src/gallium/state_trackers/nine/stateblock9.c

   1 /*
   2  * Copyright 2011 Joakim Sindholt <opensource@zhasha.com>
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * on the rights to use, copy, modify, merge, publish, distribute, sub
   8  * license, and/or sell copies of the Software, and to permit persons to whom
   9  * the Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  18  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  19  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  20  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  21  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
  22
  23 #include "stateblock9.h"
  24 #include "device9.h"
  25 #include "basetexture9.h"
  26 #include "nine_helpers.h"
  27
  28 #define DBG_CHANNEL DBG_STATEBLOCK
  29
  30 /* XXX TODO: handling of lights is broken */
  31
  32 HRESULT
  33 NineStateBlock9_ctor( struct NineStateBlock9 *This,
  34                       struct NineUnknownParams *pParams,
  35                       enum nine_stateblock_type type )
  36 {
  37     HRESULT hr = NineUnknown_ctor(&This->base, pParams);
  38
  39     DBG("This=%p pParams=%p type=%d\n", This, pParams, type);
  40
  41     if (FAILED(hr))
  42         return hr;
  43
  44     This->type = type;
  45
  46     This->state.vs_const_f = MALLOC(pParams->device->constbuf_vs->width0);
  47     This->state.ps_const_f = MALLOC(pParams->device->constbuf_ps->width0);
  48     if (!This->state.vs_const_f || !This->state.ps_const_f)
  49         return E_OUTOFMEMORY;
  50
  51     return D3D_OK;
  52 }
  53
  54 void
  55 NineStateBlock9_dtor( struct NineStateBlock9 *This )
  56 {
  57     struct nine_state *state = &This->state;
  58     struct nine_range *r;
  59     struct nine_range_pool *pool = &This->base.device->range_pool;
  60
  61     nine_state_clear(state, FALSE);
  62
  63     if (state->vs_const_f) FREE(state->vs_const_f);
  64     if (state->ps_const_f) FREE(state->ps_const_f);
  65
  66     if (state->ff.light) FREE(state->ff.light);
  67
  68     if (state->ff.transform) FREE(state->ff.transform);
  69
  70     if (This->state.changed.ps_const_f) {
  71         for (r = This->state.changed.ps_const_f; r->next; r = r->next);
  72         nine_range_pool_put_chain(pool, This->state.changed.ps_const_f, r);
  73     }
  74     if (This->state.changed.vs_const_f) {
  75         for (r = This->state.changed.vs_const_f; r->next; r = r->next);
  76         nine_range_pool_put_chain(pool, This->state.changed.vs_const_f, r);
  77     }
  78
  79     NineUnknown_dtor(&This->base);
  80 }
  81
  82 /* Copy state marked changed in @mask from @src to @dst.
  83  * If @apply is false, updating dst->changed can be omitted.
  84  * TODO: compare ?
  85  */
  86 static void
  87 nine_state_copy_common(struct nine_state *dst,
  88                        const struct nine_state *src,
  89                        struct nine_state *mask, /* aliases either src or dst */
  90                        const boolean apply,
  91                        struct nine_range_pool *pool)
  92 {
  93     unsigned i, s;
  94
  95     if (apply)
  96        dst->changed.group |= mask->changed.group;
  97
  98     if (mask->changed.group & NINE_STATE_VIEWPORT)
  99         dst->viewport = src->viewport;
 100     if (mask->changed.group & NINE_STATE_SCISSOR)
 101         dst->scissor = src->scissor;
 102
 103     if (mask->changed.group & NINE_STATE_VS)
 104         nine_bind(&dst->vs, src->vs);
 105     if (mask->changed.group & NINE_STATE_PS)
 106         nine_bind(&dst->ps, src->ps);
 107
 108     /* Vertex constants.
 109      *
 110      * Various possibilities for optimization here, like creating a per-SB
 111      * constant buffer, or memcmp'ing for changes.
 112      * Will do that later depending on what works best for specific apps.
 113      */
 114     if (mask->changed.group & NINE_STATE_VS_CONST) {
 115         struct nine_range *r;
 116         for (r = mask->changed.vs_const_f; r; r = r->next) {
 117             memcpy(&dst->vs_const_f[r->bgn * 4],
 118                    &src->vs_const_f[r->bgn * 4],
 119                    (r->end - r->bgn) * 4 * sizeof(float));
 120             if (apply)
 121                 nine_ranges_insert(&dst->changed.vs_const_f, r->bgn, r->end,
 122                                    pool);
 123         }
 124         if (mask->changed.vs_const_i) {
 125             uint16_t m = mask->changed.vs_const_i;
 126             for (i = ffs(m) - 1, m >>= i; m; ++i, m >>= 1)
 127                 if (m & 1)
 128                     memcpy(dst->vs_const_i[i], src->vs_const_i[i], 4 * sizeof(int));
 129             if (apply)
 130                 dst->changed.vs_const_i |= mask->changed.vs_const_i;
 131         }
 132         if (mask->changed.vs_const_b) {
 133             uint16_t m = mask->changed.vs_const_b;
 134             for (i = ffs(m) - 1, m >>= i; m; ++i, m >>= 1)
 135                 if (m & 1)
 136                     dst->vs_const_b[i] = src->vs_const_b[i];
 137             if (apply)
 138                 dst->changed.vs_const_b |= mask->changed.vs_const_b;
 139         }
 140     }
 141
 142     /* Pixel constants. */
 143     if (mask->changed.group & NINE_STATE_PS_CONST) {
 144         struct nine_range *r;
 145         for (r = mask->changed.ps_const_f; r; r = r->next) {
 146             memcpy(&dst->ps_const_f[r->bgn * 4],
 147                    &src->ps_const_f[r->bgn * 4],
 148                    (r->end - r->bgn) * 4 * sizeof(float));
 149             if (apply)
 150                 nine_ranges_insert(&dst->changed.ps_const_f, r->bgn, r->end,
 151                                    pool);
 152         }
 153         if (mask->changed.ps_const_i) {
 154             uint16_t m = mask->changed.ps_const_i;
 155             for (i = ffs(m) - 1, m >>= i; m; ++i, m >>= 1)
 156                 if (m & 1)
 157                     memcpy(dst->ps_const_i[i], src->ps_const_i[i], 4 * sizeof(int));
 158             if (apply)
 159                 dst->changed.ps_const_i |= mask->changed.ps_const_i;
 160         }
 161         if (mask->changed.ps_const_b) {
 162             uint16_t m = mask->changed.ps_const_b;
 163             for (i = ffs(m) - 1, m >>= i; m; ++i, m >>= 1)
 164                 if (m & 1)
 165                     dst->ps_const_b[i] = src->ps_const_b[i];
 166             if (apply)
 167                 dst->changed.ps_const_b |= mask->changed.ps_const_b;
 168         }
 169     }
 170
 171     /* Render states.
 172      * TODO: Maybe build a list ?
