[avx2] WIP check_rd_costs_avx2, almost?

This commit is contained in:
Joose Sainio 2023-04-08 18:58:40 +03:00
parent 8caabcde1a
commit c6e6f5da33

View file

@ -563,10 +563,6 @@ static INLINE void checkRdCostSkipSbbZeroOut(
static void check_rd_costs_avx2(const all_depquant_states* const state, const enum ScanPosType spt, const PQData* pqDataA, Decision* decisions, int start)
{
int32_t a[64] = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64};
__m128i offsets = _mm_set_epi32(12, 8, 4, 0);
__m128i r = _mm_i32gather_epi32(a, offsets, 1);
int64_t temp_rd_cost_a[4] = {0, 0, 0, 0};
int64_t temp_rd_cost_b[4] = {0, 0, 0, 0};
int64_t temp_rd_cost_z[4] = {0, 0, 0, 0};
@ -600,15 +596,15 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
__m128i offsets = _mm_set_epi32(18 + pqDataA->absLevel[3], 12 + pqDataA->absLevel[3], 6 + pqDataA->absLevel[0], 0 + pqDataA->absLevel[0]);
__m128i t = _mm_slli_epi32(value, 1);
offsets = _mm_sub_epi32(offsets, t);
__m128i coeff_frac_bits = _mm_i32gather_epi32(state->m_coeffFracBits[start], offsets, 1);
__m128i coeff_frac_bits = _mm_i32gather_epi32(state->m_coeffFracBits[start], offsets, 4);
__m128i max_rice = _mm_set1_epi32(15);
__m128i max_rice = _mm_set1_epi32(31);
value = _mm_min_epi32(value, max_rice);
__m128i go_rice_tab = _mm_cvtepi8_epi32(_mm_loadu_si32(&state->m_goRicePar[start]));
go_rice_tab = _mm_slli_epi32(value, 5);
value = _mm_add_epi32(value, go_rice_tab);
__m128i temp = _mm_add_epi32(coeff_frac_bits, _mm_i32gather_epi32(&g_goRiceBits[0][0], value, 1));
__m128i temp = _mm_add_epi32(coeff_frac_bits, _mm_i32gather_epi32(&g_goRiceBits[0][0], value, 4));
rd_cost_a = _mm256_add_epi64(rd_cost_a, _mm256_cvtepi32_epi64(temp));
} else {
const int pqAs[4] = {0, 0, 3, 3};
@ -629,7 +625,7 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
if (pqDataA->absLevel[1] < 4 && pqDataA->absLevel[2] < 4) {
__m128i offsets = _mm_set_epi32(18 + pqDataA->absLevel[1], 12 + pqDataA->absLevel[1], 6 + pqDataA->absLevel[2], 0 + pqDataA->absLevel[2]);
__m128i coeff_frac_bits = _mm_i32gather_epi32(state->m_coeffFracBits[start], offsets, 1);
__m128i coeff_frac_bits = _mm_i32gather_epi32(state->m_coeffFracBits[start], offsets, 4);
__m256i ext_frac_bits = _mm256_cvtepi32_epi64(coeff_frac_bits);
rd_cost_b = _mm256_add_epi64(rd_cost_b, ext_frac_bits);
} else if (pqDataA->absLevel[1] >= 4 && pqDataA->absLevel[2] >= 4) {
@ -638,28 +634,28 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
__m128i offsets = _mm_set_epi32(18 + pqDataA->absLevel[1], 12 + pqDataA->absLevel[1], 6 + pqDataA->absLevel[2], 0 + pqDataA->absLevel[2]);
__m128i t = _mm_slli_epi32(value, 1);
offsets = _mm_sub_epi32(offsets, t);
__m128i coeff_frac_bits = _mm_i32gather_epi32(state->m_coeffFracBits[start], offsets, 1);
__m128i coeff_frac_bits = _mm_i32gather_epi32(state->m_coeffFracBits[start], offsets, 4);
__m128i max_rice = _mm_set1_epi32(15);
__m128i max_rice = _mm_set1_epi32(31);
value = _mm_min_epi32(value, max_rice);
__m128i go_rice_tab = _mm_cvtepi8_epi32(_mm_loadu_si32(&state->m_goRicePar[start]));
go_rice_tab = _mm_slli_epi32(go_rice_tab, 5);
value = _mm_add_epi32(value, go_rice_tab);
__m128i temp = _mm_add_epi32(coeff_frac_bits, _mm_i32gather_epi32(&g_goRiceBits[0][0], value, 1));
__m128i temp = _mm_add_epi32(coeff_frac_bits, _mm_i32gather_epi32(&g_goRiceBits[0][0], value, 4));
rd_cost_b = _mm256_add_epi64(rd_cost_b, _mm256_cvtepi32_epi64(temp));
