Merge branch 'avx2-quant'

build/kvazaar_lib/kvazaar_lib.vcxproj

@@ -153,6 +153,14 @@
     <ClCompile Include="..\..\src\search.c" />
     <ClCompile Include="..\..\src\search_inter.c" />
     <ClCompile Include="..\..\src\search_intra.c" />
+    <ClCompile Include="..\..\src\strategies\avx2\quant-avx2.c">
+      <EnableEnhancedInstructionSet Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet Condition="'$(Configuration)|$(Platform)'=='Release|x64'">AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+      <EnableEnhancedInstructionSet Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">AdvancedVectorExtensions2</EnableEnhancedInstructionSet>
+    </ClCompile>
+    <ClCompile Include="..\..\src\strategies\generic\quant-generic.c" />
+    <ClCompile Include="..\..\src\strategies\strategies-quant.c" />
     <ClCompile Include="..\..\src\yuv_io.c" />
     <ClInclude Include="..\..\src\checkpoint.h" />
     <ClInclude Include="..\..\src\cli.h" />
@@ -201,6 +209,10 @@
     <ClInclude Include="..\..\src\kvazaar_internal.h" />
     <ClInclude Include="..\..\src\search_inter.h" />
     <ClInclude Include="..\..\src\search_intra.h" />
+    <ClInclude Include="..\..\src\strategies\strategies-common.h" />
+    <ClInclude Include="..\..\src\strategies\avx2\quant-avx2.h" />
+    <ClInclude Include="..\..\src\strategies\generic\quant-generic.h" />
+    <ClInclude Include="..\..\src\strategies\strategies-quant.h" />
     <ClInclude Include="..\..\src\yuv_io.h" />
   </ItemGroup>
   <ItemGroup>

build/kvazaar_lib/kvazaar_lib.vcxproj.filters

@@ -207,6 +207,15 @@
     <ClCompile Include="..\..\src\input_frame_buffer.c">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="..\..\src\strategies\strategies-quant.c">
+      <Filter>Source Files\strategies</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\strategies\generic\quant-generic.c">
+      <Filter>Source Files\strategies\generic</Filter>
+    </ClCompile>
+    <ClCompile Include="..\..\src\strategies\avx2\quant-avx2.c">
+      <Filter>Source Files\strategies\avx2</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\..\src\global.h">
@@ -374,6 +383,18 @@
     <ClInclude Include="..\..\src\input_frame_buffer.h">
       <Filter>Header Files</Filter>
     </ClInclude>
+    <ClInclude Include="..\..\src\strategies\strategies-common.h">
+      <Filter>Header Files\strategies</Filter>
+    </ClInclude>
+    <ClInclude Include="..\..\src\strategies\strategies-quant.h">
+      <Filter>Header Files\strategies</Filter>
+    </ClInclude>
+    <ClInclude Include="..\..\src\strategies\generic\quant-generic.h">
+      <Filter>Header Files\strategies\generic</Filter>
+    </ClInclude>
+    <ClInclude Include="..\..\src\strategies\avx2\quant-avx2.h">
+      <Filter>Header Files\strategies\avx2</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <YASM Include="..\..\src\extras\x86inc.asm">

src/Makefile

@@ -203,6 +203,7 @@ OBJS = \
   strategies/strategies-nal.o \
   strategies/strategies-dct.o \
   strategies/strategies-ipol.o \
+  strategies/strategies-quant.o \
   strategies/generic/nal-generic.o \
   strategies/generic/picture-generic.o \
   strategies/sse2/picture-sse2.o \
@@ -213,7 +214,9 @@ OBJS = \
   strategies/generic/dct-generic.o \
   strategies/avx2/dct-avx2.o \
   strategies/generic/ipol-generic.o \
-  strategies/avx2/ipol-avx2.o
+  strategies/avx2/ipol-avx2.o \
+  strategies/generic/quant-generic.o \
+  strategies/avx2/quant-avx2.o \
 
 ifndef KVZ_DISABLE_ASM
   # Compile C files in x86_asm folder with KVZ_COMPILE_ASM, which will cause

src/rdo.c

@@ -31,6 +31,7 @@
 #include "context.h"
 #include "cabac.h"
 #include "transform.h"
+#include "strategies/strategies-quant.h"
 
 
 #define QUANT_SHIFT          14

src/strategies/avx2/picture-avx2.c

@@ -26,6 +26,7 @@
 
 #if COMPILE_INTEL_AVX2
 #  include "image.h"
+#  include "strategies/strategies-common.h"
 #  include <immintrin.h>
 
 
@@ -136,6 +137,186 @@ static unsigned sad_8bit_64x64_avx2(const kvz_pixel * buf1, const kvz_pixel * bu
   return m256i_horizontal_sum(sum0);
 }
 
