p

CMakeLists.txt

@@ -143,7 +143,7 @@ target_include_directories(uvg266 PUBLIC src)
 target_include_directories(uvg266 PUBLIC src/extras)
 target_include_directories(uvg266 PUBLIC src/strategies)
 
-file(GLOB LIB_SOURCES_STRATEGIES_AVX2 RELATIVE ${PROJECT_SOURCE_DIR} "src/strategies/avx2/*.c")
+file(GLOB LIB_SOURCES_STRATEGIES_AVX2 RELATIVE ${PROJECT_SOURCE_DIR} "src/strategies/avx2/*.c" "src/dep_quant.c")
 file(GLOB LIB_SOURCES_STRATEGIES_SSE41 RELATIVE ${PROJECT_SOURCE_DIR} "src/strategies/sse41/*.c")
 
 set(CLI_SOURCES "src/encmain.c" "src/cli.c" "src/cli.h" "src/yuv_io.c" "src/yuv_io.h")

src/dep_quant.c

@@ -32,6 +32,8 @@
 
 #include "dep_quant.h"
 
+#include <immintrin.h>
+
 #include "cu.h"
 #include "encoderstate.h"
 #include "intra.h"
@@ -888,15 +890,63 @@ int uvg_dep_quant(
     height >= 4) {
     firstTestPos =((width == 4 && height == 4) || (width == 8 && height == 8)) ? 7 : 15;
   }
-  uvg_find_first_non_zero_coeff(
+  //uvg_find_first_non_zero_coeff(srcCoeff, enableScalingLists, dep_quant_context, scan, q_coeff, &firstTestPos, width, height);
-    srcCoeff,
+  const int     default_quant_coeff = dep_quant_context.m_quant->m_QScale;
-    enableScalingLists,
+  const int32_t thres = dep_quant_context.m_quant->m_thresLast;
-    dep_quant_context,
+  int           temp = firstTestPos;
-    scan,
+  if (enableScalingLists) {
-    q_coeff,
+    for (; temp >= 0; (temp)--) {
-    &firstTestPos,
+      coeff_t thresTmp = thres / (4 * q_coeff[scan[(temp)]]);
-    width, 
+      if (abs(srcCoeff[scan[(temp)]]) > thresTmp) {
-    height);
+        break;
+      }
+    }
+  } else {
+    coeff_t thresTmp = thres / (4 * default_quant_coeff);
+    if (temp >= 16 && height >= 4) {
+      __m256i th = _mm256_set1_epi16(thresTmp);
+      temp -= 15;
+      for (; temp >= 0; temp -= 16) {
+        __m256i sbb_data;
+        if (width <= 4) {
+          sbb_data = _mm256_loadu_si256((__m256i const*)&srcCoeff[scan[temp]]);
+        } else if (width == 8) {
+          uint32_t i = scan[temp];
+          __m256i  first = _mm256_loadu_si256((__m256i const*)&srcCoeff[i]);
+          __m256i  second = _mm256_loadu_si256((__m256i const*)&srcCoeff[i + 12]);
+          sbb_data = _mm256_blend_epi32(first, second, 204);
+        } else {
+          int16_t  temp_d[16];
+          uint32_t i = scan[temp];
+          memcpy(temp_d, &srcCoeff[i], 8);
+          i += width;
+          memcpy(temp_d + 4, &srcCoeff[i], 8);
+          i += width;
+          memcpy(temp_d + 8, &srcCoeff[i], 8);
+          i += width;
+          memcpy(temp_d + 12, &srcCoeff[i], 8);
+
+          sbb_data = _mm256_loadu_si256((__m256i const*)temp_d);
+        }
+        sbb_data = _mm256_abs_epi16(sbb_data);
+
+        __m256i a = _mm256_cmpgt_epi16(sbb_data, th);
+        if (!_mm256_testz_si256(a, a)) {
+          if (temp >= 0) {
+            temp += 15;
+          }
+          break;
+        }
+      }
+    }
+    for (; temp >= 0; temp--) {
+      if (abs(srcCoeff[scan[(temp)]]) > thresTmp) {
+        break;
+      }
+    }
+  }
+
+  firstTestPos = temp;
   if (firstTestPos < 0) {
     return 0;
   }
@@ -962,7 +1012,6 @@ int uvg_dep_quant(
   const uint32_t height_in_sbb = MAX(height >> 2, 1);
   const uint32_t width_in_sbb = MAX(width >> 2, 1);
   
-  const int      default_quant_coeff = dep_quant_context.m_quant->m_QScale;
   //===== populate trellis =====
   for (int scanIdx = firstTestPos; scanIdx >= 0; scanIdx--) {
     uint32_t blkpos = scan[scanIdx];

src/strategies/avx2/depquant-avx2.c

@@ -1361,7 +1361,15 @@ void uvg_dep_quant_decide_and_update_avx2(
 }
 
 
-void uvg_find_first_non_zero_avx2(const coeff_t* srcCoeff, const bool enableScalingLists, context_store dep_quant_context, const uint32_t* const scan, const int32_t* q_coeff, int* firstTestPos, const int width, const int height)
+void uvg_find_first_non_zero_avx2(
+  const coeff_t* srcCoeff,
+  const bool enableScalingLists,
+  context_store dep_quant_context, 
+  const uint32_t* const scan,
+  const int32_t* q_coeff,
+  int* firstTestPos, 
+  const int width,
+  const int height)
 {
   const int default_quant_coeff = dep_quant_context.m_quant->m_QScale;
   const int32_t thres  = dep_quant_context.m_quant->m_thresLast;