user32: Change the desktop colour and pattern to match win2k.
[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  *
10  * This library is free software; you can redistribute it and/or
11  * modify it under the terms of the GNU Lesser General Public
12  * License as published by the Free Software Foundation; either
13  * version 2.1 of the License, or (at your option) any later version.
14  *
15  * This library is distributed in the hope that it will be useful,
16  * but WITHOUT ANY WARRANTY; without even the implied warranty of
17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18  * Lesser General Public License for more details.
19  *
20  * You should have received a copy of the GNU Lesser General Public
21  * License along with this library; if not, write to the Free Software
22  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23  */
24
25 #include "config.h"
26
27 #include <math.h>
28 #include <stdio.h>
29
30 #include "wined3d_private.h"
31
32 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
33
34 #define GLINFO_LOCATION ((IWineD3DDeviceImpl *) This->baseShader.device)->adapter->gl_info
35
36 #if 0 /* Must not be 1 in cvs version */
37 # define PSTRACE(A) TRACE A
38 # define TRACE_VSVECTOR(name) TRACE( #name "=(%f, %f, %f, %f)\n", name.x, name.y, name.z, name.w)
39 #else
40 # define PSTRACE(A)
41 # define TRACE_VSVECTOR(name)
42 #endif
43
44 #define GLNAME_REQUIRE_GLSL  ((const char *)1)
45
46 static HRESULT  WINAPI IWineD3DPixelShaderImpl_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, LPVOID *ppobj) {
47     return IWineD3DBaseShaderImpl_QueryInterface((IWineD3DBaseShader *) iface, riid, ppobj);
48 }
49
50 static ULONG  WINAPI IWineD3DPixelShaderImpl_AddRef(IWineD3DPixelShader *iface) {
51     return IWineD3DBaseShaderImpl_AddRef((IWineD3DBaseShader *) iface);
52 }
53
54 static ULONG  WINAPI IWineD3DPixelShaderImpl_Release(IWineD3DPixelShader *iface) {
55     return IWineD3DBaseShaderImpl_Release((IWineD3DBaseShader *) iface);
56 }
57
58 /* *******************************************
59    IWineD3DPixelShader IWineD3DPixelShader parts follow
60    ******************************************* */
61
62 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetParent(IWineD3DPixelShader *iface, IUnknown** parent){
63     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
64
65     *parent = This->parent;
66     IUnknown_AddRef(*parent);
67     TRACE("(%p) : returning %p\n", This, *parent);
68     return WINED3D_OK;
69 }
70
71 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetDevice(IWineD3DPixelShader* iface, IWineD3DDevice **pDevice){
72     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
73     IWineD3DDevice_AddRef(This->baseShader.device);
74     *pDevice = This->baseShader.device;
75     TRACE("(%p) returning %p\n", This, *pDevice);
76     return WINED3D_OK;
77 }
78
79
80 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetFunction(IWineD3DPixelShader* impl, VOID* pData, UINT* pSizeOfData) {
81   IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)impl;
82   TRACE("(%p) : pData(%p), pSizeOfData(%p)\n", This, pData, pSizeOfData);
83
84   if (NULL == pData) {
85     *pSizeOfData = This->baseShader.functionLength;
86     return WINED3D_OK;
87   }
88   if (*pSizeOfData < This->baseShader.functionLength) {
89     /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
90      * than the required size we should write the required size and
91      * return D3DERR_MOREDATA. That's not actually true. */
92     return WINED3DERR_INVALIDCALL;
93   }
94   if (NULL == This->baseShader.function) { /* no function defined */
95     TRACE("(%p) : GetFunction no User Function defined using NULL to %p\n", This, pData);
96     (*(DWORD **) pData) = NULL;
97   } else {
98     if (This->baseShader.functionLength == 0) {
99
100     }
101     TRACE("(%p) : GetFunction copying to %p\n", This, pData);
102     memcpy(pData, This->baseShader.function, This->baseShader.functionLength);
103   }
104   return WINED3D_OK;
105 }
106
107 CONST SHADER_OPCODE IWineD3DPixelShaderImpl_shader_ins[] = {
108     /* Arithmethic */
109     {WINED3DSIO_NOP,  "nop", "NOP", 0, 0, pshader_hw_map2gl, NULL, 0, 0},
110     {WINED3DSIO_MOV,  "mov", "MOV", 1, 2, pshader_hw_map2gl, shader_glsl_mov, 0, 0},
