dlls/wined3d/vertexshader.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 void vshader_set_limits(IWineD3DVertexShaderImpl *This)
  38 {
  39     DWORD shader_version = WINED3D_SHADER_VERSION(This->baseShader.reg_maps.shader_version.major,
  40             This->baseShader.reg_maps.shader_version.minor);
  41
  42     This->baseShader.limits.texcoord = 0;
  43     This->baseShader.limits.attributes = 16;
  44     This->baseShader.limits.packed_input = 0;
  45
  46     switch (shader_version)
  47     {
  48         case WINED3D_SHADER_VERSION(1,0):
  49         case WINED3D_SHADER_VERSION(1,1):
  50             This->baseShader.limits.temporary = 12;
  51             This->baseShader.limits.constant_bool = 0;
  52             This->baseShader.limits.constant_int = 0;
  53             This->baseShader.limits.address = 1;
  54             This->baseShader.limits.packed_output = 0;
  55             This->baseShader.limits.sampler = 0;
  56             This->baseShader.limits.label = 0;
  57             /* TODO: vs_1_1 has a minimum of 96 constants. What happens if a vs_1_1 shader is used
  58              * on a vs_3_0 capable card that has 256 constants? */
  59             This->baseShader.limits.constant_float = min(256, GL_LIMITS(vshader_constantsF));
  60             break;
  61
  62         case WINED3D_SHADER_VERSION(2,0):
  63         case WINED3D_SHADER_VERSION(2,1):
  64             This->baseShader.limits.temporary = 12;
  65             This->baseShader.limits.constant_bool = 16;
  66             This->baseShader.limits.constant_int = 16;
  67             This->baseShader.limits.address = 1;
  68             This->baseShader.limits.packed_output = 0;
  69             This->baseShader.limits.sampler = 0;
  70             This->baseShader.limits.label = 16;
  71             This->baseShader.limits.constant_float = min(256, GL_LIMITS(vshader_constantsF));
  72             break;
  73
  74         case WINED3D_SHADER_VERSION(3,0):
  75             This->baseShader.limits.temporary = 32;
  76             This->baseShader.limits.constant_bool = 32;
  77             This->baseShader.limits.constant_int = 32;
  78             This->baseShader.limits.address = 1;
  79             This->baseShader.limits.packed_output = 12;
  80             This->baseShader.limits.sampler = 4;
  81             This->baseShader.limits.label = 16; /* FIXME: 2048 */
  82             /* DX10 cards on Windows advertise a d3d9 constant limit of 256 even though they are capable
  83              * of supporting much more(GL drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
  84              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0 shaders to 256.s
  85              * use constant buffers */
  86             This->baseShader.limits.constant_float = min(256, GL_LIMITS(vshader_constantsF));
  87             break;
  88
  89         default:
  90             This->baseShader.limits.temporary = 12;
  91             This->baseShader.limits.constant_bool = 16;
  92             This->baseShader.limits.constant_int = 16;
  93             This->baseShader.limits.address = 1;
  94             This->baseShader.limits.packed_output = 0;
  95             This->baseShader.limits.sampler = 0;
  96             This->baseShader.limits.label = 16;
  97             This->baseShader.limits.constant_float = min(256, GL_LIMITS(vshader_constantsF));
  98             FIXME("Unrecognized vertex shader version %u.%u\n",
  99                     This->baseShader.reg_maps.shader_version.major,
 100                     This->baseShader.reg_maps.shader_version.minor);
 101     }
 102 }
 103
 104 /* This is an internal function,
 105  * used to create fake semantics for shaders
 106  * that don't have them - d3d8 shaders where the declaration
 107  * stores the register for each input
 108  */
 109 static void vshader_set_input(
 110     IWineD3DVertexShaderImpl* This,
 111     unsigned int regnum,
 112     BYTE usage, BYTE usage_idx) {
 113
 114     This->semantics_in[regnum].usage = usage;
 115     This->semantics_in[regnum].usage_idx = usage_idx;
 116     This->semantics_in[regnum].reg.register_type = WINED3DSPR_INPUT;
 117     This->semantics_in[regnum].reg.register_idx = regnum;
 118     This->semantics_in[regnum].reg.write_mask = WINED3DSP_WRITEMASK_ALL;
 119     This->semantics_in[regnum].reg.modifiers = 0;
 120     This->semantics_in[regnum].reg.shift = 0;
 121     This->semantics_in[regnum].reg.rel_addr = NULL;
 122 }
 123
