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     return IWineD3DBaseShaderImpl_QueryInterface((IWineD3DBaseShader *) iface, riid, ppobj);
  41 }
  42
  43 static ULONG  WINAPI IWineD3DPixelShaderImpl_AddRef(IWineD3DPixelShader *iface) {
  44     return IWineD3DBaseShaderImpl_AddRef((IWineD3DBaseShader *) iface);
  45 }
  46
  47 static ULONG  WINAPI IWineD3DPixelShaderImpl_Release(IWineD3DPixelShader *iface) {
  48     return IWineD3DBaseShaderImpl_Release((IWineD3DBaseShader *) iface);
  49 }
  50
  51 /* *******************************************
  52    IWineD3DPixelShader IWineD3DPixelShader parts follow
  53    ******************************************* */
  54
  55 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetParent(IWineD3DPixelShader *iface, IUnknown** parent){
  56     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
  57
  58     *parent = This->parent;
  59     IUnknown_AddRef(*parent);
  60     TRACE("(%p) : returning %p\n", This, *parent);
  61     return WINED3D_OK;
  62 }
  63
  64 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetDevice(IWineD3DPixelShader* iface, IWineD3DDevice **pDevice){
  65     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
  66     IWineD3DDevice_AddRef(This->baseShader.device);
  67     *pDevice = This->baseShader.device;
  68     TRACE("(%p) returning %p\n", This, *pDevice);
  69     return WINED3D_OK;
  70 }
  71
  72
  73 static HRESULT  WINAPI IWineD3DPixelShaderImpl_GetFunction(IWineD3DPixelShader* impl, VOID* pData, UINT* pSizeOfData) {
  74   IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)impl;
  75   TRACE("(%p) : pData(%p), pSizeOfData(%p)\n", This, pData, pSizeOfData);
  76
  77   if (NULL == pData) {
  78     *pSizeOfData = This->baseShader.functionLength;
  79     return WINED3D_OK;
  80   }
  81   if (*pSizeOfData < This->baseShader.functionLength) {
  82     /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
  83      * than the required size we should write the required size and
  84      * return D3DERR_MOREDATA. That's not actually true. */
  85     return WINED3DERR_INVALIDCALL;
  86   }
  87   if (NULL == This->baseShader.function) { /* no function defined */
  88     TRACE("(%p) : GetFunction no User Function defined using NULL to %p\n", This, pData);
  89     (*(DWORD **) pData) = NULL;
  90   } else {
  91     if (This->baseShader.functionLength == 0) {
  92
  93     }
  94     TRACE("(%p) : GetFunction copying to %p\n", This, pData);
  95     memcpy(pData, This->baseShader.function, This->baseShader.functionLength);
  96   }
  97   return WINED3D_OK;
  98 }
  99
 100 CONST SHADER_OPCODE IWineD3DPixelShaderImpl_shader_ins[] = {
 101     /* Arithmetic */
 102     {WINED3DSIO_NOP,  "nop", "NOP", 0, 0, pshader_hw_map2gl, NULL, 0, 0},
 103     {WINED3DSIO_MOV,  "mov", "MOV", 1, 2, pshader_hw_map2gl, shader_glsl_mov, 0, 0},
 104     {WINED3DSIO_ADD,  "add", "ADD", 1, 3, pshader_hw_map2gl, shader_glsl_arith, 0, 0},
 105     {WINED3DSIO_SUB,  "sub", "SUB", 1, 3, pshader_hw_map2gl, shader_glsl_arith, 0, 0},
 106     {WINED3DSIO_MAD,  "mad", "MAD", 1, 4, pshader_hw_map2gl, shader_glsl_mad, 0, 0},
 107     {WINED3DSIO_MUL,  "mul", "MUL", 1, 3, pshader_hw_map2gl, shader_glsl_arith, 0, 0},
 108     {WINED3DSIO_RCP,  "rcp", "RCP",  1, 2, pshader_hw_map2gl, shader_glsl_rcp, 0, 0},
 109     {WINED3DSIO_RSQ,  "rsq",  "RSQ", 1, 2, pshader_hw_map2gl, shader_glsl_rsq, 0, 0},