111     {WINED3DSIO_ADD,  "add", "ADD", 1, 3, pshader_hw_map2gl, shader_glsl_arith, 0, 0},
112     {WINED3DSIO_SUB,  "sub", "SUB", 1, 3, pshader_hw_map2gl, shader_glsl_arith, 0, 0},
113     {WINED3DSIO_MAD,  "mad", "MAD", 1, 4, pshader_hw_map2gl, shader_glsl_mad, 0, 0},
114     {WINED3DSIO_MUL,  "mul", "MUL", 1, 3, pshader_hw_map2gl, shader_glsl_arith, 0, 0},
115     {WINED3DSIO_RCP,  "rcp", "RCP",  1, 2, pshader_hw_map2gl, shader_glsl_rcp, 0, 0},
116     {WINED3DSIO_RSQ,  "rsq",  "RSQ", 1, 2, pshader_hw_map2gl, shader_glsl_rsq, 0, 0},
117     {WINED3DSIO_DP3,  "dp3",  "DP3", 1, 3, pshader_hw_map2gl, shader_glsl_dot, 0, 0},
118     {WINED3DSIO_DP4,  "dp4",  "DP4", 1, 3, pshader_hw_map2gl, shader_glsl_dot, 0, 0},
119     {WINED3DSIO_MIN,  "min",  "MIN", 1, 3, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
120     {WINED3DSIO_MAX,  "max",  "MAX", 1, 3, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
121     {WINED3DSIO_SLT,  "slt",  "SLT", 1, 3, pshader_hw_map2gl, shader_glsl_compare, 0, 0},
122     {WINED3DSIO_SGE,  "sge",  "SGE", 1, 3, pshader_hw_map2gl, shader_glsl_compare, 0, 0},
123     {WINED3DSIO_ABS,  "abs",  "ABS", 1, 2, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
124     {WINED3DSIO_EXP,  "exp",  "EX2", 1, 2, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
125     {WINED3DSIO_LOG,  "log",  "LG2", 1, 2, pshader_hw_map2gl, shader_glsl_log, 0, 0},
126     {WINED3DSIO_EXPP, "expp", "EXP", 1, 2, pshader_hw_map2gl, shader_glsl_expp, 0, 0},
127     {WINED3DSIO_LOGP, "logp", "LOG", 1, 2, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
128     {WINED3DSIO_DST,  "dst",  "DST", 1, 3, pshader_hw_map2gl, shader_glsl_dst, 0, 0},
129     {WINED3DSIO_LRP,  "lrp",  "LRP", 1, 4, pshader_hw_map2gl, shader_glsl_lrp, 0, 0},
130     {WINED3DSIO_FRC,  "frc",  "FRC", 1, 2, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
131     {WINED3DSIO_CND,  "cnd",  NULL, 1, 4, pshader_hw_cnd, shader_glsl_cnd, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,4)},
132     {WINED3DSIO_CMP,  "cmp",  NULL, 1, 4, pshader_hw_cmp, shader_glsl_cmp, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(3,0)},
133     {WINED3DSIO_POW,  "pow",  "POW", 1, 3, pshader_hw_map2gl, shader_glsl_pow, 0, 0},
134     {WINED3DSIO_CRS,  "crs",  "XPD", 1, 3, pshader_hw_map2gl, shader_glsl_cross, 0, 0},
135     {WINED3DSIO_NRM,      "nrm",      NULL, 1, 2, shader_hw_nrm, shader_glsl_map2gl, 0, 0},
136     {WINED3DSIO_SINCOS,   "sincos",   NULL, 1, 4, shader_hw_sincos, shader_glsl_sincos, WINED3DPS_VERSION(2,0), WINED3DPS_VERSION(2,1)},
137     {WINED3DSIO_SINCOS,   "sincos",   "SCS", 1, 2, shader_hw_sincos, shader_glsl_sincos, WINED3DPS_VERSION(3,0), -1},
138     {WINED3DSIO_DP2ADD,   "dp2add",   NULL, 1, 4, pshader_hw_dp2add, pshader_glsl_dp2add, WINED3DPS_VERSION(2,0), -1},
139     /* Matrix */
140     {WINED3DSIO_M4x4, "m4x4", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
141     {WINED3DSIO_M4x3, "m4x3", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
142     {WINED3DSIO_M3x4, "m3x4", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
143     {WINED3DSIO_M3x3, "m3x3", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
144     {WINED3DSIO_M3x2, "m3x2", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
145     /* Register declarations */
146     {WINED3DSIO_DCL,      "dcl",      NULL, 0, 2, NULL, NULL, 0, 0},
147     /* Flow control - requires GLSL or software shaders */
148     {WINED3DSIO_REP ,     "rep",      NULL, 0, 1, NULL, shader_glsl_rep,    WINED3DPS_VERSION(2,1), -1},
149     {WINED3DSIO_ENDREP,   "endrep",   NULL, 0, 0, NULL, shader_glsl_end,    WINED3DPS_VERSION(2,1), -1},
150     {WINED3DSIO_IF,       "if",       NULL, 0, 1, NULL, shader_glsl_if,     WINED3DPS_VERSION(2,1), -1},
151     {WINED3DSIO_IFC,      "ifc",      NULL, 0, 2, NULL, shader_glsl_ifc,    WINED3DPS_VERSION(2,1), -1},
152     {WINED3DSIO_ELSE,     "else",     NULL, 0, 0, NULL, shader_glsl_else,   WINED3DPS_VERSION(2,1), -1},
153     {WINED3DSIO_ENDIF,    "endif",    NULL, 0, 0, NULL, shader_glsl_end,    WINED3DPS_VERSION(2,1), -1},
154     {WINED3DSIO_BREAK,    "break",    NULL, 0, 0, NULL, shader_glsl_break,  WINED3DPS_VERSION(2,1), -1},
