git.oblomov.eu Git - wine/blob - 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   if (NULL == This->baseShader.function) { /* no function defined */
 119     TRACE("(%p) : GetFunction no User Function defined using NULL to %p\n", This, pData);
 120     (*(DWORD **) pData) = NULL;
 121   } else {
 122     TRACE("(%p) : GetFunction copying to %p\n", This, pData);
 123     memcpy(pData, This->baseShader.function, This->baseShader.functionLength);
 124   }
 125   return WINED3D_OK;
 126 }
 127
 128 CONST SHADER_OPCODE IWineD3DPixelShaderImpl_shader_ins[] = {
 129     /* Arithmetic */
 130     {WINED3DSIO_NOP,          "nop",          "NOP",               0, 0, WINED3DSIH_NOP,          0,                      0                     },
 131     {WINED3DSIO_MOV,          "mov",          "MOV",               1, 2, WINED3DSIH_MOV,          0,                      0                     },
 132     {WINED3DSIO_ADD,          "add",          "ADD",               1, 3, WINED3DSIH_ADD,          0,                      0                     },
 133     {WINED3DSIO_SUB,          "sub",          "SUB",               1, 3, WINED3DSIH_SUB,          0,                      0                     },
 134     {WINED3DSIO_MAD,          "mad",          "MAD",               1, 4, WINED3DSIH_MAD,          0,                      0                     },
 135     {WINED3DSIO_MUL,          "mul",          "MUL",               1, 3, WINED3DSIH_MUL,          0,                      0                     },
 136     {WINED3DSIO_RCP,          "rcp",          "RCP",               1, 2, WINED3DSIH_RCP,          0,                      0                     },
 137     {WINED3DSIO_RSQ,          "rsq",          "RSQ",               1, 2, WINED3DSIH_RSQ,          0,                      0                     },
 138     {WINED3DSIO_DP3,          "dp3",          "DP3",               1, 3, WINED3DSIH_DP3,          0,                      0                     },
 139     {WINED3DSIO_DP4,          "dp4",          "DP4",               1, 3, WINED3DSIH_DP4,          0,                      0                     },
 140     {WINED3DSIO_MIN,          "min",          "MIN",               1, 3, WINED3DSIH_MIN,          0,                      0                     },
 141     {WINED3DSIO_MAX,          "max",          "MAX",               1, 3, WINED3DSIH_MAX,          0,                      0                     },
 142     {WINED3DSIO_SLT,          "slt",          "SLT",               1, 3, WINED3DSIH_SLT,          0,                      0                     },
 143     {WINED3DSIO_SGE,          "sge",          "SGE",               1, 3, WINED3DSIH_SGE,          0,                      0                     },
 144     {WINED3DSIO_ABS,          "abs",          "ABS",               1, 2, WINED3DSIH_ABS,          0,                      0                     },
 145     {WINED3DSIO_EXP,          "exp",          "EX2",               1, 2, WINED3DSIH_EXP,          0,                      0                     },
 146     {WINED3DSIO_LOG,          "log",          "LG2",               1, 2, WINED3DSIH_LOG,          0,                      0                     },
 147     {WINED3DSIO_EXPP,         "expp",         "EXP",               1, 2, WINED3DSIH_EXPP,         0,                      0                     },
 148     {WINED3DSIO_LOGP,         "logp",         "LOG",               1, 2, WINED3DSIH_LOGP,         0,                      0                     },
 149     {WINED3DSIO_DST,          "dst",          "DST",               1, 3, WINED3DSIH_DST,          0,                      0                     },
 150     {WINED3DSIO_LRP,          "lrp",          "LRP",               1, 4, WINED3DSIH_LRP,          0,                      0                     },
 151     {WINED3DSIO_FRC,          "frc",          "FRC",               1, 2, WINED3DSIH_FRC,          0,                      0                     },
