user32: Use MAKEWPARAM instead of MAKELONG for menu messages.
[wine] / 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 static HRESULT  WINAPI IWineD3DPixelShaderImpl_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, LPVOID *ppobj) {
38     TRACE("iface %p, riid %s, ppobj %p\n", iface, debugstr_guid(riid), ppobj);
39
40     if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
41             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
42             || IsEqualGUID(riid, &IID_IWineD3DBase)
43             || IsEqualGUID(riid, &IID_IUnknown))
44     {
45         IUnknown_AddRef(iface);
46         *ppobj = iface;
47         return S_OK;
48     }
49
50     WARN("%s not implemented, returning E_NOINTERFACE\n", debugstr_guid(riid));
51
52     *ppobj = NULL;
53     return E_NOINTERFACE;
54 }
55
56 static ULONG  WINAPI IWineD3DPixelShaderImpl_AddRef(IWineD3DPixelShader *iface) {
57     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
58     ULONG refcount = InterlockedIncrement(&This->baseShader.ref);
59
60     TRACE("%p increasing refcount to %u\n", This, refcount);
61
62     return refcount;
63 }
64
65 static ULONG  WINAPI IWineD3DPixelShaderImpl_Release(IWineD3DPixelShader *iface) {
66     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
67     ULONG refcount = InterlockedDecrement(&This->baseShader.ref);
68
69     TRACE("%p decreasing refcount to %u\n", This, refcount);
70
71     if (!refcount)
72     {
73         shader_cleanup((IWineD3DBaseShader *)iface);
74         HeapFree(GetProcessHeap(), 0, This);
75     }
76
77     return refcount;
78 }
79
80 /* *******************************************
81    IWineD3DPixelShader IWineD3DPixelShader parts follow
82    ******************************************* */
83
84 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetParent(IWineD3DPixelShader *iface, IUnknown** parent){
85     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
86
87     *parent = This->parent;
88     IUnknown_AddRef(*parent);
89     TRACE("(%p) : returning %p\n", This, *parent);
90     return WINED3D_OK;
91 }
92
93 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetDevice(IWineD3DPixelShader* iface, IWineD3DDevice **pDevice){
94     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
95     IWineD3DDevice_AddRef(This->baseShader.device);
96     *pDevice = This->baseShader.device;
97     TRACE("(%p) returning %p\n", This, *pDevice);
98     return WINED3D_OK;
99 }
100
101
102 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetFunction(IWineD3DPixelShader* impl, VOID* pData, UINT* pSizeOfData) {
103   IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)impl;
104   TRACE("(%p) : pData(%p), pSizeOfData(%p)\n", This, pData, pSizeOfData);
105
106   if (NULL == pData) {
107     *pSizeOfData = This->baseShader.functionLength;
108     return WINED3D_OK;
109   }
110   if (*pSizeOfData < This->baseShader.functionLength) {
111     /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
112      * than the required size we should write the required size and
113      * return D3DERR_MOREDATA. That's not actually true. */
114     return WINED3DERR_INVALIDCALL;
115   }
116
117   TRACE("(%p) : GetFunction copying to %p\n", This, pData);
118   memcpy(pData, This->baseShader.function, This->baseShader.functionLength);
119
120   return WINED3D_OK;
121 }
122
123 CONST SHADER_OPCODE IWineD3DPixelShaderImpl_shader_ins[] = {
124     /* Arithmetic */
125     {WINED3DSIO_NOP,          "nop",          0, 0, WINED3DSIH_NOP,          0,                      0                     },
126     {WINED3DSIO_MOV,          "mov",          1, 2, WINED3DSIH_MOV,          0,                      0                     },
127     {WINED3DSIO_ADD,          "add",          1, 3, WINED3DSIH_ADD,          0,                      0                     },
128     {WINED3DSIO_SUB,          "sub",          1, 3, WINED3DSIH_SUB,          0,                      0                     },
129     {WINED3DSIO_MAD,          "mad",          1, 4, WINED3DSIH_MAD,          0,                      0                     },
