dlls/wined3d/pixelshader.c

   1 /*
   2  * shaders implementation
   3  *
   4  * Copyright 2002-2003 Jason Edmeades
   5  * Copyright 2002-2003 Raphael Junqueira
   6  * Copyright 2004 Christian Costa
   7  * Copyright 2005 Oliver Stieber
   8  * Copyright 2006 Ivan Gyurdiev
   9  * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
  10  *
  11  * This library is free software; you can redistribute it and/or
  12  * modify it under the terms of the GNU Lesser General Public
  13  * License as published by the Free Software Foundation; either
  14  * version 2.1 of the License, or (at your option) any later version.
  15  *
  16  * This library is distributed in the hope that it will be useful,
  17  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * Lesser General Public License for more details.
  20  *
  21  * You should have received a copy of the GNU Lesser General Public
  22  * License along with this library; if not, write to the Free Software
  23  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  24  */
  25
  26 #include "config.h"
  27
  28 #include <math.h>
  29 #include <stdio.h>
  30
  31 #include "wined3d_private.h"
  32
  33 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  34
  35 #define GLINFO_LOCATION ((IWineD3DDeviceImpl *) This->baseShader.device)->adapter->gl_info
  36
  37 #define GLNAME_REQUIRE_GLSL  ((const char *)1)
  38
  39 static HRESULT  WINAPI IWineD3DPixelShaderImpl_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, LPVOID *ppobj) {
  40     TRACE("iface %p, riid %s, ppobj %p\n", iface, debugstr_guid(riid), ppobj);
  41
  42     if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
  43             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
  44             || IsEqualGUID(riid, &IID_IWineD3DBase)
  45             || IsEqualGUID(riid, &IID_IUnknown))
  46     {
  47         IUnknown_AddRef(iface);
  48         *ppobj = iface;
  49         return S_OK;
  50     }
  51
  52     WARN("%s not implemented, returning E_NOINTERFACE\n", debugstr_guid(riid));
  53
  54     *ppobj = NULL;
  55     return E_NOINTERFACE;
  56 }
  57
  58 static ULONG  WINAPI IWineD3DPixelShaderImpl_AddRef(IWineD3DPixelShader *iface) {
  59     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
  60     ULONG refcount = InterlockedIncrement(&This->baseShader.ref);
  61
  62     TRACE("%p increasing refcount to %u\n", This, refcount);
  63
  64     return refcount;
  65 }
  66
  67 static ULONG  WINAPI IWineD3DPixelShaderImpl_Release(IWineD3DPixelShader *iface) {
  68     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
  69     ULONG refcount = InterlockedDecrement(&This->baseShader.ref);
  70
  71     TRACE("%p decreasing refcount to %u\n", This, refcount);
  72
  73     if (!refcount)
  74     {
  75         shader_cleanup((IWineD3DBaseShader *)iface);
  76         HeapFree(GetProcessHeap(), 0, This);
  77     }
  78
  79     return refcount;
  80 }
  81
  82 /* *******************************************
  83    IWineD3DPixelShader IWineD3DPixelShader parts follow
  84    ******************************************* */
  85
  86 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetParent(IWineD3DPixelShader *iface, IUnknown** parent){
  87     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
  88
  89     *parent = This->parent;
  90     IUnknown_AddRef(*parent);
  91     TRACE("(%p) : returning %p\n", This, *parent);
  92     return WINED3D_OK;
  93 }
  94
  95 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetDevice(IWineD3DPixelShader* iface, IWineD3DDevice **pDevice){
  96     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
  97     IWineD3DDevice_AddRef(This->baseShader.device);
  98     *pDevice = This->baseShader.device;
  99     TRACE("(%p) returning %p\n", This, *pDevice);
 100     return WINED3D_OK;
 101 }
 102
 103
