Utilize registers more efficiently for 8x8 and larger blocks

src/strategies/avx2/ipol-avx2.c

@@ -34,6 +34,7 @@
 #include "strategies/generic/picture-generic.h"
 #include "strategies/strategies-ipol.h"
 #include "strategyselector.h"
+#include "strategies/strategies-common.h"
 
 
 #define FILTER_OFFSET 3
@@ -267,6 +268,22 @@ void kvz_four_tap_filter_x4_hor_avx2(int8_t *filter, kvz_pixel *data, int shift,
   _mm_storel_epi64((__m128i*)dst, temp);
 }
 
+void kvz_four_tap_filter_x8_hor_avx2(int8_t *filter, kvz_pixel *data, int shift, int16_t* dst)
+{
+  __m256i packed_data = _mm256_inserti128_si256(_mm256_castsi128_si256(_mm_loadl_epi64((__m128i*)data)), _mm_loadl_epi64((__m128i*)(data + 4)), 1);
+  __m256i packed_filter = _mm256_set1_epi32(*(int32_t*)filter);
+  __m256i idx_lookup = _mm256_broadcastsi128_si256(_mm_setr_epi8(0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6));
+
+  __m256i temp = _mm256_shuffle_epi8(packed_data, idx_lookup);
+
+  temp = _mm256_maddubs_epi16(temp, packed_filter);
+  temp = _mm256_hadd_epi16(temp, temp);
+  temp = _mm256_srai_epi16(temp, shift);
+
+  _mm_storel_epi64((__m128i*)dst, _mm256_castsi256_si128(temp));
+  _mm_storel_epi64((__m128i*)(dst + 4), _mm256_extracti128_si256(temp, 1));
+}
+
 int32_t kvz_four_tap_filter_x4_ver_16bit_avx2(int8_t *filter, int16_t *data, int16_t stride, int offset, int shift2, int shift3)
 {
 
@@ -291,6 +308,31 @@ int32_t kvz_four_tap_filter_x4_ver_16bit_avx2(int8_t *filter, int16_t *data, int
   return _mm_cvtsi128_si32(temp);
 }
 
+void kvz_four_tap_filter_x8_ver_16bit_avx2(int8_t *filter, int16_t *data, int16_t stride, int offset, int shift2, int shift3, kvz_pixel* dst)
+{
+
+  __m256i v_filter = _mm256_cvtepi8_epi16(_mm_set1_epi16(*(int16_t*)&(filter[0])));
+  __m256i v_data0 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i*)(data + stride * 0)));
+  __m256i v_data1 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i*)(data + stride * 1)));
+  __m256i v_data = _mm256_or_si256(v_data0, _mm256_slli_epi32(v_data1, 16));
+  __m256i temp =  _mm256_madd_epi16(v_filter, v_data);
+
+  v_filter = _mm256_cvtepi8_epi16(_mm_set1_epi16(*(int16_t*)&(filter[2])));
+  __m256i v_data2 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i*)(data + stride * 2)));
+  __m256i v_data3 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i*)(data + stride * 3)));
+  v_data = _mm256_or_si256(v_data2, _mm256_slli_epi32(v_data3, 16));
+  temp = _mm256_add_epi32(temp, _mm256_madd_epi16(v_filter, v_data) );
+
+  temp = _mm256_add_epi32(temp, _mm256_set1_epi32(offset));
+  temp = _mm256_srai_epi32(temp, shift2 + shift3);
+
+  temp = _mm256_packus_epi32(temp, temp);
+  temp = _mm256_packus_epi16(temp, temp);
+
+  *(int32_t*)dst = _mm_cvtsi128_si32(_mm256_castsi256_si128(temp));
+  *(int32_t*)(dst + 4) = _mm_cvtsi128_si32(_mm256_extracti128_si256(temp, 1));
+}
+
 void kvz_filter_inter_quarterpel_luma_avx2(const encoder_control_t * const encoder, kvz_pixel *src, int16_t src_stride, int width, int height, kvz_pixel *dst, int16_t dst_stride, int8_t hor_flag, int8_t ver_flag)
 {
 
@@ -538,22 +580,45 @@ void kvz_sample_octpel_chroma_avx2(const encoder_control_t * const encoder, kvz_
 #define FILTER_OFFSET_C (FILTER_OFFSET / 2)
   int16_t hor_filtered[(LCU_WIDTH_C + 1) + FILTER_SIZE_C][(LCU_WIDTH_C + 1) + FILTER_SIZE_C];
 
-  // Filter horizontally and flip x and y
+  if (width == 4) {
-  for (y = 0; y < height + FILTER_SIZE_C - 1; ++y) {
+    // Filter horizontally and flip x and y
-    for (x = 0; x < width; x+=4) {
+    for (y = 0; y < height + FILTER_SIZE_C - 1; ++y) {
-      int ypos = y - FILTER_OFFSET_C;
+      for (x = 0; x < width; x += 4) {
-      int xpos = x - FILTER_OFFSET_C;
+        int ypos = y - FILTER_OFFSET_C;
-      int16_t *dst = &(hor_filtered[y][x]);
+        int xpos = x - FILTER_OFFSET_C;
-      kvz_four_tap_filter_x4_hor_avx2(hor_filter, &src[src_stride*ypos + xpos], shift1, dst);
+        int16_t *out = &(hor_filtered[y][x]);
+        kvz_four_tap_filter_x4_hor_avx2(hor_filter, &src[src_stride*ypos + xpos], shift1, out);
+      }
     }
-  }
 
-  // Filter vertically and flip x and y
+    // Filter vertically and flip x and y
-  for (y = 0; y < height; ++y) {
+    for (y = 0; y < height; ++y) {
-    for (x = 0; x < width; x+=4) {
+      for (x = 0; x < width; x+=4) {
-      int ypos = y;
+        int ypos = y;
-      int xpos = x;
+        int xpos = x;
-      *(int32_t*)&(dst[y*dst_stride + x]) = kvz_four_tap_filter_x4_ver_16bit_avx2(ver_filter, &hor_filtered[ypos][xpos], sizeof(hor_filtered[0])/sizeof(int16_t), offset23, shift2, shift3);
+        *(int32_t*)&(dst[y*dst_stride + x]) = kvz_four_tap_filter_x4_ver_16bit_avx2(ver_filter, &hor_filtered[ypos][xpos], sizeof(hor_filtered[0])/sizeof(int16_t), offset23, shift2, shift3);
+      }
+    }
+
+  } else {
+    // Filter horizontally and flip x and y
+    for (y = 0; y < height + FILTER_SIZE_C - 1; ++y) {
+      for (x = 0; x < width; x += 8) {
+        int ypos = y - FILTER_OFFSET_C;
+        int xpos = x - FILTER_OFFSET_C;
+        int16_t *dst = &(hor_filtered[y][x]);
+        kvz_four_tap_filter_x8_hor_avx2(hor_filter, &src[src_stride*ypos + xpos], shift1, dst);
+      }
+    }
+
+    // Filter vertically and flip x and y
+    for (y = 0; y < height; ++y) {
+      for (x = 0; x < width; x+=8) {
+        int ypos = y;
+        int xpos = x;
+        kvz_pixel *out = &(dst[y*dst_stride + x]);
+        kvz_four_tap_filter_x8_ver_16bit_avx2(ver_filter, &hor_filtered[ypos][xpos], sizeof(hor_filtered[0])/sizeof(int16_t), offset23, shift2, shift3, out);
+      }
     }
   }
 }