Main: add support for Mesh & Task shaders

Joilnen · Aug 11, 2024 · bf585ac · bf585ac
1 parent 9cb9cdc
commit bf585ac
Show file tree

Hide file tree

Showing 24 changed files with 222 additions and 12 deletions.
diff --git a/OgreMain/include/OgreGpuProgram.h b/OgreMain/include/OgreGpuProgram.h
@@ -52,10 +52,12 @@ namespace Ogre {
         GPT_GEOMETRY_PROGRAM,
         GPT_DOMAIN_PROGRAM,
         GPT_HULL_PROGRAM,
-        GPT_COMPUTE_PROGRAM
+        GPT_MESH_PROGRAM, // aliases with vertex
+        GPT_COMPUTE_PROGRAM, // aliases with fragment
+        GPT_TASK_PROGRAM
     };
     enum {
-        GPT_COUNT = GPT_COMPUTE_PROGRAM + 1,
+        GPT_COUNT = GPT_TASK_PROGRAM + 1,
         /// max programs that can be active in a pipeline (e.g compute is separate)
         GPT_PIPELINE_COUNT = GPT_HULL_PROGRAM + 1
     };

diff --git a/OgreMain/include/OgreRenderSystemCapabilities.h b/OgreMain/include/OgreRenderSystemCapabilities.h
@@ -171,6 +171,8 @@ namespace Ogre
         RSC_RTT_DEPTHBUFFER_RESOLUTION_LESSEQUAL = OGRE_CAPS_VALUE(CAPS_CATEGORY_COMMON_2, 15),
         /// Supports using vertex buffers for instance data
         RSC_VERTEX_BUFFER_INSTANCE_DATA = OGRE_CAPS_VALUE(CAPS_CATEGORY_COMMON_2, 16),
+        /// Supports mesh and task programs
+        RSC_MESH_PROGRAM = OGRE_CAPS_VALUE(CAPS_CATEGORY_COMMON_2, 17),
 
         // ***** DirectX specific caps *****
         /// Is DirectX feature "per stage constants" supported

diff --git a/OgreMain/include/OgreScriptCompiler.h b/OgreMain/include/OgreScriptCompiler.h
@@ -833,7 +833,10 @@ namespace Ogre
         ID_CAMERA,
         ID_ALIGN_TO_FACE,
         ID_UNORDERED_ACCESS_MIP,
-
+        ID_MESH_PROGRAM,
+        ID_MESH_PROGRAM_REF,
+        ID_TASK_PROGRAM,
+        ID_TASK_PROGRAM_REF,
         ID_END_BUILTIN_IDS
     };
     /** @} */

diff --git a/OgreMain/src/OgreGpuProgram.cpp b/OgreMain/src/OgreGpuProgram.cpp
@@ -233,7 +233,8 @@ namespace Ogre
         if ((getType() == GPT_GEOMETRY_PROGRAM && !caps->hasCapability(RSC_GEOMETRY_PROGRAM)) ||
             ((getType() == GPT_DOMAIN_PROGRAM || getType() == GPT_HULL_PROGRAM) &&
              !caps->hasCapability(RSC_TESSELLATION_PROGRAM)) ||
-            (getType() == GPT_COMPUTE_PROGRAM && !caps->hasCapability(RSC_COMPUTE_PROGRAM)))
+            (getType() == GPT_COMPUTE_PROGRAM && !caps->hasCapability(RSC_COMPUTE_PROGRAM)) ||
+            ((getType() == GPT_MESH_PROGRAM || getType() == GPT_TASK_PROGRAM) && !caps->hasCapability(RSC_MESH_PROGRAM)))
         {
             return false;
         }
@@ -376,6 +377,10 @@ namespace Ogre
             return "hull";
         case GPT_COMPUTE_PROGRAM:
             return "compute";
+        case GPT_MESH_PROGRAM:
+            return "mesh";
+        case GPT_TASK_PROGRAM:
+            return "task";
         default:
             OGRE_EXCEPT(Exception::ERR_INVALIDPARAMS,
                 "Unexpected GPU program type",

diff --git a/OgreMain/src/OgreRenderSystemCapabilities.cpp b/OgreMain/src/OgreRenderSystemCapabilities.cpp
@@ -134,6 +134,7 @@ namespace Ogre {
             pLog->logMessage("   - Domain program constant 4-vectors: " +
                              StringConverter::toString(mConstantFloatCount[GPT_DOMAIN_PROGRAM]));
         }
+        pLog->logMessage(" * Mesh programs: " + StringConverter::toString(hasCapability(RSC_MESH_PROGRAM), true));
         pLog->logMessage(" * Compute programs: " + StringConverter::toString(hasCapability(RSC_COMPUTE_PROGRAM), true));
         if (hasCapability(RSC_COMPUTE_PROGRAM))
         {

diff --git a/OgreMain/src/OgreSceneManager.cpp b/OgreMain/src/OgreSceneManager.cpp
@@ -25,6 +25,7 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
 THE SOFTWARE.
 -----------------------------------------------------------------------------
 */
+#include "OgreGpuProgram.h"
 #include "OgreStableHeaders.h"
 
 #include "OgreControllerManager.h"
@@ -667,7 +668,8 @@ const Pass* SceneManager::_setPass(const Pass* pass, bool shadowDerivation)
     {
         bindGpuProgram(vprog->_getBindingDelegate());
     }
-    else if (!mDestRenderSystem->getCapabilities()->hasCapability(RSC_FIXED_FUNCTION))
+    else if (!mDestRenderSystem->getCapabilities()->hasCapability(RSC_FIXED_FUNCTION) &&
+             !pass->hasGpuProgram(GPT_MESH_PROGRAM))
     {
         OGRE_EXCEPT(Exception::ERR_INVALID_STATE,
                     "RenderSystem does not support FixedFunction, "
@@ -686,7 +688,7 @@ const Pass* SceneManager::_setPass(const Pass* pass, bool shadowDerivation)
         // Set fixed-function vertex parameters
     }
 
-    for(auto gptype : {GPT_DOMAIN_PROGRAM, GPT_HULL_PROGRAM, GPT_GEOMETRY_PROGRAM})
+    for(auto gptype : {GPT_DOMAIN_PROGRAM, GPT_HULL_PROGRAM, GPT_GEOMETRY_PROGRAM, GPT_MESH_PROGRAM, GPT_TASK_PROGRAM})
     {
         if (pass->hasGpuProgram(gptype))
         {

diff --git a/OgreMain/src/OgreScriptCompiler.cpp b/OgreMain/src/OgreScriptCompiler.cpp
@@ -914,6 +914,8 @@ namespace Ogre
         mIds["tessellation_hull_program"] = ID_TESSELLATION_HULL_PROGRAM;
         mIds["tessellation_domain_program"] = ID_TESSELLATION_DOMAIN_PROGRAM;
         mIds["compute_program"] = ID_COMPUTE_PROGRAM;
+        mIds["mesh_program"] = ID_MESH_PROGRAM;
+        mIds["task_program"] = ID_TASK_PROGRAM;
         mIds["technique"] = ID_TECHNIQUE;
         mIds["pass"] = ID_PASS;
         mIds["texture_unit"] = ID_TEXTURE_UNIT;
@@ -923,6 +925,8 @@ namespace Ogre
         mIds["tessellation_hull_program_ref"] = ID_TESSELLATION_HULL_PROGRAM_REF;
         mIds["tessellation_domain_program_ref"] = ID_TESSELLATION_DOMAIN_PROGRAM_REF;
         mIds["compute_program_ref"] = ID_COMPUTE_PROGRAM_REF;
+        mIds["mesh_program_ref"] = ID_MESH_PROGRAM_REF;
+        mIds["task_program_ref"] = ID_TASK_PROGRAM_REF;
         mIds["shadow_caster_vertex_program_ref"] = ID_SHADOW_CASTER_VERTEX_PROGRAM_REF;
         mIds["shadow_caster_fragment_program_ref"] = ID_SHADOW_CASTER_FRAGMENT_PROGRAM_REF;
         mIds["shadow_receiver_vertex_program_ref"] = ID_SHADOW_RECEIVER_VERTEX_PROGRAM_REF;

diff --git a/OgreMain/src/OgreScriptTranslator.cpp b/OgreMain/src/OgreScriptTranslator.cpp
@@ -26,6 +26,7 @@ THE SOFTWARE.
 -----------------------------------------------------------------------------
 */
 
+#include "OgreScriptCompiler.h"
 #include "OgreStableHeaders.h"
 #include "OgreBuiltinScriptTranslators.h"
 #include "OgreGpuProgramManager.h"
@@ -533,6 +534,12 @@ namespace Ogre{
         case ID_COMPUTE_PROGRAM:
         case ID_COMPUTE_PROGRAM_REF:
             return GPT_COMPUTE_PROGRAM;
+        case ID_MESH_PROGRAM:
+        case ID_MESH_PROGRAM_REF:
+            return GPT_MESH_PROGRAM;
+        case ID_TASK_PROGRAM:
+        case ID_TASK_PROGRAM_REF:
+            return GPT_TASK_PROGRAM;
         }
     }
 