 104 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetFunction(IWineD3DPixelShader* impl, VOID* pData, UINT* pSizeOfData) {
 105   IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)impl;
 106   TRACE("(%p) : pData(%p), pSizeOfData(%p)\n", This, pData, pSizeOfData);
 107
 108   if (NULL == pData) {
 109     *pSizeOfData = This->baseShader.functionLength;
 110     return WINED3D_OK;
 111   }
 112   if (*pSizeOfData < This->baseShader.functionLength) {
 113     /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
 114      * than the required size we should write the required size and
 115      * return D3DERR_MOREDATA. That's not actually true. */
 116     return WINED3DERR_INVALIDCALL;
 117   }
 118
 119   TRACE("(%p) : GetFunction copying to %p\n", This, pData);
 120   memcpy(pData, This->baseShader.function, This->baseShader.functionLength);
 121
 122   return WINED3D_OK;
 123 }
 124
 125 CONST SHADER_OPCODE IWineD3DPixelShaderImpl_shader_ins[] = {
 126     /* Arithmetic */
 127     {WINED3DSIO_NOP,          "nop",          0, 0, WINED3DSIH_NOP,          0,                      0                     },
 128     {WINED3DSIO_MOV,          "mov",          1, 2, WINED3DSIH_MOV,          0,                      0                     },
 129     {WINED3DSIO_ADD,          "add",          1, 3, WINED3DSIH_ADD,          0,                      0                     },
 130     {WINED3DSIO_SUB,          "sub",          1, 3, WINED3DSIH_SUB,          0,                      0                     },
 131     {WINED3DSIO_MAD,          "mad",          1, 4, WINED3DSIH_MAD,          0,                      0                     },
 132     {WINED3DSIO_MUL,          "mul",          1, 3, WINED3DSIH_MUL,          0,                      0                     },
 133     {WINED3DSIO_RCP,          "rcp",          1, 2, WINED3DSIH_RCP,          0,                      0                     },
 134     {WINED3DSIO_RSQ,          "rsq",          1, 2, WINED3DSIH_RSQ,          0,                      0                     },
 135     {WINED3DSIO_DP3,          "dp3",          1, 3, WINED3DSIH_DP3,          0,                      0                     },
 136     {WINED3DSIO_DP4,          "dp4",          1, 3, WINED3DSIH_DP4,          0,                      0                     },
 137     {WINED3DSIO_MIN,          "min",          1, 3, WINED3DSIH_MIN,          0,                      0                     },
 138     {WINED3DSIO_MAX,          "max",          1, 3, WINED3DSIH_MAX,          0,                      0                     },
 139     {WINED3DSIO_SLT,          "slt",          1, 3, WINED3DSIH_SLT,          0,                      0                     },
 140     {WINED3DSIO_SGE,          "sge",          1, 3, WINED3DSIH_SGE,          0,                      0                     },
 141     {WINED3DSIO_ABS,          "abs",          1, 2, WINED3DSIH_ABS,          0,                      0                     },
 142     {WINED3DSIO_EXP,          "exp",          1, 2, WINED3DSIH_EXP,          0,                      0                     },
 143     {WINED3DSIO_LOG,          "log",          1, 2, WINED3DSIH_LOG,          0,                      0                     },
 144     {WINED3DSIO_EXPP,         "expp",         1, 2, WINED3DSIH_EXPP,         0,                      0                     },
 145     {WINED3DSIO_LOGP,         "logp",         1, 2, WINED3DSIH_LOGP,         0,                      0                     },
 146     {WINED3DSIO_DST,          "dst",          1, 3, WINED3DSIH_DST,          0,                      0                     },
 147     {WINED3DSIO_LRP,          "lrp",          1, 4, WINED3DSIH_LRP,          0,                      0                     },
 148     {WINED3DSIO_FRC,          "frc",          1, 2, WINED3DSIH_FRC,          0,                      0                     },
