From 11b5d51b63710efbfaf913ad08570c8d9c1db36a Mon Sep 17 00:00:00 2001 From: Erik de Castro Lopo Date: Sat, 4 Oct 2014 09:42:54 +1000 Subject: Add AVX2 CPU support stream encoder. Patch-from: lvqcl --- src/libFLAC/Makefile.am | 2 + src/libFLAC/Makefile.lite | 2 + src/libFLAC/include/private/lpc.h | 5 + src/libFLAC/include/private/stream_encoder.h | 5 + src/libFLAC/libFLAC_dynamic.vcxproj | 12 + src/libFLAC/libFLAC_dynamic.vcxproj.filters | 6 + src/libFLAC/libFLAC_static.vcxproj | 12 + src/libFLAC/libFLAC_static.vcxproj.filters | 6 + src/libFLAC/lpc_intrin_avx2.c | 1120 ++++++++++++++++++++++++++ src/libFLAC/stream_encoder.c | 22 + src/libFLAC/stream_encoder_intrin_avx2.c | 142 ++++ 11 files changed, 1334 insertions(+) create mode 100644 src/libFLAC/lpc_intrin_avx2.c create mode 100644 src/libFLAC/stream_encoder_intrin_avx2.c diff --git a/src/libFLAC/Makefile.am b/src/libFLAC/Makefile.am index 473473d3..a5d0d48e 100644 --- a/src/libFLAC/Makefile.am +++ b/src/libFLAC/Makefile.am @@ -112,6 +112,7 @@ libFLAC_sources = \ lpc_intrin_sse.c \ lpc_intrin_sse2.c \ lpc_intrin_sse41.c \ + lpc_intrin_avx2.c \ md5.c \ memory.c \ metadata_iterators.c \ @@ -120,6 +121,7 @@ libFLAC_sources = \ stream_encoder.c \ stream_encoder_intrin_sse2.c \ stream_encoder_intrin_ssse3.c \ + stream_encoder_intrin_avx2.c \ stream_encoder_framing.c \ window.c \ $(extra_ogg_sources) diff --git a/src/libFLAC/Makefile.lite b/src/libFLAC/Makefile.lite index b1eee24a..a9f1d845 100644 --- a/src/libFLAC/Makefile.lite +++ b/src/libFLAC/Makefile.lite @@ -90,6 +90,7 @@ SRCS_C = \ lpc_intrin_sse.c \ lpc_intrin_sse2.c \ lpc_intrin_sse41.c \ + lpc_intrin_avx2.c \ md5.c \ memory.c \ metadata_iterators.c \ @@ -98,6 +99,7 @@ SRCS_C = \ stream_encoder.c \ stream_encoder_intrin_sse2.c \ stream_encoder_intrin_ssse3.c \ + stream_encoder_intrin_avx2.c \ stream_encoder_framing.c \ window.c \ $(OGG_SRCS) diff --git a/src/libFLAC/include/private/lpc.h b/src/libFLAC/include/private/lpc.h index 2aa604f5..7b3652fd 100644 --- a/src/libFLAC/include/private/lpc.h +++ b/src/libFLAC/include/private/lpc.h @@ -166,6 +166,11 @@ void FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse2(const FLAC__in void FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[]); void FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[]); # endif +# ifdef FLAC__AVX2_SUPPORTED +void FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_avx2(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[]); +void FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_avx2(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[]); +void FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[]); +# endif # endif #endif diff --git a/src/libFLAC/include/private/stream_encoder.h b/src/libFLAC/include/private/stream_encoder.h index 8147f9ed..3d231054 100644 --- a/src/libFLAC/include/private/stream_encoder.h +++ b/src/libFLAC/include/private/stream_encoder.h @@ -57,6 +57,11 @@ extern void FLAC__precompute_partition_info_sums_intrin_ssse3(const FLAC__int32 unsigned residual_samples, unsigned predictor_order, unsigned min_partition_order, unsigned max_partition_order, unsigned bps); #endif +#ifdef FLAC__AVX2_SUPPORTED +extern void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[], + unsigned residual_samples, unsigned predictor_order, unsigned min_partition_order, unsigned max_partition_order, unsigned bps); +#endif + #endif #endif diff --git a/src/libFLAC/libFLAC_dynamic.vcxproj b/src/libFLAC/libFLAC_dynamic.vcxproj index f15b2b57..58599dea 100644 --- a/src/libFLAC/libFLAC_dynamic.vcxproj +++ b/src/libFLAC/libFLAC_dynamic.vcxproj @@ -221,6 +221,12 @@ + + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + @@ -235,6 +241,12 @@ + + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + diff --git a/src/libFLAC/libFLAC_dynamic.vcxproj.filters b/src/libFLAC/libFLAC_dynamic.vcxproj.filters index bc18b149..fa79a8dc 100644 --- a/src/libFLAC/libFLAC_dynamic.vcxproj.filters +++ b/src/libFLAC/libFLAC_dynamic.vcxproj.filters @@ -157,6 +157,9 @@ Source Files + + Source Files + Source Files @@ -196,6 +199,9 @@ Source Files + + Source Files + Source Files diff --git a/src/libFLAC/libFLAC_static.vcxproj b/src/libFLAC/libFLAC_static.vcxproj index 3487d4a2..da5719de 100644 --- a/src/libFLAC/libFLAC_static.vcxproj +++ b/src/libFLAC/libFLAC_static.vcxproj @@ -181,6 +181,12 @@ + + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + @@ -195,6 +201,12 @@ + + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + /arch:AVX %(AdditionalOptions) + diff --git a/src/libFLAC/libFLAC_static.vcxproj.filters b/src/libFLAC/libFLAC_static.vcxproj.filters index d5c82589..bc99bf3b 100644 --- a/src/libFLAC/libFLAC_static.vcxproj.filters +++ b/src/libFLAC/libFLAC_static.vcxproj.filters @@ -157,6 +157,9 @@ Source Files + + Source Files + Source Files @@ -196,6 +199,9 @@ Source Files + + Source Files + Source Files diff --git a/src/libFLAC/lpc_intrin_avx2.c b/src/libFLAC/lpc_intrin_avx2.c new file mode 100644 index 00000000..8eec85e1 --- /dev/null +++ b/src/libFLAC/lpc_intrin_avx2.c @@ -0,0 +1,1120 @@ +/* libFLAC - Free Lossless Audio Codec library + * Copyright (C) 2000-2009 Josh Coalson + * Copyright (C) 2011-2014 Xiph.Org Foundation + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * - Neither the name of the Xiph.org Foundation nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR + * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, + * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#ifdef HAVE_CONFIG_H +# include +#endif + +#ifndef FLAC__INTEGER_ONLY_LIBRARY +#ifndef FLAC__NO_ASM +#if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && defined FLAC__HAS_X86INTRIN +#include "private/lpc.h" +#ifdef FLAC__AVX2_SUPPORTED + +#include "FLAC/assert.h" +#include "FLAC/format.h" + +#include /* AVX2 */ + +FLAC__SSE_TARGET("avx2") +void FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_avx2(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[]) +{ + int i; + FLAC__int32 sum; + __m128i cnt = _mm_cvtsi32_si128(lp_quantization); + + FLAC__ASSERT(order > 0); + FLAC__ASSERT(order <= 32); + + if(order <= 12) { + if(order > 8) { + if(order > 10) { + if(order == 12) { + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(0xffff & qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(0xffff & qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(0xffff & qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(0xffff & qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(0xffff & qlp_coeff[7 ]); + q8 = _mm256_set1_epi32(0xffff & qlp_coeff[8 ]); + q9 = _mm256_set1_epi32(0xffff & qlp_coeff[9 ]); + q10 = _mm256_set1_epi32(0xffff & qlp_coeff[10]); + q11 = _mm256_set1_epi32(0xffff & qlp_coeff[11]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q11, _mm256_loadu_si256((const __m256i*)(data+i-12))); + mull = _mm256_madd_epi16(q10, _mm256_loadu_si256((const __m256i*)(data+i-11))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q9, _mm256_loadu_si256((const __m256i*)(data+i-10))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q8, _mm256_loadu_si256((const __m256i*)(data+i-9 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q7, _mm256_loadu_si256((const __m256i*)(data+i-8 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q6, _mm256_loadu_si256((const __m256i*)(data+i-7 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q5, _mm256_loadu_si256((const __m256i*)(data+i-6 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q4, _mm256_loadu_si256((const __m256i*)(data+i-5 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q3, _mm256_loadu_si256((const __m256i*)(data+i-4 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 11 */ + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(0xffff & qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(0xffff & qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(0xffff & qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(0xffff & qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(0xffff & qlp_coeff[7 ]); + q8 = _mm256_set1_epi32(0xffff & qlp_coeff[8 ]); + q9 = _mm256_set1_epi32(0xffff & qlp_coeff[9 ]); + q10 = _mm256_set1_epi32(0xffff & qlp_coeff[10]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q10, _mm256_loadu_si256((const __m256i*)(data+i-11))); + mull = _mm256_madd_epi16(q9, _mm256_loadu_si256((const __m256i*)(data+i-10))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q8, _mm256_loadu_si256((const __m256i*)(data+i-9 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q7, _mm256_loadu_si256((const __m256i*)(data+i-8 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q6, _mm256_loadu_si256((const __m256i*)(data+i-7 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q5, _mm256_loadu_si256((const __m256i*)(data+i-6 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q4, _mm256_loadu_si256((const __m256i*)(data+i-5 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q3, _mm256_loadu_si256((const __m256i*)(data+i-4 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + else { + if(order == 10) { + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(0xffff & qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(0xffff & qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(0xffff & qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(0xffff & qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(0xffff & qlp_coeff[7 ]); + q8 = _mm256_set1_epi32(0xffff & qlp_coeff[8 ]); + q9 = _mm256_set1_epi32(0xffff & qlp_coeff[9 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q9, _mm256_loadu_si256((const __m256i*)(data+i-10))); + mull = _mm256_madd_epi16(q8, _mm256_loadu_si256((const __m256i*)(data+i-9 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q7, _mm256_loadu_si256((const __m256i*)(data+i-8 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q6, _mm256_loadu_si256((const __m256i*)(data+i-7 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q5, _mm256_loadu_si256((const __m256i*)(data+i-6 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q4, _mm256_loadu_si256((const __m256i*)(data+i-5 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q3, _mm256_loadu_si256((const __m256i*)(data+i-4 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 9 */ + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(0xffff & qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(0xffff & qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(0xffff & qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(0xffff & qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(0xffff & qlp_coeff[7 ]); + q8 = _mm256_set1_epi32(0xffff & qlp_coeff[8 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q8, _mm256_loadu_si256((const __m256i*)(data+i-9 ))); + mull = _mm256_madd_epi16(q7, _mm256_loadu_si256((const __m256i*)(data+i-8 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q6, _mm256_loadu_si256((const __m256i*)(data+i-7 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q5, _mm256_loadu_si256((const __m256i*)(data+i-6 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q4, _mm256_loadu_si256((const __m256i*)(data+i-5 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q3, _mm256_loadu_si256((const __m256i*)(data+i-4 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + } + else if(order > 4) { + if(order > 6) { + if(order == 8) { + __m256i q0, q1, q2, q3, q4, q5, q6, q7; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(0xffff & qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(0xffff & qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(0xffff & qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(0xffff & qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(0xffff & qlp_coeff[7 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q7, _mm256_loadu_si256((const __m256i*)(data+i-8 ))); + mull = _mm256_madd_epi16(q6, _mm256_loadu_si256((const __m256i*)(data+i-7 