Function to switch filters to a simplier one based on scale factors.

BUG=none
TEST=untested
R=tpsiaki@google.com

Review URL: https://webrtc-codereview.appspot.com/4989004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@894 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 892
+Version: 894
 License: BSD
 License File: LICENSE
 

include/libyuv/scale.h

@@ -18,12 +18,12 @@ namespace libyuv {
 extern "C" {
 #endif
 
-// Supported filtering
+// Supported filtering.
 enum FilterMode {
   kFilterNone = 0,  // Point sample; Fastest.
-  kFilterBilinear = 1,  // Faster than box, but lower quality scaling down.
-  kFilterBox = 2,  // Highest quality.
-  kFilterLinear = 3  // Faster than bilinear, slower than None.
+  kFilterLinear = 1, // Filter horizontally only.
+  kFilterBilinear = 2,  // Faster than box, but lower quality scaling down.
+  kFilterBox = 3  // Highest quality.
 };
 
 // Scale a YUV plane.

include/libyuv/scale_row.h

@@ -55,6 +55,10 @@ extern "C" {
 #define HAS_SCALEROWDOWN38_MIPS_DSPR2
 #endif
 
+FilterMode ScaleFilterReduce(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             FilterMode filtering);
+
 // Scale ARGB vertically with bilinear interpolation.
 void ScalePlaneVertical(int src_height,
                         int dst_width, int dst_height,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 892
+#define LIBYUV_VERSION 894
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/scale.cc

@@ -768,6 +768,11 @@ void ScalePlane(const uint8* src, int src_stride,
                 uint8* dst, int dst_stride,
                 int dst_width, int dst_height,
                 FilterMode filtering) {
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
   // Negative height means invert the image.
   if (src_height < 0) {
     src_height = -src_height;

source/scale_argb.cc

@@ -572,6 +572,16 @@ static void ScaleARGB(const uint8* src, int src_stride,
                       int dst_width, int dst_height,
                       int clip_x, int clip_y, int clip_width, int clip_height,
                       FilterMode filtering) {
+  // ARGB does not support box filter yet, but allow the user to pass it.
+  // TODO(fbarchard): Support Box filter.  Move row function to common.
+  if (filtering == kFilterBox) {
+    filtering = kFilterBilinear;
+  }
+  // Simplify filtering when possible.
+  filtering = ScaleFilterReduce(src_width, src_height,
+                                dst_width, dst_height,
+                                filtering);
+
   // Negative src_height means invert the image.
   if (src_height < 0) {
     src_height = -src_height;