155     {WINED3DSIO_BREAKC,   "breakc",   NULL, 0, 2, NULL, shader_glsl_breakc, WINED3DPS_VERSION(2,1), -1},
156     {WINED3DSIO_BREAKP,   "breakp",   GLNAME_REQUIRE_GLSL, 0, 1, NULL, NULL, 0, 0},
157     {WINED3DSIO_CALL,     "call",     NULL, 0, 1, NULL, shader_glsl_call, WINED3DPS_VERSION(2,1), -1},
158     {WINED3DSIO_CALLNZ,   "callnz",   NULL, 0, 2, NULL, shader_glsl_callnz, WINED3DPS_VERSION(2,1), -1},
159     {WINED3DSIO_LOOP,     "loop",     NULL, 0, 2, NULL, shader_glsl_loop,   WINED3DPS_VERSION(3,0), -1},
160     {WINED3DSIO_RET,      "ret",      NULL, 0, 0, NULL, NULL,               WINED3DPS_VERSION(2,1), -1},
161     {WINED3DSIO_ENDLOOP,  "endloop",  NULL, 0, 0, NULL, shader_glsl_end,    WINED3DPS_VERSION(3,0), -1},
162     {WINED3DSIO_LABEL,    "label",    NULL, 0, 1, NULL, shader_glsl_label,  WINED3DPS_VERSION(2,1), -1},
163     /* Constant definitions */
164     {WINED3DSIO_DEF,      "def",      "undefined",         1, 5, NULL, NULL, 0, 0},
165     {WINED3DSIO_DEFB,     "defb",     GLNAME_REQUIRE_GLSL, 1, 2, NULL, NULL, 0, 0},
166     {WINED3DSIO_DEFI,     "defi",     GLNAME_REQUIRE_GLSL, 1, 5, NULL, NULL, 0, 0},
167     /* Texture */
168     {WINED3DSIO_TEXCOORD, "texcoord", "undefined", 1, 1, pshader_hw_texcoord, pshader_glsl_texcoord, 0, WINED3DPS_VERSION(1,3)},
169     {WINED3DSIO_TEXCOORD, "texcrd",   "undefined", 1, 2, pshader_hw_texcoord, pshader_glsl_texcoord, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
170     {WINED3DSIO_TEXKILL,  "texkill",  "KIL",       1, 1, pshader_hw_texkill, pshader_glsl_texkill, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(3,0)},
171     {WINED3DSIO_TEX,      "tex",      "undefined", 1, 1, pshader_hw_tex, pshader_glsl_tex, 0, WINED3DPS_VERSION(1,3)},
172     {WINED3DSIO_TEX,      "texld",    "undefined", 1, 2, pshader_hw_tex, pshader_glsl_tex, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
173     {WINED3DSIO_TEX,      "texld",    "undefined", 1, 3, pshader_hw_tex, pshader_glsl_tex, WINED3DPS_VERSION(2,0), -1},
174     {WINED3DSIO_TEXBEM,   "texbem",   "undefined", 1, 2, pshader_hw_texbem, pshader_glsl_texbem, 0, WINED3DPS_VERSION(1,3)},
175     {WINED3DSIO_TEXBEML,  "texbeml",  GLNAME_REQUIRE_GLSL, 1, 2, pshader_hw_texbem, pshader_glsl_texbem, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
176     {WINED3DSIO_TEXREG2AR,"texreg2ar","undefined", 1, 2, pshader_hw_texreg2ar, pshader_glsl_texreg2ar, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
177     {WINED3DSIO_TEXREG2GB,"texreg2gb","undefined", 1, 2, pshader_hw_texreg2gb, pshader_glsl_texreg2gb, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
178     {WINED3DSIO_TEXREG2RGB,   "texreg2rgb",   "undefined", 1, 2, pshader_hw_texreg2rgb, pshader_glsl_texreg2rgb, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
179     {WINED3DSIO_TEXM3x2PAD,   "texm3x2pad",   "undefined", 1, 2, pshader_hw_texm3x2pad, pshader_glsl_texm3x2pad, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
180     {WINED3DSIO_TEXM3x2TEX,   "texm3x2tex",   "undefined", 1, 2, pshader_hw_texm3x2tex, pshader_glsl_texm3x2tex, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
181     {WINED3DSIO_TEXM3x3PAD,   "texm3x3pad",   "undefined", 1, 2, pshader_hw_texm3x3pad, pshader_glsl_texm3x3pad, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
182     {WINED3DSIO_TEXM3x3DIFF,  "texm3x3diff",  GLNAME_REQUIRE_GLSL, 1, 2, NULL, NULL, WINED3DPS_VERSION(0,0), WINED3DPS_VERSION(0,0)},
183     {WINED3DSIO_TEXM3x3SPEC,  "texm3x3spec",  "undefined", 1, 3, pshader_hw_texm3x3spec, pshader_glsl_texm3x3spec, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
184     {WINED3DSIO_TEXM3x3VSPEC, "texm3x3vspec",  "undefined", 1, 2, pshader_hw_texm3x3vspec, pshader_glsl_texm3x3vspec, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
185     {WINED3DSIO_TEXM3x3TEX,   "texm3x3tex",   "undefined", 1, 2, pshader_hw_texm3x3tex, pshader_glsl_texm3x3tex, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
186     {WINED3DSIO_TEXDP3TEX,    "texdp3tex",    NULL, 1, 2, pshader_hw_texdp3tex, pshader_glsl_texdp3tex, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
187     {WINED3DSIO_TEXM3x2DEPTH, "texm3x2depth", GLNAME_REQUIRE_GLSL, 1, 2, pshader_hw_texm3x2depth, pshader_glsl_texm3x2depth, WINED3DPS_VERSION(1,3), WINED3DPS_VERSION(1,3)},
