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          /* Fixup blend function for A8 dst buffers.
 156           * When we blend to an A8 buffer, the GPU thinks it's a G8 buffer,
 157           * and therefore we need to use the color factor for alphas. */
 158          if ((i == I915_IMMEDIATE_S6) &&
 159              (i915->current.target_fixup_format == PIPE_FORMAT_A8_UNORM)) {
 160             uint32_t imm = i915->current.immediate[i];
 161             uint32_t srcRGB = (imm >> S6_CBUF_SRC_BLEND_FACT_SHIFT) & BLENDFACT_MASK;
 162             if (srcRGB == BLENDFACT_DST_ALPHA)
 163                srcRGB = BLENDFACT_DST_COLR;
 164             else if (srcRGB == BLENDFACT_INV_DST_ALPHA)
 165                srcRGB = BLENDFACT_INV_DST_COLR;
 166             imm &= ~SRC_BLND_FACT(BLENDFACT_MASK);
 167             imm |= SRC_BLND_FACT(srcRGB);
 168             OUT_BATCH(imm);
 169          } else {
 170             OUT_BATCH(i915->current.immediate[i]);
 171          }
 172       }
 173    }
 174 }
 175
 176 static void
 177 validate_dynamic(struct i915_context *i915, unsigned *batch_space)
 178 {
 179    *batch_space = util_bitcount(i915->dynamic_dirty & ((1 << I915_MAX_DYNAMIC) - 1));
 180 }
 181
 182 static void
 183 emit_dynamic(struct i915_context *i915)
 184 {
 185    int i;
 186    for (i = 0; i < I915_MAX_DYNAMIC; i++) {
 187       if (i915->dynamic_dirty & (1 << i))
 188          OUT_BATCH(i915->current.dynamic[i]);
 189    }
 190 }
 191
 192 static void
 193 validate_static(struct i915_context *i915, unsigned *batch_space)
 194 {
 195    *batch_space = 0;
 196
 197    if (i915->current.cbuf_bo && (i915->static_dirty & I915_DST_BUF_COLOR)) {
 198       i915->validation_buffers[i915->num_validation_buffers++]
 199          = i915->current.cbuf_bo;
 200       *batch_space += 3;
 201    }
 202
 203    if (i915->current.depth_bo && (i915->static_dirty & I915_DST_BUF_DEPTH)) {
 204       i915->validation_buffers[i915->num_validation_buffers++]
 205          = i915->current.depth_bo;
 206       *batch_space += 3;
 207    }
 208
 209    if (i915->static_dirty & I915_DST_VARS)
 210       *batch_space += 2;
 211
 212    if (i915->static_dirty & I915_DST_RECT)
 213       *batch_space += 5;
 214 }
 215
 216 static void
 217 emit_static(struct i915_context *i915)
 218 {
 219    if (i915->current.cbuf_bo && (i915->static_dirty & I915_DST_BUF_COLOR)) {
 220       OUT_BATCH(_3DSTATE_BUF_INFO_CMD);
 221       OUT_BATCH(i915->current.cbuf_flags);
 222       OUT_RELOC(i915->current.cbuf_bo,
 223                 I915_USAGE_RENDER,
 224                 0);
 225    }
 226
 227    /* What happens if no zbuf??
 228     */
 229    if (i915->current.depth_bo && (i915->static_dirty & I915_DST_BUF_DEPTH)) {
 230       OUT_BATCH(_3DSTATE_BUF_INFO_CMD);
 231       OUT_BATCH(i915->current.depth_flags);
 232       OUT_RELOC(i915->current.depth_bo,
 233                 I915_USAGE_RENDER,
 234                 0);
 235    }
 236
 237    if (i915->static_dirty & I915_DST_VARS) {
 238       OUT_BATCH(_3DSTATE_DST_BUF_VARS_CMD);
 239       OUT_BATCH(i915->current.dst_buf_vars);
 240    }
 241 }
 242
 243 static void
 244 validate_map(struct i915_context *i915, unsigned *batch_space)
 245 {
 246    const uint enabled = i915->current.sampler_enable_flags;
 247    uint unit;
 248    struct i915_texture *tex;
 249
 250    *batch_space = i915->current.sampler_enable_nr ?
