src/gallium/drivers/i915/i915_state_emit.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2003 Tungsten Graphics, Inc., Cedar Park, Texas.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28
  29 #include "i915_reg.h"
  30 #include "i915_context.h"
  31 #include "i915_batch.h"
  32 #include "i915_debug.h"
  33 #include "i915_resource.h"
  34
  35 #include "pipe/p_context.h"
  36 #include "pipe/p_defines.h"
  37 #include "pipe/p_format.h"
  38
  39 #include "util/u_format.h"
  40 #include "util/u_math.h"
  41 #include "util/u_memory.h"
  42
  43 struct i915_tracked_hw_state {
  44    const char *name;
  45    void (*validate)(struct i915_context *, unsigned *batch_space);
  46    void (*emit)(struct i915_context *);
  47    unsigned dirty, batch_space;
  48 };
  49
  50
  51 static void
  52 validate_flush(struct i915_context *i915, unsigned *batch_space)
  53 {
  54    *batch_space = i915->flush_dirty ? 1 : 0;
  55 }
  56
  57 static void
  58 emit_flush(struct i915_context *i915)
  59 {
  60    /* Cache handling is very cheap atm. State handling can request to flushes:
  61     * - I915_FLUSH_CACHE which is a flush everything request and
  62     * - I915_PIPELINE_FLUSH which is specifically for the draw_offset flush.
  63     * Because the cache handling is so dumb, no explicit "invalidate map cache".
  64     * Also, the first is a strict superset of the latter, so the following logic
  65     * works. */
  66    if (i915->flush_dirty & I915_FLUSH_CACHE)
  67       OUT_BATCH(MI_FLUSH | FLUSH_MAP_CACHE);
  68    else if (i915->flush_dirty & I915_PIPELINE_FLUSH)
  69       OUT_BATCH(MI_FLUSH | INHIBIT_FLUSH_RENDER_CACHE);
  70 }
  71
  72 uint32_t invariant_state[] = {
  73    _3DSTATE_AA_CMD | AA_LINE_ECAAR_WIDTH_ENABLE | AA_LINE_ECAAR_WIDTH_1_0 |
  74              AA_LINE_REGION_WIDTH_ENABLE | AA_LINE_REGION_WIDTH_1_0,
  75
  76    _3DSTATE_DFLT_DIFFUSE_CMD, 0,
  77
  78    _3DSTATE_DFLT_SPEC_CMD, 0,
  79
  80    _3DSTATE_DFLT_Z_CMD, 0,
  81
  82    _3DSTATE_COORD_SET_BINDINGS |
  83              CSB_TCB(0, 0) |
  84              CSB_TCB(1, 1) |
  85              CSB_TCB(2, 2) |
  86              CSB_TCB(3, 3) |
  87              CSB_TCB(4, 4) |
  88              CSB_TCB(5, 5) |
  89              CSB_TCB(6, 6) |
  90              CSB_TCB(7, 7),
  91
  92    _3DSTATE_RASTER_RULES_CMD |
  93              ENABLE_POINT_RASTER_RULE |
  94              OGL_POINT_RASTER_RULE |
  95              ENABLE_LINE_STRIP_PROVOKE_VRTX |
  96              ENABLE_TRI_FAN_PROVOKE_VRTX |
  97              LINE_STRIP_PROVOKE_VRTX(1) |
  98              TRI_FAN_PROVOKE_VRTX(2) |
  99              ENABLE_TEXKILL_3D_4D |
 100              TEXKILL_4D,
 101
 102    _3DSTATE_DEPTH_SUBRECT_DISABLE,
 103
 104    /* disable indirect state for now
 105     */
 106    _3DSTATE_LOAD_INDIRECT | 0, 0};
 107
 108 static void
 109 emit_invariant(struct i915_context *i915)
 110 {
 111    i915_winsys_batchbuffer_write(i915->batch, invariant_state,
 112                                  Elements(invariant_state)*sizeof(uint32_t));
 113 }
 114
 115 static void
 116 validate_immediate(struct i915_context *i915, unsigned *batch_space)
 117 {
 118    unsigned dirty = (1 << I915_IMMEDIATE_S0 | 1 << I915_IMMEDIATE_S1 |