188     {WINED3DSIO_TEXDP3,   "texdp3",   NULL, 1, 2, pshader_hw_texdp3, pshader_glsl_texdp3, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
189     {WINED3DSIO_TEXM3x3,  "texm3x3",  NULL, 1, 2, pshader_hw_texm3x3, pshader_glsl_texm3x3, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
190     {WINED3DSIO_TEXDEPTH, "texdepth", NULL, 1, 1, pshader_hw_texdepth, pshader_glsl_texdepth, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
191     {WINED3DSIO_BEM,      "bem",      "undefined",         1, 3, pshader_hw_bem, pshader_glsl_bem, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
192     {WINED3DSIO_DSX,      "dsx",      NULL, 1, 2, NULL, shader_glsl_map2gl, WINED3DPS_VERSION(2,1), -1},
193     {WINED3DSIO_DSY,      "dsy",      NULL, 1, 2, NULL, shader_glsl_map2gl, WINED3DPS_VERSION(2,1), -1},
194     {WINED3DSIO_TEXLDD,   "texldd",   GLNAME_REQUIRE_GLSL, 1, 5, NULL, NULL, WINED3DPS_VERSION(2,1), -1},
195     {WINED3DSIO_SETP,     "setp",     GLNAME_REQUIRE_GLSL, 1, 3, NULL, NULL, 0, 0},
196     {WINED3DSIO_TEXLDL,   "texldl",   NULL, 1, 3, NULL, shader_glsl_texldl, WINED3DPS_VERSION(3,0), -1},
197     {WINED3DSIO_PHASE,    "phase",    GLNAME_REQUIRE_GLSL, 0, 0, NULL, NULL, 0, 0},
198     {0,               NULL,       NULL,   0, 0, NULL,            NULL, 0, 0}
199 };
200
201 static void pshader_set_limits(
202       IWineD3DPixelShaderImpl *This) { 
203
204       This->baseShader.limits.attributes = 0;
205       This->baseShader.limits.address = 0;
206       This->baseShader.limits.packed_output = 0;
207
208       switch (This->baseShader.hex_version) {
209           case WINED3DPS_VERSION(1,0):
210           case WINED3DPS_VERSION(1,1):
211           case WINED3DPS_VERSION(1,2):
212           case WINED3DPS_VERSION(1,3): 
213                    This->baseShader.limits.temporary = 2;
214                    This->baseShader.limits.constant_float = 8;
215                    This->baseShader.limits.constant_int = 0;
216                    This->baseShader.limits.constant_bool = 0;
217                    This->baseShader.limits.texcoord = 4;
218                    This->baseShader.limits.sampler = 4;
219                    This->baseShader.limits.packed_input = 0;
220                    This->baseShader.limits.label = 0;
221                    break;
222
223           case WINED3DPS_VERSION(1,4):
224                    This->baseShader.limits.temporary = 6;
225                    This->baseShader.limits.constant_float = 8;
226                    This->baseShader.limits.constant_int = 0;
227                    This->baseShader.limits.constant_bool = 0;
228                    This->baseShader.limits.texcoord = 6;
229                    This->baseShader.limits.sampler = 6;
230                    This->baseShader.limits.packed_input = 0;
231                    This->baseShader.limits.label = 0;
232                    break;
233                
234           /* FIXME: temporaries must match D3DPSHADERCAPS2_0.NumTemps */ 
235           case WINED3DPS_VERSION(2,0):
236                    This->baseShader.limits.temporary = 32;
237                    This->baseShader.limits.constant_float = 32;
238                    This->baseShader.limits.constant_int = 16;
239                    This->baseShader.limits.constant_bool = 16;
240                    This->baseShader.limits.texcoord = 8;
241                    This->baseShader.limits.sampler = 16;
242                    This->baseShader.limits.packed_input = 0;
243                    break;
244
245           case WINED3DPS_VERSION(2,1):
246                    This->baseShader.limits.temporary = 32;
247                    This->baseShader.limits.constant_float = 32;
248                    This->baseShader.limits.constant_int = 16;
249                    This->baseShader.limits.constant_bool = 16;
250                    This->baseShader.limits.texcoord = 8;
251                    This->baseShader.limits.sampler = 16;
252                    This->baseShader.limits.packed_input = 0;
253                    This->baseShader.limits.label = 16;
254                    break;
255
256           case WINED3DPS_VERSION(3,0):