 110     {WINED3DSIO_DP3,  "dp3",  "DP3", 1, 3, pshader_hw_map2gl, shader_glsl_dot, 0, 0},
 111     {WINED3DSIO_DP4,  "dp4",  "DP4", 1, 3, pshader_hw_map2gl, shader_glsl_dot, 0, 0},
 112     {WINED3DSIO_MIN,  "min",  "MIN", 1, 3, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
 113     {WINED3DSIO_MAX,  "max",  "MAX", 1, 3, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
 114     {WINED3DSIO_SLT,  "slt",  "SLT", 1, 3, pshader_hw_map2gl, shader_glsl_compare, 0, 0},
 115     {WINED3DSIO_SGE,  "sge",  "SGE", 1, 3, pshader_hw_map2gl, shader_glsl_compare, 0, 0},
 116     {WINED3DSIO_ABS,  "abs",  "ABS", 1, 2, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
 117     {WINED3DSIO_EXP,  "exp",  "EX2", 1, 2, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
 118     {WINED3DSIO_LOG,  "log",  "LG2", 1, 2, pshader_hw_map2gl, shader_glsl_log, 0, 0},
 119     {WINED3DSIO_EXPP, "expp", "EXP", 1, 2, pshader_hw_map2gl, shader_glsl_expp, 0, 0},
 120     {WINED3DSIO_LOGP, "logp", "LOG", 1, 2, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
 121     {WINED3DSIO_DST,  "dst",  "DST", 1, 3, pshader_hw_map2gl, shader_glsl_dst, 0, 0},
 122     {WINED3DSIO_LRP,  "lrp",  "LRP", 1, 4, pshader_hw_map2gl, shader_glsl_lrp, 0, 0},
 123     {WINED3DSIO_FRC,  "frc",  "FRC", 1, 2, pshader_hw_map2gl, shader_glsl_map2gl, 0, 0},
 124     {WINED3DSIO_CND,  "cnd",  NULL, 1, 4, pshader_hw_cnd, shader_glsl_cnd, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,4)},
 125     {WINED3DSIO_CMP,  "cmp",  NULL, 1, 4, pshader_hw_cmp, shader_glsl_cmp, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(3,0)},
 126     {WINED3DSIO_POW,  "pow",  "POW", 1, 3, pshader_hw_map2gl, shader_glsl_pow, 0, 0},
 127     {WINED3DSIO_CRS,  "crs",  "XPD", 1, 3, pshader_hw_map2gl, shader_glsl_cross, 0, 0},
 128     {WINED3DSIO_NRM,      "nrm",      NULL, 1, 2, shader_hw_nrm, shader_glsl_map2gl, 0, 0},
 129     {WINED3DSIO_SINCOS,   "sincos",   NULL, 1, 4, shader_hw_sincos, shader_glsl_sincos, WINED3DPS_VERSION(2,0), WINED3DPS_VERSION(2,1)},
 130     {WINED3DSIO_SINCOS,   "sincos",   "SCS", 1, 2, shader_hw_sincos, shader_glsl_sincos, WINED3DPS_VERSION(3,0), -1},
 131     {WINED3DSIO_DP2ADD,   "dp2add",   NULL, 1, 4, pshader_hw_dp2add, pshader_glsl_dp2add, WINED3DPS_VERSION(2,0), -1},
 132     /* Matrix */
 133     {WINED3DSIO_M4x4, "m4x4", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
 134     {WINED3DSIO_M4x3, "m4x3", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
 135     {WINED3DSIO_M3x4, "m3x4", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
 136     {WINED3DSIO_M3x3, "m3x3", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
 137     {WINED3DSIO_M3x2, "m3x2", "undefined", 1, 3, shader_hw_mnxn, shader_glsl_mnxn, 0, 0},
 138     /* Register declarations */
 139     {WINED3DSIO_DCL,      "dcl",      NULL, 0, 2, NULL, NULL, 0, 0},
 140     /* Flow control - requires GLSL or software shaders */
 141     {WINED3DSIO_REP ,     "rep",      NULL, 0, 1, NULL, shader_glsl_rep,    WINED3DPS_VERSION(2,1), -1},
 142     {WINED3DSIO_ENDREP,   "endrep",   NULL, 0, 0, NULL, shader_glsl_end,    WINED3DPS_VERSION(2,1), -1},
 143     {WINED3DSIO_IF,       "if",       NULL, 0, 1, NULL, shader_glsl_if,     WINED3DPS_VERSION(2,1), -1},
 144     {WINED3DSIO_IFC,      "ifc",      NULL, 0, 2, NULL, shader_glsl_ifc,    WINED3DPS_VERSION(2,1), -1},
 145     {WINED3DSIO_ELSE,     "else",     NULL, 0, 0, NULL, shader_glsl_else,   WINED3DPS_VERSION(2,1), -1},
