generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Remove vpx_filter_block1d16_v8_intrin_ssse3

Browse Source 
This was rewritten and moved to vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm
in 195883023bb39b5ee5c6811a316ab96d9225034d

Change-Id: I117ce983dae12006e302679ba7f175573dd9e874
This commit is contained in:
Johann 2015-09-18 15:59:12 -07:00
parent cd82f69823
commit dd4f953350
1 changed files with 0 additions and 117 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

117
vpx_dsp/x86/vpx_subpixel_8t_intrin_ssse3.c
View File

@ -291,123 +291,6 @@ void vpx_filter_block1d8_v8_intrin_ssse3(const uint8_t *src_ptr,
}
}
#if ARCH_X86_64
static void vpx_filter_block1d16_v8_intrin_ssse3(const uint8_t *src_ptr,
ptrdiff_t src_pitch,
uint8_t *output_ptr,
ptrdiff_t out_pitch,
uint32_t output_height,
const int16_t *filter) {
__m128i addFilterReg64, filtersReg, srcRegFilt1, srcRegFilt3;
__m128i firstFilters, secondFilters, thirdFilters, forthFilters;
__m128i srcRegFilt5, srcRegFilt6, srcRegFilt7, srcRegFilt8;
__m128i srcReg1, srcReg2, srcReg3, srcReg4, srcReg5, srcReg6, srcReg7;
__m128i srcReg8;
unsigned int i;
// create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
addFilterReg64 = _mm_set1_epi32((int)0x0400040u);
filtersReg = _mm_loadu_si128((const __m128i *)filter);
// converting the 16 bit (short) to 8 bit (byte) and have the same data
// in both lanes of 128 bit register.
filtersReg =_mm_packs_epi16(filtersReg, filtersReg);
// duplicate only the first 16 bits in the filter
firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x100u));
// duplicate only the second 16 bits in the filter
secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
// duplicate only the third 16 bits in the filter
thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
// duplicate only the forth 16 bits in the filter
forthFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x706u));
// load the first 7 rows of 16 bytes
srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr));
srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch));
srcReg3 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
srcReg4 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
srcReg5 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
srcReg6 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
srcReg7 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
for (i = 0; i < output_height; i++) {
// load the last 16 bytes
srcReg8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
// merge the result together
srcRegFilt5 = _mm_unpacklo_epi8(srcReg1, srcReg2);
srcRegFilt6 = _mm_unpacklo_epi8(srcReg7, srcReg8);
srcRegFilt1 = _mm_unpackhi_epi8(srcReg1, srcReg2);
srcRegFilt3 = _mm_unpackhi_epi8(srcReg7, srcReg8);
// multiply 2 adjacent elements with the filter and add the result
srcRegFilt5 = _mm_maddubs_epi16(srcRegFilt5, firstFilters);
srcRegFilt6 = _mm_maddubs_epi16(srcRegFilt6, forthFilters);
srcRegFilt1 = _mm_maddubs_epi16(srcRegFilt1, firstFilters);
srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, forthFilters);
// add and saturate the results together
srcRegFilt5 = _mm_adds_epi16(srcRegFilt5, srcRegFilt6);
srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt3);
// merge the result together
srcRegFilt3 = _mm_unpacklo_epi8(srcReg3, srcReg4);
srcRegFilt6 = _mm_unpackhi_epi8(srcReg3, srcReg4);
// multiply 2 adjacent elements with the filter and add the result
srcRegFilt3 = _mm_maddubs_epi16(srcRegFilt3, secondFilters);
srcRegFilt6 = _mm_maddubs_epi16(srcRegFilt6, secondFilters);
// merge the result together
srcRegFilt7 = _mm_unpacklo_epi8(srcReg5, srcReg6);
srcRegFilt8 = _mm_unpackhi_epi8(srcReg5, srcReg6);
// multiply 2 adjacent elements with the filter and add the result
srcRegFilt7 = _mm_maddubs_epi16(srcRegFilt7, thirdFilters);
srcRegFilt8 = _mm_maddubs_epi16(srcRegFilt8, thirdFilters);
// add and saturate the results together
srcRegFilt5 = _mm_adds_epi16(srcRegFilt5,
_mm_min_epi16(srcRegFilt3, srcRegFilt7));
srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
_mm_min_epi16(srcRegFilt6, srcRegFilt8));
// add and saturate the results together
srcRegFilt5 = _mm_adds_epi16(srcRegFilt5,
_mm_max_epi16(srcRegFilt3, srcRegFilt7));
srcRegFilt1 = _mm_adds_epi16(srcRegFilt1,
_mm_max_epi16(srcRegFilt6, srcRegFilt8));
srcRegFilt5 = _mm_adds_epi16(srcRegFilt5, addFilterReg64);
srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, addFilterReg64);
// shift by 7 bit each 16 bit
srcRegFilt5 = _mm_srai_epi16(srcRegFilt5, 7);
srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
// shrink to 8 bit each 16 bits, the first lane contain the first
// convolve result and the second lane contain the second convolve
// result
srcRegFilt1 = _mm_packus_epi16(srcRegFilt5, srcRegFilt1);
src_ptr+=src_pitch;
// shift down a row
srcReg1 = srcReg2;
srcReg2 = srcReg3;
srcReg3 = srcReg4;
srcReg4 = srcReg5;
srcReg5 = srcReg6;
srcReg6 = srcReg7;
srcReg7 = srcReg8;
// save 16 bytes convolve result
_mm_store_si128((__m128i*)output_ptr, srcRegFilt1);
output_ptr+=out_pitch;
}
}
#endif // ARCH_X86_64
filter8_1dfunction vpx_filter_block1d16_v8_ssse3;
filter8_1dfunction vpx_filter_block1d16_h8_ssse3;
filter8_1dfunction vpx_filter_block1d8_v8_ssse3;