AVX version of Polynomial

BUG=265
TEST=untested
R=thorcarpenter@google.com, yunqingwang@google.com

Review URL: https://webrtc-codereview.appspot.com/2166004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@780 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

@@ -1,6 +1,6 @@
 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 779
+Version: 780
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -142,8 +142,8 @@ extern "C" {
 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
 // Effects:
 #define HAS_ARGBCOLORTABLEROW_X86
-#define HAS_RGBCOLORTABLEROW_X86
 #define HAS_ARGBPOLYNOMIALROW_SSE2
+#define HAS_RGBCOLORTABLEROW_X86
 
 // Caveat: Visual C 2012 required for AVX2.
 #if _MSC_VER >= 1700
@@ -153,6 +153,7 @@ extern "C" {
 #define HAS_ARGBTOYROW_AVX2
 #define HAS_HALFROW_AVX2
 #define HAS_I422TOARGBROW_AVX2
+#define HAS_INTERPOLATEROW_AVX2
 #define HAS_MERGEUVROW_AVX2
 #define HAS_MIRRORROW_AVX2
 #define HAS_SPLITUVROW_AVX2
@@ -162,17 +163,17 @@ extern "C" {
 #define HAS_YUY2TOUV422ROW_AVX2
 #define HAS_YUY2TOUVROW_AVX2
 #define HAS_YUY2TOYROW_AVX2
-#define HAS_INTERPOLATEROW_AVX2
 
 // Effects:
 #define HAS_ARGBADDROW_AVX2
 #define HAS_ARGBATTENUATEROW_AVX2
 #define HAS_ARGBMIRRORROW_AVX2
 #define HAS_ARGBMULTIPLYROW_AVX2
+#define HAS_ARGBPOLYNOMIALROW_AVX2
 #define HAS_ARGBSUBTRACTROW_AVX2
 #define HAS_ARGBUNATTENUATEROW_AVX2
-#endif
+#endif  // _MSC_VER >= 1700
-#endif
+#endif  // defined(_MSC_VER)
 
 // The following are Yasm x86 only:
 // TODO(fbarchard): Port AVX2 to inline.
@@ -1549,10 +1550,12 @@ void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
 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);
+void ARGBPolynomialRow_AVX2(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);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 779
+#define LIBYUV_VERSION 780
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -2055,6 +2055,11 @@ int ARGBPolynomial(const uint8* src_argb, int src_stride_argb,
   if (TestCpuFlag(kCpuHasSSE2)) {
     ARGBPolynomialRow = ARGBPolynomialRow_SSE2;
   }
+#endif
+#if defined(HAS_ARGBPOLYNOMIALROW_AVX2)
+  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 2)) {
+    ARGBPolynomialRow = ARGBPolynomialRow_AVX2;
+  }
 #endif
   for (int y = 0; y < height; ++y) {
     ARGBPolynomialRow(src_argb, dst_argb, poly, width);

source/row_common.cc

@@ -2042,10 +2042,10 @@ void ARGBPolynomialRow_C(const uint8* src_argb,
     dr += poly[14] * r3;
     da += poly[15] * a3;
 
-    dst_argb[0]= Clamp(static_cast<int32>(db));
+    dst_argb[0] = Clamp(static_cast<int32>(db));
-    dst_argb[1]= Clamp(static_cast<int32>(dg));
+    dst_argb[1] = Clamp(static_cast<int32>(dg));
-    dst_argb[2]= Clamp(static_cast<int32>(dr));
+    dst_argb[2] = Clamp(static_cast<int32>(dr));
-    dst_argb[3]= Clamp(static_cast<int32>(da));
+    dst_argb[3] = Clamp(static_cast<int32>(da));
     src_argb += 4;
     dst_argb += 4;
   }

source/row_win.cc

@@ -6783,7 +6783,7 @@ void ARGBPolynomialRow_SSE2(const uint8* src_argb,
     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
+    pxor       xmm3, xmm3  // 0 constant for zero extending bytes to ints.
 
     align      16
  convertloop:
@@ -6814,6 +6814,66 @@ void ARGBPolynomialRow_SSE2(const uint8* src_argb,
 }
 #endif  // HAS_ARGBPOLYNOMIALROW_SSE2
 
