GPU: Clean up float conversion functions

diff --git a/libctru/source/gpu/gpu.c b/libctru/source/gpu/gpu.c
index 04294443942b1a096e165361748eb8b8e489b87e..723b65b3fa746f2162afcdd518bba357cbf794cb 100644 (file)
--- a/libctru/source/gpu/gpu.c
+++ b/libctru/source/gpu/gpu.c
@@ -229,40 +229,65 @@ void GPU_SetFloatUniform(GPU_SHADER_TYPE type, u32 startreg, u32* data, u32 numr
        GPUCMD_AddWrites(GPUREG_VSH_FLOATUNIFORM_DATA+regOffset, data, numreg*4);
 }
 
-//TODO : fix
-u32 f32tof24(float f)
-{
-       if(!f)return 0;
-       u32 v=*((u32*)&f);
-       u8 s=v>>31;
-       u32 exp=((v>>23)&0xFF)-0x40;
-       u32 man=(v>>7)&0xFFFF;
-
-       if(exp>=0)return man|(exp<<16)|(s<<23);
-       else return s<<23;
-}
-
-u32 computeInvValue(u32 val)
-{
-       //usual values
-       if(val==240)return 0x38111111;
-       if(val==480)return 0x37111111;
-       if(val==400)return 0x3747ae14;
-       //but let's not limit ourselves to the usual
-       float fval=2.0/val;
-       u32 tmp1,tmp2;
-       u32 tmp3=*((u32*)&fval);
-       tmp1=(tmp3<<9)>>9;
-       tmp2=tmp3&(~0x80000000);
-       if(tmp2)
+// f24 has:
+//  - 1 sign bit
+//  - 7 exponent bits
+//  - 16 mantissa bits
+static u32 f32tof24(float f)
+{
+       u32 i;
+       memcpy(&i, &f, 4);
+
+       u32 mantissa = (i << 9) >>  9;
+       s32 exponent = (i << 1) >> 24;
+       u32 sign     = (i << 0) >> 31;
+
+       // Truncate mantissa
+       mantissa >>= 7;
+
+       // Re-bias exponent
+       exponent = exponent - 127 + 63;
+       if (exponent < 0)
        {
-               tmp1=(tmp3<<9)>>9;
-               int tmp=((tmp3<<1)>>24)-0x40;
-               if(tmp<0)return ((tmp3>>31)<<30)<<1;
-               else tmp2=tmp;
+               // Underflow: flush to zero
+               return sign << 23;
        }
-       tmp3>>=31;
-       return (tmp1|(tmp2<<23)|(tmp3<<30))<<1;
+       else if (exponent > 0x7F)
+       {
+               // Overflow: saturate to infinity
+               return sign << 23 | 0x7F << 16;
+       }
+
+       return sign << 23 | exponent << 16 | mantissa;
+}
+
+// f31 has:
+//  - 1 sign bit
+//  - 7 exponent bits
+//  - 23 mantissa bits
+static u32 f32tof31(float f)
+{
+       u32 i;
+       memcpy(&i, &f, 4);
+
+       u32 mantissa = (i << 9) >>  9;
+       s32 exponent = (i << 1) >> 24;
+       u32 sign     = (i << 0) >> 31;
+
+       // Re-bias exponent
+       exponent = exponent - 127 + 63;
+       if (exponent < 0)
+       {
+               // Underflow: flush to zero
+               return sign << 30;
+       }
+       else if (exponent > 0x7F)
+       {
+               // Overflow: saturate to infinity
+               return sign << 30 | 0x7F << 23;
+       }
+
+       return sign << 30 | exponent << 23 | mantissa;
 }
 
 //takes PAs as arguments
@@ -288,9 +313,9 @@ void GPU_SetViewport(u32* depthBuffer, u32* colorBuffer, u32 x, u32 y, u32 w, u3
        GPUCMD_AddWrite(GPUREG_011B, 0x00000000); //?
 
        param[0x0]=f32tof24(fw/2);
-       param[0x1]=computeInvValue(fw);
+       param[0x1]=f32tof31(2.0f / fw) << 1;
        param[0x2]=f32tof24(fh/2);
-       param[0x3]=computeInvValue(fh);
+       param[0x3]=f32tof31(2.0f / fh) << 1;
        GPUCMD_AddIncrementalWrites(GPUREG_0041, param, 0x00000004);
 
        GPUCMD_AddWrite(GPUREG_0068, (y<<16)|(x&0xFFFF));