@@ -2212,6 +2219,8 @@ namespace Ogre{
                 case ID_TESSELLATION_HULL_PROGRAM_REF:
                 case ID_TESSELLATION_DOMAIN_PROGRAM_REF:
                 case ID_COMPUTE_PROGRAM_REF:
+                case ID_MESH_PROGRAM_REF:
+                case ID_TASK_PROGRAM_REF:
                     translateProgramRef(getProgramType(child->id), compiler, child);
                     break;
                 case ID_SHADOW_CASTER_VERTEX_PROGRAM_REF:
@@ -2235,6 +2244,8 @@ namespace Ogre{
                 case ID_TESSELLATION_HULL_PROGRAM:
                 case ID_TESSELLATION_DOMAIN_PROGRAM:
                 case ID_COMPUTE_PROGRAM:
+                case ID_MESH_PROGRAM:
+                case ID_TASK_PROGRAM:
                 {
                     // auto assign inline defined programs
                     processNode(compiler, i);
@@ -5264,7 +5275,9 @@ namespace Ogre{
                     obj->id == ID_GEOMETRY_PROGRAM ||
                     obj->id == ID_TESSELLATION_HULL_PROGRAM || 
                     obj->id == ID_TESSELLATION_DOMAIN_PROGRAM ||
-                    obj->id == ID_COMPUTE_PROGRAM)
+                    obj->id == ID_COMPUTE_PROGRAM ||
+                    obj->id == ID_MESH_PROGRAM ||
+                    obj->id == ID_TASK_PROGRAM)
                 translator = &mGpuProgramTranslator;
             else if(obj->id == ID_SHARED_PARAMS)
                 translator = &mSharedParamsTranslator;

diff --git a/OgreMain/src/OgreTechnique.cpp b/OgreMain/src/OgreTechnique.cpp
@@ -188,7 +188,7 @@ namespace Ogre {
             if (!caps->hasCapability(RSC_FIXED_FUNCTION) && currPass->isProgrammable() &&
                 !currPass->hasGpuProgram(GPT_COMPUTE_PROGRAM))
             {
-                if (!currPass->hasVertexProgram() ||
+                if ((!currPass->hasVertexProgram() && !currPass->hasGpuProgram(GPT_MESH_PROGRAM)) ||
                     (!currPass->hasFragmentProgram() && !currPass->hasGeometryProgram()))
                 {
                     compileErrors << "Pass " << passNum << ": RenderSystem requires both vertex and fragment programs";

diff --git a/PlugIns/CgProgramManager/src/OgreCgFxScriptLoader.cpp b/PlugIns/CgProgramManager/src/OgreCgFxScriptLoader.cpp
@@ -26,6 +26,7 @@ THE SOFTWARE.
 -----------------------------------------------------------------------------
 */
 #include "OgreCgFxScriptLoader.h"
+#include "OgreGpuProgram.h"
 #include "OgreResourceGroupManager.h"
 #include "OgreMaterialManager.h"
 #include "OgreTechnique.h"
@@ -2814,6 +2815,8 @@ namespace Ogre {
         case GPT_COMPUTE_PROGRAM:
         case GPT_DOMAIN_PROGRAM:
         case GPT_HULL_PROGRAM:
+        case GPT_MESH_PROGRAM:
+        case GPT_TASK_PROGRAM:
             break;
         }
         CGstateassignment cgStateAssignment = cgGetNamedStateAssignment(cgPass, stateName);
@@ -2833,6 +2836,8 @@ namespace Ogre {
             case GPT_COMPUTE_PROGRAM:
             case GPT_DOMAIN_PROGRAM:
             case GPT_HULL_PROGRAM:
+            case GPT_MESH_PROGRAM:
+            case GPT_TASK_PROGRAM:
                 break;
             }
 
