|
111
|
1 /* { dg-do compile { target i?86-*-* x86_64-*-* } } */
|
|
|
2 /* { dg-options "-O1 -msse2" } */
|
|
|
3
|
|
|
4 #include <xmmintrin.h>
|
|
|
5
|
|
|
6 void ConvertFloatSSE (void *inBuff, void *outBuff, int len)
|
|
|
7 {
|
|
|
8 unsigned char *inByteBuffer = reinterpret_cast < unsigned char *>(inBuff);
|
|
|
9 float *outFloatBuffer = reinterpret_cast < float *>(outBuff);
|
|
|
10
|
|
|
11 for (int i = 0; i < len / 4; i++)
|
|
|
12 {
|
|
|
13 __m128i register1 =
|
|
|
14 _mm_set_epi8 (*inByteBuffer, *(inByteBuffer + 1),
|
|
|
15 *(inByteBuffer + 2), 0,
|
|
|
16 *(inByteBuffer + 3), *(inByteBuffer + 4),
|
|
|
17 *(inByteBuffer + 5), 0,
|
|
|
18 *(inByteBuffer + 6), *(inByteBuffer + 7),
|
|
|
19 *(inByteBuffer + 8), 0,
|
|
|
20 *(inByteBuffer + 9), *(inByteBuffer + 10),
|
|
|
21 *(inByteBuffer + 11), 0);
|
|
|
22 __m128i register2 = _mm_srai_epi32 (register1, 8);
|
|
|
23 __m128 register3 = _mm_cvtepi32_ps (register2);
|
|
|
24
|
|
|
25 _mm_store_ps (outFloatBuffer, register3);
|
|
|
26 outFloatBuffer += 4;
|
|
|
27 inByteBuffer += 12;
|
|
|
28 }
|
|
|
29 }
|