 149     {WINED3DSIO_CND,          "cnd",          1, 4, WINED3DSIH_CND,          WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,4)},
 150     {WINED3DSIO_CMP,          "cmp",          1, 4, WINED3DSIH_CMP,          WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(3,0)},
 151     {WINED3DSIO_POW,          "pow",          1, 3, WINED3DSIH_POW,          0,                      0                     },
 152     {WINED3DSIO_CRS,          "crs",          1, 3, WINED3DSIH_CRS,          0,                      0                     },
 153     {WINED3DSIO_NRM,          "nrm",          1, 2, WINED3DSIH_NRM,          0,                      0                     },
 154     {WINED3DSIO_SINCOS,       "sincos",       1, 4, WINED3DSIH_SINCOS,       WINED3DPS_VERSION(2,0), WINED3DPS_VERSION(2,1)},
 155     {WINED3DSIO_SINCOS,       "sincos",       1, 2, WINED3DSIH_SINCOS,       WINED3DPS_VERSION(3,0), -1                    },
 156     {WINED3DSIO_DP2ADD,       "dp2add",       1, 4, WINED3DSIH_DP2ADD,       WINED3DPS_VERSION(2,0), -1                    },
 157     /* Matrix */
 158     {WINED3DSIO_M4x4,         "m4x4",         1, 3, WINED3DSIH_M4x4,         0,                      0                     },
 159     {WINED3DSIO_M4x3,         "m4x3",         1, 3, WINED3DSIH_M4x3,         0,                      0                     },
 160     {WINED3DSIO_M3x4,         "m3x4",         1, 3, WINED3DSIH_M3x4,         0,                      0                     },
 161     {WINED3DSIO_M3x3,         "m3x3",         1, 3, WINED3DSIH_M3x3,         0,                      0                     },
 162     {WINED3DSIO_M3x2,         "m3x2",         1, 3, WINED3DSIH_M3x2,         0,                      0                     },
 163     /* Register declarations */
 164     {WINED3DSIO_DCL,          "dcl",          0, 2, WINED3DSIH_DCL,          0,                      0                     },
 165     /* Flow control - requires GLSL or software shaders */
 166     {WINED3DSIO_REP ,         "rep",          0, 1, WINED3DSIH_REP,          WINED3DPS_VERSION(2,1), -1                    },
 167     {WINED3DSIO_ENDREP,       "endrep",       0, 0, WINED3DSIH_ENDREP,       WINED3DPS_VERSION(2,1), -1                    },
 168     {WINED3DSIO_IF,           "if",           0, 1, WINED3DSIH_IF,           WINED3DPS_VERSION(2,1), -1                    },
 169     {WINED3DSIO_IFC,          "ifc",          0, 2, WINED3DSIH_IFC,          WINED3DPS_VERSION(2,1), -1                    },
 170     {WINED3DSIO_ELSE,         "else",         0, 0, WINED3DSIH_ELSE,         WINED3DPS_VERSION(2,1), -1                    },
 171     {WINED3DSIO_ENDIF,        "endif",        0, 0, WINED3DSIH_ENDIF,        WINED3DPS_VERSION(2,1), -1                    },
 172     {WINED3DSIO_BREAK,        "break",        0, 0, WINED3DSIH_BREAK,        WINED3DPS_VERSION(2,1), -1                    },
 173     {WINED3DSIO_BREAKC,       "breakc",       0, 2, WINED3DSIH_BREAKC,       WINED3DPS_VERSION(2,1), -1                    },
 174     {WINED3DSIO_BREAKP,       "breakp",       0, 1, WINED3DSIH_BREAKP,       0,                      0                     },
 175     {WINED3DSIO_CALL,         "call",         0, 1, WINED3DSIH_CALL,         WINED3DPS_VERSION(2,1), -1                    },
 176     {WINED3DSIO_CALLNZ,       "callnz",       0, 2, WINED3DSIH_CALLNZ,       WINED3DPS_VERSION(2,1), -1                    },
 177     {WINED3DSIO_LOOP,         "loop",         0, 2, WINED3DSIH_LOOP,         WINED3DPS_VERSION(3,0), -1                    },
 178     {WINED3DSIO_RET,          "ret",          0, 0, WINED3DSIH_RET,          WINED3DPS_VERSION(2,1), -1                    },
 179     {WINED3DSIO_ENDLOOP,      "endloop",      0, 0, WINED3DSIH_ENDLOOP,      WINED3DPS_VERSION(3,0), -1                    },
