ARGBPolynomial for applying a 3 term polynomial matrix to pixels.

BUG=265
TEST=ARGBPolynomial
R=thorcarpenter@google.com

Review URL: https://webrtc-codereview.appspot.com/2159005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@778 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 777
+Version: 778
 License: BSD
 License File: LICENSE
 

include/libyuv/planar_functions.h

@@ -210,6 +210,20 @@ int RGBColorTable(uint8* dst_argb, int dst_stride_argb,
                   const uint8* table_argb,
                   int x, int y, int width, int height);
 
+// Apply a 3 term polynomial to ARGB values.
+// poly points to a 4x4 matrix.  The first row is constants.  The 2nd row is
+// coefficients for b, g, r and a.  The 3rd row is coefficients for b squared,
+// g squared, r squared and a squared.  The 4rd row is coefficients for b to
+// the 3, g to the 3, r to the 3 and a to the 3.  The values are summed and
+// result clamped to 0 to 255.
+// A polynomial approximation can be dirived using software such as 'R'.
+
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height);
+
 // Quantize a rectangle of ARGB. Alpha unaffected.
 // scale is a 16 bit fractional fixed point scaler between 0 and 65535.
 // interval_size should be a value between 1 and 255.

include/libyuv/row.h

@@ -140,8 +140,11 @@ extern "C" {
 // The following are Windows only:
 // TODO(fbarchard): Port to gcc.
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
+// Effects:
 #define HAS_ARGBCOLORTABLEROW_X86
 #define HAS_RGBCOLORTABLEROW_X86
+#define HAS_ARGBPOLYNOMIALROW_SSE2
+
 // Caveat: Visual C 2012 required for AVX2.
 #if _MSC_VER >= 1700
 #define HAS_ARGBSHUFFLEROW_AVX2
@@ -320,6 +323,7 @@ typedef __declspec(align(32)) uint32 ulvec32[8];
 typedef __declspec(align(32)) uint8 ulvec8[32];
 
 #elif defined(__GNUC__)
+// Caveat GCC 4.2 to 4.7 have a known issue using vectors with const.
 #define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
 typedef int16 __attribute__((vector_size(16))) vec16;
 typedef int32 __attribute__((vector_size(16))) vec32;
@@ -1542,6 +1546,14 @@ void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
 void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
                      uint8* dst_argb, int width);
 
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width);
+
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width);
+
 // Divide num by div and return as 16.16 fixed point result.
 int FixedDiv_C(int num, int div);
 int FixedDiv_X86(int num, int div);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 777
+#define LIBYUV_VERSION 778
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -2032,6 +2032,38 @@ int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
   return 0;
 }
 
+// Apply a 4x4 polynomial to each ARGB pixel.
+LIBYUV_API
+int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
+                   uint8* dst_argb, int dst_stride_argb,
+                   const float* poly,
+                   int width, int height) {
+  if (!src_argb || !dst_argb || !poly || width <= 0 || height <= 0) {
+    return -1;
+  }
+  // Coalesce contiguous rows.
+  if (src_stride_argb == width * 4 && dst_stride_argb == width * 4) {
+    return ARGBPolynomial(src_argb, 0,
+                          dst_argb, 0,
+                          poly,
+                          width * height, 1);
+  }
+  void (*ARGBPolynomialRow)(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) = ARGBPolynomialRow_C;
+#if defined(HAS_ARGBPOLYNOMIALROW_SSE2)
+  if (TestCpuFlag(kCpuHasSSE2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_SSE2;
+  }
+#endif
+  for (int y = 0; y < height; ++y) {
+    ARGBPolynomialRow(src_argb, dst_argb, poly, width);
+    src_argb += src_stride_argb;
+    dst_argb += dst_stride_argb;
+  }
+  return 0;
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv

source/row_common.cc

@@ -2009,8 +2009,48 @@ void UYVYToARGBRow_Unaligned_SSSE3(const uint8* src_uyvy,
   UYVYToYRow_Unaligned_SSE2(src_uyvy, row_y, width);
   I422ToARGBRow_Unaligned_SSSE3(row_y, row_u, row_v, dst_argb, width);
 }
+
 #endif  // defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)
 #endif  // !defined(LIBYUV_DISABLE_X86)
+
+void ARGBPolynomialRow_C(const uint8* src_argb,
+                         uint8* dst_argb, const float* poly,
+                         int width) {
+  for (int i = 0; i < width; ++i) {
+    float b = static_cast<float>(src_argb[0]);
+    float g = static_cast<float>(src_argb[1]);
+    float r = static_cast<float>(src_argb[2]);
+    float a = static_cast<float>(src_argb[3]);
+    float b2 = b * b;
+    float g2 = g * g;
+    float r2 = r * r;
+    float a2 = a * a;
+    float db = poly[0] + poly[4] * b;
+    float dg = poly[1] + poly[5] * g;
+    float dr = poly[2] + poly[6] * r;
+    float da = poly[3] + poly[7] * a;
+    db += poly[8] * b2;
+    dg += poly[9] * g2;
+    dr += poly[10] * r2;
+    da += poly[11] * a2;
+    float b3 = b2 * b;
+    float g3 = g2 * g;
+    float r3 = r2 * r;
+    float a3 = a2 * a;
+    db += poly[12] * b3;
+    dg += poly[13] * g3;
+    dr += poly[14] * r3;
+    da += poly[15] * a3;
+
+    dst_argb[0]= Clamp(static_cast<int32>(db));
+    dst_argb[1]= Clamp(static_cast<int32>(dg));
+    dst_argb[2]= Clamp(static_cast<int32>(dr));
+    dst_argb[3]= Clamp(static_cast<int32>(da));
+    src_argb += 4;
+    dst_argb += 4;
+  }
+}
+
 #undef clamp0
 #undef clamp255
 