source/scale_common.cc

@@ -460,18 +460,20 @@ void ScalePlaneVertical(int src_height,
                         const uint8* src_argb, uint8* dst_argb,
                         int x, int y, int dy,
                         int bpp, FilterMode filtering) {
-  int dst_widthx4 = dst_width * bpp;
-  src_argb += (x >> 16) * bpp;
+  // TODO(fbarchard): Allow higher bpp.
+  assert(bpp >= 1 && bpp <= 4);
   assert(src_height != 0);
   assert(dst_width > 0);
   assert(dst_height > 0);
+  int dst_width_bytes = dst_width * bpp;
+  src_argb += (x >> 16) * bpp;
   void (*InterpolateRow)(uint8* dst_argb, const uint8* src_argb,
       ptrdiff_t src_stride, int dst_width, int source_y_fraction) =
       InterpolateRow_C;
 #if defined(HAS_INTERPOLATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && dst_widthx4 >= 16) {
+  if (TestCpuFlag(kCpuHasSSE2) && dst_width_bytes >= 16) {
     InterpolateRow = InterpolateRow_Any_SSE2;
-    if (IS_ALIGNED(dst_widthx4, 16)) {
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
       InterpolateRow = InterpolateRow_Unaligned_SSE2;
       if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
           IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
@@ -481,9 +483,9 @@ void ScalePlaneVertical(int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_SSSE3)
-  if (TestCpuFlag(kCpuHasSSSE3) && dst_widthx4 >= 16) {
+  if (TestCpuFlag(kCpuHasSSSE3) && dst_width_bytes >= 16) {
     InterpolateRow = InterpolateRow_Any_SSSE3;
-    if (IS_ALIGNED(dst_widthx4, 16)) {
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
       InterpolateRow = InterpolateRow_Unaligned_SSSE3;
       if (IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride, 16) &&
           IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride, 16)) {
@@ -493,27 +495,27 @@ void ScalePlaneVertical(int src_height,
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_AVX2)
-  if (TestCpuFlag(kCpuHasAVX2) && dst_widthx4 >= 32) {
+  if (TestCpuFlag(kCpuHasAVX2) && dst_width_bytes >= 32) {
     InterpolateRow = InterpolateRow_Any_AVX2;
-    if (IS_ALIGNED(dst_widthx4, 32)) {
+    if (IS_ALIGNED(dst_width_bytes, 32)) {
       InterpolateRow = InterpolateRow_AVX2;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROW_NEON)
-  if (TestCpuFlag(kCpuHasNEON) && dst_widthx4 >= 16) {
+  if (TestCpuFlag(kCpuHasNEON) && dst_width_bytes >= 16) {
     InterpolateRow = InterpolateRow_Any_NEON;
-    if (IS_ALIGNED(dst_widthx4, 16)) {
+    if (IS_ALIGNED(dst_width_bytes, 16)) {
       InterpolateRow = InterpolateRow_NEON;
     }
   }
 #endif
 #if defined(HAS_INTERPOLATEROWS_MIPS_DSPR2)
-  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_widthx4 >= 4 &&
+  if (TestCpuFlag(kCpuHasMIPS_DSPR2) && dst_width_bytes >= 4 &&
       IS_ALIGNED(src_argb, 4) && IS_ALIGNED(src_stride, 4) &&
       IS_ALIGNED(dst_argb, 4) && IS_ALIGNED(dst_stride, 4)) {
     InterpolateRow = InterpolateRow_Any_MIPS_DSPR2;
-    if (IS_ALIGNED(dst_widthx4, 4)) {
+    if (IS_ALIGNED(dst_width_bytes, 4)) {
       InterpolateRow = InterpolateRow_MIPS_DSPR2;
     }
   }
@@ -526,12 +528,53 @@ void ScalePlaneVertical(int src_height,
     int yi = y >> 16;
     int yf = filtering ? ((y >> 8) & 255) : 0;
     const uint8* src = src_argb + yi * src_stride;
-    InterpolateRow(dst_argb, src, src_stride, dst_widthx4, yf);
+    InterpolateRow(dst_argb, src, src_stride, dst_width_bytes, yf);
     dst_argb += dst_stride;
     y += dy;
   }
 }
 
+// Scale plane vertically with bilinear interpolation.
+FilterMode ScaleFilterReduce(int src_width, int src_height,
+                             int dst_width, int dst_height,
+                             FilterMode filtering) {
+  if (src_width < 0) {
+    src_width = -src_width;
+  }
+  if (src_height < 0) {
+    src_height = -src_height;
+  }
+  if (filtering == kFilterBox) {
+    // If scaling both axis to 0.5 or larger, switch from Box to Bilinear.
+    if (dst_width * 2 >= src_width && dst_height * 2 >= src_height) {
+      filtering = kFilterBilinear;
+    }
+    // If scaling to larger, switch from Box to Bilinear.
+    if (dst_width >= src_width || dst_height >= src_height) {
+      filtering = kFilterBilinear;
+    }
+  }
+  if (filtering == kFilterBilinear) {
+    if (src_height == 1) {
+      filtering = kFilterLinear;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to Linear.
+    if (dst_height == src_height || dst_height * 3 == src_height) {
+      filtering = kFilterLinear;
+    }
+  }
+  if (filtering == kFilterLinear) {
+    if (src_width == 1) {
+      filtering = kFilterNone;
+    }
+    // TODO(fbarchard): Detect any odd scale factor and reduce to None.
+    if (dst_width == src_width || dst_width * 3 == src_width) {
+      filtering = kFilterNone;
+    }
+  }
+  return filtering;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv