Faster blit_with_alpha()

[HybridDog]

The blit_with_alpha function has a noticeable effect on the time it takes to join a game.

To reduce the join times, I replace the blit_with_alpha function with a new one:

• It does not uses floating-point numbers.
• It directly operates on the raw pixel data instead of using the comparatively
  slow setPixel and getPixel functions from Irrlicht.
  Only ECF_A8R8G8B8 base images are supported now.
  If the top image does not have the ECF_A8R8G8B8 colour format, it is converted;
  I assume that this happens rarely.
• There are case distinctions for fully opaque, fully transparent and semi-transparent pixels.
  This empirically increases the performance since the mixing between two semi-transparent happens rarely.
• The new function no longer has the src_pos argument since it was always the zero vector.
• The function is only documented once where it is declared.

For backwards compatibility, blit_with_alpha still mixes colours without gamma correction.
blit_with_alpha nonetheless behaves slightly different than before:
If a semi-transparent pixel is drawn on top of another semi-transparent pixel,
the colour is mixed in a way which we can consider to be more correct now.

Related issue: #14322
Roadmap goal: 2.1 Rendering/Graphics improvements (perhaps)

To do

This PR is a Ready for Review.

How to test

• Add this Lua code to a mod:

local kick_on_join = true
if kick_on_join then
	minetest.register_on_joinplayer(function(player)
		minetest.kick_player(player:get_player_name(),
			"This is for testing purposes. Don't cry.")
	end)
end

• Start a local Minetest server
• In a Shell which supports time, join the server with the following command:

time bin/minetest --name your_name --address 127.0.0.1 --go

• Repeat the above command multiple times with and without the proposed changes and note down the duration values.

With the mods I commonly use, I could observe a speedup from 2.5 s to 2.2 s real time and a speedup from 1.3 s to 1.0 s user CPU time.
The speedup may be different with large textures and/or a different selection of mods.
Joining the game may not be the only situation which is faster with the new blit_with_alpha function.

[HybridDog]

I have tested the performance with different implementation decisions now by collecting the blit_with_alpha inputs and measuring the time it takes to re-execute all the blit_with_alphas.

I tried it 9 times and there were 1687 blit_with_alpha inputs each time.
for f in {1..9}; do bin/minetest --name your_nameaaa --address 127.0.0.1 --go --quiet; done

With the branches the median duration was ca. 0.037 s.
All durations: 0.0353195, 0.0358201, 0.0363832, 0.0364065, 0.0372696, 0.0373504, 0.0381849, 0.0395097, 0.040282
Code: https://github.com/HybridDog/minetest/tree/8cbd75b076bb5d91caed2aaf90121b7753675d13

Without the branches the median was ca. 0.082 s.
All durations: 0.0795838, 0.0796296, 0.0812277, 0.08147, 0.0819993, 0.0820157, 0.082015, 0.0822578, 0.089137
Code: https://github.com/HybridDog/minetest/tree/efbce26

With the branches, a small reordering, and using SColor, the median was ca. 0.026 s.
The small reordering is the movement of the branch src_a == 0 before the branch src_a == 255 || dst_a == 0; without this the median was ca. 0.035 s.
I also tried u32 integers instead of static_casts to u8 and it was noticeable faster with u8 integers.
The code from this case is used in this Pull Request.
All durations: 0.0247669, 0.0249853, 0.0254656, 0.0255048, 0.0257042, 0.026188, 0.026427, 0.0265187, 0.0272248
Code: https://github.com/HybridDog/minetest/tree/96505ee7f

With blit_with_alpha from the master branch, the median was ca. 0.384 s.
All durations: 0.361643, 0.368806, 0.373106, 0.380362, 0.384423, 0.386407, 0.393304, 0.394426, 0.401159
Code: https://github.com/HybridDog/minetest/tree/6c7009384

[Desour]

blit_with_alpha nonetheless behaves slightly different than before:
If a semi-transparent pixel is drawn on top of another semi-transparent pixel,
the colour is mixed in a way which we can consider to be more correct now.

I also noticed how texture overlay is not alpha blending if both are semitransparent.
(This also makes ^ non-associative.)

I've made test mod for this:

code

local function trigger_hud(player)
	local row_height = 0.05
	local base_pos = {x=0.2,y=0.2}
	local ncolumns = 4

	local tex_checker = "[png:"..minetest.encode_base64(minetest.encode_png(ncolumns, 1, {
			{a=255, r=200, g=200, b=200},
			{a=255, r=55, g=55, b=55},
			{a=255, r=200, g=200, b=200},
			{a=255, r=55, g=55, b=55},
		}))

	local tex_a = "[png:"..minetest.encode_base64(minetest.encode_png(ncolumns, 1, {
			{a=255, r=255, g=255, b=255},
			{a=250, r=255, g=255, b=255},
			{a=1, r=0, g=255, b=0},
			{a=0, r=0, g=255, b=0},
		}))

	local tex_b = "[png:"..minetest.encode_base64(minetest.encode_png(ncolumns, 1, {
			{a=8, r=255, g=0, b=0},
			{a=8, r=255, g=0, b=0},
			{a=127, r=255, g=0, b=0},
			{a=127, r=255, g=0, b=0},
		}))

	-- name, texturestring
	local rows = {
		{"test_tmp:checker", tex_checker},
		{"test_tmp:a", tex_a},
		{"test_tmp:b", tex_b},
		{"test_tmp:a_b", tex_a.."^"..tex_b},
		{"test_tmp:a_b_hud", tex_a},
	}

	local function draw_row_at(i, at, z_offset)
		player:hud_add({
			type = "image",
			position = {x=base_pos.x, y=base_pos.y+at*row_height},
			name = rows[i][1],
			scale = {x=-100*row_height*ncolumns, y=-100*row_height},
			text = rows[i][2],
			alignment = {x=1, y=1},
			offset = {x=0, y=0},
			z_index = 1000 + (z_offset or 0),
		})
	end

	for row = 1, #rows do
		draw_row_at(row, row)
	end

	draw_row_at(3, 5, 1)
end

minetest.register_node("test_tmp:node1", {
	description = "Node1",
	tiles = {"default_pine_tree_top.png"},
	groups = { dig_immediate = 2 },
	on_punch = function(_pos, _node, puncher, _pointed_thing)
		if minetest.is_player(puncher) then
			trigger_hud(puncher)
		end
	end,
})

But the alpha test node you added also shows this. In master, there are some texels that are different.

I agree that we can consider this change in behaviour a bugfix though.

[Desour]

Code looks fine otherwise.

[Desour]

Tested, works. 👍

Joined your land server. It was about 1/3 (5 s) faster. And blit_with_alpha is in the flamegraph basically completely gone.

Thanks!

[SmallJoker]

Visual comparison: (1 color difference)

This is what I get when mixing the two textures in GIMP (100% layer opacity). The resulting pixel is of alpha=135.

.. that's also roughly what I get in Minetest:

The PR looks more correct, thus I approve.