Subtract 2 images.

BUG=179
TEST=planar unittest
Review URL: https://webrtc-codereview.appspot.com/1073004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@551 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 550
+Version: 551
 License: BSD
 License File: LICENSE
 

include/libyuv/planar_functions.h

@@ -209,20 +209,27 @@ int ARGBBlend(const uint8* src_argb0, int src_stride_argb0,
               uint8* dst_argb, int dst_stride_argb,
               int width, int height);
 
-// Multiply ARGB image by ARGB image.
+// Multiply ARGB image by ARGB image. Shifted down by 8. Saturates to 255.
 LIBYUV_API
 int ARGBMultiply(const uint8* src_argb0, int src_stride_argb0,
                  const uint8* src_argb1, int src_stride_argb1,
                  uint8* dst_argb, int dst_stride_argb,
                  int width, int height);
 
-// Add ARGB image with ARGB image.
+// Add ARGB image with ARGB image. Saturates to 255.
 LIBYUV_API
 int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
             const uint8* src_argb1, int src_stride_argb1,
             uint8* dst_argb, int dst_stride_argb,
             int width, int height);
 
+// Subtract ARGB image (argb1) from ARGB image (argb0). Saturates to 0.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height);
+
 // Convert I422 to YUY2.
 LIBYUV_API
 int I422ToYUY2(const uint8* src_y, int src_stride_y,

include/libyuv/row.h

@@ -111,6 +111,7 @@ extern "C" {
 #define HAS_ARGBQUANTIZEROW_SSE2
 #define HAS_ARGBSEPIAROW_SSSE3
 #define HAS_ARGBSHADEROW_SSE2
+#define HAS_ARGBSUBTRACTROW_SSE2
 #define HAS_ARGBUNATTENUATEROW_SSE2
 #define HAS_COMPUTECUMULATIVESUMROW_SSE2
 #define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
@@ -969,7 +970,7 @@ void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1,
 void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
                     uint8* dst_argb, int width);
 
-// ARGB multiply images.  Same API as Blend, but these require
+// ARGB multiply images. Same API as Blend, but these require
 // pointer and width alignment for SSE2.
 void ARGBMultiplyRow_C(const uint8* src_argb, const uint8* src_argb1,
                        uint8* dst_argb, int width);
@@ -994,6 +995,19 @@ void ARGBAddRow_NEON(const uint8* src_argb, const uint8* src_argb1,
 void ARGBAddRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
                          uint8* dst_argb, int width);
 
+// ARGB subtract images. Same API as Blend, but these require
+// pointer and width alignment for SSE2.
+void ARGBSubtractRow_C(const uint8* src_argb, const uint8* src_argb1,
+                       uint8* dst_argb, int width);
+void ARGBSubtractRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+void ARGBSubtractRow_NEON(const uint8* src_argb, const uint8* src_argb1,
+                          uint8* dst_argb, int width);
+void ARGBSubtractRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
+                              uint8* dst_argb, int width);
+
 void ARGBToRGB24Row_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
 void ARGBToRAWRow_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
 void ARGBToRGB565Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 550
+#define LIBYUV_VERSION 551
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -498,6 +498,53 @@ int ARGBAdd(const uint8* src_argb0, int src_stride_argb0,
   return 0;
 }
 
