Commit 8be4b289 authored by fbarchard@google.com's avatar fbarchard@google.com

ARGBSobelToPlane which produces a planar output.

BUG=none
TEST=none
R=ryanpetrie@google.com

Review URL: https://webrtc-codereview.appspot.com/2415005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@818 16f28f9a-4ce2-e073-06de-1de4eb20be90
parent e35422d9
Name: libyuv
URL: http://code.google.com/p/libyuv/
Version: 817
Version: 818
License: BSD
License File: LICENSE
......
......@@ -402,6 +402,12 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
uint8* dst_argb, int dst_stride_argb,
const uint8* shuffler, int width, int height);
// Sobel ARGB effect with planar output.
LIBYUV_API
int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
uint8* dst_y, int dst_stride_y,
int width, int height);
// Sobel ARGB effect.
LIBYUV_API
int ARGBSobel(const uint8* src_argb, int src_stride_argb,
......
......@@ -59,6 +59,7 @@ extern "C" {
#define HAS_INTERPOLATEROW_SSE2
#define HAS_INTERPOLATEROW_SSSE3
#define HAS_SOBELROW_SSE2
#define HAS_SOBELTOPLANEROW_SSE2
#define HAS_SOBELXROW_SSSE3
#define HAS_SOBELXYROW_SSE2
#define HAS_SOBELYROW_SSSE3
......@@ -313,6 +314,7 @@ extern "C" {
#define HAS_ARGBSHADEROW_NEON
#define HAS_ARGBSUBTRACTROW_NEON
#define HAS_SOBELROW_NEON
#define HAS_SOBELTOPLANEROW_NEON
#define HAS_SOBELXYROW_NEON
#define HAS_SOBELXROW_NEON
#define HAS_SOBELYROW_NEON
......@@ -1578,6 +1580,12 @@ void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_argb, int width);
void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_argb, int width);
void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_y, int width);
void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_y, int width);
void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_y, int width);
void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_argb, int width);
void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
......
......@@ -11,6 +11,6 @@
#ifndef INCLUDE_LIBYUV_VERSION_H_ // NOLINT
#define INCLUDE_LIBYUV_VERSION_H_
#define LIBYUV_VERSION 817
#define LIBYUV_VERSION 818
#endif // INCLUDE_LIBYUV_VERSION_H_ NOLINT
......@@ -1861,10 +1861,12 @@ int ARGBShuffle(const uint8* src_bgra, int src_stride_bgra,
}
// Sobel ARGB effect.
LIBYUV_API
int ARGBSobel(const uint8* src_argb, int src_stride_argb,
static int ARGBSobelize(const uint8* src_argb, int src_stride_argb,
uint8* dst_argb, int dst_stride_argb,
int width, int height) {
int width, int height,
void (*SobelRow)(const uint8* src_sobelx,
const uint8* src_sobely,
uint8* dst, int width)) {
const int kMaxRow = kMaxStride / 4;
const int kEdge = 16; // Extra pixels at start of row for extrude/align.
if (!src_argb || !dst_argb ||
......@@ -1921,20 +1923,6 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
SobelXRow = SobelXRow_NEON;
}
#endif
void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_argb, int width) = SobelRow_C;
#if defined(HAS_SOBELROW_SSE2)
if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
SobelRow = SobelRow_SSE2;
}
#endif
#if defined(HAS_SOBELROW_NEON)
if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
SobelRow = SobelRow_NEON;
}
#endif
// 3 rows with edges before/after.
SIMD_ALIGNED(uint8 row_y[kEdge + kMaxRow * 3]);
SIMD_ALIGNED(uint8 row_sobelx[kMaxRow]);
......@@ -1976,68 +1964,56 @@ int ARGBSobel(const uint8* src_argb, int src_stride_argb,
return 0;
}
// SobelXY ARGB effect.
// Similar to Sobel, but also stores Sobel X in R and Sobel Y in B. G = Sobel.
// Sobel ARGB effect.
