src/gallium/drivers/r300/r300_hyperz.c

   1 /*
   2  * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
   3  * Copyright 2009 Marek Olšák <maraeo@gmail.com>
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * on the rights to use, copy, modify, merge, publish, distribute, sub
   9  * license, and/or sell copies of the Software, and to permit persons to whom
  10  * the Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  22  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
  23
  24 #include "r300_context.h"
  25 #include "r300_hyperz.h"
  26 #include "r300_reg.h"
  27 #include "r300_fs.h"
  28 #include "r300_winsys.h"
  29
  30 #include "util/u_format.h"
  31 #include "util/u_mm.h"
  32
  33 /*
  34   HiZ rules - taken from various docs
  35    1. HiZ only works on depth values
  36    2. Cannot HiZ if stencil fail or zfail is !KEEP
  37    3. on R300/400, HiZ is disabled if depth test is EQUAL
  38    4. comparison changes without clears usually mean disabling HiZ
  39 */
  40 /*****************************************************************************/
  41 /* The HyperZ setup                                                          */
  42 /*****************************************************************************/
  43
  44 static bool r300_get_sc_hz_max(struct r300_context *r300)
  45 {
  46     struct r300_dsa_state *dsa_state = r300->dsa_state.state;
  47     int func = dsa_state->z_stencil_control & R300_ZS_MASK;
  48     int ret = R300_SC_HYPERZ_MIN;
  49
  50     if (func >= R300_ZS_GEQUAL && func <= R300_ZS_ALWAYS)
  51        ret = R300_SC_HYPERZ_MAX;
  52     return ret;
  53 }
  54
  55 static bool r300_zfunc_same_direction(int func1, int func2)
  56 {
  57     /* func1 is less/lessthan */
  58     if ((func1 == R300_ZS_LESS || func1 == R300_ZS_LEQUAL) &&
  59         (func2 == R300_ZS_EQUAL || func2 == R300_ZS_GEQUAL ||
  60          func2 == R300_ZS_GREATER))
  61             return FALSE;
  62
  63     /* func1 is greater/greaterthan */
  64     if ((func1 == R300_ZS_GEQUAL || func1 == R300_ZS_GREATER) &&
  65         (func2 == R300_ZS_LESS || func2 == R300_ZS_LEQUAL))
  66             return FALSE;
  67
  68     return TRUE;
  69 }
  70
  71 static int r300_get_hiz_min(struct r300_context *r300)
  72 {
  73     struct r300_dsa_state *dsa_state = r300->dsa_state.state;
  74     int func = dsa_state->z_stencil_control & R300_ZS_MASK;
  75     int ret = R300_HIZ_MIN;
  76
  77     if (func == R300_ZS_LESS || func == R300_ZS_LEQUAL)
  78        ret = R300_HIZ_MAX;
  79     return ret;
  80 }
  81
  82 static boolean r300_dsa_stencil_op_not_keep(struct pipe_stencil_state *s)
  83 {
  84     if (s->enabled && (s->fail_op != PIPE_STENCIL_OP_KEEP ||
  85                        s->zfail_op != PIPE_STENCIL_OP_KEEP))
  86         return TRUE;
  87     return FALSE;
  88 }
  89
  90 static boolean r300_can_hiz(struct r300_context *r300)
  91 {
  92     struct r300_dsa_state *dsa_state = r300->dsa_state.state;
  93     struct pipe_depth_stencil_alpha_state *dsa = &dsa_state->dsa;
  94     struct r300_screen* r300screen = r300->screen;
  95     struct r300_hyperz_state *z = r300->hyperz_state.state;
  96
  97     /* shader writes depth - no HiZ */
  98     if (r300_fragment_shader_writes_depth(r300_fs(r300))) /* (5) */
  99         return FALSE;
 100
 101     if (r300->query_current)
 102         return FALSE;
 103     /* if stencil fail/zfail op is not KEEP */
 104     if (r300_dsa_stencil_op_not_keep(&dsa->stencil[0]) ||
 105         r300_dsa_stencil_op_not_keep(&dsa->stencil[1]))
 106         return FALSE;
 107
 108     if (dsa->depth.enabled) {
 109         /* if depth func is EQUAL pre-r500 */
