src/gallium/auxiliary/vl/vl_zscan.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2011 Christian König
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include <assert.h>
  29
  30 #include <pipe/p_screen.h>
  31 #include <pipe/p_context.h>
  32
  33 #include <util/u_draw.h>
  34 #include <util/u_sampler.h>
  35 #include <util/u_inlines.h>
  36
  37 #include <tgsi/tgsi_ureg.h>
  38
  39 #include <vl/vl_defines.h>
  40 #include <vl/vl_types.h>
  41
  42 #include "vl_zscan.h"
  43 #include "vl_vertex_buffers.h"
  44
  45 enum VS_OUTPUT
  46 {
  47    VS_O_VPOS,
  48    VS_O_VTEX
  49 };
  50
  51 const int vl_zscan_linear[] =
  52 {
  53    /* Linear scan pattern */
  54     0, 1, 2, 3, 4, 5, 6, 7,
  55     8, 9,10,11,12,13,14,15,
  56    16,17,18,19,20,21,22,23,
  57    24,25,26,27,28,29,30,31,
  58    32,33,34,35,36,37,38,39,
  59    40,41,42,43,44,45,46,47,
  60    48,49,50,51,52,53,54,55,
  61    56,57,58,59,60,61,62,63
  62 };
  63
  64 const int vl_zscan_normal[] =
  65 {
  66    /* Zig-Zag scan pattern */
  67     0, 1, 8,16, 9, 2, 3,10,
  68    17,24,32,25,18,11, 4, 5,
  69    12,19,26,33,40,48,41,34,
  70    27,20,13, 6, 7,14,21,28,
  71    35,42,49,56,57,50,43,36,
  72    29,22,15,23,30,37,44,51,
  73    58,59,52,45,38,31,39,46,
  74    53,60,61,54,47,55,62,63
  75 };
  76
  77 const int vl_zscan_alternate[] =
  78 {
  79    /* Alternate scan pattern */
  80     0, 8,16,24, 1, 9, 2,10,
  81    17,25,32,40,48,56,57,49,
  82    41,33,26,18, 3,11, 4,12,
  83    19,27,34,42,50,58,35,43,
  84    51,59,20,28, 5,13, 6,14,
  85    21,29,36,44,52,60,37,45,
  86    53,61,22,30, 7,15,23,31,
  87    38,46,54,62,39,47,55,63
  88 };
  89
  90 static void *
  91 create_vert_shader(struct vl_zscan *zscan)
  92 {
  93    struct ureg_program *shader;
  94
  95    struct ureg_src scale, instance;
  96    struct ureg_src vrect, vpos;
  97
  98    struct ureg_dst tmp;
  99    struct ureg_dst o_vpos, o_vtex[zscan->num_channels];
 100
 101    signed i;
 102
 103    shader = ureg_create(TGSI_PROCESSOR_VERTEX);
 104    if (!shader)
 105       return NULL;
 106
 107    scale = ureg_imm2f(shader,
 108       (float)BLOCK_WIDTH / zscan->buffer_width,
 109       (float)BLOCK_HEIGHT / zscan->buffer_height);
 110
 111    instance = ureg_DECL_system_value(shader, 0, TGSI_SEMANTIC_INSTANCEID, 0);
 112
 113    vrect = ureg_DECL_vs_input(shader, VS_I_RECT);
 114    vpos = ureg_DECL_vs_input(shader, VS_I_VPOS);
 115
 116    tmp = ureg_DECL_temporary(shader);
 117
 118    o_vpos = ureg_DECL_output(shader, TGSI_SEMANTIC_POSITION, VS_O_VPOS);
 119
 120    for (i = 0; i < zscan->num_channels; ++i)
 121       o_vtex[i] = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX + i);
 122
 123    /*
 124     * o_vpos.xy = (vpos + vrect) * scale
 125     * o_vpos.zw = 1.0f
 126     *
 127     * tmp.xy = InstanceID / blocks_per_line
 128     * tmp.x = frac(tmp.x)
 129     * tmp.y = floor(tmp.y)
 130     *
 131     * o_vtex.x = vrect.x / blocks_per_line + tmp.x
 132     * o_vtex.y = vrect.y
 133     * o_vtex.z = tmp.z * blocks_per_line / blocks_total
 134     */
 135    ureg_ADD(shader, ureg_writemask(tmp, TGSI_WRITEMASK_XY), vpos, vrect);
 136    ureg_MUL(shader, ureg_writemask(o_vpos, TGSI_WRITEMASK_XY), ureg_src(tmp), scale);
 137    ureg_MOV(shader, ureg_writemask(o_vpos, TGSI_WRITEMASK_ZW), ureg_imm1f(shader, 1.0f));
 138
 139    ureg_MUL(shader, ureg_writemask(tmp, TGSI_WRITEMASK_XZ), ureg_scalar(instance, TGSI_SWIZZLE_X),
 140             ureg_imm1f(shader, 1.0f / zscan->blocks_per_line));
 141
 142    ureg_FRC(shader, ureg_writemask(tmp, TGSI_WRITEMASK_Y), ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_X));
 143    ureg_FLR(shader, ureg_writemask(tmp, TGSI_WRITEMASK_Z), ureg_src(tmp));
 144
 145    for (i = 0; i < zscan->num_channels; ++i) {
