fix D3D11 shaders and samplers

src/gapi/d3d11.h

@@ -52,6 +52,8 @@ void D3DCHECK(HRESULT res) {
     #define D3DCHECK(res) res
 #endif
 
+#define MAX_SAMPLERS 5
+
 namespace GAPI {
     using namespace Core;
 
@@ -73,6 +75,8 @@ namespace GAPI {
     ID3D11RasterizerState   *RS[cmMAX];    // [cullMode]
     ID3D11DepthStencilState *DS[2][2];     // [depthTest][depthWrite]
 
+    ID3D11SamplerState      *samplers[MAX_SAMPLERS];
+
 // Shader
     #include "shaders/d3d11/shaders.h"
 
@@ -272,13 +276,12 @@ namespace GAPI {
         ID3D11ShaderResourceView *SRV;
         ID3D11RenderTargetView   *RTV;
         ID3D11DepthStencilView   *DSV;
-        ID3D11SamplerState       *sampler;
 
         int       width, height, depth, origWidth, origHeight, origDepth;
         TexFormat fmt;
         uint32    opt;
 
-        Texture(int width, int height, int depth, uint32 opt) : tex2D(NULL), tex3D(NULL), SRV(NULL), RTV(NULL), DSV(NULL), sampler(NULL), width(width), height(height), depth(depth), origWidth(width), origHeight(height), origDepth(depth), fmt(FMT_RGBA), opt(opt) {}
+        Texture(int width, int height, int depth, uint32 opt) : tex2D(NULL), tex3D(NULL), SRV(NULL), RTV(NULL), DSV(NULL), width(width), height(height), depth(depth), origWidth(width), origHeight(height), origDepth(depth), fmt(FMT_RGBA), opt(opt) {}
 
         void init(void *data) {
             ASSERT((opt & OPT_PROXY) == 0);
@@ -389,8 +392,6 @@ namespace GAPI {
             }
 
             ASSERT(SRV);
-
-            initSampler();
         }
 
         void deinit() {
@@ -399,48 +400,6 @@ namespace GAPI {
             SAFE_RELEASE(SRV);
             SAFE_RELEASE(RTV);
             SAFE_RELEASE(DSV);
-            SAFE_RELEASE(sampler);
-        }
-
-        void initSampler() {
-            SAFE_RELEASE(sampler);
-
-            D3D11_SAMPLER_DESC desc;
-            memset(&desc, 0, sizeof(desc));
-
-            bool mipmaps = (opt & OPT_MIPMAPS) != 0;
-            bool filter  = (opt & OPT_NEAREST) == 0;
-
-            if (mipmaps && (opt & (OPT_VOLUME | OPT_CUBEMAP | OPT_NEAREST)) == 0 && (Core::support.maxAniso > 0)) {
-                desc.MaxAnisotropy = min(int(Core::support.maxAniso), 8);
-            } else {
-                desc.MaxAnisotropy = 1;
-            }
-
-            if (desc.MaxAnisotropy > 1) {
-                desc.Filter = D3D11_FILTER_ANISOTROPIC;
-            } else {
-                if (filter) {
-                    desc.Filter = D3D11_FILTER_MIN_MAG_MIP_LINEAR;
-                } else {
-                    desc.Filter = D3D11_FILTER_MIN_MAG_MIP_POINT;
-                }
-            }
-
-            if (fmt == FMT_SHADOW) {
-                desc.Filter = D3D11_FILTER_COMPARISON_MIN_MAG_LINEAR_MIP_POINT;
-                desc.ComparisonFunc = D3D11_COMPARISON_LESS;
-            } else {
-                desc.ComparisonFunc = D3D11_COMPARISON_NEVER;
-            }
-
-            desc.AddressU       =
-            desc.AddressV       =
-            desc.AddressW       = (opt & OPT_REPEAT) ? D3D11_TEXTURE_ADDRESS_WRAP : D3D11_TEXTURE_ADDRESS_CLAMP;
-            
-            desc.MinLOD         = 0;
-            desc.MaxLOD         = D3D11_FLOAT32_MAX;
-            device->CreateSamplerState(&desc, &sampler);
         }
 
         void generateMipMap() {
@@ -466,10 +425,8 @@ namespace GAPI {
 
                 if (opt & OPT_VERTEX) {
                     deviceContext->VSSetShaderResources(sampler, 1, &SRV);
-                    deviceContext->VSSetSamplers(sampler, 1, &this->sampler);
                 }
                 deviceContext->PSSetShaderResources(sampler, 1, &SRV);
-                deviceContext->PSSetSamplers(sampler, 1, &this->sampler);
             }
         }
 