 110         if (dsa->depth.func == PIPE_FUNC_EQUAL && !r300screen->caps.is_r500)
 111             return FALSE;
 112         /* if depth func is NOTEQUAL */
 113         if (dsa->depth.func == PIPE_FUNC_NOTEQUAL)
 114             return FALSE;
 115     }
 116     /* depth comparison function - if just cleared save and return okay */
 117     if (z->current_func == -1) {
 118         int func = dsa_state->z_stencil_control & R300_ZS_MASK;
 119         if (func != 0 && func != 7)
 120             z->current_func = dsa_state->z_stencil_control & R300_ZS_MASK;
 121     } else {
 122         /* simple don't change */
 123         if (!r300_zfunc_same_direction(z->current_func,
 124                                        (dsa_state->z_stencil_control & R300_ZS_MASK))) {
 125             DBG(r300, DBG_HYPERZ,
 126                 "z func changed direction - disabling hyper-z %d -> %d\n",
 127                 z->current_func, dsa_state->z_stencil_control);
 128             return FALSE;
 129         }
 130     }
 131     return TRUE;
 132 }
 133
 134 static void r300_update_hyperz(struct r300_context* r300)
 135 {
 136     struct r300_hyperz_state *z =
 137         (struct r300_hyperz_state*)r300->hyperz_state.state;
 138     struct pipe_framebuffer_state *fb =
 139         (struct pipe_framebuffer_state*)r300->fb_state.state;
 140     struct r300_texture *zstex =
 141             fb->zsbuf ? r300_texture(fb->zsbuf->texture) : NULL;
 142     boolean zmask_in_use = FALSE;
 143     boolean hiz_in_use = FALSE;
 144
 145     z->gb_z_peq_config = 0;
 146     z->zb_bw_cntl = 0;
 147     z->sc_hyperz = R300_SC_HYPERZ_ADJ_2;
 148     z->flush = 0;
 149
 150     if (r300->cbzb_clear) {
 151         z->zb_bw_cntl |= R300_ZB_CB_CLEAR_CACHE_LINE_WRITE_ONLY;
 152         return;
 153     }
 154
 155     if (!zstex)
 156         return;
 157
 158     if (!r300->rws->get_value(r300->rws, R300_CAN_HYPERZ))
 159         return;
 160
 161     zmask_in_use = zstex->zmask_in_use[fb->zsbuf->level];
 162     hiz_in_use = zstex->hiz_in_use[fb->zsbuf->level];
 163
 164     /* Z fastfill. */
 165     if (zmask_in_use) {
 166         z->zb_bw_cntl |= R300_FAST_FILL_ENABLE; /*  | R300_FORCE_COMPRESSED_STENCIL_VALUE_ENABLE;*/
 167     }
 168
 169     /* Zbuffer compression. */
 170     if (zmask_in_use && r300->z_compression) {
 171         z->zb_bw_cntl |= R300_RD_COMP_ENABLE;
 172         if (r300->z_decomp_rd == false)
 173             z->zb_bw_cntl |= R300_WR_COMP_ENABLE;
 174     }
 175     /* RV350 and up optimizations. */
 176     /* The section 10.4.9 in the docs is a lie. */
 177     if (r300->z_compression == RV350_Z_COMPRESS_88)
 178         z->gb_z_peq_config |= R300_GB_Z_PEQ_CONFIG_Z_PEQ_SIZE_8_8;
 179
 180     if (hiz_in_use) {
 181         bool can_hiz = r300_can_hiz(r300);
 182         if (can_hiz) {
 183             z->zb_bw_cntl |= R300_HIZ_ENABLE;
 184             z->sc_hyperz |= R300_SC_HYPERZ_ENABLE;
 185             z->sc_hyperz |= r300_get_sc_hz_max(r300);
 186             z->zb_bw_cntl |= r300_get_hiz_min(r300);
 187         }
 188     }
 189
 190     /* R500-specific features and optimizations. */
 191     if (r300->screen->caps.is_r500) {
 192         z->zb_bw_cntl |= R500_HIZ_FP_EXP_BITS_3;
 193         z->zb_bw_cntl |=
 194                 R500_HIZ_EQUAL_REJECT_ENABLE |
 195                 R500_PEQ_PACKING_ENABLE |
 196                 R500_COVERED_PTR_MASKING_ENABLE;
 197     }
 198 }
 199
 200 /*****************************************************************************/
 201 /* The ZTOP state                                                            */
 202 /*****************************************************************************/
 203
 204 static boolean r300_dsa_writes_stencil(