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q5, _mm256_loadu_si256((const __m256i*)(data+i-6 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q4, _mm256_loadu_si256((const __m256i*)(data+i-5 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q3, _mm256_loadu_si256((const __m256i*)(data+i-4 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 7 */ + __m256i q0, q1, q2, q3, q4, q5, q6; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(0xffff & qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(0xffff & qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(0xffff & qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(0xffff & qlp_coeff[6 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q6, _mm256_loadu_si256((const __m256i*)(data+i-7 ))); + mull = _mm256_madd_epi16(q5, _mm256_loadu_si256((const __m256i*)(data+i-6 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q4, _mm256_loadu_si256((const __m256i*)(data+i-5 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q3, _mm256_loadu_si256((const __m256i*)(data+i-4 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + else { + if(order == 6) { + __m256i q0, q1, q2, q3, q4, q5; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(0xffff & qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(0xffff & qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(0xffff & qlp_coeff[5 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q5, _mm256_loadu_si256((const __m256i*)(data+i-6 ))); + mull = _mm256_madd_epi16(q4, _mm256_loadu_si256((const __m256i*)(data+i-5 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q3, _mm256_loadu_si256((const __m256i*)(data+i-4 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 5 */ + __m256i q0, q1, q2, q3, q4; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(0xffff & qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(0xffff & qlp_coeff[4 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q4, _mm256_loadu_si256((const __m256i*)(data+i-5 ))); + mull = _mm256_madd_epi16(q3, _mm256_loadu_si256((const __m256i*)(data+i-4 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + } + else { + if(order > 2) { + if(order == 4) { + __m256i q0, q1, q2, q3; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(0xffff & qlp_coeff[3 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q3, _mm256_loadu_si256((const __m256i*)(data+i-4 ))); + mull = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 3 */ + __m256i q0, q1, q2; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(0xffff & qlp_coeff[2 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q2, _mm256_loadu_si256((const __m256i*)(data+i-3 ))); + mull = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + else { + if(order == 2) { + __m256i q0, q1; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(0xffff & qlp_coeff[1 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_madd_epi16(q1, _mm256_loadu_si256((const __m256i*)(data+i-2 ))); + mull = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 1 */ + __m256i q0; + q0 = _mm256_set1_epi32(0xffff & qlp_coeff[0 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ; + summ = _mm256_madd_epi16(q0, _mm256_loadu_si256((const __m256i*)(data+i-1 ))); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + } + for(; i < (int)data_len; i++) { + sum = 0; + switch(order) { + case 12: sum += qlp_coeff[11] * data[i-12]; + case 11: sum += qlp_coeff[10] * data[i-11]; + case 10: sum += qlp_coeff[ 9] * data[i-10]; + case 9: sum += qlp_coeff[ 8] * data[i- 9]; + case 8: sum += qlp_coeff[ 7] * data[i- 8]; + case 7: sum += qlp_coeff[ 6] * data[i- 7]; + case 6: sum += qlp_coeff[ 5] * data[i- 6]; + case 5: sum += qlp_coeff[ 4] * data[i- 5]; + case 4: sum += qlp_coeff[ 3] * data[i- 4]; + case 3: sum += qlp_coeff[ 2] * data[i- 3]; + case 2: sum += qlp_coeff[ 1] * data[i- 2]; + case 1: sum += qlp_coeff[ 0] * data[i- 1]; + } + residual[i] = data[i] - (sum >> lp_quantization); + } + } + else { /* order > 12 */ + for(i = 0; i < (int)data_len; i++) { + sum = 0; + switch(order) { + case 32: sum += qlp_coeff[31] * data[i-32]; + case 31: sum += qlp_coeff[30] * data[i-31]; + case 30: sum += qlp_coeff[29] * data[i-30]; + case 29: sum += qlp_coeff[28] * data[i-29]; + case 28: sum += qlp_coeff[27] * data[i-28]; + case 27: sum += qlp_coeff[26] * data[i-27]; + case 26: sum += qlp_coeff[25] * data[i-26]; + case 25: sum += qlp_coeff[24] * data[i-25]; + case 24: sum += qlp_coeff[23] * data[i-24]; + case 23: sum += qlp_coeff[22] * data[i-23]; + case 22: sum += qlp_coeff[21] * data[i-22]; + case 21: sum += qlp_coeff[20] * data[i-21]; + case 20: sum += qlp_coeff[19] * data[i-20]; + case 19: sum += qlp_coeff[18] * data[i-19]; + case 18: sum += qlp_coeff[17] * data[i-18]; + case 17: sum += qlp_coeff[16] * data[i-17]; + case 16: sum += qlp_coeff[15] * data[i-16]; + case 15: sum += qlp_coeff[14] * data[i-15]; + case 14: sum += qlp_coeff[13] * data[i-14]; + case 13: sum += qlp_coeff[12] * data[i-13]; + sum += qlp_coeff[11] * data[i-12]; + sum += qlp_coeff[10] * data[i-11]; + sum += qlp_coeff[ 9] * data[i-10]; + sum += qlp_coeff[ 8] * data[i- 9]; + sum += qlp_coeff[ 7] * data[i- 8]; + sum += qlp_coeff[ 6] * data[i- 7]; + sum += qlp_coeff[ 5] * data[i- 6]; + sum += qlp_coeff[ 4] * data[i- 5]; + sum += qlp_coeff[ 3] * data[i- 4]; + sum += qlp_coeff[ 2] * data[i- 3]; + sum += qlp_coeff[ 1] * data[i- 2]; + sum += qlp_coeff[ 0] * data[i- 1]; + } + residual[i] = data[i] - (sum >> lp_quantization); + } + } + _mm256_zeroupper(); +} + +FLAC__SSE_TARGET("avx2") +void FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_avx2(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[]) +{ + int i; + FLAC__int32 sum; + __m128i cnt = _mm_cvtsi32_si128(lp_quantization); + + FLAC__ASSERT(order > 0); + FLAC__ASSERT(order <= 32); + + if(order <= 12) { + if(order > 8) { + if(order > 10) { + if(order == 12) { + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(qlp_coeff[7 ]); + q8 = _mm256_set1_epi32(qlp_coeff[8 ]); + q9 = _mm256_set1_epi32(qlp_coeff[9 ]); + q10 = _mm256_set1_epi32(qlp_coeff[10]); + q11 = _mm256_set1_epi32(qlp_coeff[11]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q11, _mm256_loadu_si256((const __m256i*)(data+i-12))); + mull = _mm256_mullo_epi32(q10, _mm256_loadu_si256((const __m256i*)(data+i-11))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q9, _mm256_loadu_si256((const __m256i*)(data+i-10))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q8, _mm256_loadu_si256((const __m256i*)(data+i-9))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q7, _mm256_loadu_si256((const __m256i*)(data+i-8))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q6, _mm256_loadu_si256((const __m256i*)(data+i-7))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q5, _mm256_loadu_si256((const __m256i*)(data+i-6))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q4, _mm256_loadu_si256((const __m256i*)(data+i-5))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q3, _mm256_loadu_si256((const __m256i*)(data+i-4))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 11 */ + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(qlp_coeff[7 ]); + q8 = _mm256_set1_epi32(qlp_coeff[8 ]); + q9 = _mm256_set1_epi32(qlp_coeff[9 ]); + q10 = _mm256_set1_epi32(qlp_coeff[10]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q10, _mm256_loadu_si256((const __m256i*)(data+i-11))); + mull = _mm256_mullo_epi32(q9, _mm256_loadu_si256((const __m256i*)(data+i-10))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q8, _mm256_loadu_si256((const __m256i*)(data+i-9))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q7, _mm256_loadu_si256((const __m256i*)(data+i-8))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q6, _mm256_loadu_si256((const __m256i*)(data+i-7))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q5, _mm256_loadu_si256((const __m256i*)(data+i-6))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q4, _mm256_loadu_si256((const __m256i*)(data+i-5))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q3, _mm256_loadu_si256((const __m256i*)(data+i-4))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + else { + if(order == 10) { + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(qlp_coeff[7 ]); + q8 = _mm256_set1_epi32(qlp_coeff[8 ]); + q9 = _mm256_set1_epi32(qlp_coeff[9 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q9, _mm256_loadu_si256((const __m256i*)(data+i-10))); + mull = _mm256_mullo_epi32(q8, _mm256_loadu_si256((const __m256i*)(data+i-9))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q7, _mm256_loadu_si256((const __m256i*)(data+i-8))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q6, _mm256_loadu_si256((const __m256i*)(data+i-7))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q5, _mm256_loadu_si256((const __m256i*)(data+i-6))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q4, _mm256_loadu_si256((const __m256i*)(data+i-5))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q3, _mm256_loadu_si256((const __m256i*)(data+i-4))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 9 */ + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(qlp_coeff[7 ]); + q8 = _mm256_set1_epi32(qlp_coeff[8 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q8, _mm256_loadu_si256((const __m256i*)(data+i-9))); + mull = _mm256_mullo_epi32(q7, _mm256_loadu_si256((const __m256i*)(data+i-8))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q6, _mm256_loadu_si256((const __m256i*)(data+i-7))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q5, _mm256_loadu_si256((const __m256i*)(data+i-6))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q4, _mm256_loadu_si256((const __m256i*)(data+i-5))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q3, _mm256_loadu_si256((const __m256i*)(data+i-4))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + } + else if(order > 4) { + if(order > 6) { + if(order == 8) { + __m256i q0, q1, q2, q3, q4, q5, q6, q7; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(qlp_coeff[6 ]); + q7 = _mm256_set1_epi32(qlp_coeff[7 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q7, _mm256_loadu_si256((const __m256i*)(data+i-8))); + mull = _mm256_mullo_epi32(q6, _mm256_loadu_si256((const __m256i*)(data+i-7))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q5, _mm256_loadu_si256((const __m256i*)(data+i-6))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q4, _mm256_loadu_si256((const __m256i*)(data+i-5))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q3, _mm256_loadu_si256((const __m256i*)(data+i-4))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 7 */ + __m256i q0, q1, q2, q3, q4, q5, q6; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(qlp_coeff[5 ]); + q6 = _mm256_set1_epi32(qlp_coeff[6 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q6, _mm256_loadu_si256((const __m256i*)(data+i-7))); + mull = _mm256_mullo_epi32(q5, _mm256_loadu_si256((const __m256i*)(data+i-6))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q4, _mm256_loadu_si256((const __m256i*)(data+i-5))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q3, _mm256_loadu_si256((const __m256i*)(data+i-4))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + else { + if(order == 6) { + __m256i q0, q1, q2, q3, q4, q5; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(qlp_coeff[4 ]); + q5 = _mm256_set1_epi32(qlp_coeff[5 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q5, _mm256_loadu_si256((const __m256i*)(data+i-6))); + mull = _mm256_mullo_epi32(q4, _mm256_loadu_si256((const __m256i*)(data+i-5))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q3, _mm256_loadu_si256((const __m256i*)(data+i-4))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 5 */ + __m256i q0, q1, q2, q3, q4; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(qlp_coeff[3 ]); + q4 = _mm256_set1_epi32(qlp_coeff[4 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q4, _mm256_loadu_si256((const __m256i*)(data+i-5))); + mull = _mm256_mullo_epi32(q3, _mm256_loadu_si256((const __m256i*)(data+i-4))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + } + else { + if(order > 2) { + if(order == 4) { + __m256i q0, q1, q2, q3; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + q3 = _mm256_set1_epi32(qlp_coeff[3 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q3, _mm256_loadu_si256((const __m256i*)(data+i-4))); + mull = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 3 */ + __m256i q0, q1, q2; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + q2 = _mm256_set1_epi32(qlp_coeff[2 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q2, _mm256_loadu_si256((const __m256i*)(data+i-3))); + mull = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); summ = _mm256_add_epi32(summ, mull); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + else { + if(order == 2) { + __m256i q0, q1; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + q1 = _mm256_set1_epi32(qlp_coeff[1 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ, mull; + summ = _mm256_mullo_epi32(q1, _mm256_loadu_si256((const __m256i*)(data+i-2))); + mull = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); summ = _mm256_add_epi32(summ, mull); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + else { /* order == 1 */ + __m256i q0; + q0 = _mm256_set1_epi32(qlp_coeff[0 ]); + + for(i = 0; i < (int)data_len-7; i+=8) { + __m256i summ; + summ = _mm256_mullo_epi32(q0, _mm256_loadu_si256((const __m256i*)(data+i-1))); + summ = _mm256_sra_epi32(summ, cnt); + _mm256_storeu_si256((__m256i*)(residual+i), _mm256_sub_epi32(_mm256_loadu_si256((const __m256i*)(data+i)), summ)); + } + } + } + } + for(; i < (int)data_len; i++) { + sum = 0; + switch(order) { + case 12: sum += qlp_coeff[11] * data[i-12]; + case 11: sum += qlp_coeff[10] * data[i-11]; + case 10: sum += qlp_coeff[ 9] * data[i-10]; + case 9: sum += qlp_coeff[ 8] * data[i- 9]; + case 8: sum += qlp_coeff[ 7] * data[i- 8]; + case 7: sum += qlp_coeff[ 6] * data[i- 7]; + case 6: sum += qlp_coeff[ 5] * data[i- 6]; + case 5: sum += qlp_coeff[ 4] * data[i- 5]; + case 4: sum += qlp_coeff[ 3] * data[i- 4]; + case 3: sum += qlp_coeff[ 2] * data[i- 3]; + case 2: sum += qlp_coeff[ 1] * data[i- 2]; + case 1: sum += qlp_coeff[ 0] * data[i- 1]; + } + residual[i] = data[i] - (sum >> lp_quantization); + } + } + else { /* order > 12 */ + for(i = 0; i < (int)data_len; i++) { + sum = 0; + switch(order) { + case 32: sum += qlp_coeff[31] * data[i-32]; + case 31: sum += qlp_coeff[30] * data[i-31]; + case 30: sum += qlp_coeff[29] * data[i-30]; + case 29: sum += qlp_coeff[28] * data[i-29]; + case 28: sum += qlp_coeff[27] * data[i-28]; + case 27: sum += qlp_coeff[26] * data[i-27]; + case 26: sum += qlp_coeff[25] * data[i-26]; + case 25: sum += qlp_coeff[24] * data[i-25]; + case 24: sum += qlp_coeff[23] * data[i-24]; + case 23: sum += qlp_coeff[22] * data[i-23]; + case 22: sum += qlp_coeff[21] * data[i-22]; + case 21: sum += qlp_coeff[20] * data[i-21]; + case 20: sum += qlp_coeff[19] * data[i-20]; + case 19: sum += qlp_coeff[18] * data[i-19]; + case 18: sum += qlp_coeff[17] * data[i-18]; + case 17: sum += qlp_coeff[16] * data[i-17]; + case 16: sum += qlp_coeff[15] * data[i-16]; + case 15: sum += qlp_coeff[14] * data[i-15]; + case 14: sum += qlp_coeff[13] * data[i-14]; + case 13: sum += qlp_coeff[12] * data[i-13]; + sum += qlp_coeff[11] * data[i-12]; + sum += qlp_coeff[10] * data[i-11]; + sum += qlp_coeff[ 9] * data[i-10]; + sum += qlp_coeff[ 8] * data[i- 9]; + sum += qlp_coeff[ 7] * data[i- 8]; + sum += qlp_coeff[ 6] * data[i- 7]; + sum += qlp_coeff[ 5] * data[i- 6]; + sum += qlp_coeff[ 4] * data[i- 5]; + sum += qlp_coeff[ 3] * data[i- 4]; + sum += qlp_coeff[ 2] * data[i- 3]; + sum += qlp_coeff[ 1] * data[i- 2]; + sum += qlp_coeff[ 0] * data[i- 1]; + } + residual[i] = data[i] - (sum >> lp_quantization); + } + } + _mm256_zeroupper(); +} + +static FLAC__int32 pack_arr[8] = { 0, 2, 4, 6, 1, 3, 5, 7 }; + +FLAC__SSE_TARGET("avx2") +void FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2(const FLAC__int32 *data, unsigned data_len, const FLAC__int32 qlp_coeff[], unsigned order, int lp_quantization, FLAC__int32 residual[]) +{ + int i; + FLAC__int64 sum; + __m128i cnt = _mm_cvtsi32_si128(lp_quantization); + __m256i pack = _mm256_loadu_si256((const __m256i *)pack_arr); + + FLAC__ASSERT(order > 0); + FLAC__ASSERT(order <= 32); + FLAC__ASSERT(lp_quantization <= 32); /* there's no _mm256_sra_epi64() so we have to use _mm256_srl_epi64() */ + + if(order <= 12) { + if(order > 8) { + if(order > 10) { + if(order == 12) { + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10, q11; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + q3 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[3 ])); + q4 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[4 ])); + q5 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[5 ])); + q6 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[6 ])); + q7 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[7 ])); + q8 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[8 ])); + q9 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[9 ])); + q10 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[10])); + q11 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[11])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q11, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-12)))); + mull = _mm256_mul_epi32(q10, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-11)))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q9, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-10)))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q8, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-9 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q7, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-8 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q6, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-7 