130     {WINED3DSIO_MUL,          "mul",          1, 3, WINED3DSIH_MUL,          0,                      0                     },
131     {WINED3DSIO_RCP,          "rcp",          1, 2, WINED3DSIH_RCP,          0,                      0                     },
132     {WINED3DSIO_RSQ,          "rsq",          1, 2, WINED3DSIH_RSQ,          0,                      0                     },
133     {WINED3DSIO_DP3,          "dp3",          1, 3, WINED3DSIH_DP3,          0,                      0                     },
134     {WINED3DSIO_DP4,          "dp4",          1, 3, WINED3DSIH_DP4,          0,                      0                     },
135     {WINED3DSIO_MIN,          "min",          1, 3, WINED3DSIH_MIN,          0,                      0                     },
136     {WINED3DSIO_MAX,          "max",          1, 3, WINED3DSIH_MAX,          0,                      0                     },
137     {WINED3DSIO_SLT,          "slt",          1, 3, WINED3DSIH_SLT,          0,                      0                     },
138     {WINED3DSIO_SGE,          "sge",          1, 3, WINED3DSIH_SGE,          0,                      0                     },
139     {WINED3DSIO_ABS,          "abs",          1, 2, WINED3DSIH_ABS,          0,                      0                     },
140     {WINED3DSIO_EXP,          "exp",          1, 2, WINED3DSIH_EXP,          0,                      0                     },
141     {WINED3DSIO_LOG,          "log",          1, 2, WINED3DSIH_LOG,          0,                      0                     },
142     {WINED3DSIO_EXPP,         "expp",         1, 2, WINED3DSIH_EXPP,         0,                      0                     },
143     {WINED3DSIO_LOGP,         "logp",         1, 2, WINED3DSIH_LOGP,         0,                      0                     },
144     {WINED3DSIO_DST,          "dst",          1, 3, WINED3DSIH_DST,          0,                      0                     },
145     {WINED3DSIO_LRP,          "lrp",          1, 4, WINED3DSIH_LRP,          0,                      0                     },
146     {WINED3DSIO_FRC,          "frc",          1, 2, WINED3DSIH_FRC,          0,                      0                     },
147     {WINED3DSIO_CND,          "cnd",          1, 4, WINED3DSIH_CND,          WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,4)},
148     {WINED3DSIO_CMP,          "cmp",          1, 4, WINED3DSIH_CMP,          WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(3,0)},
149     {WINED3DSIO_POW,          "pow",          1, 3, WINED3DSIH_POW,          0,                      0                     },
150     {WINED3DSIO_CRS,          "crs",          1, 3, WINED3DSIH_CRS,          0,                      0                     },
151     {WINED3DSIO_NRM,          "nrm",          1, 2, WINED3DSIH_NRM,          0,                      0                     },
152     {WINED3DSIO_SINCOS,       "sincos",       1, 4, WINED3DSIH_SINCOS,       WINED3DPS_VERSION(2,0), WINED3DPS_VERSION(2,1)},
153     {WINED3DSIO_SINCOS,       "sincos",       1, 2, WINED3DSIH_SINCOS,       WINED3DPS_VERSION(3,0), -1                    },
154     {WINED3DSIO_DP2ADD,       "dp2add",       1, 4, WINED3DSIH_DP2ADD,       WINED3DPS_VERSION(2,0), -1                    },
155     /* Matrix */
156     {WINED3DSIO_M4x4,         "m4x4",         1, 3, WINED3DSIH_M4x4,         0,                      0                     },
157     {WINED3DSIO_M4x3,         "m4x3",         1, 3, WINED3DSIH_M4x3,         0,                      0                     },
158     {WINED3DSIO_M3x4,         "m3x4",         1, 3, WINED3DSIH_M3x4,         0,                      0                     },
159     {WINED3DSIO_M3x3,         "m3x3",         1, 3, WINED3DSIH_M3x3,         0,                      0                     },
160     {WINED3DSIO_M3x2,         "m3x2",         1, 3, WINED3DSIH_M3x2,         0,                      0                     },
161     /* Register declarations */
162     {WINED3DSIO_DCL,          "dcl",          0, 2, WINED3DSIH_DCL,          0,                      0                     },
163     /* Flow control - requires GLSL or software shaders */