@@ -562,6 +519,66 @@ namespace GAPI {
         } items[MAX_RENDER_BUFFERS];
     } rtCache;
 
+    void deinitSamplers() {
+        for (int i = 0; i < COUNT(samplers); i++) {
+            SAFE_RELEASE(samplers[i]);
+        }
+    }
+
+    ID3D11SamplerState* initSampler(bool filter, bool aniso, bool wrap, bool cmp) {
+        D3D11_SAMPLER_DESC desc;
+        memset(&desc, 0, sizeof(desc));
+
+        if (aniso && (Core::support.maxAniso > 0)) {
+            desc.MaxAnisotropy = min(int(Core::support.maxAniso), 8);
+        } else {
+            desc.MaxAnisotropy = 1;
+        }
+
+        if (desc.MaxAnisotropy > 1) {
+            desc.Filter = D3D11_FILTER_ANISOTROPIC;
+        } else {
+            if (filter) {
+                desc.Filter = D3D11_FILTER_MIN_MAG_MIP_LINEAR;
+            } else {
+                desc.Filter = D3D11_FILTER_MIN_MAG_MIP_POINT;
+            }
+        }
+
+        if (cmp) {
+            desc.Filter = D3D11_FILTER_COMPARISON_MIN_MAG_LINEAR_MIP_POINT;
+            desc.ComparisonFunc = D3D11_COMPARISON_LESS;
+        } else {
+            desc.ComparisonFunc = D3D11_COMPARISON_NEVER;
+        }
+
+        desc.AddressU =
+        desc.AddressV =
+        desc.AddressW = wrap ? D3D11_TEXTURE_ADDRESS_WRAP : D3D11_TEXTURE_ADDRESS_CLAMP;
+        desc.MinLOD   = 0;
+        desc.MaxLOD   = D3D11_FLOAT32_MAX;
+
+        ID3D11SamplerState *sampler;
+        device->CreateSamplerState(&desc, &sampler);
+        return sampler;
+    }
+
+    void initSamplers() {
+        deinitSamplers();
+    /*
+        0 - smpDefault
+        1 - smpPoint
+        2 - smpPointWrap
+        3 - smpLinear
+        4 - smpCmp
+    */
+        samplers[0] = initSampler(true,  true,  false, false); // TODO settings dependent
+        samplers[1] = initSampler(false, false, false, false);
+        samplers[2] = initSampler(false, false, true,  false);
+        samplers[3] = initSampler(true,  false, false, false);
+        samplers[4] = initSampler(true,  false, false, true);
+    }
+
     void init() {
         memset(&rtCache, 0, sizeof(rtCache));
         
@@ -628,6 +645,7 @@ namespace GAPI {
             #undef BLEND_FUNC
         }
 
+    // init raster state
         {
             D3D11_RASTERIZER_DESC desc;
             memset(&desc, 0, sizeof(desc));
@@ -641,6 +659,7 @@ namespace GAPI {
             device->CreateRasterizerState(&desc, &RS[cmFront]);
         }
 
+    // init depth stencil states
         {
             D3D11_DEPTH_STENCIL_DESC desc;
             memset(&desc, 0, sizeof(desc));
@@ -656,6 +675,10 @@ namespace GAPI {
                 }
             }
         }
+
+    // init samplers
+        memset(samplers, 0, sizeof(samplers));
+        initSamplers();
     }
 
     void resetDevice() {
@@ -666,6 +689,9 @@ namespace GAPI {
             SAFE_RELEASE(rtCache.items[i].RTV);
             SAFE_RELEASE(rtCache.items[i].DSV);
         }
+
+        deinitSamplers();
+
         rtCache.count = 0;
     }
 
@@ -747,6 +773,9 @@ namespace GAPI {
         deviceContext->RSSetState(RS[cmNone]);
         depthTest = depthWrite = dirtyDepthState = true;
         colorWrite = dirtyBlendState = true;
+
+        deviceContext->VSGetSamplers(0, COUNT(samplers), samplers);
+        deviceContext->PSGetSamplers(0, COUNT(samplers), samplers);
     }
 
     void cacheRenderTarget(ID3D11RenderTargetView **RTV, ID3D11DepthStencilView **DSV, int width, int height) {

src/shaders/common.hlsl

@@ -39,8 +39,8 @@ struct VS_INPUT {
 #ifdef _GAPI_D3D11
 	SamplerState           smpDefault   : register(s0);
 	SamplerState           smpPoint     : register(s1);
-	SamplerState           smpLinear  : register(s2);
+	SamplerState           smpPointWrap : register(s2);
-	SamplerState           smpAniso   : register(s3);
+	SamplerState           smpLinear    : register(s3);
 	SamplerComparisonState smpCmp       : register(s4);
 