 146     {WINED3DSIO_ENDIF,    "endif",    NULL, 0, 0, NULL, shader_glsl_end,    WINED3DPS_VERSION(2,1), -1},
 147     {WINED3DSIO_BREAK,    "break",    NULL, 0, 0, NULL, shader_glsl_break,  WINED3DPS_VERSION(2,1), -1},
 148     {WINED3DSIO_BREAKC,   "breakc",   NULL, 0, 2, NULL, shader_glsl_breakc, WINED3DPS_VERSION(2,1), -1},
 149     {WINED3DSIO_BREAKP,   "breakp",   GLNAME_REQUIRE_GLSL, 0, 1, NULL, NULL, 0, 0},
 150     {WINED3DSIO_CALL,     "call",     NULL, 0, 1, NULL, shader_glsl_call, WINED3DPS_VERSION(2,1), -1},
 151     {WINED3DSIO_CALLNZ,   "callnz",   NULL, 0, 2, NULL, shader_glsl_callnz, WINED3DPS_VERSION(2,1), -1},
 152     {WINED3DSIO_LOOP,     "loop",     NULL, 0, 2, NULL, shader_glsl_loop,   WINED3DPS_VERSION(3,0), -1},
 153     {WINED3DSIO_RET,      "ret",      NULL, 0, 0, NULL, NULL,               WINED3DPS_VERSION(2,1), -1},
 154     {WINED3DSIO_ENDLOOP,  "endloop",  NULL, 0, 0, NULL, shader_glsl_end,    WINED3DPS_VERSION(3,0), -1},
 155     {WINED3DSIO_LABEL,    "label",    NULL, 0, 1, NULL, shader_glsl_label,  WINED3DPS_VERSION(2,1), -1},
 156     /* Constant definitions */
 157     {WINED3DSIO_DEF,      "def",      "undefined",         1, 5, NULL, NULL, 0, 0},
 158     {WINED3DSIO_DEFB,     "defb",     GLNAME_REQUIRE_GLSL, 1, 2, NULL, NULL, 0, 0},
 159     {WINED3DSIO_DEFI,     "defi",     GLNAME_REQUIRE_GLSL, 1, 5, NULL, NULL, 0, 0},
 160     /* Texture */
 161     {WINED3DSIO_TEXCOORD, "texcoord", "undefined", 1, 1, pshader_hw_texcoord, pshader_glsl_texcoord, 0, WINED3DPS_VERSION(1,3)},
 162     {WINED3DSIO_TEXCOORD, "texcrd",   "undefined", 1, 2, pshader_hw_texcoord, pshader_glsl_texcoord, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 163     {WINED3DSIO_TEXKILL,  "texkill",  "KIL",       1, 1, pshader_hw_texkill, pshader_glsl_texkill, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(3,0)},
 164     {WINED3DSIO_TEX,      "tex",      "undefined", 1, 1, pshader_hw_tex, pshader_glsl_tex, 0, WINED3DPS_VERSION(1,3)},
 165     {WINED3DSIO_TEX,      "texld",    "undefined", 1, 2, pshader_hw_tex, pshader_glsl_tex, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 166     {WINED3DSIO_TEX,      "texld",    "undefined", 1, 3, pshader_hw_tex, pshader_glsl_tex, WINED3DPS_VERSION(2,0), -1},
 167     {WINED3DSIO_TEXBEM,   "texbem",   "undefined", 1, 2, pshader_hw_texbem, pshader_glsl_texbem, 0, WINED3DPS_VERSION(1,3)},
 168     {WINED3DSIO_TEXBEML,  "texbeml",  GLNAME_REQUIRE_GLSL, 1, 2, pshader_hw_texbem, pshader_glsl_texbem, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 169     {WINED3DSIO_TEXREG2AR,"texreg2ar","undefined", 1, 2, pshader_hw_texreg2ar, pshader_glsl_texreg2ar, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 170     {WINED3DSIO_TEXREG2GB,"texreg2gb","undefined", 1, 2, pshader_hw_texreg2gb, pshader_glsl_texreg2gb, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 171     {WINED3DSIO_TEXREG2RGB,   "texreg2rgb",   "undefined", 1, 2, pshader_hw_texreg2rgb, pshader_glsl_texreg2rgb, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 172     {WINED3DSIO_TEXM3x2PAD,   "texm3x2pad",   "undefined", 1, 2, pshader_hw_texm3x2pad, pshader_glsl_texm3x2pad, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 173     {WINED3DSIO_TEXM3x2TEX,   "texm3x2tex",   "undefined", 1, 2, pshader_hw_texm3x2tex, pshader_glsl_texm3x2tex, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 174     {WINED3DSIO_TEXM3x3PAD,   "texm3x3pad",   "undefined", 1, 2, pshader_hw_texm3x3pad, pshader_glsl_texm3x3pad, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 175     {WINED3DSIO_TEXM3x3DIFF,  "texm3x3diff",  GLNAME_REQUIRE_GLSL, 1, 2, NULL, NULL, WINED3DPS_VERSION(0,0), WINED3DPS_VERSION(0,0)},