@@ -2906,6 +2911,8 @@ namespace Ogre {
             case GPT_DOMAIN_PROGRAM:
             case GPT_COMPUTE_PROGRAM:
             case GPT_HULL_PROGRAM:
+            case GPT_MESH_PROGRAM:
+            case GPT_TASK_PROGRAM:
                 break;
             }
 

diff --git a/PlugIns/GLSLang/src/OgreGLSLang.cpp b/PlugIns/GLSLang/src/OgreGLSLang.cpp
@@ -317,6 +317,8 @@ static EShLanguage getShLanguage(GpuProgramType type)
     case GPT_HULL_PROGRAM:      return EShLangTessControl;
     case GPT_DOMAIN_PROGRAM:    return EShLangTessEvaluation;
     case GPT_COMPUTE_PROGRAM:   return EShLangCompute;
+    case GPT_MESH_PROGRAM:      return EShLangMeshNV;
+    case GPT_TASK_PROGRAM:      return EShLangTaskNV;
         // clang-format on
     }
 
@@ -331,6 +333,13 @@ GLSLangProgram::GLSLangProgram(ResourceManager* creator, const String& name, Res
     {
         setupBaseParamDictionary();
         memset(&DefaultTBuiltInResource.limits, 1, sizeof(TLimits));
+
+        if(sizeof(TBuiltInResource) == 420) // replace with build_info.h, when it is universally available
+        {
+            // copy VK_NV_mesh_shader limits to VK_EXT_mesh_shader
+            memcpy(&DefaultTBuiltInResource.maxMeshViewCountNV + 1, &DefaultTBuiltInResource.maxMeshOutputVerticesNV,
+                   9 * sizeof(int));
+        }
     }
 }
 
@@ -412,8 +421,16 @@ void GLSLangProgram::prepareImpl()
     if(mSyntaxCode == "gl_spirv")
         shader.setEnvClient(glslang::EShClientOpenGL, glslang::EShTargetOpenGL_450);
     else if(mSyntaxCode == "spirv")
+    {
         shader.setEnvClient(glslang::EShClientVulkan, glslang::EShTargetVulkan_1_0);
 
+        if(mType == GPT_MESH_PROGRAM || mType == GPT_TASK_PROGRAM)
+        {
+            shader.setEnvClient(glslang::EShClientVulkan, glslang::EShTargetVulkan_1_2);
+            shader.setEnvTarget(glslang::EShTargetSpv, glslang::EShTargetSpv_1_4);
+        }
+    }
+
     // minimal version is 430 for explicit uniform location, but we use latest to get all features
     if (!shader.parse(&DefaultTBuiltInResource, 460, false, EShMsgSpvRules))
     {

diff --git a/RenderSystems/Direct3D11/src/OgreD3D11HLSLProgram.cpp b/RenderSystems/Direct3D11/src/OgreD3D11HLSLProgram.cpp
@@ -859,6 +859,10 @@ namespace Ogre {
         case GPT_COMPUTE_PROGRAM:
             CreateComputeShader();
             break;
+        case GPT_MESH_PROGRAM:
+        case GPT_TASK_PROGRAM:
+            OgreAssertDbg(false, "should never get here");
+            break;
         }
     }
 

diff --git a/RenderSystems/Direct3D11/src/OgreD3D11RenderSystem.cpp b/RenderSystems/Direct3D11/src/OgreD3D11RenderSystem.cpp
@@ -2250,6 +2250,10 @@ namespace Ogre
                 mBoundComputeProgram = static_cast<D3D11HLSLProgram*>(prg);
             }
             break;
+        case GPT_MESH_PROGRAM:
+        case GPT_TASK_PROGRAM:
+            OGRE_EXCEPT(Exception::ERR_RENDERINGAPI_ERROR, "Mesh and task shaders are not supported in D3D11");
+            break;
         };
 
         RenderSystem::bindGpuProgram(prg);

diff --git a/RenderSystems/GL/src/GLSL/src/OgreGLSLProgram.cpp b/RenderSystems/GL/src/GLSL/src/OgreGLSLProgram.cpp
@@ -208,6 +208,8 @@ namespace Ogre {
             case GPT_COMPUTE_PROGRAM:
             case GPT_DOMAIN_PROGRAM:
             case GPT_HULL_PROGRAM:
+            case GPT_MESH_PROGRAM:
+            case GPT_TASK_PROGRAM:
                 break;
             }
             mGLShaderHandle = (size_t)glCreateShaderObjectARB(shaderType);

diff --git a/RenderSystems/GL/src/OgreGLRenderSystem.cpp b/RenderSystems/GL/src/OgreGLRenderSystem.cpp
@@ -2389,6 +2389,8 @@ namespace Ogre {
         case GPT_COMPUTE_PROGRAM:
         case GPT_DOMAIN_PROGRAM:
         case GPT_HULL_PROGRAM:
+        case GPT_MESH_PROGRAM:
+        case GPT_TASK_PROGRAM:
             break;
         }
 