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q5, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-6 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q4, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-5 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + else { /* order == 11 */ + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9, q10; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + q3 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[3 ])); + q4 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[4 ])); + q5 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[5 ])); + q6 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[6 ])); + q7 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[7 ])); + q8 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[8 ])); + q9 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[9 ])); + q10 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[10])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q10, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-11)))); + mull = _mm256_mul_epi32(q9, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-10)))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q8, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-9 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q7, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-8 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q6, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-7 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q5, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-6 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q4, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-5 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + } + else { + if(order == 10) { + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8, q9; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + q3 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[3 ])); + q4 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[4 ])); + q5 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[5 ])); + q6 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[6 ])); + q7 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[7 ])); + q8 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[8 ])); + q9 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[9 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q9, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-10)))); + mull = _mm256_mul_epi32(q8, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-9 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q7, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-8 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q6, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-7 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q5, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-6 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q4, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-5 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + else { /* order == 9 */ + __m256i q0, q1, q2, q3, q4, q5, q6, q7, q8; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + q3 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[3 ])); + q4 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[4 ])); + q5 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[5 ])); + q6 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[6 ])); + q7 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[7 ])); + q8 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[8 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q8, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-9 )))); + mull = _mm256_mul_epi32(q7, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-8 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q6, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-7 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q5, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-6 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q4, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-5 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + } + } + else if(order > 4) { + if(order > 6) { + if(order == 8) { + __m256i q0, q1, q2, q3, q4, q5, q6, q7; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + q3 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[3 ])); + q4 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[4 ])); + q5 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[5 ])); + q6 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[6 ])); + q7 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[7 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q7, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-8 )))); + mull = _mm256_mul_epi32(q6, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-7 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q5, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-6 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q4, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-5 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + else { /* order == 7 */ + __m256i q0, q1, q2, q3, q4, q5, q6; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + q3 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[3 ])); + q4 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[4 ])); + q5 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[5 ])); + q6 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[6 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q6, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-7 )))); + mull = _mm256_mul_epi32(q5, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-6 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q4, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-5 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + } + else { + if(order == 