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 void *
  85 v3d_create_rasterizer_state(struct pipe_context *pctx,
  86                             const struct pipe_rasterizer_state *cso)
  87 {
  88         struct v3d_rasterizer_state *so;
  89
  90         so = CALLOC_STRUCT(v3d_rasterizer_state);
  91         if (!so)
  92                 return NULL;
  93
  94         so->base = *cso;
  95
  96         /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
  97          * BCM21553).
  98          */
  99         so->point_size = MAX2(cso->point_size, .125f);
 100
 101         STATIC_ASSERT(sizeof(so->depth_offset) >=
 102                       cl_packet_length(DEPTH_OFFSET));
 103         v3dx_pack(&so->depth_offset, DEPTH_OFFSET, depth) {
 104                 depth.depth_offset_factor = cso->offset_scale;
 105                 depth.depth_offset_units = cso->offset_units;
 106         }
 107
 108         /* The HW treats polygon offset units based on a Z24 buffer, so we
 109          * need to scale up offset_units if we're only Z16.
 110          */
 111         v3dx_pack(&so->depth_offset_z16, DEPTH_OFFSET, depth) {
 112                 depth.depth_offset_factor = cso->offset_scale;
 113                 depth.depth_offset_units = cso->offset_units * 256.0;
 114         }
 115
 116         return so;
 117 }
 118
 119 /* Blend state is baked into shaders. */
 120 static void *
 121 v3d_create_blend_state(struct pipe_context *pctx,
 122                        const struct pipe_blend_state *cso)
 123 {
 124         struct v3d_blend_state *so;
 125
 126         so = CALLOC_STRUCT(v3d_blend_state);
 127         if (!so)
 128                 return NULL;
 129
 130         so->base = *cso;
 131
 132         if (cso->independent_blend_enable) {
 133                 for (int i = 0; i < VC5_MAX_DRAW_BUFFERS; i++) {
 134                         so->blend_enables |= cso->rt[i].blend_enable << i;
 135
 136                         /* V3D 4.x is when we got independent blend enables. */
 137                         assert(V3D_VERSION >= 40 ||
 138                                cso->rt[i].blend_enable == cso->rt[0].blend_enable);
 139                 }
 140         } else {
 141                 if (cso->rt[0].blend_enable)
 142                         so->blend_enables = (1 << VC5_MAX_DRAW_BUFFERS) - 1;
 143         }
 144
 145         return so;
 146 }
 147
 148 static uint32_t
 149 translate_stencil_op(enum pipe_stencil_op op)
 150 {
 151         switch (op) {
 152         case PIPE_STENCIL_OP_KEEP:      return V3D_STENCIL_OP_KEEP;
 153         case PIPE_STENCIL_OP_ZERO:      return V3D_STENCIL_OP_ZERO;
 154         case PIPE_STENCIL_OP_REPLACE:   return V3D_STENCIL_OP_REPLACE;
 155         case PIPE_STENCIL_OP_INCR:      return V3D_STENCIL_OP_INCR;
 156         case PIPE_STENCIL_OP_DECR:      return V3D_STENCIL_OP_DECR;
 157         case PIPE_STENCIL_OP_INCR_WRAP: return V3D_STENCIL_OP_INCWRAP;
 158         case PIPE_STENCIL_OP_DECR_WRAP: return V3D_STENCIL_OP_DECWRAP;
 159         case PIPE_STENCIL_OP_INVERT:    return V3D_STENCIL_OP_INVERT;
 160         }
 161         unreachable("bad stencil op");
 162 }
 163
 164 static void *
 165 v3d_create_depth_stencil_alpha_state(struct pipe_context *pctx,
 166                                      const struct pipe_depth_stencil_alpha_state *cso)
 167 {
 168         struct v3d_depth_stencil_alpha_state *so;
 169
 170         so = CALLOC_STRUCT(v3d_depth_stencil_alpha_state);
 171         if (!so)
 172                 return NULL;
 173
 174         so->base = *cso;
 175
 176         if (cso->depth.enabled) {
 177                 switch (cso->depth.func) {
 178                 case PIPE_FUNC_LESS:
 179                 case PIPE_FUNC_LEQUAL:
 180                         so->ez_state = VC5_EZ_LT_LE;
 181                         break;
 182                 case PIPE_FUNC_GREATER:
 183                 case PIPE_FUNC_GEQUAL:
 184                         so->ez_state = VC5_EZ_GT_GE;
 185                         break;
 186                 case PIPE_FUNC_NEVER:
 187                 case PIPE_FUNC_EQUAL:
 188                         so->ez_state = VC5_EZ_UNDECIDED;
 189                         break;
 190                 default:
 191                         so->ez_state = VC5_EZ_DISABLED;
 192                         break;
 193                 }
 194
 195                 /* If stencil is enabled and it's not a no-op, then it would
 196                  * break EZ updates.