+static unsigned satd_8bit_4x4_avx2(const kvz_pixel *org, const kvz_pixel *cur)
+{
+
+  __m128i original = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)org));
+  __m128i current = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)cur));
+
+  __m128i diff_lo = _mm_sub_epi16(current, original);
+
+  original = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)(org + 8)));
+  current = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)(cur + 8)));
+
+  __m128i diff_hi = _mm_sub_epi16(current, original);
+
+
+  //Hor
+  __m128i row0 = _mm_hadd_epi16(diff_lo, diff_hi);
+  __m128i row1 = _mm_hsub_epi16(diff_lo, diff_hi);
+
+  __m128i row2 = _mm_hadd_epi16(row0, row1);
+  __m128i row3 = _mm_hsub_epi16(row0, row1);
+
+  //Ver
+  row0 = _mm_hadd_epi16(row2, row3);
+  row1 = _mm_hsub_epi16(row2, row3);
+
+  row2 = _mm_hadd_epi16(row0, row1);
+  row3 = _mm_hsub_epi16(row0, row1);
+
+  //Abs and sum
+  row2 = _mm_abs_epi16(row2);
+  row3 = _mm_abs_epi16(row3);
+
+  row3 = _mm_add_epi16(row2, row3);
+
+  row3 = _mm_add_epi16(row3, _mm_shuffle_epi32(row3, KVZ_PERMUTE(2, 3, 0, 1) ));
+  row3 = _mm_add_epi16(row3, _mm_shuffle_epi32(row3, KVZ_PERMUTE(1, 0, 1, 0) ));
+  row3 = _mm_add_epi16(row3, _mm_shufflelo_epi16(row3, KVZ_PERMUTE(1, 0, 1, 0) ));
+
+  unsigned sum = _mm_extract_epi16(row3, 0);
+  unsigned satd = (sum + 1) >> 1;
+
+  return satd;
+}
+
+static void hor_add_sub_avx2(__m128i *row0, __m128i *row1){
+
+  __m128i a = _mm_hadd_epi16(*row0, *row1);
+  __m128i b = _mm_hsub_epi16(*row0, *row1);
+
+  __m128i c = _mm_hadd_epi16(a, b);
+  __m128i d = _mm_hsub_epi16(a, b);
+
+  *row0 = _mm_hadd_epi16(c, d);
+  *row1 = _mm_hsub_epi16(c, d);
+}
+
+static INLINE void ver_add_sub_avx2(__m128i temp_hor[8], __m128i temp_ver[8]){
+
+  // First stage
+  for (int i = 0; i < 8; i += 2){
+    temp_ver[i+0] = _mm_hadd_epi16(temp_hor[i + 0], temp_hor[i + 1]);
+    temp_ver[i+1] = _mm_hsub_epi16(temp_hor[i + 0], temp_hor[i + 1]);
+  }
+
+  // Second stage
+  for (int i = 0; i < 8; i += 4){
+    temp_hor[i + 0] = _mm_add_epi16(temp_ver[i + 0], temp_ver[i + 2]);
+    temp_hor[i + 1] = _mm_add_epi16(temp_ver[i + 1], temp_ver[i + 3]);
+    temp_hor[i + 2] = _mm_sub_epi16(temp_ver[i + 0], temp_ver[i + 2]);
+    temp_hor[i + 3] = _mm_sub_epi16(temp_ver[i + 1], temp_ver[i + 3]);
+  }
+
+  // Third stage
+  for (int i = 0; i < 4; ++i){
+    temp_ver[i + 0] = _mm_add_epi16(temp_hor[0 + i], temp_hor[4 + i]);
+    temp_ver[i + 4] = _mm_sub_epi16(temp_hor[0 + i], temp_hor[4 + i]);
+  }
+}
+
+INLINE static void haddwd_accumulate_avx2(__m128i *accumulate, __m128i *ver_row)
+{
+  __m128i abs_value = _mm_abs_epi16(*ver_row);
+  *accumulate = _mm_add_epi32(*accumulate, _mm_madd_epi16(abs_value, _mm_set1_epi16(1)));
+}
+
+INLINE static unsigned sum_block_avx2(__m128i *ver_row)
+{
+  __m128i sad = _mm_setzero_si128();
+  haddwd_accumulate_avx2(&sad, ver_row + 0);
+  haddwd_accumulate_avx2(&sad, ver_row + 1);
+  haddwd_accumulate_avx2(&sad, ver_row + 2);
+  haddwd_accumulate_avx2(&sad, ver_row + 3); 
+  haddwd_accumulate_avx2(&sad, ver_row + 4);
+  haddwd_accumulate_avx2(&sad, ver_row + 5);
+  haddwd_accumulate_avx2(&sad, ver_row + 6);
+  haddwd_accumulate_avx2(&sad, ver_row + 7);
+
+  sad = _mm_add_epi32(sad, _mm_shuffle_epi32(sad, KVZ_PERMUTE(2, 3, 0, 1)));
+  sad = _mm_add_epi32(sad, _mm_shuffle_epi32(sad, KVZ_PERMUTE(1, 0, 1, 0)));
+
+  return _mm_cvtsi128_si32(sad);
+}
+
+INLINE static __m128i diff_row_avx2(const kvz_pixel *buf1, const kvz_pixel *buf2)
+{
+  __m128i buf1_row = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)buf1));
+  __m128i buf2_row = _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)buf2));
+  return _mm_sub_epi16(buf1_row, buf2_row);
+}
+
+INLINE static void diff_blocks_and_hor_transform_avx2(__m128i row_diff[8], const kvz_pixel * buf1, unsigned stride1, const kvz_pixel * buf2, unsigned stride2)
+{
+  row_diff[0] = diff_row_avx2(buf1 + 0 * stride1, buf2 + 0 * stride2);
+  row_diff[1] = diff_row_avx2(buf1 + 1 * stride1, buf2 + 1 * stride2);
+  hor_add_sub_avx2(row_diff + 0, row_diff + 1);
+
+  row_diff[2] = diff_row_avx2(buf1 + 2 * stride1, buf2 + 2 * stride2);
+  row_diff[3] = diff_row_avx2(buf1 + 3 * stride1, buf2 + 3 * stride2);
+  hor_add_sub_avx2(row_diff + 2, row_diff + 3);
+
+  row_diff[4] = diff_row_avx2(buf1 + 4 * stride1, buf2 + 4 * stride2);
+  row_diff[5] = diff_row_avx2(buf1 + 5 * stride1, buf2 + 5 * stride2);
+  hor_add_sub_avx2(row_diff + 4, row_diff + 5);
+
+  row_diff[6] = diff_row_avx2(buf1 + 6 * stride1, buf2 + 6 * stride2);
+  row_diff[7] = diff_row_avx2(buf1 + 7 * stride1, buf2 + 7 * stride2);
+  hor_add_sub_avx2(row_diff + 6, row_diff + 7);
+}
+
+static unsigned kvz_satd_8bit_8x8_general_avx2(const kvz_pixel * buf1, unsigned stride1, const kvz_pixel * buf2, unsigned stride2)
+{
+  __m128i temp_hor[8];
+  __m128i temp_ver[8];
+
+  diff_blocks_and_hor_transform_avx2(temp_hor, buf1, stride1, buf2, stride2);
+
+  ver_add_sub_avx2(temp_hor, temp_ver);
+  
+  unsigned sad = sum_block_avx2(temp_ver);
+
+  unsigned result = (sad + 2) >> 2;
+  return result;
+}
+
+// Function macro for defining hadamard calculating functions
+// for fixed size blocks. They calculate hadamard for integer
+// multiples of 8x8 with the 8x8 hadamard function.
+#define SATD_NXN_AVX2(n) \
+static unsigned satd_8bit_ ## n ## x ## n ## _avx2( \
+  const kvz_pixel * const block1, const kvz_pixel * const block2) \
+{ \
+  unsigned x, y; \
+  unsigned sum = 0; \
+  for (y = 0; y < (n); y += 8) { \
+  unsigned row = y * (n); \
+  for (x = 0; x < (n); x += 8) { \
+  sum += kvz_satd_8bit_8x8_general_avx2(&block1[row + x], (n), &block2[row + x], (n)); \
+    } \
+    } \
+  return sum>>(KVZ_BIT_DEPTH-8); \
+}
+
+static unsigned satd_8bit_8x8_avx2(
+  const kvz_pixel * const block1, const kvz_pixel * const block2) 
+{ 
+  unsigned x, y; 
+  unsigned sum = 0; 
+  for (y = 0; y < (8); y += 8) { 
+  unsigned row = y * (8); 
+  for (x = 0; x < (8); x += 8) { 
+  sum += kvz_satd_8bit_8x8_general_avx2(&block1[row + x], (8), &block2[row + x], (8)); 
+      } 
+      } 
+  return sum>>(KVZ_BIT_DEPTH-8); \
+}
+
+//SATD_NXN_AVX2(8) //Use the non-macro version
+SATD_NXN_AVX2(16)
+SATD_NXN_AVX2(32)
+SATD_NXN_AVX2(64)
 