 124 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2) {
 125     if (usage_idx1 != usage_idx2) return FALSE;
 126     if (usage1 == usage2) return TRUE;
 127     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
 128     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
 129
 130     return FALSE;
 131 }
 132
 133 BOOL vshader_get_input(
 134     IWineD3DVertexShader* iface,
 135     BYTE usage_req, BYTE usage_idx_req,
 136     unsigned int* regnum) {
 137
 138     IWineD3DVertexShaderImpl* This = (IWineD3DVertexShaderImpl*) iface;
 139     int i;
 140
 141     for (i = 0; i < MAX_ATTRIBS; i++) {
 142         if (!This->baseShader.reg_maps.attributes[i]) continue;
 143
 144         if (match_usage(This->semantics_in[i].usage,
 145                 This->semantics_in[i].usage_idx, usage_req, usage_idx_req))
 146         {
 147             *regnum = i;
 148             return TRUE;
 149         }
 150     }
 151     return FALSE;
 152 }
 153
 154 /* *******************************************
 155    IWineD3DVertexShader IUnknown parts follow
 156    ******************************************* */
 157 static HRESULT  WINAPI IWineD3DVertexShaderImpl_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, LPVOID *ppobj) {
 158     TRACE("iface %p, riid %s, ppobj %p\n", iface, debugstr_guid(riid), ppobj);
 159
 160     if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
 161             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
 162             || IsEqualGUID(riid, &IID_IWineD3DBase)
 163             || IsEqualGUID(riid, &IID_IUnknown))
 164     {
 165         IUnknown_AddRef(iface);
 166         *ppobj = iface;
 167         return S_OK;
 168     }
 169
 170     WARN("%s not implemented, returning E_NOINTERFACE\n", debugstr_guid(riid));
 171
 172     *ppobj = NULL;
 173     return E_NOINTERFACE;
 174 }
 175
 176 static ULONG  WINAPI IWineD3DVertexShaderImpl_AddRef(IWineD3DVertexShader *iface) {
 177     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
 178     ULONG refcount = InterlockedIncrement(&This->baseShader.ref);
 179
 180     TRACE("%p increasing refcount to %u\n", This, refcount);
 181
 182     return refcount;
 183 }
 184
 185 static ULONG WINAPI IWineD3DVertexShaderImpl_Release(IWineD3DVertexShader *iface) {
 186     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
 187     ULONG refcount = InterlockedDecrement(&This->baseShader.ref);
 188
 189     TRACE("%p decreasing refcount to %u\n", This, refcount);
 190
 191     if (!refcount)
 192     {
 193         shader_cleanup((IWineD3DBaseShader *)iface);
 194         HeapFree(GetProcessHeap(), 0, This);
 195     }
 196
 197     return refcount;
 198 }
 199
 200 /* *******************************************
 201    IWineD3DVertexShader IWineD3DVertexShader parts follow
 202    ******************************************* */
 203
 204 static HRESULT WINAPI IWineD3DVertexShaderImpl_GetParent(IWineD3DVertexShader *iface, IUnknown** parent){
 205     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
 206
 207     *parent = This->parent;
 208     IUnknown_AddRef(*parent);
 209     TRACE("(%p) : returning %p\n", This, *parent);
 210     return WINED3D_OK;
 211 }
 212
 213 static HRESULT WINAPI IWineD3DVertexShaderImpl_GetDevice(IWineD3DVertexShader* iface, IWineD3DDevice **pDevice){
 214     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
 215     IWineD3DDevice_AddRef(This->baseShader.device);
 216     *pDevice = This->baseShader.device;
 217     TRACE("(%p) returning %p\n", This, *pDevice);
 218     return WINED3D_OK;
 219 }
 220
 221 static HRESULT WINAPI IWineD3DVertexShaderImpl_GetFunction(IWineD3DVertexShader* impl, VOID* pData, UINT* pSizeOfData) {
 222     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)impl;
 223     TRACE("(%p) : pData(%p), pSizeOfData(%p)\n", This, pData, pSizeOfData);
 224
 225     if (NULL == pData) {
 226         *pSizeOfData = This->baseShader.functionLength;
 227         return WINED3D_OK;
 228     }
 229     if (*pSizeOfData < This->baseShader.functionLength) {
 230         /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
 231          * than the required size we should write the required size and
 232          * return D3DERR_MOREDATA. That's not actually true. */
 233         return WINED3DERR_INVALIDCALL;
 234     }
 235
 236     TRACE("(%p) : GetFunction copying to %p\n", This, pData);
 237     memcpy(pData, This->baseShader.function, This->baseShader.functionLength);
