src/gallium/drivers/nouveau/nvc0/nvc0_screen.c

   1 /*
   2  * Copyright 2010 Christoph Bumiller
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #include <xf86drm.h>
  24 #include <nouveau_drm.h>
  25 #include <nvif/class.h>
  26 #include "util/u_format.h"
  27 #include "util/u_format_s3tc.h"
  28 #include "pipe/p_screen.h"
  29
  30 #include "vl/vl_decoder.h"
  31 #include "vl/vl_video_buffer.h"
  32
  33 #include "nouveau_vp3_video.h"
  34
  35 #include "nvc0/nvc0_context.h"
  36 #include "nvc0/nvc0_screen.h"
  37
  38 #include "nvc0/mme/com9097.mme.h"
  39 #include "nvc0/mme/com90c0.mme.h"
  40
  41 static boolean
  42 nvc0_screen_is_format_supported(struct pipe_screen *pscreen,
  43                                 enum pipe_format format,
  44                                 enum pipe_texture_target target,
  45                                 unsigned sample_count,
  46                                 unsigned bindings)
  47 {
  48    const struct util_format_description *desc = util_format_description(format);
  49
  50    if (sample_count > 8)
  51       return false;
  52    if (!(0x117 & (1 << sample_count))) /* 0, 1, 2, 4 or 8 */
  53       return false;
  54
  55    /* Short-circuit the rest of the logic -- this is used by the state tracker
  56     * to determine valid MS levels in a no-attachments scenario.
  57     */
  58    if (format == PIPE_FORMAT_NONE && bindings & PIPE_BIND_RENDER_TARGET)
  59       return true;
  60
  61    if (!util_format_is_supported(format, bindings))
  62       return false;
  63
  64    if ((bindings & PIPE_BIND_SAMPLER_VIEW) && (target != PIPE_BUFFER))
  65       if (util_format_get_blocksizebits(format) == 3 * 32)
  66          return false;
  67
  68    if (bindings & PIPE_BIND_LINEAR)
  69       if (util_format_is_depth_or_stencil(format) ||
  70           (target != PIPE_TEXTURE_1D &&
  71            target != PIPE_TEXTURE_2D &&
  72            target != PIPE_TEXTURE_RECT) ||
  73           sample_count > 1)
  74          return false;
  75
  76    /* Restrict ETC2 and ASTC formats here. These are only supported on GK20A.
  77     */
  78    if ((desc->layout == UTIL_FORMAT_LAYOUT_ETC ||
  79         desc->layout == UTIL_FORMAT_LAYOUT_ASTC) &&
  80        /* The claim is that this should work on GM107 but it doesn't. Need to
  81         * test further and figure out if it's a nouveau issue or a HW one.
  82        nouveau_screen(pscreen)->class_3d < GM107_3D_CLASS &&
  83         */
  84        nouveau_screen(pscreen)->class_3d != NVEA_3D_CLASS)
  85       return false;
  86
  87    /* transfers & shared are always supported */
  88    bindings &= ~(PIPE_BIND_TRANSFER_READ |
  89                  PIPE_BIND_TRANSFER_WRITE |
  90                  PIPE_BIND_LINEAR |
  91                  PIPE_BIND_SHARED);
  92
  93    return (( nvc0_format_table[format].usage |
  94             nvc0_vertex_format[format].usage) & bindings) == bindings;
  95 }
  96
  97 static int
  98 nvc0_screen_get_param(struct pipe_screen *pscreen, enum pipe_cap param)
  99 {
 100    const uint16_t class_3d = nouveau_screen(pscreen)->class_3d;
 101    struct nouveau_device *dev = nouveau_screen(pscreen)->device;
 102
 103    switch (param) {
 104    /* non-boolean caps */
 105    case PIPE_CAP_MAX_TEXTURE_2D_LEVELS:
 106    case PIPE_CAP_MAX_TEXTURE_CUBE_LEVELS:
 107       return 15;
 108    case PIPE_CAP_MAX_TEXTURE_3D_LEVELS:
 109       return 12;
 110    case PIPE_CAP_MAX_TEXTURE_ARRAY_LAYERS:
 111       return 2048;
 112    case PIPE_CAP_MIN_TEXEL_OFFSET:
 113       return -8;
 114    case PIPE_CAP_MAX_TEXEL_OFFSET:
 115       return 7;
 116    case PIPE_CAP_MIN_TEXTURE_GATHER_OFFSET:
 117       return -32;
 118    case PIPE_CAP_MAX_TEXTURE_GATHER_OFFSET:
 119       return 31;
 120    case PIPE_CAP_MAX_TEXTURE_BUFFER_SIZE:
 121       return 128 * 1024 * 1024;
 122    case PIPE_CAP_GLSL_FEATURE_LEVEL:
 123       if (class_3d <= NVF0_3D_CLASS)
 124          return 430;
 125       return 410;
 126    case PIPE_CAP_MAX_RENDER_TARGETS:
 127       return 8;
 128    case PIPE_CAP_MAX_DUAL_SOURCE_RENDER_TARGETS:
 129       return 1;
 130    case PIPE_CAP_MAX_STREAM_OUTPUT_BUFFERS:
 131       return 4;
 132    case PIPE_CAP_MAX_STREAM_OUTPUT_SEPARATE_COMPONENTS:
 133    case PIPE_CAP_MAX_STREAM_OUTPUT_INTERLEAVED_COMPONENTS:
 134       return 128;
 135    case PIPE_CAP_MAX_GEOMETRY_OUTPUT_VERTICES:
 136    case PIPE_CAP_MAX_GEOMETRY_TOTAL_OUTPUT_COMPONENTS:
 137       return 1024;
 138    case PIPE_CAP_MAX_VERTEX_STREAMS:
 139       return 4;
 140    case PIPE_CAP_MAX_VERTEX_ATTRIB_STRIDE:
 141       return 2048;
 142    case PIPE_CAP_CONSTANT_BUFFER_OFFSET_ALIGNMENT:
 143       return 256;
 144    case PIPE_CAP_TEXTURE_BUFFER_OFFSET_ALIGNMENT:
 145       return 16; /* 256 for binding as RT, but that's not possible in GL */
 146    case PIPE_CAP_SHADER_BUFFER_OFFSET_ALIGNMENT:
 147       return 16;
 148    case PIPE_CAP_MIN_MAP_BUFFER_ALIGNMENT:
 149       return NOUVEAU_MIN_BUFFER_MAP_ALIGN;
 150    case PIPE_CAP_MAX_VIEWPORTS:
 151       return NVC0_MAX_VIEWPORTS;
 152    case PIPE_CAP_MAX_TEXTURE_GATHER_COMPONENTS:
 153       return 4;
 154    case PIPE_CAP_TEXTURE_BORDER_COLOR_QUIRK:
 155       return PIPE_QUIRK_TEXTURE_BORDER_COLOR_SWIZZLE_NV50;
 156    case PIPE_CAP_ENDIANNESS:
 157       return PIPE_ENDIAN_LITTLE;
 158    case PIPE_CAP_MAX_SHADER_PATCH_VARYINGS:
 159       return 30;
 160    case PIPE_CAP_MAX_WINDOW_RECTANGLES:
 161       return NVC0_MAX_WINDOW_RECTANGLES;
 162
 163    /* supported caps */
 164    case PIPE_CAP_TEXTURE_MIRROR_CLAMP:
 165    case PIPE_CAP_TEXTURE_SWIZZLE:
 166    case PIPE_CAP_TEXTURE_SHADOW_MAP:
 167    case PIPE_CAP_NPOT_TEXTURES:
 168    case PIPE_CAP_MIXED_FRAMEBUFFER_SIZES:
 169    case PIPE_CAP_ANISOTROPIC_FILTER:
 170    case PIPE_CAP_SEAMLESS_CUBE_MAP:
 171    case PIPE_CAP_CUBE_MAP_ARRAY:
 172    case PIPE_CAP_TEXTURE_BUFFER_OBJECTS:
 173    case PIPE_CAP_TEXTURE_MULTISAMPLE:
 174    case PIPE_CAP_TWO_SIDED_STENCIL:
 175    case PIPE_CAP_DEPTH_CLIP_DISABLE:
 176    case PIPE_CAP_POINT_SPRITE:
 177    case PIPE_CAP_TGSI_TEXCOORD:
 178    case PIPE_CAP_SM3:
 179    case PIPE_CAP_FRAGMENT_COLOR_CLAMPED:
 180    case PIPE_CAP_VERTEX_COLOR_UNCLAMPED:
 181    case PIPE_CAP_VERTEX_COLOR_CLAMPED:
 182    case PIPE_CAP_QUERY_TIMESTAMP:
 183    case PIPE_CAP_QUERY_TIME_ELAPSED:
 184    case PIPE_CAP_OCCLUSION_QUERY:
 185    case PIPE_CAP_STREAM_OUTPUT_PAUSE_RESUME:
 186    case PIPE_CAP_QUERY_PIPELINE_STATISTICS:
 187    case PIPE_CAP_BLEND_EQUATION_SEPARATE:
 188    case PIPE_CAP_INDEP_BLEND_ENABLE:
 189    case PIPE_CAP_INDEP_BLEND_FUNC:
 190    case PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT:
 191    case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER:
 192    case PIPE_CAP_PRIMITIVE_RESTART:
 193    case PIPE_CAP_TGSI_INSTANCEID:
 194    case PIPE_CAP_VERTEX_ELEMENT_INSTANCE_DIVISOR:
 195    case PIPE_CAP_MIXED_COLORBUFFER_FORMATS:
 196    case PIPE_CAP_CONDITIONAL_RENDER:
 197    case PIPE_CAP_TEXTURE_BARRIER:
 198    case PIPE_CAP_QUADS_FOLLOW_PROVOKING_VERTEX_CONVENTION:
 199    case PIPE_CAP_START_INSTANCE:
 200    case PIPE_CAP_BUFFER_MAP_PERSISTENT_COHERENT:
 201    case PIPE_CAP_DRAW_INDIRECT:
 202    case PIPE_CAP_USER_CONSTANT_BUFFERS:
 203    case PIPE_CAP_USER_INDEX_BUFFERS:
 204    case PIPE_CAP_USER_VERTEX_BUFFERS:
 205    case PIPE_CAP_TEXTURE_QUERY_LOD:
 206    case PIPE_CAP_SAMPLE_SHADING:
 207    case PIPE_CAP_TEXTURE_GATHER_OFFSETS:
 208    case PIPE_CAP_TEXTURE_GATHER_SM5:
 209    case PIPE_CAP_TGSI_FS_FINE_DERIVATIVE:
 210    case PIPE_CAP_CONDITIONAL_RENDER_INVERTED:
 211    case PIPE_CAP_SAMPLER_VIEW_TARGET:
 212    case PIPE_CAP_CLIP_HALFZ:
 213    case PIPE_CAP_POLYGON_OFFSET_CLAMP:
 214    case PIPE_CAP_MULTISAMPLE_Z_RESOLVE:
 215    case PIPE_CAP_TEXTURE_FLOAT_LINEAR:
 216    case PIPE_CAP_TEXTURE_HALF_FLOAT_LINEAR:
 217    case PIPE_CAP_DEPTH_BOUNDS_TEST:
 218    case PIPE_CAP_TGSI_TXQS:
 219    case PIPE_CAP_COPY_BETWEEN_COMPRESSED_AND_PLAIN_FORMATS:
 220    case PIPE_CAP_FORCE_PERSAMPLE_INTERP:
 221    case PIPE_CAP_SHAREABLE_SHADERS:
 222    case PIPE_CAP_CLEAR_TEXTURE:
 223    case PIPE_CAP_DRAW_PARAMETERS:
 224    case PIPE_CAP_TGSI_PACK_HALF_FLOAT:
 225    case PIPE_CAP_MULTI_DRAW_INDIRECT:
 226    case PIPE_CAP_MULTI_DRAW_INDIRECT_PARAMS:
 227    case PIPE_CAP_TGSI_FS_FACE_IS_INTEGER_SYSVAL:
 228    case PIPE_CAP_QUERY_BUFFER_OBJECT:
 229    case PIPE_CAP_INVALIDATE_BUFFER:
 230    case PIPE_CAP_STRING_MARKER:
 231    case PIPE_CAP_FRAMEBUFFER_NO_ATTACHMENT:
 232    case PIPE_CAP_CULL_DISTANCE:
 233    case PIPE_CAP_PRIMITIVE_RESTART_FOR_PATCHES:
 234    case PIPE_CAP_ROBUST_BUFFER_ACCESS_BEHAVIOR:
 235    case PIPE_CAP_COMPUTE:
 236    case PIPE_CAP_TGSI_VOTE:
 237    case PIPE_CAP_POLYGON_OFFSET_UNITS_UNSCALED:
 238       return 1;
 239    case PIPE_CAP_SEAMLESS_CUBE_MAP_PER_TEXTURE:
 240       return (class_3d >= NVE4_3D_CLASS) ? 1 : 0;
 241    case PIPE_CAP_PREFER_BLIT_BASED_TEXTURE_TRANSFER:
 242       return nouveau_screen(pscreen)->vram_domain & NOUVEAU_BO_VRAM ? 1 : 0;
 243
 244    /* unsupported caps */
 245    case PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT:
 246    case PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER:
 247    case PIPE_CAP_SHADER_STENCIL_EXPORT:
 248    case PIPE_CAP_TGSI_CAN_COMPACT_CONSTANTS:
 249    case PIPE_CAP_VERTEX_BUFFER_OFFSET_4BYTE_ALIGNED_ONLY:
 250    case PIPE_CAP_VERTEX_BUFFER_STRIDE_4BYTE_ALIGNED_ONLY:
 251    case PIPE_CAP_VERTEX_ELEMENT_SRC_OFFSET_4BYTE_ALIGNED_ONLY:
 252    case PIPE_CAP_TGSI_VS_LAYER_VIEWPORT:
 253    case PIPE_CAP_FAKE_SW_MSAA:
 254    case PIPE_CAP_TGSI_VS_WINDOW_SPACE_POSITION:
 255    case PIPE_CAP_VERTEXID_NOBASE:
 256    case PIPE_CAP_RESOURCE_FROM_USER_MEMORY:
 257    case PIPE_CAP_DEVICE_RESET_STATUS_QUERY:
 258    case PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL:
 259    case PIPE_CAP_GENERATE_MIPMAP:
 260    case PIPE_CAP_BUFFER_SAMPLER_VIEW_RGBA_ONLY:
 261    case PIPE_CAP_SURFACE_REINTERPRET_BLOCKS:
 262    case PIPE_CAP_QUERY_MEMORY_INFO:
 263    case PIPE_CAP_PCI_GROUP:
 264    case PIPE_CAP_PCI_BUS:
 265    case PIPE_CAP_PCI_DEVICE:
 266    case PIPE_CAP_PCI_FUNCTION:
 267       return 0;
 268
 269    case PIPE_CAP_VENDOR_ID:
 270       return 0x10de;
 271    case PIPE_CAP_DEVICE_ID: {
 272       uint64_t device_id;
 273       if (nouveau_getparam(dev, NOUVEAU_GETPARAM_PCI_DEVICE, &device_id)) {
 274          NOUVEAU_ERR("NOUVEAU_GETPARAM_PCI_DEVICE failed.\n");
 275          return -1;
 276       }
 277       return device_id;
 278    }
 279    case PIPE_CAP_ACCELERATED:
 280       return 1;
 281    case PIPE_CAP_VIDEO_MEMORY:
 282       return dev->vram_size >> 20;
 283    case PIPE_CAP_UMA:
 284       return 0;
 285    }
 286
 287    NOUVEAU_ERR("unknown PIPE_CAP %d\n", param);
 288    return 0;
 289 }
 290
 291 static int
 292 nvc0_screen_get_shader_param(struct pipe_screen *pscreen, unsigned shader,
 293                              enum pipe_shader_cap param)
 294 {
 295    const uint16_t class_3d = nouveau_screen(pscreen)->class_3d;
 296
 297    switch (shader) {
 298    case PIPE_SHADER_VERTEX:
 299    case PIPE_SHADER_GEOMETRY:
 300    case PIPE_SHADER_FRAGMENT:
 301    case PIPE_SHADER_COMPUTE:
 302       break;
 303    case PIPE_SHADER_TESS_CTRL:
 304    case PIPE_SHADER_TESS_EVAL:
 305       if (class_3d >= GM107_3D_CLASS)
 306          return 0;
 307       break;
 308    default:
 309       return 0;
 310    }
 311
 312    switch (param) {
 313    case PIPE_SHADER_CAP_PREFERRED_IR:
 314       return PIPE_SHADER_IR_TGSI;
 315    case PIPE_SHADER_CAP_SUPPORTED_IRS:
 316       return 1 << PIPE_SHADER_IR_TGSI;
 317    case PIPE_SHADER_CAP_MAX_INSTRUCTIONS:
 318    case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS:
 319    case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS:
 320    case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
 321       return 16384;
 322    case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
 323       return 16;
 324    case PIPE_SHADER_CAP_MAX_INPUTS:
 325       if (shader == PIPE_SHADER_VERTEX)
 326          return 32;
 327       /* NOTE: These only count our slots for GENERIC varyings.