 197                  */
 198                 if (cso->stencil[0].enabled &&
 199                     (cso->stencil[0].zfail_op != PIPE_STENCIL_OP_KEEP ||
 200                      cso->stencil[0].func != PIPE_FUNC_ALWAYS ||
 201                      (cso->stencil[1].enabled &&
 202                       (cso->stencil[1].zfail_op != PIPE_STENCIL_OP_KEEP &&
 203                        cso->stencil[1].func != PIPE_FUNC_ALWAYS)))) {
 204                         so->ez_state = VC5_EZ_DISABLED;
 205                 }
 206         }
 207
 208         const struct pipe_stencil_state *front = &cso->stencil[0];
 209         const struct pipe_stencil_state *back = &cso->stencil[1];
 210
 211         if (front->enabled) {
 212                 STATIC_ASSERT(sizeof(so->stencil_front) >=
 213                               cl_packet_length(STENCIL_CFG));
 214                 v3dx_pack(&so->stencil_front, STENCIL_CFG, config) {
 215                         config.front_config = true;
 216                         /* If !back->enabled, then the front values should be
 217                          * used for both front and back-facing primitives.
 218                          */
 219                         config.back_config = !back->enabled;
 220
 221                         config.stencil_write_mask = front->writemask;
 222                         config.stencil_test_mask = front->valuemask;
 223
 224                         config.stencil_test_function = front->func;
 225                         config.stencil_pass_op =
 226                                 translate_stencil_op(front->zpass_op);
 227                         config.depth_test_fail_op =
 228                                 translate_stencil_op(front->zfail_op);
 229                         config.stencil_test_fail_op =
 230                                 translate_stencil_op(front->fail_op);
 231                 }
 232         }
 233         if (back->enabled) {
 234                 STATIC_ASSERT(sizeof(so->stencil_back) >=
 235                               cl_packet_length(STENCIL_CFG));
 236                 v3dx_pack(&so->stencil_back, STENCIL_CFG, config) {
 237                         config.front_config = false;
 238                         config.back_config = true;
 239
 240                         config.stencil_write_mask = back->writemask;
 241                         config.stencil_test_mask = back->valuemask;
 242
 243                         config.stencil_test_function = back->func;
 244                         config.stencil_pass_op =
 245                                 translate_stencil_op(back->zpass_op);
 246                         config.depth_test_fail_op =
 247                                 translate_stencil_op(back->zfail_op);
 248                         config.stencil_test_fail_op =
 249                                 translate_stencil_op(back->fail_op);
 250                 }
 251         }
 252
 253         return so;
 254 }
 255
 256 static void
 257 v3d_set_polygon_stipple(struct pipe_context *pctx,
 258                         const struct pipe_poly_stipple *stipple)
 259 {
 260         struct v3d_context *v3d = v3d_context(pctx);
 261         v3d->stipple = *stipple;
 262         v3d->dirty |= VC5_DIRTY_STIPPLE;
 263 }
 264
 265 static void
 266 v3d_set_scissor_states(struct pipe_context *pctx,
 267                        unsigned start_slot,
 268                        unsigned num_scissors,
 269                        const struct pipe_scissor_state *scissor)
 270 {
 271         struct v3d_context *v3d = v3d_context(pctx);
 272
 273         v3d->scissor = *scissor;
 274         v3d->dirty |= VC5_DIRTY_SCISSOR;
 275 }
 276
 277 static void
 278 v3d_set_viewport_states(struct pipe_context *pctx,
 279                         unsigned start_slot,
 280                         unsigned num_viewports,
 281                         const struct pipe_viewport_state *viewport)
 282 {
 283         struct v3d_context *v3d = v3d_context(pctx);
 284         v3d->viewport = *viewport;
 285         v3d->dirty |= VC5_DIRTY_VIEWPORT;
 286 }
 287
 288 static void
 289 v3d_set_vertex_buffers(struct pipe_context *pctx,
 290                        unsigned start_slot, unsigned count,
 291                        const struct pipe_vertex_buffer *vb)
 292 {
 293         struct v3d_context *v3d = v3d_context(pctx);
 294         struct v3d_vertexbuf_stateobj *so = &v3d->vertexbuf;
 295
 296         util_set_vertex_buffers_mask(so->vb, &so->enabled_mask, vb,
 297                                      start_slot, count);
 298         so->count = util_last_bit(so->enabled_mask);
 299
 300         v3d->dirty |= VC5_DIRTY_VTXBUF;
 301 }
 302
 303 static void
 304 v3d_blend_state_bind(struct pipe_context *pctx, void *hwcso)
 305 {
 306         struct v3d_context *v3d = v3d_context(pctx);
 307         v3d->blend = hwcso;
 308         v3d->dirty |= VC5_DIRTY_BLEND;
 309 }
 310
 311 static void
 312 v3d_rasterizer_state_bind(struct pipe_context *pctx, void *hwcso)
 313 {
 314         struct v3d_context *v3d = v3d_context(pctx);
 315         v3d->rasterizer = hwcso;
 316         v3d->dirty |= VC5_DIRTY_RASTERIZER;
 317 }
 318
 319 static void
 320 v3d_zsa_state_bind(struct pipe_context *pctx, void *hwcso)
 321 {
 322         struct v3d_context *v3d = v3d_context(pctx);
 323         v3d->zsa = hwcso;
 324         v3d->dirty |= VC5_DIRTY_ZSA;
 325 }
 326
 327 static void *
 328 v3d_vertex_state_create(struct pipe_context *pctx, unsigned num_elements,
 329                         const struct pipe_vertex_element *elements)
 330 {
 331         struct v3d_context *v3d = v3d_context(pctx);
