why skia on windows has bad efficiency - gdi

Question

I compared skia with gdi painting on windows. both drawing 98000 random lines. to my suprise that skia is far low efficiency than gdi(the skia painting cost 1600ms, while gdi cost 0ms). my testing code was paste below. any suggestions?
bool PaintCompare() {
//generate ramdon points
std::default_random_engine e(std::chrono::high_resolution_clock::now().time_since_epoch().count());
std::uniform_real_distribution<float> u(10, 500);
SkPoint pts[100];
for (int i = 0; i<100; i++)
pts[i].set(u(e), u(e));
SkPaint paint;
paint.setColor(SkColorSetRGB(255, 0, 0));
//create skia canvas
sk_sp<SkSurface> rasterSurface(
SkSurface::MakeRasterN32Premul(600, 600));
SkCanvas* canvas = rasterSurface->getCanvas();
//draw lines with skia
auto start = std::chrono::high_resolution_clock::now();
for (int i = 0; i<1000; i++)
{
for (int j = 1; j<99; j++)
{
canvas->drawLine(pts[j].fX, pts[j].fY, pts[j + 1].fX, pts[j + 1].fY, paint);
}
}
auto cost = std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::high_resolution_clock::now() - start);
sk_sp<SkImage> img(rasterSurface->makeImageSnapshot());
if (!img) { return false; }
SkBitmap skBmp;
if (!img->asLegacyBitmap(&skBmp, SkImage::kRO_LegacyBitmapMode)) {
return false;
}
//show bitmap on hdc
BITMAPINFO bmi;
memset(&bmi, 0, sizeof(bmi));
bmi.bmiHeader.biSize = sizeof(BITMAPINFOHEADER);
bmi.bmiHeader.biWidth = 600;
bmi.bmiHeader.biHeight = -600; // top-down image
bmi.bmiHeader.biPlanes = 1;
bmi.bmiHeader.biBitCount = 32;
bmi.bmiHeader.biCompression = BI_RGB;
bmi.bmiHeader.biSizeImage = 0;
HDC hdc = GetDC();
LPVOID pBits = NULL;
HBITMAP hBmp = CreateDIBSection(hdc, &bmi, DIB_RGB_COLORS, &pBits, 0, 0);
skBmp.copyPixelsTo(pBits, skBmp.getSize());
CDC memdc;
memdc.CreateCompatibleDC(hdc);
memdc.SelectBitmap(hBmp);
BitBlt(hdc, 0, 0, 600, 600, memdc, 0, 0, SRCCOPY);
memdc.DeleteDC();
//draw with gdi
CPen pen;
pen.CreatePen(PS_SOLID, 1, RGB(0, 255, 0));
RECT rc{ 0,0,600,600 };
CBitmap bmp;
bmp.CreateCompatibleBitmap(hdc, 600, 600);
memdc.CreateCompatibleDC(hdc);
memdc.SelectBitmap(bmp);
memdc.FillSolidRect(&rc, RGB(0, 0, 0));
memdc.SelectPen(pen);
start = std::chrono::high_resolution_clock::now();
for (int i = 0; i<1000; i++)
{
for (int j = 1; j<99; j++)
{
memdc.MoveTo(pts[j].fX, pts[j].fY);
memdc.LineTo(pts[j + 1].fX, pts[j + 1].fY);
}
}
auto cost2 = std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::high_resolution_clock::now() - start);
//copy bitmap to window
BitBlt(hdc, 700, 0, 600, 600, memdc, 0, 0, SRCCOPY);
ReleaseDC(hdc);
memdc.DeleteDC();
//wchar_t buf[256];
//wsprintf(buf, L"left cost=%I64d, right cost=%I64d", cost.count(), cost2.count());
//GetParent().SetWindowText(buf);
//cost == 1596615 microseconds
//cost2 == 107253 microseconds
}

Answer 1

I found the problem finally. i give the result in debug mode!
in debug mode that skia with raster backend is 20 times slower than gdi.
however in release mode skia with raster backend is 4-5 times slower than gdi.
i had another test that skia uses opengl as backend. the result shows skia and gdi spend almost the same time. skia is about 15% slower than gdi.