 238
 239     return WINED3D_OK;
 240 }
 241
 242 /* Note that for vertex shaders CompileShader isn't called until the
 243  * shader is first used. The reason for this is that we need the vertex
 244  * declaration the shader will be used with in order to determine if
 245  * the data in a register is of type D3DCOLOR, and needs swizzling. */
 246 static HRESULT WINAPI IWineD3DVertexShaderImpl_SetFunction(IWineD3DVertexShader *iface, CONST DWORD *pFunction) {
 247
 248     IWineD3DVertexShaderImpl *This =(IWineD3DVertexShaderImpl *)iface;
 249     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
 250     const struct wined3d_shader_frontend *fe;
 251     HRESULT hr;
 252     shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
 253
 254     TRACE("(%p) : pFunction %p\n", iface, pFunction);
 255
 256     fe = shader_select_frontend(*pFunction);
 257     if (!fe)
 258     {
 259         FIXME("Unable to find frontend for shader.\n");
 260         return WINED3DERR_INVALIDCALL;
 261     }
 262     This->baseShader.frontend = fe;
 263     This->baseShader.frontend_data = fe->shader_init(pFunction);
 264     if (!This->baseShader.frontend_data)
 265     {
 266         FIXME("Failed to initialize frontend.\n");
 267         return WINED3DERR_INVALIDCALL;
 268     }
 269
 270     /* First pass: trace shader */
 271     if (TRACE_ON(d3d_shader)) shader_trace_init(fe, This->baseShader.frontend_data, pFunction);
 272
 273     /* Initialize immediate constant lists */
 274     list_init(&This->baseShader.constantsF);
 275     list_init(&This->baseShader.constantsB);
 276     list_init(&This->baseShader.constantsI);
 277
 278     /* Second pass: figure out registers used, semantics, etc.. */
 279     This->min_rel_offset = GL_LIMITS(vshader_constantsF);
 280     This->max_rel_offset = 0;
 281     hr = shader_get_registers_used((IWineD3DBaseShader*) This, fe,
 282             reg_maps, This->semantics_in, This->semantics_out, pFunction);
 283     if (hr != WINED3D_OK) return hr;
 284
 285     vshader_set_limits(This);
 286
 287     This->baseShader.shader_mode = deviceImpl->vs_selected_mode;
 288
 289     if(deviceImpl->vs_selected_mode == SHADER_ARB &&
 290        (GLINFO_LOCATION).arb_vs_offset_limit      &&
 291        This->min_rel_offset <= This->max_rel_offset) {
 292
 293         if(This->max_rel_offset - This->min_rel_offset > 127) {
 294             FIXME("The difference between the minimum and maximum relative offset is > 127\n");
 295             FIXME("Which this OpenGL implementation does not support. Try using GLSL\n");
 296             FIXME("Min: %d, Max: %d\n", This->min_rel_offset, This->max_rel_offset);
 297         } else if(This->max_rel_offset - This->min_rel_offset > 63) {
 298             This->rel_offset = This->min_rel_offset + 63;
 299         } else if(This->max_rel_offset > 63) {
 300             This->rel_offset = This->min_rel_offset;
 301         } else {
 302             This->rel_offset = 0;
 303         }
 304     }
 305     This->baseShader.load_local_constsF = This->baseShader.reg_maps.usesrelconstF && !list_empty(&This->baseShader.constantsF);
 306
 307     /* copy the function ... because it will certainly be released by application */
 308     This->baseShader.function = HeapAlloc(GetProcessHeap(), 0, This->baseShader.functionLength);
 309     if (!This->baseShader.function) return E_OUTOFMEMORY;
 310     memcpy(This->baseShader.function, pFunction, This->baseShader.functionLength);
 311
 312     return WINED3D_OK;
 313 }
 314
 315 /* Preload semantics for d3d8 shaders */
 316 static void WINAPI IWineD3DVertexShaderImpl_FakeSemantics(IWineD3DVertexShader *iface, IWineD3DVertexDeclaration *vertex_declaration) {
 317     IWineD3DVertexShaderImpl *This =(IWineD3DVertexShaderImpl *)iface;
 318     IWineD3DVertexDeclarationImpl* vdecl = (IWineD3DVertexDeclarationImpl*)vertex_declaration;
 319
 320     unsigned int i;
 321     for (i = 0; i < vdecl->element_count; ++i)
 322     {
 323         const struct wined3d_vertex_declaration_element *e = &vdecl->elements[i];
 324         vshader_set_input(This, e->output_slot, e->usage, e->usage_idx);
 325     }
 326 }
 327
 328 /* Set local constants for d3d8 shaders */
 329 static HRESULT WINAPI IWIneD3DVertexShaderImpl_SetLocalConstantsF(IWineD3DVertexShader *iface,
 330         UINT start_idx, const float *src_data, UINT count) {
 331     IWineD3DVertexShaderImpl *This =(IWineD3DVertexShaderImpl *)iface;
 332     UINT i, end_idx;