} else {
const int pqAs[4] = {0, 0, 3, 3};
const int pqBs[4] = {2, 2, 1, 1};
int64_t rd_costs[4] = {0, 0, 0, 0};
for (int i = 0; i < 4; i++) {
const int state_offset = start + i;
const int pqA = pqAs[i];
const int pqB = pqBs[i];
const int32_t* goRiceTab = g_goRiceBits[state->m_goRicePar[state_offset]];
if (pqDataA->absLevel[pqA] < 4) {
rd_costs[i] = state->m_coeffFracBits[state_offset][pqDataA->absLevel[pqA]];
if (pqDataA->absLevel[pqB] < 4) {
rd_costs[i] = state->m_coeffFracBits[state_offset][pqDataA->absLevel[pqB]];
} else {
const coeff_t value = (pqDataA->absLevel[pqA] - 4) >> 1;
rd_costs[i] += state->m_coeffFracBits[state_offset][pqDataA->absLevel[pqA] - (value << 1)] + goRiceTab[value < RICEMAX ? value : RICEMAX - 1];
const coeff_t value = (pqDataA->absLevel[pqB] - 4) >> 1;
rd_costs[i] += state->m_coeffFracBits[state_offset][pqDataA->absLevel[pqB] - (value << 1)] + goRiceTab[value < RICEMAX ? value : RICEMAX - 1];
}
}
rd_cost_b = _mm256_add_epi64(rd_cost_b, _mm256_loadu_si256(&rd_costs[0]));
@ -672,7 +668,7 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
__m256i even = _mm256_permutevar8x32_epi32(original, even_mask);
__m256i odd = _mm256_permutevar8x32_epi32(original, odd_mask);
__m256i even_64 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(even, 0));
__m256i odd_64 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(odd, 1));
__m256i odd_64 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(odd, 0));
rd_cost_a = _mm256_add_epi64(rd_cost_a, odd_64);
rd_cost_b = _mm256_add_epi64(rd_cost_b, odd_64);
rd_cost_z = _mm256_add_epi64(rd_cost_z, even_64);
@ -683,11 +679,11 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
__m256i even = _mm256_permutevar8x32_epi32(original, even_mask);
__m256i odd = _mm256_permutevar8x32_epi32(original, odd_mask);
__m256i m_sigFracBits_0 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(even, 0));
__m256i m_sigFracBits_1 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(odd, 1));
__m256i m_sigFracBits_1 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(odd, 0));
original = _mm256_loadu_si256((__m256i const*)state->m_sbbFracBits[start]);
odd = _mm256_permutevar8x32_epi32(original, odd_mask);
__m256i m_sbbFracBits_1 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(odd, 1));
__m256i m_sbbFracBits_1 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(odd, 0));
rd_cost_a = _mm256_add_epi64(rd_cost_a, m_sbbFracBits_1);
@ -706,22 +702,26 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
__m256i even = _mm256_permutevar8x32_epi32(original, even_mask);
__m256i odd = _mm256_permutevar8x32_epi32(original, odd_mask);
__m256i even_64 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(even, 0));
__m256i odd_64 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(odd, 1));
__m256i odd_64 = _mm256_cvtepi32_epi64(_mm256_extracti128_si256(odd, 0));
rd_cost_a = _mm256_add_epi64(rd_cost_a, odd_64);
rd_cost_b = _mm256_add_epi64(rd_cost_b, odd_64);
rd_cost_z = _mm256_add_epi64(rd_cost_z, even_64);
rd_cost_z = _mm256_add_epi64(rd_cost_z, even_64);
_mm256_storeu_epi64(temp_rd_cost_a, rd_cost_a);
_mm256_storeu_epi64(temp_rd_cost_b, rd_cost_b);
_mm256_storeu_epi64(temp_rd_cost_z, rd_cost_z);
}
else if (!state->m_numSigSbb[start] && !state->m_numSigSbb[start + 1] && !state->m_numSigSbb[start + 2] && !state->m_numSigSbb[start + 3]) {