 #endif //COMPILE_INTEL_AVX2
 
@@ -153,6 +334,12 @@ int kvz_strategy_register_picture_avx2(void* opaque, uint8_t bitdepth)
     success &= kvz_strategyselector_register(opaque, "sad_16x16", "avx2", 40, &sad_8bit_16x16_avx2);
     success &= kvz_strategyselector_register(opaque, "sad_32x32", "avx2", 40, &sad_8bit_32x32_avx2);
     success &= kvz_strategyselector_register(opaque, "sad_64x64", "avx2", 40, &sad_8bit_64x64_avx2);
+
+    success &= kvz_strategyselector_register(opaque, "satd_4x4", "avx2", 40, &satd_8bit_4x4_avx2);
+    success &= kvz_strategyselector_register(opaque, "satd_8x8", "avx2", 40, &satd_8bit_8x8_avx2);
+    success &= kvz_strategyselector_register(opaque, "satd_16x16", "avx2", 40, &satd_8bit_16x16_avx2);
+    success &= kvz_strategyselector_register(opaque, "satd_32x32", "avx2", 40, &satd_8bit_32x32_avx2);
+    success &= kvz_strategyselector_register(opaque, "satd_64x64", "avx2", 40, &satd_8bit_64x64_avx2);
   }
 #endif
   return success;