 332         struct v3d_vertex_stateobj *so = CALLOC_STRUCT(v3d_vertex_stateobj);
 333
 334         if (!so)
 335                 return NULL;
 336
 337         memcpy(so->pipe, elements, sizeof(*elements) * num_elements);
 338         so->num_elements = num_elements;
 339
 340         for (int i = 0; i < so->num_elements; i++) {
 341                 const struct pipe_vertex_element *elem = &elements[i];
 342                 const struct util_format_description *desc =
 343                         util_format_description(elem->src_format);
 344                 uint32_t r_size = desc->channel[0].size;
 345
 346                 const uint32_t size =
 347                         cl_packet_length(GL_SHADER_STATE_ATTRIBUTE_RECORD);
 348
 349                 v3dx_pack(&so->attrs[i * size],
 350                           GL_SHADER_STATE_ATTRIBUTE_RECORD, attr) {
 351                         /* vec_size == 0 means 4 */
 352                         attr.vec_size = desc->nr_channels & 3;
 353                         attr.signed_int_type = (desc->channel[0].type ==
 354                                                 UTIL_FORMAT_TYPE_SIGNED);
 355
 356                         attr.normalized_int_type = desc->channel[0].normalized;
 357                         attr.read_as_int_uint = desc->channel[0].pure_integer;
 358                         attr.instance_divisor = MIN2(elem->instance_divisor,
 359                                                      0xffff);
 360
 361                         switch (desc->channel[0].type) {
 362                         case UTIL_FORMAT_TYPE_FLOAT:
 363                                 if (r_size == 32) {
 364                                         attr.type = ATTRIBUTE_FLOAT;
 365                                 } else {
 366                                         assert(r_size == 16);
 367                                         attr.type = ATTRIBUTE_HALF_FLOAT;
 368                                 }
 369                                 break;
 370
 371                         case UTIL_FORMAT_TYPE_SIGNED:
 372                         case UTIL_FORMAT_TYPE_UNSIGNED:
 373                                 switch (r_size) {
 374                                 case 32:
 375                                         attr.type = ATTRIBUTE_INT;
 376                                         break;
 377                                 case 16:
 378                                         attr.type = ATTRIBUTE_SHORT;
 379                                         break;
 380                                 case 10:
 381                                         attr.type = ATTRIBUTE_INT2_10_10_10;
 382                                         break;
 383                                 case 8:
 384                                         attr.type = ATTRIBUTE_BYTE;
 385                                         break;
 386                                 default:
 387                                         fprintf(stderr,
 388                                                 "format %s unsupported\n",
 389                                                 desc->name);
 390                                         attr.type = ATTRIBUTE_BYTE;
 391                                         abort();
 392                                 }
 393                                 break;
 394
 395                         default:
 396                                 fprintf(stderr,
 397                                         "format %s unsupported\n",
 398                                         desc->name);
 399                                 abort();
 400                         }
 401                 }
 402         }
 403
 404         /* Set up the default attribute values in case any of the vertex
 405          * elements use them.
 406          */
 407         so->default_attribute_values = v3d_bo_alloc(v3d->screen,
 408                                                     VC5_MAX_ATTRIBUTES *
 409                                                     4 * sizeof(float),
 410                                                     "default_attributes");
 411         uint32_t *attrs = v3d_bo_map(so->default_attribute_values);
 412         for (int i = 0; i < VC5_MAX_ATTRIBUTES; i++) {
 413                 attrs[i * 4 + 0] = 0;
 414                 attrs[i * 4 + 1] = 0;
 415                 attrs[i * 4 + 2] = 0;
 416                 if (i < so->num_elements &&
 417                     util_format_is_pure_integer(so->pipe[i].src_format)) {
 418                         attrs[i * 4 + 3] = 1;
 419                 } else {
 420                         attrs[i * 4 + 3] = fui(1.0);
 421                 }
 422         }
 423
 424         return so;
 425 }
 426
 427 static void
 428 v3d_vertex_state_delete(struct pipe_context *pctx, void *hwcso)
 429 {
 430         struct v3d_vertex_stateobj *so = hwcso;
 431
 432         v3d_bo_unreference(&so->default_attribute_values);
 433         free(so);
 434 }
 435
 436 static void
 437 v3d_vertex_state_bind(struct pipe_context *pctx, void *hwcso)
 438 {
 439         struct v3d_context *v3d = v3d_context(pctx);
 440         v3d->vtx = hwcso;
 441         v3d->dirty |= VC5_DIRTY_VTXSTATE;
 442 }
 443
 444 static void
 445 v3d_set_constant_buffer(struct pipe_context *pctx, uint shader, uint index,
 446                         const struct pipe_constant_buffer *cb)
 447 {
 448         struct v3d_context *v3d = v3d_context(pctx);
 449         struct v3d_constbuf_stateobj *so = &v3d->constbuf[shader];
 450
 451         util_copy_constant_buffer(&so->cb[index], cb);
 452
 453         /* Note that the state tracker can unbind constant buffers by
 454          * passing NULL here.
 455          */
 456         if (unlikely(!cb)) {
 457                 so->enabled_mask &= ~(1 << index);
 458                 so->dirty_mask &= ~(1 << index);
 459                 return;
 460         }
 461
 462         so->enabled_mask |= 1 << index;
 463         so->dirty_mask |= 1 << index;
 464         v3d->dirty |= VC5_DIRTY_CONSTBUF;
 465 }
 466
 467 static void
 468 v3d_set_framebuffer_state(struct pipe_context *pctx,
 469                           const struct pipe_framebuffer_state *framebuffer)
 470 {
 471         struct v3d_context *v3d = v3d_context(pctx);
 472         struct pipe_framebuffer_state *cso = &v3d->framebuffer;
 473
 474         v3d->job = NULL;
 475
 476         util_copy_framebuffer_state(cso, framebuffer);
 477
 478         v3d->swap_color_rb = 0;
 479         v3d->blend_dst_alpha_one = 0;
 480         for (int i = 0; i < v3d->framebuffer.nr_cbufs; i++) {
 481                 struct pipe_surface *cbuf = v3d->framebuffer.cbufs[i];
 482                 if (!cbuf)
 483                         continue;
 484
 485                 const struct util_format_description *desc =
 486                         util_format_description(cbuf->format);
 487
 488                 /* For BGRA8 formats (DRI window system default format), we
 489                  * need to swap R and B, since the HW's format is RGBA8.
 490                  */
 491                 if (desc->swizzle[0] == PIPE_SWIZZLE_Z &&
 492                     cbuf->format != PIPE_FORMAT_B5G6R5_UNORM) {
 493                         v3d->swap_color_rb |= 1 << i;
 494                 }
 495
 496                 if (desc->swizzle[3] == PIPE_SWIZZLE_1)
 497                         v3d->blend_dst_alpha_one |= 1 << i;
 498         }
 499
 500         v3d->dirty |= VC5_DIRTY_FRAMEBUFFER;
 501 }
 502
 503 static struct v3d_texture_stateobj *
 504 v3d_get_stage_tex(struct v3d_context *v3d, enum pipe_shader_type shader)
 505 {
 506         switch (shader) {
 507         case PIPE_SHADER_FRAGMENT:
 508                 v3d->dirty |= VC5_DIRTY_FRAGTEX;
 509                 return &v3d->fragtex;
 510                 break;
 511         case PIPE_SHADER_VERTEX:
 512                 v3d->dirty |= VC5_DIRTY_VERTTEX;
 513                 return &v3d->verttex;
 514                 break;
 515         default:
 516                 fprintf(stderr, "Unknown shader target %d\n", shader);
 517                 abort();
 518         }
 519 }
 520
 521 static uint32_t translate_wrap(uint32_t pipe_wrap, bool using_nearest)
 522 {
 523         switch (pipe_wrap) {
 524         case PIPE_TEX_WRAP_REPEAT:
 525                 return 0;
 526         case PIPE_TEX_WRAP_CLAMP_TO_EDGE:
 527                 return 1;
 528         case PIPE_TEX_WRAP_MIRROR_REPEAT:
 529                 return 2;
 530         case PIPE_TEX_WRAP_CLAMP_TO_BORDER:
 531                 return 3;
 532         case PIPE_TEX_WRAP_CLAMP:
 533                 return (using_nearest ? 1 : 3);
 534         default:
 535                 unreachable("Unknown wrap mode");
 536         }
 537 }
 538
 539
 540 static void *
 541 v3d_create_sampler_state(struct pipe_context *pctx,
 542                          const struct pipe_sampler_state *cso)
 543 {
 544         MAYBE_UNUSED struct v3d_context *v3d = v3d_context(pctx);
 545         struct v3d_sampler_state *so = CALLOC_STRUCT(v3d_sampler_state);
 546
 547         if (!so)
 548                 return NULL;
 549
 550         memcpy(so, cso, sizeof(*cso));
 551
 552         bool either_nearest =
 553                 (cso->mag_img_filter == PIPE_TEX_MIPFILTER_NEAREST ||
 554                  cso->min_img_filter == PIPE_TEX_MIPFILTER_NEAREST);
 555
 556 #if V3D_VERSION >= 40
 557         so->bo = v3d_bo_alloc(v3d->screen, cl_packet_length(SAMPLER_STATE),
 558                               "sampler");
 559         void *map = v3d_bo_map(so->bo);
 560
 561         v3dx_pack(map, SAMPLER_STATE, sampler) {
 562                 sampler.wrap_i_border = false;
 563
 564                 sampler.wrap_s = translate_wrap(cso->wrap_s, either_nearest);
 565                 sampler.wrap_t = translate_wrap(cso->wrap_t, either_nearest);
 566                 sampler.wrap_r = translate_wrap(cso->wrap_r, either_nearest);
 567
 568                 sampler.fixed_bias = cso->lod_bias;
 569                 sampler.depth_compare_function = cso->compare_func;
 570
 571                 sampler.min_filter_nearest =
 572                         cso->min_img_filter == PIPE_TEX_FILTER_NEAREST;
 573                 sampler.mag_filter_nearest =
 574                         cso->mag_img_filter == PIPE_TEX_FILTER_NEAREST;
 575                 sampler.mip_filter_nearest =
 576                         cso->min_mip_filter != PIPE_TEX_MIPFILTER_LINEAR;
 577
 578                 sampler.min_level_of_detail = MIN2(MAX2(0, cso->min_lod),
 579                                                    15);
 580                 sampler.max_level_of_detail = MIN2(cso->max_lod, 15);
 581
 582                 /* If we're not doing inter-miplevel filtering, we need to
 583                  * clamp the LOD so that we only sample from baselevel.
 584                  * However, we need to still allow the calculated LOD to be
 585                  * fractionally over the baselevel, so that the HW can decide
 586                  * between the min and mag filters.
 587                  */
 588                 if (cso->min_mip_filter == PIPE_TEX_MIPFILTER_NONE) {
 589                         sampler.min_level_of_detail =
 590                                 MIN2(sampler.min_level_of_detail, 1.0 / 256.0);
 591                         sampler.max_level_of_detail =
 592                                 MIN2(sampler.max_level_of_detail, 1.0 / 256.0);
 593                 }
 594
 595                 if (cso->max_anisotropy) {
 596                         sampler.anisotropy_enable = true;
 597
 598                         if (cso->max_anisotropy > 8)
 599                                 sampler.maximum_anisotropy = 3;
 600                         else if (cso->max_anisotropy > 4)
 601                                 sampler.maximum_anisotropy = 2;
 602                         else if (cso->max_anisotropy > 2)
 603                                 sampler.maximum_anisotropy = 1;
 604                 }
 605
 606                 sampler.border_color_mode = V3D_BORDER_COLOR_FOLLOWS;
 607                 /* XXX: The border color field is in the TMU blending format
 608                  * (32, f16, or i16), and we need to customize it based on
 609                  * that.
 610                  *
 611                  * XXX: for compat alpha formats, we need the alpha field to
 612                  * be in the red channel.
 613                  */
 614                 sampler.border_color_red =
 615                         util_float_to_half(cso->border_color.f[0]);
 616                 sampler.border_color_green =
 617                         util_float_to_half(cso->border_color.f[1]);
 618                 sampler.border_color_blue =
 619                         util_float_to_half(cso->border_color.f[2]);
 620                 sampler.border_color_alpha =
 621                         util_float_to_half(cso->border_color.f[3]);
 622         }
 623
 624 #else /* V3D_VERSION < 40 */
 625         v3dx_pack(&so->p0, TEXTURE_UNIFORM_PARAMETER_0_CFG_MODE1, p0) {
 626                 p0.s_wrap_mode = translate_wrap(cso->wrap_s, either_nearest);
 627                 p0.t_wrap_mode = translate_wrap(cso->wrap_t, either_nearest);
 628                 p0.r_wrap_mode = translate_wrap(cso->wrap_r, either_nearest);
 629         }
 630
 631         v3dx_pack(&so->texture_shader_state, TEXTURE_SHADER_STATE, tex) {
 632                 tex.depth_compare_function = cso->compare_func;
 633                 tex.fixed_bias = cso->lod_bias;
 634         }
 635 #endif /* V3D_VERSION < 40 */
 636         return so;
 637 }
 638
 639 static void
 640 v3d_sampler_states_bind(struct pipe_context *pctx,
 641                         enum pipe_shader_type shader, unsigned start,
 642                         unsigned nr, void **hwcso)
 643 {
 644         struct v3d_context *v3d = v3d_context(pctx);
 645         struct v3d_texture_stateobj *stage_tex = v3d_get_stage_tex(v3d, shader);
 646
 647         assert(start == 0);
 648         unsigned i;
 649         unsigned new_nr = 0;
 650
 651         for (i = 0; i < nr; i++) {
 652                 if (hwcso[i])
 653                         new_nr = i + 1;
 654                 stage_tex->samplers[i] = hwcso[i];
 655         }
 656
 657         for (; i < stage_tex->num_samplers; i++) {
 658                 stage_tex->samplers[i] = NULL;
 659         }
 660
 661         stage_tex->num_samplers = new_nr;
 662 }
 663
 664 static void
 665 v3d_sampler_state_delete(struct pipe_context *pctx,
 666                          void *hwcso)
 667 {
 668         struct pipe_sampler_state *psampler = hwcso;
 669         struct v3d_sampler_state *sampler = v3d_sampler_state(psampler);
 670
 671         v3d_bo_unreference(&sampler->bo);
 672         free(psampler);
 673 }
 674
 675 #if V3D_VERSION >= 40
 676 static uint32_t
 677 translate_swizzle(unsigned char pipe_swizzle)
 678 {
 679         switch (pipe_swizzle) {
 680         case PIPE_SWIZZLE_0:
 681                 return 0;
 682         case PIPE_SWIZZLE_1:
 683                 return 1;
 684         case PIPE_SWIZZLE_X:
 685         case PIPE_SWIZZLE_Y:
 686         case PIPE_SWIZZLE_Z:
 687         case PIPE_SWIZZLE_W:
 688                 return 2 + pipe_swizzle;
 689         default:
 690                 unreachable("unknown swizzle");
 691         }
 692 }
 693 #endif
 694
 695 static struct pipe_sampler_view *
 696 v3d_create_sampler_view(struct pipe_context *pctx, struct pipe_resource *prsc,
 697                         const struct pipe_sampler_view *cso)
 698 {
 699         struct v3d_context *v3d = v3d_context(pctx);
 700         struct v3d_screen *screen = v3d->screen;
 701         struct v3d_sampler_view *so = CALLOC_STRUCT(v3d_sampler_view);
 702         struct v3d_resource *rsc = v3d_resource(prsc);
 703
 704         if (!so)
 705                 return NULL;
 706
 707         so->base = *cso;
 708
 709         pipe_reference(NULL, &prsc->reference);
 710
 711         /* Compute the sampler view's swizzle up front. This will be plugged
 712          * into either the sampler (for 16-bit returns) or the shader's
 713          * texture key (for 32)
 714          */
 715         uint8_t view_swizzle[4] = {
 716                 cso->swizzle_r,
 717                 cso->swizzle_g,
 718                 cso->swizzle_b,
 719                 cso->swizzle_a
 720         };
 721         const uint8_t *fmt_swizzle =
 722                 v3d_get_format_swizzle(&screen->devinfo, so->base.format);
 723         util_format_compose_swizzles(fmt_swizzle, view_swizzle, so->swizzle);
 724
 725         so->base.texture = prsc;
 726         so->base.reference.count = 1;
 727         so->base.context = pctx;
 728
 729         int msaa_scale = prsc->nr_samples > 1 ? 2 : 1;
 730
 731 #if V3D_VERSION >= 40
 732         so->bo = v3d_bo_alloc(v3d->screen,
 733                               cl_packet_length(TEXTURE_SHADER_STATE), "sampler");
 734         void *map = v3d_bo_map(so->bo);
 735
 736         v3dx_pack(map, TEXTURE_SHADER_STATE, tex) {
 737 #else /* V3D_VERSION < 40 */
 738         STATIC_ASSERT(sizeof(so->texture_shader_state) >=
 739                       cl_packet_length(TEXTURE_SHADER_STATE));
 740         v3dx_pack(&so->texture_shader_state, TEXTURE_SHADER_STATE, tex) {
 741 #endif
 742
 743                 tex.image_width = prsc->width0 * msaa_scale;
 744                 tex.image_height = prsc->height0 * msaa_scale;
 745
 746 #if V3D_VERSION >= 40
 747                 /* On 4.x, the height of a 1D texture is redefined to be the
 748                  * upper 14 bits of the width (which is only usable with txf).
 749                  */
 750                 if (prsc->target == PIPE_TEXTURE_1D ||
 751                     prsc->target == PIPE_TEXTURE_1D_ARRAY) {
 752                         tex.image_height = tex.image_width >> 14;
 753                 }
 754 #endif
 755
 756                 if (prsc->target == PIPE_TEXTURE_3D) {
 757                         tex.image_depth = prsc->depth0;
 758                 } else {
 759                         tex.image_depth = (cso->u.tex.last_layer -
 760                                            cso->u.tex.first_layer) + 1;
 761                 }
 762
 763                 tex.srgb = util_format_is_srgb(cso->format);
 764
 765                 tex.base_level = cso->u.tex.first_level;
 766 #if V3D_VERSION >= 40
 767                 tex.max_level = cso->u.tex.last_level;
 768                 /* Note that we don't have a job to reference the texture's sBO
 769                  * at state create time, so any time this sampler view is used
 770                  * we need to add the texture to the job.