@@ -2444,6 +2446,8 @@ namespace Ogre {
         case GPT_COMPUTE_PROGRAM:
         case GPT_DOMAIN_PROGRAM:
         case GPT_HULL_PROGRAM:
+        case GPT_MESH_PROGRAM:
+        case GPT_TASK_PROGRAM:
             break;
         }
     }

diff --git a/RenderSystems/GL3Plus/src/GLSL/src/OgreGLSLSeparableProgram.cpp b/RenderSystems/GL3Plus/src/GLSL/src/OgreGLSLSeparableProgram.cpp
@@ -67,12 +67,14 @@ namespace Ogre
             }
         }
 
+            #define GL_MESH_SHADER_BIT_NV 0x00000040
             GLenum ogre2gltype[GPT_COUNT] = {
                 GL_VERTEX_SHADER_BIT,
                 GL_FRAGMENT_SHADER_BIT,
                 GL_GEOMETRY_SHADER_BIT,
                 GL_TESS_EVALUATION_SHADER_BIT,
                 GL_TESS_CONTROL_SHADER_BIT,
+                GL_MESH_SHADER_BIT_NV,
                 GL_COMPUTE_SHADER_BIT
             };
 

diff --git a/RenderSystems/GL3Plus/src/GLSL/src/OgreGLSLShader.cpp b/RenderSystems/GL3Plus/src/GLSL/src/OgreGLSLShader.cpp
@@ -235,6 +235,8 @@ namespace Ogre {
     /// Get OpenGL GLSL shader type from OGRE GPU program type.
     static GLenum getGLShaderType(GpuProgramType programType)
     {
+        #define GL_MESH_SHADER_NV 0x9559
+        #define GL_TASK_SHADER_NV 0x955A
         switch (programType)
         {
         case GPT_VERTEX_PROGRAM:
@@ -249,6 +251,10 @@ namespace Ogre {
             return GL_FRAGMENT_SHADER;
         case GPT_COMPUTE_PROGRAM:
             return GL_COMPUTE_SHADER;
+        case GPT_MESH_PROGRAM:
+            return GL_MESH_SHADER_NV;
+        case GPT_TASK_PROGRAM:
+            return GL_TASK_SHADER_NV;
         }
 
         return 0;
@@ -352,6 +358,8 @@ namespace Ogre {
                         break;
                     case GPT_FRAGMENT_PROGRAM:
                     case GPT_COMPUTE_PROGRAM:
+                    case GPT_MESH_PROGRAM:
+                    case GPT_TASK_PROGRAM:
                         // Fragment and compute shaders do
                         // not have standard blocks.
                         break;

diff --git a/RenderSystems/GL3Plus/src/OgreGL3PlusRenderSystem.cpp b/RenderSystems/GL3Plus/src/OgreGL3PlusRenderSystem.cpp
@@ -418,6 +418,11 @@ namespace Ogre {
             // OGRE_CHECK_GL_ERROR(glGetIntegerv(GL_MAX_COMPUTE_WORK_GROUP_INVOCATIONS, &workgroupInvocations));
         }
 
+        if (checkExtension("GL_NV_mesh_shader"))
+        {
+            rsc->setCapability(RSC_MESH_PROGRAM);
+        }
+
         if (hasMinGLVersion(4, 1) || checkExtension("GL_ARB_get_program_binary"))
         {
             GLint formats = 0;
@@ -1102,6 +1107,14 @@ namespace Ogre {
                 //                OGRE_CHECK_GL_ERROR(glDrawArraysInstanced(GL_PATCHES, 0, primCount, 1));
             }
         }
+        else if (mCurrentShader[GPT_MESH_PROGRAM])
+        {
+            OgreAssert(op.indexData, "indexData required for mesh shader");
+
+            typedef void (APIENTRYP PFNGLDRAWMESHTASKSNVPROC) (GLuint first, GLuint count);
+            auto glDrawMeshTasksNV = (PFNGLDRAWMESHTASKSNVPROC)mGLSupport->getProcAddress("glDrawMeshTasksNV");
+            OGRE_CHECK_GL_ERROR(glDrawMeshTasksNV(0, op.indexData->indexCount));
+        }
         else if (op.useIndexes)
         {
             void *pBufferData = VBO_BUFFER_OFFSET(op.indexData->indexStart *