hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-30 20:54:07 +00:00
Code Issues Actions Projects Releases Wiki Activity

Added case 12 to bipred_recon no mov

Browse source
This commit is contained in:
Reima Hyvönen 2018-11-29 09:48:22 +02:00
parent f8696b54a4
commit e4a10880f3
1 changed files with 153 additions and 54 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

117
src/strategies/avx2/picture-avx2.c
View file

@ -729,8 +729,6 @@ static void inter_recon_bipred_no_mov_avx2(
kvz_pixel* temp_lcu_v) {
// This function is used only when kvazaar can't find any movement from the current block
int y_in_lcu, x_in_lcu;
__m256i sample0_epi8, sample1_epi8, temp_y_epi8;
int32_t * pointer = 0;
@ -769,6 +767,27 @@ static void inter_recon_bipred_no_mov_avx2(
break;
case 12:
sample0_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
temp_y_epi8 = _mm256_avg_epu8(sample0_epi8, sample1_epi8);
// Store 64-bits from vector to memory
_mm_storel_epi64((__m128i*)&(lcu->rec.y[(y_in_lcu)* LCU_WIDTH + x_in_lcu]), _mm256_castsi256_si128(temp_y_epi8));
x_in_lcu = ((xpos + temp_x + 8) & ((LCU_WIDTH)-1));
sample0_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
temp_y_epi8 = _mm256_avg_epu8(sample0_epi8, sample1_epi8);
pointer = (int32_t*)&(lcu->rec.y[(y_in_lcu)* LCU_WIDTH + x_in_lcu]);
*pointer = _mm_cvtsi128_si32(_mm256_castsi256_si128(temp_y_epi8));
break;
case 16:
sample0_epi8 = _mm256_castsi128_si256(_mm_loadu_si128((__m128i*)&temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
@ -781,7 +800,7 @@ static void inter_recon_bipred_no_mov_avx2(
break;
default:
case 32:
sample0_epi8 = _mm256_loadu_si256((__m256i*) &(temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
sample1_epi8 = _mm256_loadu_si256((__m256i*) &(lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu]));
@ -793,6 +812,18 @@ static void inter_recon_bipred_no_mov_avx2(
_mm256_storeu_si256((__m256i*)&(lcu->rec.y[(y_in_lcu)* LCU_WIDTH + x_in_lcu]), temp_y_epi8);
break;
default:
// If width is something strange size, use this
for (int temp_i = 0; temp_i < width; ++temp_i) {
x_in_lcu = ((xpos + temp_i) & ((LCU_WIDTH)-1));
int sample0_y = (temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu] << (14 - KVZ_BIT_DEPTH));
int sample1_y = (lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu] << (14 - KVZ_BIT_DEPTH));
lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu] = (kvz_pixel)kvz_fast_clip_32bit_to_pixel((sample0_y + sample1_y) >> 1);
}
}
if (temp_x < width >> 1 && temp_y < height >> 1) {
@ -825,6 +856,36 @@ static void inter_recon_bipred_no_mov_avx2(
break;
case 12:
sample0_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&temp_lcu_u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&lcu->rec.u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
temp_u_epi8 = _mm256_avg_epu8(sample0_epi8, sample1_epi8);
sample0_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&temp_lcu_v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample1_epi8 = _mm256_castsi128_si256(_mm_cvtsi32_si128(*(int32_t*)&lcu->rec.v[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
temp_v_epi8 = _mm256_avg_epu8(sample0_epi8, sample1_epi8);
pointer = (int32_t*)&(lcu->rec.u[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu]);
*pointer = _mm_cvtsi128_si32(_mm256_castsi256_si128(temp_u_epi8));
pointer = (int32_t*)&(lcu->rec.v[(y_in_lcu)* LCU_WIDTH_C + x_in_lcu]);
*pointer = _mm_cvtsi128_si32(_mm256_castsi256_si128(temp_v_epi8));
// This is used only with odd shaped objects
for (int temp_i = 4; temp_i < width >> 1; ++temp_i) {
int temp_x_in_lcu = (((xpos >> 1) + temp_i) & (LCU_WIDTH_C - 1));
int16_t sample0_u = (temp_lcu_u[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] << (14 - KVZ_BIT_DEPTH));
int16_t sample1_u = (lcu->rec.u[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] << (14 - KVZ_BIT_DEPTH));
lcu->rec.u[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] = (kvz_pixel)kvz_fast_clip_32bit_to_pixel((sample0_u + sample1_u) >> 1);
int16_t sample0_v = (temp_lcu_v[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] << (14 - KVZ_BIT_DEPTH));
int16_t sample1_v = (lcu->rec.v[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] << (14 - KVZ_BIT_DEPTH));
lcu->rec.v[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] = (kvz_pixel)kvz_fast_clip_32bit_to_pixel((sample0_v + sample1_v) >> 1);
}
break;
case 16:
sample0_epi8 = _mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)&temp_lcu_u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
@ -862,7 +923,7 @@ static void inter_recon_bipred_no_mov_avx2(
break;
default:
case 64:
sample0_epi8 = _mm256_loadu_si256((__m256i*) &(temp_lcu_u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