 205         struct pipe_stencil_state *s)
 206 {
 207     return s->enabled && s->writemask &&
 208            (s->fail_op  != PIPE_STENCIL_OP_KEEP ||
 209             s->zfail_op != PIPE_STENCIL_OP_KEEP ||
 210             s->zpass_op != PIPE_STENCIL_OP_KEEP);
 211 }
 212
 213 static boolean r300_dsa_writes_depth_stencil(
 214         struct pipe_depth_stencil_alpha_state *dsa)
 215 {
 216     /* We are interested only in the cases when a depth or stencil value
 217      * can be changed. */
 218
 219     if (dsa->depth.enabled && dsa->depth.writemask &&
 220         dsa->depth.func != PIPE_FUNC_NEVER)
 221         return TRUE;
 222
 223     if (r300_dsa_writes_stencil(&dsa->stencil[0]) ||
 224         r300_dsa_writes_stencil(&dsa->stencil[1]))
 225         return TRUE;
 226
 227     return FALSE;
 228 }
 229
 230 static boolean r300_dsa_alpha_test_enabled(
 231         struct pipe_depth_stencil_alpha_state *dsa)
 232 {
 233     /* We are interested only in the cases when alpha testing can kill
 234      * a fragment. */
 235
 236     return dsa->alpha.enabled && dsa->alpha.func != PIPE_FUNC_ALWAYS;
 237 }
 238
 239 static void r300_update_ztop(struct r300_context* r300)
 240 {
 241     struct r300_ztop_state* ztop_state =
 242         (struct r300_ztop_state*)r300->ztop_state.state;
 243     uint32_t old_ztop = ztop_state->z_buffer_top;
 244
 245     /* This is important enough that I felt it warranted a comment.
 246      *
 247      * According to the docs, these are the conditions where ZTOP must be
 248      * disabled:
 249      * 1) Alpha testing enabled
 250      * 2) Texture kill instructions in fragment shader
 251      * 3) Chroma key culling enabled
 252      * 4) W-buffering enabled
 253      *
 254      * The docs claim that for the first three cases, if no ZS writes happen,
 255      * then ZTOP can be used.
 256      *
 257      * (3) will never apply since we do not support chroma-keyed operations.
 258      * (4) will need to be re-examined (and this comment updated) if/when
 259      * Hyper-Z becomes supported.
 260      *
 261      * Additionally, the following conditions require disabled ZTOP:
 262      * 5) Depth writes in fragment shader
 263      * 6) Outstanding occlusion queries
 264      *
 265      * This register causes stalls all the way from SC to CB when changed,
 266      * but it is buffered on-chip so it does not hurt to write it if it has
 267      * not changed.
 268      *
 269      * ~C.
 270      */
 271
 272     /* ZS writes */
 273     if (r300_dsa_writes_depth_stencil(r300->dsa_state.state) &&
 274            (r300_dsa_alpha_test_enabled(r300->dsa_state.state) ||  /* (1) */
 275             r300_fs(r300)->shader->info.uses_kill)) {              /* (2) */
 276         ztop_state->z_buffer_top = R300_ZTOP_DISABLE;
 277     } else if (r300_fragment_shader_writes_depth(r300_fs(r300))) { /* (5) */
 278         ztop_state->z_buffer_top = R300_ZTOP_DISABLE;
 279     } else if (r300->query_current) {                              /* (6) */
 280         ztop_state->z_buffer_top = R300_ZTOP_DISABLE;
 281     } else {
 282         ztop_state->z_buffer_top = R300_ZTOP_ENABLE;
 283     }
 284     if (ztop_state->z_buffer_top != old_ztop)
 285         r300->ztop_state.dirty = TRUE;
 286 }
 287
 288 #define ALIGN_DIVUP(x, y) (((x) + (y) - 1) / (y))
 289
 290 static void r300_update_hiz_clear(struct r300_context *r300)
 291 {
 292     struct pipe_framebuffer_state *fb =
 293         (struct pipe_framebuffer_state*)r300->fb_state.state;
 294     uint32_t height;
 295
 296     height = ALIGN_DIVUP(fb->zsbuf->height, 4);
 297     r300->hiz_clear.size = height * 4;
 298 }
 299