 152     {WINED3DSIO_CND,          "cnd",          NULL,                1, 4, WINED3DSIH_CND,          WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,4)},
 153     {WINED3DSIO_CMP,          "cmp",          NULL,                1, 4, WINED3DSIH_CMP,          WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(3,0)},
 154     {WINED3DSIO_POW,          "pow",          "POW",               1, 3, WINED3DSIH_POW,          0,                      0                     },
 155     {WINED3DSIO_CRS,          "crs",          "XPD",               1, 3, WINED3DSIH_CRS,          0,                      0                     },
 156     {WINED3DSIO_NRM,          "nrm",          NULL,                1, 2, WINED3DSIH_NRM,          0,                      0                     },
 157     {WINED3DSIO_SINCOS,       "sincos",       NULL,                1, 4, WINED3DSIH_SINCOS,       WINED3DPS_VERSION(2,0), WINED3DPS_VERSION(2,1)},
 158     {WINED3DSIO_SINCOS,       "sincos",       "SCS",               1, 2, WINED3DSIH_SINCOS,       WINED3DPS_VERSION(3,0), -1                    },
 159     {WINED3DSIO_DP2ADD,       "dp2add",       NULL,                1, 4, WINED3DSIH_DP2ADD,       WINED3DPS_VERSION(2,0), -1                    },
 160     /* Matrix */
 161     {WINED3DSIO_M4x4,         "m4x4",         "undefined",         1, 3, WINED3DSIH_M4x4,         0,                      0                     },
 162     {WINED3DSIO_M4x3,         "m4x3",         "undefined",         1, 3, WINED3DSIH_M4x3,         0,                      0                     },
 163     {WINED3DSIO_M3x4,         "m3x4",         "undefined",         1, 3, WINED3DSIH_M3x4,         0,                      0                     },
 164     {WINED3DSIO_M3x3,         "m3x3",         "undefined",         1, 3, WINED3DSIH_M3x3,         0,                      0                     },
 165     {WINED3DSIO_M3x2,         "m3x2",         "undefined",         1, 3, WINED3DSIH_M3x2,         0,                      0                     },
 166     /* Register declarations */
 167     {WINED3DSIO_DCL,          "dcl",          NULL,                0, 2, WINED3DSIH_DCL,          0,                      0                     },
 168     /* Flow control - requires GLSL or software shaders */
 169     {WINED3DSIO_REP ,         "rep",          NULL,                0, 1, WINED3DSIH_REP,          WINED3DPS_VERSION(2,1), -1                    },
 170     {WINED3DSIO_ENDREP,       "endrep",       NULL,                0, 0, WINED3DSIH_ENDREP,       WINED3DPS_VERSION(2,1), -1                    },
 171     {WINED3DSIO_IF,           "if",           NULL,                0, 1, WINED3DSIH_IF,           WINED3DPS_VERSION(2,1), -1                    },
 172     {WINED3DSIO_IFC,          "ifc",          NULL,                0, 2, WINED3DSIH_IFC,          WINED3DPS_VERSION(2,1), -1                    },
 173     {WINED3DSIO_ELSE,         "else",         NULL,                0, 0, WINED3DSIH_ELSE,         WINED3DPS_VERSION(2,1), -1                    },
 174     {WINED3DSIO_ENDIF,        "endif",        NULL,                0, 0, WINED3DSIH_ENDIF,        WINED3DPS_VERSION(2,1), -1                    },
 175     {WINED3DSIO_BREAK,        "break",        NULL,                0, 0, WINED3DSIH_BREAK,        WINED3DPS_VERSION(2,1), -1                    },
 176     {WINED3DSIO_BREAKC,       "breakc",       NULL,                0, 2, WINED3DSIH_BREAKC,       WINED3DPS_VERSION(2,1), -1                    },
 177     {WINED3DSIO_BREAKP,       "breakp",       GLNAME_REQUIRE_GLSL, 0, 1, WINED3DSIH_BREAKP,       0,                      0                     },
 178     {WINED3DSIO_CALL,         "call",         NULL,                0, 1, WINED3DSIH_CALL,         WINED3DPS_VERSION(2,1), -1                    },
 179     {WINED3DSIO_CALLNZ,       "callnz",       NULL,                0, 2, WINED3DSIH_CALLNZ,       WINED3DPS_VERSION(2,1), -1                    },