 119                      1 << I915_IMMEDIATE_S2 | 1 << I915_IMMEDIATE_S3 |
 120                      1 << I915_IMMEDIATE_S3 | 1 << I915_IMMEDIATE_S4 |
 121                      1 << I915_IMMEDIATE_S5 | 1 << I915_IMMEDIATE_S6) &
 122                     i915->immediate_dirty;
 123
 124    if (i915->immediate_dirty & (1 << I915_IMMEDIATE_S0) && i915->vbo)
 125       i915->validation_buffers[i915->num_validation_buffers++] = i915->vbo;
 126
 127    *batch_space = 1 + util_bitcount(dirty);
 128 }
 129
 130 static void
 131 emit_immediate(struct i915_context *i915)
 132 {
 133    /* remove unwanted bits and S7 */
 134    unsigned dirty = (1 << I915_IMMEDIATE_S0 | 1 << I915_IMMEDIATE_S1 |
 135                      1 << I915_IMMEDIATE_S2 | 1 << I915_IMMEDIATE_S3 |
 136                      1 << I915_IMMEDIATE_S3 | 1 << I915_IMMEDIATE_S4 |
 137                      1 << I915_IMMEDIATE_S5 | 1 << I915_IMMEDIATE_S6) &
 138                     i915->immediate_dirty;
 139    int i, num = util_bitcount(dirty);
 140    assert(num && num <= I915_MAX_IMMEDIATE);
 141
 142    OUT_BATCH(_3DSTATE_LOAD_STATE_IMMEDIATE_1 |
 143              dirty << 4 | (num - 1));
 144
 145    if (i915->immediate_dirty & (1 << I915_IMMEDIATE_S0)) {
 146       if (i915->vbo)
 147          OUT_RELOC(i915->vbo, I915_USAGE_VERTEX,
 148                    i915->current.immediate[I915_IMMEDIATE_S0]);
 149       else
 150          OUT_BATCH(0);
 151    }
 152
 153    for (i = 1; i < I915_MAX_IMMEDIATE; i++) {
 154       if (dirty & (1 << i))
 155          OUT_BATCH(i915->current.immediate[i]);
 156    }
 157 }
 158
 159 static void
 160 validate_dynamic(struct i915_context *i915, unsigned *batch_space)
 161 {
 162    *batch_space = util_bitcount(i915->dynamic_dirty & ((1 << I915_MAX_DYNAMIC) - 1));
 163 }
 164
 165 static void
 166 emit_dynamic(struct i915_context *i915)
 167 {
 168    int i;
 169    for (i = 0; i < I915_MAX_DYNAMIC; i++) {
 170       if (i915->dynamic_dirty & (1 << i))
 171          OUT_BATCH(i915->current.dynamic[i]);
 172    }
 173 }
 174
 175 static void
 176 validate_static(struct i915_context *i915, unsigned *batch_space)
 177 {
 178    *batch_space = 0;
 179
 180    if (i915->current.cbuf_bo && (i915->static_dirty & I915_DST_BUF_COLOR)) {
 181       i915->validation_buffers[i915->num_validation_buffers++]
 182          = i915->current.cbuf_bo;
 183       *batch_space += 3;
 184    }
 185
 186    if (i915->current.depth_bo && (i915->static_dirty & I915_DST_BUF_DEPTH)) {
 187       i915->validation_buffers[i915->num_validation_buffers++]
 188          = i915->current.depth_bo;
 189       *batch_space += 3;
 190    }
 191
 192    if (i915->static_dirty & I915_DST_VARS)
 193       *batch_space += 2;
 194
 195    if (i915->static_dirty & I915_DST_RECT)
 196       *batch_space += 5;
 197 }
 198
 199 static void
 200 emit_static(struct i915_context *i915)
 201 {
 202    if (i915->current.cbuf_bo && (i915->static_dirty & I915_DST_BUF_COLOR)) {
 203       OUT_BATCH(_3DSTATE_BUF_INFO_CMD);
 204       OUT_BATCH(i915->current.cbuf_flags);
 205       OUT_RELOC(i915->current.cbuf_bo,
 206                 I915_USAGE_RENDER,
 207                 0);
 208    }
 209
 210    /* What happens if no zbuf??