 146       ureg_ADD(shader, ureg_writemask(tmp, TGSI_WRITEMASK_X), ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_Y),
 147                ureg_imm1f(shader, 1.0f / (zscan->blocks_per_line * BLOCK_WIDTH) * (i - (signed)zscan->num_channels / 2)));
 148
 149       ureg_MAD(shader, ureg_writemask(o_vtex[i], TGSI_WRITEMASK_X), vrect,
 150                ureg_imm1f(shader, 1.0f / zscan->blocks_per_line), ureg_src(tmp));
 151       ureg_MOV(shader, ureg_writemask(o_vtex[i], TGSI_WRITEMASK_Y), vrect);
 152       ureg_MUL(shader, ureg_writemask(o_vtex[i], TGSI_WRITEMASK_Z), ureg_src(tmp),
 153                ureg_imm1f(shader, (float)zscan->blocks_per_line / zscan->blocks_total));
 154    }
 155
 156    ureg_release_temporary(shader, tmp);
 157    ureg_END(shader);
 158
 159    return ureg_create_shader_and_destroy(shader, zscan->pipe);
 160 }
 161
 162 static void *
 163 create_frag_shader(struct vl_zscan *zscan)
 164 {
 165    struct ureg_program *shader;
 166    struct ureg_src vtex[zscan->num_channels];
 167
 168    struct ureg_src src, scan, quant;
 169
 170    struct ureg_dst tmp[zscan->num_channels];
 171    struct ureg_dst fragment;
 172
 173    unsigned i;
 174
 175    shader = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 176    if (!shader)
 177       return NULL;
 178
 179    for (i = 0; i < zscan->num_channels; ++i)
 180       vtex[i] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX + i, TGSI_INTERPOLATE_LINEAR);
 181
 182    src = ureg_DECL_sampler(shader, 0);
 183    scan = ureg_DECL_sampler(shader, 1);
 184    quant = ureg_DECL_sampler(shader, 2);
 185
 186    for (i = 0; i < zscan->num_channels; ++i)
 187       tmp[i] = ureg_DECL_temporary(shader);
 188
 189    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
 190
 191    /*
 192     * tmp.x = tex(vtex, 1)
 193     * tmp.y = vtex.z
 194     * fragment = tex(tmp, 0) * quant
 195     */
 196    for (i = 0; i < zscan->num_channels; ++i)
 197       ureg_TEX(shader, ureg_writemask(tmp[i], TGSI_WRITEMASK_X), TGSI_TEXTURE_2D, vtex[i], scan);
 198
 199    for (i = 0; i < zscan->num_channels; ++i)
 200       ureg_MOV(shader, ureg_writemask(tmp[i], TGSI_WRITEMASK_Y), ureg_scalar(vtex[i], TGSI_SWIZZLE_Z));
 201
 202    for (i = 0; i < zscan->num_channels; ++i)
 203       ureg_TEX(shader, tmp[i], TGSI_TEXTURE_2D, ureg_src(tmp[i]), src);
 204
 205    // TODO: Fetch quant and use it
 206    for (i = 0; i < zscan->num_channels; ++i)
 207       ureg_MUL(shader, ureg_writemask(fragment, TGSI_WRITEMASK_X << i), ureg_src(tmp[i]), ureg_imm1f(shader, 1.0f));
 208
 209    for (i = 0; i < zscan->num_channels; ++i)
 210       ureg_release_temporary(shader, tmp[i]);
 211    ureg_END(shader);
 212
 213    return ureg_create_shader_and_destroy(shader, zscan->pipe);
 214 }
 215
 216 static bool
 217 init_shaders(struct vl_zscan *zscan)
 218 {
 219    assert(zscan);
 220
 221    zscan->vs = create_vert_shader(zscan);
 222    if (!zscan->vs)
 223       goto error_vs;
 224
 225    zscan->fs = create_frag_shader(zscan);
 226    if (!zscan->fs)
 227       goto error_fs;
 228
 229    return true;
 230
 231 error_fs:
 232    zscan->pipe->delete_vs_state(zscan->pipe, zscan->vs);
 233
 234 error_vs:
 235    return false;
 236 }
 237
 238 static void
 239 cleanup_shaders(struct vl_zscan *zscan)
 240 {
 241    assert(zscan);
 242
 243    zscan->pipe->delete_vs_state(zscan->pipe, zscan->vs);
 244    zscan->pipe->delete_fs_state(zscan->pipe, zscan->fs);
 245 }
 246
 247 static bool
 248 init_state(struct vl_zscan *zscan)
 249 {
 250    struct pipe_blend_state blend;
 251    struct pipe_rasterizer_state rs_state;
 252    struct pipe_sampler_state sampler;
 253    unsigned i;
 254
 255    assert(zscan);
 256
 257    memset(&rs_state, 0, sizeof(rs_state));
 258    rs_state.gl_rasterization_rules = true;