rd_cost_z = _mm256_setr_epi64x(decisions->rdCost[3], decisions->rdCost[3], decisions->rdCost[0], decisions->rdCost[0]);
rd_cost_z = _mm256_setr_epi64x(decisions->rdCost[0], decisions->rdCost[0], decisions->rdCost[3], decisions->rdCost[3]);
_mm256_storeu_epi64(temp_rd_cost_a, rd_cost_a);
_mm256_storeu_epi64(temp_rd_cost_b, rd_cost_b);
_mm256_storeu_epi64(temp_rd_cost_z, rd_cost_z);
}
else {
const int pqAs[4] = {0, 0, 3, 3};
int64_t temp_rd_cost_a[4] = {0, 0, 0, 0};
int64_t temp_rd_cost_b[4] = {0, 0, 0, 0};
int64_t temp_rd_cost_z[4] = {0, 0, 0, 0};
int64_t z_out[4] = {0, 0, 0, 0};
_mm256_storeu_epi64(z_out, rd_cost_z);
_mm256_storeu_epi64(temp_rd_cost_a, rd_cost_a);
_mm256_storeu_epi64(temp_rd_cost_b, rd_cost_b);
_mm256_storeu_epi64(temp_rd_cost_z, rd_cost_z);
for (int i = 0; i < 4; i++) {
const int state_offset = start + i;
if (state->m_numSigSbb[state_offset]) {
@ -729,13 +729,9 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
temp_rd_cost_b[i] += state->m_sigFracBits[state_offset][1];
temp_rd_cost_z[i] += state->m_sigFracBits[state_offset][0];
} else {
z_out[i] = decisions->rdCost[pqAs[i]];
temp_rd_cost_z[i] = decisions->rdCost[pqAs[i]];
}
}
rd_cost_z = _mm256_loadu_epi64(z_out);
rd_cost_a = _mm256_add_epi64(rd_cost_a, _mm256_loadu_epi64(temp_rd_cost_a));
rd_cost_b = _mm256_add_epi64(rd_cost_b, _mm256_loadu_epi64(temp_rd_cost_b));
rd_cost_z = _mm256_add_epi64(rd_cost_z, _mm256_loadu_epi64(temp_rd_cost_z));
}
}
_mm256_storeu_epi64(temp_rd_cost_a, rd_cost_a);
@ -743,25 +739,25 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
_mm256_storeu_epi64(temp_rd_cost_z, rd_cost_z);
} else if (all_under_four) {
__m128i scale_bits = _mm_set1_epi32(1 << SCALE_BITS);
__m128i max_rice = _mm_set1_epi32(15);
__m128i max_rice = _mm_set1_epi32(31);
__m128i go_rice_zero = _mm_cvtepi8_epi32(_mm_loadu_epi8(&state->m_goRiceZero[start]));
// RD cost A
{
__m128i pq_abs_a = _mm_set_epi32(pqDataA->absLevel[3], pqDataA->absLevel[3], pqDataA->absLevel[0], pqDataA->absLevel[0]);
__m128i cmp = _mm_cmplt_epi32(go_rice_zero, pq_abs_a);
__m128i cmp = _mm_cmpgt_epi32(pq_abs_a, go_rice_zero);
__m128i go_rice_smaller = _mm_min_epi32(pq_abs_a, max_rice);
__m128i other = _mm_sub_epi32(pq_abs_a, _mm_set1_epi32(1));
__m128i selected = _mm_blendv_epi8(go_rice_smaller, other, cmp);
__m128i selected = _mm_blendv_epi8(other, go_rice_smaller, cmp);
__m128i go_rice_offset = _mm_cvtepi8_epi32(_mm_loadu_si32(&state->m_goRicePar[start]));
go_rice_offset = _mm_slli_epi32(go_rice_offset, 5);
__m128i offsets = _mm_add_epi32(selected, go_rice_offset);
__m128i go_rice_tab = _mm_i32gather_epi32(&g_goRiceBits[0][0], offsets, 1);
__m128i go_rice_tab = _mm_i32gather_epi32(&g_goRiceBits[0][0], offsets, 4);
__m128i temp = _mm_add_epi32(go_rice_tab, scale_bits);
rd_cost_a = _mm256_add_epi64(rd_cost_a, _mm256_cvtepi32_epi64(temp));
@ -769,20 +765,20 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
// RD cost b
{
__m128i pq_abs_b = _mm_set_epi32(pqDataA->absLevel[1], pqDataA->absLevel[1], pqDataA->absLevel[2], pqDataA->absLevel[2]);
__m128i cmp = _mm_cmplt_epi32(go_rice_zero, pq_abs_b);
__m128i cmp = _mm_cmpgt_epi32(pq_abs_b, go_rice_zero);
__m128i go_rice_smaller = _mm_min_epi32(pq_abs_b, max_rice);
__m128i other = _mm_sub_epi32(pq_abs_b, _mm_set1_epi32(1));
__m128i selected = _mm_blendv_epi8(go_rice_smaller, other, cmp);
__m128i selected = _mm_blendv_epi8(other, go_rice_smaller, cmp);
__m128i go_rice_offset = _mm_cvtepi8_epi32(_mm_loadu_si32(&state->m_goRicePar[start]));
go_rice_offset = _mm_slli_epi32(go_rice_offset, 5);
