LuaVBO updating cost tester

dbg_luavbo_drawing.lua

function widget:GetInfo()
	return {
		name    = "LuaVBO Drawing Tester",
		desc    = "Simplest proof-of-concept VBO Drawer",
		author  = "Beherith GL4",
		date    = "2024.06.15",
		license  = "GNU GPL, v2 or later",
		layer   = 2,
		enabled = true,
	}
end
-- =================================================================================
-- The goal of this widget is to test the performance cost of 6 things
-- [ ] A. Cost of updating (or adding) a single instance to a VBO
-- [ ] B. Cost of updating (or adding) a single Unit to an instanceVBO
	-- 17 us
-- [ ] C. Cost of uploading 1-10 elements to a VBO
-- [ ] D. Cost of uploadint 1-10 units to an instanceVBO
-- [ ] E. Cost of doing ALL uploads in one pass for single units
-- [ ] F. Cost of doing ALL uploads in one pass for a unitVBO
--
--
-- General notes:
--  - A single call costs about 13 us
--	- Adding a unitID costs 33% more (17us), and this is true across the board, even for batched uploads
--	- Each successive hot cached call costs only 40% of the first call (~5 us)
--  - The most expensive part of single calls is the upload to VBO part, account for 70% of the total time
--  - Batched uploading costs about 20% of the time of individual uploads (!)
--		- E.g if we have more than 100 items, and modify 20% of them , then batched uploads are faster than individual
--
--
--
--
--
-- =================================================================================

local testVBO = nil
local testInstanceVBO = nil
local testShader = nil

local luaShaderDir = "LuaUI/Widgets/Include/"
local LuaShader = VFS.Include(luaShaderDir.."LuaShader.lua")
VFS.Include(luaShaderDir.."instancevbotable.lua")
local rand = math.random

local isUnitVBO = true	
local totalcnt = 100

local allunits = {}

local instanceCache = {}
for i=1,20 do instanceCache[i] = 0 end
if isUnitVBO then instanceCache[5] = 1 end 
instanceCache[10] = 1.0

local vsSrc = [[
#version 420
#extension GL_ARB_uniform_buffer_object : require
#extension GL_ARB_shader_storage_buffer_object : require
#extension GL_ARB_shading_language_420pack: require
layout (location = 0) in vec3 position;
layout (location = 1) in vec3 normals;
layout (location = 2) in vec2 uvs;

layout (location = 3) in vec4 worldposrad;
layout (location = 4) in vec4 params; // Contains isworld shit
layout (location = 5) in vec4 color1;
layout (location = 6) in vec4 color2;
layout (location = 7) in uvec4 instData; // unitID, teamID

//__ENGINEUNIFORMBUFFERDEFS__

struct SUniformsBuffer {
    uint composite; //     u8 drawFlag; u8 unused1; u16 id;

    uint unused2;
    uint unused3;
    uint unused4;

    float maxHealth;
    float health;
    float unused5;
    float unused6;

    vec4 drawPos;
    vec4 speed;
    vec4[4] userDefined; //can't use float[16] because float in arrays occupies 4 * float space
};

layout(std140, binding=1) readonly buffer UniformsBuffer {
    SUniformsBuffer uni[];
};

out DataVS {
	vec4 vs_color;
};
void main() {
	vec4 vertexPosition = vec4(worldposrad.xyz + position.xyz * worldposrad.w, 1.0);
	if (params.x > 0.5) {
		vertexPosition.xyz += uni[instData.y].drawPos.xyz;
	}
	vs_color = color1;
	gl_Position = cameraViewProj * vec4(vertexPosition.xyz, 1.0);
}]]


local fsSrc =  [[
#version 330
#extension GL_ARB_uniform_buffer_object : require
#extension GL_ARB_shading_language_420pack: require

uniform sampler2D myTexture;

//__ENGINEUNIFORMBUFFERDEFS__

in DataVS {
	vec4 vs_color;
};

out vec4 fragColor;

void main() {
	fragColor.rgba = vec4(vs_color.rgb, 1.0);
}]]

local function AddTestElement(instanceID, unitID, px,py,pz,r, noupload)
	instanceCache[1] = px
	instanceCache[2] = py
	instanceCache[3] = pz
	instanceCache[4] = r
	pushElementInstance(testInstanceVBO, instanceCache, instanceID, true, noupload, unitID)
end