 333
 334     TRACE("(%p) : start_idx %u, src_data %p, count %u\n", This, start_idx, src_data, count);
 335
 336     end_idx = start_idx + count;
 337     if (end_idx > GL_LIMITS(vshader_constantsF)) {
 338         WARN("end_idx %u > float constants limit %u\n", end_idx, GL_LIMITS(vshader_constantsF));
 339         end_idx = GL_LIMITS(vshader_constantsF);
 340     }
 341
 342     for (i = start_idx; i < end_idx; ++i) {
 343         local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 344         if (!lconst) return E_OUTOFMEMORY;
 345
 346         lconst->idx = i;
 347         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
 348         list_add_head(&This->baseShader.constantsF, &lconst->entry);
 349     }
 350
 351     return WINED3D_OK;
 352 }
 353
 354 static GLuint vertexshader_compile(IWineD3DVertexShaderImpl *This, const struct vs_compile_args *args) {
 355     IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
 356     SHADER_BUFFER buffer;
 357     GLuint ret;
 358
 359     /* Generate the HW shader */
 360     TRACE("(%p) : Generating hardware program\n", This);
 361     shader_buffer_init(&buffer);
 362     This->cur_args = args;
 363     ret = deviceImpl->shader_backend->shader_generate_vshader((IWineD3DVertexShader *)This, &buffer, args);
 364     This->cur_args = NULL;
 365     shader_buffer_free(&buffer);
 366
 367     return ret;
 368 }
 369
 370 const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
 371 {
 372     /*** IUnknown methods ***/
 373     IWineD3DVertexShaderImpl_QueryInterface,
 374     IWineD3DVertexShaderImpl_AddRef,
 375     IWineD3DVertexShaderImpl_Release,
 376     /*** IWineD3DBase methods ***/
 377     IWineD3DVertexShaderImpl_GetParent,
 378     /*** IWineD3DBaseShader methods ***/
 379     IWineD3DVertexShaderImpl_SetFunction,
 380     /*** IWineD3DVertexShader methods ***/
 381     IWineD3DVertexShaderImpl_GetDevice,
 382     IWineD3DVertexShaderImpl_GetFunction,
 383     IWineD3DVertexShaderImpl_FakeSemantics,
 384     IWIneD3DVertexShaderImpl_SetLocalConstantsF
 385 };
 386
 387 void find_vs_compile_args(IWineD3DVertexShaderImpl *shader, IWineD3DStateBlockImpl *stateblock, struct vs_compile_args *args) {
 388     args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
 389     args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
 390 }
 391
 392 static inline BOOL vs_args_equal(const struct vs_compile_args *stored, const struct vs_compile_args *new,
 393                                  const DWORD use_map) {
 394     if((stored->swizzle_map & use_map) != new->swizzle_map) return FALSE;
 395     return stored->fog_src == new->fog_src;
 396 }
 397
 398 GLuint find_gl_vshader(IWineD3DVertexShaderImpl *shader, const struct vs_compile_args *args)
 399 {
 400     UINT i;
 401     DWORD new_size = shader->shader_array_size;
 402     struct vs_compiled_shader *new_array;
 403     DWORD use_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.use_map;
 404
 405     /* Usually we have very few GL shaders for each d3d shader(just 1 or maybe 2),
 406      * so a linear search is more performant than a hashmap or a binary search
 407      * (cache coherency etc)
 408      */
 409     for(i = 0; i < shader->num_gl_shaders; i++) {
 410         if(vs_args_equal(&shader->gl_shaders[i].args, args, use_map)) {
 411             return shader->gl_shaders[i].prgId;
 412         }
 413     }
 414
 415     TRACE("No matching GL shader found, compiling a new shader\n");
 416
 417     if(shader->shader_array_size == shader->num_gl_shaders) {
 418         if (shader->num_gl_shaders)
 419         {
 420             new_size = shader->shader_array_size + max(1, shader->shader_array_size / 2);
 421             new_array = HeapReAlloc(GetProcessHeap(), 0, shader->gl_shaders,
 422                                     new_size * sizeof(*shader->gl_shaders));
 423         } else {
 424             new_array = HeapAlloc(GetProcessHeap(), 0, sizeof(*shader->gl_shaders));
 425             new_size = 1;
 426         }
 427
 428         if(!new_array) {
 429             ERR("Out of memory\n");
 430             return 0;
 431         }
 432         shader->gl_shaders = new_array;
 433         shader->shader_array_size = new_size;
 434     }
 435
 436     shader->gl_shaders[shader->num_gl_shaders].args = *args;
 437     shader->gl_shaders[shader->num_gl_shaders].prgId = vertexshader_compile(shader, args);
 438     return shader->gl_shaders[shader->num_gl_shaders++].prgId;
 439 }