src/strategies/avx2/quant-avx2.c

@@ -0,0 +1,220 @@
+/*****************************************************************************
+* This file is part of Kvazaar HEVC encoder.
+*
+* Copyright (C) 2013-2015 Tampere University of Technology and others (see
+* COPYING file).
+*
+* Kvazaar is free software: you can redistribute it and/or modify it under
+* the terms of the GNU Lesser General Public License as published by the
+* Free Software Foundation; either version 2.1 of the License, or (at your
+* option) any later version.
+*
+* Kvazaar is distributed in the hope that it will be useful, but WITHOUT ANY
+* WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+* FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
+* more details.
+*
+* You should have received a copy of the GNU General Public License along
+* with Kvazaar.  If not, see <http://www.gnu.org/licenses/>.
+****************************************************************************/
+
+/*
+* \file
+*/
+
+#include <stdlib.h>
+
+#include "quant-avx2.h"
+#include "../generic/quant-generic.h"
+#include "../strategies-common.h"
+#include "strategyselector.h"
+#include "encoder.h"
+#include "transform.h"
+
+#if COMPILE_INTEL_AVX2
+#include <immintrin.h>
+
+/**
+* \brief quantize transformed coefficents
+*
+*/
+
+void kvz_quant_flat_avx2(const encoder_state_t * const state, coeff_t *coef, coeff_t *q_coef, int32_t width,
+  int32_t height, int8_t type, int8_t scan_idx, int8_t block_type)
+{
+  const encoder_control_t * const encoder = state->encoder_control;
+  const uint32_t log2_block_size = kvz_g_convert_to_bit[width] + 2;
+  const uint32_t * const scan = kvz_g_sig_last_scan[scan_idx][log2_block_size - 1];
+
+  int32_t qp_scaled = kvz_get_scaled_qp(type, state->global->QP, (encoder->bitdepth - 8) * 6);
+  const uint32_t log2_tr_size = kvz_g_convert_to_bit[width] + 2;
+  const int32_t scalinglist_type = (block_type == CU_INTRA ? 0 : 3) + (int8_t)("\0\3\1\2"[type]);
+  const int32_t *quant_coeff = encoder->scaling_list.quant_coeff[log2_tr_size - 2][scalinglist_type][qp_scaled % 6];
+  const int32_t transform_shift = MAX_TR_DYNAMIC_RANGE - encoder->bitdepth - log2_tr_size; //!< Represents scaling through forward transform
+  const int32_t q_bits = QUANT_SHIFT + qp_scaled / 6 + transform_shift;
+  const int32_t add = ((state->global->slicetype == KVZ_SLICE_I) ? 171 : 85) << (q_bits - 9);
+  const int32_t q_bits8 = q_bits - 8;
+
+  assert(quant_coeff[0] <= (1 << 15) - 1 && quant_coeff[0] >= -(1 << 15)); //Assuming flat values to fit int16_t
+
+  uint32_t ac_sum = 0;
+
+  __m256i v_ac_sum = _mm256_setzero_si256();
+  __m256i v_quant_coeff = _mm256_set1_epi16(quant_coeff[0]);
+
+  for (int32_t n = 0; n < width * height; n += 16) {
+
+    __m256i v_level = _mm256_loadu_si256((__m256i*)&(coef[n]));
+    __m256i v_sign = _mm256_cmpgt_epi16(_mm256_setzero_si256(), v_level);
+    v_sign = _mm256_or_si256(v_sign, _mm256_set1_epi16(1));
+
+    v_level = _mm256_abs_epi16(v_level);
+    __m256i low_a = _mm256_unpacklo_epi16(v_level, _mm256_set1_epi16(0));
+    __m256i high_a = _mm256_unpackhi_epi16(v_level, _mm256_set1_epi16(0));
+
+    __m256i low_b = _mm256_unpacklo_epi16(v_quant_coeff, _mm256_set1_epi16(0));
+    __m256i high_b = _mm256_unpackhi_epi16(v_quant_coeff, _mm256_set1_epi16(0));
+
+    __m256i v_level32_a = _mm256_madd_epi16(low_a, low_b);
+    __m256i v_level32_b = _mm256_madd_epi16(high_a, high_b);
+
+    v_level32_a = _mm256_add_epi32(v_level32_a, _mm256_set1_epi32(add));
+    v_level32_b = _mm256_add_epi32(v_level32_b, _mm256_set1_epi32(add));
+
+    v_level32_a = _mm256_srai_epi32(v_level32_a, q_bits);
+    v_level32_b = _mm256_srai_epi32(v_level32_b, q_bits);
+
+    v_level = _mm256_packs_epi32(v_level32_a, v_level32_b);
+    v_level = _mm256_sign_epi16(v_level, v_sign);
+
+    _mm256_storeu_si256((__m256i*)&(q_coef[n]), v_level);
+
+    v_ac_sum = _mm256_add_epi32(v_ac_sum, v_level32_a);
+    v_ac_sum = _mm256_add_epi32(v_ac_sum, v_level32_b);
+  }
+
+  __m128i temp = _mm_add_epi32(_mm256_castsi256_si128(v_ac_sum), _mm256_extracti128_si256(v_ac_sum, 1));