 251      2 + 3*i915->current.sampler_enable_nr : 0;
 252
 253    for (unit = 0; unit < I915_TEX_UNITS; unit++) {
 254       if (enabled & (1 << unit)) {
 255          tex = i915_texture(i915->fragment_sampler_views[unit]->texture);
 256          i915->validation_buffers[i915->num_validation_buffers++] = tex->buffer;
 257       }
 258    }
 259 }
 260
 261 static void
 262 emit_map(struct i915_context *i915)
 263 {
 264    const uint nr = i915->current.sampler_enable_nr;
 265    if (nr) {
 266       const uint enabled = i915->current.sampler_enable_flags;
 267       uint unit;
 268       uint count = 0;
 269       OUT_BATCH(_3DSTATE_MAP_STATE | (3 * nr));
 270       OUT_BATCH(enabled);
 271       for (unit = 0; unit < I915_TEX_UNITS; unit++) {
 272          if (enabled & (1 << unit)) {
 273             struct i915_texture *texture = i915_texture(i915->fragment_sampler_views[unit]->texture);
 274             struct i915_winsys_buffer *buf = texture->buffer;
 275             assert(buf);
 276
 277             count++;
 278
 279             OUT_RELOC(buf, I915_USAGE_SAMPLER, 0);
 280             OUT_BATCH(i915->current.texbuffer[unit][0]); /* MS3 */
 281             OUT_BATCH(i915->current.texbuffer[unit][1]); /* MS4 */
 282          }
 283       }
 284       assert(count == nr);
 285    }
 286 }
 287
 288 static void
 289 validate_sampler(struct i915_context *i915, unsigned *batch_space)
 290 {
 291    *batch_space = i915->current.sampler_enable_nr ?
 292      2 + 3*i915->current.sampler_enable_nr : 0;
 293 }
 294
 295 static void
 296 emit_sampler(struct i915_context *i915)
 297 {
 298    if (i915->current.sampler_enable_nr) {
 299       int i;
 300
 301       OUT_BATCH( _3DSTATE_SAMPLER_STATE |
 302                  (3 * i915->current.sampler_enable_nr) );
 303
 304       OUT_BATCH( i915->current.sampler_enable_flags );
 305
 306       for (i = 0; i < I915_TEX_UNITS; i++) {
 307          if (i915->current.sampler_enable_flags & (1<<i)) {
 308             OUT_BATCH( i915->current.sampler[i][0] );
 309             OUT_BATCH( i915->current.sampler[i][1] );
 310             OUT_BATCH( i915->current.sampler[i][2] );
 311          }
 312       }
 313    }
 314 }
 315
 316 static void
 317 validate_constants(struct i915_context *i915, unsigned *batch_space)
 318 {
 319    *batch_space = i915->fs->num_constants ?
 320       2 + 4*i915->fs->num_constants : 0;
 321 }
 322
 323 static void
 324 emit_constants(struct i915_context *i915)
 325 {
 326    /* Collate the user-defined constants with the fragment shader's
 327     * immediates according to the constant_flags[] array.
 328     */
 329    const uint nr = i915->fs->num_constants;
 330    if (nr) {
 331       uint i;
 332
 333       OUT_BATCH( _3DSTATE_PIXEL_SHADER_CONSTANTS | (nr * 4) );
 334       OUT_BATCH((1 << nr) - 1);
 335
 336       for (i = 0; i < nr; i++) {
 337          const uint *c;
 338          if (i915->fs->constant_flags[i] == I915_CONSTFLAG_USER) {
 339             /* grab user-defined constant */
 340             c = (uint *) i915_buffer(i915->constants[PIPE_SHADER_FRAGMENT])->data;
 341             c += 4 * i;
 342          }
 343          else {
 344             /* emit program constant */
 345             c = (uint *) i915->fs->constants[i];
 346          }
 347 #if 0 /* debug */
 348          {
 349             float *f = (float *) c;
 350             printf("Const %2d: %f %f %f %f %s\n", i, f[0], f[1], f[2], f[3],
 351                    (i915->fs->constant_flags[i] == I915_CONSTFLAG_USER
 352                     ? "user" : "immediate"));
 353          }
 354 #endif
 355          OUT_BATCH(*c++);
 356          OUT_BATCH(*c++);
 357          OUT_BATCH(*c++);
 358          OUT_BATCH(*c++);
 359       }
 360    }
 361 }
 362
 363 static void
 364 validate_program(struct i915_context *i915, unsigned *batch_space)
 365 {
 366    uint additional_size = i915->current.target_fixup_format ? 1 : 0;
 367
 368    /* we need more batch space if we want to emulate rgba framebuffers */
 369    *batch_space = i915->fs->program_len + 3 * additional_size;
 370 }
 371
 372 static void
 373 emit_program(struct i915_context *i915)
 374 {
 375    uint need_target_fixup = i915->current.target_fixup_format ? 1 : 0;
 376    uint i;
 377
 378    /* we should always have, at least, a pass-through program */