 180     {WINED3DSIO_LOOP,         "loop",         NULL,                0, 2, WINED3DSIH_LOOP,         WINED3DPS_VERSION(3,0), -1                    },
 181     {WINED3DSIO_RET,          "ret",          NULL,                0, 0, WINED3DSIH_RET,          WINED3DPS_VERSION(2,1), -1                    },
 182     {WINED3DSIO_ENDLOOP,      "endloop",      NULL,                0, 0, WINED3DSIH_ENDLOOP,      WINED3DPS_VERSION(3,0), -1                    },
 183     {WINED3DSIO_LABEL,        "label",        NULL,                0, 1, WINED3DSIH_LABEL,        WINED3DPS_VERSION(2,1), -1                    },
 184     /* Constant definitions */
 185     {WINED3DSIO_DEF,          "def",          "undefined",         1, 5, WINED3DSIH_DEF,          0,                      0                     },
 186     {WINED3DSIO_DEFB,         "defb",         GLNAME_REQUIRE_GLSL, 1, 2, WINED3DSIH_DEFB,         0,                      0                     },
 187     {WINED3DSIO_DEFI,         "defi",         GLNAME_REQUIRE_GLSL, 1, 5, WINED3DSIH_DEFI,         0,                      0                     },
 188     /* Texture */
 189     {WINED3DSIO_TEXCOORD,     "texcoord",     "undefined",         1, 1, WINED3DSIH_TEXCOORD,     0,                      WINED3DPS_VERSION(1,3)},
 190     {WINED3DSIO_TEXCOORD,     "texcrd",       "undefined",         1, 2, WINED3DSIH_TEXCOORD,     WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 191     {WINED3DSIO_TEXKILL,      "texkill",      "KIL",               1, 1, WINED3DSIH_TEXKILL,      WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(3,0)},
 192     {WINED3DSIO_TEX,          "tex",          "undefined",         1, 1, WINED3DSIH_TEX,          0,                      WINED3DPS_VERSION(1,3)},
 193     {WINED3DSIO_TEX,          "texld",        "undefined",         1, 2, WINED3DSIH_TEX,          WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 194     {WINED3DSIO_TEX,          "texld",        "undefined",         1, 3, WINED3DSIH_TEX,          WINED3DPS_VERSION(2,0), -1                    },
 195     {WINED3DSIO_TEXBEM,       "texbem",       "undefined",         1, 2, WINED3DSIH_TEXBEM,       0,                      WINED3DPS_VERSION(1,3)},
 196     {WINED3DSIO_TEXBEML,      "texbeml",      GLNAME_REQUIRE_GLSL, 1, 2, WINED3DSIH_TEXBEML,      WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 197     {WINED3DSIO_TEXREG2AR,    "texreg2ar",    "undefined",         1, 2, WINED3DSIH_TEXREG2AR,    WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 198     {WINED3DSIO_TEXREG2GB,    "texreg2gb",    "undefined",         1, 2, WINED3DSIH_TEXREG2GB,    WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 199     {WINED3DSIO_TEXREG2RGB,   "texreg2rgb",   "undefined",         1, 2, WINED3DSIH_TEXREG2RGB,   WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 200     {WINED3DSIO_TEXM3x2PAD,   "texm3x2pad",   "undefined",         1, 2, WINED3DSIH_TEXM3x2PAD,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 201     {WINED3DSIO_TEXM3x2TEX,   "texm3x2tex",   "undefined",         1, 2, WINED3DSIH_TEXM3x2TEX,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 202     {WINED3DSIO_TEXM3x3PAD,   "texm3x3pad",   "undefined",         1, 2, WINED3DSIH_TEXM3x3PAD,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 203     {WINED3DSIO_TEXM3x3DIFF,  "texm3x3diff",  GLNAME_REQUIRE_GLSL, 1, 2, WINED3DSIH_TEXM3x3DIFF,  WINED3DPS_VERSION(0,0), WINED3DPS_VERSION(0,0)},
 204     {WINED3DSIO_TEXM3x3SPEC,  "texm3x3spec",  "undefined",         1, 3, WINED3DSIH_TEXM3x3SPEC,  WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 205     {WINED3DSIO_TEXM3x3VSPEC, "texm3x3vspec", "undefined",         1, 2, WINED3DSIH_TEXM3x3VSPEC, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 206     {WINED3DSIO_TEXM3x3TEX,   "texm3x3tex",   "undefined",         1, 2, WINED3DSIH_TEXM3x3TEX,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 207     {WINED3DSIO_TEXDP3TEX,    "texdp3tex",    NULL,                1, 2, WINED3DSIH_TEXDP3TEX,    WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 208     {WINED3DSIO_TEXM3x2DEPTH, "texm3x2depth", GLNAME_REQUIRE_GLSL, 1, 2, WINED3DSIH_TEXM3x2DEPTH, WINED3DPS_VERSION(1,3), WINED3DPS_VERSION(1,3)},