 173      */
 174     for (i = 0; i < Elements(dst->changed.rs); ++i) {
 175         uint32_t m = mask->changed.rs[i];
 176         if (apply)
 177             dst->changed.rs[i] |= m;
 178         while (m) {
 179             const int r = ffs(m) - 1;
 180             m &= ~(1 << r);
 181             dst->rs[i * 32 + r] = src->rs[i * 32 + r];
 182         }
 183     }
 184
 185
 186     /* Clip planes. */
 187     if (mask->changed.ucp) {
 188         for (i = 0; i < PIPE_MAX_CLIP_PLANES; ++i)
 189             if (mask->changed.ucp & (1 << i))
 190                 memcpy(dst->clip.ucp[i],
 191                        src->clip.ucp[i], sizeof(src->clip.ucp[0]));
 192         if (apply)
 193            dst->changed.ucp |= mask->changed.ucp;
 194     }
 195
 196     /* Sampler state. */
 197     if (mask->changed.group & NINE_STATE_SAMPLER) {
 198         for (s = 0; s < NINE_MAX_SAMPLERS; ++s) {
 199             if (mask->changed.sampler[s] == 0x3ffe) {
 200                 memcpy(&dst->samp[s], &src->samp[s], sizeof(dst->samp[s]));
 201             } else {
 202                 uint32_t m = mask->changed.sampler[s];
 203                 while (m) {
 204                     const int i = ffs(m) - 1;
 205                     m &= ~(1 << i);
 206                     dst->samp[s][i] = src->samp[s][i];
 207                 }
 208             }
 209             if (apply)
 210                 dst->changed.sampler[s] |= mask->changed.sampler[s];
 211         }
 212     }
 213
 214     /* Index buffer. */
 215     if (mask->changed.group & NINE_STATE_IDXBUF)
 216         nine_bind(&dst->idxbuf, src->idxbuf);
 217
 218     /* Vertex streams. */
 219     if (mask->changed.vtxbuf | mask->changed.stream_freq) {
 220         uint32_t m = mask->changed.vtxbuf | mask->changed.stream_freq;
 221         for (i = 0; m; ++i, m >>= 1) {
 222             if (mask->changed.vtxbuf & (1 << i)) {
 223                 nine_bind(&dst->stream[i], src->stream[i]);
 224                 if (src->stream[i]) {
 225                     dst->vtxbuf[i].buffer_offset = src->vtxbuf[i].buffer_offset;
 226                     dst->vtxbuf[i].buffer = src->vtxbuf[i].buffer;
 227                     dst->vtxbuf[i].stride = src->vtxbuf[i].stride;
 228                 }
 229             }
 230             if (mask->changed.stream_freq & (1 << i))
 231                 dst->stream_freq[i] = src->stream_freq[i];
 232         }
 233         dst->stream_instancedata_mask &= ~mask->changed.stream_freq;
 234         dst->stream_instancedata_mask |=
 235             src->stream_instancedata_mask & mask->changed.stream_freq;
 236         if (apply) {
 237             dst->changed.vtxbuf |= mask->changed.vtxbuf;
 238             dst->changed.stream_freq |= mask->changed.stream_freq;
 239         }
 240     }
 241
 242     if (!(mask->changed.group & NINE_STATE_FF))
 243         return;
 244     WARN_ONCE("Fixed function state not handled properly by StateBlocks.\n");
 245
 246     /* Fixed function state. */
 247     if (apply)
 248         dst->ff.changed.group |= src->ff.changed.group;
 249
 250     if (mask->changed.group & NINE_STATE_FF_MATERIAL)
 251         dst->ff.material = src->ff.material;
 252
 253     if (mask->changed.group & NINE_STATE_FF_PSSTAGES) {
 254         for (s = 0; s < NINE_MAX_SAMPLERS; ++s) {
 255             for (i = 0; i < NINED3DTSS_COUNT; ++i)
 256                 if (mask->ff.changed.tex_stage[s][i / 32] & (1 << (i % 32)))
 257                     dst->ff.tex_stage[s][i] = src->ff.tex_stage[s][i];
 258             if (apply) {
 259                 /* TODO: it's 32 exactly, just offset by 1 as 0 is unused */
 260                 dst->ff.changed.tex_stage[s][0] |=
