Use full screen triangles instead of quads for blits, clears etc.

[Novum]

I noticed that all full screen operations are done using two triangles. This is inefficient because of cache locality and empty helper threads on the triangle boundaries

Here are a couple references discussing this, the effect is known and measurable:
https://wallisc.github.io/rendering/2021/04/18/Fullscreen-Pass.html
https://michaldrobot.com/2014/04/01/gcn-execution-patterns-in-full-screen-passes/
https://stackoverflow.com/questions/55197347/webgl-full-screen-quad-or-triangle-for-invoking-fragment-shader-for-every-pixel/64686565#64686565

It's a minor saving, but it's actually simpler than drawing two triangles.

[bkaradzic]

Can you point to examples you're talking about?!

Most of examples use this code which uses single triangle:

bgfx/examples/09-hdr/hdr.cpp

Lines 40 to 96 in 4581f14

	void screenSpaceQuad(float _textureWidth, float _textureHeight, bool _originBottomLeft = false, float _width = 1.0f, float _height = 1.0f)
	{
	if (3 == bgfx::getAvailTransientVertexBuffer(3, PosColorTexCoord0Vertex::ms_layout) )
	{
	bgfx::TransientVertexBuffer vb;
	bgfx::allocTransientVertexBuffer(&vb, 3, PosColorTexCoord0Vertex::ms_layout);
	PosColorTexCoord0Vertex* vertex = (PosColorTexCoord0Vertex*)vb.data;
	const float zz = 0.0f;
	const float minx = -_width;
	const float maxx = _width;
	const float miny = 0.0f;
	const float maxy = _height*2.0f;
	const float texelHalfW = s_texelHalf/_textureWidth;
	const float texelHalfH = s_texelHalf/_textureHeight;
	const float minu = -1.0f + texelHalfW;
	const float maxu = 1.0f + texelHalfW;
	float minv = texelHalfH;
	float maxv = 2.0f + texelHalfH;
	if (_originBottomLeft)
	{
	float temp = minv;
	minv = maxv;
	maxv = temp;
	minv -= 1.0f;
	maxv -= 1.0f;
	}
	vertex[0].m_x = minx;
	vertex[0].m_y = miny;
	vertex[0].m_z = zz;
	vertex[0].m_rgba = 0xffffffff;
	vertex[0].m_u = minu;
	vertex[0].m_v = minv;
	vertex[1].m_x = maxx;
	vertex[1].m_y = miny;
	vertex[1].m_z = zz;
	vertex[1].m_rgba = 0xffffffff;
	vertex[1].m_u = maxu;
	vertex[1].m_v = minv;
	vertex[2].m_x = maxx;
	vertex[2].m_y = maxy;
	vertex[2].m_z = zz;
	vertex[2].m_rgba = 0xffffffff;
	vertex[2].m_u = maxu;
	vertex[2].m_v = maxv;
	bgfx::setVertexBuffer(0, &vb);
	}
	}

[Novum]

I saw a clear use two triangles in RenderDoc with DX11. I appologize that I extrapolated this to other cases.

[bkaradzic]

Ah yeah... In case of clear the issue is not that big. The main problem is when full screen pass shader does more than setting single fragment color. I don't remember exactly why I did it this way originally, but in examples where full screen shaders are more heavy it's using screenSpaceQuad that does single triangle method.

[Novum]

Tangential question about that, why is it using triangles to clear? At least on AMD HW fast clear only works if the proper clear method of the API is called.

[bkaradzic]

Quad is used only in cases when there is partial clear of frame buffer. ClearRenderTargetView and similar don't respect view port settings. If view rect matches whole frame buffer it will go thru ClearRenderTargetView code, and if it's partial clear it will go thru clear quad code.

[Novum]

Hm, strange, I'm pretty sure I was clearing the entire frame buffer. I'll check what's going on at some point and let you know if there is a potential problem. Thank you.