 176     {WINED3DSIO_TEXM3x3SPEC,  "texm3x3spec",  "undefined", 1, 3, pshader_hw_texm3x3spec, pshader_glsl_texm3x3spec, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 177     {WINED3DSIO_TEXM3x3VSPEC, "texm3x3vspec",  "undefined", 1, 2, pshader_hw_texm3x3vspec, pshader_glsl_texm3x3vspec, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 178     {WINED3DSIO_TEXM3x3TEX,   "texm3x3tex",   "undefined", 1, 2, pshader_hw_texm3x3tex, pshader_glsl_texm3x3tex, WINED3DPS_VERSION(1,0), WINED3DPS_VERSION(1,3)},
 179     {WINED3DSIO_TEXDP3TEX,    "texdp3tex",    NULL, 1, 2, pshader_hw_texdp3tex, pshader_glsl_texdp3tex, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 180     {WINED3DSIO_TEXM3x2DEPTH, "texm3x2depth", GLNAME_REQUIRE_GLSL, 1, 2, pshader_hw_texm3x2depth, pshader_glsl_texm3x2depth, WINED3DPS_VERSION(1,3), WINED3DPS_VERSION(1,3)},
 181     {WINED3DSIO_TEXDP3,   "texdp3",   NULL, 1, 2, pshader_hw_texdp3, pshader_glsl_texdp3, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 182     {WINED3DSIO_TEXM3x3,  "texm3x3",  NULL, 1, 2, pshader_hw_texm3x3, pshader_glsl_texm3x3, WINED3DPS_VERSION(1,2), WINED3DPS_VERSION(1,3)},
 183     {WINED3DSIO_TEXDEPTH, "texdepth", NULL, 1, 1, pshader_hw_texdepth, pshader_glsl_texdepth, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 184     {WINED3DSIO_BEM,      "bem",      "undefined",         1, 3, pshader_hw_bem, pshader_glsl_bem, WINED3DPS_VERSION(1,4), WINED3DPS_VERSION(1,4)},
 185     {WINED3DSIO_DSX,      "dsx",      NULL, 1, 2, NULL, shader_glsl_map2gl, WINED3DPS_VERSION(2,1), -1},
 186     {WINED3DSIO_DSY,      "dsy",      NULL, 1, 2, NULL, shader_glsl_map2gl, WINED3DPS_VERSION(2,1), -1},
 187     {WINED3DSIO_TEXLDD,   "texldd",   GLNAME_REQUIRE_GLSL, 1, 5, NULL, NULL, WINED3DPS_VERSION(2,1), -1},
 188     {WINED3DSIO_SETP,     "setp",     GLNAME_REQUIRE_GLSL, 1, 3, NULL, NULL, 0, 0},
 189     {WINED3DSIO_TEXLDL,   "texldl",   NULL, 1, 3, NULL, shader_glsl_texldl, WINED3DPS_VERSION(3,0), -1},
 190     {WINED3DSIO_PHASE,    "phase",    GLNAME_REQUIRE_GLSL, 0, 0, NULL, NULL, 0, 0},
 191     {0,               NULL,       NULL,   0, 0, NULL,            NULL, 0, 0}
 192 };
 193
 194 static void pshader_set_limits(
 195       IWineD3DPixelShaderImpl *This) {
 196
 197       This->baseShader.limits.attributes = 0;
 198       This->baseShader.limits.address = 0;
 199       This->baseShader.limits.packed_output = 0;
 200
 201       switch (This->baseShader.hex_version) {
 202           case WINED3DPS_VERSION(1,0):
 203           case WINED3DPS_VERSION(1,1):
 204           case WINED3DPS_VERSION(1,2):
 205           case WINED3DPS_VERSION(1,3):
 206                    This->baseShader.limits.temporary = 2;
 207                    This->baseShader.limits.constant_float = 8;
 208                    This->baseShader.limits.constant_int = 0;
 209                    This->baseShader.limits.constant_bool = 0;
 210                    This->baseShader.limits.texcoord = 4;
 211                    This->baseShader.limits.sampler = 4;
 212                    This->baseShader.limits.packed_input = 0;
 213                    This->baseShader.limits.label = 0;
 214                    break;
 215
 216           case WINED3DPS_VERSION(1,4):