LIBYUV_API
int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
int ARGBSobel(const uint8* src_argb, int src_stride_argb,
uint8* dst_argb, int dst_stride_argb,
int width, int height) {
const int kMaxRow = kMaxStride / 4;
const int kEdge = 16; // Extra pixels at start of row for extrude/align.
if (!src_argb || !dst_argb ||
width <= 0 || height == 0 || width > (kMaxRow - kEdge)) {
return -1;
}
// Negative height means invert the image.
if (height < 0) {
height = -height;
src_argb = src_argb + (height - 1) * src_stride_argb;
src_stride_argb = -src_stride_argb;
}
// ARGBToBayer used to select G channel from ARGB.
void (*ARGBToBayerRow)(const uint8* src_argb, uint8* dst_bayer,
uint32 selector, int pix) = ARGBToBayerRow_C;
#if defined(HAS_ARGBTOBAYERROW_SSSE3)
if (TestCpuFlag(kCpuHasSSSE3) && width >= 8 &&
IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16)) {
ARGBToBayerRow = ARGBToBayerRow_Any_SSSE3;
if (IS_ALIGNED(width, 8)) {
ARGBToBayerRow = ARGBToBayerRow_SSSE3;
}
}
#elif defined(HAS_ARGBTOBAYERROW_NEON)
if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
ARGBToBayerRow = ARGBToBayerRow_Any_NEON;
if (IS_ALIGNED(width, 8)) {
ARGBToBayerRow = ARGBToBayerRow_NEON;
}
}
#endif
void (*SobelYRow)(const uint8* src_y0, const uint8* src_y1,
uint8* dst_sobely, int width) = SobelYRow_C;
#if defined(HAS_SOBELYROW_SSSE3)
if (TestCpuFlag(kCpuHasSSSE3)) {
SobelYRow = SobelYRow_SSSE3;
void (*SobelRow)(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_argb, int width) = SobelRow_C;
#if defined(HAS_SOBELROW_SSE2)
if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
SobelRow = SobelRow_SSE2;
}
#endif
#if defined(HAS_SOBELYROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
SobelYRow = SobelYRow_NEON;
#if defined(HAS_SOBELROW_NEON)
if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 8)) {
SobelRow = SobelRow_NEON;
}
#endif
void (*SobelXRow)(const uint8* src_y0, const uint8* src_y1,
const uint8* src_y2, uint8* dst_sobely, int width) =
SobelXRow_C;
#if defined(HAS_SOBELXROW_SSSE3)
if (TestCpuFlag(kCpuHasSSSE3)) {
SobelXRow = SobelXRow_SSSE3;
return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
width, height, SobelRow);
}
// Sobel ARGB effect with planar output.
LIBYUV_API
int ARGBSobelToPlane(const uint8* src_argb, int src_stride_argb,
uint8* dst_y, int dst_stride_y,
int width, int height) {
void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_, int width) = SobelToPlaneRow_C;
#if defined(HAS_SOBELTOPLANEROW_SSE2)
if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 16) &&
IS_ALIGNED(dst_y, 16) && IS_ALIGNED(dst_stride_y, 16)) {
SobelToPlaneRow = SobelToPlaneRow_SSE2;
}
#endif
#if defined(HAS_SOBELXROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
SobelXRow = SobelXRow_NEON;
#if defined(HAS_SOBELTOPLANEROW_NEON)
if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(width, 16)) {
SobelToPlaneRow = SobelToPlaneRow_NEON;
}
#endif
return ARGBSobelize(src_argb, src_stride_argb, dst_y, dst_stride_y,
width, height, SobelToPlaneRow);
}
// SobelXY ARGB effect.
// Similar to Sobel, but also stores Sobel X in R and Sobel Y in B. G = Sobel.
