Merge branch '10bit_fixes'

ultravideo · Sep 5, 2024 · 5cc93dc · 5cc93dc
2 parents 2f9a214 + 9e61ac3
commit 5cc93dc
Show file tree

Hide file tree

Showing 5 changed files with 54 additions and 38 deletions.
diff --git a/src/image.c b/src/image.c
@@ -472,7 +472,7 @@ unsigned kvz_image_calc_satd(const kvz_picture *pic,
                              pic_data,
                              pic->stride,
                              ref_data,
-                             ref->stride) >> (KVZ_BIT_DEPTH - 8);
+                             ref->stride);
   } else {
     // Extrapolate pixels from outside the frame.
 
@@ -514,7 +514,7 @@ unsigned kvz_image_calc_satd(const kvz_picture *pic,
       pic_data,
       pic->stride,
       ext_origin,
-      ext_s) >> (KVZ_BIT_DEPTH - 8);
+      ext_s);
 
     return satd;
   }

diff --git a/src/rate_control.c b/src/rate_control.c
@@ -716,7 +716,7 @@ static double qp_to_lambda(encoder_state_t* const state, int qp)
     alpha = -state->frame->c_para[index] * state->frame->k_para[index];
     beta = state->frame->k_para[index] - 1;
   }
-
+  alpha *= (double)(1 << (KVZ_BIT_DEPTH - 8));
   double est_lambda;
   int est_qp;
   if (state->frame->is_irap && encoder->cfg.intra_bit_allocation) {

diff --git a/src/strategies/avx2/dct-avx2.c b/src/strategies/avx2/dct-avx2.c
@@ -38,7 +38,6 @@
 
 #if COMPILE_INTEL_AVX2
 #include "kvazaar.h"
-#if KVZ_BIT_DEPTH == 8
 #include <immintrin.h>
 
 #include "strategyselector.h"
@@ -938,30 +937,28 @@ static void matrix_i ## type ## _## n ## x ## n ## _avx2(int8_t bitdepth, const
 TRANSFORM(dct, 32);
 ITRANSFORM(dct, 32);
 
-#endif // KVZ_BIT_DEPTH == 8
+
 #endif //COMPILE_INTEL_AVX2
 
 int kvz_strategy_register_dct_avx2(void* opaque, uint8_t bitdepth)
 {
   bool success = true;
 #if COMPILE_INTEL_AVX2
-#if KVZ_BIT_DEPTH == 8
-  if (bitdepth == 8){
-    success &= kvz_strategyselector_register(opaque, "fast_forward_dst_4x4", "avx2", 40, &matrix_dst_4x4_avx2);
+  // Coefficients are the same for all bitdepths, no need to disable for 10-bit
+  success &= kvz_strategyselector_register(opaque, "fast_forward_dst_4x4", "avx2", 40, &matrix_dst_4x4_avx2);
 
-    success &= kvz_strategyselector_register(opaque, "dct_4x4", "avx2", 40, &matrix_dct_4x4_avx2);
-    success &= kvz_strategyselector_register(opaque, "dct_8x8", "avx2", 40, &matrix_dct_8x8_avx2);
-    success &= kvz_strategyselector_register(opaque, "dct_16x16", "avx2", 40, &matrix_dct_16x16_avx2);
-    success &= kvz_strategyselector_register(opaque, "dct_32x32", "avx2", 40, &matrix_dct_32x32_avx2);
+  success &= kvz_strategyselector_register(opaque, "dct_4x4", "avx2", 40, &matrix_dct_4x4_avx2);
+  success &= kvz_strategyselector_register(opaque, "dct_8x8", "avx2", 40, &matrix_dct_8x8_avx2);
+  success &= kvz_strategyselector_register(opaque, "dct_16x16", "avx2", 40, &matrix_dct_16x16_avx2);
+  success &= kvz_strategyselector_register(opaque, "dct_32x32", "avx2", 40, &matrix_dct_32x32_avx2);
 
-    success &= kvz_strategyselector_register(opaque, "fast_inverse_dst_4x4", "avx2", 40, &matrix_idst_4x4_avx2);
+  success &= kvz_strategyselector_register(opaque, "fast_inverse_dst_4x4", "avx2", 40, &matrix_idst_4x4_avx2);
 