164     {WINED3DSIO_REP ,         "rep",          0, 1, WINED3DSIH_REP,          WINED3DPS_VERSION(2,1), -1                    },
165     {WINED3DSIO_ENDREP,       "endrep",       0, 0, WINED3DSIH_ENDREP,       WINED3DPS_VERSION(2,1), -1                    },
166     {WINED3DSIO_IF,           "if",           0, 1, WINED3DSIH_IF,           WINED3DPS_VERSION(2,1), -1                    },
167     {WINED3DSIO_IFC,          "ifc",          0, 2, WINED3DSIH_IFC,          WINED3DPS_VERSION(2,1), -1                    },
168     {WINED3DSIO_ELSE,         "else",         0, 0, WINED3DSIH_ELSE,         WINED3DPS_VERSION(2,1), -1                    },
169     {WINED3DSIO_ENDIF,        "endif",        0, 0, WINED3DSIH_ENDIF,        WINED3DPS_VERSION(2,1), -1                    },
170     {WINED3DSIO_BREAK,        "break",        0, 0, WINED3DSIH_BREAK,        WINED3DPS_VERSION(2,1), -1                    },
171     {WINED3DSIO_BREAKC,       "breakc",       0, 2, WINED3DSIH_BREAKC,       WINED3DPS_VERSION(2,1), -1                    },
172     {WINED3DSIO_BREAKP,       "breakp",       0, 1, WINED3DSIH_BREAKP,       0,                      0                     },
173     {WINED3DSIO_CALL,         "call",         0, 1, WINED3DSIH_CALL,         WINED3DPS_VERSION(2,1), -1                    },
174     {WINED3DSIO_CALLNZ,       "callnz",       0, 2, WINED3DSIH_CALLNZ,       WINED3DPS_VERSION(2,1), -1                    },
175     {WINED3DSIO_LOOP,         "loop",         0, 2, WINED3DSIH_LOOP,         WINED3DPS_VERSION(3,0), -1                    },
176     {WINED3DSIO_RET,          "ret",          0, 0, WINED3DSIH_RET,          WINED3DPS_VERSION(2,1), -1                    },
177     {WINED3DSIO_ENDLOOP,      "endloop",      0, 0, WINED3DSIH_ENDLOOP,      WINED3DPS_VERSION(3,0), -1                    },
178     {WINED3DSIO_LABEL,        "label",        0, 1, WINED3DSIH_LABEL,        WINED3DPS_VERSION(2,1), -1                    },
179     /* Constant definitions */
180     {WINED3DSIO_DEF,          "def",          1, 5, WINED3DSIH_DEF,          0,                      0                     },
181     {WINED3DSIO_DEFB,         "defb",         1, 2, WINED3DSIH_DEFB,         0,                      0                     },
182     {WINED3DSIO_DEFI,         "defi",         1, 5, WINED3DSIH_DEFI,         0,                      0                     },
183     /* Texture */
184     {WINED3DSIO_TEXCOORD,     "texcoord",     1, 1, WINED3DSIH_TEXCOORD,     0,                      WINED3DPS_VERSION(1,3)},
185     {WINED3DSIO_TEXCOORD,     "texcrd",       1, 2, WINED3DSIH_TEXCOORD,     WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
186     {WINED3DSIO_TEXKILL,      "texkill",      1, 1, WINED3DSIH_TEXKILL,      WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(3,0)},
187     {WINED3DSIO_TEX,          "tex",          1, 1, WINED3DSIH_TEX,          0,                      WINED3DPS_VERSION(1,3)},
188     {WINED3DSIO_TEX,          "texld",        1, 2, WINED3DSIH_TEX,          WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
189     {WINED3DSIO_TEX,          "texld",        1, 3, WINED3DSIH_TEX,          WINED3DPS_VERSION(2,0), -1                    },
190     {WINED3DSIO_TEXBEM,       "texbem",       1, 2, WINED3DSIH_TEXBEM,       0,                      WINED3DPS_VERSION(1,3)},
191     {WINED3DSIO_TEXBEML,      "texbeml",      1, 2, WINED3DSIH_TEXBEML,      WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
192     {WINED3DSIO_TEXREG2AR,    "texreg2ar",    1, 2, WINED3DSIH_TEXREG2AR,    WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
193     {WINED3DSIO_TEXREG2GB,    "texreg2gb",    1, 2, WINED3DSIH_TEXREG2GB,    WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
194     {WINED3DSIO_TEXREG2RGB,   "texreg2rgb",   1, 2, WINED3DSIH_TEXREG2RGB,   WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
195     {WINED3DSIO_TEXM3x2PAD,   "texm3x2pad",   1, 2, WINED3DSIH_TEXM3x2PAD,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
196     {WINED3DSIO_TEXM3x2TEX,   "texm3x2tex",   1, 2, WINED3DSIH_TEXM3x2TEX,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
197     {WINED3DSIO_TEXM3x3PAD,   "texm3x3pad",   1, 2, WINED3DSIH_TEXM3x3PAD,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