 217                    This->baseShader.limits.temporary = 6;
 218                    This->baseShader.limits.constant_float = 8;
 219                    This->baseShader.limits.constant_int = 0;
 220                    This->baseShader.limits.constant_bool = 0;
 221                    This->baseShader.limits.texcoord = 6;
 222                    This->baseShader.limits.sampler = 6;
 223                    This->baseShader.limits.packed_input = 0;
 224                    This->baseShader.limits.label = 0;
 225                    break;
 226
 227           /* FIXME: temporaries must match D3DPSHADERCAPS2_0.NumTemps */
 228           case WINED3DPS_VERSION(2,0):
 229                    This->baseShader.limits.temporary = 32;
 230                    This->baseShader.limits.constant_float = 32;
 231                    This->baseShader.limits.constant_int = 16;
 232                    This->baseShader.limits.constant_bool = 16;
 233                    This->baseShader.limits.texcoord = 8;
 234                    This->baseShader.limits.sampler = 16;
 235                    This->baseShader.limits.packed_input = 0;
 236                    break;
 237
 238           case WINED3DPS_VERSION(2,1):
 239                    This->baseShader.limits.temporary = 32;
 240                    This->baseShader.limits.constant_float = 32;
 241                    This->baseShader.limits.constant_int = 16;
 242                    This->baseShader.limits.constant_bool = 16;
 243                    This->baseShader.limits.texcoord = 8;
 244                    This->baseShader.limits.sampler = 16;
 245                    This->baseShader.limits.packed_input = 0;
 246                    This->baseShader.limits.label = 16;
 247                    break;
 248
 249           case WINED3DPS_VERSION(3,0):
 250                    This->baseShader.limits.temporary = 32;
 251                    This->baseShader.limits.constant_float = 224;
 252                    This->baseShader.limits.constant_int = 16;
 253                    This->baseShader.limits.constant_bool = 16;
 254                    This->baseShader.limits.texcoord = 0;
 255                    This->baseShader.limits.sampler = 16;
 256                    This->baseShader.limits.packed_input = 12;
 257                    This->baseShader.limits.label = 16; /* FIXME: 2048 */
 258                    break;
 259
 260           default: This->baseShader.limits.temporary = 32;
 261                    This->baseShader.limits.constant_float = 32;
 262                    This->baseShader.limits.constant_int = 16;
 263                    This->baseShader.limits.constant_bool = 16;
 264                    This->baseShader.limits.texcoord = 8;
 265                    This->baseShader.limits.sampler = 16;
 266                    This->baseShader.limits.packed_input = 0;
 267                    This->baseShader.limits.label = 0;
 268                    FIXME("Unrecognized pixel shader version %#x\n",
 269                        This->baseShader.hex_version);
 270       }
 271 }
 272
 273 /** Generate a pixel shader string using either GL_FRAGMENT_PROGRAM_ARB
 274     or GLSL and send it to the card */
 275 static inline VOID IWineD3DPixelShaderImpl_GenerateShader(
 276     IWineD3DPixelShader *iface) {
 277     IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
 278     SHADER_BUFFER buffer;
 279
 280 #if 0 /* FIXME: Use the buffer that is held by the device, this is ok since fixups will be skipped for software shaders