 211     */
 212    if (i915->current.depth_bo && (i915->static_dirty & I915_DST_BUF_DEPTH)) {
 213       OUT_BATCH(_3DSTATE_BUF_INFO_CMD);
 214       OUT_BATCH(i915->current.depth_flags);
 215       OUT_RELOC(i915->current.depth_bo,
 216                 I915_USAGE_RENDER,
 217                 0);
 218    }
 219
 220    if (i915->static_dirty & I915_DST_VARS) {
 221       OUT_BATCH(_3DSTATE_DST_BUF_VARS_CMD);
 222       OUT_BATCH(i915->current.dst_buf_vars);
 223    }
 224 }
 225
 226 static void
 227 validate_map(struct i915_context *i915, unsigned *batch_space)
 228 {
 229    const uint enabled = i915->current.sampler_enable_flags;
 230    uint unit;
 231    struct i915_texture *tex;
 232
 233    *batch_space = i915->current.sampler_enable_nr ?
 234      2 + 3*i915->current.sampler_enable_nr : 0;
 235
 236    for (unit = 0; unit < I915_TEX_UNITS; unit++) {
 237       if (enabled & (1 << unit)) {
 238          tex = i915_texture(i915->fragment_sampler_views[unit]->texture);
 239          i915->validation_buffers[i915->num_validation_buffers++] = tex->buffer;
 240       }
 241    }
 242 }
 243
 244 static void
 245 emit_map(struct i915_context *i915)
 246 {
 247    const uint nr = i915->current.sampler_enable_nr;
 248    if (nr) {
 249       const uint enabled = i915->current.sampler_enable_flags;
 250       uint unit;
 251       uint count = 0;
 252       OUT_BATCH(_3DSTATE_MAP_STATE | (3 * nr));
 253       OUT_BATCH(enabled);
 254       for (unit = 0; unit < I915_TEX_UNITS; unit++) {
 255          if (enabled & (1 << unit)) {
 256             struct i915_texture *texture = i915_texture(i915->fragment_sampler_views[unit]->texture);
 257             struct i915_winsys_buffer *buf = texture->buffer;
 258             assert(buf);
 259
 260             count++;
 261
 262             OUT_RELOC(buf, I915_USAGE_SAMPLER, 0);
 263             OUT_BATCH(i915->current.texbuffer[unit][0]); /* MS3 */
 264             OUT_BATCH(i915->current.texbuffer[unit][1]); /* MS4 */
 265          }
 266       }
 267       assert(count == nr);
 268    }
 269 }
 270
 271 static void
 272 validate_sampler(struct i915_context *i915, unsigned *batch_space)
 273 {
 274    *batch_space = i915->current.sampler_enable_nr ?
 275      2 + 3*i915->current.sampler_enable_nr : 0;
 276 }
 277
 278 static void
 279 emit_sampler(struct i915_context *i915)
 280 {
 281    if (i915->current.sampler_enable_nr) {
 282       int i;
 283
 284       OUT_BATCH( _3DSTATE_SAMPLER_STATE |
 285                  (3 * i915->current.sampler_enable_nr) );
 286
 287       OUT_BATCH( i915->current.sampler_enable_flags );
 288
 289       for (i = 0; i < I915_TEX_UNITS; i++) {
 290          if (i915->current.sampler_enable_flags & (1<<i)) {
 291             OUT_BATCH( i915->current.sampler[i][0] );
 292             OUT_BATCH( i915->current.sampler[i][1] );
 293             OUT_BATCH( i915->current.sampler[i][2] );
 294          }
 295       }
 296    }
 297 }
 298
 299 static void
 300 validate_constants(struct i915_context *i915, unsigned *batch_space)
 301 {
 302    *batch_space = i915->fs->num_constants ?
 303       2 + 4*i915->fs->num_constants : 0;
 304 }
 305
 306 static void
 307 emit_constants(struct i915_context *i915)
 308 {
 309    /* Collate the user-defined constants with the fragment shader's
 310     * immediates according to the constant_flags[] array.