 328        * The address space may be larger, but the actual hard limit seems to be
 329        * less than what the address space layout permits, so don't add TEXCOORD,
 330        * COLOR, etc. here.
 331        */
 332       if (shader == PIPE_SHADER_FRAGMENT)
 333          return 0x1f0 / 16;
 334       /* Actually this counts CLIPVERTEX, which occupies the last generic slot,
 335        * and excludes 0x60 per-patch inputs.
 336        */
 337       return 0x200 / 16;
 338    case PIPE_SHADER_CAP_MAX_OUTPUTS:
 339       return 32;
 340    case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE:
 341       return 65536;
 342    case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
 343       return NVC0_MAX_PIPE_CONSTBUFS;
 344    case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
 345       return shader != PIPE_SHADER_FRAGMENT;
 346    case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
 347       return shader != PIPE_SHADER_FRAGMENT || class_3d < GM107_3D_CLASS;
 348    case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
 349    case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
 350       return 1;
 351    case PIPE_SHADER_CAP_MAX_PREDS:
 352       return 0;
 353    case PIPE_SHADER_CAP_MAX_TEMPS:
 354       return NVC0_CAP_MAX_PROGRAM_TEMPS;
 355    case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED:
 356       return 1;
 357    case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED:
 358       return 1;
 359    case PIPE_SHADER_CAP_SUBROUTINES:
 360       return 1;
 361    case PIPE_SHADER_CAP_INTEGERS:
 362       return 1;
 363    case PIPE_SHADER_CAP_DOUBLES:
 364       return 1;
 365    case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
 366       return 1;
 367    case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
 368       return 1;
 369    case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
 370    case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
 371       return 0;
 372    case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
 373       return NVC0_MAX_BUFFERS;
 374    case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
 375       return (class_3d >= NVE4_3D_CLASS) ? 32 : 16;
 376    case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
 377       return (class_3d >= NVE4_3D_CLASS) ? 32 : 16;
 378    case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
 379       return 32;
 380    case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
 381       if (class_3d == NVE4_3D_CLASS || class_3d == NVF0_3D_CLASS)
 382          return NVC0_MAX_IMAGES;
 383       if (class_3d < NVE4_3D_CLASS)
 384          if (shader == PIPE_SHADER_FRAGMENT || shader == PIPE_SHADER_COMPUTE)
 385             return NVC0_MAX_IMAGES;
 386       return 0;
 387    default:
 388       NOUVEAU_ERR("unknown PIPE_SHADER_CAP %d\n", param);
 389       return 0;
 390    }
 391 }
 392
 393 static float
 394 nvc0_screen_get_paramf(struct pipe_screen *pscreen, enum pipe_capf param)
 395 {
 396    switch (param) {
 397    case PIPE_CAPF_MAX_LINE_WIDTH:
 398    case PIPE_CAPF_MAX_LINE_WIDTH_AA:
 399       return 10.0f;
 400    case PIPE_CAPF_MAX_POINT_WIDTH:
 401       return 63.0f;
 402    case PIPE_CAPF_MAX_POINT_WIDTH_AA:
 403       return 63.375f;
 404    case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY:
 405       return 16.0f;
 406    case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS:
 407       return 15.0f;
 408    case PIPE_CAPF_GUARD_BAND_LEFT:
 409    case PIPE_CAPF_GUARD_BAND_TOP:
 410       return 0.0f;
 411    case PIPE_CAPF_GUARD_BAND_RIGHT:
 412    case PIPE_CAPF_GUARD_BAND_BOTTOM:
 413       return 0.0f; /* that or infinity */
 414    }
 415
 416    NOUVEAU_ERR("unknown PIPE_CAPF %d\n", param);
 417    return 0.0f;
 418 }
 419
 420 static int
 421 nvc0_screen_get_compute_param(struct pipe_screen *pscreen,
 422                               enum pipe_shader_ir ir_type,
 423                               enum pipe_compute_cap param, void *data)
 424 {
 425    struct nvc0_screen *screen = nvc0_screen(pscreen);
 426    const uint16_t obj_class = screen->compute->oclass;
 427
 428 #define RET(x) do {                  \
 429    if (data)                         \
 430       memcpy(data, x, sizeof(x));    \
 431    return sizeof(x);                 \
 432 } while (0)
 433
 434    switch (param) {
 435    case PIPE_COMPUTE_CAP_GRID_DIMENSION:
 436       RET((uint64_t []) { 3 });
 437    case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
 438       if (obj_class >= NVE4_COMPUTE_CLASS) {
 439          RET(((uint64_t []) { 0x7fffffff, 65535, 65535 }));
 440       } else {
 441          RET(((uint64_t []) { 65535, 65535, 65535 }));
 442       }
 443    case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
 444       RET(((uint64_t []) { 1024, 1024, 64 }));
 445    case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
 446       RET((uint64_t []) { 1024 });
 447    case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE: /* g[] */
 448       RET((uint64_t []) { 1ULL << 40 });
 449    case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE: /* s[] */
 450       switch (obj_class) {
 451       case GM200_COMPUTE_CLASS:
 452          RET((uint64_t []) { 96 << 10 });
 453          break;
 454       case GM107_COMPUTE_CLASS:
 455          RET((uint64_t []) { 64 << 10 });
 456          break;
 457       default:
 458          RET((uint64_t []) { 48 << 10 });
 459          break;
 460       }
 461    case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE: /* l[] */
 462       RET((uint64_t []) { 512 << 10 });
 463    case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE: /* c[], arbitrary limit */
 464       RET((uint64_t []) { 4096 });
 465    case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
 466       RET((uint32_t []) { 32 });
 467    case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
 468       RET((uint64_t []) { 1ULL << 40 });
 469    case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
 470       RET((uint32_t []) { 0 });
 471    case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
 472       RET((uint32_t []) { screen->mp_count_compute });
 473    case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
 474       RET((uint32_t []) { 512 }); /* FIXME: arbitrary limit */
 475    default:
 476       return 0;
 477    }
 478
 479 #undef RET
 480 }
 481
 482 static void
 483 nvc0_screen_destroy(struct pipe_screen *pscreen)
 484 {
 485    struct nvc0_screen *screen = nvc0_screen(pscreen);
 486
 487    if (!nouveau_drm_screen_unref(&screen->base))
 488       return;
 489
 490    if (screen->base.fence.current) {
 491       struct nouveau_fence *current = NULL;
 492
 493       /* nouveau_fence_wait will create a new current fence, so wait on the
 494        * _current_ one, and remove both.