 771                  */
 772                 tex.texture_base_pointer = cl_address(NULL,
 773                                                       rsc->bo->offset +
 774                                                       rsc->slices[0].offset +
 775                                                       cso->u.tex.first_layer *
 776                                                       rsc->cube_map_stride),
 777
 778                 tex.swizzle_r = translate_swizzle(so->swizzle[0]);
 779                 tex.swizzle_g = translate_swizzle(so->swizzle[1]);
 780                 tex.swizzle_b = translate_swizzle(so->swizzle[2]);
 781                 tex.swizzle_a = translate_swizzle(so->swizzle[3]);
 782 #endif
 783                 tex.array_stride_64_byte_aligned = rsc->cube_map_stride / 64;
 784
 785                 if (prsc->nr_samples > 1 && V3D_VERSION < 40) {
 786                         /* Using texture views to reinterpret formats on our
 787                          * MSAA textures won't work, because we don't lay out
 788                          * the bits in memory as it's expected -- for example,
 789                          * RGBA8 and RGB10_A2 are compatible in the
 790                          * ARB_texture_view spec, but in HW we lay them out as
 791                          * 32bpp RGBA8 and 64bpp RGBA16F.  Just assert for now
 792                          * to catch failures.
 793                          *
 794                          * We explicitly allow remapping S8Z24 to RGBA8888 for
 795                          * v3d_blit.c's stencil blits.
 796                          */
 797                         assert((util_format_linear(cso->format) ==
 798                                 util_format_linear(prsc->format)) ||
 799                                (prsc->format == PIPE_FORMAT_S8_UINT_Z24_UNORM &&
 800                                 cso->format == PIPE_FORMAT_R8G8B8A8_UNORM));
 801                         uint32_t output_image_format =
 802                                 v3d_get_rt_format(&screen->devinfo, cso->format);
 803                         uint32_t internal_type;
 804                         uint32_t internal_bpp;
 805                         v3d_get_internal_type_bpp_for_output_format(&screen->devinfo,
 806                                                                     output_image_format,
 807                                                                     &internal_type,
 808                                                                     &internal_bpp);
 809
 810                         switch (internal_type) {
 811                         case V3D_INTERNAL_TYPE_8:
 812                                 tex.texture_type = TEXTURE_DATA_FORMAT_RGBA8;
 813                                 break;
 814                         case V3D_INTERNAL_TYPE_16F:
 815                                 tex.texture_type = TEXTURE_DATA_FORMAT_RGBA16F;
 816                                 break;
 817                         default:
 818                                 unreachable("Bad MSAA texture type");
 819                         }
 820
 821                         /* sRGB was stored in the tile buffer as linear and
 822                          * would have been encoded to sRGB on resolved tile
 823                          * buffer store.  Note that this means we would need
 824                          * shader code if we wanted to read an MSAA sRGB
 825                          * texture without sRGB decode.
 826                          */
 827                         tex.srgb = false;
 828                 } else {
 829                         tex.texture_type = v3d_get_tex_format(&screen->devinfo,
 830                                                               cso->format);
 831                 }
 832
 833                 /* Since other platform devices may produce UIF images even
 834                  * when they're not big enough for V3D to assume they're UIF,
 835                  * we force images with level 0 as UIF to be always treated
 836                  * that way.
 837                  */
 838                 tex.level_0_is_strictly_uif = (rsc->slices[0].tiling ==
 839                                                VC5_TILING_UIF_XOR ||
 840                                                rsc->slices[0].tiling ==
 841                                                VC5_TILING_UIF_NO_XOR);
 842                 tex.level_0_xor_enable = (rsc->slices[0].tiling ==
 843                                           VC5_TILING_UIF_XOR);
 844
 845                 if (tex.level_0_is_strictly_uif)
 846                         tex.level_0_ub_pad = rsc->slices[0].ub_pad;
 847
 848 #if V3D_VERSION >= 40
 849                 if (tex.uif_xor_disable ||
 850                     tex.level_0_is_strictly_uif) {
 851                         tex.extended = true;
 852                 }
 853 #endif /* V3D_VERSION >= 40 */
 854         };
 855
 856         return &so->base;
 857 }
 858
 859 static void
 860 v3d_sampler_view_destroy(struct pipe_context *pctx,
 861                          struct pipe_sampler_view *psview)
 862 {
 863         struct v3d_sampler_view *sview = v3d_sampler_view(psview);
 864
 865         v3d_bo_unreference(&sview->bo);
 866         pipe_resource_reference(&psview->texture, NULL);