 311     */
 312    const uint nr = i915->fs->num_constants;
 313    if (nr) {
 314       uint i;
 315
 316       OUT_BATCH( _3DSTATE_PIXEL_SHADER_CONSTANTS | (nr * 4) );
 317       OUT_BATCH((1 << nr) - 1);
 318
 319       for (i = 0; i < nr; i++) {
 320          const uint *c;
 321          if (i915->fs->constant_flags[i] == I915_CONSTFLAG_USER) {
 322             /* grab user-defined constant */
 323             c = (uint *) i915_buffer(i915->constants[PIPE_SHADER_FRAGMENT])->data;
 324             c += 4 * i;
 325          }
 326          else {
 327             /* emit program constant */
 328             c = (uint *) i915->fs->constants[i];
 329          }
 330 #if 0 /* debug */
 331          {
 332             float *f = (float *) c;
 333             printf("Const %2d: %f %f %f %f %s\n", i, f[0], f[1], f[2], f[3],
 334                    (i915->fs->constant_flags[i] == I915_CONSTFLAG_USER
 335                     ? "user" : "immediate"));
 336          }
 337 #endif
 338          OUT_BATCH(*c++);
 339          OUT_BATCH(*c++);
 340          OUT_BATCH(*c++);
 341          OUT_BATCH(*c++);
 342       }
 343    }
 344 }
 345
 346 static const struct
 347 {
 348    enum pipe_format format;
 349    uint hw_swizzle;
 350 } fixup_formats[] = {
 351    { PIPE_FORMAT_R8G8B8A8_UNORM, 0x21030000 /* BGRA */},
 352    { PIPE_FORMAT_L8_UNORM,       0x00030000 /* RRRA */},
 353    { PIPE_FORMAT_I8_UNORM,       0x00030000 /* RRRA */},
 354    { PIPE_FORMAT_A8_UNORM,       0x33330000 /* AAAA */},
 355    { PIPE_FORMAT_NONE,           0x00000000},
 356 };
 357
 358 static uint need_target_fixup(struct pipe_surface* p)
 359 {
 360    enum pipe_format f;
 361    /* if we don't have a surface bound yet, we don't need to fixup the shader */
 362    if (!p)
 363       return 0;
 364
 365    f = p->format;
 366    for(int i=0; fixup_formats[i].format != PIPE_FORMAT_NONE; i++)
 367       if (fixup_formats[i].format == f)
 368          return 1;
 369
 370    return 0;
 371 }
 372
 373 static uint fixup_swizzle(enum pipe_format f)
 374 {
 375    for(int i=0; fixup_formats[i].format != PIPE_FORMAT_NONE; i++)
 376       if (fixup_formats[i].format == f)
 377          return fixup_formats[i].hw_swizzle;
 378
 379    return 0;
 380 }
 381
 382 static void
 383 validate_program(struct i915_context *i915, unsigned *batch_space)
 384 {
 385    struct pipe_surface *cbuf_surface = i915->framebuffer.cbufs[0];
 386    uint additional_size = need_target_fixup(cbuf_surface);
 387
 388    /* we need more batch space if we want to emulate rgba framebuffers */
 389    *batch_space = i915->fs->program_len + 3 * additional_size;
 390 }
 391
 392 static void
 393 emit_program(struct i915_context *i915)
 394 {
 395    struct pipe_surface *cbuf_surface = i915->framebuffer.cbufs[0];
 396    uint target_fixup = need_target_fixup(cbuf_surface);
 397    uint i;
 398
 399    /* we should always have, at least, a pass-through program */
 400    assert(i915->fs->program_len > 0);
 401
 402    {
 403       /* first word has the size, we have to adjust that */
 404       uint size = (i915->fs->program[0]);
 405       size += target_fixup * 3;
 406       OUT_BATCH(size);
 407    }
 408
 409    /* output the declarations of the program */
 410    for (i=1 ; i < i915->fs->program_len; i++)
 411       OUT_BATCH(i915->fs->program[i]);
 412
 413    /* we emit an additional mov with swizzle to fake RGBA framebuffers */
 414    if (target_fixup) {
 415       /* mov out_color, out_color.zyxw */
 416       OUT_BATCH(A0_MOV |
 417                 (REG_TYPE_OC << A0_DEST_TYPE_SHIFT) |
 418                 A0_DEST_CHANNEL_ALL |
 419                 (REG_TYPE_OC << A0_SRC0_TYPE_SHIFT) |
 420                 (T_DIFFUSE << A0_SRC0_NR_SHIFT));
 421       OUT_BATCH(fixup_swizzle(cbuf_surface->format));
 422       OUT_BATCH(0);
 423    }
 424 }
 425
 426 static void
 427 emit_draw_rect(struct i915_context *i915)
 428 {
 429    if (i915->static_dirty & I915_DST_RECT) {
 430       OUT_BATCH(_3DSTATE_DRAW_RECT_CMD);
 431       OUT_BATCH(DRAW_RECT_DIS_DEPTH_OFS);
 432       OUT_BATCH(i915->current.draw_offset);
 433       OUT_BATCH(i915->current.draw_size);
 434       OUT_BATCH(i915->current.draw_offset);
 435    }
 436 }
 437
 438 static boolean
 439 i915_validate_state(struct i915_context *i915, unsigned *batch_space)
 440 {
 441    unsigned tmp;
 442
 443    i915->num_validation_buffers = 0;
 444    if (i915->hardware_dirty & I915_HW_INVARIANT)
 445       *batch_space = Elements(invariant_state);
 446    else
 447       *batch_space = 0;
 448
 449 #define VALIDATE_ATOM(atom, hw_dirty) \
 450    if (i915->hardware_dirty & hw_dirty) { \
 451       validate_##atom(i915, &tmp); \
 452       *batch_space += tmp; }
 453    VALIDATE_ATOM(flush, I915_HW_FLUSH);
 454    VALIDATE_ATOM(immediate, I915_HW_IMMEDIATE);
 455    VALIDATE_ATOM(dynamic, I915_HW_DYNAMIC);
 456    VALIDATE_ATOM(static, I915_HW_STATIC);
 457    VALIDATE_ATOM(map, I915_HW_MAP);
 458    VALIDATE_ATOM(sampler, I915_HW_SAMPLER);
 459    VALIDATE_ATOM(constants, I915_HW_CONSTANTS);
 460    VALIDATE_ATOM(program, I915_HW_PROGRAM);
 461 #undef VALIDATE_ATOM
 462
 463    if (i915->num_validation_buffers == 0)
 464       return TRUE;
 465
 466    if (!i915_winsys_validate_buffers(i915->batch, i915->validation_buffers,
 467                                      i915->num_validation_buffers))
 468       return FALSE;
 469
 470    return TRUE;
 471 }
 472
 473 /* Push the state into the sarea and/or texture memory.
 474  */
 475 void
 476 i915_emit_hardware_state(struct i915_context *i915 )
 477 {
 478    unsigned batch_space;
 479    uintptr_t save_ptr;
 480
 481    assert(i915->dirty == 0);
 482
 483    if (I915_DBG_ON(DBG_ATOMS))
 484       i915_dump_hardware_dirty(i915, __FUNCTION__);
 485
 486    if (!i915_validate_state(i915, &batch_space)) {
 487       FLUSH_BATCH(NULL);
 488       assert(i915_validate_state(i915, &batch_space));
 489    }
 490
 491    if(!BEGIN_BATCH(batch_space)) {
 492       FLUSH_BATCH(NULL);
 493       assert(i915_validate_state(i915, &batch_space));
 494       assert(BEGIN_BATCH(batch_space));
 495    }
 496
 497    save_ptr = (uintptr_t)i915->batch->ptr;
 498
 499 #define EMIT_ATOM(atom, hw_dirty) \
 500    if (i915->hardware_dirty & hw_dirty) \
 501       emit_##atom(i915);
 502    EMIT_ATOM(flush, I915_HW_FLUSH);
 503    EMIT_ATOM(invariant, I915_HW_INVARIANT);
 504    EMIT_ATOM(immediate, I915_HW_IMMEDIATE);
 505    EMIT_ATOM(dynamic, I915_HW_DYNAMIC);
 506    EMIT_ATOM(static, I915_HW_STATIC);
 507    EMIT_ATOM(map, I915_HW_MAP);
 508    EMIT_ATOM(sampler, I915_HW_SAMPLER);
 509    EMIT_ATOM(constants, I915_HW_CONSTANTS);
 510    EMIT_ATOM(program, I915_HW_PROGRAM);
 511    EMIT_ATOM(draw_rect, I915_HW_STATIC);
 512 #undef EMIT_ATOM
 513
 514    I915_DBG(DBG_EMIT, "%s: used %d dwords, %d dwords reserved\n", __FUNCTION__,
 515             ((uintptr_t)i915->batch->ptr - save_ptr) / 4,
 516             batch_space);
 517    assert(((uintptr_t)i915->batch->ptr - save_ptr) / 4 == batch_space);
 518
 519    i915->hardware_dirty = 0;
 520    i915->immediate_dirty = 0;
 521    i915->dynamic_dirty = 0;
 522    i915->static_dirty = 0;
 523    i915->flush_dirty = 0;
 524 }