 495        */
 496       nouveau_fence_ref(screen->base.fence.current, &current);
 497       nouveau_fence_wait(current, NULL);
 498       nouveau_fence_ref(NULL, &current);
 499       nouveau_fence_ref(NULL, &screen->base.fence.current);
 500    }
 501    if (screen->base.pushbuf)
 502       screen->base.pushbuf->user_priv = NULL;
 503
 504    if (screen->blitter)
 505       nvc0_blitter_destroy(screen);
 506    if (screen->pm.prog) {
 507       screen->pm.prog->code = NULL; /* hardcoded, don't FREE */
 508       nvc0_program_destroy(NULL, screen->pm.prog);
 509       FREE(screen->pm.prog);
 510    }
 511
 512    nouveau_bo_ref(NULL, &screen->text);
 513    nouveau_bo_ref(NULL, &screen->uniform_bo);
 514    nouveau_bo_ref(NULL, &screen->tls);
 515    nouveau_bo_ref(NULL, &screen->txc);
 516    nouveau_bo_ref(NULL, &screen->fence.bo);
 517    nouveau_bo_ref(NULL, &screen->poly_cache);
 518
 519    nouveau_heap_destroy(&screen->lib_code);
 520    nouveau_heap_destroy(&screen->text_heap);
 521
 522    FREE(screen->tic.entries);
 523
 524    nouveau_object_del(&screen->eng3d);
 525    nouveau_object_del(&screen->eng2d);
 526    nouveau_object_del(&screen->m2mf);
 527    nouveau_object_del(&screen->compute);
 528    nouveau_object_del(&screen->nvsw);
 529
 530    nouveau_screen_fini(&screen->base);
 531
 532    FREE(screen);
 533 }
 534
 535 static int
 536 nvc0_graph_set_macro(struct nvc0_screen *screen, uint32_t m, unsigned pos,
 537                      unsigned size, const uint32_t *data)
 538 {
 539    struct nouveau_pushbuf *push = screen->base.pushbuf;
 540
 541    size /= 4;
 542
 543    assert((pos + size) <= 0x800);
 544
 545    BEGIN_NVC0(push, SUBC_3D(NVC0_GRAPH_MACRO_ID), 2);
 546    PUSH_DATA (push, (m - 0x3800) / 8);
 547    PUSH_DATA (push, pos);
 548    BEGIN_1IC0(push, SUBC_3D(NVC0_GRAPH_MACRO_UPLOAD_POS), size + 1);
 549    PUSH_DATA (push, pos);
 550    PUSH_DATAp(push, data, size);
 551
 552    return pos + size;
 553 }
 554
 555 static void
 556 nvc0_magic_3d_init(struct nouveau_pushbuf *push, uint16_t obj_class)
 557 {
 558    BEGIN_NVC0(push, SUBC_3D(0x10cc), 1);
 559    PUSH_DATA (push, 0xff);
 560    BEGIN_NVC0(push, SUBC_3D(0x10e0), 2);
 561    PUSH_DATA (push, 0xff);
 562    PUSH_DATA (push, 0xff);
 563    BEGIN_NVC0(push, SUBC_3D(0x10ec), 2);
 564    PUSH_DATA (push, 0xff);
 565    PUSH_DATA (push, 0xff);
 566    BEGIN_NVC0(push, SUBC_3D(0x074c), 1);
 567    PUSH_DATA (push, 0x3f);
 568
 569    BEGIN_NVC0(push, SUBC_3D(0x16a8), 1);
 570    PUSH_DATA (push, (3 << 16) | 3);
 571    BEGIN_NVC0(push, SUBC_3D(0x1794), 1);
 572    PUSH_DATA (push, (2 << 16) | 2);
 573
 574    if (obj_class < GM107_3D_CLASS) {
 575       BEGIN_NVC0(push, SUBC_3D(0x12ac), 1);
 576       PUSH_DATA (push, 0);
 577    }
 578    BEGIN_NVC0(push, SUBC_3D(0x0218), 1);
 579    PUSH_DATA (push, 0x10);
 580    BEGIN_NVC0(push, SUBC_3D(0x10fc), 1);
 581    PUSH_DATA (push, 0x10);
 582    BEGIN_NVC0(push, SUBC_3D(0x1290), 1);
 583    PUSH_DATA (push, 0x10);
 584    BEGIN_NVC0(push, SUBC_3D(0x12d8), 2);
 585    PUSH_DATA (push, 0x10);
 586    PUSH_DATA (push, 0x10);
 587    BEGIN_NVC0(push, SUBC_3D(0x1140), 1);
 588    PUSH_DATA (push, 0x10);
 589    BEGIN_NVC0(push, SUBC_3D(0x1610), 1);
 590    PUSH_DATA (push, 0xe);
 591
 592    BEGIN_NVC0(push, NVC0_3D(VERTEX_ID_GEN_MODE), 1);
 593    PUSH_DATA (push, NVC0_3D_VERTEX_ID_GEN_MODE_DRAW_ARRAYS_ADD_START);
 594    BEGIN_NVC0(push, SUBC_3D(0x030c), 1);
 595    PUSH_DATA (push, 0);
 596    BEGIN_NVC0(push, SUBC_3D(0x0300), 1);
 597    PUSH_DATA (push, 3);
 598
 599    BEGIN_NVC0(push, SUBC_3D(0x02d0), 1);
 600    PUSH_DATA (push, 0x3fffff);
 601    BEGIN_NVC0(push, SUBC_3D(0x0fdc), 1);
 602    PUSH_DATA (push, 1);
 603    BEGIN_NVC0(push, SUBC_3D(0x19c0), 1);
 604    PUSH_DATA (push, 1);
 605
 606    if (obj_class < GM107_3D_CLASS) {
 607       BEGIN_NVC0(push, SUBC_3D(0x075c), 1);
 608       PUSH_DATA (push, 3);
 609
 610       if (obj_class >= NVE4_3D_CLASS) {
 611          BEGIN_NVC0(push, SUBC_3D(0x07fc), 1);
 612          PUSH_DATA (push, 1);
 613       }
 614    }
 615
 616    /* TODO: find out what software methods 0x1528, 0x1280 and (on nve4) 0x02dc
 617     * are supposed to do */
 618 }
 619
 620 static void
 621 nvc0_screen_fence_emit(struct pipe_screen *pscreen, u32 *sequence)
 622 {
 623    struct nvc0_screen *screen = nvc0_screen(pscreen);
 624    struct nouveau_pushbuf *push = screen->base.pushbuf;
 625
 626    /* we need to do it after possible flush in MARK_RING */
 627    *sequence = ++screen->base.fence.sequence;
 628
 629    assert(PUSH_AVAIL(push) + push->rsvd_kick >= 5);
 630    PUSH_DATA (push, NVC0_FIFO_PKHDR_SQ(NVC0_3D(QUERY_ADDRESS_HIGH), 4));
 631    PUSH_DATAh(push, screen->fence.bo->offset);
 632    PUSH_DATA (push, screen->fence.bo->offset);
 633    PUSH_DATA (push, *sequence);
 634    PUSH_DATA (push, NVC0_3D_QUERY_GET_FENCE | NVC0_3D_QUERY_GET_SHORT |
 635               (0xf << NVC0_3D_QUERY_GET_UNIT__SHIFT));
 636 }
 637
 638 static u32
 639 nvc0_screen_fence_update(struct pipe_screen *pscreen)
 640 {
 641    struct nvc0_screen *screen = nvc0_screen(pscreen);
 642    return screen->fence.map[0];
 643 }
 644
 645 static int
 646 nvc0_screen_init_compute(struct nvc0_screen *screen)
 647 {
 648    screen->base.base.get_compute_param = nvc0_screen_get_compute_param;
 649
