diff --git a/src/libFLAC/stream_encoder.c b/src/libFLAC/stream_encoder.c index 24e1ed4..a7e1191 100644 --- a/src/libFLAC/stream_encoder.c +++ b/src/libFLAC/stream_encoder.c @@ -3984,29 +3984,25 @@ void precompute_partition_info_sums_( /* first do max_partition_order */ { + const unsigned threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples); unsigned partition, residual_sample, end = (unsigned)(-(int)predictor_order); - /* WATCHOUT: "+ bps + FLAC__MAX_EXTRA_RESIDUAL_BPS" is the maximum - * assumed size of the average residual magnitude */ - if(FLAC__bitmath_ilog2(default_partition_samples) + bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < 32) { - FLAC__uint32 abs_residual_partition_sum; - + /* WATCHOUT: "bps + FLAC__MAX_EXTRA_RESIDUAL_BPS" is the maximum assumed size of the average residual magnitude */ + if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) { for(partition = residual_sample = 0; partition < partitions; partition++) { + FLAC__uint32 abs_residual_partition_sum = 0; end += default_partition_samples; - abs_residual_partition_sum = 0; for( ; residual_sample < end; residual_sample++) abs_residual_partition_sum += abs(residual[residual_sample]); /* abs(INT_MIN) is undefined, but if the residual is INT_MIN we have bigger problems */ abs_residual_partition_sums[partition] = abs_residual_partition_sum; } } else { /* have to pessimistically use 64 bits for accumulator */ - FLAC__uint64 abs_residual_partition_sum; - for(partition = residual_sample = 0; partition < partitions; partition++) { + FLAC__uint64 abs_residual_partition_sum64 = 0; end += default_partition_samples; - abs_residual_partition_sum = 0; for( ; residual_sample < end; residual_sample++) - abs_residual_partition_sum += abs(residual[residual_sample]); /* abs(INT_MIN) is undefined, but if the residual is INT_MIN we have bigger problems */ - abs_residual_partition_sums[partition] = abs_residual_partition_sum; + abs_residual_partition_sum64 += abs(residual[residual_sample]); /* abs(INT_MIN) is undefined, but if the residual is INT_MIN we have bigger problems */ + abs_residual_partition_sums[partition] = abs_residual_partition_sum64; } } } diff --git a/src/libFLAC/stream_encoder_intrin_avx2.c b/src/libFLAC/stream_encoder_intrin_avx2.c index 3aa3197..4381bdb 100644 --- a/src/libFLAC/stream_encoder_intrin_avx2.c +++ b/src/libFLAC/stream_encoder_intrin_avx2.c @@ -55,29 +55,29 @@ void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual /* first do max_partition_order */ { + const unsigned threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples); unsigned partition, residual_sample, end = (unsigned)(-(int)predictor_order); - __m256i res256, sum256; - __m128i res128, sum128; - if(FLAC__bitmath_ilog2(default_partition_samples) + bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < 32) { + 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; - sum256 = _mm256_setzero_si256(); for( ; (int)residual_sample < (int)end-7; residual_sample+=8) { - res256 = _mm256_abs_epi32(_mm256_loadu_si256((const __m256i*)(residual+residual_sample))); + __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) { - res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample))); + __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++) { - res128 = _mm_cvtsi32_si128(residual[residual_sample]); + __m128i res128 = _mm_cvtsi32_si128(residual[residual_sample]); res128 = _mm_abs_epi32(res128); sum128 = _mm_add_epi32(sum128, res128); } @@ -89,26 +89,27 @@ void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual } 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; - sum256 = _mm256_setzero_si256(); for( ; (int)residual_sample < (int)end-3; residual_sample+=4) { - res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample))); - res256 = _mm256_cvtepu32_epi64(res128); + __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) { - res128 = _mm_loadl_epi64((const __m128i*)(residual+residual_sample)); + __m128i res128 = _mm_loadl_epi64((const __m128i*)(residual+residual_sample)); res128 = _mm_abs_epi32(res128); res128 = _mm_cvtepu32_epi64(res128); sum128 = _mm_add_epi64(sum128, res128); } for( ; residual_sample < end; residual_sample++) { - res128 = _mm_cvtsi32_si128(residual[residual_sample]); + __m128i res128 = _mm_cvtsi32_si128(residual[residual_sample]); res128 = _mm_abs_epi32(res128); sum128 = _mm_add_epi64(sum128, res128); } diff --git a/src/libFLAC/stream_encoder_intrin_sse2.c b/src/libFLAC/stream_encoder_intrin_sse2.c index ec5541c..be8395a 100644 --- a/src/libFLAC/stream_encoder_intrin_sse2.c +++ b/src/libFLAC/stream_encoder_intrin_sse2.c @@ -55,14 +55,14 @@ void FLAC__precompute_partition_info_sums_intrin_sse2(const FLAC__int32 residual /* first do max_partition_order */ { + const unsigned threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples); unsigned partition, residual_sample, end = (unsigned)(-(int)predictor_order); - unsigned e1, e3; - __m128i mm_res, mm_sum, mm_mask; - if(FLAC__bitmath_ilog2(default_partition_samples) + bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < 32) { + if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) { for(partition = residual_sample = 0; partition < partitions; partition++) { + __m128i mm_sum = _mm_setzero_si128(); + unsigned e1, e3; end += default_partition_samples; - mm_sum = _mm_setzero_si128(); e1 = (residual_sample + 3) & ~3; e3 = end & ~3; if(e1 > end) @@ -70,24 +70,24 @@ void FLAC__precompute_partition_info_sums_intrin_sse2(const FLAC__int32 residual /* assumption: residual[] is properly aligned so (residual + e1) is properly aligned too and _mm_loadu_si128() is fast */ for( ; residual_sample < e1; residual_sample++) { - mm_res = _mm_cvtsi32_si128(residual[residual_sample]); - mm_mask = _mm_srai_epi32(mm_res, 31); + __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); + __m128i mm_mask = _mm_srai_epi32(mm_res, 31); mm_res = _mm_xor_si128(mm_res, mm_mask); mm_res = _mm_sub_epi32(mm_res, mm_mask); /* abs(INT_MIN) is undefined, but if the residual is INT_MIN we have bigger problems */ mm_sum = _mm_add_epi32(mm_sum, mm_res); } for( ; residual_sample < e3; residual_sample+=4) { - mm_res = _mm_loadu_si128((const __m128i*)(residual+residual_sample)); - mm_mask = _mm_srai_epi32(mm_res, 31); + __m128i mm_res = _mm_loadu_si128((const __m128i*)(residual+residual_sample)); + __m128i mm_mask = _mm_srai_epi32(mm_res, 31); mm_res = _mm_xor_si128(mm_res, mm_mask); mm_res = _mm_sub_epi32(mm_res, mm_mask); mm_sum = _mm_add_epi32(mm_sum, mm_res); } for( ; residual_sample < end; residual_sample++) { - mm_res = _mm_cvtsi32_si128(residual[residual_sample]); - mm_mask = _mm_srai_epi32(mm_res, 31); + __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); + __m128i mm_mask = _mm_srai_epi32(mm_res, 31); mm_res = _mm_xor_si128(mm_res, mm_mask); mm_res = _mm_sub_epi32(mm_res, mm_mask); mm_sum = _mm_add_epi32(mm_sum, mm_res); @@ -100,23 +100,24 @@ void FLAC__precompute_partition_info_sums_intrin_sse2(const FLAC__int32 residual } else { /* have to pessimistically use 64 bits for accumulator */ for(partition = residual_sample = 0; partition < partitions; partition++) { + __m128i mm_sum = _mm_setzero_si128(); + unsigned e1, e3; end += default_partition_samples; - mm_sum = _mm_setzero_si128(); e1 = (residual_sample + 1) & ~1; e3 = end & ~1; FLAC__ASSERT(e1 <= end); for( ; residual_sample < e1; residual_sample++) { - mm_res = _mm_cvtsi32_si128(residual[residual_sample]); /* 0 0 0 r0 */ - mm_mask = _mm_srai_epi32(mm_res, 31); + __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); /* 0 0 0 r0 */ + __m128i mm_mask = _mm_srai_epi32(mm_res, 31); mm_res = _mm_xor_si128(mm_res, mm_mask); mm_res = _mm_sub_epi32(mm_res, mm_mask); /* 0 0 0 |r0| == 00 |r0_64| */ mm_sum = _mm_add_epi64(mm_sum, mm_res); } for( ; residual_sample < e3; residual_sample+=2) { - mm_res = _mm_loadl_epi64((const __m128i*)(residual+residual_sample)); /* 0 0 r1 r0 */ - mm_mask = _mm_srai_epi32(mm_res, 31); + __m128i mm_res = _mm_loadl_epi64((const __m128i*)(residual+residual_sample)); /* 0 0 r1 r0 */ + __m128i mm_mask = _mm_srai_epi32(mm_res, 31); mm_res = _mm_xor_si128(mm_res, mm_mask); mm_res = _mm_sub_epi32(mm_res, mm_mask); /* 