LIBYUV_API
int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
uint8* dst_argb, int dst_stride_argb,
int width, int height) {
void (*SobelXYRow)(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_argb, int width) = SobelXYRow_C;
#if defined(HAS_SOBELXYROW_SSE2)
......@@ -2051,45 +2027,8 @@ int ARGBSobelXY(const uint8* src_argb, int src_stride_argb,
SobelXYRow = SobelXYRow_NEON;
}
#endif
SIMD_ALIGNED(uint8 row_y[kEdge + kMaxRow * 3]);
SIMD_ALIGNED(uint8 row_sobelx[kMaxRow]);
SIMD_ALIGNED(uint8 row_sobely[kMaxRow]);
// Convert first row.
uint8* row_y0 = row_y + kEdge;
uint8* row_y1 = row_y0 + kMaxRow;
uint8* row_y2 = row_y1 + kMaxRow;
ARGBToBayerRow(src_argb, row_y0, 0x0d090501, width);
row_y0[-1] = row_y0[0];
memset(row_y0 + width, row_y0[width - 1], 16); // extrude 16 pixels.
ARGBToBayerRow(src_argb, row_y1, 0x0d090501, width);
row_y1[-1] = row_y1[0];
memset(row_y1 + width, row_y1[width - 1], 16);
memset(row_y2 + width, 0, 16);
for (int y = 0; y < height; ++y) {
// Convert next row of ARGB to Y.
if (y < (height - 1)) {
src_argb += src_stride_argb;
}
ARGBToBayerRow(src_argb, row_y2, 0x0d090501, width);
row_y2[-1] = row_y2[0];
row_y2[width] = row_y2[width - 1];
SobelXRow(row_y0 - 1, row_y1 - 1, row_y2 - 1, row_sobelx, width);
SobelYRow(row_y0 - 1, row_y2 - 1, row_sobely, width);
SobelXYRow(row_sobelx, row_sobely, dst_argb, width);
// Cycle thru circular queue of 3 row_y buffers.
uint8* row_yt = row_y0;
row_y0 = row_y1;
row_y1 = row_y2;
row_y2 = row_yt;
dst_argb += dst_stride_argb;
}
return 0;
return ARGBSobelize(src_argb, src_stride_argb, dst_argb, dst_stride_argb,
width, height, SobelXYRow);
}
// Apply a 4x4 polynomial to each ARGB pixel.
......
......@@ -869,6 +869,16 @@ void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
}
}
void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_y, int width) {
for (int i = 0; i < width; ++i) {
int r = src_sobelx[i];
int b = src_sobely[i];
int s = clamp255(r + b);
dst_y[i] = static_cast<uint8>(s);
}
}
void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_argb, int width) {
for (int i = 0; i < width; ++i) {
......
......@@ -2644,6 +2644,28 @@ void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
);
}
// Adds Sobel X and Sobel Y and stores Sobel into plane.
void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_y, int width) {
asm volatile (
// 16 pixel loop.
".p2align 2 \n"
"1: \n"
"vld1.8 {q0}, [%0]! \n" // load 16 sobelx.
"vld1.8 {q1}, [%1]! \n" // load 16 sobely.
"subs %3, %3, #16 \n" // 16 processed per loop.
"vqadd.u8 q0, q0, q1 \n" // add
"vst1.8 {q0}, [%2]! \n" // store 16 pixels.
"bgt 1b \n"
: "+r"(src_sobelx), // %0
"+r"(src_sobely), // %1
"+r"(dst_y), // %2
"+r"(width) // %3
:
: "cc", "memory", "q0", "q1"
);
}
// Mixes Sobel X, Sobel Y and Sobel into ARGB.
// A = 255
// R = Sobel X
......
......@@ -4767,6 +4767,43 @@ void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
}
#endif // HAS_SOBELROW_SSE2
#ifdef HAS_SOBELTOPLANEROW_SSE2
// Adds Sobel X and Sobel Y and stores Sobel into a plane.
void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_y, int width) {
asm volatile (
"sub %0,%1 \n"
"pcmpeqb %%xmm5,%%xmm5 \n"
"pslld $0x18,%%xmm5 \n"
// 8 pixel loop.