 650    switch (screen->base.device->chipset & ~0xf) {
 651    case 0xc0:
 652    case 0xd0:
 653       return nvc0_screen_compute_setup(screen, screen->base.pushbuf);
 654    case 0xe0:
 655    case 0xf0:
 656    case 0x100:
 657    case 0x110:
 658    case 0x120:
 659       return nve4_screen_compute_setup(screen, screen->base.pushbuf);
 660    default:
 661       return -1;
 662    }
 663 }
 664
 665 bool
 666 nvc0_screen_resize_tls_area(struct nvc0_screen *screen,
 667                             uint32_t lpos, uint32_t lneg, uint32_t cstack)
 668 {
 669    struct nouveau_bo *bo = NULL;
 670    int ret;
 671    uint64_t size = (lpos + lneg) * 32 + cstack;
 672
 673    if (size >= (1 << 20)) {
 674       NOUVEAU_ERR("requested TLS size too large: 0x%"PRIx64"\n", size);
 675       return false;
 676    }
 677
 678    size *= (screen->base.device->chipset >= 0xe0) ? 64 : 48; /* max warps */
 679    size  = align(size, 0x8000);
 680    size *= screen->mp_count;
 681
 682    size = align(size, 1 << 17);
 683
 684    ret = nouveau_bo_new(screen->base.device, NV_VRAM_DOMAIN(&screen->base), 1 << 17, size,
 685                         NULL, &bo);
 686    if (ret) {
 687       NOUVEAU_ERR("failed to allocate TLS area, size: 0x%"PRIx64"\n", size);
 688       return false;
 689    }
 690    nouveau_bo_ref(NULL, &screen->tls);
 691    screen->tls = bo;
 692    return true;
 693 }
 694
 695 #define FAIL_SCREEN_INIT(str, err)                    \
 696    do {                                               \
 697       NOUVEAU_ERR(str, err);                          \
 698       goto fail;                                      \
 699    } while(0)
 700
 701 struct nouveau_screen *
 702 nvc0_screen_create(struct nouveau_device *dev)
 703 {
 704    struct nvc0_screen *screen;
 705    struct pipe_screen *pscreen;
 706    struct nouveau_object *chan;
 707    struct nouveau_pushbuf *push;
 708    uint64_t value;
 709    uint32_t obj_class;
 710    uint32_t flags;
 711    int ret;
 712    unsigned i;
 713
 714    switch (dev->chipset & ~0xf) {
 715    case 0xc0:
 716    case 0xd0:
 717    case 0xe0:
 718    case 0xf0:
 719    case 0x100:
 720    case 0x110:
 721    case 0x120:
 722       break;
 723    default:
 724       return NULL;
 725    }
 726
 727    screen = CALLOC_STRUCT(nvc0_screen);
 728    if (!screen)
 729       return NULL;
 730    pscreen = &screen->base.base;
 731    pscreen->destroy = nvc0_screen_destroy;
 732
 733    ret = nouveau_screen_init(&screen->base, dev);
 734    if (ret) {
 735       nvc0_screen_destroy(pscreen);
 736       return NULL;
 737    }
 738    chan = screen->base.channel;
 739    push = screen->base.pushbuf;
 740    push->user_priv = screen;
 741    push->rsvd_kick = 5;
 742
 743    screen->base.vidmem_bindings |= PIPE_BIND_CONSTANT_BUFFER |
 744       PIPE_BIND_SHADER_BUFFER |
 745       PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER |
 746       PIPE_BIND_COMMAND_ARGS_BUFFER | PIPE_BIND_QUERY_BUFFER;
 747    screen->base.sysmem_bindings |=
 748       PIPE_BIND_VERTEX_BUFFER | PIPE_BIND_INDEX_BUFFER;
 749
 750    if (screen->base.vram_domain & NOUVEAU_BO_GART) {
 751       screen->base.sysmem_bindings |= screen->base.vidmem_bindings;
 752       screen->base.vidmem_bindings = 0;
 753    }
 754
 755    pscreen->context_create = nvc0_create;
 756    pscreen->is_format_supported = nvc0_screen_is_format_supported;
 757    pscreen->get_param = nvc0_screen_get_param;
 758    pscreen->get_shader_param = nvc0_screen_get_shader_param;
 759    pscreen->get_paramf = nvc0_screen_get_paramf;
 760    pscreen->get_driver_query_info = nvc0_screen_get_driver_query_info;
 761    pscreen->get_driver_query_group_info = nvc0_screen_get_driver_query_group_info;
 762
 763    nvc0_screen_init_resource_functions(pscreen);
 764
 765    screen->base.base.get_video_param = nouveau_vp3_screen_get_video_param;
 766    screen->base.base.is_video_format_supported = nouveau_vp3_screen_video_supported;
 767
 768    flags = NOUVEAU_BO_GART | NOUVEAU_BO_MAP;
 769    if (screen->base.drm->version >= 0x01000202)
 770       flags |= NOUVEAU_BO_COHERENT;
 771
 772    ret = nouveau_bo_new(dev, flags, 0, 4096, NULL, &screen->fence.bo);
 773    if (ret)
 774       goto fail;
 775    nouveau_bo_map(screen->fence.bo, 0, NULL);
 776    screen->fence.map = screen->fence.bo->map;
 777    screen->base.fence.emit = nvc0_screen_fence_emit;
 778    screen->base.fence.update = nvc0_screen_fence_update;
 779
 780
 781    ret = nouveau_object_new(chan, (dev->chipset < 0xe0) ? 0x1f906e : 0x906e,
 782                             NVIF_CLASS_SW_GF100, NULL, 0, &screen->nvsw);
 783    if (ret)
 784       FAIL_SCREEN_INIT("Error creating SW object: %d\n", ret);
 785
 786    BEGIN_NVC0(push, SUBC_SW(NV01_SUBCHAN_OBJECT), 1);
 787    PUSH_DATA (push, screen->nvsw->handle);
 788
 789    switch (dev->chipset & ~0xf) {
 790    case 0x120:
 791    case 0x110:
 792    case 0x100:
 793    case 0xf0:
 794       obj_class = NVF0_P2MF_CLASS;
 795       break;
 796    case 0xe0:
 797       obj_class = NVE4_P2MF_CLASS;
 798       break;
 799    default:
 800       obj_class = NVC0_M2MF_CLASS;
 801       break;
 802    }
 803    ret = nouveau_object_new(chan, 0xbeef323f, obj_class, NULL, 0,
 804                             &screen->m2mf);
 805    if (ret)
 806       FAIL_SCREEN_INIT("Error allocating PGRAPH context for M2MF: %d\n", ret);