257                    This->baseShader.limits.temporary = 32;
258                    This->baseShader.limits.constant_float = 224;
259                    This->baseShader.limits.constant_int = 16;
260                    This->baseShader.limits.constant_bool = 16;
261                    This->baseShader.limits.texcoord = 0;
262                    This->baseShader.limits.sampler = 16;
263                    This->baseShader.limits.packed_input = 12;
264                    This->baseShader.limits.label = 16; /* FIXME: 2048 */
265                    break;
266
267           default: 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 = 0;
275                    FIXME("Unrecognized pixel shader version %#x\n", 
276                        This->baseShader.hex_version);
277       }
278 }
279
280 /** Generate a pixel shader string using either GL_FRAGMENT_PROGRAM_ARB
281     or GLSL and send it to the card */
282 static inline VOID IWineD3DPixelShaderImpl_GenerateShader(
283     IWineD3DPixelShader *iface,
284     shader_reg_maps* reg_maps,
285     CONST DWORD *pFunction) {
286
287     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
288     SHADER_BUFFER buffer;
289
290 #if 0 /* FIXME: Use the buffer that is held by the device, this is ok since fixups will be skipped for software shaders
291         it also requires entering a critical section but cuts down the runtime footprint of wined3d and any memory fragmentation that may occur... */
292     if (This->device->fixupVertexBufferSize < SHADER_PGMSIZE) {
293         HeapFree(GetProcessHeap(), 0, This->fixupVertexBuffer);
294         This->fixupVertexBuffer = HeapAlloc(GetProcessHeap() , 0, SHADER_PGMSIZE);
295         This->fixupVertexBufferSize = PGMSIZE;
296         This->fixupVertexBuffer[0] = 0;
297     }
298     buffer.buffer = This->device->fixupVertexBuffer;
299 #else
300     buffer.buffer = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, SHADER_PGMSIZE); 
301 #endif
302     buffer.bsize = 0;
303     buffer.lineNo = 0;
304     buffer.newline = TRUE;
305
306     if (This->baseShader.shader_mode == SHADER_GLSL) {
307
308         /* Create the hw GLSL shader object and assign it as the baseShader.prgId */
309         GLhandleARB shader_obj = GL_EXTCALL(glCreateShaderObjectARB(GL_FRAGMENT_SHADER_ARB));
310
311         if (GL_SUPPORT(ARB_DRAW_BUFFERS)) {
312             shader_addline(&buffer, "#extension GL_ARB_draw_buffers : enable\n");
313         }
314
315         /* Base Declarations */
316         shader_generate_glsl_declarations( (IWineD3DBaseShader*) This, reg_maps, &buffer, &GLINFO_LOCATION);
317
318         /* Pack 3.0 inputs */
319         if (This->baseShader.hex_version >= WINED3DPS_VERSION(3,0)) {
320
321             if(((IWineD3DDeviceImpl *) This->baseShader.device)->strided_streams.u.s.position_transformed) {
322                 This->vertexprocessing = pretransformed;
323                 pshader_glsl_input_pack(&buffer, This->semantics_in, iface);
324             } else if(!use_vs((IWineD3DDeviceImpl *) This->baseShader.device)) {
325                 This->vertexprocessing = fixedfunction;
326                 pshader_glsl_input_pack(&buffer, This->semantics_in, iface);
327             } else {
328                 This->vertexprocessing = vertexshader;
329             }
330         }
331
332         /* Base Shader Body */
333         shader_generate_main( (IWineD3DBaseShader*) This, &buffer, reg_maps, pFunction);
334
335         /* Pixel shaders < 2.0 place the resulting color in R0 implicitly */
336         if (This->baseShader.hex_version < WINED3DPS_VERSION(2,0)) {
337             /* Some older cards like GeforceFX ones don't support multiple buffers, so also not gl_FragData */
338             if(GL_SUPPORT(ARB_DRAW_BUFFERS))
339                 shader_addline(&buffer, "gl_FragData[0] = R0;\n");
340             else
341                 shader_addline(&buffer, "gl_FragColor = R0;\n");
342         }
343
344         /* Pixel shader < 3.0 do not replace the fog stage.
345          * This implements linear fog computation and blending.
346          * TODO: non linear fog
347          * NOTE: gl_Fog.start and gl_Fog.end don't hold fog start s and end e but
348          * -1/(e-s) and e/(e-s) respectively.
