diff --git a/src/libFLAC/stream_encoder_intrin_avx2.c b/src/libFLAC/stream_encoder_intrin_avx2.c index 4381bdb..d0cd0e4 100644 --- a/src/libFLAC/stream_encoder_intrin_avx2.c +++ b/src/libFLAC/stream_encoder_intrin_avx2.c @@ -77,8 +77,7 @@ void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual } for( ; residual_sample < end; residual_sample++) { - __m128i res128 = _mm_cvtsi32_si128(residual[residual_sample]); - res128 = _mm_abs_epi32(res128); + __m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); sum128 = _mm_add_epi32(sum128, res128); } @@ -102,15 +101,13 @@ void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual 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_loadl_epi64((const __m128i*)(residual+residual_sample)); - res128 = _mm_abs_epi32(res128); + __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_cvtsi32_si128(residual[residual_sample]); - res128 = _mm_abs_epi32(res128); + __m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); 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 be8395a..ed01978 100644 --- a/src/libFLAC/stream_encoder_intrin_sse2.c +++ b/src/libFLAC/stream_encoder_intrin_sse2.c @@ -43,6 +43,17 @@ #include /* for abs() */ #include /* SSE2 */ #include "FLAC/assert.h" +#include "share/compat.h" + +FLAC__SSE_TARGET("sse2") +static inline __m128i local_abs_epi32(__m128i val) +{ + __m128i mask = _mm_srai_epi32(val, 31); + val = _mm_xor_si128(val, mask); + val = _mm_sub_epi32(val, mask); + return val; +} + FLAC__SSE_TARGET("sse2") void FLAC__precompute_partition_info_sums_intrin_sse2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[], @@ -70,26 +81,17 @@ 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++) { - __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 */ + __m128i mm_res = local_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); mm_sum = _mm_add_epi32(mm_sum, mm_res); } for( ; residual_sample < e3; residual_sample+=4) { - __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); + __m128i mm_res = local_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample))); mm_sum = _mm_add_epi32(mm_sum, mm_res); } for( ; residual_sample < end; residual_sample++) { - __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); + __m128i mm_res = local_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); mm_sum = _mm_add_epi32(mm_sum, mm_res); } @@ -108,27 +110,18 @@ void FLAC__precompute_partition_info_sums_intrin_sse2(const FLAC__int32 residual FLAC__ASSERT(e1 <= end); for( ; residual_sample < e1; residual_sample++) { - __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| */ + __m128i mm_res = local_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); /* 0 0 0 |r0| == 00 |r0_64| */ mm_sum = _mm_add_epi64(mm_sum, mm_res); } for( ; residual_sample < e3; residual_sample+=2) { - __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| */ + __m128i mm_res = local_abs_epi32(_mm_loadl_epi64((const __m128i*)(residual+residual_sample))); /* 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++) { - __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); + __m128i mm_res = local_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); 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 547ef95..465950b 100644 --- a/src/libFLAC/stream_encoder_intrin_ssse3.c +++ b/src/libFLAC/stream_encoder_intrin_ssse3.c @@ -70,20 +70,17 @@ 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++) { - __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 */ + __m128i mm_res = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); mm_sum = _mm_add_epi32(mm_sum, mm_res); } for( ; residual_sample < e3; residual_sample+=4) { - __m128i mm_res = _mm_loadu_si128((const __m128i*)(residual+residual_sample)); - mm_res = _mm_abs_epi32(mm_res); + __m128i mm_res = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample))); mm_sum = _mm_add_epi32(mm_sum, mm_res); } for( ; residual_sample < end; residual_sample++) { - __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); - mm_res = _mm_abs_epi32(mm_res); + __m128i mm_res = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); mm_sum = _mm_add_epi32(mm_sum, mm_res); } @@ -102,21 +99,18 @@ void FLAC__precompute_partition_info_sums_intrin_ssse3(const FLAC__int32 residua FLAC__ASSERT(e1 <= end); for( ; residual_sample < e1; residual_sample++) { - __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| */ + __m128i mm_res = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); /* 0 0 0 |r0| == 00 |r0_64| */ mm_sum = _mm_add_epi64(mm_sum, mm_res); } for( ; residual_sample < e3; residual_sample+=2) { - __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| */ + __m128i mm_res = _mm_abs_epi32(_mm_loadl_epi64((const __m128i*)(residual+residual_sample))); /* 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++) { - __m128i mm_res = _mm_cvtsi32_si128(residual[residual_sample]); - mm_res = _mm_abs_epi32(mm_res); + __m128i mm_res = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample])); mm_sum = _mm_add_epi64(mm_sum, mm_res); }