Refactor getting host D3DFORMAT, use it for COLORSIGN, and add signed-to-unsigned mapping

src/core/hle/D3D8/Direct3D9/CxbxPixelShaderTemplate.hlsl

@@ -253,14 +253,23 @@ static bool alphakill[4] = ALPHAKILL;
 
 static const float4 WarningColor = float4(0, 1, 1, 1); // Returned when unhandled scenario is encountered
 
+#define unsigned_to_signed(x) (((x) * 2) - 1) // Shifts range from [0..1] to [-1..1] (just like s_bx2)
+#define signed_to_unsigned(x) (((x) + 1) / 2) // Shifts range from [-1..1] to [0..1]
+
 float4 PerformColorSign(const float4 ColorSign, float4 t)
 {
-	// Per color channel, optionally convert the value range into two's complement signed values (from (0, +1) to (-1, +1), using s_bx2):
+	// Per color channel, based on the ColorSign setting :
-	// This is often used for bumpmaps
+	// either keep the value range as-is (when ColorSign is zero)
-	if (ColorSign.r > 0) t.r = s_bx2(t.r);
+	// or convert from [0..1] to [-1..+1] (when ColorSign is more than zero, often used for bumpmaps),
-	if (ColorSign.g > 0) t.g = s_bx2(t.g);
+	// or convert from [-1..1] to [0..1] (when ColorSign is less than zero):
-	if (ColorSign.b > 0) t.b = s_bx2(t.b);
+	if (ColorSign.r > 0) t.r = unsigned_to_signed(t.r);
-	if (ColorSign.a > 0) t.a = s_bx2(t.a);
+	if (ColorSign.g > 0) t.g = unsigned_to_signed(t.g);
+	if (ColorSign.b > 0) t.b = unsigned_to_signed(t.b);
+	if (ColorSign.a > 0) t.a = unsigned_to_signed(t.a);
+	if (ColorSign.r < 0) t.r = signed_to_unsigned(t.r);
+	if (ColorSign.g < 0) t.g = signed_to_unsigned(t.g);
+	if (ColorSign.b < 0) t.b = signed_to_unsigned(t.b);
+	if (ColorSign.a < 0) t.a = signed_to_unsigned(t.a);
 	// TODO : Instead of the above, create a mirror texture with a host format that has identical component layout, but with all components signed.
 	// Then, in here, when any component has to be read as signed, sample the signed texture (ouch : with what dimension and coordinate?!)
 	// and replace the components that we read from the unsigned texture, but which have to be signed, with the signed components read from the signed mirror texture.

src/core/hle/D3D8/Direct3D9/Direct3D9.cpp

@@ -184,6 +184,7 @@ static xbox::PVOID                   g_pXbox_Palette_Data[xbox::X_D3DTS_STAGECOU
 static unsigned                     g_Xbox_Palette_Size[xbox::X_D3DTS_STAGECOUNT] = { 0 }; // cached palette size
 
 
+             D3DFORMAT               g_HostTextureFormats[xbox::X_D3DTS_STAGECOUNT]; // Updated by CxbxUpdateHostTextures(), read by CxbxCalcColorSign
        xbox::X_D3DBaseTexture       *g_pXbox_SetTexture[xbox::X_D3DTS_STAGECOUNT] = {0,0,0,0}; // Set by our D3DDevice_SetTexture and D3DDevice_SwitchTexture patches
 static xbox::X_D3DBaseTexture        CxbxActiveTextureCopies[xbox::X_D3DTS_STAGECOUNT] = {}; // Set by D3DDevice_SwitchTexture. Cached active texture
 
@@ -905,8 +906,13 @@ void *GetDataFromXboxResource(xbox::X_D3DResource *pXboxResource)
 	return (uint8_t*)pData;
 }
 