-    success &= kvz_strategyselector_register(opaque, "idct_4x4", "avx2", 40, &matrix_idct_4x4_avx2);
-    success &= kvz_strategyselector_register(opaque, "idct_8x8", "avx2", 40, &matrix_idct_8x8_avx2);
-    success &= kvz_strategyselector_register(opaque, "idct_16x16", "avx2", 40, &matrix_idct_16x16_avx2);
-    success &= kvz_strategyselector_register(opaque, "idct_32x32", "avx2", 40, &matrix_idct_32x32_avx2);
-  }
-#endif // KVZ_BIT_DEPTH == 8
+  success &= kvz_strategyselector_register(opaque, "idct_4x4", "avx2", 40, &matrix_idct_4x4_avx2);
+  success &= kvz_strategyselector_register(opaque, "idct_8x8", "avx2", 40, &matrix_idct_8x8_avx2);
+  success &= kvz_strategyselector_register(opaque, "idct_16x16", "avx2", 40, &matrix_idct_16x16_avx2);
+  success &= kvz_strategyselector_register(opaque, "idct_32x32", "avx2", 40, &matrix_idct_32x32_avx2);
+
 #endif //COMPILE_INTEL_AVX2  
   return success;
 }
diff --git a/src/strategies/generic/sao_shared_generics.h b/src/strategies/generic/sao_shared_generics.h
@@ -67,10 +67,10 @@ static int sao_edge_ddistortion_generic(const kvz_pixel *orig_data,
       uint32_t a_pos = (y + a_ofs.y) * block_width + x + a_ofs.x;
       uint32_t b_pos = (y + b_ofs.y) * block_width + x + b_ofs.x;
 
-      uint8_t   a    =  rec_data[a_pos];
-      uint8_t   b    =  rec_data[b_pos];
-      uint8_t   c    =  rec_data[c_pos];
-      uint8_t   orig = orig_data[c_pos];
+      kvz_pixel a    =  rec_data[a_pos];
+      kvz_pixel b    =  rec_data[b_pos];
+      kvz_pixel c    =  rec_data[c_pos];
+      kvz_pixel orig = orig_data[c_pos];
 
       int32_t eo_cat = sao_calc_eo_cat(a, b, c);
       int32_t offset = offsets[eo_cat];

diff --git a/src/yuv_io.c b/src/yuv_io.c
@@ -57,24 +57,43 @@ static int read_and_fill_frame_data(FILE *file,
                                     unsigned width, unsigned height, unsigned bytes_per_sample,
                                     unsigned array_width, kvz_pixel *data)
 {
-  kvz_pixel* p = data;
-  kvz_pixel* end = data + array_width * height;
-  kvz_pixel fill_char;
-  unsigned i;
 
-  while (p < end) {
-    // Read the beginning of the line from input.
-    if (width != fread(p, bytes_per_sample, width, file))
-      return 0;
-
-    // Fill the rest with the last pixel value.
-    fill_char = p[width - 1];
+  unsigned   i;
+  // Handle separately the case where we use KVZ_BIT_DEPTH 10+ but the input is 8-bit.
+  if (bytes_per_sample != sizeof(kvz_pixel)) {
+    uint8_t* p = (uint8_t*)data;
+    uint8_t* end = (uint8_t*)data + array_width * height;
+    uint8_t  fill_char;
+    while (p < end) {
+      // Read the beginning of the line from input.
+      if (width != fread(p, bytes_per_sample, width, file)) return 0;
+      // Fill the rest with the last pixel value.
+      // Fill the rest with the last pixel value.
+      fill_char = p[width - 1];
+
+      for (i = width; i < array_width; ++i) {
+        p[i] = fill_char;
+      }
 
-    for (i = width; i < array_width; ++i) {
-      p[i] = fill_char;
+      p += array_width;
     }
+  }
+  else {
+    kvz_pixel* p = data;
+    kvz_pixel* end = data + array_width * height;
+    kvz_pixel  fill_char;
+    while (p < end) {
+      // Read the beginning of the line from input.
+      if (width != fread(p, bytes_per_sample, width, file)) return 0;
+      // Fill the rest with the last pixel value.
+      fill_char = p[width - 1];
+
+      for (i = width; i < array_width; ++i) {
+        p[i] = fill_char;
+      }
 
-    p += array_width;
+      p += array_width;
+    }
   }
   return 1;
 }
@@ -313,7 +332,7 @@ int yuv_io_seek(FILE* file, unsigned frames,
 
     // Seek failed. Skip data by reading.
     error = 0;
-    unsigned char* tmp[4096];
+    unsigned char tmp[4096];
     size_t bytes_left = skip_bytes;
     while (bytes_left > 0 && !error) {
       const size_t skip = MIN(4096, bytes_left);