source/row_win.cc

@@ -6767,6 +6767,53 @@ int FixedDiv_X86(int num, int div) {
   }
 }
 #endif  // HAS_FIXEDDIV_X86
+
+#ifdef HAS_ARGBPOLYNOMIALROW_SSE2
+__declspec(naked) __declspec(align(16))
+void ARGBPolynomialRow_SSE2(const uint8* src_argb,
+                            uint8* dst_argb, const float* poly,
+                            int width) {
+  __asm {
+    mov        eax, [esp + 12]   /* poly */
+    movdqu     xmm4, [eax]
+    movdqu     xmm5, [eax + 16]
+    movdqu     xmm6, [eax + 32]
+    movdqu     xmm7, [eax + 48]
+
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 16]  /* width */
+    pxor       xmm3, xmm3  // 4 bytes to 4 ints
+
+    align      16
+ convertloop:
+    movd       xmm0, [eax]  // BGRA
+    lea        eax, [eax + 4]
+    punpcklbw  xmm0, xmm3
+    punpcklwd  xmm0, xmm3
+    cvtdq2ps   xmm0, xmm0  // 4 floats
+    movdqa     xmm1, xmm0  // X
+    mulps      xmm0, xmm5  // C1 * X
+    addps      xmm0, xmm4  // result = C0 + C1 * X
+    movdqa     xmm2, xmm1
+    mulps      xmm2, xmm1  // X * X
+    mulps      xmm1, xmm2  // X * X * X
+    mulps      xmm2, xmm6  // C2 * X * X
+    mulps      xmm1, xmm7  // C3 * X * X * X
+    addps      xmm0, xmm2  // result += C2 * X * X
+    addps      xmm0, xmm1  // result += C3 * X * X * X
+    cvttps2dq  xmm0, xmm0
+    packuswb   xmm0, xmm0
+    packuswb   xmm0, xmm0
+    sub        ecx, 1
+    movd       [edx], xmm0
+    lea        edx, [edx + 4]
+    jg         convertloop
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_SSE2
+
 #endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
 
 #ifdef __cplusplus

unit_test/planar_test.cc

@@ -1656,4 +1656,68 @@ TEST_F(libyuvTest, ARGBBlur_Opt) {
   EXPECT_LE(max_diff, 1);
 }
 
+TEST_F(libyuvTest, TestARGBPolynomial) {
+  SIMD_ALIGNED(uint8 orig_pixels[1280][4]);
+  SIMD_ALIGNED(uint8 dst_pixels[1280][4]);
+
+  static const float kWarmifyPolynomial[16] = {
+    0.94230f,  -3.03300f,    -2.92500f,  0.f,  // C0
+    0.584500f,  1.112000f,    1.535000f, 1.f,  // C1 x
+    0.001313f, -0.002503f,   -0.004496f, 0.f,  // C2 x * x
+    0.0f,       0.000006965f, 0.000008781f, 0.f,  // C3 x * x * x
+ };
+
+  // Test blue
+  orig_pixels[0][0] = 255u;
+  orig_pixels[0][1] = 0u;
+  orig_pixels[0][2] = 0u;
+  orig_pixels[0][3] = 128u;
+  // Test green
+  orig_pixels[1][0] = 0u;
+  orig_pixels[1][1] = 255u;
+  orig_pixels[1][2] = 0u;
+  orig_pixels[1][3] = 0u;
+  // Test red
+  orig_pixels[2][0] = 0u;
+  orig_pixels[2][1] = 0u;
+  orig_pixels[2][2] = 255u;
+  orig_pixels[2][3] = 255u;
+  // Test color
+  orig_pixels[3][0] = 16u;
+  orig_pixels[3][1] = 64u;
+  orig_pixels[3][2] = 192u;
+  orig_pixels[3][3] = 224u;
+  // Do 16 to test asm version.
+  ARGBPolynomial(&orig_pixels[0][0], 0, &dst_pixels[0][0], 0,
+                 &kWarmifyPolynomial[0], 16, 1);
+  EXPECT_EQ(235u, dst_pixels[0][0]);
+  EXPECT_EQ(0u, dst_pixels[0][1]);
+  EXPECT_EQ(0u, dst_pixels[0][2]);
+  EXPECT_EQ(128u, dst_pixels[0][3]);
+  EXPECT_EQ(0u, dst_pixels[1][0]);
+  EXPECT_EQ(233u, dst_pixels[1][1]);
+  EXPECT_EQ(0u, dst_pixels[1][2]);
+  EXPECT_EQ(0u, dst_pixels[1][3]);
+  EXPECT_EQ(0u, dst_pixels[2][0]);
+  EXPECT_EQ(0u, dst_pixels[2][1]);
+  EXPECT_EQ(241u, dst_pixels[2][2]);
+  EXPECT_EQ(255u, dst_pixels[2][3]);
+  EXPECT_EQ(10u, dst_pixels[3][0]);
+  EXPECT_EQ(59u, dst_pixels[3][1]);
+  EXPECT_EQ(188u, dst_pixels[3][2]);
+  EXPECT_EQ(224u, dst_pixels[3][3]);
+
+  for (int i = 0; i < 1280; ++i) {
+    orig_pixels[i][0] = i;
+    orig_pixels[i][1] = i / 2;
+    orig_pixels[i][2] = i / 3;
+    orig_pixels[i][3] = i;
+  }
+  for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
+    ARGBPolynomial(&orig_pixels[0][0], 0, &dst_pixels[0][0], 0,
+                   &kWarmifyPolynomial[0], 1280, 1);
+  }
+}
+
+
 }  // namespace libyuv