From 837ccc2f31614495a7ffe535d86f42a70af96664 Mon Sep 17 00:00:00 2001 From: shenlei Date: Mon, 12 Oct 2026 16:12:14 +0900 Subject: [PATCH] Share GPU skinning palette per instance; push it once per change Render3DDraw/HardwareBatch hold the SDK's shared palette instead of copying 16 floats per bone per draw, and HardwareScene re-pushes the uniform only when the palette pointer changes. Co-Authored-By: Claude Opus 5.5 --- sdk | 2 +- src/eter_render/Gr2ModelRenderer.cpp | 4 ++-- src/eter_render/HardwareScene.cpp | 3 ++- src/eter_render/HardwareScene.h | 2 +- src/eter_render/RenderCommands3D.h | 3 +-- tests/ava2_render_3d_test.cpp | 10 ++++++++-- tests/hardware_scene_test.cpp | 8 ++++---- 7 files changed, 19 insertions(+), 13 deletions(-) diff --git a/sdk b/sdk index de8fa400..a1b39381 160000 --- a/sdk +++ b/sdk @@ -1 +1 @@ -Subproject commit de8fa4009831c954d99bc5d186b5e5bf82068f86 +Subproject commit a1b3938159e59453c311e4a00839da6b8890b4bf diff --git a/src/eter_render/Gr2ModelRenderer.cpp b/src/eter_render/Gr2ModelRenderer.cpp index 5bd4e541..04f777a6 100644 --- a/src/eter_render/Gr2ModelRenderer.cpp +++ b/src/eter_render/Gr2ModelRenderer.cpp @@ -67,7 +67,7 @@ Render3DDraw ToRender3DDraw(gr2_render::Gr2MeshDraw&& mesh) { draw.indices = std::move(mesh.indices); draw.bone_indices = std::move(mesh.bone_indices); draw.bone_weights = std::move(mesh.bone_weights); - draw.bone_matrices = std::move(mesh.bone_matrices); + draw.palette = std::move(mesh.palette); draw.alpha_test = mesh.alpha_test; draw.alpha_ref = mesh.alpha_ref; draw.cull_mode = mesh.double_sided ? 1 : 2; @@ -914,7 +914,7 @@ void Gr2SceneRenderer::RenderSceneGPU(SDL_Renderer* renderer, const std::vector< t.fogColor[0]=fog_r;t.fogColor[1]=fog_g;t.fogColor[2]=fog_b; t.fogParams[0]=m_fog_density;t.fogParams[1]=m_fog_near;t.fogParams[2]=m_fog_far; t.modes[0]=1;t.modes[1]=draw.lighting ? 1.f : 0.f;t.modes[2]=m_fog_enabled&&draw.fog ? (m_fog_exp ? 2.f : 1.f) : 0.f;t.modes[3]=draw.z_enable ? 0.f : 1.f; - if(draw.gpu_skinning){batch.palette=draw.bone_matrices;t.skin[0]=batch.palette.size()/16;} + if(draw.gpu_skinning && draw.palette){batch.palette=draw.palette;t.skin[0]=batch.palette->size()/16;} batch.indexCount=mesh.indices.size(); if(!batch.geometry) { const uint32_t base=m_hardwareVertices.size();batch.firstIndex=m_hardwareIndices.size(); diff --git a/src/eter_render/HardwareScene.cpp b/src/eter_render/HardwareScene.cpp index 364fc7fc..5a361be8 100644 --- a/src/eter_render/HardwareScene.cpp +++ b/src/eter_render/HardwareScene.cpp @@ -268,6 +268,7 @@ bool HardwareScene::Render(SDL_Renderer* renderer,int width,int height,const std auto* pass=SDL_BeginGPURenderPass(cmd,&color,1,&depth);if(!pass){SDL_CancelGPUCommandBuffer(cmd);return false;} SDL_BindGPUVertexStorageBuffers(pass,0,&instances_,1); float palette[128*16]{};SDL_PushGPUVertexUniformData(cmd,1,palette,sizeof(palette)); + const std::vector* pushedPalette=nullptr; // draws of one instance share a palette; push once for(size_t i=0;i1)transform.modes[0]=2; SDL_PushGPUVertexUniformData(cmd,0,&transform,sizeof(transform)); - if(transform.skin[0]>0){std::fill(std::begin(palette),std::end(palette),0);std::copy_n(batch.palette.data(),std::min(batch.palette.size(),size_t(128*16)),palette);SDL_PushGPUVertexUniformData(cmd,1,palette,sizeof(palette));} + if(transform.skin[0]>0 && batch.palette && batch.palette.get()!=pushedPalette){pushedPalette=batch.palette.get();std::fill(std::begin(palette),std::end(palette),0);std::copy_n(batch.palette->data(),std::min(batch.palette->size(),size_t(128*16)),palette);SDL_PushGPUVertexUniformData(cmd,1,palette,sizeof(palette));} SDL_DrawGPUIndexedPrimitives(pass,batch.indexCount,count,mesh ? 