+DWORD D3DUSAGE_INVALID = -1;
+
 typedef struct _resource_info_t {
+	
 	ComPtr<IDirect3DResource> pHostResource;
+	D3DFORMAT HostFormat = D3DFMT_UNKNOWN;
+	DWORD HostUsage = D3DUSAGE_INVALID;
 	DWORD dwXboxResourceType = 0;
 	void* pXboxData = xbox::zeroptr;
 	size_t szXboxDataSize = 0;
@@ -1149,6 +1155,42 @@ void SetHostResource(xbox::X_D3DResource* pXboxResource, IDirect3DResource* pHos
 	resourceInfo.lastUpdate = std::chrono::steady_clock::now();
 	resourceInfo.nextHashTime = resourceInfo.lastUpdate + resourceInfo.hashLifeTime;
 	resourceInfo.forceRehash = false;
+
+	HRESULT hRet = STATUS_INVALID_PARAMETER; // Default to an error condition, so we can use D3D_OK to check for success
+	D3DSURFACE_DESC surfaceDesc;
+	D3DVOLUME_DESC volumeDesc;
+	UINT Level = 0; // TODO : When should Level every be something other than zero, and if so : what other value?
+	switch (resourceInfo.dwXboxResourceType) {// TODO : Better check pHostResource class type
+	case xbox::X_D3DRTYPE_SURFACE:
+		hRet = ((IDirect3DSurface*)pHostResource)->GetDesc(&surfaceDesc);
+		break;
+	case xbox::X_D3DRTYPE_TEXTURE:
+		hRet = ((IDirect3DTexture*)pHostResource)->GetLevelDesc(Level, &surfaceDesc);
+		break;
+	case xbox::X_D3DRTYPE_VOLUMETEXTURE: {
+		hRet = ((IDirect3DVolumeTexture*)pHostResource)->GetLevelDesc(Level, &volumeDesc);
+		break; }
+	case xbox::X_D3DRTYPE_CUBETEXTURE:
+		hRet = ((IDirect3DCubeTexture*)pHostResource)->GetLevelDesc(Level, &surfaceDesc);
+		break;
+	}
+
+	D3DFORMAT hostFormat = D3DFMT_UNKNOWN;
+	DWORD hostUsage = D3DUSAGE_INVALID;
+
+	if (SUCCEEDED(hRet)) {
+		if (resourceInfo.dwXboxResourceType == xbox::X_D3DRTYPE_VOLUMETEXTURE) {
+			hostFormat = volumeDesc.Format;
+			hostUsage = volumeDesc.Usage;
+		}
+		else {
+			hostFormat = surfaceDesc.Format;
+			hostUsage = surfaceDesc.Usage;
+		}
+	}
+
+	resourceInfo.HostFormat = hostFormat;
+	resourceInfo.HostUsage = hostUsage;
 }
 
 IDirect3DSurface *GetHostSurface(xbox::X_D3DResource *pXboxResource, DWORD D3DUsage = 0)
@@ -2280,30 +2322,20 @@ static void EmuVerifyResourceIsRegistered(xbox::X_D3DResource *pResource, DWORD
 	auto& ResourceCache = GetResourceCache(key);
 	auto it = ResourceCache.find(key);
 	if (it != ResourceCache.end()) {
+		// TODO : Should this check be (D3DUsage & D3DUSAGE_RENDERTARGET) instead?
 		if (D3DUsage == D3DUSAGE_RENDERTARGET && IsResourceAPixelContainer(pResource) && EmuXBFormatIsRenderTarget(GetXboxPixelContainerFormat((xbox::X_D3DPixelContainer*)pResource))) {
             // Render targets have special behavior: We can't trash them on guest modification
             // this fixes an issue where CubeMaps were broken because the surface Set in GetCubeMapSurface
             // would be overwritten by the surface created in SetRenderTarget
             // However, if a non-rendertarget surface is used here, we'll need to recreate it as a render target!
-            auto hostResource = it->second.pHostResource.Get();
+            auto hostResource = it->second.pHostResource;
-            auto xboxSurface = (xbox::X_D3DSurface*)pResource;
+            assert(xboxResourceType == GetXboxD3DResourceType(pResource));
-            auto xboxTexture = (xbox::X_D3DTexture*)pResource;
-            auto xboxResourceType = GetXboxD3DResourceType(pResource);
-
-            // Determine if the associated host resource is a render target already, if so, do nothing
-            HRESULT hRet = STATUS_INVALID_PARAMETER; // Default to an error condition, so we can use D3D_OK to check for success
-            D3DSURFACE_DESC surfaceDesc;
-            if (xboxResourceType == xbox::X_D3DRTYPE_SURFACE) {
-                hRet = ((IDirect3DSurface*)hostResource)->GetDesc(&surfaceDesc);
-            } else if (xboxResourceType == xbox::X_D3DRTYPE_TEXTURE) {
-                hRet = ((IDirect3DTexture*)hostResource)->GetLevelDesc(0, &surfaceDesc);
-            }
 