349          */
350         if(This->baseShader.hex_version < WINED3DPS_VERSION(3,0)) {
351             shader_addline(&buffer, "float Fog = clamp(gl_FogFragCoord * gl_Fog.start + gl_Fog.end, 0.0, 1.0);\n");
352             if(GL_SUPPORT(ARB_DRAW_BUFFERS))
353                 shader_addline(&buffer, "gl_FragData[0].xyz = mix(gl_Fog.color.xyz, gl_FragData[0].xyz, Fog);\n");
354             else
355                 shader_addline(&buffer, "gl_FragColor.xyz = mix(gl_Fog.color.xyz, gl_FragColor.xyz, Fog);\n");
356         }
357         if(This->srgb_enabled) {
358             const char *fragcolor;
359
360             if(GL_SUPPORT(ARB_DRAW_BUFFERS)) {
361                 fragcolor = "gl_FragData[0]";
362             } else {
363                 fragcolor = "gl_FragColor";
364             }
365             shader_addline(&buffer, "tmp0.xyz = pow(%s.xyz, vec3(%f, %f, %f)) * vec3(%f, %f, %f) - vec3(%f, %f, %f);\n",
366                             fragcolor, srgb_pow, srgb_pow, srgb_pow, srgb_mul_high, srgb_mul_high, srgb_mul_high,
367                             srgb_sub_high, srgb_sub_high, srgb_sub_high);
368             shader_addline(&buffer, "tmp1.xyz = %s.xyz * srgb_mul_low.xyz;\n", fragcolor);
369             shader_addline(&buffer, "%s.x = %s.x < srgb_comparison.x ? tmp1.x : tmp0.x;\n", fragcolor, fragcolor);
370             shader_addline(&buffer, "%s.y = %s.y < srgb_comparison.y ? tmp1.y : tmp0.y;\n", fragcolor, fragcolor);
371             shader_addline(&buffer, "%s.z = %s.z < srgb_comparison.z ? tmp1.z : tmp0.z;\n", fragcolor, fragcolor);
372             shader_addline(&buffer, "%s = clamp(%s, 0.0, 1.0);\n", fragcolor, fragcolor);
373         }
374
375         shader_addline(&buffer, "}\n");
376
377         TRACE("Compiling shader object %u\n", shader_obj);
378         GL_EXTCALL(glShaderSourceARB(shader_obj, 1, (const char**)&buffer.buffer, NULL));
379         GL_EXTCALL(glCompileShaderARB(shader_obj));
380         print_glsl_info_log(&GLINFO_LOCATION, shader_obj);
381
382         /* Store the shader object */
383         This->baseShader.prgId = shader_obj;
384
385     } else if (This->baseShader.shader_mode == SHADER_ARB) {
386         /*  Create the hw ARB shader */
387         shader_addline(&buffer, "!!ARBfp1.0\n");
388
389         shader_addline(&buffer, "TEMP TMP;\n");     /* Used in matrix ops */
390         shader_addline(&buffer, "TEMP TMP2;\n");    /* Used in matrix ops */
391         shader_addline(&buffer, "TEMP TA;\n");      /* Used for modifiers */
392         shader_addline(&buffer, "TEMP TB;\n");      /* Used for modifiers */
393         shader_addline(&buffer, "TEMP TC;\n");      /* Used for modifiers */
394         shader_addline(&buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
395         shader_addline(&buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
396         shader_addline(&buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
397
398         /* Base Declarations */
399         shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, &buffer, &GLINFO_LOCATION);
400
401         /* We need two variables for fog blending */
402         shader_addline(&buffer, "TEMP TMP_FOG;\n");
403         if (This->baseShader.hex_version >= WINED3DPS_VERSION(2,0)) {
404             shader_addline(&buffer, "TEMP TMP_COLOR;\n");
405         }
406
407         /* Base Shader Body */
408         shader_generate_main( (IWineD3DBaseShader*) This, &buffer, reg_maps, pFunction);
409
410         /* calculate fog and blend it
411          * NOTE: state.fog.params.y and state.fog.params.z don't hold fog start s and end e but
412          * -1/(e-s) and e/(e-s) respectively.
413          */
414         shader_addline(&buffer, "MAD_SAT TMP_FOG, fragment.fogcoord, state.fog.params.y, state.fog.params.z;\n");
415
416         if(This->srgb_enabled) {
417             if (This->baseShader.hex_version < WINED3DPS_VERSION(2,0)) {
418                 shader_addline(&buffer, "LRP TMP_COLOR.rgb, TMP_FOG.x, R0, state.fog.color;\n");
419                 shader_addline(&buffer, "MOV result.color.a, R0.a;\n");
420             } else {
421                 shader_addline(&buffer, "LRP TMP_COLOR.rgb, TMP_FOG.x, TMP_COLOR, state.fog.color;\n");
422                 shader_addline(&buffer, "MOV result.color.a, TMP_COLOR.a;\n");
423             }
424             /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
425
426             /* Calculate the > 0.0031308 case */
427             shader_addline(&buffer, "POW TMP.x, TMP_COLOR.x, srgb_pow.x;\n");
428             shader_addline(&buffer, "POW TMP.y, TMP_COLOR.y, srgb_pow.y;\n");
429             shader_addline(&buffer, "POW TMP.z, TMP_COLOR.z, srgb_pow.z;\n");
430             shader_addline(&buffer, "MUL TMP, TMP, srgb_mul_hi;\n");
431             shader_addline(&buffer, "SUB TMP, TMP, srgb_sub_hi;\n");
432             /* Calculate the < case */
433             shader_addline(&buffer, "MUL TMP2, srgb_mul_low, TMP_COLOR;\n");
434             /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
435             shader_addline(&buffer, "SLT TA, srgb_comparison, TMP_COLOR;\n");
436             shader_addline(&buffer, "SGE TB, srgb_comparison, TMP_COLOR;\n");
437             /* Store the components > 0.0031308 in the destination */
438             shader_addline(&buffer, "MUL TMP_COLOR, TMP, TA;\n");
439             /* Add the components that are < 0.0031308 */
440             shader_addline(&buffer, "MAD result.color.xyz, TMP2, TB, TMP_COLOR;\n");
441             /* [0.0;1.0] clamping. Not needed, this is done implicitly */
442         } else {