0 : batch.firstIndex,0,count>1 ? i : 0);++lastDrawCalls_;i+=count; } SDL_EndGPURenderPass(pass);const auto beforeSubmit=SDL_GetPerformanceCounter(); diff --git a/src/eter_render/HardwareScene.h b/src/eter_render/HardwareScene.h index 8d1969c4..82782ea4 100644 --- a/src/eter_render/HardwareScene.h +++ b/src/eter_render/HardwareScene.h @@ -31,7 +31,7 @@ struct HardwareBatch { float alphaRef; HardwareTransform transform{}; std::shared_ptr geometry; - std::vector palette; + std::shared_ptr> palette; }; class HardwareScene { diff --git a/src/eter_render/RenderCommands3D.h b/src/eter_render/RenderCommands3D.h index 079763e5..f4472954 100644 --- a/src/eter_render/RenderCommands3D.h +++ b/src/eter_render/RenderCommands3D.h @@ -52,7 +52,6 @@ struct Render3DDraw { command.indices.clear(); command.indices.shrink_to_fit(); command.bone_indices.clear(); command.bone_indices.shrink_to_fit(); command.bone_weights.clear(); command.bone_weights.shrink_to_fit(); - command.bone_matrices.clear(); command.bone_matrices.shrink_to_fit(); command.geometry = std::make_shared(std::move(source)); return command; } @@ -83,7 +82,7 @@ struct Render3DDraw { // Skeletal skinning streams std::vector bone_indices; // 4 uint8 per vertex std::vector bone_weights; // 4 floats per vertex - std::vector bone_matrices; // 16 floats per bone (palette) + std::shared_ptr> palette; // GPU skinning, 16 floats per bone; shared per instance // Render states bool z_enable = true; diff --git a/tests/ava2_render_3d_test.cpp b/tests/ava2_render_3d_test.cpp index 0f0b911e..59f948ae 100644 --- a/tests/ava2_render_3d_test.cpp +++ b/tests/ava2_render_3d_test.cpp @@ -18,6 +18,7 @@ #include #include #include +#include #include #include #include "test_rundir.h" @@ -716,8 +717,13 @@ int main(int argc, char** argv) { assert(cpu.size()==gpu.size()); auto again=instance.GetImpl()->BuildDrawCommands(time+.1f,nullptr,true); assert(again.size()==gpu.size()); + // One palette per instance per frame: draws of a mesh/part share it instead of copying. + std::set palettes;size_t skinned=0; + for(const auto& g:gpu)if(g.gpu_skinning){palettes.insert(g.palette.get());++skinned;} + assert(skinned<2 || palettes.size()=g.bone_matrices.size())continue; - const auto* m=g.bone_matrices.data()+bone*16; + if(w<=0 || size_t(bone)*16+15>=g.palette->size())continue; + const auto* m=g.palette->data()+bone*16; for(int k=0;k<3;++k){p[k]+=(mesh.positions[v*3]*m[k]+mesh.positions[v*3+1]*m[k+4]+mesh.positions[v*3+2]*m[k+8]+m[k+12])*w; n[k]+=(mesh.normals[v*3]*m[k]+mesh.normals[v*3+1]*m[k+4]+mesh.normals[v*3+2]*m[k+8])*w;} total+=w; diff --git a/tests/hardware_scene_test.cpp b/tests/hardware_scene_test.cpp index 90750bc9..242dd4ff 100644 --- a/tests/hardware_scene_test.cpp +++ b/tests/hardware_scene_test.cpp @@ -139,20 +139,20 @@ int main() { assert(resident.GetLastDrawCallCount()==2 && redAt(16)>245 && redAt(48)>120 && redAt(48)<135); SDL_DestroySurface(pixels); // Bone translation moves the raw triangle to the center; no CPU-deformed stream is uploaded. - left.transform.skin[0]=1;left.palette={1,0,0,0,0,1,0,0,0,0,1,0,.5f,0,0,1}; + left.transform.skin[0]=1;left.palette=std::make_shared>(std::vector{1,0,0,0,0,1,0,0,0,0,1,0,.5f,0,0,1}); auto skinned=std::make_shared(*mesh);skinned->bone_indices=std::vector(12,0);skinned->bone_weights={1,0,0,0,1,0,0,0,1,0,0,0};left.geometry=skinned; pixels=frame({left});assert(redAt(32)>245 && redAt(16)<5);SDL_DestroySurface(pixels); // Two weighted, rotated bones: compare every rendered pixel to CPU-transformed vertices. left.transform.skin[0]=2; - left.palette={1,0,0,0,0,1,0,0,0,0,1,0,.2f,0,0,1, - 0,1,0,0,-1,0,0,0,0,0,1,0,.7f,0,0,1}; + left.palette=std::make_shared>(std::vector{1,0,0,0,0,1,0,0,0,0,1,0,.2f,0,0,1, + 0,1,0,0,-1,0,0,0,0,0,1,0,.7f,0,0,1}); skinned=std::make_shared(*mesh); skinned->bone_indices={0,1,0,0,0,1,0,0,0,1,0,0}; skinned->bone_weights={.4f,.6f,0,0,.4f,.6f,0,0,.4f,.6f,0,0}; left.geometry=skinned;pixels=frame({left}); std::vector reference; for(size_t v=0;v<3;++v){auto vertex=ava2::MakeHardwareVertex(*skinned,v);float pos[4]{}; - for(size_t b=0;b<2;++b){const auto* matrix=left.palette.data()+b*16;float w=vertex.weights[b]; + for(size_t b=0;b<2;++b){const auto* matrix=left.palette->data()+b*16;float w=vertex.weights[b]; for(int k=0;k<4;++k)pos[k]+=(vertex.position[0]*matrix[k]+vertex.position[1]*matrix[k+4]+vertex.position[2]*matrix[k+8]+matrix[k+12])*w;} std::copy_n(pos,4,vertex.position);vertex.position[0]-=.5f;reference.push_back(vertex);} ava2::HardwareBatch cpu{};cpu.indexCount=3;cpu.cull=1;cpu.depth=true;cpu.depthWrite=true;cpu.alphaRef=-1;