 180     {WINED3DSIO_LABEL,        "label",        0, 1, WINED3DSIH_LABEL,        WINED3DPS_VERSION(2,1), -1                    },
 181     /* Constant definitions */
 182     {WINED3DSIO_DEF,          "def",          1, 5, WINED3DSIH_DEF,          0,                      0                     },
 183     {WINED3DSIO_DEFB,         "defb",         1, 2, WINED3DSIH_DEFB,         0,                      0                     },
 184     {WINED3DSIO_DEFI,         "defi",         1, 5, WINED3DSIH_DEFI,         0,                      0                     },
 185     /* Texture */
 186     {WINED3DSIO_TEXCOORD,     "texcoord",     1, 1, WINED3DSIH_TEXCOORD,     0,                      WINED3DPS_VERSION(1,3)},
 187     {WINED3DSIO_TEXCOORD,     "texcrd",       1, 2, WINED3DSIH_TEXCOORD,     WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 188     {WINED3DSIO_TEXKILL,      "texkill",      1, 1, WINED3DSIH_TEXKILL,      WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(3,0)},
 189     {WINED3DSIO_TEX,          "tex",          1, 1, WINED3DSIH_TEX,          0,                      WINED3DPS_VERSION(1,3)},
 190     {WINED3DSIO_TEX,          "texld",        1, 2, WINED3DSIH_TEX,          WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 191     {WINED3DSIO_TEX,          "texld",        1, 3, WINED3DSIH_TEX,          WINED3DPS_VERSION(2,0), -1                    },
 192     {WINED3DSIO_TEXBEM,       "texbem",       1, 2, WINED3DSIH_TEXBEM,       0,                      WINED3DPS_VERSION(1,3)},
 193     {WINED3DSIO_TEXBEML,      "texbeml",      1, 2, WINED3DSIH_TEXBEML,      WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 194     {WINED3DSIO_TEXREG2AR,    "texreg2ar",    1, 2, WINED3DSIH_TEXREG2AR,    WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 195     {WINED3DSIO_TEXREG2GB,    "texreg2gb",    1, 2, WINED3DSIH_TEXREG2GB,    WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 196     {WINED3DSIO_TEXREG2RGB,   "texreg2rgb",   1, 2, WINED3DSIH_TEXREG2RGB,   WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 197     {WINED3DSIO_TEXM3x2PAD,   "texm3x2pad",   1, 2, WINED3DSIH_TEXM3x2PAD,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 198     {WINED3DSIO_TEXM3x2TEX,   "texm3x2tex",   1, 2, WINED3DSIH_TEXM3x2TEX,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 199     {WINED3DSIO_TEXM3x3PAD,   "texm3x3pad",   1, 2, WINED3DSIH_TEXM3x3PAD,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 200     {WINED3DSIO_TEXM3x3DIFF,  "texm3x3diff",  1, 2, WINED3DSIH_TEXM3x3DIFF,  WINED3DPS_VERSION(0,0), WINED3DPS_VERSION(0,0)},
 201     {WINED3DSIO_TEXM3x3SPEC,  "texm3x3spec",  1, 3, WINED3DSIH_TEXM3x3SPEC,  WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 202     {WINED3DSIO_TEXM3x3VSPEC, "texm3x3vspec", 1, 2, WINED3DSIH_TEXM3x3VSPEC, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 203     {WINED3DSIO_TEXM3x3TEX,   "texm3x3tex",   1, 2, WINED3DSIH_TEXM3x3TEX,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 204     {WINED3DSIO_TEXDP3TEX,    "texdp3tex",    1, 2, WINED3DSIH_TEXDP3TEX,    WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 205     {WINED3DSIO_TEXM3x2DEPTH, "texm3x2depth", 1, 2, WINED3DSIH_TEXM3x2DEPTH, WINED3DPS_VERSION(1,3), WINED3DPS_VERSION(1,3)},
 206     {WINED3DSIO_TEXDP3,       "texdp3",       1, 2, WINED3DSIH_TEXDP3,       WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 207     {WINED3DSIO_TEXM3x3,      "texm3x3",      1, 2, WINED3DSIH_TEXM3x3,      WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 208     {WINED3DSIO_TEXDEPTH,     "texdepth",     1, 1, WINED3DSIH_TEXDEPTH,     WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 209     {WINED3DSIO_BEM,          "bem",          1, 3, WINED3DSIH_BEM,          WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 210     {WINED3DSIO_DSX,          "dsx",          1, 2, WINED3DSIH_DSX,          WINED3DPS_VERSION(2,1), -1                    },