 281         it also requires entering a critical section but cuts down the runtime footprint of wined3d and any memory fragmentation that may occur... */
 282     if (This->device->fixupVertexBufferSize < SHADER_PGMSIZE) {
 283         HeapFree(GetProcessHeap(), 0, This->fixupVertexBuffer);
 284         This->fixupVertexBuffer = HeapAlloc(GetProcessHeap() , 0, SHADER_PGMSIZE);
 285         This->fixupVertexBufferSize = PGMSIZE;
 286         This->fixupVertexBuffer[0] = 0;
 287     }
 288     buffer.buffer = This->device->fixupVertexBuffer;
 289 #else
 290     buffer.buffer = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, SHADER_PGMSIZE);
 291 #endif
 292     buffer.bsize = 0;
 293     buffer.lineNo = 0;
 294     buffer.newline = TRUE;
 295
 296     ((IWineD3DDeviceImpl *)This->baseShader.device)->shader_backend->shader_generate_pshader(iface, &buffer);
 297
 298 #if 1 /* if were using the data buffer of device then we don't need to free it */
 299   HeapFree(GetProcessHeap(), 0, buffer.buffer);
 300 #endif
 301 }
 302
 303 static HRESULT WINAPI IWineD3DPixelShaderImpl_SetFunction(IWineD3DPixelShader *iface, CONST DWORD *pFunction) {
 304
 305     IWineD3DPixelShaderImpl *This =(IWineD3DPixelShaderImpl *)iface;
 306     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
 307
 308     TRACE("(%p) : pFunction %p\n", iface, pFunction);
 309
 310     /* First pass: trace shader */
 311     shader_trace_init((IWineD3DBaseShader*) This, pFunction);
 312     pshader_set_limits(This);
 313
 314     /* Initialize immediate constant lists */
 315     list_init(&This->baseShader.constantsF);
 316     list_init(&This->baseShader.constantsB);
 317     list_init(&This->baseShader.constantsI);
 318
 319     if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) > 1) {
 320         shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
 321         HRESULT hr;
 322         unsigned int i, j, highest_reg_used = 0, num_regs_used = 0;
 323
 324         /* Second pass: figure out which registers are used, what the semantics are, etc.. */
 325         memset(reg_maps, 0, sizeof(shader_reg_maps));
 326         hr = shader_get_registers_used((IWineD3DBaseShader*) This, reg_maps,
 327             This->semantics_in, NULL, pFunction, NULL);
 328         if (FAILED(hr)) return hr;
 329         /* FIXME: validate reg_maps against OpenGL */
 330
 331         for(i = 0; i < MAX_REG_INPUT; i++) {
 332             if(This->input_reg_used[i]) {
 333                 num_regs_used++;
 334                 highest_reg_used = i;
 335             }
 336         }
 337
 338         /* Don't do any register mapping magic if it is not needed, or if we can't
 339          * achieve anything anyway
 340          */
 341         if(highest_reg_used < (GL_LIMITS(glsl_varyings) / 4) ||
 342            num_regs_used > (GL_LIMITS(glsl_varyings) / 4) ) {
 343             if(num_regs_used > (GL_LIMITS(glsl_varyings) / 4)) {
 344                 /* This happens with relative addressing. The input mapper function
 345                  * warns about this if the higher registers are declared too, so
 346                  * don't write a FIXME here
 347                  */
 348                 WARN("More varying registers used than supported\n");
 349             }
 350
 351             for(i = 0; i < MAX_REG_INPUT; i++) {
 352                 This->input_reg_map[i] = i;
 353             }
 354             This->declared_in_count = highest_reg_used + 1;