443             if (This->baseShader.hex_version < WINED3DPS_VERSION(2,0)) {
444                 shader_addline(&buffer, "LRP result.color.rgb, TMP_FOG.x, R0, state.fog.color;\n");
445                 shader_addline(&buffer, "MOV result.color.a, R0.a;\n");
446             } else {
447                 shader_addline(&buffer, "LRP result.color.rgb, TMP_FOG.x, TMP_COLOR, state.fog.color;\n");
448                 shader_addline(&buffer, "MOV result.color.a, TMP_COLOR.a;\n");
449             }
450         }
451
452         shader_addline(&buffer, "END\n"); 
453
454         /* TODO: change to resource.glObjectHandle or something like that */
455         GL_EXTCALL(glGenProgramsARB(1, &This->baseShader.prgId));
456
457         TRACE("Creating a hw pixel shader, prg=%d\n", This->baseShader.prgId);
458         GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, This->baseShader.prgId));
459
460         TRACE("Created hw pixel shader, prg=%d\n", This->baseShader.prgId);
461         /* Create the program and check for errors */
462         GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
463             buffer.bsize, buffer.buffer));
464
465         if (glGetError() == GL_INVALID_OPERATION) {
466             GLint errPos;
467             glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
468             FIXME("HW PixelShader Error at position %d: %s\n",
469                   errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
470             This->baseShader.prgId = -1;
471         }
472     }
473
474     This->needsbumpmat = reg_maps->bumpmat;
475
476 #if 1 /* if were using the data buffer of device then we don't need to free it */
477   HeapFree(GetProcessHeap(), 0, buffer.buffer);
478 #endif
479 }
480
481 static HRESULT WINAPI IWineD3DPixelShaderImpl_SetFunction(IWineD3DPixelShader *iface, CONST DWORD *pFunction) {
482
483     IWineD3DPixelShaderImpl *This =(IWineD3DPixelShaderImpl *)iface;
484     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
485
486     TRACE("(%p) : pFunction %p\n", iface, pFunction);
487
488     /* First pass: trace shader */
489     shader_trace_init((IWineD3DBaseShader*) This, pFunction);
490     pshader_set_limits(This);
491
492     /* Initialize immediate constant lists */
493     list_init(&This->baseShader.constantsF);
494     list_init(&This->baseShader.constantsB);
495     list_init(&This->baseShader.constantsI);
496
497     if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) > 1) {
498         shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
499         HRESULT hr;
500         unsigned int i, j, highest_reg_used = 0, num_regs_used = 0;
501
502         /* Second pass: figure out which registers are used, what the semantics are, etc.. */
503         memset(reg_maps, 0, sizeof(shader_reg_maps));
504         hr = shader_get_registers_used((IWineD3DBaseShader*) This, reg_maps,
505             This->semantics_in, NULL, pFunction, NULL);
506         if (FAILED(hr)) return hr;
507         /* FIXME: validate reg_maps against OpenGL */
508
509         for(i = 0; i < MAX_REG_INPUT; i++) {
510             if(This->input_reg_used[i]) {
511                 num_regs_used++;
512                 highest_reg_used = i;
513             }
514         }
515
516         /* Don't do any register mapping magic if it is not needed, or if we can't
517          * achive anything anyway
518          */
519         if(highest_reg_used < (GL_LIMITS(glsl_varyings) / 4) ||
520            num_regs_used > (GL_LIMITS(glsl_varyings) / 4) ) {
521             if(num_regs_used > (GL_LIMITS(glsl_varyings) / 4)) {
522                 /* This happens with relative addressing. The input mapper function
523                  * warns about this if the higher registers are declared too, so
524                  * don't write a FIXME here
525                  */
526                 WARN("More varying registers used than supported\n");
527             }
528
529             for(i = 0; i < MAX_REG_INPUT; i++) {
530                 This->input_reg_map[i] = i;
531             }
532         } else {
533             j = 0;
534             for(i = 0; i < MAX_REG_INPUT; i++) {
535                 if(This->input_reg_used[i]) {
536                     This->input_reg_map[i] = j;
537                     j++;
538                 } else {
539                     This->input_reg_map[i] = -1;
540                 }
541             }
542         }
543     }
544     This->baseShader.load_local_constsF = FALSE;
545
546     This->baseShader.shader_mode = deviceImpl->ps_selected_mode;
547
548     TRACE("(%p) : Copying the function\n", This);
549     if (NULL != pFunction) {
550         void *function;
551
552         function = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, This->baseShader.functionLength);
553         if (!function) return E_OUTOFMEMORY;
554         memcpy(function, pFunction, This->baseShader.functionLength);
555         This->baseShader.function = function;
556     } else {
557         This->baseShader.function = NULL;
558     }
559
560     return WINED3D_OK;
561 }
562
563 static HRESULT WINAPI IWineD3DPixelShaderImpl_CompileShader(IWineD3DPixelShader *iface) {
564
565     IWineD3DPixelShaderImpl *This =(IWineD3DPixelShaderImpl *)iface;
566     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
567     CONST DWORD *function = This->baseShader.function;
568     UINT i, sampler;
569     IWineD3DBaseTextureImpl *texture;
570
571     TRACE("(%p) : function %p\n", iface, function);
572
573     /* We're already compiled, but check if any of the hardcoded stateblock assumptions
574      * changed.