6) { + __m256i q0, q1, q2, q3, q4, q5; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + q3 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[3 ])); + q4 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[4 ])); + q5 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[5 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q5, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-6 )))); + mull = _mm256_mul_epi32(q4, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-5 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + else { /* order == 5 */ + __m256i q0, q1, q2, q3, q4; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + q3 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[3 ])); + q4 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[4 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q4, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-5 )))); + mull = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + } + } + else { + if(order > 2) { + if(order == 4) { + __m256i q0, q1, q2, q3; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + q3 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[3 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q3, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-4 )))); + mull = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + else { /* order == 3 */ + __m256i q0, q1, q2; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + q2 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[2 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q2, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-3 )))); + mull = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); summ = _mm256_add_epi64(summ, mull); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + } + else { + if(order == 2) { + __m256i q0, q1; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + q1 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[1 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ, mull; + summ = _mm256_mul_epi32(q1, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-2 )))); + mull = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); summ = _mm256_add_epi64(summ, mull); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + else { /* order == 1 */ + __m256i q0; + q0 = _mm256_cvtepu32_epi64(_mm_set1_epi32(qlp_coeff[0 ])); + + for(i = 0; i < (int)data_len-3; i+=4) { + __m256i summ; + summ = _mm256_mul_epi32(q0, _mm256_cvtepu32_epi64(_mm_loadu_si128((const __m128i*)(data+i-1 )))); + summ = _mm256_permutevar8x32_epi32(_mm256_srl_epi64(summ, cnt), pack); + _mm_storeu_si128((__m128i*)(residual+i), _mm_sub_epi32(_mm_loadu_si128((const __m128i*)(data+i)), _mm256_castsi256_si128(summ))); + } + } + } + } + for(; i < (int)data_len; i++) { + sum = 0; + switch(order) { + case 12: sum += qlp_coeff[11] * (FLAC__int64)data[i-12]; + case 11: sum += qlp_coeff[10] * (FLAC__int64)data[i-11]; + case 10: sum += qlp_coeff[ 9] * (FLAC__int64)data[i-10]; + case 9: sum += qlp_coeff[ 8] * (FLAC__int64)data[i- 9]; + case 8: sum += qlp_coeff[ 7] * (FLAC__int64)data[i- 8]; + case 7: sum += qlp_coeff[ 6] * (FLAC__int64)data[i- 7]; + case 6: sum += qlp_coeff[ 5] * (FLAC__int64)data[i- 6]; + case 5: sum += qlp_coeff[ 4] * (FLAC__int64)data[i- 5]; + case 4: sum += qlp_coeff[ 3] * (FLAC__int64)data[i- 4]; + case 3: sum += qlp_coeff[ 2] * (FLAC__int64)data[i- 3]; + case 2: sum += qlp_coeff[ 1] * (FLAC__int64)data[i- 2]; + case 1: sum += qlp_coeff[ 0] * (FLAC__int64)data[i- 1]; + } + residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization); + } + } + else { /* order > 12 */ + for(i = 0; i < (int)data_len; i++) { + sum = 0; + switch(order) { + case 32: sum += qlp_coeff[31] * (FLAC__int64)data[i-32]; + case 31: sum += qlp_coeff[30] * (FLAC__int64)data[i-31]; + case 30: sum += qlp_coeff[29] * (FLAC__int64)data[i-30]; + case 29: sum += qlp_coeff[28] * (FLAC__int64)data[i-29]; + case 28: sum += qlp_coeff[27] * (FLAC__int64)data[i-28]; + case 27: sum += qlp_coeff[26] * (FLAC__int64)data[i-27]; + case 26: sum += qlp_coeff[25] * (FLAC__int64)data[i-26]; + case 25: sum += qlp_coeff[24] * (FLAC__int64)data[i-25]; + case 24: sum += qlp_coeff[23] * (FLAC__int64)data[i-24]; + case 23: sum += qlp_coeff[22] * (FLAC__int64)data[i-23]; + case 22: sum += qlp_coeff[21] * (FLAC__int64)data[i-22]; + case 21: sum += qlp_coeff[20] * (FLAC__int64)data[i-21]; + case 20: sum += qlp_coeff[19] * (FLAC__int64)data[i-20]; + case 19: sum += qlp_coeff[18] * (FLAC__int64)data[i-19]; + case 18: sum += qlp_coeff[17] * (FLAC__int64)data[i-18]; + case 17: sum += qlp_coeff[16] * (FLAC__int64)data[i-17]; + case 16: sum += qlp_coeff[15] * (FLAC__int64)data[i-16]; + case 15: sum += qlp_coeff[14] * (FLAC__int64)data[i-15]; + case 14: sum += qlp_coeff[13] * (FLAC__int64)data[i-14]; + case 13: sum += qlp_coeff[12] * (FLAC__int64)data[i-13]; + sum += qlp_coeff[11] * (FLAC__int64)data[i-12]; + sum += qlp_coeff[10] * (FLAC__int64)data[i-11]; + sum += qlp_coeff[ 9] * (FLAC__int64)data[i-10]; + sum += qlp_coeff[ 8] * (FLAC__int64)data[i- 9]; + sum += qlp_coeff[ 7] * (FLAC__int64)data[i- 8]; + sum += qlp_coeff[ 6] * (FLAC__int64)data[i- 7]; + sum += qlp_coeff[ 5] * (FLAC__int64)data[i- 6]; + sum += qlp_coeff[ 4] * (FLAC__int64)data[i- 5]; + sum += qlp_coeff[ 3] * (FLAC__int64)data[i- 4]; + sum += qlp_coeff[ 2] * (FLAC__int64)data[i- 3]; + sum += qlp_coeff[ 1] * (FLAC__int64)data[i- 2]; + sum += qlp_coeff[ 0] * (FLAC__int64)data[i- 1]; + } + residual[i] = data[i] - (FLAC__int32)(sum >> lp_quantization); + } + } + _mm256_zeroupper(); +} + +#endif /* FLAC__AVX2_SUPPORTED */ +#endif /* (FLAC__CPU_IA32 || FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN */ +#endif /* FLAC__NO_ASM */ +#endif /* FLAC__INTEGER_ONLY_LIBRARY */ diff --git a/src/libFLAC/stream_encoder.c b/src/libFLAC/stream_encoder.c index b860bfd8..c0430def 100644 --- a/src/libFLAC/stream_encoder.c +++ b/src/libFLAC/stream_encoder.c @@ -950,6 +950,13 