Merge branch 'nahim_cjit' of github.com:Devsh-Graphics-Programming/Nabla

Author: devshgraphicsprogramming

include/nbl/asset/utils/CJITIncludeLoader.h

@@ -0,0 +1,29 @@
+#ifndef _NBL_VIDEO_C_JIT_INCLUDE_LOADER_H_INCLUDED_
+#define _NBL_VIDEO_C_JIT_INCLUDE_LOADER_H_INCLUDED_
+
+#include "nbl/asset/utils/IShaderCompiler.h"
+
+#include "nbl/video/SPhysicalDeviceFeatures.h"
+#include "nbl/video/SPhysicalDeviceLimits.h"
+
+#include <string>
+
+
+namespace nbl::video
+{
+
+class NBL_API2 CJITIncludeLoader : public asset::IShaderCompiler::IIncludeLoader
+{
+    public:
+        CJITIncludeLoader(const SPhysicalDeviceLimits& limits, const SPhysicalDeviceFeatures& features);
+        std::optional<std::string> getInclude(const system::path& searchPath, const std::string& includeName) const override;
+
+    private:
+        core::unordered_map<system::path, std::string> m_includes;
+        std::string collectDeviceCaps(const SPhysicalDeviceLimits& limits, const SPhysicalDeviceFeatures& features);
+
+};
+
+} //nbl::video
+
+#endif // CJITINCLUDELOADER_H

include/nbl/video/ILogicalDevice.h

@@ -29,7 +29,7 @@
 #include "nbl/video/CThreadSafeGPUQueueAdapter.h"
 #include "nbl/video/IDeviceMemoryAllocator.h"
 
-#include "nbl/video/SPhysicalDeviceFeatures.h"
+#include "nbl/video/CJITIncludeLoader.h"
 
 namespace nbl::video
 {
@@ -502,8 +502,12 @@ class NBL_API2 ILogicalDevice : public core::IReferenceCounted, public IDeviceMe
                 (*m_offsets)[qci.familyIndex + 1u] = qci.count;
             }
             std::inclusive_scan(m_offsets->begin(),m_offsets->end(),m_offsets->begin());
+
+            addJITIncludeLoader();
         }
 
+        void addJITIncludeLoader();
+
         // must be called by implementations of mapMemory()
         static void post_mapMemory(IDeviceMemoryAllocation* memory, void* ptr, IDeviceMemoryAllocation::MemoryRange rng, core::bitflag<IDeviceMemoryAllocation::E_MAPPING_CPU_ACCESS_FLAGS> access) 
         {

src/nbl/CMakeLists.txt

@@ -198,7 +198,7 @@ set(NBL_ASSET_SOURCES
 	${NBL_ROOT_PATH}/src/nbl/asset/interchange/CGLSLLoader.cpp
 	${NBL_ROOT_PATH}/src/nbl/asset/interchange/CHLSLLoader.cpp
 	${NBL_ROOT_PATH}/src/nbl/asset/interchange/CSPVLoader.cpp
-	
+
 # Pipeline loaders
 	${NBL_ROOT_PATH}/src/nbl/asset/interchange/CGraphicsPipelineLoaderMTL.cpp
 
@@ -280,6 +280,7 @@ set(NBL_VIDEO_SOURCES
 	${NBL_ROOT_PATH}/src/nbl/video/IGPUDescriptorSet.cpp
 	${NBL_ROOT_PATH}/src/nbl/video/IDeviceMemoryAllocation.cpp
 	${NBL_ROOT_PATH}/src/nbl/video/IDeviceMemoryBacked.cpp
+	${NBL_ROOT_PATH}/src/nbl/video/CJITIncludeLoader.cpp
 
 # Vulkan
 	${NBL_ROOT_PATH}/src/nbl/video/CVulkanSwapchain.cpp

src/nbl/asset/utils/CJITIncludeLoader.cpp

@@ -0,0 +1,58 @@
+#include "nbl/asset/utils/CJITIncludeLoader.h"
+
+namespace nbl::video
+{
+
+CJITIncludeLoader::CJITIncludeLoader(const SPhysicalDeviceLimits& limits, const SPhysicalDeviceFeatures& features)
+{
+    m_includes["nbl/builtin/hlsl/jit/device_capabilities.hlsl"] = collectDeviceCaps(limits, features);
+}
+
+std::optional<std::string> CJITIncludeLoader::getInclude(const system::path& searchPath, const std::string& includeName) const
+{
+    system::path path = searchPath / includeName;
+
+    assert(searchPath == "nbl/builtin/hlsl/jit");
+    
+    // Look up the content in m_includes map
+    auto it = m_includes.find(path);
+    if (it != m_includes.end())
+    {
+        // Return the content of the specified include file
+        return it->second;
+    }
+
+    return std::nullopt;
+}
+
+
+std::string CJITIncludeLoader::collectDeviceCaps(const SPhysicalDeviceLimits& limits, const SPhysicalDeviceFeatures& features)
+{
+    return R"===(
+        #ifndef _NBL_BUILTIN_HLSL_JIT_DEVICE_CAPABILITIES_INCLUDED_
+        #define _NBL_BUILTIN_HLSL_JIT_DEVICE_CAPABILITIES_INCLUDED_
+
+        namespace nbl
+        {
+        namespace hlsl
+        {
+        namespace jit
+        {
+            struct device_capabilities
+            {
+                NBL_CONSTEXPR_STATIC_INLINE bool shaderFloat64 = )===" + std::to_string(features.shaderFloat64) + R"===(;
+                NBL_CONSTEXPR_STATIC_INLINE bool shaderDrawParameters = )===" + std::to_string(features.shaderDrawParameters) + R"===(;
+                NBL_CONSTEXPR_STATIC_INLINE bool subgroupArithmetic = )===" + std::to_string(limits.shaderSubgroupArithmetic) + R"===(;
+                NBL_CONSTEXPR_STATIC_INLINE bool fragmentShaderPixelInterlock = )===" + std::to_string(features.fragmentShaderPixelInterlock) + R"===(;
+
+                NBL_CONSTEXPR_STATIC_INLINE uint16_t maxOptimallyResidentWorkgroupInvocations = )===" + std::to_string(limits.maxOptimallyResidentWorkgroupInvocations) + R"===(;
+            };
+        }
+        }
+        }
+
+        #endif  // _NBL_BUILTIN_HLSL_JIT_DEVICE_CAPABILITIES_INCLUDED_
+    )===";
+}
+
+}

src/nbl/video/ILogicalDevice.cpp

@@ -6,6 +6,17 @@ using namespace nbl::video;
 
 E_API_TYPE ILogicalDevice::getAPIType() const { return m_physicalDevice->getAPIType(); }
 
+void ILogicalDevice::addJITIncludeLoader()
+{
+    if(auto cc = (m_physicalDevice && m_compilerSet) ? m_compilerSet->getShaderCompiler(asset::IShader::E_CONTENT_TYPE::ECT_HLSL) : nullptr)
+    {
+        if(auto finder = cc->getDefaultIncludeFinder())
+        {
+            finder->addSearchPath("nbl/builtin/hlsl/jit", core::make_smart_refctd_ptr<CJITIncludeLoader>(m_physicalDevice->getLimits(), m_physicalDevice->getFeatures()));
+        }
+    }
+}
+
 core::smart_refctd_ptr<IGPUDescriptorSetLayout> ILogicalDevice::createDescriptorSetLayout(const IGPUDescriptorSetLayout::SBinding* _begin, const IGPUDescriptorSetLayout::SBinding* _end)
 {
     uint32_t dynamicSSBOCount=0u,dynamicUBOCount=0u;