sample1_epi8 = _mm256_loadu_si256((__m256i*) &(lcu->rec.u[y_in_lcu * LCU_WIDTH_C + x_in_lcu]));
@ -876,6 +937,20 @@ static void inter_recon_bipred_no_mov_avx2(
_mm256_storeu_si256((__m256i*)&(lcu->rec.y[(y_in_lcu)* LCU_WIDTH + x_in_lcu]), temp_v_epi8);
break;
default:
// This is used only with odd shaped objects
for (int temp_i = 0; temp_i < width >> 1; ++temp_i) {
int temp_x_in_lcu = (((xpos >> 1) + temp_i) & (LCU_WIDTH_C - 1));
int16_t sample0_u = (temp_lcu_u[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] << (14 - KVZ_BIT_DEPTH));
int16_t sample1_u = (lcu->rec.u[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] << (14 - KVZ_BIT_DEPTH));
lcu->rec.u[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] = (kvz_pixel)kvz_fast_clip_32bit_to_pixel((sample0_u + sample1_u) >> 1);
int16_t sample0_v = (temp_lcu_v[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] << (14 - KVZ_BIT_DEPTH));
int16_t sample1_v = (lcu->rec.v[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] << (14 - KVZ_BIT_DEPTH));
lcu->rec.v[y_in_lcu * LCU_WIDTH_C + temp_x_in_lcu] = (kvz_pixel)kvz_fast_clip_32bit_to_pixel((sample0_v + sample1_v) >> 1);
}
break;
}
y_in_lcu = ((ypos + temp_y) & ((LCU_WIDTH)-1));
@ -929,11 +1004,26 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
{
case false:
if (width < 4) {
// If width is smaller than 4 there's no need to use SIMD
for (int temp_i = 0; temp_i < width; ++temp_i) {
x_in_lcu = ((xpos + temp_i) & ((LCU_WIDTH)-1));
int sample0_y = (hi_prec_luma_rec0 ? high_precision_rec0->y[y_in_lcu * LCU_WIDTH + x_in_lcu] : (temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu] << (14 - KVZ_BIT_DEPTH)));
int sample1_y = (hi_prec_luma_rec1 ? high_precision_rec1->y[y_in_lcu * LCU_WIDTH + x_in_lcu] : (lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu] << (14 - KVZ_BIT_DEPTH)));
lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu] = (kvz_pixel)kvz_fast_clip_32bit_to_pixel((sample0_y + sample1_y + offset) >> shift);
}
}
else{
// Load total of 4 elements from memory to vector
sample0_epi32 = hi_prec_luma_rec0 ? _mm256_cvtepi16_epi32(_mm_loadl_epi64((__m128i*) &(high_precision_rec0->y[y_in_lcu * LCU_WIDTH + x_in_lcu]))) :
_mm256_slli_epi32(_mm256_cvtepu8_epi32(_mm_cvtsi32_si128(*(int32_t*)&(temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu]))), 14 - KVZ_BIT_DEPTH);
sample1_epi32 = hi_prec_luma_rec1 ? _mm256_cvtepu16_epi32(_mm_loadl_epi64((__m128i*) &(high_precision_rec1->y[y_in_lcu * LCU_WIDTH + x_in_lcu]))) :
sample1_epi32 = hi_prec_luma_rec1 ? _mm256_cvtepi16_epi32(_mm_loadl_epi64((__m128i*) &(high_precision_rec1->y[y_in_lcu * LCU_WIDTH + x_in_lcu]))) :
_mm256_slli_epi32(_mm256_cvtepu8_epi32(_mm_cvtsi32_si128(*(int32_t*) &(lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu]))), 14 - KVZ_BIT_DEPTH);
@ -947,13 +1037,24 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
temp_epi16 = _mm256_permute4x64_epi64(temp_epi16, _MM_SHUFFLE(3, 1, 2, 0));
temp_epi8 = _mm256_packus_epi16(temp_epi16, temp_epi16);
pointer = (int32_t*)&(lcu->rec.y[(y_in_lcu)* LCU_WIDTH + x_in_lcu]);
*pointer = _mm_cvtsi128_si32(_mm256_castsi256_si128(temp_epi8));
for (int temp_i = temp_x + 4; temp_i < width; ++temp_i) {
x_in_lcu = ((xpos + temp_i) & ((LCU_WIDTH)-1));
int16_t sample0_y = (hi_prec_luma_rec0 ? high_precision_rec0->y[y_in_lcu * LCU_WIDTH + x_in_lcu] : (temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu] << (14 - KVZ_BIT_DEPTH)));
int16_t sample1_y = (hi_prec_luma_rec1 ? high_precision_rec1->y[y_in_lcu * LCU_WIDTH + x_in_lcu] : (lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu] << (14 - KVZ_BIT_DEPTH)));
lcu->rec.y[y_in_lcu * LCU_WIDTH + x_in_lcu] = (kvz_pixel)kvz_fast_clip_32bit_to_pixel((sample0_y + sample1_y + offset) >> shift);
}
}
break;
default:
// Load total of 8 elements from memory to vector
sample0_epi32 = hi_prec_luma_rec0 ? _mm256_cvtepi16_epi32(_mm_loadu_si128((__m128i*) &(high_precision_rec0->y[y_in_lcu * LCU_WIDTH + x_in_lcu]))) :
_mm256_slli_epi32(_mm256_cvtepu8_epi32((_mm_loadl_epi64((__m128i*) &(temp_lcu_y[y_in_lcu * LCU_WIDTH + x_in_lcu])))), 14 - KVZ_BIT_DEPTH);
@ -971,7 +1072,6 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
temp_epi16 = _mm256_permute4x64_epi64(temp_epi16, _MM_SHUFFLE(3, 1, 2, 0));
temp_epi8 = _mm256_packus_epi16(temp_epi16, temp_epi16);
// Store 64-bits from vector to memory
_mm_storel_epi64((__m128i*)&(lcu->rec.y[(y_in_lcu)* LCU_WIDTH + x_in_lcu]), _mm256_castsi256_si128(temp_epi8));
@ -981,7 +1081,6 @@ static void inter_recon_bipred_avx2(const int hi_prec_luma_rec0,
}
}
for (int temp_y = 0; temp_y < height >> 1; ++temp_y) {
int y_in_lcu = (((ypos >> 1) + temp_y) & (LCU_WIDTH_C - 1));