             // Only continue checking if we were able to get the surface desc, if it failed, we fall-through
             // to previous resource management behavior
-            if (SUCCEEDED(hRet)) {
+            if (it->second.HostUsage != D3DUSAGE_INVALID) {
                 // If this resource is already created as a render target on the host, simply return
-                if (surfaceDesc.Usage & D3DUSAGE_RENDERTARGET) {
+                if (it->second.HostUsage & D3DUSAGE_RENDERTARGET) {
                     return;
                 }
 
@@ -2311,6 +2343,8 @@ static void EmuVerifyResourceIsRegistered(xbox::X_D3DResource *pResource, DWORD
                 // We need to re-create it as a render target
                 switch (xboxResourceType) {
                     case xbox::X_D3DRTYPE_SURFACE: {
+                        auto xboxSurface = (xbox::X_D3DSurface*)pResource;
+
                         // Free the host surface
                         FreeHostResource(key);
 
@@ -7301,7 +7335,6 @@ void CxbxUpdateHostTextures()
 		auto pXboxBaseTexture = g_pXbox_SetTexture[stage];
 		IDirect3DBaseTexture* pHostBaseTexture = nullptr;
 		bool bNeedRelease = false;
-
 		if (pXboxBaseTexture != xbox::zeroptr) {
 			DWORD XboxResourceType = GetXboxCommonResourceType(pXboxBaseTexture);
 			switch (XboxResourceType) {
@@ -7320,6 +7353,15 @@ void CxbxUpdateHostTextures()
 				LOG_TEST_CASE("ActiveTexture set to an unhandled resource type!");
 				break;
 			}
+
+			// Read HostFormat from GetResourceCache :
+			// TODO : Optimize this, as we're doing the lookup twice (once in GetHostBaseTexture, once here)
+			auto key = GetHostResourceKey(pXboxBaseTexture, stage);
+			auto& ResourceCache = GetResourceCache(key);
+			auto it = ResourceCache.find(key);
+			if (it != ResourceCache.end()) {
+				g_HostTextureFormats[stage] = it->second.HostFormat;
+			}
 		}
 
 		HRESULT hRet = g_pD3DDevice->SetTexture(stage, pHostBaseTexture);

src/core/hle/D3D8/Direct3D9/Direct3D9.h

@@ -59,6 +59,8 @@ extern xbox::dword_xt g_Xbox_VertexShader_Handle;
 
 extern xbox::X_PixelShader *g_pXbox_PixelShader;
 
+extern D3DFORMAT g_HostTextureFormats[xbox::X_D3DTS_STAGECOUNT];
+
 extern xbox::X_D3DBaseTexture *g_pXbox_SetTexture[xbox::X_D3DTS_STAGECOUNT];
 
 namespace xbox {

src/core/hle/D3D8/Direct3D9/FixedFunctionPixelShader.hlsl

@@ -4,17 +4,23 @@
 
 static const float4 WarningColor = float4(0, 1, 1, 1); // Returned when unhandled scenario is encountered
 
