inline __m256i avx2_uv_to_ringid(const __m256i u, const __m256i v) { // return static_cast<unsigned>(std::max({std::abs(u), std::abs(v), // std::abs(u+v)})); __m256i ringid = _mm256_abs_epi32(u); ringid = _mm256_max_epu32(ringid, _mm256_abs_epi32(v)); ringid = _mm256_max_epu32(ringid, _mm256_abs_epi32(_mm256_add_epi32(u,v))); return ringid; }
int vpx_highbd_satd_avx2(const tran_low_t *coeff, int length) { __m256i accum = _mm256_setzero_si256(); int i; for (i = 0; i < length; i += 8, coeff += 8) { const __m256i src_line = _mm256_loadu_si256((const __m256i *)coeff); const __m256i abs = _mm256_abs_epi32(src_line); accum = _mm256_add_epi32(accum, abs); } { // 32 bit horizontal add const __m256i a = _mm256_srli_si256(accum, 8); const __m256i b = _mm256_add_epi32(accum, a); const __m256i c = _mm256_srli_epi64(b, 32); const __m256i d = _mm256_add_epi32(b, c); const __m128i accum_128 = _mm_add_epi32(_mm256_castsi256_si128(d), _mm256_extractf128_si256(d, 1)); return _mm_cvtsi128_si32(accum_128); } }
static INLINE void quantize(const __m256i *qp, __m256i *c, const int16_t *iscan_ptr, int log_scale, tran_low_t *qcoeff, tran_low_t *dqcoeff, __m256i *eob) { const __m256i abs_coeff = _mm256_abs_epi32(*c); __m256i q = _mm256_add_epi32(abs_coeff, qp[0]); __m256i q_lo = _mm256_mul_epi32(q, qp[1]); __m256i q_hi = _mm256_srli_epi64(q, 32); const __m256i qp_hi = _mm256_srli_epi64(qp[1], 32); q_hi = _mm256_mul_epi32(q_hi, qp_hi); q_lo = _mm256_srli_epi64(q_lo, 16 - log_scale); q_hi = _mm256_srli_epi64(q_hi, 16 - log_scale); q_hi = _mm256_slli_epi64(q_hi, 32); q = _mm256_or_si256(q_lo, q_hi); const __m256i abs_s = _mm256_slli_epi32(abs_coeff, 1 + log_scale); const __m256i mask = _mm256_cmpgt_epi32(qp[2], abs_s); q = _mm256_andnot_si256(mask, q); __m256i dq = _mm256_mullo_epi32(q, qp[2]); dq = _mm256_srai_epi32(dq, log_scale); q = _mm256_sign_epi32(q, *c); dq = _mm256_sign_epi32(dq, *c); _mm256_storeu_si256((__m256i *)qcoeff, q); _mm256_storeu_si256((__m256i *)dqcoeff, dq); const __m128i isc = _mm_loadu_si128((const __m128i *)iscan_ptr); const __m128i zr = _mm_setzero_si128(); const __m128i lo = _mm_unpacklo_epi16(isc, zr); const __m128i hi = _mm_unpackhi_epi16(isc, zr); const __m256i iscan = _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1); const __m256i zero = _mm256_setzero_si256(); const __m256i zc = _mm256_cmpeq_epi32(dq, zero); const __m256i nz = _mm256_cmpeq_epi32(zc, zero); __m256i cur_eob = _mm256_sub_epi32(iscan, nz); cur_eob = _mm256_and_si256(cur_eob, nz); *eob = _mm256_max_epi32(cur_eob, *eob); }
__m256i test_mm256_abs_epi32(__m256i a) { // CHECK: @llvm.x86.avx2.pabs.d return _mm256_abs_epi32(a); }
__m256i test_mm256_abs_epi32(__m256i a) { // CHECK-LABEL: test_mm256_abs_epi32 // CHECK: call <8 x i32> @llvm.x86.avx2.pabs.d(<8 x i32> %{{.*}}) return _mm256_abs_epi32(a); }
void extern avx2_test (void) { x = _mm256_abs_epi32 (x); }
void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[], uint32_t residual_samples, uint32_t predictor_order, uint32_t min_partition_order, uint32_t max_partition_order, uint32_t bps) { const uint32_t default_partition_samples = (residual_samples + predictor_order) >> max_partition_order; uint32_t partitions = 1u << max_partition_order; FLAC__ASSERT(default_partition_samples > predictor_order); /* first do max_partition_order */ { const uint32_t threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples); uint32_t partition, residual_sample, end = (uint32_t)(-(int32_t)predictor_order); if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) { for(partition = residual_sample = 0; partition < partitions; partition++) { __m256i sum256 = _mm256_setzero_si256(); __m128i sum128; end += default_partition_samples; for( ; (int)residual_sample < (int)end-7; residual_sample+=8) { __m256i res256 = _mm256_abs_epi32(_mm256_loadu_si256((const __m256i*)(residual+residual_sample))); sum256 = _mm256_add_epi32(sum256, res256); } sum128 = _mm_add_epi32(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256)); for( ; (int)residual_sample < (int)end-3; residual_sample+=4) { __m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample))); sum128 = _mm_add_epi32(sum128, res128); } for( ; residual_sample < end; residual_sample++) { __m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); sum128 = _mm_add_epi32(sum128, res128); } sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_SHUFFLE(1,0,3,2))); sum128 = _mm_add_epi32(sum128, _mm_shufflelo_epi16(sum128, _MM_SHUFFLE(1,0,3,2))); abs_residual_partition_sums[partition] = (FLAC__uint32)_mm_cvtsi128_si32(sum128); /* workaround for MSVC bugs (at least versions 2015 and 2017 are affected) */ #if (defined _MSC_VER) && (defined FLAC__CPU_X86_64) abs_residual_partition_sums[partition] &= 0xFFFFFFFF; /**/ #endif } } else { /* have to pessimistically use 64 bits for accumulator */ for(partition = residual_sample = 0; partition < partitions; partition++) { __m256i