Skip to content

A fast 2D and 3D game engine — TypeScript SDK, C++/WebAssembly core, visual editor. Ship one project to web, desktop, WeChat MiniGames, playable ads, and native Android / iOS.

perf(render): Voxel GI reads its cache plainly where nothing else writes it, and skips empty slots when it ages them, so a phone spends far less on its upkeep and its rays

yhh0ef5a567f4ca
diff --git a/src/esengine/renderer/frame/VoxelGI.cpp b/src/esengine/renderer/frame/VoxelGI.cpp
index 2e8e37038..350ba9477 100644
--- a/src/esengine/renderer/frame/VoxelGI.cpp
+++ b/src/esengine/renderer/frame/VoxelGI.cpp
@@ -108,14 +108,15 @@ fn main(@builtin(global_invocation_id) id : vec3u) {
let cell = gi.origin.xyz + rel;
let tag = hash3(cell);
let base = i * WORDS;
+ // Each invocation owns its slot, so the words are read and written plainly (readWriteCache): ageing a million slots through atomics
+ // every frame was most of what the cache cost a phone. An empty slot (no samples, no area) has nothing to age and is left alone.
if ((tags[i] & ~SEEN) != tag) {
tags[i] = tag;
- for (var k = 0u; k < WORDS; k = k + 1u) { atomicStore(&grid[base + k], 0u); }
+ for (var k = 0u; k < WORDS; k = k + 1u) { grid[base + k] = 0u; }
} else {
tags[i] = tag;
- for (var k = 0u; k < WORDS; k = k + 1u) {
- atomicStore(&grid[base + k], u32(f32(atomicLoad(&grid[base + k])) * gi.a.y));
- }
+ if ((grid[base] | grid[base + 1u] | grid[base + 2u] | grid[base + 3u] | grid[base + 7u]) == 0u) { return; }
+ for (var k = 0u; k < WORDS; k = k + 1u) { grid[base + k] = u32(f32(grid[base + k]) * gi.a.y); }
}
}
)WGSL";
@@ -429,6 +430,25 @@ std::string join(const char* a, const char* b, const char* c = "") {
return std::string(a) + b + c;
}
+/// The gather only reads the cache, so it gets it as a plain read-only array. Every word read through atomicLoad cost a trip past the
+/// cache on a mobile GPU (Adreno under Vulkan): the gather of a 720p frame took over a second a frame on a phone, against a few
+/// milliseconds read plainly. The writers' atomics stay, and the gather's dispatch follows them in the graph, so it sees their sums.
+std::string plainCache(std::string s, const char* access) {
+ const auto replace = [&s](const std::string& from, const std::string& to) {
+ for (size_t at = s.find(from); at != std::string::npos; at = s.find(from, at + to.size())) s.replace(at, from.size(), to);
+ };
+ replace("var<storage, read_write> grid : array<atomic<u32>>;", std::string("var<storage, ") + access + "> grid : array<u32>;");
+ replace("var<storage, read_write> tags : array<u32>;", std::string("var<storage, ") + access + "> tags : array<u32>;");
+ // addArea adds atomically; neither kernel calls it, but a plain array cannot even name an atomicAdd.
+ const size_t from = s.find("fn addArea(");
+ if (from != std::string::npos) s.erase(from, s.find("\n}", from) + 2 - from);
+ replace("atomicLoad(&grid[", "(grid[");
+ return s;
+}
+std::string readOnlyCache(const std::string& s) { return plainCache(s, "read"); }
+/// For the maintain, whose every invocation owns one slot outright.
+std::string readWriteCache(const std::string& s) { return plainCache(s, "read_write"); }
+
struct ParamsBlock {
i32 origin[4];
u32 dims[4];
@@ -503,8 +523,8 @@ bool VoxelGI::ensure() {
}
return true;
};
- return make(maintain_, join(kCommon, kMaintain), "maintain") && make(inject_, join(kCommon, kSurface, kInject), "inject")
- && make(gather_, join(kCommon, kSurface, kGather), "gather") && make(apply_, join(kCommon, kApply), "apply")
+ return make(maintain_, readWriteCache(join(kCommon, kMaintain)), "maintain") && make(inject_, join(kCommon, kSurface, kInject), "inject")
+ && make(gather_, readOnlyCache(join(kCommon, kSurface, kGather)), "gather") && make(apply_, join(kCommon, kApply), "apply")
&& make(sunInject_, join(kCommon, kSun), "sun");
}
@@ -722,8 +742,8 @@ rg::ResourceId VoxelGI::declare(rg::RenderGraph& graph, rg::ResourceId input, rg
const GfxComputeBinding bindings[8] = {
GfxComputeBinding::uniform(0, context_.frameUbo()),
GfxComputeBinding::uniform(1, params_),
- GfxComputeBinding::storage(2, grid_),
- GfxComputeBinding::storage(3, tags_),
+ GfxComputeBinding::storageRead(2, grid_),
+ GfxComputeBinding::storageRead(3, tags_),
GfxComputeBinding::depth(5, depth),
GfxComputeBinding::sampled(6, gbuffer),
GfxComputeBinding::storageTexture(7, history_[cur]),