minor fixes and debug

Author: devshgraphicsprogramming

include/nbl/builtin/hlsl/glsl_compat/core.hlsl

@@ -14,6 +14,7 @@ namespace hlsl
 namespace glsl
 {
 
+#ifdef __HLSL_VERSION
 template<typename T>
 T atomicAdd(NBL_REF_ARG(T) ptr, T value)
 {
@@ -59,7 +60,8 @@ T atomicCompSwap(NBL_REF_ARG(T) ptr, T comparator, T value)
  * For Compute Shaders
  */
 
- // TODO (Future): Its annoying we have to forward declare those, but accessing gl_NumSubgroups and other gl_* values is not yet possible due to https://github.com/microsoft/DirectXShaderCompiler/issues/4217
+// TODO (Future): Its annoying we have to forward declare those, but accessing gl_NumSubgroups and other gl_* values is not yet possible due to https://github.com/microsoft/DirectXShaderCompiler/issues/4217
+// also https://github.com/microsoft/DirectXShaderCompiler/issues/5280
 uint32_t gl_LocalInvocationIndex();
 uint32_t3 gl_GlobalInvocationID();
 uint32_t3 gl_WorkGroupID();
@@ -78,6 +80,7 @@ void tess_ctrl_barrier() {
 void memoryBarrierShared() {
     spirv::memoryBarrier(spv::ScopeDevice, spv::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsWorkgroupMemoryMask);
 }
+#endif
 
 }
 }

include/nbl/builtin/hlsl/glsl_compat/subgroup_basic.hlsl

@@ -14,6 +14,7 @@ namespace hlsl
 namespace glsl
 {
 
+#ifdef __HLSL_VERSION
 uint32_t gl_SubgroupSize() {
     return WaveGetLaneCount();
 }
@@ -55,6 +56,7 @@ void subgroupMemoryBarrierShared() {
 void subgroupMemoryBarrierImage() {
     spirv::memoryBarrier(spv::ScopeSubgroup, spv::MemorySemanticsAcquireReleaseMask | spv::MemorySemanticsImageMemoryMask);
 }
+#endif
 
 }
 }

include/nbl/builtin/hlsl/subgroup/arithmetic_portability.hlsl

@@ -17,10 +17,10 @@ namespace hlsl
 namespace subgroup
 {
 
-#ifdef NBL_GL_KHR_shader_subgroup_arithmetic
-#define IMPL portability
-#else
+#ifdef NBL_GLSL_LIMIT_SHADER_SUBGROUP_ARITHMETIC
 #define IMPL native
+#else
+#define IMPL portability
 #endif
 
 template<class Binop>

include/nbl/builtin/hlsl/subgroup/arithmetic_portability_impl.hlsl

@@ -83,10 +83,10 @@ struct inclusive_scan
         const uint subgroupInvocation = glsl::gl_SubgroupInvocationID();
         const uint halfSubgroupSize = glsl::gl_SubgroupSize() >> 1u;
 
-        uint rhs = glsl::subgroupShuffleUp<type_t>(value, 1u); // all invocations must execute the shuffle, even if we don't apply the op() to all of them
+        type_t rhs = glsl::subgroupShuffleUp<type_t>(value, 1u); // all invocations must execute the shuffle, even if we don't apply the op() to all of them
         value = op(value, subgroupInvocation<1u ? Binop::identity:rhs);
 
-        [[unroll(MinSubgroupSizeLog2-1)]]
+        [[unroll(nbl::hlsl::subgroup::MinSubgroupSizeLog2-1)]]
         for (uint step=2u; step<=halfSubgroupSize; step <<= 1u)
         {
             rhs = glsl::subgroupShuffleUp<type_t>(value, step);

src/nbl/asset/utils/CHLSLCompiler.cpp

@@ -313,6 +313,18 @@ std::string CHLSLCompiler::preprocessShader(std::string&& code, IShader::E_SHADE
     auto resolvedString = proc.Process();
     IShaderCompiler::reenableDirectives(resolvedString);
 
+    // for debugging cause MSVC doesn't like to show more than 21k LoC in TextVisualizer
+    if constexpr (true)
+    {
+        system::ISystem::future_t<core::smart_refctd_ptr<system::IFile>> future;
+        m_system->createFile(future,system::path(preprocessOptions.sourceIdentifier).parent_path()/"preprocessed.hlsl",system::IFileBase::ECF_WRITE);
+        if (auto file=future.acquire(); file&&bool(*file))
+        {
+            system::IFile::success_t succ;
+            (*file)->write(succ,resolvedString.data(),0,resolvedString.size()+1);
+        }
+    }
+
     return resolvedString;
 }
 

src/nbl/video/CVulkanLogicalDevice.h

@@ -733,6 +733,18 @@ class CVulkanLogicalDevice final : public ILogicalDevice
         vk_createInfo.flags = static_cast<VkShaderModuleCreateFlags>(0u); // reserved for future use by Vulkan
         vk_createInfo.codeSize = spirv->getSize();
         vk_createInfo.pCode = static_cast<const uint32_t*>(spirv->getPointer());
+
+        // for debugging 
+        if constexpr (true)
+        {
+            system::ISystem::future_t<core::smart_refctd_ptr<system::IFile>> future;
+            m_physicalDevice->getSystem()->createFile(future,system::path(cpushader->getFilepathHint()).parent_path()/"compiled.spv", system::IFileBase::ECF_WRITE);
+            if (auto file = future.acquire(); file && bool(*file))
+            {
+                system::IFile::success_t succ;
+                (*file)->write(succ,vk_createInfo.pCode,0,vk_createInfo.codeSize);
+            }
+        }
 
         VkShaderModule vk_shaderModule;
         if (m_devf.vk.vkCreateShaderModule(m_vkdev, &vk_createInfo, nullptr, &vk_shaderModule) == VK_SUCCESS)