__m128i offsets = _mm_add_epi32(selected, go_rice_offset);
__m128i go_rice_tab = _mm_i32gather_epi32(&g_goRiceBits[0][0], offsets, 1);
__m128i go_rice_tab = _mm_i32gather_epi32(&g_goRiceBits[0][0], offsets, 4);
__m128i temp = _mm_add_epi32(go_rice_tab, scale_bits);
rd_cost_b = _mm256_add_epi64(rd_cost_b, _mm256_cvtepi32_epi64(temp));
@ -793,7 +789,8 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
go_rice_offset = _mm_slli_epi32(go_rice_offset, 5);
go_rice_offset = _mm_add_epi32(go_rice_offset, go_rice_zero);
rd_cost_z = _mm256_add_epi64(rd_cost_z, _mm256_cvtepi32_epi64(go_rice_offset));
__m128i go_rice_tab = _mm_i32gather_epi32(&g_goRiceBits[0][0], go_rice_offset, 4);
rd_cost_z = _mm256_add_epi64(rd_cost_z, _mm256_cvtepi32_epi64(go_rice_tab));
}
_mm256_storeu_epi64(temp_rd_cost_a, rd_cost_a);
_mm256_storeu_epi64(temp_rd_cost_b, rd_cost_b);
@ -868,7 +865,7 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
decisions->prevId[0] = state->m_stateId[start + 1];
}
// Decision 1
// Decision 2
if (temp_rd_cost_a[1] < decisions->rdCost[2]) {
decisions->rdCost[2] = temp_rd_cost_a[1];
decisions->absLevel[2] = pqDataA->absLevel[0];
@ -885,35 +882,35 @@ static void check_rd_costs_avx2(const all_depquant_states* const state, const en
decisions->prevId[2] = state->m_stateId[start];
}
// Decision 2
if (temp_rd_cost_a[2] < decisions->rdCost[0]) {
decisions->rdCost[2] = temp_rd_cost_a[2];
decisions->absLevel[2] = pqDataA->absLevel[3];
decisions->prevId[2] = state->m_stateId[start + 2];
// Decision 1
if (temp_rd_cost_a[2] < decisions->rdCost[1]) {
decisions->rdCost[1] = temp_rd_cost_a[2];
decisions->absLevel[1] = pqDataA->absLevel[3];
decisions->prevId[1] = state->m_stateId[start + 2];
}
if (temp_rd_cost_z[2] < decisions->rdCost[0]) {
decisions->rdCost[2] = temp_rd_cost_z[2];
decisions->absLevel[2] = 0;
decisions->prevId[2] = state->m_stateId[start + 2];
if (temp_rd_cost_z[2] < decisions->rdCost[1]) {
decisions->rdCost[1] = temp_rd_cost_z[2];
decisions->absLevel[1] = 0;
decisions->prevId[1] = state->m_stateId[start + 2];
}
if (temp_rd_cost_b[3] < decisions->rdCost[0]) {
decisions->rdCost[2] = temp_rd_cost_b[3];
decisions->absLevel[2] = pqDataA->absLevel[1];
decisions->prevId[2] = state->m_stateId[start + 3];
if (temp_rd_cost_b[3] < decisions->rdCost[1]) {
decisions->rdCost[1] = temp_rd_cost_b[3];
decisions->absLevel[1] = pqDataA->absLevel[1];
decisions->prevId[1] = state->m_stateId[start + 3];
}
// Decision 3
if (temp_rd_cost_a[3] < decisions->rdCost[1]) {
if (temp_rd_cost_a[3] < decisions->rdCost[3]) {
decisions->rdCost[3] = temp_rd_cost_a[3];
decisions->absLevel[3] = pqDataA->absLevel[3];
decisions->prevId[3] = state->m_stateId[start + 3];
}
if (temp_rd_cost_z[3] < decisions->rdCost[1]) {
if (temp_rd_cost_z[3] < decisions->rdCost[3]) {
decisions->rdCost[3] = temp_rd_cost_z[3];
decisions->absLevel[3] = 0;
decisions->prevId[3] = state->m_stateId[start + 3];
}
if (temp_rd_cost_b[2] < decisions->rdCost[1]) {
if (temp_rd_cost_b[2] < decisions->rdCost[3]) {
decisions->rdCost[3] = temp_rd_cost_b[2];
decisions->absLevel[3] = pqDataA->absLevel[1];
decisions->prevId[3] = state->m_stateId[start + 2];