 211     {WINED3DSIO_DSY,          "dsy",          1, 2, WINED3DSIH_DSY,          WINED3DPS_VERSION(2,1), -1                    },
 212     {WINED3DSIO_TEXLDD,       "texldd",       1, 5, WINED3DSIH_TEXLDD,       WINED3DPS_VERSION(2,1), -1                    },
 213     {WINED3DSIO_SETP,         "setp",         1, 3, WINED3DSIH_SETP,         0,                      0                     },
 214     {WINED3DSIO_TEXLDL,       "texldl",       1, 3, WINED3DSIH_TEXLDL,       WINED3DPS_VERSION(3,0), -1                    },
 215     {WINED3DSIO_PHASE,        "phase",        0, 0, WINED3DSIH_PHASE,        0,                      0                     },
 216     {0,                       NULL,           0, 0, 0,                       0,                      0                     }
 217 };
 218
 219 static void pshader_set_limits(
 220       IWineD3DPixelShaderImpl *This) {
 221
 222       This->baseShader.limits.attributes = 0;
 223       This->baseShader.limits.address = 0;
 224       This->baseShader.limits.packed_output = 0;
 225
 226       switch (This->baseShader.reg_maps.shader_version)
 227       {
 228           case WINED3DPS_VERSION(1,0):
 229           case WINED3DPS_VERSION(1,1):
 230           case WINED3DPS_VERSION(1,2):
 231           case WINED3DPS_VERSION(1,3):
 232                    This->baseShader.limits.temporary = 2;
 233                    This->baseShader.limits.constant_float = 8;
 234                    This->baseShader.limits.constant_int = 0;
 235                    This->baseShader.limits.constant_bool = 0;
 236                    This->baseShader.limits.texcoord = 4;
 237                    This->baseShader.limits.sampler = 4;
 238                    This->baseShader.limits.packed_input = 0;
 239                    This->baseShader.limits.label = 0;
 240                    break;
 241
 242           case WINED3DPS_VERSION(1,4):
 243                    This->baseShader.limits.temporary = 6;
 244                    This->baseShader.limits.constant_float = 8;
 245                    This->baseShader.limits.constant_int = 0;
 246                    This->baseShader.limits.constant_bool = 0;
 247                    This->baseShader.limits.texcoord = 6;
 248                    This->baseShader.limits.sampler = 6;
 249                    This->baseShader.limits.packed_input = 0;
 250                    This->baseShader.limits.label = 0;
 251                    break;
 252
 253           /* FIXME: temporaries must match D3DPSHADERCAPS2_0.NumTemps */
 254           case WINED3DPS_VERSION(2,0):
 255                    This->baseShader.limits.temporary = 32;
 256                    This->baseShader.limits.constant_float = 32;
 257                    This->baseShader.limits.constant_int = 16;
 258                    This->baseShader.limits.constant_bool = 16;
 259                    This->baseShader.limits.texcoord = 8;
 260                    This->baseShader.limits.sampler = 16;
 261                    This->baseShader.limits.packed_input = 0;
 262                    break;
 263
 264           case WINED3DPS_VERSION(2,1):
 265                    This->baseShader.limits.temporary = 32;
 266                    This->baseShader.limits.constant_float = 32;
 267                    This->baseShader.limits.constant_int = 16;
 268                    This->baseShader.limits.constant_bool = 16;
 269                    This->baseShader.limits.texcoord = 8;
 270                    This->baseShader.limits.sampler = 16;
 271                    This->baseShader.limits.packed_input = 0;
 272                    This->baseShader.limits.label = 16;
 273                    break;
 274
 275           case WINED3DPS_VERSION(3,0):
 276                    This->baseShader.limits.temporary = 32;
 277                    This->baseShader.limits.constant_float = 224;
 278                    This->baseShader.limits.constant_int = 16;