 867         free(psview);
 868 }
 869
 870 static void
 871 v3d_set_sampler_views(struct pipe_context *pctx,
 872                       enum pipe_shader_type shader,
 873                       unsigned start, unsigned nr,
 874                       struct pipe_sampler_view **views)
 875 {
 876         struct v3d_context *v3d = v3d_context(pctx);
 877         struct v3d_texture_stateobj *stage_tex = v3d_get_stage_tex(v3d, shader);
 878         unsigned i;
 879         unsigned new_nr = 0;
 880
 881         assert(start == 0);
 882
 883         for (i = 0; i < nr; i++) {
 884                 if (views[i])
 885                         new_nr = i + 1;
 886                 pipe_sampler_view_reference(&stage_tex->textures[i], views[i]);
 887         }
 888
 889         for (; i < stage_tex->num_textures; i++) {
 890                 pipe_sampler_view_reference(&stage_tex->textures[i], NULL);
 891         }
 892
 893         stage_tex->num_textures = new_nr;
 894 }
 895
 896 static struct pipe_stream_output_target *
 897 v3d_create_stream_output_target(struct pipe_context *pctx,
 898                                 struct pipe_resource *prsc,
 899                                 unsigned buffer_offset,
 900                                 unsigned buffer_size)
 901 {
 902         struct pipe_stream_output_target *target;
 903
 904         target = CALLOC_STRUCT(pipe_stream_output_target);
 905         if (!target)
 906                 return NULL;
 907
 908         pipe_reference_init(&target->reference, 1);
 909         pipe_resource_reference(&target->buffer, prsc);
 910
 911         target->context = pctx;
 912         target->buffer_offset = buffer_offset;
 913         target->buffer_size = buffer_size;
 914
 915         return target;
 916 }
 917
 918 static void
 919 v3d_stream_output_target_destroy(struct pipe_context *pctx,
 920                                  struct pipe_stream_output_target *target)
 921 {
 922         pipe_resource_reference(&target->buffer, NULL);
 923         free(target);
 924 }
 925
 926 static void
 927 v3d_set_stream_output_targets(struct pipe_context *pctx,
 928                               unsigned num_targets,
 929                               struct pipe_stream_output_target **targets,
 930                               const unsigned *offsets)
 931 {
 932         struct v3d_context *ctx = v3d_context(pctx);
 933         struct v3d_streamout_stateobj *so = &ctx->streamout;
 934         unsigned i;
 935
 936         assert(num_targets <= ARRAY_SIZE(so->targets));
 937
 938         for (i = 0; i < num_targets; i++) {
 939                 if (offsets[i] != -1)
 940                         so->offsets[i] = offsets[i];
 941
 942                 pipe_so_target_reference(&so->targets[i], targets[i]);
 943         }
 944
 945         for (; i < so->num_targets; i++)
 946                 pipe_so_target_reference(&so->targets[i], NULL);
 947
 948         so->num_targets = num_targets;
 949
 950         ctx->dirty |= VC5_DIRTY_STREAMOUT;
 951 }
 952
 953 void
 954 v3dX(state_init)(struct pipe_context *pctx)
 955 {
 956         pctx->set_blend_color = v3d_set_blend_color;
 957         pctx->set_stencil_ref = v3d_set_stencil_ref;
 958         pctx->set_clip_state = v3d_set_clip_state;
 959         pctx->set_sample_mask = v3d_set_sample_mask;
 960         pctx->set_constant_buffer = v3d_set_constant_buffer;
 961         pctx->set_framebuffer_state = v3d_set_framebuffer_state;
 962         pctx->set_polygon_stipple = v3d_set_polygon_stipple;
 963         pctx->set_scissor_states = v3d_set_scissor_states;
 964         pctx->set_viewport_states = v3d_set_viewport_states;
 965
 966         pctx->set_vertex_buffers = v3d_set_vertex_buffers;
 967
 968         pctx->create_blend_state = v3d_create_blend_state;
 969         pctx->bind_blend_state = v3d_blend_state_bind;
 970         pctx->delete_blend_state = v3d_generic_cso_state_delete;
 971
 972         pctx->create_rasterizer_state = v3d_create_rasterizer_state;
 973         pctx->bind_rasterizer_state = v3d_rasterizer_state_bind;
 974         pctx->delete_rasterizer_state = v3d_generic_cso_state_delete;
 975
 976         pctx->create_depth_stencil_alpha_state = v3d_create_depth_stencil_alpha_state;
 977         pctx->bind_depth_stencil_alpha_state = v3d_zsa_state_bind;
 978         pctx->delete_depth_stencil_alpha_state = v3d_generic_cso_state_delete;
 979
 980         pctx->create_vertex_elements_state = v3d_vertex_state_create;
 981         pctx->delete_vertex_elements_state = v3d_vertex_state_delete;
 982         pctx->bind_vertex_elements_state = v3d_vertex_state_bind;
 983
 984         pctx->create_sampler_state = v3d_create_sampler_state;
 985         pctx->delete_sampler_state = v3d_sampler_state_delete;
 986         pctx->bind_sampler_states = v3d_sampler_states_bind;
 987
 988         pctx->create_sampler_view = v3d_create_sampler_view;
 989         pctx->sampler_view_destroy = v3d_sampler_view_destroy;
 990         pctx->set_sampler_views = v3d_set_sampler_views;
 991
 992         pctx->create_stream_output_target = v3d_create_stream_output_target;
 993         pctx->stream_output_target_destroy = v3d_stream_output_target_destroy;
 994         pctx->set_stream_output_targets = v3d_set_stream_output_targets;
 995 }