-#define s_bx2(x)    (( 2 * max(0, x)) - 1)  // PS_INPUTMAPPING_EXPAND_NORMAL=     0x40L, // invalid for final combiner // Shifts range from [0..1] to [-1..1]
+#define unsigned_to_signed(x) (((x) * 2) - 1) // Shifts range from [0..1] to [-1..1] (just like s_bx2)
-#define s_bias(x)         (max(0, x) - 0.5) // PS_INPUTMAPPING_HALFBIAS_NORMAL=   0x80L, // invalid for final combiner // Clamps negative x to 0 and then subtracts 0.5
+#define signed_to_unsigned(x) (((x) + 1) / 2) // Shifts range from [-1..1] to [0..1]
 
 float4 PerformColorSign(const float4 ColorSign, float4 t)
 {
-	// Per color channel, optionally convert the value range into two's complement signed values (from (0, +1) to (-1, +1), using s_bx2):
+	// Per color channel, based on the ColorSign setting :
-	// This is often used for bumpmaps
+	// either keep the value range as-is (when ColorSign is zero)
-	if (ColorSign.r > 0) t.r = s_bx2(t.r);
+	// or convert from [0..1] to [-1..+1] (when ColorSign is more than zero, often used for bumpmaps),
-	if (ColorSign.g > 0) t.g = s_bx2(t.g);
+	// or convert from [-1..1] to [0..1] (when ColorSign is less than zero):
-	if (ColorSign.b > 0) t.b = s_bx2(t.b);
+	if (ColorSign.r > 0) t.r = unsigned_to_signed(t.r);
-	if (ColorSign.a > 0) t.a = s_bx2(t.a);
+	if (ColorSign.g > 0) t.g = unsigned_to_signed(t.g);
+	if (ColorSign.b > 0) t.b = unsigned_to_signed(t.b);
+	if (ColorSign.a > 0) t.a = unsigned_to_signed(t.a);
+	if (ColorSign.r < 0) t.r = signed_to_unsigned(t.r);
+	if (ColorSign.g < 0) t.g = signed_to_unsigned(t.g);
+	if (ColorSign.b < 0) t.b = signed_to_unsigned(t.b);
+	if (ColorSign.a < 0) t.a = signed_to_unsigned(t.a);
 	// TODO : Instead of the above, create a mirror texture with a host format that has identical component layout, but with all components signed.
 	// Then, in here, when any component has to be read as signed, sample the signed texture (ouch : with what dimension and coordinate?!)
 	// and replace the components that we read from the unsigned texture, but which have to be signed, with the signed components read from the signed mirror texture.

src/core/hle/D3D8/XbPixelShader.cpp

@@ -968,28 +968,54 @@ float AsFloat(uint32_t value)
 	return *(float*)&v;
 }
 
-// This mimicks behaviour of XDK LazySetShaderStageProgram, which we bypass due to our drawing patches without trampolines.
+// Determines the Cxbx ColorSign requirement, as handled in the HLSL shaders by PerformColorSign()
-DWORD CxbxGetColorSign(int stage_nr)
+float CxbxComponentColorSignFromXboxAndHost(bool XboxMarksComponentSigned, bool HostComponentIsSigned)
 {
-	// When bump environment mapping is enabled
+	// Equal "signedness" between Xbox and host implies we must not convert the component scale :
-	if (XboxTextureStates.Get(stage_nr, xbox::X_D3DTSS_COLOROP) >= xbox::X_D3DTOP_BUMPENVMAP)
+	if (XboxMarksComponentSigned == HostComponentIsSigned)
-		// Always mark the blue (alias for U) and green (alias for  V) color channels as signed :
+		return 0.0f;
-		return xbox::X_D3DTSIGN_GSIGNED | xbox::X_D3DTSIGN_BSIGNED;
+
+	// Xbox wants the components to be signed (even though host has them unsigned)
+	if (XboxMarksComponentSigned)
+		return 1.0f; // Mark the component for scaling from unsigned_to_signed
+
+	// Xbox doesn't want signed values, but host has them signed :
+	return -1.0f; // Mark the component for scaling from signed_to_unsigned
+}
+
+D3DXCOLOR CxbxCalcColorSign(int stage_nr)
+{
+	// Without overrides, just use what the running executable put in COLORSIGN :
+	DWORD XboxColorSign = XboxTextureStates.Get(stage_nr, xbox::X_D3DTSS_COLORSIGN);
+
+	{ // This mimicks behaviour of XDK LazySetShaderStageProgram, which we bypass due to our drawing patches without trampolines.
+		// When bump environment mapping is enabled
+		if (XboxTextureStates.Get(stage_nr, xbox::X_D3DTSS_COLOROP) >= xbox::X_D3DTOP_BUMPENVMAP)
+			// Always mark the blue (alias for U) and green (alias for  V) color channels as signed :
+			XboxColorSign |= xbox::X_D3DTSIGN_GSIGNED | xbox::X_D3DTSIGN_BSIGNED;
+	}
 
