hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-23 18:14:06 +00:00
Code Issues Actions Projects Releases Wiki Activity

[avx2] AVX2 version of depquant now exactly matches scalar version

Browse source
This commit is contained in:
Joose Sainio 2023-04-22 11:53:54 +03:00
parent 49a36dd455
commit 473b1b8bb1
3 changed files with 14 additions and 10 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/dep_quant.c
View file

@ -664,7 +664,7 @@ void uvg_dep_quant_update_state_eos(
memset(state->m_absLevelsAndCtxInit[curr_state_offset], 0, 16 * sizeof(uint8_t)); memset(state->m_absLevelsAndCtxInit[curr_state_offset], 0, 16 * sizeof(uint8_t));
} }
uint8_t* temp = (uint8_t*)(&state->m_absLevelsAndCtxInit[curr_state_offset][scan_pos & 15]); uint8_t* temp = (uint8_t*)(&state->m_absLevelsAndCtxInit[curr_state_offset][scan_pos & 15]);
*temp = (uint8_t)MIN(32, decisions->absLevel[decision_id]); *temp = (uint8_t)MIN(51, decisions->absLevel[decision_id]);
update_common_context(ctxs, state->m_commonCtx, scan_pos, cg_pos, width_in_sbb, height_in_sbb, next_sbb_right, update_common_context(ctxs, state->m_commonCtx, scan_pos, cg_pos, width_in_sbb, height_in_sbb, next_sbb_right,
next_sbb_below, prvState, ctxs->m_curr_state_offset + decision_id); next_sbb_below, prvState, ctxs->m_curr_state_offset + decision_id);

20
src/strategies/avx2/depquant-avx2.c
View file

@ -35,6 +35,7 @@
*/ */
#include "strategies/avx2/depquant-avx2.h" #include "strategies/avx2/depquant-avx2.h"
#include "strategyselector.h"
#if COMPILE_INTEL_AVX2 && defined X86_64 #if COMPILE_INTEL_AVX2 && defined X86_64
#include "dep_quant.h" #include "dep_quant.h"
@ -352,13 +353,13 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
__m256i cheaper_first = _mm256_blendv_epi8(rd_cost_a, rd_cost_b, a_vs_b); __m256i cheaper_first = _mm256_blendv_epi8(rd_cost_a, rd_cost_b, a_vs_b);
__m256i cheaper_first_data = _mm256_blendv_epi8(a_data, b_data, a_vs_b); __m256i cheaper_first_data = _mm256_blendv_epi8(a_data, b_data, a_vs_b);
__m256i z_vs_decision = _mm256_cmpgt_epi64(rd_cost_z, rd_cost_decision); __m256i z_vs_decision = _mm256_cmpgt_epi64(rd_cost_decision, rd_cost_z);
__m256i cheaper_second = _mm256_blendv_epi8(rd_cost_z, rd_cost_decision, z_vs_decision); __m256i cheaper_second = _mm256_blendv_epi8(rd_cost_decision, rd_cost_z, z_vs_decision);
__m256i cheaper_second_data = _mm256_blendv_epi8(z_data, decision_data, z_vs_decision); __m256i cheaper_second_data = _mm256_blendv_epi8(decision_data, z_data, z_vs_decision);
__m256i final_decision = _mm256_cmpgt_epi64(cheaper_first, cheaper_second); __m256i final_decision = _mm256_cmpgt_epi64(cheaper_second, cheaper_first);
__m256i final_rd_cost = _mm256_blendv_epi8(cheaper_first, cheaper_second, final_decision); __m256i final_rd_cost = _mm256_blendv_epi8(cheaper_second, cheaper_first, final_decision);
__m256i final_data = _mm256_blendv_epi8(cheaper_first_data, cheaper_second_data, final_decision); __m256i final_data = _mm256_blendv_epi8(cheaper_second_data, cheaper_first_data, final_decision);
_mm256_store_si256((__m256i*)decisions->rdCost, final_rd_cost); _mm256_store_si256((__m256i*)decisions->rdCost, final_rd_cost);
final_data = _mm256_permutevar8x32_epi32(final_data, _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0)); final_data = _mm256_permutevar8x32_epi32(final_data, _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0));
@ -952,7 +953,7 @@ static INLINE void update_states_avx2(
} }
} }
uint32_t level_offset = scan_pos & 15; uint32_t level_offset = scan_pos & 15;
__m128i max_abs = _mm_min_epi32(abs_level, _mm_set1_epi32(32)); __m128i max_abs = _mm_min_epi32(abs_level, _mm_set1_epi32(51));
uint32_t max_abs_s[4]; uint32_t max_abs_s[4];
_mm_storeu_si128((__m128i*)max_abs_s, max_abs); _mm_storeu_si128((__m128i*)max_abs_s, max_abs);
for (int i = 0; i < 4; ++i) { for (int i = 0; i < 4; ++i) {
@ -1094,7 +1095,7 @@ static INLINE void update_states_avx2(
} }
__m128i sum_abs = _mm_srli_epi32(tinit, 8); __m128i sum_abs = _mm_srli_epi32(tinit, 8);
sum_abs = _mm_min_epi32(sum_abs, _mm_set1_epi32(32)); sum_abs = _mm_min_epi32(sum_abs, _mm_set1_epi32(51));
switch (numIPos) { switch (numIPos) {
case 5: case 5:
{ {
@ -1103,6 +1104,9 @@ static INLINE void update_states_avx2(
_mm_add_epi32(levels_start_offsets, _mm_set1_epi32(next_nb_info_ssb.inPos[4])), _mm_add_epi32(levels_start_offsets, _mm_set1_epi32(next_nb_info_ssb.inPos[4])),
1); 1);
sum_abs = _mm_add_epi32(t, sum_abs); sum_abs = _mm_add_epi32(t, sum_abs);
// Need this to make sure we don't go beyond 255
sum_abs = _mm_and_si128(sum_abs, first_byte);
sum_abs = _mm_min_epi32(sum_abs, _mm_set1_epi32(51));
} }
case 4: case 4:
{ {

2
src/transform.c
View file

@ -437,7 +437,7 @@ static void quantize_chroma(
int8_t height = cu_loc->chroma_height; int8_t height = cu_loc->chroma_height;
if(state->encoder_control->cfg.dep_quant && transform != CHROMA_TS) { if(state->encoder_control->cfg.dep_quant && transform != CHROMA_TS) {
int abs_sum = 0; int abs_sum = 0;
state->quant_blocks[2].needs_init = state->encoder_control->cfg.jccr; state->quant_blocks[2].needs_init |= state->encoder_control->cfg.jccr;
uvg_dep_quant( uvg_dep_quant(
state, state,
cur_tu, cur_tu,