git.oblomov.eu Git - wine/blob - 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  * Copyright 2009 Henri Verbeet for CodeWeavers
  11  *
  12  * This library is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU Lesser General Public
  14  * License as published by the Free Software Foundation; either
  15  * version 2.1 of the License, or (at your option) any later version.
  16  *
  17  * This library is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  20  * Lesser General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Lesser General Public
  23  * License along with this library; if not, write to the Free Software
  24  * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
  25  */
  26
  27 #include "config.h"
  28
  29 #include <math.h>
  30 #include <stdio.h>
  31
  32 #include "wined3d_private.h"
  33
  34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
  35
  36 #define GLINFO_LOCATION ((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info
  37
  38 static void vshader_set_limits(IWineD3DVertexShaderImpl *This)
  39 {
  40     DWORD shader_version = WINED3D_SHADER_VERSION(This->baseShader.reg_maps.shader_version.major,
  41             This->baseShader.reg_maps.shader_version.minor);
  42
  43     This->baseShader.limits.texcoord = 0;
  44     This->baseShader.limits.attributes = 16;
  45     This->baseShader.limits.packed_input = 0;
  46
  47     switch (shader_version)
  48     {
  49         case WINED3D_SHADER_VERSION(1,0):
  50         case WINED3D_SHADER_VERSION(1,1):
  51             This->baseShader.limits.temporary = 12;
  52             This->baseShader.limits.constant_bool = 0;
  53             This->baseShader.limits.constant_int = 0;
  54             This->baseShader.limits.address = 1;
  55             This->baseShader.limits.packed_output = 0;
  56             This->baseShader.limits.sampler = 0;
  57             This->baseShader.limits.label = 0;
  58             /* TODO: vs_1_1 has a minimum of 96 constants. What happens if a vs_1_1 shader is used
  59              * on a vs_3_0 capable card that has 256 constants? */
  60             This->baseShader.limits.constant_float = min(256, GL_LIMITS(vshader_constantsF));
  61             break;
  62
  63         case WINED3D_SHADER_VERSION(2,0):
  64         case WINED3D_SHADER_VERSION(2,1):
  65             This->baseShader.limits.temporary = 12;
  66             This->baseShader.limits.constant_bool = 16;
  67             This->baseShader.limits.constant_int = 16;
  68             This->baseShader.limits.address = 1;
  69             This->baseShader.limits.packed_output = 0;
  70             This->baseShader.limits.sampler = 0;
  71             This->baseShader.limits.label = 16;
  72             This->baseShader.limits.constant_float = min(256, GL_LIMITS(vshader_constantsF));
  73             break;
  74
  75         case WINED3D_SHADER_VERSION(4,0):
  76             FIXME("Using 3.0 limits for 4.0 shader\n");
  77             /* Fall through */
  78
  79         case WINED3D_SHADER_VERSION(3,0):
  80             This->baseShader.limits.temporary = 32;
  81             This->baseShader.limits.constant_bool = 32;
  82             This->baseShader.limits.constant_int = 32;
  83             This->baseShader.limits.address = 1;
  84             This->baseShader.limits.packed_output = 12;
  85             This->baseShader.limits.sampler = 4;
  86             This->baseShader.limits.label = 16; /* FIXME: 2048 */
  87             /* DX10 cards on Windows advertise a d3d9 constant limit of 256 even though they are capable
  88              * of supporting much more(GL drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
  89              * wined3d-advertised maximum. Clamp the constant limit for <= 3.0 shaders to 256.s
  90              * use constant buffers */
  91             This->baseShader.limits.constant_float = min(256, GL_LIMITS(vshader_constantsF));
  92             break;
  93
  94         default:
  95             This->baseShader.limits.temporary = 12;
  96             This->baseShader.limits.constant_bool = 16;
  97             This->baseShader.limits.constant_int = 16;
  98             This->baseShader.limits.address = 1;
  99             This->baseShader.limits.packed_output = 0;
 100             This->baseShader.limits.sampler = 0;
 101             This->baseShader.limits.label = 16;
 102             This->baseShader.limits.constant_float = min(256, GL_LIMITS(vshader_constantsF));
 103             FIXME("Unrecognized vertex shader version %u.%u\n",
 104                     This->baseShader.reg_maps.shader_version.major,
 105                     This->baseShader.reg_maps.shader_version.minor);
 106     }
 107 }
 108
 109 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2) {
 110     if (usage_idx1 != usage_idx2) return FALSE;
 111     if (usage1 == usage2) return TRUE;
 112     if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
 113     if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
 114
 115     return FALSE;
 116 }
 117
 118 BOOL vshader_get_input(IWineD3DVertexShader* iface, BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
 119 {
 120     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
 121     WORD map = This->baseShader.reg_maps.input_registers;
 122     unsigned int i;
 123
 124     for (i = 0; map; map >>= 1, ++i)
 125     {
 126         if (!(map & 1)) continue;
 127
 128         if (match_usage(This->attributes[i].usage,
 129                 This->attributes[i].usage_idx, usage_req, usage_idx_req))
 130         {
 131             *regnum = i;
 132             return TRUE;
 133         }
 134     }
 135     return FALSE;
 136 }
 137
 138 /* *******************************************
 139    IWineD3DVertexShader IUnknown parts follow
 140    ******************************************* */
 141 static HRESULT  WINAPI IWineD3DVertexShaderImpl_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, LPVOID *ppobj) {
 142     TRACE("iface %p, riid %s, ppobj %p\n", iface, debugstr_guid(riid), ppobj);
 143
 144     if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
 145             || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
 146             || IsEqualGUID(riid, &IID_IWineD3DBase)
 147             || IsEqualGUID(riid, &IID_IUnknown))
 148     {
 149         IUnknown_AddRef(iface);
 150         *ppobj = iface;
 151         return S_OK;
 152     }
 153
 154     WARN("%s not implemented, returning E_NOINTERFACE\n", debugstr_guid(riid));
 155
 156     *ppobj = NULL;
 157     return E_NOINTERFACE;
 158 }
 159
 160 static ULONG  WINAPI IWineD3DVertexShaderImpl_AddRef(IWineD3DVertexShader *iface) {
 161     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
 162     ULONG refcount = InterlockedIncrement(&This->baseShader.ref);
 163
 164     TRACE("%p increasing refcount to %u\n", This, refcount);
 165
 166     return refcount;
 167 }
 168
 169 static ULONG WINAPI IWineD3DVertexShaderImpl_Release(IWineD3DVertexShader *iface) {
 170     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
 171     ULONG refcount = InterlockedDecrement(&This->baseShader.ref);
 172
 173     TRACE("%p decreasing refcount to %u\n", This, refcount);
 174
 175     if (!refcount)
 176     {
 177         shader_cleanup((IWineD3DBaseShader *)iface);
 178         This->parent_ops->wined3d_object_destroyed(This->parent);
 179         HeapFree(GetProcessHeap(), 0, This);
 180     }
 181
 182     return refcount;
 183 }
 184
 185 /* *******************************************
 186    IWineD3DVertexShader IWineD3DVertexShader parts follow
 187    ******************************************* */
 188
 189 static HRESULT WINAPI IWineD3DVertexShaderImpl_GetParent(IWineD3DVertexShader *iface, IUnknown** parent){
 190     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
 191
 192     *parent = This->parent;
 193     IUnknown_AddRef(*parent);
 194     TRACE("(%p) : returning %p\n", This, *parent);
 195     return WINED3D_OK;
 196 }
 197
 198 static HRESULT WINAPI IWineD3DVertexShaderImpl_GetDevice(IWineD3DVertexShader* iface, IWineD3DDevice **pDevice){
 199     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
 200     IWineD3DDevice_AddRef(This->baseShader.device);
 201     *pDevice = This->baseShader.device;
 202     TRACE("(%p) returning %p\n", This, *pDevice);
 203     return WINED3D_OK;
 204 }
 205
 206 static HRESULT WINAPI IWineD3DVertexShaderImpl_GetFunction(IWineD3DVertexShader* impl, VOID* pData, UINT* pSizeOfData) {
 207     IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)impl;
 208     TRACE("(%p) : pData(%p), pSizeOfData(%p)\n", This, pData, pSizeOfData);
 209
 210     if (NULL == pData) {
 211         *pSizeOfData = This->baseShader.functionLength;
 212         return WINED3D_OK;
 213     }
 214     if (*pSizeOfData < This->baseShader.functionLength) {
 215         /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
 216          * than the required size we should write the required size and
 217          * return D3DERR_MOREDATA. That's not actually true. */
 218         return WINED3DERR_INVALIDCALL;
 219     }
 220
 221     TRACE("(%p) : GetFunction copying to %p\n", This, pData);
 222     memcpy(pData, This->baseShader.function, This->baseShader.functionLength);
 223
 224     return WINED3D_OK;
 225 }
 226
 227 static HRESULT vertexshader_set_function(IWineD3DVertexShaderImpl *shader,
 228         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature)
 229 {
 230     IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
 231     const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
 232     const struct wined3d_shader_frontend *fe;
 233     unsigned int i;
 234     HRESULT hr;
 235     shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
 236
 237     TRACE("shader %p, byte_code %p, output_signature %p.\n", shader, byte_code, output_signature);
 238
 239     fe = shader_select_frontend(*byte_code);
 240     if (!fe)
 241     {
 242         FIXME("Unable to find frontend for shader.\n");
 243         return WINED3DERR_INVALIDCALL;
 244     }
 245     shader->baseShader.frontend = fe;
 246     shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
 247     if (!shader->baseShader.frontend_data)
 248     {
 249         FIXME("Failed to initialize frontend.\n");
 250         return WINED3DERR_INVALIDCALL;
 251     }
 252
 253     /* First pass: trace shader */
 254     if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
 255
 256     /* Initialize immediate constant lists */
 257     list_init(&shader->baseShader.constantsF);