sum256 = _mm256_setzero_si256(); __m128i sum128; end += default_partition_samples; for( ; (int)residual_sample < (int)end-3; residual_sample+=4) { __m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample))); __m256i res256 = _mm256_cvtepu32_epi64(res128); sum256 = _mm256_add_epi64(sum256, res256); } sum128 = _mm_add_epi64(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256)); for( ; (int)residual_sample < (int)end-1; residual_sample+=2) { __m128i res128 = _mm_abs_epi32(_mm_loadl_epi64((const __m128i*)(residual+residual_sample))); res128 = _mm_cvtepu32_epi64(res128); sum128 = _mm_add_epi64(sum128, res128); } for( ; residual_sample < end; residual_sample++) { __m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); sum128 = _mm_add_epi64(sum128, res128); } sum128 = _mm_add_epi64(sum128, _mm_srli_si128(sum128, 8)); _mm_storel_epi64((__m128i*)(abs_residual_partition_sums+partition), sum128); } } } /* now merge partitions for lower orders */ { uint32_t from_partition = 0, to_partition = partitions; int partition_order; for(partition_order = (int)max_partition_order - 1; partition_order >= (int)min_partition_order; partition_order--) { uint32_t i; partitions >>= 1; for(i = 0; i < partitions; i++) { abs_residual_partition_sums[to_partition++] = abs_residual_partition_sums[from_partition ] + abs_residual_partition_sums[from_partition+1]; from_partition += 2; } } } _mm256_zeroupper(); }
void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[], uint32_t residual_samples, uint32_t predictor_order, uint32_t min_partition_order, uint32_t max_partition_order, uint32_t bps) { const uint32_t default_partition_samples = (residual_samples + predictor_order) >> max_partition_order; uint32_t partitions = 1u << max_partition_order; FLAC__ASSERT(default_partition_samples > predictor_order); /* first do max_partition_order */ { const uint32_t threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples); uint32_t partition, residual_sample, end = (uint32_t)(-(int32_t)predictor_order); if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) { for(partition = residual_sample = 0; partition < partitions; partition++) { __m256i sum256 = _mm256_setzero_si256(); __m128i sum128; end += default_partition_samples; for( ; (int)residual_sample < (int)end-7; residual_sample+=8) { __m256i res256 = _mm256_abs_epi32(_mm256_loadu_si256((const __m256i*)(residual+residual_sample))); sum256 = _mm256_add_epi32(sum256, res256); } sum128 = _mm_add_epi32(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256)); for( ; (int)residual_sample < (int)end-3; residual_sample+=4) { __m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample))); sum128 = _mm_add_epi32(sum128, res128); } for( ; residual_sample < end; residual_sample++) { __m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); sum128 = _mm_add_epi32(sum128, res128); } sum128 = _mm_hadd_epi32(sum128, sum128); sum128 = _mm_hadd_epi32(sum128, sum128); abs_residual_partition_sums[partition] = (FLAC__uint32)_mm_cvtsi128_si32(sum128); /* workaround for a bug in MSVC2015U2 - see https://connect.microsoft.com/VisualStudio/feedback/details/2659191/incorrect-code-generation-for-x86-64 */ #if (defined _MSC_VER) && (_MSC_FULL_VER == 190023918) && (defined FLAC__CPU_X86_64) abs_residual_partition_sums[partition] &= 0xFFFFFFFF; /**/ #endif } } else { /* have to pessimistically use 64 bits for accumulator */ for(partition = residual_sample = 0; partition < partitions; partition++) { __m256i sum256 = _mm256_setzero_si256(); __m128i sum128; end += default_partition_samples; for( ; (int)residual_sample < (int)end-3; residual_sample+=4) { __m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample))); __m256i res256 = _mm256_cvtepu32_epi64(res128); sum256 = _mm256_add_epi64(sum256, res256); } sum128 = _mm_add_epi64(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256)); for( ; (int)residual_sample < (int)end-1; residual_sample+=2) { __m128i res128 = _mm_abs_epi32(_mm_loadl_epi64((const __m128i*)(residual+residual_sample))); res128 = _mm_cvtepu32_epi64(res128); sum128 = _mm_add_epi64(sum128, res128); } for( ; residual_sample < end; residual_sample++) { __m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); sum128 = _mm_add_epi64(sum128, res128); } sum128 = _mm_add_epi64(sum128, _mm_srli_si128(sum128, 8)); _mm_storel_epi64((__m128i*)(abs_residual_partition_sums+partition), sum128); } } } /* now merge partitions for lower orders */ { uint32_t from_partition = 0, to_partition = partitions; int partition_order; for(partition_order = (int)max_partition_order - 1; partition_order >= (int)min_partition_order; partition_order--) { uint32_t i; partitions >>= 1; for(i = 0; i < partitions; i++) { abs_residual_partition_sums[to_partition++] = abs_residual_partition_sums[from_partition ] + abs_residual_partition_sums[from_partition+1]; from_partition += 2; } } } _mm256_zeroupper(); }