 209     {WINED3DSIO_TEXDP3,       "texdp3",       NULL,                1, 2, WINED3DSIH_TEXDP3,       WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 210     {WINED3DSIO_TEXM3x3,      "texm3x3",      NULL,                1, 2, WINED3DSIH_TEXM3x3,      WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 211     {WINED3DSIO_TEXDEPTH,     "texdepth",     NULL,                1, 1, WINED3DSIH_TEXDEPTH,     WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 212     {WINED3DSIO_BEM,          "bem",          "undefined",         1, 3, WINED3DSIH_BEM,          WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 213     {WINED3DSIO_DSX,          "dsx",          NULL,                1, 2, WINED3DSIH_DSX,          WINED3DPS_VERSION(2,1), -1                    },
 214     {WINED3DSIO_DSY,          "dsy",          NULL,                1, 2, WINED3DSIH_DSY,          WINED3DPS_VERSION(2,1), -1                    },
 215     {WINED3DSIO_TEXLDD,       "texldd",       GLNAME_REQUIRE_GLSL, 1, 5, WINED3DSIH_TEXLDD,       WINED3DPS_VERSION(2,1), -1                    },
 216     {WINED3DSIO_SETP,         "setp",         GLNAME_REQUIRE_GLSL, 1, 3, WINED3DSIH_SETP,         0,                      0                     },
 217     {WINED3DSIO_TEXLDL,       "texldl",       NULL,                1, 3, WINED3DSIH_TEXLDL,       WINED3DPS_VERSION(3,0), -1                    },
 218     {WINED3DSIO_PHASE,        "phase",        GLNAME_REQUIRE_GLSL, 0, 0, WINED3DSIH_PHASE,        0,                      0                     },
 219     {0,                       NULL,           NULL,                0, 0, 0,                       0,                      0                     }
 220 };
 221
 222 static void pshader_set_limits(
 223       IWineD3DPixelShaderImpl *This) {
 224
 225       This->baseShader.limits.attributes = 0;
 226       This->baseShader.limits.address = 0;
 227       This->baseShader.limits.packed_output = 0;
 228
 229       switch (This->baseShader.hex_version) {
 230           case WINED3DPS_VERSION(1,0):
 231           case WINED3DPS_VERSION(1,1):
 232           case WINED3DPS_VERSION(1,2):
 233           case WINED3DPS_VERSION(1,3):
 234                    This->baseShader.limits.temporary = 2;
 235                    This->baseShader.limits.constant_float = 8;
 236                    This->baseShader.limits.constant_int = 0;
 237                    This->baseShader.limits.constant_bool = 0;
 238                    This->baseShader.limits.texcoord = 4;
 239                    This->baseShader.limits.sampler = 4;
 240                    This->baseShader.limits.packed_input = 0;
 241                    This->baseShader.limits.label = 0;
 242                    break;
 243
 244           case WINED3DPS_VERSION(1,4):
 245                    This->baseShader.limits.temporary = 6;
 246                    This->baseShader.limits.constant_float = 8;
 247                    This->baseShader.limits.constant_int = 0;
 248                    This->baseShader.limits.constant_bool = 0;
 249                    This->baseShader.limits.texcoord = 6;
 250                    This->baseShader.limits.sampler = 6;
 251                    This->baseShader.limits.packed_input = 0;
 252                    This->baseShader.limits.label = 0;
 253                    break;
 254
 255           /* FIXME: temporaries must match D3DPSHADERCAPS2_0.NumTemps */
 256           case WINED3DPS_VERSION(2,0):
 257                    This->baseShader.limits.temporary = 32;
 258                    This->baseShader.limits.constant_float = 32;
