src/gallium/drivers/v3d/v3dx_state.c

   1 /*
   2  * Copyright © 2014-2017 Broadcom
   3  * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
   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  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * 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 NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  22  * IN THE SOFTWARE.
  23  */
  24
  25 #include "pipe/p_state.h"
  26 #include "util/u_format.h"
  27 #include "util/u_framebuffer.h"
  28 #include "util/u_inlines.h"
  29 #include "util/u_math.h"
  30 #include "util/u_memory.h"
  31 #include "util/u_half.h"
  32 #include "util/u_helpers.h"
  33
  34 #include "v3d_context.h"
  35 #include "v3d_tiling.h"
  36 #include "broadcom/common/v3d_macros.h"
  37 #include "broadcom/cle/v3dx_pack.h"
  38
  39 static void
  40 v3d_generic_cso_state_delete(struct pipe_context *pctx, void *hwcso)
  41 {
  42         free(hwcso);
  43 }
  44
  45 static void
  46 v3d_set_blend_color(struct pipe_context *pctx,
  47                     const struct pipe_blend_color *blend_color)
  48 {
  49         struct v3d_context *v3d = v3d_context(pctx);
  50         v3d->blend_color.f = *blend_color;
  51         for (int i = 0; i < 4; i++) {
  52                 v3d->blend_color.hf[i] =
  53                         util_float_to_half(blend_color->color[i]);
  54         }
  55         v3d->dirty |= VC5_DIRTY_BLEND_COLOR;
  56 }
  57
  58 static void
  59 v3d_set_stencil_ref(struct pipe_context *pctx,
  60                     const struct pipe_stencil_ref *stencil_ref)
  61 {
  62         struct v3d_context *v3d = v3d_context(pctx);
  63         v3d->stencil_ref = *stencil_ref;
  64         v3d->dirty |= VC5_DIRTY_STENCIL_REF;
  65 }
  66
  67 static void
  68 v3d_set_clip_state(struct pipe_context *pctx,
  69                    const struct pipe_clip_state *clip)
  70 {
  71         struct v3d_context *v3d = v3d_context(pctx);
  72         v3d->clip = *clip;
  73         v3d->dirty |= VC5_DIRTY_CLIP;
  74 }
  75
  76 static void
  77 v3d_set_sample_mask(struct pipe_context *pctx, unsigned sample_mask)
  78 {
  79         struct v3d_context *v3d = v3d_context(pctx);
  80         v3d->sample_mask = sample_mask & ((1 << VC5_MAX_SAMPLES) - 1);
  81         v3d->dirty |= VC5_DIRTY_SAMPLE_STATE;
  82 }
  83
  84 static uint16_t
  85 float_to_187_half(float f)
  86 {
  87         return fui(f) >> 16;
  88 }
  89
  90 static void *
  91 v3d_create_rasterizer_state(struct pipe_context *pctx,
  92                             const struct pipe_rasterizer_state *cso)
  93 {
  94         struct v3d_rasterizer_state *so;
  95
  96         so = CALLOC_STRUCT(v3d_rasterizer_state);
  97         if (!so)
  98                 return NULL;
  99
 100         so->base = *cso;
 101
 102         /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
 103          * BCM21553).
 104          */
 105         so->point_size = MAX2(cso->point_size, .125f);
 106
 107         if (cso->offset_tri) {
 108                 so->offset_units = float_to_187_half(cso->offset_units);
 109                 so->z16_offset_units = float_to_187_half(cso->offset_units * 256.0);
 110                 so->offset_factor = float_to_187_half(cso->offset_scale);
 111         }
 112
 113         return so;
 114 }
 115
 116 /* Blend state is baked into shaders. */
 117 static void *
 118 v3d_create_blend_state(struct pipe_context *pctx,
 119                        const struct pipe_blend_state *cso)
 120 {
 121         struct v3d_blend_state *so;
 122
 123         so = CALLOC_STRUCT(v3d_blend_state);
 124         if (!so)
 125                 return NULL;
 126
 127         so->base = *cso;
 128
 129         if (cso->independent_blend_enable) {
 130                 for (int i = 0; i < VC5_MAX_DRAW_BUFFERS; i++) {
 131                         so->blend_enables |= cso->rt[i].blend_enable << i;
 132
 133                         /* V3D 4.x is when we got independent blend enables. */
 134                         assert(V3D_VERSION >= 40 ||
 135                                cso->rt[i].blend_enable == cso->rt[0].blend_enable);
 136                 }
 137         } else {
 138                 if (cso->rt[0].blend_enable)
 139                         so->blend_enables = (1 << VC5_MAX_DRAW_BUFFERS) - 1;
 140         }
 141
 142         return so;
 143 }
 144
 145 static uint32_t
 146 translate_stencil_op(enum pipe_stencil_op op)
 147 {
 148         switch (op) {
 149         case PIPE_STENCIL_OP_KEEP:      return V3D_STENCIL_OP_KEEP;
 150         case PIPE_STENCIL_OP_ZERO:      return V3D_STENCIL_OP_ZERO;
 151         case PIPE_STENCIL_OP_REPLACE:   return V3D_STENCIL_OP_REPLACE;
 152         case PIPE_STENCIL_OP_INCR:      return V3D_STENCIL_OP_INCR;
 153         case PIPE_STENCIL_OP_DECR:      return V3D_STENCIL_OP_DECR;
 154         case PIPE_STENCIL_OP_INCR_WRAP: return V3D_STENCIL_OP_INCWRAP;
 155         case PIPE_STENCIL_OP_DECR_WRAP: return V3D_STENCIL_OP_DECWRAP;
 156         case PIPE_STENCIL_OP_INVERT:    return V3D_STENCIL_OP_INVERT;
 157         }
 158         unreachable("bad stencil op");
 159 }
 160
 161 static void *
 162 v3d_create_depth_stencil_alpha_state(struct pipe_context *pctx,
 163                                      const struct pipe_depth_stencil_alpha_state *cso)
