212			int16x8_t result1_vec, result2_vec;
213			int16x8_t zeros;
214			zeros = vdupq_n_s16(0);
215			for (i = 0; i < eighth_points; ++i) {
216			src0_vec = vld1q_s16(src0);
217			src1_vec = vld1q_s16(src1);
218			src2_vec = vld1q_s16(src2);
219			src3_vec = vld1q_s16(src3);
220			diff12 = vsubq_s16(src0_vec, src1_vec);
221			diff34 = vsubq_s16(src2_vec, src3_vec);
222			comp0 = (int16x8_t)vcgeq_s16(diff12, zeros);
223			comp1 = (int16x8_t)vcltq_s16(diff12, zeros);
224			comp2 = (int16x8_t)vcgeq_s16(diff34, zeros);
225			comp3 = (int16x8_t)vcltq_s16(diff34, zeros);
226			comp0 = vandq_s16(src0_vec, comp0);
227			comp1 = vandq_s16(src1_vec, comp1);
228			comp2 = vandq_s16(src2_vec, comp2);
229			comp3 = vandq_s16(src3_vec, comp3);
230
231			result1_vec = vaddq_s16(comp0, comp1);
232			result2_vec = vaddq_s16(comp2, comp3);
233
234			diff12 = vsubq_s16(result1_vec, result2_vec);
235			comp0 = (int16x8_t)vcgeq_s16(diff12, zeros);
236			comp1 = (int16x8_t)vcltq_s16(diff12, zeros);
237			comp0 = vandq_s16(result1_vec, comp0);
238			comp1 = vandq_s16(result2_vec, comp1);
239			result1_vec = vaddq_s16(comp0, comp1);
240			vst1q_s16(target, result1_vec);
241			src0 += 8;
242			src1 += 8;
243			src2 += 8;
244			src3 += 8;
245			target += 8;
246			}
247
248			short temp0 = 0;
249			short temp1 = 0;
250			for (i = eighth_points * 8; i < num_points; ++i) {
251			temp0 = ((short)(*src0 - *src1) > 0) ? *src0 : *src1;
252			temp1 = ((short)(*src2 - *src3) > 0) ? *src2 : *src3;
253			*target++ = ((short)(temp0 - temp1) > 0) ? temp0 : temp1;
254			src0++;
255			src1++;
256			src2++;
257			src3++;
258			}
259			}
260			#endif /* LV_HAVE_NEON */
261
262
263			#ifdef LV_HAVE_GENERIC
264		✗	static inline void volk_16i_x4_quad_max_star_16i_generic(short* target,
265			short* src0,
266			short* src1,
267			short* src2,
268			short* src3,
269			unsigned int num_points)
270			{
271		✗	const unsigned int num_bytes = num_points * 2;
272
273		✗	int i = 0;
274
275		✗	int bound = num_bytes >> 1;
276
277		✗	short temp0 = 0;
278		✗	short temp1 = 0;
279		✗	for (i = 0; i < bound; ++i) {
280		✗	temp0 = ((short)(src0[i] - src1[i]) > 0) ? src0[i] : src1[i];
281		✗	temp1 = ((short)(src2[i] - src3[i]) > 0) ? src2[i] : src3[i];
282		✗	target[i] = ((short)(temp0 - temp1) > 0) ? temp0 : temp1;
283			}
284		✗	}
285
286			#endif /*LV_HAVE_GENERIC*/
287
288			#endif /*INCLUDED_volk_16i_x4_quad_max_star_16i_a_H*/
289