198     {WINED3DSIO_TEXM3x3DIFF,  "texm3x3diff",  1, 2, WINED3DSIH_TEXM3x3DIFF,  WINED3DPS_VERSION(0,0), WINED3DPS_VERSION(0,0)},
199     {WINED3DSIO_TEXM3x3SPEC,  "texm3x3spec",  1, 3, WINED3DSIH_TEXM3x3SPEC,  WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
200     {WINED3DSIO_TEXM3x3VSPEC, "texm3x3vspec", 1, 2, WINED3DSIH_TEXM3x3VSPEC, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
201     {WINED3DSIO_TEXM3x3TEX,   "texm3x3tex",   1, 2, WINED3DSIH_TEXM3x3TEX,   WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
202     {WINED3DSIO_TEXDP3TEX,    "texdp3tex",    1, 2, WINED3DSIH_TEXDP3TEX,    WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
203     {WINED3DSIO_TEXM3x2DEPTH, "texm3x2depth", 1, 2, WINED3DSIH_TEXM3x2DEPTH, WINED3DPS_VERSION(1,3), WINED3DPS_VERSION(1,3)},
204     {WINED3DSIO_TEXDP3,       "texdp3",       1, 2, WINED3DSIH_TEXDP3,       WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
205     {WINED3DSIO_TEXM3x3,      "texm3x3",      1, 2, WINED3DSIH_TEXM3x3,      WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
206     {WINED3DSIO_TEXDEPTH,     "texdepth",     1, 1, WINED3DSIH_TEXDEPTH,     WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
207     {WINED3DSIO_BEM,          "bem",          1, 3, WINED3DSIH_BEM,          WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
208     {WINED3DSIO_DSX,          "dsx",          1, 2, WINED3DSIH_DSX,          WINED3DPS_VERSION(2,1), -1                    },
209     {WINED3DSIO_DSY,          "dsy",          1, 2, WINED3DSIH_DSY,          WINED3DPS_VERSION(2,1), -1                    },
210     {WINED3DSIO_TEXLDD,       "texldd",       1, 5, WINED3DSIH_TEXLDD,       WINED3DPS_VERSION(2,1), -1                    },
211     {WINED3DSIO_SETP,         "setp",         1, 3, WINED3DSIH_SETP,         0,                      0                     },
212     {WINED3DSIO_TEXLDL,       "texldl",       1, 3, WINED3DSIH_TEXLDL,       WINED3DPS_VERSION(3,0), -1                    },
213     {WINED3DSIO_PHASE,        "phase",        0, 0, WINED3DSIH_PHASE,        0,                      0                     },
214     {0,                       NULL,           0, 0, 0,                       0,                      0                     }
215 };
216
217 static void pshader_set_limits(
218       IWineD3DPixelShaderImpl *This) { 
219
220       This->baseShader.limits.attributes = 0;
221       This->baseShader.limits.address = 0;
222       This->baseShader.limits.packed_output = 0;
223
224       switch (This->baseShader.reg_maps.shader_version)
225       {
226           case WINED3DPS_VERSION(1,0):
227           case WINED3DPS_VERSION(1,1):
228           case WINED3DPS_VERSION(1,2):
229           case WINED3DPS_VERSION(1,3): 
230                    This->baseShader.limits.temporary = 2;
231                    This->baseShader.limits.constant_float = 8;
232                    This->baseShader.limits.constant_int = 0;
233                    This->baseShader.limits.constant_bool = 0;
234                    This->baseShader.limits.texcoord = 4;
235                    This->baseShader.limits.sampler = 4;
236                    This->baseShader.limits.packed_input = 0;
237                    This->baseShader.limits.label = 0;
238                    break;
239
240           case WINED3DPS_VERSION(1,4):
241                    This->baseShader.limits.temporary = 6;
242                    This->baseShader.limits.constant_float = 8;
243                    This->baseShader.limits.constant_int = 0;
244                    This->baseShader.limits.constant_bool = 0;
245                    This->baseShader.limits.texcoord = 6;
246                    This->baseShader.limits.sampler = 6;
247                    This->baseShader.limits.packed_input = 0;
248                    This->baseShader.limits.label = 0;
249                    break;
250                
251           /* FIXME: temporaries must match D3DPSHADERCAPS2_0.NumTemps */ 
252           case WINED3DPS_VERSION(2,0):
253                    This->baseShader.limits.temporary = 32;