 807
 808    BEGIN_NVC0(push, SUBC_M2MF(NV01_SUBCHAN_OBJECT), 1);
 809    PUSH_DATA (push, screen->m2mf->oclass);
 810    if (screen->m2mf->oclass == NVE4_P2MF_CLASS) {
 811       BEGIN_NVC0(push, SUBC_COPY(NV01_SUBCHAN_OBJECT), 1);
 812       PUSH_DATA (push, 0xa0b5);
 813    }
 814
 815    ret = nouveau_object_new(chan, 0xbeef902d, NVC0_2D_CLASS, NULL, 0,
 816                             &screen->eng2d);
 817    if (ret)
 818       FAIL_SCREEN_INIT("Error allocating PGRAPH context for 2D: %d\n", ret);
 819
 820    BEGIN_NVC0(push, SUBC_2D(NV01_SUBCHAN_OBJECT), 1);
 821    PUSH_DATA (push, screen->eng2d->oclass);
 822    BEGIN_NVC0(push, SUBC_2D(NVC0_2D_SINGLE_GPC), 1);
 823    PUSH_DATA (push, 0);
 824    BEGIN_NVC0(push, NVC0_2D(OPERATION), 1);
 825    PUSH_DATA (push, NV50_2D_OPERATION_SRCCOPY);
 826    BEGIN_NVC0(push, NVC0_2D(CLIP_ENABLE), 1);
 827    PUSH_DATA (push, 0);
 828    BEGIN_NVC0(push, NVC0_2D(COLOR_KEY_ENABLE), 1);
 829    PUSH_DATA (push, 0);
 830    BEGIN_NVC0(push, SUBC_2D(0x0884), 1);
 831    PUSH_DATA (push, 0x3f);
 832    BEGIN_NVC0(push, SUBC_2D(0x0888), 1);
 833    PUSH_DATA (push, 1);
 834    BEGIN_NVC0(push, NVC0_2D(COND_MODE), 1);
 835    PUSH_DATA (push, NV50_2D_COND_MODE_ALWAYS);
 836
 837    BEGIN_NVC0(push, SUBC_2D(NVC0_GRAPH_NOTIFY_ADDRESS_HIGH), 2);
 838    PUSH_DATAh(push, screen->fence.bo->offset + 16);
 839    PUSH_DATA (push, screen->fence.bo->offset + 16);
 840
 841    switch (dev->chipset & ~0xf) {
 842    case 0x120:
 843       obj_class = GM200_3D_CLASS;
 844       break;
 845    case 0x110:
 846       obj_class = GM107_3D_CLASS;
 847       break;
 848    case 0x100:
 849    case 0xf0:
 850       obj_class = NVF0_3D_CLASS;
 851       break;
 852    case 0xe0:
 853       switch (dev->chipset) {
 854       case 0xea:
 855          obj_class = NVEA_3D_CLASS;
 856          break;
 857       default:
 858          obj_class = NVE4_3D_CLASS;
 859          break;
 860       }
 861       break;
 862    case 0xd0:
 863       obj_class = NVC8_3D_CLASS;
 864       break;
 865    case 0xc0:
 866    default:
 867       switch (dev->chipset) {
 868       case 0xc8:
 869          obj_class = NVC8_3D_CLASS;
 870          break;
 871       case 0xc1:
 872          obj_class = NVC1_3D_CLASS;
 873          break;
 874       default:
 875          obj_class = NVC0_3D_CLASS;
 876          break;
 877       }
 878       break;
 879    }
 880    ret = nouveau_object_new(chan, 0xbeef003d, obj_class, NULL, 0,
 881                             &screen->eng3d);
 882    if (ret)
 883       FAIL_SCREEN_INIT("Error allocating PGRAPH context for 3D: %d\n", ret);
 884    screen->base.class_3d = obj_class;
 885
 886    BEGIN_NVC0(push, SUBC_3D(NV01_SUBCHAN_OBJECT), 1);
 887    PUSH_DATA (push, screen->eng3d->oclass);
 888
 889    BEGIN_NVC0(push, NVC0_3D(COND_MODE), 1);
 890    PUSH_DATA (push, NVC0_3D_COND_MODE_ALWAYS);
 891
 892    if (debug_get_bool_option("NOUVEAU_SHADER_WATCHDOG", true)) {
 893       /* kill shaders after about 1 second (at 100 MHz) */
 894       BEGIN_NVC0(push, NVC0_3D(WATCHDOG_TIMER), 1);
 895       PUSH_DATA (push, 0x17);
 896    }
 897
 898    IMMED_NVC0(push, NVC0_3D(ZETA_COMP_ENABLE),
 899                     screen->base.drm->version >= 0x01000101);
 900    BEGIN_NVC0(push, NVC0_3D(RT_COMP_ENABLE(0)), 8);
 901    for (i = 0; i < 8; ++i)
 902            PUSH_DATA(push, screen->base.drm->version >= 0x01000101);
 903
 904    BEGIN_NVC0(push, NVC0_3D(RT_CONTROL), 1);
 905    PUSH_DATA (push, 1);
 906
 907    BEGIN_NVC0(push, NVC0_3D(CSAA_ENABLE), 1);
 908    PUSH_DATA (push, 0);
 909    BEGIN_NVC0(push, NVC0_3D(MULTISAMPLE_ENABLE), 1);
 910    PUSH_DATA (push, 0);
 911    BEGIN_NVC0(push, NVC0_3D(MULTISAMPLE_MODE), 1);
 912    PUSH_DATA (push, NVC0_3D_MULTISAMPLE_MODE_MS1);
 913    BEGIN_NVC0(push, NVC0_3D(MULTISAMPLE_CTRL), 1);
 914    PUSH_DATA (push, 0);
 915    BEGIN_NVC0(push, NVC0_3D(LINE_WIDTH_SEPARATE), 1);
 916    PUSH_DATA (push, 1);
 917    BEGIN_NVC0(push, NVC0_3D(PRIM_RESTART_WITH_DRAW_ARRAYS), 1);
 918    PUSH_DATA (push, 1);
 919    BEGIN_NVC0(push, NVC0_3D(BLEND_SEPARATE_ALPHA), 1);
 920    PUSH_DATA (push, 1);
 921    BEGIN_NVC0(push, NVC0_3D(BLEND_ENABLE_COMMON), 1);
 922    PUSH_DATA (push, 0);
 923    BEGIN_NVC0(push, NVC0_3D(SHADE_MODEL), 1);
 924    PUSH_DATA (push, NVC0_3D_SHADE_MODEL_SMOOTH);
 925    if (screen->eng3d->oclass < NVE4_3D_CLASS) {
 926       IMMED_NVC0(push, NVC0_3D(TEX_MISC), 0);
 927    } else {
 928       BEGIN_NVC0(push, NVE4_3D(TEX_CB_INDEX), 1);
 929       PUSH_DATA (push, 15);
 930    }
 931    BEGIN_NVC0(push, NVC0_3D(CALL_LIMIT_LOG), 1);
 932    PUSH_DATA (push, 8); /* 128 */
 933    BEGIN_NVC0(push, NVC0_3D(ZCULL_STATCTRS_ENABLE), 1);
 934    PUSH_DATA (push, 1);
 935    if (screen->eng3d->oclass >= NVC1_3D_CLASS) {
 936       BEGIN_NVC0(push, NVC0_3D(CACHE_SPLIT), 1);
 937       PUSH_DATA (push, NVC0_3D_CACHE_SPLIT_48K_SHARED_16K_L1);
 938    }
 939
 940    nvc0_magic_3d_init(push, screen->eng3d->oclass);
 941
 942    ret = nouveau_bo_new(dev, NV_VRAM_DOMAIN(&screen->base), 1 << 17, 1 << 20, NULL,
 943                         &screen->text);
 944    if (ret)
 945       goto fail;
 946
 947    /* XXX: getting a page fault at the end of the code buffer every few
 948     *  launches, don't use the last 256 bytes to work around them - prefetch ?