+// Subtract 2 ARGB images and store to destination.
+LIBYUV_API
+int ARGBSubtract(const uint8* src_argb0, int src_stride_argb0,
+                 const uint8* src_argb1, int src_stride_argb1,
+                 uint8* dst_argb, int dst_stride_argb,
+                 int width, int height) {
+  if (!src_argb0 || !src_argb1 || !dst_argb || width <= 0 || height == 0) {
+    return -1;
+  }
+  // Negative height means invert the image.
+  if (height < 0) {
+    height = -height;
+    dst_argb = dst_argb + (height - 1) * dst_stride_argb;
+    dst_stride_argb = -dst_stride_argb;
+  }
+
+  void (*ARGBSubtractRow)(const uint8* src0, const uint8* src1, uint8* dst,
+                          int width) = ARGBSubtractRow_C;
+#if defined(HAS_ARGBSUBTRACTROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
+      IS_ALIGNED(src_argb0, 16) && IS_ALIGNED(src_stride_argb0, 16) &&
+      IS_ALIGNED(src_argb1, 16) && IS_ALIGNED(src_stride_argb1, 16) &&
+      IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBSubtractRow = ARGBSubtractRow_SSE2;
+    }
+  }
+#elif defined(HAS_ARGBSUBTRACTROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBSubtractRow = ARGBSubtractRow_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBSubtractRow = ARGBSubtractRow_NEON;
+    }
+  }
+#endif
+
+  // Subtract plane
+  for (int y = 0; y < height; ++y) {
+    ARGBSubtractRow(src_argb0, src_argb1, dst_argb, width);
+    src_argb0 += src_stride_argb0;
+    src_argb1 += src_stride_argb1;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
 // Convert I422 to BGRA.
 LIBYUV_API
 int I422ToBGRA(const uint8* src_y, int src_stride_y,

source/row_any.cc

@@ -390,6 +390,10 @@ MATHROW_ANY(ARGBMultiplyRow_Any_SSE2, ARGBMultiplyRow_SSE2, ARGBMultiplyRow_C,
 #ifdef HAS_ARGBADDROW_SSE2
 MATHROW_ANY(ARGBAddRow_Any_SSE2, ARGBAddRow_SSE2, ARGBAddRow_C, 3)
 #endif
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+MATHROW_ANY(ARGBSubtractRow_Any_SSE2, ARGBSubtractRow_SSE2, ARGBSubtractRow_C,
+            3)
+#endif
 #ifdef HAS_ARGBMULTIPLYROW_NEON
 MATHROW_ANY(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, ARGBMultiplyRow_C,
             7)
@@ -397,6 +401,10 @@ MATHROW_ANY(ARGBMultiplyRow_Any_NEON, ARGBMultiplyRow_NEON, ARGBMultiplyRow_C,
 #ifdef HAS_ARGBADDROW_NEON
 MATHROW_ANY(ARGBAddRow_Any_NEON, ARGBAddRow_NEON, ARGBAddRow_C, 7)
 #endif
+#ifdef HAS_ARGBSUBTRACTROW_NEON
+MATHROW_ANY(ARGBSubtractRow_Any_NEON, ARGBSubtractRow_NEON, ARGBSubtractRow_C,
+            7)
+#endif
 #undef MATHROW_ANY
 
 #ifdef __cplusplus

source/row_common.cc

@@ -751,6 +751,30 @@ void ARGBAddRow_C(const uint8* src_argb0, const uint8* src_argb1,
 }
 #undef SHADE
 
+#define SHADE(f, v) (v >= f) ? 0 : (f - v)
+
+void ARGBSubtractRow_C(const uint8* src_argb0, const uint8* src_argb1,
+                       uint8* dst_argb, int width) {
+  for (int i = 0; i < width; ++i) {
+    const uint32 b = src_argb0[0];
+    const uint32 g = src_argb0[1];
+    const uint32 r = src_argb0[2];
+    const uint32 a = src_argb0[3];
+    const uint32 b_sub = src_argb1[0];
+    const uint32 g_sub = src_argb1[1];
+    const uint32 r_sub = src_argb1[2];
+    const uint32 a_sub = src_argb1[3];
+    dst_argb[0] = SHADE(b, b_sub);
+    dst_argb[1] = SHADE(g, g_sub);
+    dst_argb[2] = SHADE(r, r_sub);
+    dst_argb[3] = SHADE(a, a_sub);
+    src_argb0 += 4;
+    src_argb1 += 4;
+    dst_argb += 4;
+  }
+}
+#undef SHADE
+
 void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width) {
   // Copy a Y to RGB.
   for (int x = 0; x < width; ++x) {

source/row_neon.cc

@@ -2552,7 +2552,31 @@ void ARGBAddRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
     "+r"(dst_argb),   // %2
     "+r"(width)       // %3
   :
-  : "cc", "memory", "q0", "q1"
+  : "cc", "memory", "q0", "q1", "q2", "q3"
+  );
+}
+
+// Subtract 2 rows of ARGB pixels, 8 pixels at a time.
+void ARGBSubtractRow_NEON(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    // 8 pixel loop.
+    ".p2align  2                               \n"
+  "1:                                          \n"
+    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "vld4.8     {d4, d5, d6, d7}, [%1]!        \n"  // load 8 more ARGB pixels.
+    "subs       %3, %3, #8                     \n"  // 8 processed per loop.
+    "vqsub.u8   q0, q0, q2                     \n"  // subtract B, G
+    "vqsub.u8   q1, q1, q3                     \n"  // subtract R, A
+    "vst4.8     {d0, d1, d2, d3}, [%2]!        \n"  // store 8 ARGB pixels.
+    "bgt        1b                             \n"
+
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "cc", "memory", "q0", "q1", "q2", "q3"
   );
 }
 

source/row_posix.cc

@@ -4007,7 +4007,6 @@ void ARGBMultiplyRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
 void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
                      uint8* dst_argb, int width) {
   asm volatile (
-    "pxor      %%xmm5,%%xmm5                   \n"
     "sub       %0,%1                           \n"
     "sub       %0,%2                           \n"
 
@@ -4028,12 +4027,44 @@ void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
   :
   : "memory", "cc"
 #if defined(__SSE2__)
-    , "xmm0", "xmm1", "xmm5"
+    , "xmm0", "xmm1"
 #endif
   );
 }
 #endif  // HAS_ARGBADDROW_SSE2
 
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels, 4 pixels at a time.
+// Aligned to 16 bytes.
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  asm volatile (
+    "sub       %0,%1                           \n"
+    "sub       %0,%2                           \n"
+
+    // 4 pixel loop.
+    ".p2align  4                               \n"
+  "1:                                          \n"
+    "movdqa    (%0),%%xmm0                     \n"
+    "movdqa    (%0,%1),%%xmm1                  \n"
+    "psubusb   %%xmm1,%%xmm0                   \n"
+    "sub       $0x4,%3                         \n"
+    "movdqa    %%xmm0,(%0,%2,1)                \n"
+    "lea       0x10(%0),%0                     \n"
+    "jg        1b                              \n"
+  : "+r"(src_argb0),  // %0
+    "+r"(src_argb1),  // %1
+    "+r"(dst_argb),   // %2
+    "+r"(width)       // %3
+  :
+  : "memory", "cc"
+#if defined(__SSE2__)
+    , "xmm0", "xmm1"
+#endif
+  );
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
 #ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2
 // Creates a table of cumulative sums where each value is a sum of all values
 // above and to the left of the value, inclusive of the value.

source/row_win.cc

@@ -4348,6 +4348,37 @@ void ARGBAddRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
 }
 #endif  // HAS_ARGBADDROW_SSE2
 
+#ifdef HAS_ARGBSUBTRACTROW_SSE2
+// Subtract 2 rows of ARGB pixels together, 4 pixels at a time.
+// Aligned to 16 bytes.
+__declspec(naked) __declspec(align(16))
+void ARGBSubtractRow_SSE2(const uint8* src_argb0, const uint8* src_argb1,
+                          uint8* dst_argb, int width) {
+  __asm {
+    push       esi
+    mov        eax, [esp + 4 + 4]   // src_argb0
+    mov        esi, [esp + 4 + 8]   // src_argb1
+    mov        edx, [esp + 4 + 12]  // dst_argb
+    mov        ecx, [esp + 4 + 16]  // width
+    sub        esi, eax
+    sub        edx, eax
+
+    align      16
+ convertloop:
+    movdqa     xmm0, [eax]        // read 4 pixels from src_argb0
+    movdqa     xmm1, [eax + esi]  // read 4 pixels from src_argb1
+    psubusb    xmm0, xmm1         // src_argb0 - src_argb1
+    sub        ecx, 4
+    movdqa     [eax + edx], xmm0
+    lea        eax, [eax + 16]
+    jg         convertloop
+
+    pop        esi
+    ret
+  }
+}
+#endif  // HAS_ARGBSUBTRACTROW_SSE2
+
 #ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
 // Consider float CumulativeSum.
 // Consider calling CumulativeSum one row at time as needed.