+  temp = _mm_add_epi32(temp, _mm_shuffle_epi32(temp, KVZ_PERMUTE(2, 3, 0, 1)));
+  temp = _mm_add_epi32(temp, _mm_shuffle_epi32(temp, KVZ_PERMUTE(1, 0, 1, 0)));
+  ac_sum += _mm_cvtsi128_si32(temp);
+
+  if (!(encoder->sign_hiding && ac_sum >= 2)) return;
+
+  int32_t delta_u[LCU_WIDTH*LCU_WIDTH >> 2];
+
+  for (int32_t n = 0; n < width * height; n++) {
+    int32_t level;
+    level = coef[n];
+    level = ((int64_t)abs(level) * quant_coeff[n] + add) >> q_bits;
+    delta_u[n] = (int32_t)(((int64_t)abs(coef[n]) * quant_coeff[n] - (level << q_bits)) >> q_bits8);
+  }
+
+  if (ac_sum >= 2) {
+#define SCAN_SET_SIZE 16
+#define LOG2_SCAN_SET_SIZE 4
+    int32_t n, last_cg = -1, abssum = 0, subset, subpos;
+    for (subset = (width*height - 1) >> LOG2_SCAN_SET_SIZE; subset >= 0; subset--) {
+      int32_t first_nz_pos_in_cg = SCAN_SET_SIZE, last_nz_pos_in_cg = -1;
+      subpos = subset << LOG2_SCAN_SET_SIZE;
+      abssum = 0;
+
+      // Find last coeff pos
+      for (n = SCAN_SET_SIZE - 1; n >= 0; n--)  {
+        if (q_coef[scan[n + subpos]])  {
+          last_nz_pos_in_cg = n;
+          break;
+        }
+      }
+
+      // First coeff pos
+      for (n = 0; n <SCAN_SET_SIZE; n++) {
+        if (q_coef[scan[n + subpos]]) {
+          first_nz_pos_in_cg = n;
+          break;
+        }
+      }
+
+      // Sum all kvz_quant coeffs between first and last
+      for (n = first_nz_pos_in_cg; n <= last_nz_pos_in_cg; n++) {
+        abssum += q_coef[scan[n + subpos]];
+      }
+
+      if (last_nz_pos_in_cg >= 0 && last_cg == -1) {
+        last_cg = 1;
+      }
+
+      if (last_nz_pos_in_cg - first_nz_pos_in_cg >= 4) {
+        int32_t signbit = (q_coef[scan[subpos + first_nz_pos_in_cg]] > 0 ? 0 : 1);
+        if (signbit != (abssum & 0x1)) { // compare signbit with sum_parity
+          int32_t min_cost_inc = 0x7fffffff, min_pos = -1, cur_cost = 0x7fffffff;
+          int16_t final_change = 0, cur_change = 0;
+          for (n = (last_cg == 1 ? last_nz_pos_in_cg : SCAN_SET_SIZE - 1); n >= 0; n--) {
+            uint32_t blkPos = scan[n + subpos];
+            if (q_coef[blkPos] != 0) {
+              if (delta_u[blkPos] > 0) {
+                cur_cost = -delta_u[blkPos];
+                cur_change = 1;
+              }
+              else if (n == first_nz_pos_in_cg && abs(q_coef[blkPos]) == 1) {
+                cur_cost = 0x7fffffff;
+              }
+              else {
+                cur_cost = delta_u[blkPos];
+                cur_change = -1;
+              }
+            }
+            else if (n < first_nz_pos_in_cg && ((coef[blkPos] >= 0) ? 0 : 1) != signbit) {
+              cur_cost = 0x7fffffff;
+            }
+            else {
+              cur_cost = -delta_u[blkPos];
+              cur_change = 1;
+            }
+
+            if (cur_cost < min_cost_inc) {
+              min_cost_inc = cur_cost;
+              final_change = cur_change;
+              min_pos = blkPos;
+            }
+          } // CG loop
+
+          if (q_coef[min_pos] == 32767 || q_coef[min_pos] == -32768) {
+            final_change = -1;
+          }
+
+          if (coef[min_pos] >= 0) q_coef[min_pos] += final_change;
+          else q_coef[min_pos] -= final_change;
+        } // Hide
+      }
+      if (last_cg == 1) last_cg = 0;
+    }
+
+#undef SCAN_SET_SIZE
+#undef LOG2_SCAN_SET_SIZE
+  }
+}
+
+void kvz_quant_avx2(const encoder_state_t * const state, coeff_t *coef, coeff_t *q_coef, int32_t width,
+  int32_t height, int8_t type, int8_t scan_idx, int8_t block_type)
+{
+  if (state->encoder_control->scaling_list.enable){
+    kvz_quant_generic(state, coef, q_coef, width, height, type, scan_idx, block_type);
+  }
+  else {
+    kvz_quant_flat_avx2(state, coef, q_coef, width, height, type, scan_idx, block_type);
+  }
+}
+
+#endif //COMPILE_INTEL_AVX2
+
+
+int kvz_strategy_register_quant_avx2(void* opaque, uint8_t bitdepth)
+{
+  bool success = true;
+
+#if COMPILE_INTEL_AVX2
+  success &= kvz_strategyselector_register(opaque, "quant", "avx2", 40, &kvz_quant_avx2);
+#endif //COMPILE_INTEL_AVX2
+
+  return success;
+}