 949     */
 950    nouveau_heap_init(&screen->text_heap, 0, (1 << 20) - 0x100);
 951
 952    ret = nouveau_bo_new(dev, NV_VRAM_DOMAIN(&screen->base), 1 << 12, 7 << 16, NULL,
 953                         &screen->uniform_bo);
 954    if (ret)
 955       goto fail;
 956
 957    PUSH_REFN (push, screen->uniform_bo, NV_VRAM_DOMAIN(&screen->base) | NOUVEAU_BO_WR);
 958
 959    for (i = 0; i < 5; ++i) {
 960       /* TIC and TSC entries for each unit (nve4+ only) */
 961       /* auxiliary constants (6 user clip planes, base instance id) */
 962       BEGIN_NVC0(push, NVC0_3D(CB_SIZE), 3);
 963       PUSH_DATA (push, 2048);
 964       PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(i));
 965       PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_AUX_INFO(i));
 966       BEGIN_NVC0(push, NVC0_3D(CB_BIND(i)), 1);
 967       PUSH_DATA (push, (15 << 4) | 1);
 968       if (screen->eng3d->oclass >= NVE4_3D_CLASS) {
 969          unsigned j;
 970          BEGIN_1IC0(push, NVC0_3D(CB_POS), 9);
 971          PUSH_DATA (push, NVC0_CB_AUX_UNK_INFO);
 972          for (j = 0; j < 8; ++j)
 973             PUSH_DATA(push, j);
 974       } else {
 975          BEGIN_NVC0(push, NVC0_3D(TEX_LIMITS(i)), 1);
 976          PUSH_DATA (push, 0x54);
 977       }
 978    }
 979    BEGIN_NVC0(push, NVC0_3D(LINKED_TSC), 1);
 980    PUSH_DATA (push, 0);
 981
 982    /* return { 0.0, 0.0, 0.0, 0.0 } for out-of-bounds vtxbuf access */
 983    BEGIN_NVC0(push, NVC0_3D(CB_SIZE), 3);
 984    PUSH_DATA (push, 256);
 985    PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_AUX_RUNOUT_INFO);
 986    PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_AUX_RUNOUT_INFO);
 987    BEGIN_1IC0(push, NVC0_3D(CB_POS), 5);
 988    PUSH_DATA (push, 0);
 989    PUSH_DATAf(push, 0.0f);
 990    PUSH_DATAf(push, 0.0f);
 991    PUSH_DATAf(push, 0.0f);
 992    PUSH_DATAf(push, 0.0f);
 993    BEGIN_NVC0(push, NVC0_3D(VERTEX_RUNOUT_ADDRESS_HIGH), 2);
 994    PUSH_DATAh(push, screen->uniform_bo->offset + NVC0_CB_AUX_RUNOUT_INFO);
 995    PUSH_DATA (push, screen->uniform_bo->offset + NVC0_CB_AUX_RUNOUT_INFO);
 996
 997    if (screen->base.drm->version >= 0x01000101) {
 998       ret = nouveau_getparam(dev, NOUVEAU_GETPARAM_GRAPH_UNITS, &value);
 999       if (ret) {
1000          NOUVEAU_ERR("NOUVEAU_GETPARAM_GRAPH_UNITS failed.\n");
1001          goto fail;
1002       }
1003    } else {
1004       if (dev->chipset >= 0xe0 && dev->chipset < 0xf0)
1005          value = (8 << 8) | 4;
1006       else
1007          value = (16 << 8) | 4;
1008    }
1009    screen->gpc_count = value & 0x000000ff;
1010    screen->mp_count = value >> 8;
1011    screen->mp_count_compute = screen->mp_count;
1012
1013    nvc0_screen_resize_tls_area(screen, 128 * 16, 0, 0x200);
1014
1015    BEGIN_NVC0(push, NVC0_3D(CODE_ADDRESS_HIGH), 2);
1016    PUSH_DATAh(push, screen->text->offset);
1017    PUSH_DATA (push, screen->text->offset);
1018    BEGIN_NVC0(push, NVC0_3D(TEMP_ADDRESS_HIGH), 4);
1019    PUSH_DATAh(push, screen->tls->offset);
1020    PUSH_DATA (push, screen->tls->offset);
1021    PUSH_DATA (push, screen->tls->size >> 32);
1022    PUSH_DATA (push, screen->tls->size);
1023    BEGIN_NVC0(push, NVC0_3D(WARP_TEMP_ALLOC), 1);
1024    PUSH_DATA (push, 0);
1025    /* Reduce likelihood of collision with real buffers by placing the hole at
1026     * the top of the 4G area. This will have to be dealt with for real
1027     * eventually by blocking off that area from the VM.
1028     */
1029    BEGIN_NVC0(push, NVC0_3D(LOCAL_BASE), 1);
1030    PUSH_DATA (push, 0xff << 24);
1031
1032    if (screen->eng3d->oclass < GM107_3D_CLASS) {
1033       ret = nouveau_bo_new(dev, NV_VRAM_DOMAIN(&screen->base), 1 << 17, 1 << 20, NULL,
1034                            &screen->poly_cache);
1035       if (ret)
1036          goto fail;
1037
1038       BEGIN_NVC0(push, NVC0_3D(VERTEX_QUARANTINE_ADDRESS_HIGH), 3);
1039       PUSH_DATAh(push, screen->poly_cache->offset);
1040       PUSH_DATA (push, screen->poly_cache->offset);
1041       PUSH_DATA (push, 3);
1042    }
1043
1044    ret = nouveau_bo_new(dev, NV_VRAM_DOMAIN(&screen->base), 1 << 17, 1 << 17, NULL,
1045                         &screen->txc);
1046    if (ret)
1047       goto fail;
1048
1049    BEGIN_NVC0(push, NVC0_3D(TIC_ADDRESS_HIGH), 3);
1050    PUSH_DATAh(push, screen->txc->offset);
1051    PUSH_DATA (push, screen->txc->offset);
1052    PUSH_DATA (push, NVC0_TIC_MAX_ENTRIES - 1);
1053    if (screen->eng3d->oclass >= GM107_3D_CLASS) {
1054       screen->tic.maxwell = true;
1055       if (screen->eng3d->oclass == GM107_3D_CLASS) {
1056          screen->tic.maxwell =
1057             debug_get_bool_option("NOUVEAU_MAXWELL_TIC", true);
1058          IMMED_NVC0(push, SUBC_3D(0x0f10), screen->tic.maxwell);
1059       }
1060    }
1061
1062    BEGIN_NVC0(push, NVC0_3D(TSC_ADDRESS_HIGH), 3);
1063    PUSH_DATAh(push, screen->txc->offset + 65536);
1064    PUSH_DATA (push, screen->txc->offset + 65536);
1065    PUSH_DATA (push, NVC0_TSC_MAX_ENTRIES - 1);
1066
1067    BEGIN_NVC0(push, NVC0_3D(SCREEN_Y_CONTROL), 1);
1068    PUSH_DATA (push, 0);
1069    BEGIN_NVC0(push, NVC0_3D(WINDOW_OFFSET_X), 2);
1070    PUSH_DATA (push, 0);
1071    PUSH_DATA (push, 0);
1072    BEGIN_NVC0(push, NVC0_3D(ZCULL_REGION), 1); /* deactivate ZCULL */
1073    PUSH_DATA (push, 0x3f);
1074
1075    BEGIN_NVC0(push, NVC0_3D(CLIP_RECTS_MODE), 1);
1076    PUSH_DATA (push, NVC0_3D_CLIP_RECTS_MODE_INSIDE_ANY);
1077    BEGIN_NVC0(push, NVC0_3D(CLIP_RECT_HORIZ(0)), 8 * 2);
1078    for (i = 0; i < 8 * 2; ++i)
1079       PUSH_DATA(push, 0);
1080    BEGIN_NVC0(push, NVC0_3D(CLIP_RECTS_EN), 1);
1081    PUSH_DATA (push, 0);
1082    BEGIN_NVC0(push, NVC0_3D(CLIPID_ENABLE), 1);
1083    PUSH_DATA (push, 0);
1084
1085    /* neither scissors, viewport nor stencil mask should affect clears */
1086    BEGIN_NVC0(push, NVC0_3D(CLEAR_FLAGS), 1);
1087    PUSH_DATA (push, 0);
1088
1089    BEGIN_NVC0(push, NVC0_3D(VIEWPORT_TRANSFORM_EN), 1);
1090    PUSH_DATA (push, 1);
1091    for (i = 0; i < NVC0_MAX_VIEWPORTS; i++) {
1092       BEGIN_NVC0(push, NVC0_3D(DEPTH_RANGE_NEAR(i)), 2);
1093       PUSH_DATAf(push, 0.0f);
1094       PUSH_DATAf(push, 1.0f);
1095    }
1096    BEGIN_NVC0(push, NVC0_3D(VIEW_VOLUME_CLIP_CTRL), 1);
1097    PUSH_DATA (push, NVC0_3D_VIEW_VOLUME_CLIP_CTRL_UNK1_UNK1);
1098
1099    /* We use scissors instead of exact view volume clipping,
1100     * so they're always enabled.