unit_test/planar_test.cc

@@ -1037,4 +1037,72 @@ TEST_F(libyuvTest, ARGBAdd_Opt) {
   EXPECT_LE(max_diff, 1);
 }
 
+static int TestSubtract(int width, int height, int benchmark_iterations,
+                        int invert, int off) {
+  const int kBpp = 4;
+  const int kStride = (width * kBpp + 15) & ~15;
+  align_buffer_64(src_argb_a, kStride * height + off);
+  align_buffer_64(src_argb_b, kStride * height + off);
+  align_buffer_64(dst_argb_c, kStride * height);
+  align_buffer_64(dst_argb_opt, kStride * height);
+  srandom(time(NULL));
+  for (int i = 0; i < kStride * height; ++i) {
+    src_argb_a[i + off] = (random() & 0xff);
+    src_argb_b[i + off] = (random() & 0xff);
+  }
+  memset(dst_argb_c, 0, kStride * height);
+  memset(dst_argb_opt, 0, kStride * height);
+
+  MaskCpuFlags(0);
+  ARGBSubtract(src_argb_a + off, kStride,
+               src_argb_b + off, kStride,
+               dst_argb_c, kStride,
+               width, invert * height);
+  MaskCpuFlags(-1);
+  for (int i = 0; i < benchmark_iterations; ++i) {
+    ARGBSubtract(src_argb_a + off, kStride,
+                 src_argb_b + off, kStride,
+                 dst_argb_opt, kStride,
+                 width, invert * height);
+  }
+  int max_diff = 0;
+  for (int i = 0; i < kStride * height; ++i) {
+    int abs_diff =
+        abs(static_cast<int>(dst_argb_c[i]) -
+            static_cast<int>(dst_argb_opt[i]));
+    if (abs_diff > max_diff) {
+      max_diff = abs_diff;
+    }
+  }
+  free_aligned_buffer_64(src_argb_a)
+  free_aligned_buffer_64(src_argb_b)
+  free_aligned_buffer_64(dst_argb_c)
+  free_aligned_buffer_64(dst_argb_opt)
+  return max_diff;
+}
+
+TEST_F(libyuvTest, ARGBSubtract_Any) {
+  int max_diff = TestSubtract(benchmark_width_ - 1, benchmark_height_,
+                              benchmark_iterations_, +1, 0);
+  EXPECT_LE(max_diff, 1);
+}
+
+TEST_F(libyuvTest, ARGBSubtract_Unaligned) {
+  int max_diff = TestSubtract(benchmark_width_, benchmark_height_,
+                              benchmark_iterations_, +1, 1);
+  EXPECT_LE(max_diff, 1);
+}
+
+TEST_F(libyuvTest, ARGBSubtract_Invert) {
+  int max_diff = TestSubtract(benchmark_width_, benchmark_height_,
+                              benchmark_iterations_, -1, 0);
+  EXPECT_LE(max_diff, 1);
+}
+
+TEST_F(libyuvTest, ARGBSubtract_Opt) {
+  int max_diff = TestSubtract(benchmark_width_, benchmark_height_,
+                              benchmark_iterations_, +1, 0);
+  EXPECT_LE(max_diff, 1);
+}
+
 }  // namespace libyuv