yet faster VSDCT instruction sequences

git-svn-id: svn://svn.icculus.org/twilight/trunk/darkplaces@9947 d7cf8633-e32d-0410-b094-e92efae38249

diff --git a/gl_rmain.c b/gl_rmain.c
index a5f3513f3a671cf574328fa2fa14614ba79eb738..2ccf7b1928a1079bd17cfe271157890207f61433 100644 (file)
--- a/gl_rmain.c
+++ b/gl_rmain.c
@@ -1064,11 +1064,9 @@ static const char *builtinshaderstring =
 "vec3 GetShadowMapTC2D(vec3 dir)\n"
 "{\n"
 "      vec3 adir = abs(dir);\n"
-"      float ma = max(max(adir.x, adir.y), adir.z);\n"
+"      vec2 aparams = ShadowMap_Parameters.xy / max(max(adir.x, adir.y), adir.z);\n"
 "      vec4 proj = textureCube(Texture_CubeProjection, dir);\n"
-"      vec3 stc = vec3(mix(dir.xy, dir.zz, proj.xy) * ShadowMap_Parameters.x, ShadowMap_Parameters.w) / ma;\n"
-"      stc += vec3(proj.zw * ShadowMap_Parameters.y, ShadowMap_Parameters.z);\n"
-"      return stc;\n"
+"      return vec3(mix(dir.xy, dir.zz, proj.xy) * aparams.x + proj.zw * ShadowMap_Parameters.z, aparams.y + ShadowMap_Parameters.w);\n"
 "}\n"
 "#  else\n"
 "vec3 GetShadowMapTC2D(vec3 dir)\n"
@@ -1078,9 +1076,8 @@ static const char *builtinshaderstring =
 "      vec4 proj = vec4(dir, 2.5);\n"
 "      if (adir.x > ma) { ma = adir.x; proj = vec4(dir.zyx, 0.5); }\n"
 "      if (adir.y > ma) { ma = adir.y; proj = vec4(dir.xzy, 1.5); }\n"
-"      vec3 stc = vec3(proj.xy * ShadowMap_Parameters.x, ShadowMap_Parameters.w) / ma;\n"
-"      stc += vec3(vec2(proj.z < 0.0 ? 1.5 : 0.5, proj.w) * ShadowMap_Parameters.y, ShadowMap_Parameters.z);\n"
-"      return stc;\n"
+"      vec2 aparams = ShadowMap_Parameters.xy / ma;\n"
+"      return vec3(proj.xy * aparams.x + vec2(proj.z < 0.0 ? 1.5 : 0.5, proj.w) * ShadowMap_Parameters.z, aparams.y + ShadowMap_Parameters.w);\n"
 "}\n"
 "#  endif\n"
 "# endif\n"
@@ -2190,11 +2187,9 @@ const char *builtincgshaderstring =
 "float3 GetShadowMapTC2D(float3 dir, float4 ShadowMap_Parameters, samplerCUBE Texture_CubeProjection)\n"
 "{\n"
 "      float3 adir = abs(dir);\n"
-"      float ma = max(max(adir.x, adir.y), adir.z);\n"
-"      float4 proj = texCUBE(Texture_CubeProjection, dir);\n"
-"      float3 stc = float3(lerp(dir.xy, dir.zz, proj.xy) * ShadowMap_Parameters.x, ShadowMap_Parameters.w) / ma;\n"
-"      stc += float3(proj.zw * ShadowMap_Parameters.y, ShadowMap_Parameters.z);\n"
-"      return stc;\n"
+"      float2 aparams = ShadowMap_Parameters.xy / max(max(adir.x, adir.y), adir.z);\n"
+"      float4 proj = texCUBEe(Texture_CubeProjection, dir);\n"
+"      return float3(mix(dir.xy, dir.zz, proj.xy) * aparams.x + proj.zw * ShadowMap_Parameters.z, aparams.y + ShadowMap_Parameters.w);\n"
 "}\n"
 "#  else\n"
 "float3 GetShadowMapTC2D(float3 dir, float4 ShadowMap_Parameters)\n"
@@ -2204,9 +2199,8 @@ const char *builtincgshaderstring =
 "      float4 proj = float4(dir, 2.5);\n"
 "      if (adir.x > ma) { ma = adir.x; proj = float4(dir.zyx, 0.5); }\n"
 "      if (adir.y > ma) { ma = adir.y; proj = float4(dir.xzy, 1.5); }\n"
-"      float3 stc = float3(proj.xy * ShadowMap_Parameters.x, ShadowMap_Parameters.w) / ma;\n"
-"      stc += float3(vec2(proj.z < 0.0 ? 1.5 : 0.5, proj.w) * ShadowMap_Parameters.y, ShadowMap_Parameters.z);\n"
-"      return stc;\n"
+"      float2 aparams = ShadowMap_Parameters.xy / ma;\n"
+"      return float3(proj.xy * aparams.x + float2(proj.z < 0.0 ? 1.5 : 0.5, proj.w) * ShadowMap_Parameters.z, aparams.y + ShadowMap_Parameters.w);\n"
 "}\n"
 "#  endif\n"
 "# endif\n"