 279                    This->baseShader.limits.constant_bool = 16;
 280                    This->baseShader.limits.texcoord = 0;
 281                    This->baseShader.limits.sampler = 16;
 282                    This->baseShader.limits.packed_input = 12;
 283                    This->baseShader.limits.label = 16; /* FIXME: 2048 */
 284                    break;
 285
 286           default: This->baseShader.limits.temporary = 32;
 287                    This->baseShader.limits.constant_float = 32;
 288                    This->baseShader.limits.constant_int = 16;
 289                    This->baseShader.limits.constant_bool = 16;
 290                    This->baseShader.limits.texcoord = 8;
 291                    This->baseShader.limits.sampler = 16;
 292                    This->baseShader.limits.packed_input = 0;
 293                    This->baseShader.limits.label = 0;
 294                    FIXME("Unrecognized pixel shader version %#x\n",
 295                            This->baseShader.reg_maps.shader_version);
 296       }
 297 }
 298
 299 static HRESULT WINAPI IWineD3DPixelShaderImpl_SetFunction(IWineD3DPixelShader *iface, CONST DWORD *pFunction) {
 300
 301     IWineD3DPixelShaderImpl *This =(IWineD3DPixelShaderImpl *)iface;
 302     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
 303     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
 304     shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
 305     HRESULT hr;
 306
 307     TRACE("(%p) : pFunction %p\n", iface, pFunction);
 308
 309     /* First pass: trace shader */
 310     if (TRACE_ON(d3d_shader)) shader_trace_init(pFunction, This->baseShader.shader_ins);
 311
 312     /* Initialize immediate constant lists */
 313     list_init(&This->baseShader.constantsF);
 314     list_init(&This->baseShader.constantsB);
 315     list_init(&This->baseShader.constantsI);
 316
 317     /* Second pass: figure out which registers are used, what the semantics are, etc.. */
 318     memset(reg_maps, 0, sizeof(shader_reg_maps));
 319     hr = shader_get_registers_used((IWineD3DBaseShader *)This, reg_maps, This->semantics_in, NULL, pFunction);
 320     if (FAILED(hr)) return hr;
 321
 322     pshader_set_limits(This);
 323
 324     for (i = 0; i < MAX_REG_INPUT; ++i)
 325     {
 326         if (This->input_reg_used[i])
 327         {
 328             ++num_regs_used;
 329             highest_reg_used = i;
 330         }
 331     }
 332
 333     /* Don't do any register mapping magic if it is not needed, or if we can't
 334      * achieve anything anyway */
 335     if (highest_reg_used < (GL_LIMITS(glsl_varyings) / 4)
 336             || num_regs_used > (GL_LIMITS(glsl_varyings) / 4))
 337     {
 338         if (num_regs_used > (GL_LIMITS(glsl_varyings) / 4))
 339         {
 340             /* This happens with relative addressing. The input mapper function
 341              * warns about this if the higher registers are declared too, so
 342              * don't write a FIXME here */
 343             WARN("More varying registers used than supported\n");
 344         }
 345
 346         for (i = 0; i < MAX_REG_INPUT; ++i)
 347         {
 348             This->input_reg_map[i] = i;
 349         }
 350
 351         This->declared_in_count = highest_reg_used + 1;
 352     }
 353     else
 354     {
 355         This->declared_in_count = 0;
 356         for (i = 0; i < MAX_REG_INPUT; ++i)
 357         {
 358             if (This->input_reg_used[i]) This->input_reg_map[i] = This->declared_in_count++;
 359             else This->input_reg_map[i] = -1;
 360         }
 361     }
 362
 363     This->baseShader.load_local_constsF = FALSE;
 364
 365     This->baseShader.shader_mode = deviceImpl->ps_selected_mode;
 366
 367     TRACE("(%p) : Copying the function\n", This);
 368