 355         } else {
 356             j = 0;
 357             for(i = 0; i < MAX_REG_INPUT; i++) {
 358                 if(This->input_reg_used[i]) {
 359                     This->input_reg_map[i] = j;
 360                     j++;
 361                 } else {
 362                     This->input_reg_map[i] = -1;
 363                 }
 364             }
 365             This->declared_in_count = j;
 366         }
 367     }
 368     This->baseShader.load_local_constsF = FALSE;
 369
 370     This->baseShader.shader_mode = deviceImpl->ps_selected_mode;
 371
 372     TRACE("(%p) : Copying the function\n", This);
 373     if (NULL != pFunction) {
 374         void *function;
 375
 376         function = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, This->baseShader.functionLength);
 377         if (!function) return E_OUTOFMEMORY;
 378         memcpy(function, pFunction, This->baseShader.functionLength);
 379         This->baseShader.function = function;
 380     } else {
 381         This->baseShader.function = NULL;
 382     }
 383
 384     return WINED3D_OK;
 385 }
 386
 387 static HRESULT WINAPI IWineD3DPixelShaderImpl_CompileShader(IWineD3DPixelShader *iface) {
 388
 389     IWineD3DPixelShaderImpl *This =(IWineD3DPixelShaderImpl *)iface;
 390     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
 391     CONST DWORD *function = This->baseShader.function;
 392     UINT i, sampler;
 393     IWineD3DBaseTextureImpl *texture;
 394
 395     TRACE("(%p) : function %p\n", iface, function);
 396
 397     /* We're already compiled, but check if any of the hardcoded stateblock assumptions
 398      * changed.
 399      */
 400     if (This->baseShader.is_compiled) {
 401         char srgbenabled = deviceImpl->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
 402         for(i = 0; i < This->baseShader.num_sampled_samplers; i++) {
 403             sampler = This->baseShader.sampled_samplers[i];
 404             texture = (IWineD3DBaseTextureImpl *) deviceImpl->stateBlock->textures[sampler];
 405             if(texture && texture->baseTexture.shader_conversion_group != This->baseShader.sampled_format[sampler]) {
 406                 WARN("Recompiling shader %p due to format change on sampler %d\n", This, sampler);
 407                 WARN("Old format group %s, new is %s\n",
 408                      debug_d3dformat(This->baseShader.sampled_format[sampler]),
 409                      debug_d3dformat(texture->baseTexture.shader_conversion_group));
 410                 goto recompile;
 411             }
 412         }
 413
 414         /* TODO: Check projected textures */
 415         /* TODO: Check texture types(2D, Cube, 3D) */
 416
 417         if(srgbenabled != This->srgb_enabled && This->srgb_mode_hardcoded) {
 418             WARN("Recompiling shader because srgb correction is different and hardcoded\n");
 419             goto recompile;
 420         }
 421         if(This->baseShader.reg_maps.vpos && !This->vpos_uniform) {
 422             if(This->render_offscreen != deviceImpl->render_offscreen ||
 423                This->height != ((IWineD3DSurfaceImpl *) deviceImpl->render_targets[0])->currentDesc.Height) {
 424                 WARN("Recompiling shader because vpos is used, hard compiled and changed\n");
 425                 goto recompile;
 426             }
 427         }
 428         if(This->baseShader.reg_maps.usesdsy && !This->vpos_uniform) {