diff --git a/r_shadow.c b/r_shadow.c
index 1a426936b1a112f208b9e9dca4c04c7a04ac16d4..6410fec5666ad29640eec452300d4053b88927d3 100644 (file)
--- a/r_shadow.c
+++ b/r_shadow.c
@@ -2108,15 +2108,15 @@ void R_Shadow_RenderMode_ShadowMap(int side, int clear, int size)
        nearclip = r_shadow_shadowmapping_nearclip.value / rsurface.rtlight->radius;
        farclip = 1.0f;
        bias = r_shadow_shadowmapping_bias.value * nearclip * (1024.0f / size);// * rsurface.rtlight->radius;
-       r_shadow_shadowmap_parameters[2] = 0.5f + 0.5f * (farclip + nearclip) / (farclip - nearclip);
-       r_shadow_shadowmap_parameters[3] = -nearclip * farclip / (farclip - nearclip) - 0.5f * bias;
+       r_shadow_shadowmap_parameters[1] = -nearclip * farclip / (farclip - nearclip) - 0.5f * bias;
+       r_shadow_shadowmap_parameters[3] = 0.5f + 0.5f * (farclip + nearclip) / (farclip - nearclip);
        r_shadow_shadowmapside = side;
        r_shadow_shadowmapsize = size;
        switch (r_shadow_shadowmode)
        {
        case R_SHADOW_SHADOWMODE_SHADOWMAP2D:
                r_shadow_shadowmap_parameters[0] = 0.5f * (size - r_shadow_shadowmapborder);
-               r_shadow_shadowmap_parameters[1] = r_shadow_shadowmapvsdct ? 2.5f*size : size;
+               r_shadow_shadowmap_parameters[2] = r_shadow_shadowmapvsdct ? 2.5f*size : size;
                R_Viewport_InitRectSideView(&viewport, &rsurface.rtlight->matrix_lighttoworld, side, size, r_shadow_shadowmapborder, nearclip, farclip, NULL);
                if (r_shadow_rendermode == R_SHADOW_RENDERMODE_SHADOWMAP2D) goto init_done;
 
@@ -2133,7 +2133,7 @@ void R_Shadow_RenderMode_ShadowMap(int side, int clear, int size)
                break;
        case R_SHADOW_SHADOWMODE_SHADOWMAPRECTANGLE:
                r_shadow_shadowmap_parameters[0] = 0.5f * (size - r_shadow_shadowmapborder);
-               r_shadow_shadowmap_parameters[1] = r_shadow_shadowmapvsdct ? 2.5f*size : size;
+               r_shadow_shadowmap_parameters[2] = r_shadow_shadowmapvsdct ? 2.5f*size : size;
                R_Viewport_InitRectSideView(&viewport, &rsurface.rtlight->matrix_lighttoworld, side, size, r_shadow_shadowmapborder, nearclip, farclip, NULL);
                if (r_shadow_rendermode == R_SHADOW_RENDERMODE_SHADOWMAPRECTANGLE) goto init_done;
 
@@ -2150,7 +2150,7 @@ void R_Shadow_RenderMode_ShadowMap(int side, int clear, int size)
                break;
        case R_SHADOW_SHADOWMODE_SHADOWMAPCUBESIDE:
                r_shadow_shadowmap_parameters[0] = 1.0f;
-               r_shadow_shadowmap_parameters[1] = 1.0f;
+               r_shadow_shadowmap_parameters[2] = 1.0f;
                R_Viewport_InitCubeSideView(&viewport, &rsurface.rtlight->matrix_lighttoworld, side, size, nearclip, farclip, NULL);
                if (r_shadow_rendermode == R_SHADOW_RENDERMODE_SHADOWMAPCUBESIDE) goto init_done;