575      */
576     if (This->baseShader.is_compiled) {
577         char srgbenabled = deviceImpl->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
578         for(i = 0; i < This->baseShader.num_sampled_samplers; i++) {
579             sampler = This->baseShader.sampled_samplers[i];
580             texture = (IWineD3DBaseTextureImpl *) deviceImpl->stateBlock->textures[sampler];
581             if(texture && texture->baseTexture.shader_conversion_group != This->baseShader.sampled_format[sampler]) {
582                 WARN("Recompiling shader %p due to format change on sampler %d\n", This, sampler);
583                 WARN("Old format group %s, new is %s\n",
584                      debug_d3dformat(This->baseShader.sampled_format[sampler]),
585                      debug_d3dformat(texture->baseTexture.shader_conversion_group));
586                 goto recompile;
587             }
588         }
589
590         /* TODO: Check projected textures */
591         /* TODO: Check texture types(2D, Cube, 3D) */
592
593         if(srgbenabled != This->srgb_enabled && This->srgb_mode_hardcoded) {
594             WARN("Recompiling shader because srgb correction is different and hardcoded\n");
595             goto recompile;
596         }
597         if(This->baseShader.reg_maps.vpos && !This->vpos_uniform) {
598             if(This->render_offscreen != deviceImpl->render_offscreen ||
599                This->height != ((IWineD3DSurfaceImpl *) deviceImpl->render_targets[0])->currentDesc.Height) {
600                 WARN("Recompiling shader because vpos is used, hard compiled and changed\n");
601                 goto recompile;
602             }
603         }
604         if(This->baseShader.reg_maps.usesdsy && !This->vpos_uniform) {
605             if(This->render_offscreen ? 0 : 1 != deviceImpl->render_offscreen ? 0 : 1) {
606                 WARN("Recompiling shader because dsy is used, hard compiled and render_offscreen changed\n");
607                 goto recompile;
608             }
609         }
610         if(This->baseShader.hex_version >= WINED3DPS_VERSION(3,0)) {
611             if(((IWineD3DDeviceImpl *) This->baseShader.device)->strided_streams.u.s.position_transformed &&
612                  This->vertexprocessing != pretransformed) {
613                 WARN("Recompiling shader because pretransformed vertices are provided, which wasn't the case before\n");
614                 goto recompile;
615             } else if(!use_vs((IWineD3DDeviceImpl *) This->baseShader.device) &&
616                        This->vertexprocessing != fixedfunction) {
617                 WARN("Recompiling shader because fixed function vp is in use, which wasn't the case before\n");
618                 goto recompile;
619             } else if(This->vertexprocessing != vertexshader) {
620                 WARN("Recompiling shader because vertex shaders are in use, which wasn't the case before\n");
621                 goto recompile;
622             }
623         }
624
625         return WINED3D_OK;
626
627         recompile:
628         if(This->baseShader.recompile_count > 50) {
629             FIXME("Shader %p recompiled more than 50 times\n", This);
630         } else {
631             This->baseShader.recompile_count++;
632         }
633
634         deviceImpl->shader_backend->shader_destroy((IWineD3DBaseShader *) iface);
635     }
636
637     /* We don't need to compile */
638     if (!function) {
639         This->baseShader.is_compiled = TRUE;
640         return WINED3D_OK;
641     }
642
643     if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1) {
644         shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
645         HRESULT hr;
646
647         /* Second pass: figure out which registers are used, what the semantics are, etc.. */
648         memset(reg_maps, 0, sizeof(shader_reg_maps));
649         hr = shader_get_registers_used((IWineD3DBaseShader*)This, reg_maps,
650             This->semantics_in, NULL, This->baseShader.function, deviceImpl->stateBlock);
651         if (FAILED(hr)) return hr;
652         /* FIXME: validate reg_maps against OpenGL */
653     }
654
655     /* Reset fields tracking stateblock values being hardcoded in the shader */
656     This->baseShader.num_sampled_samplers = 0;
657
658     /* Generate the HW shader */
659     TRACE("(%p) : Generating hardware program\n", This);
660     IWineD3DPixelShaderImpl_GenerateShader(iface, &This->baseShader.reg_maps, function);
661
662     This->baseShader.is_compiled = TRUE;
663
664     return WINED3D_OK;
665 }
666
667 const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
668 {
669     /*** IUnknown methods ***/
670     IWineD3DPixelShaderImpl_QueryInterface,
671     IWineD3DPixelShaderImpl_AddRef,
672     IWineD3DPixelShaderImpl_Release,
673     /*** IWineD3DBase methods ***/
674     IWineD3DPixelShaderImpl_GetParent,
675     /*** IWineD3DBaseShader methods ***/
676     IWineD3DPixelShaderImpl_SetFunction,
677     IWineD3DPixelShaderImpl_CompileShader,
678     /*** IWineD3DPixelShader methods ***/
679     IWineD3DPixelShaderImpl_GetDevice,
680     IWineD3DPixelShaderImpl_GetFunction
681 };