254                    This->baseShader.limits.constant_float = 32;
255                    This->baseShader.limits.constant_int = 16;
256                    This->baseShader.limits.constant_bool = 16;
257                    This->baseShader.limits.texcoord = 8;
258                    This->baseShader.limits.sampler = 16;
259                    This->baseShader.limits.packed_input = 0;
260                    break;
261
262           case WINED3DPS_VERSION(2,1):
263                    This->baseShader.limits.temporary = 32;
264                    This->baseShader.limits.constant_float = 32;
265                    This->baseShader.limits.constant_int = 16;
266                    This->baseShader.limits.constant_bool = 16;
267                    This->baseShader.limits.texcoord = 8;
268                    This->baseShader.limits.sampler = 16;
269                    This->baseShader.limits.packed_input = 0;
270                    This->baseShader.limits.label = 16;
271                    break;
272
273           case WINED3DPS_VERSION(3,0):
274                    This->baseShader.limits.temporary = 32;
275                    This->baseShader.limits.constant_float = 224;
276                    This->baseShader.limits.constant_int = 16;
277                    This->baseShader.limits.constant_bool = 16;
278                    This->baseShader.limits.texcoord = 0;
279                    This->baseShader.limits.sampler = 16;
280                    This->baseShader.limits.packed_input = 12;
281                    This->baseShader.limits.label = 16; /* FIXME: 2048 */
282                    break;
283
284           default: This->baseShader.limits.temporary = 32;
285                    This->baseShader.limits.constant_float = 32;
286                    This->baseShader.limits.constant_int = 16;
287                    This->baseShader.limits.constant_bool = 16;
288                    This->baseShader.limits.texcoord = 8;
289                    This->baseShader.limits.sampler = 16;
290                    This->baseShader.limits.packed_input = 0;
291                    This->baseShader.limits.label = 0;
292                    FIXME("Unrecognized pixel shader version %#x\n",
293                            This->baseShader.reg_maps.shader_version);
294       }
295 }
296
297 static HRESULT WINAPI IWineD3DPixelShaderImpl_SetFunction(IWineD3DPixelShader *iface, CONST DWORD *pFunction) {
298
299     IWineD3DPixelShaderImpl *This =(IWineD3DPixelShaderImpl *)iface;
300     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
301     unsigned int i, highest_reg_used = 0, num_regs_used = 0;
302     shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
303     HRESULT hr;
304
305     TRACE("(%p) : pFunction %p\n", iface, pFunction);
306
307     /* First pass: trace shader */
308     if (TRACE_ON(d3d_shader)) shader_trace_init(pFunction, This->baseShader.shader_ins);
309
310     /* Initialize immediate constant lists */
311     list_init(&This->baseShader.constantsF);
312     list_init(&This->baseShader.constantsB);
313     list_init(&This->baseShader.constantsI);
314
315     /* Second pass: figure out which registers are used, what the semantics are, etc.. */
316     memset(reg_maps, 0, sizeof(shader_reg_maps));
317     hr = shader_get_registers_used((IWineD3DBaseShader *)This, reg_maps, This->semantics_in, NULL, pFunction);
318     if (FAILED(hr)) return hr;
319
320     pshader_set_limits(This);
321
322     for (i = 0; i < MAX_REG_INPUT; ++i)
323     {
324         if (This->input_reg_used[i])
325         {
326             ++num_regs_used;
327             highest_reg_used = i;
328         }
329     }
330
331     /* Don't do any register mapping magic if it is not needed, or if we can't
332      * achieve anything anyway */
333     if (highest_reg_used < (GL_LIMITS(glsl_varyings) / 4)
334             || num_regs_used > (GL_LIMITS(glsl_varyings) / 4))
335     {
336         if (num_regs_used > (GL_LIMITS(glsl_varyings) / 4))
337         {
338             /* This happens with relative addressing. The input mapper function
339              * warns about this if the higher registers are declared too, so
340              * don't write a FIXME here */