0 0 |r1| |r0| */ mm_res = _mm_shuffle_epi32(mm_res, _MM_SHUFFLE(3,1,2,0)); /* 0 |r1| 0 |r0| == |r1_64| |r0_64| */ @@ -124,8 +125,8 @@ void FLAC__precompute_partition_info_sums_intrin_sse2(const FLAC__int32 residual } for( ; residual_sample < end; residual_sample++) { - mm_res = _mm_cvtsi32_si128(residual[residual_sample]); - mm_mask = _mm_srai_epi32(mm_res, 31); + __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); + __m128i mm_mask = _mm_srai_epi32(mm_res, 31); mm_res = _mm_xor_si128(mm_res, mm_mask); mm_res = _mm_sub_epi32(mm_res, mm_mask); mm_sum = _mm_add_epi64(mm_sum, mm_res); diff --git a/src/libFLAC/stream_encoder_intrin_ssse3.c b/src/libFLAC/stream_encoder_intrin_ssse3.c index 2dbd18c..547ef95 100644 --- a/src/libFLAC/stream_encoder_intrin_ssse3.c +++ b/src/libFLAC/stream_encoder_intrin_ssse3.c @@ -55,14 +55,14 @@ void FLAC__precompute_partition_info_sums_intrin_ssse3(const FLAC__int32 residua /* first do max_partition_order */ { + const unsigned threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples); unsigned partition, residual_sample, end = (unsigned)(-(int)predictor_order); - unsigned e1, e3; - __m128i mm_res, mm_sum; - if(FLAC__bitmath_ilog2(default_partition_samples) + bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < 32) { + if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) { for(partition = residual_sample = 0; partition < partitions; partition++) { + __m128i mm_sum = _mm_setzero_si128(); + unsigned e1, e3; end += default_partition_samples; - mm_sum = _mm_setzero_si128(); e1 = (residual_sample + 3) & ~3; e3 = end & ~3; if(e1 > end) @@ -70,19 +70,19 @@ void FLAC__precompute_partition_info_sums_intrin_ssse3(const FLAC__int32 residua /* assumption: residual[] is properly aligned so (residual + e1) is properly aligned too and _mm_loadu_si128() is fast */ for( ; residual_sample < e1; residual_sample++) { - mm_res = _mm_cvtsi32_si128(residual[residual_sample]); + __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); mm_res = _mm_abs_epi32(mm_res); /* abs(INT_MIN) is undefined, but if the residual is INT_MIN we have bigger problems */ mm_sum = _mm_add_epi32(mm_sum, mm_res); } for( ; residual_sample < e3; residual_sample+=4) { - mm_res = _mm_loadu_si128((const __m128i*)(residual+residual_sample)); + __m128i mm_res = _mm_loadu_si128((const __m128i*)(residual+residual_sample)); mm_res = _mm_abs_epi32(mm_res); mm_sum = _mm_add_epi32(mm_sum, mm_res); } for( ; residual_sample < end; residual_sample++) { - mm_res = _mm_cvtsi32_si128(residual[residual_sample]); + __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); mm_res = _mm_abs_epi32(mm_res); mm_sum = _mm_add_epi32(mm_sum, mm_res); } @@ -94,27 +94,28 @@ void FLAC__precompute_partition_info_sums_intrin_ssse3(const FLAC__int32 residua } else { /* have to pessimistically use 64 bits for accumulator */ for(partition = residual_sample = 0; partition < partitions; partition++) { + __m128i mm_sum = _mm_setzero_si128(); + unsigned e1, e3; end += default_partition_samples; - mm_sum = _mm_setzero_si128(); e1 = (residual_sample + 1) & ~1; e3 = end & ~1; FLAC__ASSERT(e1 <= end); for( ; residual_sample < e1; residual_sample++) { - mm_res = _mm_cvtsi32_si128(residual[residual_sample]); /* 0 0 0 r0 */ + __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); /* 0 0 0 r0 */ mm_res = _mm_abs_epi32(mm_res); /* 0 0 0 |r0| == 00 |r0_64| */ mm_sum = _mm_add_epi64(mm_sum, mm_res); } for( ; residual_sample < e3; residual_sample+=2) { - mm_res = _mm_loadl_epi64((const __m128i*)(residual+residual_sample)); /* 0 0 r1 r0 */ + __m128i mm_res = _mm_loadl_epi64((const __m128i*)(residual+residual_sample)); /* 0 0 r1 r0 */ mm_res = _mm_abs_epi32(mm_res); /* 0 0 |r1| |r0| */ mm_res = _mm_shuffle_epi32(mm_res, _MM_SHUFFLE(3,1,2,0)); /* 0 |r1| 0 |r0| == |r1_64| |r0_64| */ mm_sum = _mm_add_epi64(mm_sum, mm_res); } for( ; residual_sample < end; residual_sample++) { - mm_res = _mm_cvtsi32_si128(residual[residual_sample]); + __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); mm_res = _mm_abs_epi32(mm_res); mm_sum = _mm_add_epi64(mm_sum, mm_res); }