 #if 0 // When this block is enabled, XDK samples BumpEarth and BumpLens turn red-ish, so keep this off for now...
-	// TODO : Perhaps turn COLORSIGN off when the active host texture is already signed (like D3DFMT_L6V5U5 if supported)
-	//              and enable COLORSIGN when the active host texture was converted from signed to unsigned?
-
 	// Check if the pixel shader specifies bump mapping for this stage (TODO : How to handle this with the fixed function shader?)
 	DWORD PSTextureModes = XboxRenderStates.GetXboxRenderState(xbox::X_D3DRS_PSTEXTUREMODES);
 	PS_TEXTUREMODES StageTextureMode = (PS_TEXTUREMODES)((PSTextureModes >> (stage_nr * 5)) & PS_TEXTUREMODES_MASK);
 	if (StageTextureMode == PS_TEXTUREMODES_BUMPENVMAP || StageTextureMode == PS_TEXTUREMODES_BUMPENVMAP_LUM)
-		return xbox::X_D3DTSIGN_GSIGNED | xbox::X_D3DTSIGN_BSIGNED;
+		XboxColorSign |= xbox::X_D3DTSIGN_GSIGNED | xbox::X_D3DTSIGN_BSIGNED;
 
 #endif
+	// Host D3DFMT's with one or more signed components : D3DFMT_V8U8, D3DFMT_Q8W8V8U8, D3DFMT_V16U16, D3DFMT_Q16W16V16U16, D3DFMT_CxV8U8
+	D3DFORMAT HostTextureFormat = g_HostTextureFormats[stage_nr];
+	bool HostTextureFormatIsSignedForA = (HostTextureFormat == D3DFMT_Q8W8V8U8); // No need to check for unused formats : D3DFMT_Q16W16V16U16, D3DFMT_CxV8U8, D3DFMT_A2W10V10U10
+	bool HostTextureFormatIsSignedForR = (HostTextureFormat == D3DFMT_V8U8) || (HostTextureFormat == D3DFMT_V16U16) || (HostTextureFormat == D3DFMT_L6V5U5) || (HostTextureFormat == D3DFMT_X8L8V8U8) || HostTextureFormatIsSignedForA;
+	bool HostTextureFormatIsSignedForG = HostTextureFormatIsSignedForR;
+	bool HostTextureFormatIsSignedForB = HostTextureFormatIsSignedForA;
 
-	// Without overrides, just return what the running executable put in COLORSIGN :
+	D3DXCOLOR CxbxColorSign;
-	return XboxTextureStates.Get(stage_nr, xbox::X_D3DTSS_COLORSIGN);
+	CxbxColorSign.r = CxbxComponentColorSignFromXboxAndHost(XboxColorSign & xbox::X_D3DTSIGN_RSIGNED, HostTextureFormatIsSignedForR); // Maps to COLORSIGN.r
+	CxbxColorSign.g = CxbxComponentColorSignFromXboxAndHost(XboxColorSign & xbox::X_D3DTSIGN_GSIGNED, HostTextureFormatIsSignedForG); // Maps to COLORSIGN.g
+	CxbxColorSign.b = CxbxComponentColorSignFromXboxAndHost(XboxColorSign & xbox::X_D3DTSIGN_BSIGNED, HostTextureFormatIsSignedForB); // Maps to COLORSIGN.b
+	CxbxColorSign.a = CxbxComponentColorSignFromXboxAndHost(XboxColorSign & xbox::X_D3DTSIGN_ASIGNED, HostTextureFormatIsSignedForA); // Maps to COLORSIGN.a
+	return CxbxColorSign;
 }
 