341             WARN("More varying registers used than supported\n");
342         }
343
344         for (i = 0; i < MAX_REG_INPUT; ++i)
345         {
346             This->input_reg_map[i] = i;
347         }
348
349         This->declared_in_count = highest_reg_used + 1;
350     }
351     else
352     {
353         This->declared_in_count = 0;
354         for (i = 0; i < MAX_REG_INPUT; ++i)
355         {
356             if (This->input_reg_used[i]) This->input_reg_map[i] = This->declared_in_count++;
357             else This->input_reg_map[i] = -1;
358         }
359     }
360
361     This->baseShader.load_local_constsF = FALSE;
362
363     This->baseShader.shader_mode = deviceImpl->ps_selected_mode;
364
365     TRACE("(%p) : Copying the function\n", This);
366
367     This->baseShader.function = HeapAlloc(GetProcessHeap(), 0, This->baseShader.functionLength);
368     if (!This->baseShader.function) return E_OUTOFMEMORY;
369     memcpy(This->baseShader.function, pFunction, This->baseShader.functionLength);
370
371     return WINED3D_OK;
372 }
373
374 static void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
375 {
376     DWORD shader_version = reg_maps->shader_version;
377     DWORD *samplers = reg_maps->samplers;
378     unsigned int i;
379
380     if (WINED3DSHADER_VERSION_MAJOR(shader_version) != 1) return;
381
382     for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
383     {
384         /* We don't sample from this sampler */
385         if (!samplers[i]) continue;
386
387         if (!textures[i])
388         {
389             ERR("No texture bound to sampler %u, using 2D\n", i);
390             samplers[i] = (0x1 << 31) | WINED3DSTT_2D;
391             continue;
392         }
393
394         switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
395         {
396             case GL_TEXTURE_RECTANGLE_ARB:
397             case GL_TEXTURE_2D:
398                 /* We have to select between texture rectangles and 2D textures later because 2.0 and
399                  * 3.0 shaders only have WINED3DSTT_2D as well */
400                 samplers[i] = (1 << 31) | WINED3DSTT_2D;
401                 break;
402
403             case GL_TEXTURE_3D:
404                 samplers[i] = (1 << 31) | WINED3DSTT_VOLUME;
405                 break;
406
407             case GL_TEXTURE_CUBE_MAP_ARB:
408                 samplers[i] = (1 << 31) | WINED3DSTT_CUBE;
409                 break;
410
411             default:
412                 FIXME("Unrecognized texture type %#x, using 2D\n",
413                         IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
414                 samplers[i] = (0x1 << 31) | WINED3DSTT_2D;
415         }
416     }
417 }
418
419 static GLuint pixelshader_compile(IWineD3DPixelShaderImpl *This, const struct ps_compile_args *args)
420 {
421     CONST DWORD *function = This->baseShader.function;
422     GLuint retval;
423     SHADER_BUFFER buffer;
424     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
425
426     TRACE("(%p) : function %p\n", This, function);
427
428     pixelshader_update_samplers(&This->baseShader.reg_maps,
429             ((IWineD3DDeviceImpl *)This->baseShader.device)->stateBlock->textures);
430
431     /* Generate the HW shader */
432     TRACE("(%p) : Generating hardware program\n", This);
433     This->cur_args = args;
434     shader_buffer_init(&buffer);
435     retval = device->shader_backend->shader_generate_pshader((IWineD3DPixelShader *)This, &buffer, args);
436     shader_buffer_free(&buffer);
437     This->cur_args = NULL;
438
439     return retval;
440 }
441
442 const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
443 {
444     /*** IUnknown methods ***/
445     IWineD3DPixelShaderImpl_QueryInterface,
446     IWineD3DPixelShaderImpl_AddRef,
447     IWineD3DPixelShaderImpl_Release,
448     /*** IWineD3DBase methods ***/
449     IWineD3DPixelShaderImpl_GetParent,
450     /*** IWineD3DBaseShader methods ***/
451     IWineD3DPixelShaderImpl_SetFunction,
452     /*** IWineD3DPixelShader methods ***/
453     IWineD3DPixelShaderImpl_GetDevice,
454     IWineD3DPixelShaderImpl_GetFunction
455 };
456