 164 {
 165         struct v3d_depth_stencil_alpha_state *so;
 166
 167         so = CALLOC_STRUCT(v3d_depth_stencil_alpha_state);
 168         if (!so)
 169                 return NULL;
 170
 171         so->base = *cso;
 172
 173         if (cso->depth.enabled) {
 174                 switch (cso->depth.func) {
 175                 case PIPE_FUNC_LESS:
 176                 case PIPE_FUNC_LEQUAL:
 177                         so->ez_state = VC5_EZ_LT_LE;
 178                         break;
 179                 case PIPE_FUNC_GREATER:
 180                 case PIPE_FUNC_GEQUAL:
 181                         so->ez_state = VC5_EZ_GT_GE;
 182                         break;
 183                 case PIPE_FUNC_NEVER:
 184                 case PIPE_FUNC_EQUAL:
 185                         so->ez_state = VC5_EZ_UNDECIDED;
 186                         break;
 187                 default:
 188                         so->ez_state = VC5_EZ_DISABLED;
 189                         break;
 190                 }
 191
 192                 /* If stencil is enabled and it's not a no-op, then it would
 193                  * break EZ updates.
 194                  */
 195                 if (cso->stencil[0].enabled &&
 196                     (cso->stencil[0].zfail_op != PIPE_STENCIL_OP_KEEP ||
 197                      cso->stencil[0].func != PIPE_FUNC_ALWAYS ||
 198                      (cso->stencil[1].enabled &&
 199                       (cso->stencil[1].zfail_op != PIPE_STENCIL_OP_KEEP &&
 200                        cso->stencil[1].func != PIPE_FUNC_ALWAYS)))) {
 201                         so->ez_state = VC5_EZ_DISABLED;
 202                 }
 203         }
 204
 205         const struct pipe_stencil_state *front = &cso->stencil[0];
 206         const struct pipe_stencil_state *back = &cso->stencil[1];
 207
 208         if (front->enabled) {
 209                 STATIC_ASSERT(sizeof(so->stencil_front) >=
 210                               cl_packet_length(STENCIL_CONFIG));
 211                 v3dx_pack(&so->stencil_front, STENCIL_CONFIG, config) {
 212                         config.front_config = true;
 213                         /* If !back->enabled, then the front values should be
 214                          * used for both front and back-facing primitives.
 215                          */
 216                         config.back_config = !back->enabled;
 217
 218                         config.stencil_write_mask = front->writemask;
 219                         config.stencil_test_mask = front->valuemask;
 220
 221                         config.stencil_test_function = front->func;
 222                         config.stencil_pass_op =
 223                                 translate_stencil_op(front->zpass_op);
 224                         config.depth_test_fail_op =
 225                                 translate_stencil_op(front->zfail_op);
 226                         config.stencil_test_fail_op =
 227                                 translate_stencil_op(front->fail_op);
 228                 }
 229         }
 230         if (back->enabled) {
 231                 STATIC_ASSERT(sizeof(so->stencil_back) >=
 232                               cl_packet_length(STENCIL_CONFIG));
 233                 v3dx_pack(&so->stencil_back, STENCIL_CONFIG, config) {
 234                         config.front_config = false;
 235                         config.back_config = true;
 236
 237                         config.stencil_write_mask = back->writemask;
 238                         config.stencil_test_mask = back->valuemask;
 239
 240                         config.stencil_test_function = back->func;
 241                         config.stencil_pass_op =
 242                                 translate_stencil_op(back->zpass_op);
 243                         config.depth_test_fail_op =
 244                                 translate_stencil_op(back->zfail_op);
 245                         config.stencil_test_fail_op =
 246                                 translate_stencil_op(back->fail_op);
 247                 }
 248         }
 249
 250         return so;
 251 }
 252
 253 static void
 254 v3d_set_polygon_stipple(struct pipe_context *pctx,
 255                         const struct pipe_poly_stipple *stipple)
 256 {
 257         struct v3d_context *v3d = v3d_context(pctx);
 258         v3d->stipple = *stipple;
 259         v3d->dirty |= VC5_DIRTY_STIPPLE;
 260 }
 261
 262 static void
 263 v3d_set_scissor_states(struct pipe_context *pctx,
 264                        unsigned start_slot,
 265                        unsigned num_scissors,
 266                        const struct pipe_scissor_state *scissor)
 267 {
 268         struct v3d_context *v3d = v3d_context(pctx);
 269
 270         v3d->scissor = *scissor;
 271         v3d->dirty |= VC5_DIRTY_SCISSOR;
 272 }
 273
 274 static void
 275 v3d_set_viewport_states(struct pipe_context *pctx,
 276                         unsigned start_slot,
 277                         unsigned num_viewports,
 278                         const struct pipe_viewport_state *viewport)
 279 {
 280         struct v3d_context *v3d = v3d_context(pctx);
 281         v3d->viewport = *viewport;
 282         v3d->dirty |= VC5_DIRTY_VIEWPORT;
 283 }
 284
 285 static void
 286 v3d_set_vertex_buffers(struct pipe_context *pctx,
 287                        unsigned start_slot, unsigned count,
 288                        const struct pipe_vertex_buffer *vb)
 289 {
 290         struct v3d_context *v3d = v3d_context(pctx);
 291         struct v3d_vertexbuf_stateobj *so = &v3d->vertexbuf;
 292
 293         util_set_vertex_buffers_mask(so->vb, &so->enabled_mask, vb,
 294                                      start_slot, count);
 295         so->count = util_last_bit(so->enabled_mask);
 296
 297         v3d->dirty |= VC5_DIRTY_VTXBUF;
 298 }
 299
 300 static void
 301 v3d_blend_state_bind(struct pipe_context *pctx, void *hwcso)
 302 {
 303         struct v3d_context *v3d = v3d_context(pctx);
 304         v3d->blend = hwcso;
 305         v3d->dirty |= VC5_DIRTY_BLEND;
 306 }
 307
 308 static void
 309 v3d_rasterizer_state_bind(struct pipe_context *pctx, void *hwcso)
 310 {
 311         struct v3d_context *v3d = v3d_context(pctx);
 312         v3d->rasterizer = hwcso;
 313         v3d->dirty |= VC5_DIRTY_RASTERIZER;
 314 }
 315
 316 static void
 317 v3d_zsa_state_bind(struct pipe_context *pctx, void *hwcso)
 318 {
 319         struct v3d_context *v3d = v3d_context(pctx);
 320         v3d->zsa = hwcso;
 321         v3d->dirty |= VC5_DIRTY_ZSA;
 322 }
 323
 324 static void *
 325 v3d_vertex_state_create(struct pipe_context *pctx, unsigned num_elements,
 326                         const struct pipe_vertex_element *elements)
 327 {
 328         struct v3d_context *v3d = v3d_context(pctx);
 329         struct v3d_vertex_stateobj *so = CALLOC_STRUCT(v3d_vertex_stateobj);
 330
 331         if (!so)
 332                 return NULL;
 333
 334         memcpy(so->pipe, elements, sizeof(*elements) * num_elements);
 335         so->num_elements = num_elements;
 336
 337         for (int i = 0; i < so->num_elements; i++) {
 338                 const struct pipe_vertex_element *elem = &elements[i];
 339                 const struct util_format_description *desc =
 340                         util_format_description(elem->src_format);
 341                 uint32_t r_size = desc->channel[0].size;
 342
 343                 const uint32_t size =
 344                         cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD);
 345
 346                 v3dx_pack(&so->attrs[i * size],
 347                           GL_SHADER_STATE_ATTRIBUTE_RECORD, attr) {
 348                         /* vec_size == 0 means 4 */
 349                         attr.vec_size = desc->nr_channels & 3;
 350                         attr.signed_int_type = (desc->channel[0].type ==
 351                                                 UTIL_FORMAT_TYPE_SIGNED);
 352
 353                         attr.normalized_int_type = desc->channel[0].normalized;
 354                         attr.read_as_int_uint = desc->channel[0].pure_integer;
 355                         attr.instance_divisor = MIN2(elem->instance_divisor,
 356                                                      0xffff);
 357
 358                         switch (desc->channel[0].type) {
 359                         case UTIL_FORMAT_TYPE_FLOAT:
 360                                 if (r_size == 32) {
 361                                         attr.type = ATTRIBUTE_FLOAT;
 362                                 } else {
 363                                         assert(r_size == 16);
 364                                         attr.type = ATTRIBUTE_HALF_FLOAT;
 365                                 }
 366                                 break;
 367
 368                         case UTIL_FORMAT_TYPE_SIGNED:
 369                         case UTIL_FORMAT_TYPE_UNSIGNED:
 370                                 switch (r_size) {
 371                                 case 32:
 372                                         attr.type = ATTRIBUTE_INT;
 373                                         break;
 374                                 case 16:
 375                                         attr.type = ATTRIBUTE_SHORT;
 376                                         break;
 377                                 case 10:
 378                                         attr.type = ATTRIBUTE_INT2_10_10_10;
 379                                         break;
 380                                 case 8:
 381                                         attr.type = ATTRIBUTE_BYTE;
 382                                         break;
 383                                 default:
 384                                         fprintf(stderr,
 385                                                 "format %s unsupported\n",
 386                                                 desc->name);
 387                                         attr.type = ATTRIBUTE_BYTE;
 388                                         abort();
 389                                 }
 390                                 break;
 391
 392                         default:
 393                                 fprintf(stderr,
 394                                         "format %s unsupported\n",
 395                                         desc->name);
 396                                 abort();
 397                         }
 398                 }
 399         }
 400
 401         /* Set up the default attribute values in case any of the vertex
 402          * elements use them.
 403          */
 404         so->default_attribute_values = v3d_bo_alloc(v3d->screen,
 405                                                     VC5_MAX_ATTRIBUTES *
 406                                                     4 * sizeof(float),
 407                                                     "default_attributes");
 408         uint32_t *attrs = v3d_bo_map(so->default_attribute_values);
 409         for (int i = 0; i < VC5_MAX_ATTRIBUTES; i++) {
 410                 attrs[i * 4 + 0] = 0;
 411                 attrs[i * 4 + 1] = 0;
 412                 attrs[i * 4 + 2] = 0;
 413                 if (i < so->num_elements &&
 414                     util_format_is_pure_integer(so->pipe[i].src_format)) {
 415                         attrs[i * 4 + 3] = 1;
 416                 } else {
 417                         attrs[i * 4 + 3] = fui(1.0);
 418                 }
 419         }
 420
 421         return so;
 422 }
 423
 424 static void
 425 v3d_vertex_state_delete(struct pipe_context *pctx, void *hwcso)
 426 {
 427         struct v3d_vertex_stateobj *so = hwcso;
 428
 429         v3d_bo_unreference(&so->default_attribute_values);
 430         free(so);
 431 }
 432
 433 static void
 434 v3d_vertex_state_bind(struct pipe_context *pctx, void *hwcso)
 435 {
 436         struct v3d_context *v3d = v3d_context(pctx);
 437         v3d->vtx = hwcso;
 438         v3d->dirty |= VC5_DIRTY_VTXSTATE;
 439 }
 440
 441 static void
 442 v3d_set_constant_buffer(struct pipe_context *pctx, uint shader, uint index,
 443                         const struct pipe_constant_buffer *cb)
 444 {
 445         struct v3d_context *v3d = v3d_context(pctx);
 446         struct v3d_constbuf_stateobj *so = &v3d->constbuf[shader];
 447
 448         util_copy_constant_buffer(&so->cb[index], cb);
 449
 450         /* Note that the state tracker can unbind constant buffers by
 451          * passing NULL here.
 452          */
 453         if (unlikely(!cb)) {
 454                 so->enabled_mask &= ~(1 << index);
 455                 so->dirty_mask &= ~(1 << index);
 456                 return;
 457         }
 458
 459         so->enabled_mask |= 1 << index;
 460         so->dirty_mask |= 1 << index;
 461         v3d->dirty |= VC5_DIRTY_CONSTBUF;
 462 }
 463
 464 static void
 465 v3d_set_framebuffer_state(struct pipe_context *pctx,
 466                           const struct pipe_framebuffer_state *framebuffer)
 467 {
 468         struct v3d_context *v3d = v3d_context(pctx);
 469         struct pipe_framebuffer_state *cso = &v3d->framebuffer;
 470
 471         v3d->job = NULL;
 472
 473         util_copy_framebuffer_state(cso, framebuffer);
 474
 475         v3d->swap_color_rb = 0;
 476         v3d->blend_dst_alpha_one = 0;
 477         for (int i = 0; i < v3d->framebuffer.nr_cbufs; i++) {
 478                 struct pipe_surface *cbuf = v3d->framebuffer.cbufs[i];
 479                 if (!cbuf)
 480                         continue;
 481
 482                 const struct util_format_description *desc =
 483                         util_format_description(cbuf->format);
 484
 485                 /* For BGRA8 formats (DRI window system default format), we
 486                  * need to swap R and B, since the HW's format is RGBA8.
 487                  */
 488                 if (desc->swizzle[0] == PIPE_SWIZZLE_Z &&
 489                     cbuf->format != PIPE_FORMAT_B5G6R5_UNORM) {
 490                         v3d->swap_color_rb |= 1 << i;
 491                 }
 492
 493                 if (desc->swizzle[3] == PIPE_SWIZZLE_1)
 494                         v3d->blend_dst_alpha_one |= 1 << i;
 495         }
 496
 497         v3d->dirty |= VC5_DIRTY_FRAMEBUFFER;
 498 }
 499
 500 static struct v3d_texture_stateobj *
 501 v3d_get_stage_tex(struct v3d_context *v3d, enum pipe_shader_type shader)
 502 {
 503         switch (shader) {
 504         case PIPE_SHADER_FRAGMENT:
 505                 v3d->dirty |= VC5_DIRTY_FRAGTEX;
 506                 return &v3d->fragtex;
 507                 break;
 508         case PIPE_SHADER_VERTEX:
 509                 v3d->dirty |= VC5_DIRTY_VERTTEX;
 510                 return &v3d->verttex;
 511                 break;
 512         default:
 513                 fprintf(stderr, "Unknown shader target %d\n", shader);
 514                 abort();
 515         }
 516 }
 517
 518 static uint32_t translate_wrap(uint32_t pipe_wrap, bool using_nearest)
 519 {
 520         switch (pipe_wrap) {
 521         case PIPE_TEX_WRAP_REPEAT:
 522                 return 0;
 523         case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 524                 return 1;
 525         case PIPE_TEX_WRAP_MIRROR_REPEAT:
 526                 return 2;
 527         case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
 528                 return 3;
 529         case PIPE_TEX_WRAP_CLAMP:
 530                 return (using_nearest ? 1 : 3);
 531         default:
 532                 unreachable("Unknown wrap mode");
 533         }
 534 }
 535
 536
 537 static void *
 538 v3d_create_sampler_state(struct pipe_context *pctx,
 539                          const struct pipe_sampler_state *cso)
 540 {
 541         MAYBE_UNUSED struct v3d_context *v3d = v3d_context(pctx);
 542         struct v3d_sampler_state *so = CALLOC_STRUCT(v3d_sampler_state);
 543
 544         if (!so)
 545                 return NULL;
 546
 547         memcpy(so, cso, sizeof(*cso));
 548
 549         bool either_nearest =
 550                 (cso->mag_img_filter == PIPE_TEX_MIPFILTER_NEAREST ||
 551                  cso->min_img_filter == PIPE_TEX_MIPFILTER_NEAREST);
 552
 553 #if V3D_VERSION >= 40
 554         so->bo = v3d_bo_alloc(v3d->screen, cl_packet_length(SAMPLER_STATE),
 555                               "sampler");
 556         void *map = v3d_bo_map(so->bo);
 557
 558         v3dx_pack(map, SAMPLER_STATE, sampler) {
 559                 sampler.wrap_i_border = false;
 560
 561                 sampler.wrap_s = translate_wrap(cso->wrap_s, either_nearest);
 562                 sampler.wrap_t = translate_wrap(cso->wrap_t, either_nearest);
 563                 sampler.wrap_r = translate_wrap(cso->wrap_r, either_nearest);
 564
 565                 sampler.fixed_bias = cso->lod_bias;
 566                 sampler.depth_compare_function = cso->compare_func;
 567
 568                 sampler.min_filter_nearest =
 569                         cso->min_img_filter == PIPE_TEX_FILTER_NEAREST;
 570                 sampler.mag_filter_nearest =
 571                         cso->mag_img_filter == PIPE_TEX_FILTER_NEAREST;
 572                 sampler.mip_filter_nearest =
 573                         cso->min_mip_filter != PIPE_TEX_MIPFILTER_LINEAR;
 574
 575                 sampler.min_level_of_detail = MIN2(MAX2(0, cso->min_lod),
 576                                                    15);
 577                 sampler.max_level_of_detail = MIN2(cso->max_lod, 15);
 578
 579                 /* If we're not doing inter-miplevel filtering, we need to
 580                  * clamp the LOD so that we only sample from baselevel.
 581                  * However, we need to still allow the calculated LOD to be
 582                  * fractionally over the baselevel, so that the HW can decide
 583                  * between the min and mag filters.
 584                  */
 585                 if (cso->min_mip_filter == PIPE_TEX_MIPFILTER_NONE) {
 586                         sampler.min_level_of_detail =
 587                                 MIN2(sampler.min_level_of_detail, 1.0 / 256.0);
 588                         sampler.max_level_of_detail =
 589                                 MIN2(sampler.max_level_of_detail, 1.0 / 256.0);
 590                 }
 591
 592                 if (cso->max_anisotropy) {
 593                         sampler.anisotropy_enable = true;
 594
 595                         if (cso->max_anisotropy > 8)
 596                                 sampler.maximum_anisotropy = 3;
 597                         else if (cso->max_anisotropy > 4)
 598                                 sampler.maximum_anisotropy = 2;
 599                         else if (cso->max_anisotropy > 2)
 600                                 sampler.maximum_anisotropy = 1;
 601                 }
 602
 603                 sampler.border_colour_mode = V3D_BORDER_COLOUR_FOLLOWS;
 604                 /* XXX: The border colour field is in the TMU blending format
 605                  * (32, f16, or i16), and we need to customize it based on
 606                  * that.
 607                  *
 608                  * XXX: for compat alpha formats, we need the alpha field to
 609                  * be in the red channel.
 610                  */
 611                 sampler.border_colour_red =
 612                         util_float_to_half(cso->border_color.f[0]);
 613                 sampler.border_colour_green =
 614                         util_float_to_half(cso->border_color.f[1]);
 615                 sampler.border_colour_blue =
 616                         util_float_to_half(cso->border_color.f[2]);
 617                 sampler.border_colour_alpha =
 618                         util_float_to_half(cso->border_color.f[3]);
 619         }
 620
 621 #else /* V3D_VERSION < 40 */
 622         v3dx_pack(&so->p0, TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1, p0) {
 623                 p0.s_wrap_mode = translate_wrap(cso->wrap_s, either_nearest);
 624                 p0.t_wrap_mode = translate_wrap(cso->wrap_t, either_nearest);
 625                 p0.r_wrap_mode = translate_wrap(cso->wrap_r, either_nearest);
 626         }
 627
 628         v3dx_pack(&so->texture_shader_state, TEXTURE_SHADER_STATE, tex) {
 629                 tex.depth_compare_function = cso->compare_func;
 630                 tex.fixed_bias = cso->lod_bias;
 631         }
 632 #endif /* V3D_VERSION < 40 */
 633         return so;
 634 }
 635
 636 static void
 637 v3d_sampler_states_bind(struct pipe_context *pctx,
 638                         enum pipe_shader_type shader, unsigned start,
 639                         unsigned nr, void **hwcso)
 640 {
 641         struct v3d_context *v3d = v3d_context(pctx);
 642         struct v3d_texture_stateobj *stage_tex = v3d_get_stage_tex(v3d, shader);
 643
 644         assert(start == 0);
 645         unsigned i;
 646         unsigned new_nr = 0;
 647
 648         for (i = 0; i < nr; i++) {
 649                 if (hwcso[i])
 650                         new_nr = i + 1;
 651                 stage_tex->samplers[i] = hwcso[i];
 652         }
 653
 654         for (; i < stage_tex->num_samplers; i++) {
 655                 stage_tex->samplers[i] = NULL;
 656         }
 657
 658         stage_tex->num_samplers = new_nr;
 659 }
 660
 661 static void
 662 v3d_sampler_state_delete(struct pipe_context *pctx,
 663                          void *hwcso)
 664 {
 665         struct pipe_sampler_state *psampler = hwcso;
 666         struct v3d_sampler_state *sampler = v3d_sampler_state(psampler);
 667
 668         v3d_bo_unreference(&sampler->bo);
 669         free(psampler);
 670 }
 671
 672 #if V3D_VERSION >= 40
 673 static uint32_t
 674 translate_swizzle(unsigned char pipe_swizzle)
 675 {
 676         switch (pipe_swizzle) {
 677         case PIPE_SWIZZLE_0:
 678                 return 0;
 679         case PIPE_SWIZZLE_1:
 680                 return 1;
 681         case PIPE_SWIZZLE_X:
 682         case PIPE_SWIZZLE_Y:
 683         case PIPE_SWIZZLE_Z:
 684         case PIPE_SWIZZLE_W:
 685                 return 2 + pipe_swizzle;
 686         default:
 687                 unreachable("unknown swizzle");
 688         }
 689 }
 690 #endif
 691
 692 static struct pipe_sampler_view *
 693 v3d_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
 694                         const struct pipe_sampler_view *cso)
 695 {
 696         struct v3d_context *v3d = v3d_context(pctx);
 697         struct v3d_screen *screen = v3d->screen;
 698         struct v3d_sampler_view *so = CALLOC_STRUCT(v3d_sampler_view);
 699         struct v3d_resource *rsc = v3d_resource(prsc);
 700
 701         if (!so)
 702                 return NULL;
 703
 704         so->base = *cso;
 705
 706         pipe_reference(NULL, &prsc->reference);
 707
 708         /* Compute the sampler view's swizzle up front. This will be plugged
 709          * into either the sampler (for 16-bit returns) or the shader's
 710          * texture key (for 32)
 711          */
 712         uint8_t view_swizzle[4] = {
 713                 cso->swizzle_r,
 714                 cso->swizzle_g,
 715                 cso->swizzle_b,
 716                 cso->swizzle_a
 717         };
 718         const uint8_t *fmt_swizzle =
 719                 v3d_get_format_swizzle(&screen->devinfo, so->base.format);
 720         util_format_compose_swizzles(fmt_swizzle, view_swizzle, so->swizzle);
 721
 722         so->base.texture = prsc;
 723         so->base.reference.count = 1;
 724         so->base.context = pctx;
 725
 726         int msaa_scale = prsc->nr_samples > 1 ? 2 : 1;
 727
 728 #if V3D_VERSION >= 40
 729         so->bo = v3d_bo_alloc(v3d->screen,
 730                               cl_packet_length(TEXTURE_SHADER_STATE), "sampler");
 731         void *map = v3d_bo_map(so->bo);
 732
 733         v3dx_pack(map, TEXTURE_SHADER_STATE, tex) {
 734 #else /* V3D_VERSION < 40 */
 735         STATIC_ASSERT(sizeof(so->texture_shader_state) >=
 736                       cl_packet_length(TEXTURE_SHADER_STATE));
 737         v3dx_pack(&so->texture_shader_state, TEXTURE_SHADER_STATE, tex) {
 738 #endif
 739
 740                 tex.image_width = prsc->width0 * msaa_scale;
 741                 tex.image_height = prsc->height0 * msaa_scale;
 742
 743 #if V3D_VERSION >= 40
 744                 /* On 4.x, the height of a 1D texture is redefined to be the
 745                  * upper 14 bits of the width (which is only usable with txf).
 746                  */
 747                 if (prsc->target == PIPE_TEXTURE_1D ||
 748                     prsc->target == PIPE_TEXTURE_1D_ARRAY) {
 749                         tex.image_height = tex.image_width >> 14;
 750                 }
 751 #endif
 752
 753                 if (prsc->target == PIPE_TEXTURE_3D) {
 754                         tex.image_depth = prsc->depth0;
 755                 } else {
 756                         tex.image_depth = (cso->u.tex.last_layer -
 757                                            cso->u.tex.first_layer) + 1;
 758                 }
 759
 760                 tex.srgb = util_format_is_srgb(cso->format);
 761
 762                 tex.base_level = cso->u.tex.first_level;
 763 #if V3D_VERSION >= 40
 764                 tex.max_level = cso->u.tex.last_level;
 765                 /* Note that we don't have a job to reference the texture's sBO
 766                  * at state create time, so any time this sampler view is used
 767                  * we need to add the texture to the job.
 768                  */
 769                 tex.texture_base_pointer = cl_address(NULL,
 770                                                       rsc->bo->offset +
 771                                                       rsc->slices[0].offset +
 772                                                       cso->u.tex.first_layer *
 773                                                       rsc->cube_map_stride),
 774
 775                 tex.swizzle_r = translate_swizzle(so->swizzle[0]);
 776                 tex.swizzle_g = translate_swizzle(so->swizzle[1]);
 777                 tex.swizzle_b = translate_swizzle(so->swizzle[2]);
 778                 tex.swizzle_a = translate_swizzle(so->swizzle[3]);
 779 #endif
 780                 tex.array_stride_64_byte_aligned = rsc->cube_map_stride / 64;
 781
 782                 if (prsc->nr_samples > 1 && V3D_VERSION < 40) {
 783                         /* Using texture views to reinterpret formats on our
 784                          * MSAA textures won't work, because we don't lay out
 785                          * the bits in memory as it's expected -- for example,
 786                          * RGBA8 and RGB10_A2 are compatible in the
 787                          * ARB_texture_view spec, but in HW we lay them out as
 788                          * 32bpp RGBA8 and 64bpp RGBA16F.  Just assert for now
 789                          * to catch failures.
 790                          *
 791                          * We explicitly allow remapping S8Z24 to RGBA8888 for
 792                          * v3d_blit.c's stencil blits.
 793                          */
 794                         assert((util_format_linear(cso->format) ==
 795                                 util_format_linear(prsc->format)) ||
 796                                (prsc->format == PIPE_FORMAT_S8_UINT_Z24_UNORM &&
 797                                 cso->format == PIPE_FORMAT_R8G8B8A8_UNORM));
 798                         uint32_t output_image_format =
 799                                 v3d_get_rt_format(&screen->devinfo, cso->format);
 800                         uint32_t internal_type;
 801                         uint32_t internal_bpp;
 802                         v3d_get_internal_type_bpp_for_output_format(&screen->devinfo,
 803                                                                     output_image_format,
 804                                                                     &internal_type,
 805                                                                     &internal_bpp);
 806
 807                         switch (internal_type) {
 808                         case V3D_INTERNAL_TYPE_8:
 809                                 tex.texture_type = TEXTURE_DATA_FORMAT_RGBA8;
 810                                 break;
 811                         case V3D_INTERNAL_TYPE_16F:
 812                                 tex.texture_type = TEXTURE_DATA_FORMAT_RGBA16F;
 813                                 break;
 814                         default:
 815                                 unreachable("Bad MSAA texture type");
 816                         }
 817
 818                         /* sRGB was stored in the tile buffer as linear and
 819                          * would have been encoded to sRGB on resolved tile
 820                          * buffer store.  Note that this means we would need
 821                          * shader code if we wanted to read an MSAA sRGB
 822                          * texture without sRGB decode.
 823                          */
 824                         tex.srgb = false;
 825                 } else {
 826                         tex.texture_type = v3d_get_tex_format(&screen->devinfo,
 827                                                               cso->format);
 828                 }
 829
 830                 /* Since other platform devices may produce UIF images even
 831                  * when they're not big enough for V3D to assume they're UIF,
 832                  * we force images with level 0 as UIF to be always treated
 833                  * that way.
 834                  */
 835                 tex.level_0_is_strictly_uif = (rsc->slices[0].tiling ==
 836                                                VC5_TILING_UIF_XOR ||
 837                                                rsc->slices[0].tiling ==
 838                                                VC5_TILING_UIF_NO_XOR);
 839                 tex.level_0_xor_enable = (rsc->slices[0].tiling ==
 840                                           VC5_TILING_UIF_XOR);
 841
 842                 if (tex.level_0_is_strictly_uif)
 843                         tex.level_0_ub_pad = rsc->slices[0].ub_pad;
 844
 845 #if V3D_VERSION >= 40
 846                 if (tex.uif_xor_disable ||
 847                     tex.level_0_is_strictly_uif) {
 848                         tex.extended = true;
 849                 }
 850 #endif /* V3D_VERSION >= 40 */
 851         };
 852
 853         return &so->base;
 854 }
 855
 856 static void
 857 v3d_sampler_view_destroy(struct pipe_context *pctx,
 858                          struct pipe_sampler_view *psview)
 859 {
 860         struct v3d_sampler_view *sview = v3d_sampler_view(psview);
 861
 862         v3d_bo_unreference(&sview->bo);
 863         pipe_resource_reference(&psview->texture, NULL);
 864         free(psview);
 865 }
 866
 867 static void
 868 v3d_set_sampler_views(struct pipe_context *pctx,
 869                       enum pipe_shader_type shader,
 870                       unsigned start, unsigned nr,
 871                       struct pipe_sampler_view **views)
 872 {
 873         struct v3d_context *v3d = v3d_context(pctx);
 874         struct v3d_texture_stateobj *stage_tex = v3d_get_stage_tex(v3d, shader);
 875         unsigned i;
 876         unsigned new_nr = 0;
 877
 878         assert(start == 0);
 879
 880         for (i = 0; i < nr; i++) {
 881                 if (views[i])
 882                         new_nr = i + 1;
 883                 pipe_sampler_view_reference(&stage_tex->textures[i], views[i]);
 884         }
 885
 886         for (; i < stage_tex->num_textures; i++) {
 887                 pipe_sampler_view_reference(&stage_tex->textures[i], NULL);
 888         }
 889
 890         stage_tex->num_textures = new_nr;
 891 }
 892
 893 static struct pipe_stream_output_target *
 894 v3d_create_stream_output_target(struct pipe_context *pctx,
 895                                 struct pipe_resource *prsc,
 896                                 unsigned buffer_offset,
 897                                 unsigned buffer_size)
 898 {
 899         struct pipe_stream_output_target *target;
 900
 901         target = CALLOC_STRUCT(pipe_stream_output_target);
 902         if (!target)
 903                 return NULL;
 904
 905         pipe_reference_init(&target->reference, 1);
 906         pipe_resource_reference(&target->buffer, prsc);
 907
 908         target->context = pctx;
 909         target->buffer_offset = buffer_offset;
 910         target->buffer_size = buffer_size;
 911
 912         return target;
 913 }
 914
 915 static void
 916 v3d_stream_output_target_destroy(struct pipe_context *pctx,
 917                                  struct pipe_stream_output_target *target)
 918 {
 919         pipe_resource_reference(&target->buffer, NULL);
 920         free(target);
 921 }
 922
 923 static void
 924 v3d_set_stream_output_targets(struct pipe_context *pctx,
 925                               unsigned num_targets,
 926                               struct pipe_stream_output_target **targets,
 927                               const unsigned *offsets)
 928 {
 929         struct v3d_context *ctx = v3d_context(pctx);
 930         struct v3d_streamout_stateobj *so = &ctx->streamout;
 931         unsigned i;
 932
 933         assert(num_targets <= ARRAY_SIZE(so->targets));
 934
 935         for (i = 0; i < num_targets; i++) {
 936                 if (offsets[i] != -1)
 937                         so->offsets[i] = offsets[i];
 938
 939                 pipe_so_target_reference(&so->targets[i], targets[i]);
 940         }
 941
 942         for (; i < so->num_targets; i++)
 943                 pipe_so_target_reference(&so->targets[i], NULL);
 944
 945         so->num_targets = num_targets;
 946
 947         ctx->dirty |= VC5_DIRTY_STREAMOUT;
 948 }
 949
 950 void
 951 v3dX(state_init)(struct pipe_context *pctx)
 952 {
 953         pctx->set_blend_color = v3d_set_blend_color;
 954         pctx->set_stencil_ref = v3d_set_stencil_ref;
 955         pctx->set_clip_state = v3d_set_clip_state;
 956         pctx->set_sample_mask = v3d_set_sample_mask;
 957         pctx->set_constant_buffer = v3d_set_constant_buffer;
 958         pctx->set_framebuffer_state = v3d_set_framebuffer_state;
 959         pctx->set_polygon_stipple = v3d_set_polygon_stipple;
 960         pctx->set_scissor_states = v3d_set_scissor_states;
 961         pctx->set_viewport_states = v3d_set_viewport_states;
 962
 963         pctx->set_vertex_buffers = v3d_set_vertex_buffers;
 964
 965         pctx->create_blend_state = v3d_create_blend_state;
 966         pctx->bind_blend_state = v3d_blend_state_bind;
 967         pctx->delete_blend_state = v3d_generic_cso_state_delete;
 968
 969         pctx->create_rasterizer_state = v3d_create_rasterizer_state;
 970         pctx->bind_rasterizer_state = v3d_rasterizer_state_bind;
 971         pctx->delete_rasterizer_state = v3d_generic_cso_state_delete;
 972
 973         pctx->create_depth_stencil_alpha_state = v3d_create_depth_stencil_alpha_state;
 974         pctx->bind_depth_stencil_alpha_state = v3d_zsa_state_bind;
 975         pctx->delete_depth_stencil_alpha_state = v3d_generic_cso_state_delete;
 976
 977         pctx->create_vertex_elements_state = v3d_vertex_state_create;
 978         pctx->delete_vertex_elements_state = v3d_vertex_state_delete;
 979         pctx->bind_vertex_elements_state = v3d_vertex_state_bind;
 980
 981         pctx->create_sampler_state = v3d_create_sampler_state;
 982         pctx->delete_sampler_state = v3d_sampler_state_delete;
 983         pctx->bind_sampler_states = v3d_sampler_states_bind;
 984
 985         pctx->create_sampler_view = v3d_create_sampler_view;
 986         pctx->sampler_view_destroy = v3d_sampler_view_destroy;
 987         pctx->set_sampler_views = v3d_set_sampler_views;
 988
 989         pctx->create_stream_output_target = v3d_create_stream_output_target;
 990         pctx->stream_output_target_destroy = v3d_stream_output_target_destroy;
 991         pctx->set_stream_output_targets = v3d_set_stream_output_targets;
 992 }