 // Set constant state for the fixed function pixel shader
@@ -1010,11 +1036,11 @@ void UpdateFixedFunctionPixelShaderState()
 	for (int i = 0; i < xbox::X_D3DTS_STAGECOUNT; i++) {
 		auto stage = &ffPsState.stages[i];
 		stage->COLORKEYOP = (float)XboxTextureStates.Get(i, xbox::X_D3DTSS_COLORKEYOP);
-		DWORD colorsign = CxbxGetColorSign(i);
+		auto CxbxColorSign = CxbxCalcColorSign(i);
-		stage->COLORSIGN.x = (colorsign & xbox::X_D3DTSIGN_RSIGNED) ? 1.0f : 0.0f; // Maps to COLORSIGN.r
+		stage->COLORSIGN.x = CxbxColorSign.r;
-		stage->COLORSIGN.y = (colorsign & xbox::X_D3DTSIGN_GSIGNED) ? 1.0f : 0.0f; // Maps to COLORSIGN.g
+		stage->COLORSIGN.y = CxbxColorSign.g;
-		stage->COLORSIGN.z = (colorsign & xbox::X_D3DTSIGN_BSIGNED) ? 1.0f : 0.0f; // Maps to COLORSIGN.b
+		stage->COLORSIGN.z = CxbxColorSign.b;
-		stage->COLORSIGN.w = (colorsign & xbox::X_D3DTSIGN_ASIGNED) ? 1.0f : 0.0f; // Maps to COLORSIGN.a
+		stage->COLORSIGN.w = CxbxColorSign.a;
 		stage->ALPHAKILL = (float)XboxTextureStates.Get(i, xbox::X_D3DTSS_ALPHAKILL);
 		stage->BUMPENVMAT00 = AsFloat(XboxTextureStates.Get(i, xbox::X_D3DTSS_BUMPENVMAT00));
 		stage->BUMPENVMAT01 = AsFloat(XboxTextureStates.Get(i, xbox::X_D3DTSS_BUMPENVMAT01));
@@ -1136,11 +1162,7 @@ void DxbxUpdateActivePixelShader() // NOPATCH
 
   // Texture color sign
   for (int stage_nr = 0; stage_nr < xbox::X_D3DTS_STAGECOUNT; stage_nr++) {
-    auto colorsign = CxbxGetColorSign(stage_nr);
+    fColor[PSH_XBOX_CONSTANT_COLORSIGN + stage_nr] = CxbxCalcColorSign(stage_nr);
-    fColor[PSH_XBOX_CONSTANT_COLORSIGN + stage_nr].r = (colorsign & xbox::X_D3DTSIGN_RSIGNED) ? 1.0f : 0.0f; // Maps to COLORSIGN[stage].r (treated as bool, true when > 0.0)
-    fColor[PSH_XBOX_CONSTANT_COLORSIGN + stage_nr].g = (colorsign & xbox::X_D3DTSIGN_GSIGNED) ? 1.0f : 0.0f; // Maps to COLORSIGN[stage].g
-    fColor[PSH_XBOX_CONSTANT_COLORSIGN + stage_nr].b = (colorsign & xbox::X_D3DTSIGN_BSIGNED) ? 1.0f : 0.0f; // Maps to COLORSIGN[stage].b
-    fColor[PSH_XBOX_CONSTANT_COLORSIGN + stage_nr].a = (colorsign & xbox::X_D3DTSIGN_ASIGNED) ? 1.0f : 0.0f; // Maps to COLORSIGN[stage].a
   }
 
 #if 0 // New, doesn't work yet