457 void find_ps_compile_args(IWineD3DPixelShaderImpl *shader, IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args) {
458     UINT i;
459     IWineD3DBaseTextureImpl *tex;
460
461     memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set */
462     args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
463
464     for(i = 0; i < MAX_FRAGMENT_SAMPLERS; i++) {
465         if(shader->baseShader.reg_maps.samplers[i] == 0) continue;
466         tex = (IWineD3DBaseTextureImpl *) stateblock->textures[i];
467         if(!tex) {
468             args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
469             continue;
470         }
471         args->color_fixup[i] = tex->baseTexture.shader_color_fixup;
472     }
473     if (shader->baseShader.reg_maps.shader_version >= WINED3DPS_VERSION(3,0))
474     {
475         if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
476         {
477             args->vp_mode = pretransformed;
478         }
479         else if (use_vs(stateblock))
480         {
481             args->vp_mode = vertexshader;
482         } else {
483             args->vp_mode = fixedfunction;
484         }
485         args->fog = FOG_OFF;
486     } else {
487         args->vp_mode = vertexshader;
488         if(stateblock->renderState[WINED3DRS_FOGENABLE]) {
489             switch(stateblock->renderState[WINED3DRS_FOGTABLEMODE]) {
490                 case WINED3DFOG_NONE:
491                     if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
492                             || use_vs(stateblock))
493                     {
494                         args->fog = FOG_LINEAR;
495                         break;
496                     }
497                     switch(stateblock->renderState[WINED3DRS_FOGVERTEXMODE]) {
498                         case WINED3DFOG_NONE: /* Drop through */
499                         case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
500                         case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
501                         case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
502                     }
503                     break;
504
505                 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
506                 case WINED3DFOG_EXP:    args->fog = FOG_EXP;    break;
507                 case WINED3DFOG_EXP2:   args->fog = FOG_EXP2;   break;
508             }
509         } else {
510             args->fog = FOG_OFF;
511         }
512     }
513 }
514
515 GLuint find_gl_pshader(IWineD3DPixelShaderImpl *shader, const struct ps_compile_args *args)
516 {
517     UINT i;
518     DWORD new_size;
519     struct ps_compiled_shader *new_array;
520
521     /* Usually we have very few GL shaders for each d3d shader(just 1 or maybe 2),
522      * so a linear search is more performant than a hashmap or a binary search
523      * (cache coherency etc)
524      */
525     for(i = 0; i < shader->num_gl_shaders; i++) {
526         if(memcmp(&shader->gl_shaders[i].args, args, sizeof(*args)) == 0) {
527             return shader->gl_shaders[i].prgId;
528         }
529     }
530
531     TRACE("No matching GL shader found, compiling a new shader\n");
532     if(shader->shader_array_size == shader->num_gl_shaders) {
533         if(shader->gl_shaders) {
534             new_size = shader->shader_array_size + max(1, shader->shader_array_size / 2);
535             new_array = HeapReAlloc(GetProcessHeap(), 0, shader->gl_shaders,
536                                     new_size * sizeof(*shader->gl_shaders));
537         } else {
538             new_array = HeapAlloc(GetProcessHeap(), 0, sizeof(*shader->gl_shaders));
539             new_size = 1;
540         }
541
542         if(!new_array) {
543             ERR("Out of memory\n");
544             return 0;
545         }
546         shader->gl_shaders = new_array;
547         shader->shader_array_size = new_size;
548     }
549
550     shader->gl_shaders[shader->num_gl_shaders].args = *args;
551     shader->gl_shaders[shader->num_gl_shaders].prgId = pixelshader_compile(shader, args);
552     return shader->gl_shaders[shader->num_gl_shaders++].prgId;
553 }