+#ifdef HAS_ARGBPOLYNOMIALROW_AVX2
+__declspec(naked) __declspec(align(16))
+void ARGBPolynomialRow_AVX2(const uint8* src_argb,
+                           uint8* dst_argb, const float* poly,
+                           int width) {
+  __asm {
+    mov        eax, [esp + 12]   /* poly */
+    vmovdqu    xmm4, [eax]
+    vmovdqu    xmm5, [eax + 16]
+    vmovdqu    xmm6, [eax + 32]
+    vmovdqu    xmm7, [eax + 48]
+    vpermq     ymm4, ymm4, 0x44  // dup low qwords to high qwords
+    vpermq     ymm5, ymm5, 0x44
+    vpermq     ymm6, ymm6, 0x44
+    vpermq     ymm7, ymm7, 0x44
+
+    mov        eax, [esp + 4]   /* src_argb */
+    mov        edx, [esp + 8]   /* dst_argb */
+    mov        ecx, [esp + 16]  /* width */
+    vpxor      ymm3, ymm3, ymm3  // 0 constant for zero extending bytes to ints.
+
+    align      16
+ convertloop:
+    vmovq      xmm0, qword ptr [eax]  // 2 BGRA pixels
+    lea        eax, [eax + 8]
+
+//    vpmovzxbd  ymm0, ymm0
+// TODO(fbarchard): Consider vex256 to avoid vpermq.
+    vpunpcklbw xmm0, xmm0, xmm3  // b0g0r0a0_b0g0r0a0_00000000_00000000
+    vpermq     ymm0, ymm0, 0xd8  // b0g0r0a0_00000000_b0g0r0a0_00000000
+    vpunpcklwd ymm0, ymm0, ymm3  // b000g000_r000a000_b000g000_r000a000
+
+    vcvtdq2ps  ymm0, ymm0  // 8 floats
+    vmovdqa    ymm1, ymm0  // X
+    vmulps     ymm0, ymm0, ymm5  // C1 * X
+    vaddps     ymm0, ymm0, ymm4  // result = C0 + C1 * X
+    vmovdqa    ymm2, ymm1
+    vmulps     ymm2, ymm2, ymm1  // X * X
+    vmulps     ymm1, ymm1, ymm2  // X * X * X
+    vmulps     ymm2, ymm2, ymm6  // C2 * X * X
+    vmulps     ymm1, ymm1, ymm7  // C3 * X * X * X
+    vaddps     ymm0, ymm0, ymm2  // result += C2 * X * X
+    vaddps     ymm0, ymm0, ymm1  // result += C3 * X * X * X
+    vcvttps2dq ymm0, ymm0
+
+//    vpmovzxdb  ymm0, ymm0      // b000g000_r000a000_b000g000_r000a000
+    vpackusdw  ymm0, ymm0, ymm3  // b0g0r0a0_00000000_b0g0r0a0_00000000
+    vpermq     ymm0, ymm0, 0xd8  // b0g0r0a0_b0g0r0a0_00000000_00000000
+    vpackuswb  xmm0, xmm0, xmm3  // b0g0r0a0_b0g0r0a0_00000000_00000000
+
+    sub        ecx, 2
+    vmovq      qword ptr [edx], xmm0
+    lea        edx, [edx + 8]
+    jg         convertloop
+    vzeroupper
+    ret
+  }
+}
+#endif  // HAS_ARGBPOLYNOMIALROW_AVX2
+
 #endif  // !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && defined(_MSC_VER)
 
 #ifdef __cplusplus

unit_test/planar_test.cc

@@ -1665,7 +1665,7 @@ TEST_F(libyuvTest, TestARGBPolynomial) {
     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;

util/psnr.h

@@ -10,7 +10,7 @@
 
 // Get PSNR for video sequence. Assuming RAW 4:2:0 Y:Cb:Cr format
 
-#ifndef UTIL_PSNR_H_
+#ifndef UTIL_PSNR_H_  // NOLINT
 #define UTIL_PSNR_H_
 
 #ifdef __cplusplus
@@ -36,4 +36,4 @@ double ComputeSumSquareError(const uint8* org, const uint8* rec, int size);
 }  // extern "C"
 #endif
 
-#endif  // UTIL_PSNR_H_
+#endif  // UTIL_PSNR_H_  // NOLINT

util/ssim.cc

@@ -8,7 +8,7 @@
  *  be found in the AUTHORS file in the root of the source tree.
  */
 
-#include "./ssim.h"
+#include "../util/ssim.h"  // NOLINT
 
 #include <math.h>
 #include <string.h>

util/ssim.h

@@ -10,7 +10,7 @@
 
 // Get SSIM for video sequence. Assuming RAW 4:2:0 Y:Cb:Cr format
 
-#ifndef UTIL_SSIM_H_
+#ifndef UTIL_SSIM_H_  // NOLINT
 #define UTIL_SSIM_H_
 
 #ifdef __cplusplus
@@ -32,4 +32,4 @@ double CalcLSSIM(double ssim);
 }  // extern "C"
 #endif
 
-#endif  // UTIL_SSIM_H_
+#endif  // UTIL_SSIM_H_  // NOLINT