 258     list_init(&shader->baseShader.constantsB);
 259     list_init(&shader->baseShader.constantsI);
 260
 261     /* Second pass: figure out registers used, semantics, etc.. */
 262     shader->min_rel_offset = gl_info->max_vshader_constantsF;
 263     shader->max_rel_offset = 0;
 264     hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
 265             reg_maps, shader->attributes, NULL, shader->output_signature,
 266             byte_code, gl_info->max_vshader_constantsF);
 267     if (hr != WINED3D_OK) return hr;
 268
 269     if (output_signature)
 270     {
 271         for (i = 0; i < output_signature->element_count; ++i)
 272         {
 273             struct wined3d_shader_signature_element *e = &output_signature->elements[i];
 274             reg_maps->output_registers |= 1 << e->register_idx;
 275             shader->output_signature[e->register_idx] = *e;
 276         }
 277     }
 278
 279     vshader_set_limits(shader);
 280
 281     if (device->vs_selected_mode == SHADER_ARB
 282             && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
 283             && shader->min_rel_offset <= shader->max_rel_offset)
 284     {
 285         if (shader->max_rel_offset - shader->min_rel_offset > 127)
 286         {
 287             FIXME("The difference between the minimum and maximum relative offset is > 127\n");
 288             FIXME("Which this OpenGL implementation does not support. Try using GLSL\n");
 289             FIXME("Min: %d, Max: %d\n", shader->min_rel_offset, shader->max_rel_offset);
 290         }
 291         else if (shader->max_rel_offset - shader->min_rel_offset > 63)
 292         {
 293             shader->rel_offset = shader->min_rel_offset + 63;
 294         }
 295         else if (shader->max_rel_offset > 63)
 296         {
 297             shader->rel_offset = shader->min_rel_offset;
 298         }
 299         else
 300         {
 301             shader->rel_offset = 0;
 302         }
 303     }
 304     shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
 305             && !list_empty(&shader->baseShader.constantsF);
 306
 307     /* copy the function ... because it will certainly be released by application */
 308     shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
 309     if (!shader->baseShader.function) return E_OUTOFMEMORY;
 310     memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
 311
 312     return WINED3D_OK;
 313 }
 314
 315 /* Set local constants for d3d8 shaders */
 316 static HRESULT WINAPI IWIneD3DVertexShaderImpl_SetLocalConstantsF(IWineD3DVertexShader *iface,
 317         UINT start_idx, const float *src_data, UINT count) {
 318     IWineD3DVertexShaderImpl *This =(IWineD3DVertexShaderImpl *)iface;
 319     UINT i, end_idx;
 320
 321     TRACE("(%p) : start_idx %u, src_data %p, count %u\n", This, start_idx, src_data, count);
 322
 323     end_idx = start_idx + count;
 324     if (end_idx > GL_LIMITS(vshader_constantsF)) {
 325         WARN("end_idx %u > float constants limit %u\n", end_idx, GL_LIMITS(vshader_constantsF));
 326         end_idx = GL_LIMITS(vshader_constantsF);
 327     }
 328
 329     for (i = start_idx; i < end_idx; ++i) {
 330         local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
 331         if (!lconst) return E_OUTOFMEMORY;
 332
 333         lconst->idx = i;
 334         memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
 335         list_add_head(&This->baseShader.constantsF, &lconst->entry);
 336     }
 337
 338     return WINED3D_OK;
 339 }
 340
 341 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
 342 {
 343     /*** IUnknown methods ***/
 344     IWineD3DVertexShaderImpl_QueryInterface,
 345     IWineD3DVertexShaderImpl_AddRef,
 346     IWineD3DVertexShaderImpl_Release,
 347     /*** IWineD3DBase methods ***/
 348     IWineD3DVertexShaderImpl_GetParent,
 349     /*** IWineD3DBaseShader methods ***/
 350     IWineD3DVertexShaderImpl_GetDevice,
 351     IWineD3DVertexShaderImpl_GetFunction,
 352     /*** IWineD3DVertexShader methods ***/
 353     IWIneD3DVertexShaderImpl_SetLocalConstantsF
 354 };
 355
 356 void find_vs_compile_args(IWineD3DVertexShaderImpl *shader, IWineD3DStateBlockImpl *stateblock, struct vs_compile_args *args) {
 357     args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
 358     args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
 359 }
 360
 361 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
 362         const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
 363         IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
 364 {
 365     HRESULT hr;
 366
 367     if (!byte_code) return WINED3DERR_INVALIDCALL;
 368
 369     shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
 370     shader->parent = parent;
 371     shader->parent_ops = parent_ops;
 372     shader_init(&shader->baseShader, (IWineD3DDevice *)device);
 373     list_add_head(&device->shaders, &shader->baseShader.shader_list_entry);
 374
 375     hr = vertexshader_set_function(shader, byte_code, output_signature);
 376     if (FAILED(hr))
 377     {
 378         WARN("Failed to set function, hr %#x.\n", hr);
 379         shader_cleanup((IWineD3DBaseShader *)shader);
 380         return hr;
 381     }
 382
 383     return WINED3D_OK;
 384 }