 429             if(This->render_offscreen ? 0 : 1 != deviceImpl->render_offscreen ? 0 : 1) {
 430                 WARN("Recompiling shader because dsy is used, hard compiled and render_offscreen changed\n");
 431                 goto recompile;
 432             }
 433         }
 434         if(This->baseShader.hex_version >= WINED3DPS_VERSION(3,0)) {
 435             if(((IWineD3DDeviceImpl *) This->baseShader.device)->strided_streams.u.s.position_transformed) {
 436                 if(This->vertexprocessing != pretransformed) {
 437                     WARN("Recompiling shader because pretransformed vertices are provided, which wasn't the case before\n");
 438                     goto recompile;
 439                 }
 440             } else if(!use_vs((IWineD3DDeviceImpl *) This->baseShader.device) &&
 441                        This->vertexprocessing != fixedfunction) {
 442                 WARN("Recompiling shader because fixed function vp is in use, which wasn't the case before\n");
 443                 goto recompile;
 444             } else if(This->vertexprocessing != vertexshader) {
 445                 WARN("Recompiling shader because vertex shaders are in use, which wasn't the case before\n");
 446                 goto recompile;
 447             }
 448         }
 449
 450         return WINED3D_OK;
 451
 452         recompile:
 453         if(This->baseShader.recompile_count > 50) {
 454             FIXME("Shader %p recompiled more than 50 times\n", This);
 455         } else {
 456             This->baseShader.recompile_count++;
 457         }
 458
 459         deviceImpl->shader_backend->shader_destroy((IWineD3DBaseShader *) iface);
 460     }
 461
 462     /* We don't need to compile */
 463     if (!function) {
 464         This->baseShader.is_compiled = TRUE;
 465         return WINED3D_OK;
 466     }
 467
 468     if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1) {
 469         shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
 470         HRESULT hr;
 471
 472         /* Second pass: figure out which registers are used, what the semantics are, etc.. */
 473         memset(reg_maps, 0, sizeof(shader_reg_maps));
 474         hr = shader_get_registers_used((IWineD3DBaseShader*)This, reg_maps,
 475             This->semantics_in, NULL, This->baseShader.function, deviceImpl->stateBlock);
 476         if (FAILED(hr)) return hr;
 477         /* FIXME: validate reg_maps against OpenGL */
 478     }
 479
 480     /* Reset fields tracking stateblock values being hardcoded in the shader */
 481     This->baseShader.num_sampled_samplers = 0;
 482
 483     /* Generate the HW shader */
 484     TRACE("(%p) : Generating hardware program\n", This);
 485     IWineD3DPixelShaderImpl_GenerateShader(iface);
 486
 487     This->baseShader.is_compiled = TRUE;
 488
 489     return WINED3D_OK;
 490 }
 491
 492 const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
 493 {
 494     /*** IUnknown methods ***/
 495     IWineD3DPixelShaderImpl_QueryInterface,
 496     IWineD3DPixelShaderImpl_AddRef,
 497     IWineD3DPixelShaderImpl_Release,
 498     /*** IWineD3DBase methods ***/
 499     IWineD3DPixelShaderImpl_GetParent,
 500     /*** IWineD3DBaseShader methods ***/
 501     IWineD3DPixelShaderImpl_SetFunction,
 502     IWineD3DPixelShaderImpl_CompileShader,
 503     /*** IWineD3DPixelShader methods ***/
 504     IWineD3DPixelShaderImpl_GetDevice,
 505     IWineD3DPixelShaderImpl_GetFunction
 506 };