 261                     mask->ff.changed.tex_stage[s][0];
 262                 dst->ff.changed.tex_stage[s][1] |=
 263                     mask->ff.changed.tex_stage[s][1];
 264             }
 265         }
 266     }
 267     if (mask->changed.group & NINE_STATE_FF_LIGHTING) {
 268         if (dst->ff.num_lights < mask->ff.num_lights) {
 269             dst->ff.light = REALLOC(dst->ff.light,
 270                                     dst->ff.num_lights * sizeof(D3DLIGHT9),
 271                                     mask->ff.num_lights * sizeof(D3DLIGHT9));
 272             dst->ff.num_lights = mask->ff.num_lights;
 273         }
 274         for (i = 0; i < mask->ff.num_lights; ++i)
 275             if (mask->ff.light[i].Type != NINED3DLIGHT_INVALID)
 276                 dst->ff.light[i] = src->ff.light[i];
 277
 278         DBG("TODO: active lights\n");
 279     }
 280     if (mask->changed.group & NINE_STATE_FF_VSTRANSF) {
 281         for (i = 0; i < Elements(mask->ff.changed.transform); ++i) {
 282             if (!mask->ff.changed.transform[i])
 283                 continue;
 284             for (s = i * 32; s < (i * 32 + 32); ++s) {
 285                 if (!(mask->ff.changed.transform[i] & (1 << (s % 32))))
 286                     continue;
 287                 *nine_state_access_transform(dst, s, TRUE) =
 288                     *nine_state_access_transform( /* const because !alloc */
 289                         (struct nine_state *)src, s, FALSE);
 290             }
 291             if (apply)
 292                 dst->ff.changed.transform[i] |= mask->ff.changed.transform[i];
 293         }
 294     }
 295 }
 296
 297 static void
 298 nine_state_copy_common_all(struct nine_state *dst,
 299                            const struct nine_state *src,
 300                            struct nine_state *help,
 301                            const boolean apply,
 302                            struct nine_range_pool *pool,
 303                            const int MaxStreams)
 304 {
 305     unsigned i;
 306
 307     if (apply)
 308        dst->changed.group |= src->changed.group;
 309
 310     dst->viewport = src->viewport;
 311     dst->scissor = src->scissor;
 312
 313     nine_bind(&dst->vs, src->vs);
 314     nine_bind(&dst->ps, src->ps);
 315
 316     /* Vertex constants.
 317      *
 318      * Various possibilities for optimization here, like creating a per-SB
 319      * constant buffer, or memcmp'ing for changes.
 320      * Will do that later depending on what works best for specific apps.
 321      */
 322     if (1) {
 323         struct nine_range *r = help->changed.vs_const_f;
 324         memcpy(&dst->vs_const_f[0],
 325                &src->vs_const_f[0], (r->end - r->bgn) * 4 * sizeof(float));
 326         if (apply)
 327             nine_ranges_insert(&dst->changed.vs_const_f, r->bgn, r->end, pool);
 328
 329         memcpy(dst->vs_const_i, src->vs_const_i, sizeof(dst->vs_const_i));
 330         memcpy(dst->vs_const_b, src->vs_const_b, sizeof(dst->vs_const_b));
 331         if (apply) {
 332             dst->changed.vs_const_i |= src->changed.vs_const_i;
 333             dst->changed.vs_const_b |= src->changed.vs_const_b;
 334         }
 335     }
 336
 337     /* Pixel constants. */
 338     if (1) {
 339         struct nine_range *r = help->changed.ps_const_f;
 340         memcpy(&dst->ps_const_f[0],
 341                &src->ps_const_f[0], (r->end - r->bgn) * 4 * sizeof(float));
 342         if (apply)
 343             nine_ranges_insert(&dst->changed.ps_const_f, r->bgn, r->end, pool);
 344
 345         memcpy(dst->ps_const_i, src->ps_const_i, sizeof(dst->ps_const_i));
 346         memcpy(dst->ps_const_b, src->ps_const_b, sizeof(dst->ps_const_b));