".p2align 4 \n"
BUNDLEALIGN
"1: \n"
"movdqa "MEMACCESS(0)",%%xmm0 \n"
MEMOPREG(movdqa,0x00,0,1,1,xmm1) // movdqa (%0,%1,1),%%xmm1
"lea "MEMLEA(0x10,0)",%0 \n"
"paddusb %%xmm1,%%xmm0 \n"
"sub $0x10,%3 \n"
"movdqa %%xmm0,"MEMACCESS(2)" \n"
"lea "MEMLEA(0x10,2)",%2 \n"
"jg 1b \n"
: "+r"(src_sobelx), // %0
"+r"(src_sobely), // %1
"+r"(dst_y), // %2
"+r"(width) // %3
:
: "memory", "cc"
#if defined(__native_client__) && defined(__x86_64__)
, "r14"
#endif
#if defined(__SSE2__)
, "xmm0", "xmm1"
#endif
);
}
#endif // HAS_SOBELTOPLANEROW_SSE2
#ifdef HAS_SOBELXYROW_SSE2
// Mixes Sobel X, Sobel Y and Sobel into ARGB.
// A = 255
......
......@@ -5737,6 +5737,36 @@ void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
}
#endif // HAS_SOBELROW_SSE2
#ifdef HAS_SOBELTOPLANEROW_SSE2
// Adds Sobel X and Sobel Y and stores Sobel into a plane.
__declspec(naked) __declspec(align(16))
void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_y, int width) {
__asm {
push esi
mov eax, [esp + 4 + 4] // src_sobelx
mov esi, [esp + 4 + 8] // src_sobely
mov edx, [esp + 4 + 12] // dst_argb
mov ecx, [esp + 4 + 16] // width
sub esi, eax
align 16
convertloop:
movdqa xmm0, [eax] // read 16 pixels src_sobelx
movdqa xmm1, [eax + esi] // read 16 pixels src_sobely
lea eax, [eax + 16]
paddusb xmm0, xmm1 // sobel = sobelx + sobely
sub ecx, 16
movdqa [edx], xmm0
lea edx, [edx + 16]
jg convertloop
pop esi
ret
}
}
#endif // HAS_SOBELTOPLANEROW_SSE2
#ifdef HAS_SOBELXYROW_SSE2
// Mixes Sobel X, Sobel Y and Sobel into ARGB.
// A = 255
......
......@@ -1136,7 +1136,7 @@ TEST_F(libyuvTest, TestSobelX) {
sobel_pixels_opt, 1280);
}
for (int i = 0; i < 1280; ++i) {
EXPECT_EQ(sobel_pixels_opt[i], sobel_pixels_c[i]);
EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]);
}
}
......@@ -1172,7 +1172,7 @@ TEST_F(libyuvTest, TestSobelY) {
SobelYRow(orig_pixels_0, orig_pixels_1, sobel_pixels_opt, 1280);
}
for (int i = 0; i < 1280; ++i) {
EXPECT_EQ(sobel_pixels_opt[i], sobel_pixels_c[i]);
EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]);
}
}
......@@ -1215,8 +1215,46 @@ TEST_F(libyuvTest, TestSobel) {
for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
SobelRow(orig_sobelx, orig_sobely, sobel_pixels_opt, 1280);
}
for (int i = 0; i < 16; ++i) {
EXPECT_EQ(sobel_pixels_opt[i], sobel_pixels_c[i]);
for (int i = 0; i < 1280 * 4; ++i) {
EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]);
}
}
TEST_F(libyuvTest, TestSobelToPlane) {
SIMD_ALIGNED(uint8 orig_sobelx[1280]);
SIMD_ALIGNED(uint8 orig_sobely[1280]);
SIMD_ALIGNED(uint8 sobel_pixels_c[1280]);
SIMD_ALIGNED(uint8 sobel_pixels_opt[1280]);
for (int i = 0; i < 1280; ++i) {
orig_sobelx[i] = i;
orig_sobely[i] = i * 2;
}
SobelToPlaneRow_C(orig_sobelx, orig_sobely, sobel_pixels_c, 1280);
EXPECT_EQ(0u, sobel_pixels_c[0]);
EXPECT_EQ(3u, sobel_pixels_c[1]);
EXPECT_EQ(6u, sobel_pixels_c[2]);
EXPECT_EQ(99u, sobel_pixels_c[33]);
EXPECT_EQ(255u, sobel_pixels_c[100]);
void (*SobelToPlaneRow)(const uint8* src_sobelx, const uint8* src_sobely,