src/strategies/avx2/quant-avx2.h

@@ -0,0 +1,26 @@
+#ifndef STRATEGIES_QUANT_AVX2_H_
+#define STRATEGIES_QUANT_AVX2_H_
+/*****************************************************************************
+ * This file is part of Kvazaar HEVC encoder.
+ *
+ * Copyright (C) 2013-2015 Tampere University of Technology and others (see
+ * COPYING file).
+ *
+ * Kvazaar is free software: you can redistribute it and/or modify it under
+ * the terms of the GNU Lesser General Public License as published by the
+ * Free Software Foundation; either version 2.1 of the License, or (at your
+ * option) any later version.
+ *
+ * Kvazaar is distributed in the hope that it will be useful, but WITHOUT ANY
+ * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+ * FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with Kvazaar.  If not, see <http://www.gnu.org/licenses/>.
+ ****************************************************************************/
+#include <stdint.h>
+
+int kvz_strategy_register_quant_avx2(void* opaque, uint8_t bitdepth);
+
+#endif //STRATEGIES_QUANT_AVX2_H_

src/strategies/generic/quant-generic.c

@@ -0,0 +1,173 @@
+/*****************************************************************************
+ * This file is part of Kvazaar HEVC encoder.
+ *
+ * Copyright (C) 2013-2015 Tampere University of Technology and others (see
+ * COPYING file).
+ *
+ * Kvazaar is free software: you can redistribute it and/or modify it under
+ * the terms of the GNU Lesser General Public License as published by the
+ * Free Software Foundation; either version 2.1 of the License, or (at your
+ * option) any later version.
+ *
+ * Kvazaar is distributed in the hope that it will be useful, but WITHOUT ANY
+ * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+ * FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with Kvazaar.  If not, see <http://www.gnu.org/licenses/>.
+ ****************************************************************************/
+
+/*
+ * \file
+ */
+
+#include <stdlib.h>
+
+#include "quant-generic.h"
+#include "strategyselector.h"
+#include "encoder.h"
+#include "transform.h"
+
+#define QUANT_SHIFT 14
+/**
+* \brief quantize transformed coefficents
+*
+*/
+void kvz_quant_generic(const encoder_state_t * const state, coeff_t *coef, coeff_t *q_coef, int32_t width,
+  int32_t height, int8_t type, int8_t scan_idx, int8_t block_type)
+{
+  const encoder_control_t * const encoder = state->encoder_control;
+  const uint32_t log2_block_size = kvz_g_convert_to_bit[width] + 2;
+  const uint32_t * const scan = kvz_g_sig_last_scan[scan_idx][log2_block_size - 1];
+
+  int32_t qp_scaled = kvz_get_scaled_qp(type, state->global->QP, (encoder->bitdepth - 8) * 6);
+  const uint32_t log2_tr_size = kvz_g_convert_to_bit[width] + 2;
+  const int32_t scalinglist_type = (block_type == CU_INTRA ? 0 : 3) + (int8_t)("\0\3\1\2"[type]);
+  const int32_t *quant_coeff = encoder->scaling_list.quant_coeff[log2_tr_size - 2][scalinglist_type][qp_scaled % 6];
+  const int32_t transform_shift = MAX_TR_DYNAMIC_RANGE - encoder->bitdepth - log2_tr_size; //!< Represents scaling through forward transform
+  const int32_t q_bits = QUANT_SHIFT + qp_scaled / 6 + transform_shift;
+  const int32_t add = ((state->global->slicetype == KVZ_SLICE_I) ? 171 : 85) << (q_bits - 9);
+  const int32_t q_bits8 = q_bits - 8;
+
+  uint32_t ac_sum = 0;
+
+  for (int32_t n = 0; n < width * height; n++) {
+    int32_t level;
+    int32_t  sign;
+
+    level = coef[n];
+    sign = (level < 0 ? -1 : 1);
+
+    level = ((int64_t)abs(level) * quant_coeff[n] + add) >> q_bits;
+    ac_sum += level;
+
+    level *= sign;
+    q_coef[n] = (coeff_t)(CLIP(-32768, 32767, level));
+  }
+
+  if (!(encoder->sign_hiding && ac_sum >= 2)) return;
+
+  int32_t delta_u[LCU_WIDTH*LCU_WIDTH >> 2];
+
+  for (int32_t n = 0; n < width * height; n++) {
+    int32_t level;
+    level = coef[n];
+    level = ((int64_t)abs(level) * quant_coeff[n] + add) >> q_bits;
+    delta_u[n] = (int32_t)(((int64_t)abs(coef[n]) * quant_coeff[n] - (level << q_bits)) >> q_bits8);
+  }
+
+  if (ac_sum >= 2) {
+#define SCAN_SET_SIZE 16
+#define LOG2_SCAN_SET_SIZE 4
+    int32_t n, last_cg = -1, abssum = 0, subset, subpos;
+    for (subset = (width*height - 1) >> LOG2_SCAN_SET_SIZE; subset >= 0; subset--) {
+      int32_t first_nz_pos_in_cg = SCAN_SET_SIZE, last_nz_pos_in_cg = -1;
+      subpos = subset << LOG2_SCAN_SET_SIZE;
+      abssum = 0;
+
+      // Find last coeff pos
+      for (n = SCAN_SET_SIZE - 1; n >= 0; n--)  {
+        if (q_coef[scan[n + subpos]])  {
+          last_nz_pos_in_cg = n;
+          break;
+        }
+      }
+
+      // First coeff pos
+      for (n = 0; n <SCAN_SET_SIZE; n++) {
+        if (q_coef[scan[n + subpos]]) {
+          first_nz_pos_in_cg = n;
+          break;
+        }
+      }
+
+      // Sum all kvz_quant coeffs between first and last
+      for (n = first_nz_pos_in_cg; n <= last_nz_pos_in_cg; n++) {
+        abssum += q_coef[scan[n + subpos]];
+      }
+
+      if (last_nz_pos_in_cg >= 0 && last_cg == -1) {
+        last_cg = 1;
+      }
+
+      if (last_nz_pos_in_cg - first_nz_pos_in_cg >= 4) {
+        int32_t signbit = (q_coef[scan[subpos + first_nz_pos_in_cg]] > 0 ? 0 : 1);
+        if (signbit != (abssum & 0x1)) { // compare signbit with sum_parity
+          int32_t min_cost_inc = 0x7fffffff, min_pos = -1, cur_cost = 0x7fffffff;
+          int16_t final_change = 0, cur_change = 0;
+          for (n = (last_cg == 1 ? last_nz_pos_in_cg : SCAN_SET_SIZE - 1); n >= 0; n--) {
+            uint32_t blkPos = scan[n + subpos];
+            if (q_coef[blkPos] != 0) {
+              if (delta_u[blkPos] > 0) {
+                cur_cost = -delta_u[blkPos];
+                cur_change = 1;
+              }
+              else if (n == first_nz_pos_in_cg && abs(q_coef[blkPos]) == 1) {
+                cur_cost = 0x7fffffff;
+              }
+              else {
+                cur_cost = delta_u[blkPos];
+                cur_change = -1;
+              }
+            }
+            else if (n < first_nz_pos_in_cg && ((coef[blkPos] >= 0) ? 0 : 1) != signbit) {
+              cur_cost = 0x7fffffff;
+            }
+            else {
+              cur_cost = -delta_u[blkPos];
+              cur_change = 1;
+            }
+
+            if (cur_cost < min_cost_inc) {
+              min_cost_inc = cur_cost;
+              final_change = cur_change;
+              min_pos = blkPos;
+            }
+          } // CG loop
+
+          if (q_coef[min_pos] == 32767 || q_coef[min_pos] == -32768) {
+            final_change = -1;
+          }
+
+          if (coef[min_pos] >= 0) q_coef[min_pos] += final_change;
+          else q_coef[min_pos] -= final_change;
+        } // Hide
+      }
+      if (last_cg == 1) last_cg = 0;
+    }
+
+#undef SCAN_SET_SIZE
+#undef LOG2_SCAN_SET_SIZE
+  }
+}
+
+
+int kvz_strategy_register_quant_generic(void* opaque, uint8_t bitdepth)
+{
+  bool success = true;
+
+  success &= kvz_strategyselector_register(opaque, "quant", "generic", 0, &kvz_quant_generic);
+
+  return success;
+}

src/strategies/generic/quant-generic.h

@@ -0,0 +1,31 @@
+#ifndef STRATEGIES_QUANT_GENERIC_H_
+#define STRATEGIES_QUANT_GENERIC_H_
+/*****************************************************************************
+ * This file is part of Kvazaar HEVC encoder.
+ *
+ * Copyright (C) 2013-2015 Tampere University of Technology and others (see
+ * COPYING file).
+ *
+ * Kvazaar is free software: you can redistribute it and/or modify it under
+ * the terms of the GNU Lesser General Public License as published by the
+ * Free Software Foundation; either version 2.1 of the License, or (at your
+ * option) any later version.
+ *
+ * Kvazaar is distributed in the hope that it will be useful, but WITHOUT ANY
+ * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+ * FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with Kvazaar.  If not, see <http://www.gnu.org/licenses/>.
+ ****************************************************************************/
+#include <stdint.h>
+#include "encoderstate.h"
+
+#define QUANT_SHIFT 14
+
+int kvz_strategy_register_quant_generic(void* opaque, uint8_t bitdepth);
+void kvz_quant_generic(const encoder_state_t * const state, coeff_t *coef, coeff_t *q_coef, int32_t width,
+  int32_t height, int8_t type, int8_t scan_idx, int8_t block_type);
+
+#endif //STRATEGIES_QUANT_GENERIC_H_

src/strategies/strategies-common.h

@@ -0,0 +1,8 @@
+#ifndef STRATEGIES_COMMON_H_
+#define STRATEGIES_COMMON_H_
+
+//Use with shuffle and permutation intrinsics.
+//Parameters are indices to packed elements. Each must be 0, 1, 2 or 3.
+#define KVZ_PERMUTE(a, b, c, d) ( (a << 0) | (b << 2) | (c << 4) | (d << 6) )
+
+#endif //STRATEGIES_COMMON_H_

src/strategies/strategies-quant.c

@@ -0,0 +1,41 @@
+/*****************************************************************************
+ * This file is part of Kvazaar HEVC encoder.
+ *
+ * Copyright (C) 2013-2015 Tampere University of Technology and others (see
+ * COPYING file).
+ *
+ * Kvazaar is free software: you can redistribute it and/or modify it under
+ * the terms of the GNU Lesser General Public License as published by the
+ * Free Software Foundation; either version 2.1 of the License, or (at your
+ * option) any later version.
+ *
+ * Kvazaar is distributed in the hope that it will be useful, but WITHOUT ANY
+ * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+ * FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with Kvazaar.  If not, see <http://www.gnu.org/licenses/>.
+ ****************************************************************************/
+
+#include "strategies-quant.h"
+#include "strategyselector.h"
+
+// Define function pointers.
+quant_func *kvz_quant;
+
+// Headers for platform optimizations.
+#include "generic/quant-generic.h"
+#include "avx2/quant-avx2.h"
+
+
+int kvz_strategy_register_quant(void* opaque, uint8_t bitdepth) {
+  bool success = true;
+
+  success &= kvz_strategy_register_quant_generic(opaque, bitdepth);
+
+  if (kvz_g_hardware_flags.intel_flags.avx2) {
+    success &= kvz_strategy_register_quant_avx2(opaque, bitdepth);
+  }
+  return success;
+}

src/strategies/strategies-quant.h

@@ -0,0 +1,40 @@
+#ifndef STRATEGIES_QUANT_H_
+#define STRATEGIES_QUANT_H_
+/*****************************************************************************
+ * This file is part of Kvazaar HEVC encoder.
+ *
+ * Copyright (C) 2013-2015 Tampere University of Technology and others (see
+ * COPYING file).
+ *
+ * Kvazaar is free software: you can redistribute it and/or modify it under
+ * the terms of the GNU Lesser General Public License as published by the
+ * Free Software Foundation; either version 2.1 of the License, or (at your
+ * option) any later version.
+ *
+ * Kvazaar is distributed in the hope that it will be useful, but WITHOUT ANY
+ * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+ * FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along
+ * with Kvazaar.  If not, see <http://www.gnu.org/licenses/>.
+ ****************************************************************************/
+ 
+#include "encoderstate.h"
+
+// Declare function pointers.
+typedef unsigned (quant_func)(const encoder_state_t * const state, coeff_t *coef, coeff_t *q_coef, int32_t width,
+  int32_t height, int8_t type, int8_t scan_idx, int8_t block_type);
+
+// Declare function pointers.
+extern quant_func * kvz_quant;
+
+int kvz_strategy_register_quant(void* opaque, uint8_t bitdepth);
+
+
+#define STRATEGIES_QUANT_EXPORTS \
+  {"quant", (void**) &kvz_quant}, \
+
+
+
+#endif //STRATEGIES_QUANT_H_

src/strategyselector.c

@@ -70,6 +70,11 @@ int kvz_strategyselector_init(int32_t cpuid, uint8_t bitdepth) {
     return 0;
   }
 
+  if (!kvz_strategy_register_quant(&strategies, bitdepth)) {
+    fprintf(stderr, "kvz_strategy_register_quant failed!\n");
+    return 0;
+  }
+  
   while(cur_strategy_to_select->fptr) {
     *(cur_strategy_to_select->fptr) = strategyselector_choose_for(&strategies, cur_strategy_to_select->strategy_type);
     

src/strategyselector.h

@@ -148,12 +148,14 @@ int kvz_strategyselector_register(void *opaque, const char *type, const char *st
 #include "strategies/strategies-picture.h"
 #include "strategies/strategies-dct.h"
 #include "strategies/strategies-ipol.h"
+#include "strategies/strategies-quant.h"
 
 static const strategy_to_select_t strategies_to_select[] = {
   STRATEGIES_NAL_EXPORTS
   STRATEGIES_PICTURE_EXPORTS
   STRATEGIES_DCT_EXPORTS
   STRATEGIES_IPOL_EXPORTS
+  STRATEGIES_QUANT_EXPORTS
   { NULL, NULL },
 };
 

src/transform.c

@@ -33,6 +33,8 @@
 #include "nal.h"
 #include "rdo.h"
 #include "strategies/strategies-dct.h"
+#include "strategies/strategies-quant.h"
+#include "strategies/generic/quant-generic.h"
 
 //////////////////////////////////////////////////////////////////////////
 // INITIALIZATIONS
@@ -127,137 +129,6 @@ void kvz_itransform2d(const encoder_control_t * const encoder, int16_t *block, i
   idct_func(encoder->bitdepth, coeff, block);
 }
 