 347         if (apply) {
 348             dst->changed.ps_const_i |= src->changed.ps_const_i;
 349             dst->changed.ps_const_b |= src->changed.ps_const_b;
 350         }
 351     }
 352
 353     /* Render states. */
 354     memcpy(dst->rs, src->rs, sizeof(dst->rs));
 355     if (apply)
 356         memcpy(dst->changed.rs, src->changed.rs, sizeof(dst->changed.rs));
 357
 358
 359     /* Clip planes. */
 360     memcpy(&dst->clip, &src->clip, sizeof(dst->clip));
 361     if (apply)
 362         dst->changed.ucp = src->changed.ucp;
 363
 364     /* Sampler state. */
 365     memcpy(dst->samp, src->samp, sizeof(dst->samp));
 366     if (apply)
 367         memcpy(dst->changed.sampler,
 368                src->changed.sampler, sizeof(dst->changed.sampler));
 369
 370     /* Index buffer. */
 371     nine_bind(&dst->idxbuf, src->idxbuf);
 372
 373     /* Vertex streams. */
 374     if (1) {
 375         for (i = 0; i < Elements(dst->stream); ++i) {
 376             nine_bind(&dst->stream[i], src->stream[i]);
 377             if (src->stream[i]) {
 378                 dst->vtxbuf[i].buffer_offset = src->vtxbuf[i].buffer_offset;
 379                 dst->vtxbuf[i].buffer = src->vtxbuf[i].buffer;
 380                 dst->vtxbuf[i].stride = src->vtxbuf[i].stride;
 381             }
 382             dst->stream_freq[i] = src->stream_freq[i];
 383         }
 384         dst->stream_instancedata_mask = src->stream_instancedata_mask;
 385         if (apply) {
 386             dst->changed.vtxbuf = (1ULL << MaxStreams) - 1;
 387             dst->changed.stream_freq = (1ULL << MaxStreams) - 1;
 388         }
 389     }
 390
 391     /* keep this check in case we want to disable FF */
 392     if (!(help->changed.group & NINE_STATE_FF))
 393         return;
 394     WARN_ONCE("Fixed function state not handled properly by StateBlocks.\n");
 395
 396     /* Fixed function state. */
 397     if (apply)
 398         dst->ff.changed.group = src->ff.changed.group;
 399
 400     dst->ff.material = src->ff.material;
 401
 402     memcpy(dst->ff.tex_stage, src->ff.tex_stage, sizeof(dst->ff.tex_stage));
 403     if (apply) /* TODO: memset */
 404         memcpy(dst->ff.changed.tex_stage,
 405                src->ff.changed.tex_stage, sizeof(dst->ff.changed.tex_stage));
 406
 407     /* Lights. */
 408     if (1) {
 409         if (dst->ff.num_lights < src->ff.num_lights) {
 410             dst->ff.light = REALLOC(dst->ff.light,
 411                                     dst->ff.num_lights * sizeof(D3DLIGHT9),
 412                                     src->ff.num_lights * sizeof(D3DLIGHT9));
 413             dst->ff.num_lights = src->ff.num_lights;
 414         }
 415         memcpy(dst->ff.light,
 416                src->ff.light, src->ff.num_lights * sizeof(dst->ff.light[0]));
 417
 418         DBG("TODO: active lights\n");
 419     }
 420
 421     /* Transforms. */
 422     if (1) {
 423         if (dst->ff.num_transforms < src->ff.num_transforms) {
 424             dst->ff.transform = REALLOC(dst->ff.transform,
 425                 dst->ff.num_transforms * sizeof(dst->ff.transform[0]),
 426                 src->ff.num_transforms * sizeof(src->ff.transform[0]));
 427             dst->ff.num_transforms = src->ff.num_transforms;
 428         }
 429         memcpy(dst->ff.transform,
 430                src->ff.transform, src->ff.num_transforms * sizeof(D3DMATRIX));
 431         if (apply) /* TODO: memset */
 432             memcpy(dst->ff.changed.transform,
 433                    src->ff.changed.transform, sizeof(dst->ff.changed.transform));
 434     }
 435 }
 436
 437 /* Capture those bits of current device state that have been changed between
 438  * BeginStateBlock and EndStateBlock.