 369     This->baseShader.function = HeapAlloc(GetProcessHeap(), 0, This->baseShader.functionLength);
 370     if (!This->baseShader.function) return E_OUTOFMEMORY;
 371     memcpy(This->baseShader.function, pFunction, This->baseShader.functionLength);
 372
 373     return WINED3D_OK;
 374 }
 375
 376 static void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
 377 {
 378     DWORD shader_version = reg_maps->shader_version;
 379     DWORD *samplers = reg_maps->samplers;
 380     unsigned int i;
 381
 382     if (WINED3DSHADER_VERSION_MAJOR(shader_version) != 1) return;
 383
 384     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
 385     {
 386         /* We don't sample from this sampler */
 387         if (!samplers[i]) continue;
 388
 389         if (!textures[i])
 390         {
 391             ERR("No texture bound to sampler %u, using 2D\n", i);
 392             samplers[i] = (0x1 << 31) | WINED3DSTT_2D;
 393             continue;
 394         }
 395
 396         switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
 397         {
 398             case GL_TEXTURE_RECTANGLE_ARB:
 399             case GL_TEXTURE_2D:
 400                 /* We have to select between texture rectangles and 2D textures later because 2.0 and
 401                  * 3.0 shaders only have WINED3DSTT_2D as well */
 402                 samplers[i] = (1 << 31) | WINED3DSTT_2D;
 403                 break;
 404
 405             case GL_TEXTURE_3D:
 406                 samplers[i] = (1 << 31) | WINED3DSTT_VOLUME;
 407                 break;
 408
 409             case GL_TEXTURE_CUBE_MAP_ARB:
 410                 samplers[i] = (1 << 31) | WINED3DSTT_CUBE;
 411                 break;
 412
 413             default:
 414                 FIXME("Unrecognized texture type %#x, using 2D\n",
 415                         IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
 416                 samplers[i] = (0x1 << 31) | WINED3DSTT_2D;
 417         }
 418     }
 419 }
 420
 421 static GLuint pixelshader_compile(IWineD3DPixelShaderImpl *This, const struct ps_compile_args *args)
 422 {
 423     CONST DWORD *function = This->baseShader.function;
 424     GLuint retval;
 425     SHADER_BUFFER buffer;
 426     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
 427
 428     TRACE("(%p) : function %p\n", This, function);
 429
 430     pixelshader_update_samplers(&This->baseShader.reg_maps,
 431             ((IWineD3DDeviceImpl *)This->baseShader.device)->stateBlock->textures);
 432
 433     /* Reset fields tracking stateblock values being hardcoded in the shader */
 434     This->baseShader.num_sampled_samplers = 0;
 435
 436     /* Generate the HW shader */
 437     TRACE("(%p) : Generating hardware program\n", This);
 438     shader_buffer_init(&buffer);
 439     retval = device->shader_backend->shader_generate_pshader((IWineD3DPixelShader *)This, &buffer, args);
 440     shader_buffer_free(&buffer);
 441
 442     return retval;
 443 }
 444
 445 const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
 446 {
 447     /*** IUnknown methods ***/
 448     IWineD3DPixelShaderImpl_QueryInterface,
 449     IWineD3DPixelShaderImpl_AddRef,
 450     IWineD3DPixelShaderImpl_Release,
 451     /*** IWineD3DBase methods ***/
 452     IWineD3DPixelShaderImpl_GetParent,
 453     /*** IWineD3DBaseShader methods ***/
 454     IWineD3DPixelShaderImpl_SetFunction,
 455     /*** IWineD3DPixelShader methods ***/
 456     IWineD3DPixelShaderImpl_GetDevice,
 457     IWineD3DPixelShaderImpl_GetFunction
 458 };
 459
 460 void find_ps_compile_args(IWineD3DPixelShaderImpl *shader, IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args) {
 461     UINT i, sampler;
 462     IWineD3DBaseTextureImpl *tex;
 463
 464     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set */
 465     args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
 466