uint8* dst_y, int width) = SobelToPlaneRow_C;
#if defined(HAS_SOBELTOPLANEROW_SSE2)
if (TestCpuFlag(kCpuHasSSE2)) {
SobelToPlaneRow = SobelToPlaneRow_SSE2;
}
#endif
#if defined(HAS_SOBELTOPLANEROW_NEON)
if (TestCpuFlag(kCpuHasNEON)) {
SobelToPlaneRow = SobelToPlaneRow_NEON;
}
#endif
for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
SobelToPlaneRow(orig_sobelx, orig_sobely, sobel_pixels_opt, 1280);
}
for (int i = 0; i < 1280; ++i) {
EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]);
}
}
......@@ -1255,8 +1293,8 @@ TEST_F(libyuvTest, TestSobelXY) {
for (int i = 0; i < benchmark_pixels_div1280_; ++i) {
SobelXYRow(orig_sobelx, orig_sobely, sobel_pixels_opt, 1280);
}
for (int i = 0; i < 16; ++i) {
EXPECT_EQ(sobel_pixels_opt[i], sobel_pixels_c[i]);
for (int i = 0; i < 1280 * 4; ++i) {
EXPECT_EQ(sobel_pixels_c[i], sobel_pixels_opt[i]);
}
}
......@@ -1604,6 +1642,75 @@ TEST_F(libyuvTest, ARGBSobel_Opt) {
EXPECT_EQ(0, max_diff);
}
static int TestSobelToPlane(int width, int height, int benchmark_iterations,
int invert, int off) {
if (width < 1) {
width = 1;
}
const int kSrcBpp = 4;
const int kDstBpp = 1;
const int kSrcStride = (width * kSrcBpp + 15) & ~15;
const int kDstStride = (width * kDstBpp + 15) & ~15;
align_buffer_64(src_argb_a, kSrcStride * height + off);
align_buffer_64(dst_argb_c, kDstStride * height);
align_buffer_64(dst_argb_opt, kDstStride * height);
memset(src_argb_a, 0, kSrcStride * height + off);
srandom(time(NULL));
for (int i = 0; i < kSrcStride * height; ++i) {
src_argb_a[i + off] = (random() & 0xff);
}
memset(dst_argb_c, 0, kDstStride * height);
memset(dst_argb_opt, 0, kDstStride * height);
MaskCpuFlags(0);
ARGBSobelToPlane(src_argb_a + off, kSrcStride,
dst_argb_c, kDstStride,
width, invert * height);
MaskCpuFlags(-1);
for (int i = 0; i < benchmark_iterations; ++i) {
ARGBSobelToPlane(src_argb_a + off, kSrcStride,
dst_argb_opt, kDstStride,
width, invert * height);
}
int max_diff = 0;
for (int i = 0; i < kDstStride * 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(dst_argb_c)
free_aligned_buffer_64(dst_argb_opt)
return max_diff;
}
TEST_F(libyuvTest, ARGBSobelToPlane_Any) {
int max_diff = TestSobelToPlane(benchmark_width_ - 1, benchmark_height_,
benchmark_iterations_, +1, 0);
EXPECT_EQ(0, max_diff);
}
TEST_F(libyuvTest, ARGBSobelToPlane_Unaligned) {
int max_diff = TestSobelToPlane(benchmark_width_, benchmark_height_,
benchmark_iterations_, +1, 1);
EXPECT_EQ(0, max_diff);
}
TEST_F(libyuvTest, ARGBSobelToPlane_Invert) {
int max_diff = TestSobelToPlane(benchmark_width_, benchmark_height_,
benchmark_iterations_, -1, 0);
EXPECT_EQ(0, max_diff);
}
TEST_F(libyuvTest, ARGBSobelToPlane_Opt) {
int max_diff = TestSobelToPlane(benchmark_width_, benchmark_height_,
benchmark_iterations_, +1, 0);
EXPECT_EQ(0, max_diff);
}
static int TestSobelXY(int width, int height, int benchmark_iterations,
int invert, int off) {
if (width < 1) {
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment