notaz.gp2x.de / mupen64plus-pandora.git / blobdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | inline | side by side
rice: optimize some uv calculations
[mupen64plus-pandora.git] / source / gles2rice / src / RenderBase.cpp
diff --git a/source/gles2rice/src/RenderBase.cpp b/source/gles2rice/src/RenderBase.cpp
index ee5b76c..1ba7ec8 100644 (file)
--- a/source/gles2rice/src/RenderBase.cpp
+++ b/source/gles2rice/src/RenderBase.cpp
@@ -854,6 +854,17 @@ void ComputeLOD(bool openGL)
 bool bHalfTxtScale=false;
 extern uint32 lastSetTile;
 
+#ifndef __ARM_NEON__
+static void multiply_subtract2(float *d, const float *m1, const float *m2, const float *s)
+{
+    int i;
+    for (i = 0; i < 2; i++)
+        d[i] = m1[i] * m2[i] - s[i];
+}
+#else
+extern "C" void multiply_subtract2(float *d, const float *m1, const float *m2, const float *s);
+#endif
+
 void InitVertex(uint32 dwV, uint32 vtxIndex, bool bTexture, bool openGL)
 {
     VTX_DUMP(TRACE2("Init vertex (%d) to vtx buf[%d]:", dwV, vtxIndex));
@@ -959,22 +970,22 @@ void InitVertex(uint32 dwV, uint32 vtxIndex, bool bTexture, bool openGL)
         }
         else
         {
-            float tex0u = g_fVtxTxtCoords[dwV].x *gRSP.tex0scaleX - gRSP.tex0OffsetX ;
-            float tex0v = g_fVtxTxtCoords[dwV].y *gRSP.tex0scaleY - gRSP.tex0OffsetY ;
+            TexCord tex0;
+            multiply_subtract2(&tex0.u, &g_fVtxTxtCoords[dwV].x, &gRSP.tex0scaleX, &gRSP.tex0OffsetX);
 
             if( CRender::g_pRender->IsTexel1Enable() )
             {
-                float tex1u = g_fVtxTxtCoords[dwV].x *gRSP.tex1scaleX - gRSP.tex1OffsetX ;
-                float tex1v = g_fVtxTxtCoords[dwV].y *gRSP.tex1scaleY - gRSP.tex1OffsetY ;
+                TexCord tex1;
+                multiply_subtract2(&tex1.u, &g_fVtxTxtCoords[dwV].x, &gRSP.tex1scaleX, &gRSP.tex1OffsetX);
 
-                CRender::g_pRender->SetVertexTextureUVCoord(v, tex0u, tex0v, tex1u, tex1v);
-                VTX_DUMP(TRACE2("  (tex0): %f, %f",  tex0u,tex0v));
-                VTX_DUMP(TRACE2("  (tex1): %f, %f",  tex1u,tex1v));
+                CRender::g_pRender->SetVertexTextureUVCoord(v, tex0, tex1);
+                VTX_DUMP(TRACE2("  (tex0): %f, %f",  tex0.u,tex0.v));
+                VTX_DUMP(TRACE2("  (tex1): %f, %f",  tex1.u,tex1.v));
             }
             else
             {
-                CRender::g_pRender->SetVertexTextureUVCoord(v, tex0u, tex0v);
-                VTX_DUMP(TRACE2("  (tex0): %f, %f",  tex0u,tex0v));
+                CRender::g_pRender->SetVertexTextureUVCoord(v, tex0);
+                VTX_DUMP(TRACE2("  (tex0): %f, %f",  tex0.u,tex0.v));
             }
         }
 
@@ -1528,13 +1539,19 @@ void ProcessVertexDataNoSSE(uint32 dwAddr, uint32 dwV0, uint32 dwNum)
     DEBUGGER_PAUSE_AND_DUMP(NEXT_VERTEX_CMD,{TRACE0("Paused at Vertex Cmd");});
 }
 
+/* NEON code */
+
+#include "RenderBase_neon.h"
+
 extern "C" void pv_neon(XVECTOR4 *g_vtxTransformed, XVECTOR4 *g_vecProjected,
     uint32 *g_dwVtxDifColor, VECTOR2 *g_fVtxTxtCoords,
     float *g_fFogCoord, uint32 *g_clipFlag2,
-    uint32 dwNum, const FiddledVtx *vtx,
+    uint32 dwNum, int neon_state,
+    const FiddledVtx *vtx,
     const Light *gRSPlights, const float *fRSPAmbientLightRGBA,
     const XMATRIX *gRSPworldProject, const XMATRIX *gRSPmodelViewTop,
-    uint32 gRSPnumLights, float gRSPfFogMin);
+    uint32 gRSPnumLights, float gRSPfFogMin,
+    uint32 primitiveColor, uint32 primitiveColor_);
 
 void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum)
 {
@@ -1545,14 +1562,10 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum)
 
     // assumtions:
     // - g_clipFlag is not used at all
+    // - g_fFogCoord is not used at all
     // - g_vtxNonTransformed is not used after ProcessVertexData*() returns
     // - g_normal - same
 
-#define PV_NEON_ENABLE_LIGHT    (1 << 0)
-#define PV_NEON_ENABLE_SHADE    (1 << 1)
-#define PV_NEON_ENABLE_FOG      (1 << 2)
-#define PV_NEON_FOG_ALPHA       (1 << 3)
-
     int neon_state = 0;
     if ( gRSP.bLightingEnable )
         neon_state |= PV_NEON_ENABLE_LIGHT;
@@ -1578,20 +1591,29 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum)
     //          - g_vtxTransformed[i]
     //          - g_dwVtxDifColor[i]            -> vertex color
     //          - g_fVtxTxtCoords[i]            -> vertex texture cooridinates
-    //          - g_fFogCoord[i]
+    //          - g_fFogCoord[i]                -> unused
     //          - g_clipFlag2[i]
 
     const FiddledVtx * pVtxBase = (const FiddledVtx*)(g_pRDRAMu8 + dwAddr);
     g_pVtxBase = (FiddledVtx *)pVtxBase;
 
+    gRSPmodelViewTop._14 = gRSPmodelViewTop._24 =
+    gRSPmodelViewTop._34 = 0;
+
     // SP_Timing(RSP_GBI0_Vtx);
     status.SPCycleCount += Timing_RSP_GBI0_Vtx * dwNum;
 
-    if (!(neon_state & (PV_NEON_ENABLE_LIGHT | PV_NEON_ENABLE_SHADE))) {
-        for (i = dwV0; i < dwV0 + dwNum; i++)
-            g_dwVtxDifColor[i] = gRDP.primitiveColor; // FLAT shade
-    }
-
+#if 1
+    i = dwV0;
+    pv_neon(&g_vtxTransformed[i], &g_vecProjected[i],
+            &g_dwVtxDifColor[i], &g_fVtxTxtCoords[i],
+            &g_fFogCoord[i], &g_clipFlag2[i],
+            dwNum, neon_state, &pVtxBase[i - dwV0],
+            gRSPlights, gRSP.fAmbientColors,
+            &gRSPworldProject, &gRSPmodelViewTop,
+            gRSPnumLights, gRSPfFogMin,
+            gRDP.primitiveColor, gRDP.primitiveColor);
+#else
     for (i = dwV0; i < dwV0 + dwNum; i++)
     {
         const FiddledVtx & vert = pVtxBase[i - dwV0];
@@ -1608,13 +1630,6 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum)
         g_vecProjected[i].y = g_vtxTransformed[i].y * g_vecProjected[i].w;
         g_vecProjected[i].z = g_vtxTransformed[i].z * g_vecProjected[i].w;
 
-        if( neon_state & PV_NEON_ENABLE_FOG )
-        {
-            g_fFogCoord[i] = g_vecProjected[i].z;
-            if( g_vecProjected[i].w < 0 || g_vecProjected[i].z < 0 || g_fFogCoord[i] < gRSPfFogMin )
-                g_fFogCoord[i] = gRSPfFogMin;
-        }
-
         // RSP_Vtx_Clipping(i);
         g_clipFlag2[i] = 0;
         if( g_vecProjected[i].w > 0 )
@@ -1664,6 +1679,8 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum)
             color.r = vert.rgba.b;
             color.a = vert.rgba.a;
         }
+        else
+            g_dwVtxDifColor[i] = gRDP.primitiveColor; // FLAT shade
 
         // ReplaceAlphaWithFogFactor(i);
         if( neon_state & PV_NEON_FOG_ALPHA )
@@ -1671,7 +1688,8 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum)
             // Use fog factor to replace vertex alpha
             if( g_vecProjected[i].z > 1 )
                 *(((uint8*)&(g_dwVtxDifColor[i]))+3) = 0xFF;
-            if( g_vecProjected[i].z < 0 )
+            // missing 'else' in original code??
+            else if( g_vecProjected[i].z < 0 )
                 *(((uint8*)&(g_dwVtxDifColor[i]))+3) = 0;
             else
                 *(((uint8*)&(g_dwVtxDifColor[i]))+3) = (uint8)(g_vecProjected[i].z*255);    
@@ -1680,6 +1698,7 @@ void ProcessVertexDataNEON(uint32 dwAddr, uint32 dwV0, uint32 dwNum)
         g_fVtxTxtCoords[i].x = (float)vert.tu;
         g_fVtxTxtCoords[i].y = (float)vert.tv; 
     }
+#endif
 }
 
 bool PrepareTriangle(uint32 dwV0, uint32 dwV1, uint32 dwV2)
Unnamed repository; edit this file 'description' to name the repository.