 259                    This->baseShader.limits.constant_int = 16;
 260                    This->baseShader.limits.constant_bool = 16;
 261                    This->baseShader.limits.texcoord = 8;
 262                    This->baseShader.limits.sampler = 16;
 263                    This->baseShader.limits.packed_input = 0;
 264                    break;
 265
 266           case WINED3DPS_VERSION(2,1):
 267                    This->baseShader.limits.temporary = 32;
 268                    This->baseShader.limits.constant_float = 32;
 269                    This->baseShader.limits.constant_int = 16;
 270                    This->baseShader.limits.constant_bool = 16;
 271                    This->baseShader.limits.texcoord = 8;
 272                    This->baseShader.limits.sampler = 16;
 273                    This->baseShader.limits.packed_input = 0;
 274                    This->baseShader.limits.label = 16;
 275                    break;
 276
 277           case WINED3DPS_VERSION(3,0):
 278                    This->baseShader.limits.temporary = 32;
 279                    This->baseShader.limits.constant_float = 224;
 280                    This->baseShader.limits.constant_int = 16;
 281                    This->baseShader.limits.constant_bool = 16;
 282                    This->baseShader.limits.texcoord = 0;
 283                    This->baseShader.limits.sampler = 16;
 284                    This->baseShader.limits.packed_input = 12;
 285                    This->baseShader.limits.label = 16; /* FIXME: 2048 */
 286                    break;
 287
 288           default: This->baseShader.limits.temporary = 32;
 289                    This->baseShader.limits.constant_float = 32;
 290                    This->baseShader.limits.constant_int = 16;
 291                    This->baseShader.limits.constant_bool = 16;
 292                    This->baseShader.limits.texcoord = 8;
 293                    This->baseShader.limits.sampler = 16;
 294                    This->baseShader.limits.packed_input = 0;
 295                    This->baseShader.limits.label = 0;
 296                    FIXME("Unrecognized pixel shader version %#x\n",
 297                        This->baseShader.hex_version);
 298       }
 299 }
 300
 301 /** Generate a pixel shader string using either GL_FRAGMENT_PROGRAM_ARB
 302     or GLSL and send it to the card */
 303 static inline GLuint IWineD3DPixelShaderImpl_GenerateShader(
 304     IWineD3DPixelShaderImpl *This) {
 305     SHADER_BUFFER buffer;
 306     GLuint shader;
 307
 308     shader_buffer_init(&buffer);
 309     shader = ((IWineD3DDeviceImpl *)This->baseShader.device)->shader_backend->shader_generate_pshader((IWineD3DPixelShader *)This, &buffer);
 310     shader_buffer_free(&buffer);
 311
 312     return shader;
 313 }
 314
 315 static HRESULT WINAPI IWineD3DPixelShaderImpl_SetFunction(IWineD3DPixelShader *iface, CONST DWORD *pFunction) {
 316
 317     IWineD3DPixelShaderImpl *This =(IWineD3DPixelShaderImpl *)iface;
 318     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
 319     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
 320     shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
 321     HRESULT hr;
 322
 323     TRACE("(%p) : pFunction %p\n", iface, pFunction);
 324
 325     /* First pass: trace shader */
 326     if (TRACE_ON(d3d_shader)) shader_trace_init(pFunction, This->baseShader.shader_ins);
 327
 328     /* Initialize immediate constant lists */
 329     list_init(&This->baseShader.constantsF);
 330     list_init(&This->baseShader.constantsB);
 331     list_init(&This->baseShader.constantsI);
 332
 333     /* Second pass: figure out which registers are used, what the semantics are, etc.. */
 334     memset(reg_maps, 0, sizeof(shader_reg_maps));
 335     hr = shader_get_registers_used((IWineD3DBaseShader *)This, reg_maps, This->semantics_in, NULL, pFunction);
 336     if (FAILED(hr)) return hr;
 337
 338     pshader_set_limits(This);
 339
 340     for (i = 0; i < MAX_REG_INPUT; ++i)
 341     {
 342         if (This->input_reg_used[i])
 343         {
 344             ++num_regs_used;
 345             highest_reg_used = i;
 346         }
 347     }
 348
 349     /* Don't do any register mapping magic if it is not needed, or if we can't
 350      * achieve anything anyway */
 351     if (highest_reg_used < (GL_LIMITS(glsl_varyings) / 4)
 352             || num_regs_used > (GL_LIMITS(glsl_varyings) / 4))
 353     {
 354         if (num_regs_used > (GL_LIMITS(glsl_varyings) / 4))
 355         {
 356             /* This happens with relative addressing. The input mapper function
 357              * warns about this if the higher registers are declared too, so
 358              * don't write a FIXME here */
 359             WARN("More varying registers used than supported\n");
 360         }
 361
 362         for (i = 0; i < MAX_REG_INPUT; ++i)
 363         {
 364             This->input_reg_map[i] = i;
 365         }
 366
 367         This->declared_in_count = highest_reg_used + 1;
 368     }
 369     else
 370     {
 371         This->declared_in_count = 0;
 372         for (i = 0; i < MAX_REG_INPUT; ++i)
 373         {
 374             if (This->input_reg_used[i]) This->input_reg_map[i] = This->declared_in_count++;
 375             else This->input_reg_map[i] = -1;
 376         }
 377     }
 378
 379     This->baseShader.load_local_constsF = FALSE;
 380
 381     This->baseShader.shader_mode = deviceImpl->ps_selected_mode;
 382
 383     TRACE("(%p) : Copying the function\n", This);
 384     if (NULL != pFunction) {
 385         void *function;
 386
 387         function = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, This->baseShader.functionLength);
 388         if (!function) return E_OUTOFMEMORY;
 389         memcpy(function, pFunction, This->baseShader.functionLength);
 390         This->baseShader.function = function;
 391     } else {
 392         This->baseShader.function = NULL;
 393     }
 394
 395     return WINED3D_OK;
 396 }
 397
 398 void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures,
 399         DWORD shader_version)
 400 {
 401     DWORD *samplers = reg_maps->samplers;
 402     unsigned int i;
 403
 404     if (WINED3DSHADER_VERSION_MAJOR(shader_version) != 1) return;
 405
 406     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
 407     {
 408         /* We don't sample from this sampler */
 409         if (!samplers[i]) continue;
 410
 411         if (!textures[i])
 412         {
 413             ERR("No texture bound to sampler %u, using 2D\n", i);
 414             samplers[i] = (0x1 << 31) | WINED3DSTT_2D;
 415             continue;
 416         }
 417
 418         switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
 419         {
 420             case GL_TEXTURE_RECTANGLE_ARB:
 421             case GL_TEXTURE_2D:
 422                 /* We have to select between texture rectangles and 2D textures later because 2.0 and
 423                  * 3.0 shaders only have WINED3DSTT_2D as well */
 424                 samplers[i] = (1 << 31) | WINED3DSTT_2D;
 425                 break;
 426
 427             case GL_TEXTURE_3D:
 428                 samplers[i] = (1 << 31) | WINED3DSTT_VOLUME;
 429                 break;
 430
 431             case GL_TEXTURE_CUBE_MAP_ARB:
 432                 samplers[i] = (1 << 31) | WINED3DSTT_CUBE;
 433                 break;
 434
 435             default:
 436                 FIXME("Unrecognized texture type %#x, using 2D\n",
 437                         IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
 438                 samplers[i] = (0x1 << 31) | WINED3DSTT_2D;
 439         }
 440     }
 441 }
 442
 443 static GLuint pixelshader_compile(IWineD3DPixelShaderImpl *This, const struct ps_compile_args *args)
 444 {
 445     CONST DWORD *function = This->baseShader.function;