1101     */
1102    for (i = 0; i < NVC0_MAX_VIEWPORTS; i++) {
1103       BEGIN_NVC0(push, NVC0_3D(SCISSOR_ENABLE(i)), 3);
1104       PUSH_DATA (push, 1);
1105       PUSH_DATA (push, 8192 << 16);
1106       PUSH_DATA (push, 8192 << 16);
1107    }
1108
1109 #define MK_MACRO(m, n) i = nvc0_graph_set_macro(screen, m, i, sizeof(n), n);
1110
1111    i = 0;
1112    MK_MACRO(NVC0_3D_MACRO_VERTEX_ARRAY_PER_INSTANCE, mme9097_per_instance_bf);
1113    MK_MACRO(NVC0_3D_MACRO_BLEND_ENABLES, mme9097_blend_enables);
1114    MK_MACRO(NVC0_3D_MACRO_VERTEX_ARRAY_SELECT, mme9097_vertex_array_select);
1115    MK_MACRO(NVC0_3D_MACRO_TEP_SELECT, mme9097_tep_select);
1116    MK_MACRO(NVC0_3D_MACRO_GP_SELECT, mme9097_gp_select);
1117    MK_MACRO(NVC0_3D_MACRO_POLYGON_MODE_FRONT, mme9097_poly_mode_front);
1118    MK_MACRO(NVC0_3D_MACRO_POLYGON_MODE_BACK, mme9097_poly_mode_back);
1119    MK_MACRO(NVC0_3D_MACRO_DRAW_ARRAYS_INDIRECT, mme9097_draw_arrays_indirect);
1120    MK_MACRO(NVC0_3D_MACRO_DRAW_ELEMENTS_INDIRECT, mme9097_draw_elts_indirect);
1121    MK_MACRO(NVC0_3D_MACRO_DRAW_ARRAYS_INDIRECT_COUNT, mme9097_draw_arrays_indirect_count);
1122    MK_MACRO(NVC0_3D_MACRO_DRAW_ELEMENTS_INDIRECT_COUNT, mme9097_draw_elts_indirect_count);
1123    MK_MACRO(NVC0_3D_MACRO_QUERY_BUFFER_WRITE, mme9097_query_buffer_write);
1124    MK_MACRO(NVC0_CP_MACRO_LAUNCH_GRID_INDIRECT, mme90c0_launch_grid_indirect);
1125
1126    BEGIN_NVC0(push, NVC0_3D(RASTERIZE_ENABLE), 1);
1127    PUSH_DATA (push, 1);
1128    BEGIN_NVC0(push, NVC0_3D(RT_SEPARATE_FRAG_DATA), 1);
1129    PUSH_DATA (push, 1);
1130    BEGIN_NVC0(push, NVC0_3D(MACRO_GP_SELECT), 1);
1131    PUSH_DATA (push, 0x40);
1132    BEGIN_NVC0(push, NVC0_3D(LAYER), 1);
1133    PUSH_DATA (push, 0);
1134    BEGIN_NVC0(push, NVC0_3D(MACRO_TEP_SELECT), 1);
1135    PUSH_DATA (push, 0x30);
1136    BEGIN_NVC0(push, NVC0_3D(PATCH_VERTICES), 1);
1137    PUSH_DATA (push, 3);
1138    BEGIN_NVC0(push, NVC0_3D(SP_SELECT(2)), 1);
1139    PUSH_DATA (push, 0x20);
1140    BEGIN_NVC0(push, NVC0_3D(SP_SELECT(0)), 1);
1141    PUSH_DATA (push, 0x00);
1142    screen->save_state.patch_vertices = 3;
1143
1144    BEGIN_NVC0(push, NVC0_3D(POINT_COORD_REPLACE), 1);
1145    PUSH_DATA (push, 0);
1146    BEGIN_NVC0(push, NVC0_3D(POINT_RASTER_RULES), 1);
1147    PUSH_DATA (push, NVC0_3D_POINT_RASTER_RULES_OGL);
1148
1149    IMMED_NVC0(push, NVC0_3D(EDGEFLAG), 1);
1150
1151    if (nvc0_screen_init_compute(screen))
1152       goto fail;
1153
1154    PUSH_KICK (push);
1155
1156    screen->tic.entries = CALLOC(4096, sizeof(void *));
1157    screen->tsc.entries = screen->tic.entries + 2048;
1158
1159    if (!nvc0_blitter_create(screen))
1160       goto fail;
1161
1162    nouveau_fence_new(&screen->base, &screen->base.fence.current, false);
1163
1164    return &screen->base;
1165
1166 fail:
1167    screen->base.base.context_create = NULL;
1168    return &screen->base;
1169 }
1170
1171 int
1172 nvc0_screen_tic_alloc(struct nvc0_screen *screen, void *entry)
1173 {
1174    int i = screen->tic.next;
1175
1176    while (screen->tic.lock[i / 32] & (1 << (i % 32)))
1177       i = (i + 1) & (NVC0_TIC_MAX_ENTRIES - 1);
1178
1179    screen->tic.next = (i + 1) & (NVC0_TIC_MAX_ENTRIES - 1);
1180
1181    if (screen->tic.entries[i])
1182       nv50_tic_entry(screen->tic.entries[i])->id = -1;
1183
1184    screen->tic.entries[i] = entry;
1185    return i;
1186 }
1187
1188 int
1189 nvc0_screen_tsc_alloc(struct nvc0_screen *screen, void *entry)
1190 {
1191    int i = screen->tsc.next;
1192
1193    while (screen->tsc.lock[i / 32] & (1 << (i % 32)))
1194       i = (i + 1) & (NVC0_TSC_MAX_ENTRIES - 1);
1195
1196    screen->tsc.next = (i + 1) & (NVC0_TSC_MAX_ENTRIES - 1);
1197
1198    if (screen->tsc.entries[i])
1199       nv50_tsc_entry(screen->tsc.entries[i])->id = -1;
1200
1201    screen->tsc.entries[i] = entry;
1202    return i;
1203 }