 439  */
 440 HRESULT WINAPI
 441 NineStateBlock9_Capture( struct NineStateBlock9 *This )
 442 {
 443     struct nine_state *dst = &This->state;
 444     struct nine_state *src = &This->base.device->state;
 445     const int MaxStreams = This->base.device->caps.MaxStreams;
 446     unsigned s;
 447
 448     DBG("This=%p\n", This);
 449
 450     if (This->type == NINESBT_ALL)
 451         nine_state_copy_common_all(dst, src, dst, FALSE, NULL, MaxStreams);
 452     else
 453         nine_state_copy_common(dst, src, dst, FALSE, NULL);
 454
 455     if (dst->changed.group & NINE_STATE_VDECL)
 456         nine_bind(&dst->vdecl, src->vdecl);
 457
 458     /* Textures */
 459     if (dst->changed.texture) {
 460         uint32_t m = dst->changed.texture;
 461         for (s = 0; m; ++s, m >>= 1)
 462             if (m & 1)
 463                 nine_bind(&dst->texture[s], src->texture[s]);
 464     }
 465
 466     return D3D_OK;
 467 }
 468
 469 /* Set state managed by this StateBlock as current device state. */
 470 HRESULT WINAPI
 471 NineStateBlock9_Apply( struct NineStateBlock9 *This )
 472 {
 473     struct nine_state *dst = &This->base.device->state;
 474     struct nine_state *src = &This->state;
 475     struct nine_range_pool *pool = &This->base.device->range_pool;
 476     const int MaxStreams = This->base.device->caps.MaxStreams;
 477     unsigned s;
 478
 479     DBG("This=%p\n", This);
 480
 481     if (This->type == NINESBT_ALL)
 482         nine_state_copy_common_all(dst, src, src, TRUE, pool, MaxStreams);
 483     else
 484         nine_state_copy_common(dst, src, src, TRUE, pool);
 485
 486     if ((src->changed.group & NINE_STATE_VDECL) && src->vdecl)
 487         nine_bind(&dst->vdecl, src->vdecl);
 488
 489     /* Textures */
 490     if (src->changed.texture) {
 491         uint32_t m = src->changed.texture;
 492         dst->changed.texture |= m;
 493
 494         dst->samplers_shadow &= ~m;
 495
 496         for (s = 0; m; ++s, m >>= 1) {
 497             struct NineBaseTexture9 *tex = src->texture[s];
 498             if (!(m & 1))
 499                 continue;
 500             if (tex) {
 501                 tex->bind_count++;
 502                 if ((tex->dirty | tex->dirty_mip) && LIST_IS_EMPTY(&tex->list))
 503                     list_add(&tex->list, &This->base.device->update_textures);
 504                 dst->samplers_shadow |= tex->shadow << s;
 505             }
 506             if (src->texture[s])
 507                 src->texture[s]->bind_count--;
 508             nine_bind(&dst->texture[s], src->texture[s]);
 509         }
 510     }
 511
 512     return D3D_OK;
 513 }
 514
 515 IDirect3DStateBlock9Vtbl NineStateBlock9_vtable = {
 516     (void *)NineUnknown_QueryInterface,
 517     (void *)NineUnknown_AddRef,
 518     (void *)NineUnknown_Release,
 519     (void *)NineUnknown_GetDevice, /* actually part of StateBlock9 iface */
 520     (void *)NineStateBlock9_Capture,
 521     (void *)NineStateBlock9_Apply
 522 };
 523
 524 static const GUID *NineStateBlock9_IIDs[] = {
 525     &IID_IDirect3DStateBlock9,
 526     &IID_IUnknown,
 527     NULL
 528 };
 529
 530 HRESULT
 531 NineStateBlock9_new( struct NineDevice9 *pDevice,
 532                      struct NineStateBlock9 **ppOut,
 533                      enum nine_stateblock_type type)
 534 {
 535     NINE_DEVICE_CHILD_NEW(StateBlock9, ppOut, pDevice, type);
 536 }