function widget:Initialize()
	local engineUniformBufferDefs = LuaShader.GetEngineUniformBufferDefs()
	testShader = LuaShader( {
		vertex = vsSrc:gsub("//__ENGINEUNIFORMBUFFERDEFS__", engineUniformBufferDefs),
		fragment = fsSrc:gsub("//__ENGINEUNIFORMBUFFERDEFS__", engineUniformBufferDefs),
		uniformInt = { myTexture = 0,},
		uniformFloat = { },
		},
		"LuaVBO Drawing Tester GL4"
	)
	testShader:Initialize()
	local sphereVBO, numVerts, sphereIndexVBO, numIndices = makeSphereVBO(16,8,1)
	local testVBOLayout = {
		  {id = 3, name = 'worldposrad', size = 4}, -- widthlength
		  {id = 4, name = 'params', size = 4}, --  emit dir
		  {id = 5, name = 'color1', size = 4}, --- color
		  {id = 6, name = 'color2', size = 4}, --- color
		  {id = 7, name = 'instData', type = GL.UNSIGNED_INT, size= 4},
		}
	if isUnitVBO then 
		testInstanceVBO = makeInstanceVBOTable(testVBOLayout,256, "testVBO", 7)
	else
		testInstanceVBO = makeInstanceVBOTable(testVBOLayout,256, "testVBO", nil)
	end
	if not testInstanceVBO then Spring.Echo("failed to create testInstanceVBO") end 
	testInstanceVBO.numVertices = numIndices
	testInstanceVBO.vertexVBO = sphereVBO
	testInstanceVBO.indexVBO = sphereIndexVBO  
	testInstanceVBO.VAO = makeVAOandAttach(testInstanceVBO.vertexVBO, testInstanceVBO.instanceVBO)
	testInstanceVBO.VAO:AttachIndexBuffer(testInstanceVBO.indexVBO)
	allunits = Spring.GetAllUnits()

	
	for i= 1,totalcnt do
		if isUnitVBO then 
			local testUnitID = allunits[math.random(1,#allunits)]
			AddTestElement(i, testUnitID, 1000 * rand(), 1000* rand(), 1000*rand(), 10*rand())
		else
			AddTestElement(i, nil,        1000 * rand(), 1000* rand(), 1000*rand(), 10*rand())
		end
	end
end

local drawFrame
function widget:Update()
	-- [ ] A. Cost of updating (or adding) a single instance to a VBO
-- [ ] B. Cost of updating (or adding) a single Unit to an instanceVBO
-- [ ] C. Cost of uploading 1-10 elements to a VBO
-- [ ] D. Cost of uploadint 1-10 units to an instanceVBO
-- [ ] E. Cost of doing ALL uploads in one pass for single units
-- [ ] F. Cost of doing ALL uploads in one pass for a unitVBO


	drawFrame = Spring.GetDrawFrame()
	if isUnitVBO or true then
		if drawFrame%4 ==0 then
			-- [ ] B. Cost of updating (or adding) a single Unit to an instanceVBO
			-- 	Costs on average 16us per call
			local testUnitID = allunits[math.random(1,#allunits)]
			tracy.ZoneBeginN("testVBO:".. (isUnitVBO and "B" or "A"))
			AddTestElement(math.random(1,totalcnt-1) ,    isUnitVBO and testUnitID  or nil,   1000 * rand(), 1000* rand(), 1000*rand(), 10*rand())
			tracy.ZoneEnd()
		
		elseif drawFrame%4 ==1 then
			-- [ ] D. Cost of uploading 1-10 units to an instanceVBO
			-- Costs on average 16 us for first, + 6 us per additional unit
			local count = math.floor(drawFrame/4) % 20
			local testUnitID = allunits[math.random(1,#allunits)]
			tracy.ZoneBeginN(string.format("testVBO:%s:%03d",isUnitVBO and "D" or "C" , count))
			for j =1, count do 
				AddTestElement(math.random(1,totalcnt-1) ,     isUnitVBO and testUnitID  or nil,   1000 * rand(), 1000* rand(), 1000*rand(), 10*rand())
			end
			tracy.ZoneEnd()

		elseif drawFrame %4 == 2 then
			-- [ ] F. Cost of uploading 1-10 DIFFERENT units to an instanceVBO
			-- Costs on average 16 us for first, + 6 us per additional unit
			local count = math.floor(drawFrame/4) % 20
			tracy.ZoneBeginN(string.format("testVBO:%s:%03d",isUnitVBO and "F" or "E" , count))
			for j =1, count do 
				local testUnitID = allunits[math.random(1,#allunits)]
				AddTestElement(math.random(1,totalcnt-1) ,     isUnitVBO and testUnitID  or nil,   1000 * rand(), 1000* rand(), 1000*rand(), 10*rand())
			end
			tracy.ZoneEnd()
		else
			-- [ ] H. Cost of NOT uploading 1-10 DIFFERENT units to an instanceVBO, then uploading all!
			-- Costs on average 16 us for first, + 6 us per additional unit
			local count = math.floor(drawFrame/4) % 20
			tracy.ZoneBeginN(string.format("testVBO:%s:%03d",isUnitVBO and "H" or "G" , count))
			for j =1, count do 
				local testUnitID = allunits[math.random(1,#allunits)]
				AddTestElement(math.random(1,totalcnt-1) ,     isUnitVBO and testUnitID  or nil,   1000 * rand(), 1000* rand(), 1000*rand(), 10*rand(), true) -- note NOUPLOAD
			end
			tracy.ZoneEnd()
			
			tracy.ZoneBeginN(string.format("testVBO:%s:%04d", isUnitVBO and "H" or "G", totalcnt))
				uploadAllElements(testInstanceVBO)
			tracy.ZoneEnd()
		end
	else
		
	end
end


function widget:DrawWorld()
	if testInstanceVBO.usedElements > 0 then 
		gl.Culling(false)
		testShader:Activate()
		drawInstanceVBO(testInstanceVBO)
		testShader:Deactivate()
	end
end

function widget:Shutdown()
end