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