static void aom_filter_block1d4_h4_ssse3(
const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr,
ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) {
__m128i filtersReg;
__m128i addFilterReg32, filt1Reg, firstFilters, srcReg32b1, srcRegFilt32b1_1;
unsigned int i;
src_ptr -= 3;
addFilterReg32 = _mm_set1_epi16(32);
filtersReg = _mm_loadu_si128((const __m128i *)filter);
filtersReg = _mm_srai_epi16(filtersReg, 1);
// 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);
firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi32(0x5040302u));
filt1Reg = _mm_load_si128((__m128i const *)(filtd4));
for (i = output_height; i > 0; i -= 1) {
// load the 2 strides of source
srcReg32b1 = _mm_loadu_si128((const __m128i *)src_ptr);
// filter the source buffer
srcRegFilt32b1_1 = _mm_shuffle_epi8(srcReg32b1, filt1Reg);
// multiply 4 adjacent elements with the filter and add the result
srcRegFilt32b1_1 = _mm_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
srcRegFilt32b1_1 = _mm_hadds_epi16(srcRegFilt32b1_1, _mm_setzero_si128());
// shift by 6 bit each 16 bit
srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
srcRegFilt32b1_1 = _mm_srai_epi16(srcRegFilt32b1_1, 6);
// shrink to 8 bit each 16 bits, the first lane contain the first
// convolve result and the second lane contain the second convolve result
srcRegFilt32b1_1 = _mm_packus_epi16(srcRegFilt32b1_1, _mm_setzero_si128());
src_ptr += src_pixels_per_line;
*((uint32_t *)(output_ptr)) = _mm_cvtsi128_si32(srcRegFilt32b1_1);
output_ptr += output_pitch;
}
}
__m128i test_mm_hadds_epi16(__m128i a, __m128i b) {
// CHECK-LABEL: test_mm_hadds_epi16
// CHECK: call <8 x i16> @llvm.x86.ssse3.phadd.sw.128(<8 x i16> %{{.*}}, <8 x i16> %{{.*}})
return _mm_hadds_epi16(a, b);
}
static void aom_filter_block1d4_v4_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 filtersReg;
__m128i addFilterReg32;
__m128i srcReg2, srcReg3, srcReg23, srcReg4, srcReg34, srcReg5, srcReg45,
srcReg6, srcReg56;
__m128i srcReg23_34_lo, srcReg45_56_lo;
__m128i srcReg2345_3456_lo, srcReg2345_3456_hi;
__m128i resReglo, resReghi;
__m128i firstFilters;
unsigned int i;
ptrdiff_t src_stride, dst_stride;
addFilterReg32 = _mm_set1_epi16(32);
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_srai_epi16(filtersReg, 1);
filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi32(0x5040302u));
// multiple the size of the source and destination stride by two
src_stride = src_pitch << 1;
dst_stride = out_pitch << 1;
srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
srcReg23 = _mm_unpacklo_epi32(srcReg2, srcReg3);
srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
// have consecutive loads on the same 256 register
srcReg34 = _mm_unpacklo_epi32(srcReg3, srcReg4);
srcReg23_34_lo = _mm_unpacklo_epi8(srcReg23, srcReg34);
for (i = output_height; i > 1; i -= 2) {
srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
srcReg45 = _mm_unpacklo_epi32(srcReg4, srcReg5);
srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
srcReg56 = _mm_unpacklo_epi32(srcReg5, srcReg6);
// merge every two consecutive registers
srcReg45_56_lo = _mm_unpacklo_epi8(srcReg45, srcReg56);
srcReg2345_3456_lo = _mm_unpacklo_epi16(srcReg23_34_lo, srcReg45_56_lo);
srcReg2345_3456_hi = _mm_unpackhi_epi16(srcReg23_34_lo, srcReg45_56_lo);
// multiply 2 adjacent elements with the filter and add the result
resReglo = _mm_maddubs_epi16(srcReg2345_3456_lo, firstFilters);
resReghi = _mm_maddubs_epi16(srcReg2345_3456_hi, firstFilters);
resReglo = _mm_hadds_epi16(resReglo, _mm_setzero_si128());
resReghi = _mm_hadds_epi16(resReghi, _mm_setzero_si128());
// shift by 6 bit each 16 bit
resReglo = _mm_adds_epi16(resReglo, addFilterReg32);
resReghi = _mm_adds_epi16(resReghi, addFilterReg32);
resReglo = _mm_srai_epi16(resReglo, 6);
resReghi = _mm_srai_epi16(resReghi, 6);
// shrink to 8 bit each 16 bits, the first lane contain the first
// convolve result and the second lane contain the second convolve
// result
resReglo = _mm_packus_epi16(resReglo, resReglo);
resReghi = _mm_packus_epi16(resReghi, resReghi);
src_ptr += src_stride;
*((uint32_t *)(output_ptr)) = _mm_cvtsi128_si32(resReglo);
*((uint32_t *)(output_ptr + out_pitch)) = _mm_cvtsi128_si32(resReghi);
output_ptr += dst_stride;
// save part of the registers for next strides
srcReg23_34_lo = srcReg45_56_lo;
srcReg4 = srcReg6;
}
}