-
-#define QUANT_SHIFT 14
-/**
- * \brief quantize transformed coefficents
- *
- */
-void kvz_quant(const encoder_state_t * const state, coeff_t *coef, coeff_t *q_coef, int32_t width,
-           int32_t height, int8_t type, int8_t scan_idx, int8_t block_type )
-{
-  const encoder_control_t * const encoder = state->encoder_control;
-  const uint32_t log2_block_size = kvz_g_convert_to_bit[ width ] + 2;
-  const uint32_t * const scan = kvz_g_sig_last_scan[ scan_idx ][ log2_block_size - 1 ];
-
-  int32_t qp_scaled = kvz_get_scaled_qp(type, state->global->QP, (encoder->bitdepth-8)*6);
-
-  const uint32_t log2_tr_size = kvz_g_convert_to_bit[ width ] + 2;
-  const int32_t scalinglist_type = (block_type == CU_INTRA ? 0 : 3) + (int8_t)("\0\3\1\2"[type]);
-  const int32_t *quant_coeff = encoder->scaling_list.quant_coeff[log2_tr_size-2][scalinglist_type][qp_scaled%6];
-  const int32_t transform_shift = MAX_TR_DYNAMIC_RANGE - encoder->bitdepth - log2_tr_size; //!< Represents scaling through forward transform
-  const int32_t q_bits = QUANT_SHIFT + qp_scaled/6 + transform_shift;
-  const int32_t add = ((state->global->slicetype == KVZ_SLICE_I) ? 171 : 85) << (q_bits - 9);
-  const int32_t q_bits8 = q_bits - 8;
-
-  uint32_t ac_sum = 0;
-
-  for (int32_t n = 0; n < width * height; n++) {
-    int32_t level;
-    int32_t  sign;
-
-    level = coef[n];
-    sign  = (level < 0 ? -1: 1);
-
-    level = ((int64_t)abs(level) * quant_coeff[n] + add) >> q_bits;
-    ac_sum += level;
-
-    level *= sign;
-    q_coef[n] = (coeff_t)(CLIP( -32768, 32767, level));
-  }
-
-  if (!(encoder->sign_hiding && ac_sum >= 2)) return;
-
-  int32_t delta_u[LCU_WIDTH*LCU_WIDTH >> 2];
-
-  for (int32_t n = 0; n < width * height; n++) {
-    int32_t level;
-    level = coef[n];
-    level = ((int64_t)abs(level) * quant_coeff[n] + add) >> q_bits;
-    delta_u[n] = (int32_t)(((int64_t)abs(coef[n]) * quant_coeff[n] - (level << q_bits)) >> q_bits8);
-  }
-
-  if(ac_sum >= 2) {
-    #define SCAN_SET_SIZE 16
-    #define LOG2_SCAN_SET_SIZE 4
-    int32_t n,last_cg = -1, abssum = 0, subset, subpos;
-    for(subset = (width*height - 1)>>LOG2_SCAN_SET_SIZE; subset >= 0; subset--) {
-      int32_t first_nz_pos_in_cg = SCAN_SET_SIZE, last_nz_pos_in_cg=-1;
-      subpos = subset<<LOG2_SCAN_SET_SIZE;
-      abssum = 0;
-
-      // Find last coeff pos
-      for (n = SCAN_SET_SIZE - 1; n >= 0; n--)  {
-        if (q_coef[scan[n + subpos]])  {
-          last_nz_pos_in_cg = n;
-          break;
-        }
-      }
-
-      // First coeff pos
-      for (n = 0; n <SCAN_SET_SIZE; n++) {
-        if (q_coef[scan[n + subpos]]) {
-          first_nz_pos_in_cg = n;
-          break;
-        }
-      }
-
-      // Sum all kvz_quant coeffs between first and last
-      for(n = first_nz_pos_in_cg; n <= last_nz_pos_in_cg; n++) {
-        abssum += q_coef[scan[n + subpos]];
-      }
-
-      if(last_nz_pos_in_cg >= 0 && last_cg == -1) {
-        last_cg = 1;
-      }
-
-      if(last_nz_pos_in_cg - first_nz_pos_in_cg >= 4) {
-        int32_t signbit = (q_coef[scan[subpos + first_nz_pos_in_cg]] > 0 ? 0 : 1) ;
-        if(signbit != (abssum&0x1)) { // compare signbit with sum_parity
-          int32_t min_cost_inc = 0x7fffffff,  min_pos =-1, cur_cost=0x7fffffff;
-          int16_t final_change = 0, cur_change=0;
-          for(n = (last_cg == 1 ? last_nz_pos_in_cg : SCAN_SET_SIZE - 1); n >= 0; n--) {
-            uint32_t blkPos  = scan[n + subpos];
-            if(q_coef[blkPos] != 0) {
-              if(delta_u[blkPos] > 0) {
-                cur_cost = -delta_u[blkPos];
-                cur_change=1;
-              } else if(n == first_nz_pos_in_cg && abs(q_coef[blkPos]) == 1) {
-                cur_cost=0x7fffffff;
-              } else {
-                cur_cost = delta_u[blkPos];
-                cur_change =-1;
-              }
-            } else if(n < first_nz_pos_in_cg && ((coef[blkPos] >= 0)?0:1) != signbit) {
-              cur_cost = 0x7fffffff;
-            } else {
-              cur_cost   = -delta_u[blkPos];
-              cur_change = 1;
-            }
-
-            if(cur_cost < min_cost_inc) {
-              min_cost_inc = cur_cost;
-              final_change = cur_change;
-              min_pos      = blkPos;
-            }
-          } // CG loop
-
-          if(q_coef[min_pos] == 32767 || q_coef[min_pos] == -32768) {
-            final_change = -1;
-          }
-
-          if(coef[min_pos] >= 0) q_coef[min_pos] += final_change;
-          else q_coef[min_pos] -= final_change;
-        } // Hide
-      }
-      if (last_cg == 1) last_cg=0;
-    }
-
-    #undef SCAN_SET_SIZE
-    #undef LOG2_SCAN_SET_SIZE
-  }
-}
-
 /**
  * \brief inverse quantize transformed and quantized coefficents
  *

src/transform.h

@@ -35,8 +35,6 @@ extern const int16_t kvz_g_inv_quant_scales[6];
 
 
 
-void kvz_quant(const encoder_state_t *state, coeff_t *coef, coeff_t *q_coef, int32_t width,
-           int32_t height, int8_t type, int8_t scan_idx, int8_t block_type);
 void kvz_dequant(const encoder_state_t *state, coeff_t *q_coef, coeff_t *coef, int32_t width, int32_t height, int8_t type, int8_t block_type);
 
 void kvz_transformskip(const encoder_control_t *encoder, int16_t *block,int16_t *coeff, int8_t block_size);