 	Texture2D    sDiffuse     : register(t0);
@@ -54,13 +54,14 @@ struct VS_INPUT {
 	TextureCube  sEnvironment : register(t4);
 	Texture2D    sMask        : register(t5);
 	
-	#define SAMPLE_2D        (T,uv) T.Sample(smpDefault, uv)
+	#define SAMPLE_2D(T,uv)            T.Sample(smpDefault,   uv)
-	#define SAMPLE_2D_POINT  (T,uv) T.Sample(smpPoint,   uv)
+	#define SAMPLE_2D_POINT(T,uv)      T.Sample(smpPoint,     uv)
-	#define SAMPLE_2D_LINEAR (T,uv) T.Sample(smpLinear,  uv)
+	#define SAMPLE_2D_POINT_WRAP(T,uv) T.Sample(smpPointWrap, uv)
-	#define SAMPLE_2D_ANISO  (T,uv) T.Sample(smpAniso,   uv)
+	#define SAMPLE_2D_LINEAR(T,uv)     T.Sample(smpLinear,    uv)
-	#define SAMPLE_2D_CMP    (T,uv) T.SampleCmp(smpCmp,  uv.xy, uv.z)
+	#define SAMPLE_2D_CMP(T,uv)        T.SampleCmp(smpCmp,    uv.xy, uv.z)
-	#define SAMPLE_3D        (T,uv) T.Sample(smpLinear,  uv)
+	#define SAMPLE_2D_LOD0(T,uv)       T.SampleLevel(smpLinear, uv, 0)
-	#define SAMPLE_CUBE      (T,uv) T.Sample(smpLinear,  uv)
+	#define SAMPLE_3D(T,uv)            T.Sample(smpLinear,    uv)
+	#define SAMPLE_CUBE(T,uv)          T.Sample(smpLinear,    uv)
 #else
 	sampler2D    sDiffuse     : register(s0);
 	sampler2D    sNormal      : register(s1);
@@ -69,13 +70,14 @@ struct VS_INPUT {
 	samplerCUBE  sEnvironment : register(s4);
 	sampler2D    sMask        : register(s5);
 	
-	#define SAMPLE_2D        (T,uv) tex2D(T, uv)
+	#define SAMPLE_2D(T,uv)             tex2D(T, uv)
-	#define SAMPLE_2D_POINT  (T,uv) tex2D(T, uv)
+	#define SAMPLE_2D_POINT(T,uv)       tex2D(T, uv)
-	#define SAMPLE_2D_LINEAR (T,uv) tex2D(T, uv)
+	#define SAMPLE_2D_POINT_WRAP(T,uv)  tex2D(T, uv)
-	#define SAMPLE_2D_ANISO  (T,uv) tex2D(T, uv)
+	#define SAMPLE_2D_LINEAR(T,uv)      tex2D(T, uv)
-	#define SAMPLE_2D_CMP    (T,uv) ((tex2Dlod(T, uv.xy) => uv.z) ? 1 : 0)
+	#define SAMPLE_2D_LOD0(T,uv)        tex2Dlod(T, float4(uv.xy, 0, 0))
-	#define SAMPLE_3D        (T,uv) tex3D(T, uv)
+	#define SAMPLE_2D_CMP(T,uv)         ((tex2D(T, uv.xy) => uv.z) ? 1 : 0)
-	#define SAMPLE_CUBE      (T,uv) texCUBE(T, uv)
+	#define SAMPLE_3D(T,uv)             tex3D(T, uv)
+	#define SAMPLE_CUBE(T,uv)           texCUBE(T, uv)
 #endif
 
 float4      uParam                  : register(  c0 );
@@ -143,9 +145,9 @@ float calcCausticsV(float3 coord) {
 	return 0.5 + abs(sin(dot(coord.xyz, 1.0 / 1024.0) + uParam.x)) * 0.75;
 }
 
-float3 calcNormalF(float2 tcR, float2 tcB, float base) {
+float3 calcHeightMapNormal(float2 tcR, float2 tcB, float base) {
-	float dx = SAMPLE_2D_LINEAR(sNormal, tcR).x - base;
+	float dx = SAMPLE_2D_LOD0(sNormal, tcR).x - base;
-	float dz = SAMPLE_2D_LINEAR(sNormal, tcB).x - base;
+	float dz = SAMPLE_2D_LOD0(sNormal, tcB).x - base;
 	return normalize( float3(dx, 64.0 / (1024.0 * 8.0), dz) );
 }
 