 379    assert(i915->fs->program_len > 0);
 380
 381    {
 382       /* first word has the size, we have to adjust that */
 383       uint size = (i915->fs->program[0]);
 384       size += need_target_fixup * 3;
 385       OUT_BATCH(size);
 386    }
 387
 388    /* output the declarations of the program */
 389    for (i=1 ; i < i915->fs->program_len; i++)
 390       OUT_BATCH(i915->fs->program[i]);
 391
 392    /* we emit an additional mov with swizzle to fake RGBA framebuffers */
 393    if (need_target_fixup) {
 394       /* mov out_color, out_color.zyxw */
 395       OUT_BATCH(A0_MOV |
 396                 (REG_TYPE_OC << A0_DEST_TYPE_SHIFT) |
 397                 A0_DEST_CHANNEL_ALL |
 398                 (REG_TYPE_OC << A0_SRC0_TYPE_SHIFT) |
 399                 (T_DIFFUSE << A0_SRC0_NR_SHIFT));
 400       OUT_BATCH(i915->current.fixup_swizzle);
 401       OUT_BATCH(0);
 402    }
 403 }
 404
 405 static void
 406 emit_draw_rect(struct i915_context *i915)
 407 {
 408    if (i915->static_dirty & I915_DST_RECT) {
 409       OUT_BATCH(_3DSTATE_DRAW_RECT_CMD);
 410       OUT_BATCH(DRAW_RECT_DIS_DEPTH_OFS);
 411       OUT_BATCH(i915->current.draw_offset);
 412       OUT_BATCH(i915->current.draw_size);
 413       OUT_BATCH(i915->current.draw_offset);
 414    }
 415 }
 416
 417 static boolean
 418 i915_validate_state(struct i915_context *i915, unsigned *batch_space)
 419 {
 420    unsigned tmp;
 421
 422    i915->num_validation_buffers = 0;
 423    if (i915->hardware_dirty & I915_HW_INVARIANT)
 424       *batch_space = Elements(invariant_state);
 425    else
 426       *batch_space = 0;
 427
 428 #define VALIDATE_ATOM(atom, hw_dirty) \
 429    if (i915->hardware_dirty & hw_dirty) { \
 430       validate_##atom(i915, &tmp); \
 431       *batch_space += tmp; }
 432    VALIDATE_ATOM(flush, I915_HW_FLUSH);
 433    VALIDATE_ATOM(immediate, I915_HW_IMMEDIATE);
 434    VALIDATE_ATOM(dynamic, I915_HW_DYNAMIC);
 435    VALIDATE_ATOM(static, I915_HW_STATIC);
 436    VALIDATE_ATOM(map, I915_HW_MAP);
 437    VALIDATE_ATOM(sampler, I915_HW_SAMPLER);
 438    VALIDATE_ATOM(constants, I915_HW_CONSTANTS);
 439    VALIDATE_ATOM(program, I915_HW_PROGRAM);
 440 #undef VALIDATE_ATOM
 441
 442    if (i915->num_validation_buffers == 0)
 443       return TRUE;
 444
 445    if (!i915_winsys_validate_buffers(i915->batch, i915->validation_buffers,
 446                                      i915->num_validation_buffers))
 447       return FALSE;
 448
 449    return TRUE;
 450 }
 451
 452 /* Push the state into the sarea and/or texture memory.
 453  */
 454 void
 455 i915_emit_hardware_state(struct i915_context *i915 )
 456 {
 457    unsigned batch_space;
 458    uintptr_t save_ptr;
 459
 460    assert(i915->dirty == 0);
 461
 462    if (I915_DBG_ON(DBG_ATOMS))
 463       i915_dump_hardware_dirty(i915, __FUNCTION__);
 464
 465    if (!i915_validate_state(i915, &batch_space)) {
 466       FLUSH_BATCH(NULL);
 467       assert(i915_validate_state(i915, &batch_space));
 468    }
 469
 470    if(!BEGIN_BATCH(batch_space)) {
 471       FLUSH_BATCH(NULL);
 472       assert(i915_validate_state(i915, &batch_space));
 473       assert(BEGIN_BATCH(batch_space));
 474    }
 475
 476    save_ptr = (uintptr_t)i915->batch->ptr;
 477
 478 #define EMIT_ATOM(atom, hw_dirty) \
 479    if (i915->hardware_dirty & hw_dirty) \
 480       emit_##atom(i915);
 481    EMIT_ATOM(flush, I915_HW_FLUSH);
 482    EMIT_ATOM(invariant, I915_HW_INVARIANT);
 483    EMIT_ATOM(immediate, I915_HW_IMMEDIATE);
 484    EMIT_ATOM(dynamic, I915_HW_DYNAMIC);
 485    EMIT_ATOM(static, I915_HW_STATIC);
 486    EMIT_ATOM(map, I915_HW_MAP);
 487    EMIT_ATOM(sampler, I915_HW_SAMPLER);
 488    EMIT_ATOM(constants, I915_HW_CONSTANTS);
 489    EMIT_ATOM(program, I915_HW_PROGRAM);
 490    EMIT_ATOM(draw_rect, I915_HW_STATIC);
 491 #undef EMIT_ATOM
 492
 493    I915_DBG(DBG_EMIT, "%s: used %d dwords, %d dwords reserved\n", __FUNCTION__,
 494             ((uintptr_t)i915->batch->ptr - save_ptr) / 4,
 495             batch_space);
 496    assert(((uintptr_t)i915->batch->ptr - save_ptr) / 4 == batch_space);
 497
 498    i915->hardware_dirty = 0;
 499    i915->immediate_dirty = 0;
 500    i915->dynamic_dirty = 0;
 501    i915->static_dirty = 0;
 502    i915->flush_dirty = 0;
 503 }