 467     for(i = 0; i < shader->baseShader.num_sampled_samplers; i++) {
 468         sampler = shader->baseShader.sampled_samplers[i];
 469         tex = (IWineD3DBaseTextureImpl *) stateblock->textures[sampler];
 470         if(!tex) {
 471             args->color_fixup[sampler] = COLOR_FIXUP_IDENTITY;
 472             continue;
 473         }
 474         args->color_fixup[sampler] = tex->baseTexture.shader_color_fixup;
 475     }
 476     if (shader->baseShader.reg_maps.shader_version >= WINED3DPS_VERSION(3,0))
 477     {
 478         if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
 479         {
 480             args->vp_mode = pretransformed;
 481         }
 482         else if (use_vs(stateblock))
 483         {
 484             args->vp_mode = vertexshader;
 485         } else {
 486             args->vp_mode = fixedfunction;
 487         }
 488         args->fog = FOG_OFF;
 489     } else {
 490         args->vp_mode = vertexshader;
 491         if(stateblock->renderState[WINED3DRS_FOGENABLE]) {
 492             switch(stateblock->renderState[WINED3DRS_FOGTABLEMODE]) {
 493                 case WINED3DFOG_NONE:
 494                     if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
 495                             || use_vs(stateblock))
 496                     {
 497                         args->fog = FOG_LINEAR;
 498                         break;
 499                     }
 500                     switch(stateblock->renderState[WINED3DRS_FOGVERTEXMODE]) {
 501                         case WINED3DFOG_NONE: /* Drop through */
 502                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
 503                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
 504                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
 505                     }
 506                     break;
 507
 508                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
 509                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
 510                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
 511             }
 512         } else {
 513             args->fog = FOG_OFF;
 514         }
 515     }
 516 }
 517
 518 GLuint find_gl_pshader(IWineD3DPixelShaderImpl *shader, const struct ps_compile_args *args)
 519 {
 520     UINT i;
 521     DWORD new_size;
 522     struct ps_compiled_shader *new_array;
 523
 524     /* Usually we have very few GL shaders for each d3d shader(just 1 or maybe 2),
 525      * so a linear search is more performant than a hashmap or a binary search
 526      * (cache coherency etc)
 527      */
 528     for(i = 0; i < shader->num_gl_shaders; i++) {
 529         if(memcmp(&shader->gl_shaders[i].args, args, sizeof(*args)) == 0) {
 530             return shader->gl_shaders[i].prgId;
 531         }
 532     }
 533
 534     TRACE("No matching GL shader found, compiling a new shader\n");
 535     if(shader->shader_array_size == shader->num_gl_shaders) {
 536         if(shader->gl_shaders) {
 537             new_size = shader->shader_array_size + max(1, shader->shader_array_size / 2);
 538             new_array = HeapReAlloc(GetProcessHeap(), 0, shader->gl_shaders,
 539                                     new_size * sizeof(*shader->gl_shaders));
 540         } else {
 541             new_array = HeapAlloc(GetProcessHeap(), 0, sizeof(*shader->gl_shaders));
 542             new_size = 1;
 543         }
 544
 545         if(!new_array) {
 546             ERR("Out of memory\n");
 547             return 0;
 548         }
 549         shader->gl_shaders = new_array;
 550         shader->shader_array_size = new_size;
 551     }
 552
 553     shader->gl_shaders[shader->num_gl_shaders].args = *args;
 554     shader->gl_shaders[shader->num_gl_shaders].prgId = pixelshader_compile(shader, args);
 555     return shader->gl_shaders[shader->num_gl_shaders++].prgId;
 556 }