 300 static void r300_update_zmask_clear(struct r300_context *r300)
 301 {
 302     struct pipe_framebuffer_state *fb =
 303         (struct pipe_framebuffer_state*)r300->fb_state.state;
 304     uint32_t height;
 305     int mult;
 306
 307     if (r300->z_compression == RV350_Z_COMPRESS_88)
 308         mult = 8;
 309     else
 310         mult = 4;
 311
 312     height = ALIGN_DIVUP(fb->zsbuf->height, mult);
 313
 314     r300->zmask_clear.size = height * 4;
 315 }
 316
 317 void r300_update_hyperz_state(struct r300_context* r300)
 318 {
 319     r300_update_ztop(r300);
 320     if (r300->hyperz_state.dirty) {
 321         r300_update_hyperz(r300);
 322     }
 323
 324     if (r300->hiz_clear.dirty) {
 325        r300_update_hiz_clear(r300);
 326     }
 327     if (r300->zmask_clear.dirty) {
 328        r300_update_zmask_clear(r300);
 329     }
 330 }
 331
 332 void r300_hiz_alloc_block(struct r300_context *r300, struct r300_surface *surf)
 333 {
 334     struct r300_texture *tex;
 335     uint32_t zsize, ndw;
 336     int level = surf->base.level;
 337
 338     tex = r300_texture(surf->base.texture);
 339
 340     if (tex->hiz_mem[level])
 341         return;
 342
 343     zsize = tex->desc.layer_size_in_bytes[level];
 344     zsize /= util_format_get_blocksize(tex->desc.b.b.format);
 345     ndw = ALIGN_DIVUP(zsize, 64);
 346
 347     tex->hiz_mem[level] = u_mmAllocMem(r300->hiz_mm, ndw, 0, 0);
 348     return;
 349 }
 350
 351 void r300_zmask_alloc_block(struct r300_context *r300, struct r300_surface *surf, int compress)
 352 {
 353     int bsize = 256;
 354     uint32_t zsize, ndw;
 355     int level = surf->base.level;
 356     struct r300_texture *tex;
 357
 358     tex = r300_texture(surf->base.texture);
 359
 360     /* We currently don't handle decompression for 3D textures and cubemaps
 361      * correctly. */
 362     if (tex->desc.b.b.target != PIPE_TEXTURE_1D &&
 363         tex->desc.b.b.target != PIPE_TEXTURE_2D &&
 364         tex->desc.b.b.target != PIPE_TEXTURE_RECT)
 365         return;
 366
 367     /* Cannot flush zmask of 16-bit zbuffers. */
 368     if (util_format_get_blocksizebits(tex->desc.b.b.format) == 16)
 369         return;
 370
 371     if (tex->zmask_mem[level])
 372         return;
 373
 374     zsize = tex->desc.layer_size_in_bytes[level];
 375     zsize /= util_format_get_blocksize(tex->desc.b.b.format);
 376
 377     /* each zmask dword represents 16 4x4 blocks - which is 256 pixels
 378        or 16 8x8 depending on the gb peq flag = 1024 pixels */
 379     if (compress == RV350_Z_COMPRESS_88)
 380         bsize = 1024;
 381
 382     ndw = ALIGN_DIVUP(zsize, bsize);
 383     tex->zmask_mem[level] = u_mmAllocMem(r300->zmask_mm, ndw, 0, 0);
 384     return;
 385 }
 386
 387 boolean r300_hyperz_init_mm(struct r300_context *r300)
 388 {
 389     struct r300_screen* r300screen = r300->screen;
 390     int frag_pipes = r300screen->caps.num_frag_pipes;
 391
 392     r300->zmask_mm = u_mmInit(0, r300screen->caps.zmask_ram * frag_pipes);
 393     if (!r300->zmask_mm)
 394       return FALSE;
 395
 396     if (r300screen->caps.hiz_ram) {
 397       r300->hiz_mm = u_mmInit(0, r300screen->caps.hiz_ram * frag_pipes);
 398       if (!r300->hiz_mm) {
 399         u_mmDestroy(r300->zmask_mm);
 400         r300->zmask_mm = NULL;
 401         return FALSE;
 402       }
 403     }
 404
 405     return TRUE;
 406 }
 407
 408 void r300_hyperz_destroy_mm(struct r300_context *r300)
 409 {
 410     struct r300_screen* r300screen = r300->screen;
 411
 412     if (r300screen->caps.hiz_ram) {
 413       u_mmDestroy(r300->hiz_mm);
 414       r300->hiz_mm = NULL;
 415     }
 416
 417     u_mmDestroy(r300->zmask_mm);
 418     r300->zmask_mm = NULL;
 419 }