enabled fp16 and fma cpu support

WickedEngine/CommonInclude.h

@@ -69,6 +69,21 @@ constexpr float bilinear(float4 gather, float2 pixel_frac)
 	return lerp(top_row, bottom_row, pixel_frac.y);
 }
 
+// Stack allocated string utility:
+template <unsigned capacity = 256>
+struct StackString
+{
+	char chars[capacity] = {};
+	unsigned cnt = 0;
+	static_assert(capacity > 1);
+	constexpr operator const char* () const { return chars; }
+	constexpr const char* const c_str() const { return chars; }
+	constexpr void push_back(const char* str) { while (*str != 0 && (cnt < (capacity - 1))) { chars[cnt++] = *str; str++; } }
+	constexpr unsigned size() const { return capacity; }
+	constexpr unsigned length() const { return cnt; }
+	constexpr bool empty() const { return cnt == 0; }
+};
+
 // CPU intrinsics:
 #if defined(_WIN32)
 // Windows, Xbox:
@@ -349,17 +364,8 @@ constexpr const char* relative_path(const char* path)
 // Extract function name from a string at compile-time
 constexpr auto extract_function_name(const char* str)
 {
-	struct ReturnString
-	{
-		char chars[256] = {};
-		constexpr operator const char* () const { return chars; }
-		constexpr const char* const c_str() const { return chars; }
-	} ret;
-	int i = 0;
-	for (const char* currentCharacter = str; *currentCharacter != '\0' && *currentCharacter != '(' && (i < sizeof(ret.chars) - 1); ++currentCharacter)
-	{
-		ret.chars[i++] = *currentCharacter;
-	}
+	StackString ret;
+	ret.push_back(str);
 	return ret;
 }
 

WickedEngine/wiInitializer.cpp

@@ -1,7 +1,6 @@
 #include "wiInitializer.h"
 #include "WickedEngine.h"
 
-#include <string>
 #include <thread>
 #include <atomic>
 
@@ -40,16 +39,50 @@ namespace wi::initializer
 		static constexpr const char* platform_string = "Xbox";
 #endif // PLATFORM
 
-		wilog("\n[wi::initializer] Initializing Wicked Engine, please wait...\nVersion: %s\nPlatform: %s", wi::version::GetVersionString(), platform_string)
-		
+		wilog("\n[wi::initializer] Initializing Wicked Engine, please wait...\nVersion: %s\nPlatform: %s", wi::version::GetVersionString(), platform_string);
+
+		StackString cpustring;
+		cpustring.push_back("\nCPU features used by this build: ");
+#ifdef _XM_SSE_INTRINSICS_
+		cpustring.push_back("SSE 2; ");
+#endif // _XM_SSE_INTRINSICS_
+#ifdef _XM_SSE3_INTRINSICS_
+		cpustring.push_back("SSE 3; ");
+#endif // _XM_SSE3_INTRINSICS_
+#ifdef _XM_SSE4_INTRINSICS_
+		cpustring.push_back("SSE 4; ");
+#endif // _XM_SSE4_INTRINSICS_
+#ifdef _XM_AVX_INTRINSICS_
+		cpustring.push_back("AVX; ");
+#endif // _XM_AVX_INTRINSICS_
+#ifdef _XM_AVX2_INTRINSICS_
+		cpustring.push_back("AVX 2; ");
+#endif // _XM_AVX2_INTRINSICS_
+#ifdef _XM_ARM_NEON_INTRINSICS_
+		cpustring.push_back("NEON; ");
+#endif // _XM_ARM_NEON_INTRINSICS_
+#ifdef _XM_F16C_INTRINSICS_
+		cpustring.push_back("F16C; ");
+#endif // _XM_F16C_INTRINSICS_
+#ifdef _XM_FMA3_INTRINSICS_
+		cpustring.push_back("FMA3; ");
+#endif // _XM_FMA3_INTRINSICS_
+
+		wilog(cpustring);
+
+		if (!XMVerifyCPUSupport())
+		{
+			wilog_messagebox("XMVerifyCPUSupport() failed! This means that your CPU doesn't support a required feature! %s", cpustring.c_str());
+		}
+
 		size_t shaderdump_count = wi::renderer::GetShaderDumpCount();
 		if (shaderdump_count > 0)
 		{
-			wi::backlog::post("\nEmbedded shaders found: " + std::to_string(shaderdump_count));
+			wilog("\nEmbedded shaders found: %d", (int)shaderdump_count);
 		}
 		else
 		{
-			wi::backlog::post("\nNo embedded shaders found, shaders will be compiled at runtime if needed.\n\tShader source path: " + wi::renderer::GetShaderSourcePath() + "\n\tShader binary path: " + wi::renderer::GetShaderPath());
+			wilog("\nNo embedded shaders found, shaders will be compiled at runtime if needed.\n\tShader source path: %s\n\tShader binary path: %s", wi::renderer::GetShaderSourcePath().c_str(), wi::renderer::GetShaderPath().c_str());
 		}
 
 		wi::backlog::post("");

WickedEngine/wiMath.h

@@ -5,6 +5,9 @@
 #include <algorithm>
 #include <limits>
 
+#define _XM_F16C_INTRINSICS_
+#define _XM_FMA3_INTRINSICS_
+
 #if __has_include("DirectXMath.h")
 // In this case, DirectXMath is coming from Windows SDK.
 //	It is better to use this on Windows as some Windows libraries could depend on the same

WickedEngine/wiRenderer.cpp

@@ -194,7 +194,7 @@ Texture texture_curlNoise;
 Texture texture_weatherMap;
 
 // Direct reference to a renderable instance:
-struct RenderBatch
+struct alignas(16) RenderBatch
 {
 	uint32_t meshIndex;
 	uint32_t instanceIndex;

WickedEngine/wiVersion.cpp

@@ -9,7 +9,7 @@ namespace wi::version
 	// minor features, major updates, breaking compatibility changes
 	const int minor = 71;
 	// minor bug fixes, alterations, refactors, updates
-	const int revision = 743;
+	const int revision = 744;
 
 	const std::string version_string = std::to_string(major) + "." + std::to_string(minor) + "." + std::to_string(revision);