@@ -181,7 +183,7 @@ float SHADOW(float2 p) {
 	#elif SHADOW_DEPTH
 		return SAMPLE_2D_POINT(sShadow, float4(p, 0, 0)).x;
 	#else
-		return unpack(SAMPLE_2D_POINT(sShadow, float4(p, 0, 0)));
+		return unpack(SAMPLE_2D_POINT(sShadow, p));
 	#endif
 }
 

src/shaders/compose_entity.hlsl

@@ -43,7 +43,7 @@ VS_OUTPUT main(VS_INPUT In) {
 	if (OPT_AMBIENT) {
 		Out.ambient = calcAmbient(Out.normal.xyz);
 	} else {
-		Out.ambient = min(uMaterial.yyy, In.aLight);
+		Out.ambient = min(uMaterial.yyy, In.aLight.xyz);
 	}
 
 	float4 lum, att, light;
@@ -69,7 +69,7 @@ VS_OUTPUT main(VS_INPUT In) {
 		Out.light.xyz += Out.ambient + uLightColor[0].xyz * light.x;
 	}
 
-	Out.diffuse = float4(In.aColor * (uMaterial.x * 1.8), 1.0);
+	Out.diffuse = float4(In.aColor.xyz * (uMaterial.x * 1.8), 1.0);
 
 	Out.diffuse *= uMaterial.w;
 

src/shaders/water_caustics.hlsl

@@ -27,8 +27,8 @@ VS_OUTPUT main(VS_INPUT In) {
 	float3 rCoord = float3(coord.x, coord.y, 0.0) * uPosScale[1].xzy;
 
 	float2 uv     = getInvUV(rCoord.xy, uTexParam).xy;
-	float2 info   = SAMPLE_2D_LINEAR(sNormal, float4(uv, 0, 0)).xy;
+	float2 info   = SAMPLE_2D_LOD0(sNormal, uv).xy;
-	float3 normal = calcNormalV(uv, info.x).xzy;
+	float3 normal = calcHeightMapNormal(float2(uv.x + uTexParam.x, uv.y), float2(uv.x, uv.y - uTexParam.y), info.x).xzy;
 
 	float3 light  = float3(0.0, 0.0, 1.0);
 	float3 refOld = refract(-light, float3(0.0, 0.0, 1.0), 0.75);

src/shaders/water_compose.hlsl

@@ -47,7 +47,7 @@ half4 main(VS_OUTPUT In) : COLOR0 {
 	float3 viewVec = normalize(In.viewVec.xyz);
 	
 	float  base   = SAMPLE_2D_LINEAR(sNormal, In.texCoord).x;
-	float3 normal = calcNormalF(In.texCoordR, In.texCoordB, base);
+	float3 normal = calcHeightMapNormal(In.texCoordR, In.texCoordB, base);
 	
 	float2 dudv = mul(uViewProj, float4(normal.x, 0.0, normal.z, 0.0)).xy * uParam.z;
 	float3 rv = reflect(-viewVec, normal);

src/shaders/water_rays.hlsl

@@ -44,7 +44,7 @@ float4 main(VS_OUTPUT In/*, float2 pixelCoord: VPOS*/) : COLOR0 {
 	float3 p0 = uViewPos.xyz - viewVec * t;
 	float3 p1 = In.coord.xyz;
 
-	float dither = SAMPLE_2D_POINT(sMask, float4(pixelCoord * (1.0 / 8.0), 0.0, 0.0)).x;
+	float dither = SAMPLE_2D_POINT_WRAP(sMask, pixelCoord * (1.0 / 8.0)).x;
 	
 	float3 delta = (p1 - p0) / RAY_STEPS;
 	float3 pos	= p0 + delta * dither;
@@ -53,7 +53,7 @@ float4 main(VS_OUTPUT In/*, float2 pixelCoord: VPOS*/) : COLOR0 {
 	for (float i = 0.0; i < RAY_STEPS; i++) {
 		float3 wpos = (pos - uPosScale[0].xyz) / uPosScale[1].xyz;
 		float2 tc = wpos.xz * 0.5 + 0.5;
-		float light = SAMPLE_2D_LINEAR(sReflect, float4(tc, 0, 0)).x;
+		float light = SAMPLE_2D_LINEAR(sReflect, tc).x;
 		sum += light * (1.0 - (clamp(wpos.y, -1.0, 1.0) * 0.5 + 0.5));
 		pos += delta;
 	}