 446     GLuint retval;
 447
 448     TRACE("(%p) : function %p\n", This, function);
 449
 450     pixelshader_update_samplers(&This->baseShader.reg_maps,
 451             ((IWineD3DDeviceImpl *)This->baseShader.device)->stateBlock->textures, This->baseShader.hex_version);
 452
 453     /* Reset fields tracking stateblock values being hardcoded in the shader */
 454     This->baseShader.num_sampled_samplers = 0;
 455
 456     /* Generate the HW shader */
 457     TRACE("(%p) : Generating hardware program\n", This);
 458     retval = IWineD3DPixelShaderImpl_GenerateShader(This);
 459
 460     This->baseShader.is_compiled = TRUE;
 461
 462     return retval;
 463 }
 464
 465 const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
 466 {
 467     /*** IUnknown methods ***/
 468     IWineD3DPixelShaderImpl_QueryInterface,
 469     IWineD3DPixelShaderImpl_AddRef,
 470     IWineD3DPixelShaderImpl_Release,
 471     /*** IWineD3DBase methods ***/
 472     IWineD3DPixelShaderImpl_GetParent,
 473     /*** IWineD3DBaseShader methods ***/
 474     IWineD3DPixelShaderImpl_SetFunction,
 475     /*** IWineD3DPixelShader methods ***/
 476     IWineD3DPixelShaderImpl_GetDevice,
 477     IWineD3DPixelShaderImpl_GetFunction
 478 };
 479
 480 void find_ps_compile_args(IWineD3DPixelShaderImpl *shader, IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args) {
 481     UINT i, sampler;
 482     IWineD3DBaseTextureImpl *tex;
 483
 484     args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
 485
 486     memset(args->color_fixup, 0, sizeof(args->color_fixup));
 487     for(i = 0; i < shader->baseShader.num_sampled_samplers; i++) {
 488         sampler = shader->baseShader.sampled_samplers[i];
 489         tex = (IWineD3DBaseTextureImpl *) stateblock->textures[sampler];
 490         if(!tex) {
 491             args->color_fixup[sampler] = COLOR_FIXUP_IDENTITY;
 492             continue;
 493         }
 494         args->color_fixup[sampler] = tex->baseTexture.shader_color_fixup;
 495     }
 496     if(shader->baseShader.hex_version >= WINED3DPS_VERSION(3,0)) {
 497         if(((IWineD3DDeviceImpl *) shader->baseShader.device)->strided_streams.u.s.position_transformed) {
 498             args->vp_mode = pretransformed;
 499         } else if(use_vs((IWineD3DDeviceImpl *) shader->baseShader.device)) {
 500             args->vp_mode = vertexshader;
 501         } else {
 502             args->vp_mode = fixedfunction;
 503         }
 504     } else {
 505         args->vp_mode = vertexshader;
 506     }
 507 }
 508
 509 GLuint find_gl_pshader(IWineD3DPixelShaderImpl *shader, const struct ps_compile_args *args)
 510 {
 511     UINT i;
 512     struct ps_compiled_shader *old_array;
 513
 514     /* Usually we have very few GL shaders for each d3d shader(just 1 or maybe 2),
 515      * so a linear search is more performant than a hashmap
 516      */
 517     for(i = 0; i < shader->num_gl_shaders; i++) {
 518         if(memcmp(&shader->gl_shaders[i].args, args, sizeof(*args)) == 0) {
 519             return shader->gl_shaders[i].prgId;
 520         }
 521     }
 522
 523     TRACE("No matching GL shader found, compiling a new shader\n");
 524     old_array = shader->gl_shaders;
 525     if(old_array) {
 526         shader->gl_shaders = HeapReAlloc(GetProcessHeap(), 0, old_array,
 527                                          (shader->num_gl_shaders + 1) * sizeof(*shader->gl_shaders));
 528     } else {
 529         shader->gl_shaders = HeapAlloc(GetProcessHeap(), 0, sizeof(*shader->gl_shaders));
 530     }
 531
 532     if(!shader->gl_shaders) {
 533         ERR("Out of memory\n");
 534         return 0;
 535     }
 536
 537     shader->gl_shaders[shader->num_gl_shaders].args = *args;
 538     shader->gl_shaders[shader->num_gl_shaders].prgId = pixelshader_compile(shader, args);
 539     return shader->gl_shaders[shader->num_gl_shaders++].prgId;
 540 }