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_( encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_64bit = FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_sse41; } # endif +# ifdef FLAC__AVX2_SUPPORTED + if(encoder->private_->cpuinfo.ia32.avx2) { + encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_16bit = FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_avx2; + encoder->private_->local_lpc_compute_residual_from_qlp_coefficients = FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_avx2; + encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_64bit = FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2; + } +# endif # ifdef FLAC__SSE2_SUPPORTED if (encoder->private_->cpuinfo.ia32.sse2) { @@ -986,6 +993,13 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_( encoder->private_->local_lpc_compute_residual_from_qlp_coefficients = FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_sse41; } # endif +# ifdef FLAC__AVX2_SUPPORTED + if(encoder->private_->cpuinfo.x86.avx2) { + encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_16bit = FLAC__lpc_compute_residual_from_qlp_coefficients_16_intrin_avx2; + encoder->private_->local_lpc_compute_residual_from_qlp_coefficients = FLAC__lpc_compute_residual_from_qlp_coefficients_intrin_avx2; + encoder->private_->local_lpc_compute_residual_from_qlp_coefficients_64bit = FLAC__lpc_compute_residual_from_qlp_coefficients_wide_intrin_avx2; + } +# endif # ifdef FLAC__SSE2_SUPPORTED encoder->private_->local_fixed_compute_best_predictor = FLAC__fixed_compute_best_predictor_intrin_sse2; @@ -1013,6 +1027,10 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_( if(encoder->private_->cpuinfo.ia32.ssse3) encoder->private_->local_precompute_partition_info_sums = FLAC__precompute_partition_info_sums_intrin_ssse3; # endif +# ifdef FLAC__AVX2_SUPPORTED + if(encoder->private_->cpuinfo.ia32.avx2) + encoder->private_->local_precompute_partition_info_sums = FLAC__precompute_partition_info_sums_intrin_avx2; +# endif # elif defined FLAC__CPU_X86_64 # ifdef FLAC__SSE2_SUPPORTED encoder->private_->local_precompute_partition_info_sums = FLAC__precompute_partition_info_sums_intrin_sse2; @@ -1021,6 +1039,10 @@ static FLAC__StreamEncoderInitStatus init_stream_internal_( if(encoder->private_->cpuinfo.x86.ssse3) encoder->private_->local_precompute_partition_info_sums = FLAC__precompute_partition_info_sums_intrin_ssse3; # endif +# ifdef FLAC__AVX2_SUPPORTED + if(encoder->private_->cpuinfo.x86.avx2) + encoder->private_->local_precompute_partition_info_sums = FLAC__precompute_partition_info_sums_intrin_avx2; +# endif # endif /* FLAC__CPU_... */ } #endif /* !FLAC__NO_ASM && FLAC__HAS_X86INTRIN */ diff --git a/src/libFLAC/stream_encoder_intrin_avx2.c b/src/libFLAC/stream_encoder_intrin_avx2.c new file mode 100644 index 00000000..3aa31972 --- /dev/null +++ b/src/libFLAC/stream_encoder_intrin_avx2.c @@ -0,0 +1,142 @@ +/* libFLAC - Free Lossless Audio Codec library + * Copyright (C) 2000-2009 Josh Coalson + * Copyright (C) 2011-2014 Xiph.Org Foundation + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions + * are met: + * + * - Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * + * - Redistributions in binary form must reproduce the above copyright + * notice, this list of conditions and the following disclaimer in the + * documentation and/or other materials provided with the distribution. + * + * - Neither the name of the Xiph.org Foundation nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR + * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, + * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, + * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR + * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF + * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING + * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS + * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + */ + +#ifdef HAVE_CONFIG_H +# include +#endif + +#ifndef FLAC__NO_ASM +#if (defined FLAC__CPU_IA32 || defined FLAC__CPU_X86_64) && defined FLAC__HAS_X86INTRIN +#include "private/stream_encoder.h" +#include "private/bitmath.h" +#ifdef FLAC__AVX2_SUPPORTED + +#include /* for abs() */ +#include /* AVX2 */ +#include "FLAC/assert.h" + +FLAC__SSE_TARGET("avx2") +void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[], + unsigned residual_samples, unsigned predictor_order, unsigned min_partition_order, unsigned max_partition_order, unsigned bps) +{ + const unsigned default_partition_samples = (residual_samples + predictor_order) >> max_partition_order; + unsigned partitions = 1u << max_partition_order; + + FLAC__ASSERT(default_partition_samples > predictor_order); + + /* first do max_partition_order */ + { + 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) { + for(partition = residual_sample = 0; partition < partitions; partition++) { + 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))); + 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))); + sum128 = _mm_add_epi32(sum128, res128); + } + + for( ; residual_sample < end; residual_sample++) { + res128 = _mm_cvtsi32_si128(residual[residual_sample]); + res128 = _mm_abs_epi32(res128); + 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); + } + } + else { /* have to pessimistically use 64 bits for accumulator */ + for(partition = residual_sample = 0; partition < partitions; partition++) { + 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); + 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)); + 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]); + res128 = _mm_abs_epi32(res128); + 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 */ + { + unsigned 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--) { + unsigned 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(); +} + +#endif /* FLAC__AVX2_SUPPORTED */ +#endif /* (FLAC__CPU_IA32 || FLAC__CPU_X86_64) && FLAC__HAS_X86INTRIN */ +#endif /* FLAC__NO_ASM */ -- cgit v1.2.1