 259    zscan->rs_state = zscan->pipe->create_rasterizer_state(zscan->pipe, &rs_state);
 260    if (!zscan->rs_state)
 261       goto error_rs_state;
 262
 263    memset(&blend, 0, sizeof blend);
 264
 265    blend.independent_blend_enable = 0;
 266    blend.rt[0].blend_enable = 0;
 267    blend.rt[0].rgb_func = PIPE_BLEND_ADD;
 268    blend.rt[0].rgb_src_factor = PIPE_BLENDFACTOR_ONE;
 269    blend.rt[0].rgb_dst_factor = PIPE_BLENDFACTOR_ONE;
 270    blend.rt[0].alpha_func = PIPE_BLEND_ADD;
 271    blend.rt[0].alpha_src_factor = PIPE_BLENDFACTOR_ONE;
 272    blend.rt[0].alpha_dst_factor = PIPE_BLENDFACTOR_ONE;
 273    blend.logicop_enable = 0;
 274    blend.logicop_func = PIPE_LOGICOP_CLEAR;
 275    /* Needed to allow color writes to FB, even if blending disabled */
 276    blend.rt[0].colormask = PIPE_MASK_RGBA;
 277    blend.dither = 0;
 278    zscan->blend = zscan->pipe->create_blend_state(zscan->pipe, &blend);
 279    if (!zscan->blend)
 280       goto error_blend;
 281
 282    for (i = 0; i < 3; ++i) {
 283       memset(&sampler, 0, sizeof(sampler));
 284       sampler.wrap_s = PIPE_TEX_WRAP_REPEAT;
 285       sampler.wrap_t = PIPE_TEX_WRAP_REPEAT;
 286       sampler.wrap_r = PIPE_TEX_WRAP_REPEAT;
 287       sampler.min_img_filter = PIPE_TEX_FILTER_NEAREST;
 288       sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
 289       sampler.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
 290       sampler.compare_mode = PIPE_TEX_COMPARE_NONE;
 291       sampler.compare_func = PIPE_FUNC_ALWAYS;
 292       sampler.normalized_coords = 1;
 293       zscan->samplers[i] = zscan->pipe->create_sampler_state(zscan->pipe, &sampler);
 294       if (!zscan->samplers[i])
 295          goto error_samplers;
 296    }
 297
 298    return true;
 299
 300 error_samplers:
 301    for (i = 0; i < 2; ++i)
 302       if (zscan->samplers[i])
 303          zscan->pipe->delete_sampler_state(zscan->pipe, zscan->samplers[i]);
 304
 305    zscan->pipe->delete_rasterizer_state(zscan->pipe, zscan->rs_state);
 306
 307 error_blend:
 308    zscan->pipe->delete_blend_state(zscan->pipe, zscan->blend);
 309
 310 error_rs_state:
 311    return false;
 312 }
 313
 314 static void
 315 cleanup_state(struct vl_zscan *zscan)
 316 {
 317    unsigned i;
 318
 319    assert(zscan);
 320
 321    for (i = 0; i < 3; ++i)
 322       zscan->pipe->delete_sampler_state(zscan->pipe, zscan->samplers[i]);
 323
 324    zscan->pipe->delete_rasterizer_state(zscan->pipe, zscan->rs_state);
 325    zscan->pipe->delete_blend_state(zscan->pipe, zscan->blend);
 326 }
 327
 328 struct pipe_sampler_view *
 329 vl_zscan_layout(struct pipe_context *pipe, const int layout[64], unsigned blocks_per_line)
 330 {
 331    const unsigned total_size = blocks_per_line * BLOCK_WIDTH * BLOCK_HEIGHT;
 332
 333    int patched_layout[64];
 334
 335    struct pipe_resource res_tmpl, *res;
 336    struct pipe_sampler_view sv_tmpl, *sv;
 337    struct pipe_transfer *buf_transfer;
 338    unsigned x, y, i, pitch;
 339    float *f;
 340
 341    struct pipe_box rect =
 342    {
 343       0, 0, 0,
 344       BLOCK_WIDTH * blocks_per_line,
 345       BLOCK_HEIGHT,
 346       1
 347    };
 348
 349    assert(pipe && layout && blocks_per_line);
 350
 351    for (i = 0; i < 64; ++i)
 352       patched_layout[layout[i]] = i;
 353
 354    memset(&res_tmpl, 0, sizeof(res_tmpl));
 355    res_tmpl.target = PIPE_TEXTURE_2D;
 356    res_tmpl.format = PIPE_FORMAT_R32_FLOAT;
 357    res_tmpl.width0 = BLOCK_WIDTH * blocks_per_line;
 358    res_tmpl.height0 = BLOCK_HEIGHT;
 359    res_tmpl.depth0 = 1;
 360    res_tmpl.array_size = 1;
 361    res_tmpl.usage = PIPE_USAGE_IMMUTABLE;
 362    res_tmpl.bind = PIPE_BIND_SAMPLER_VIEW;
 363
 364    res = pipe->screen->resource_create(pipe->screen, &res_tmpl);
 365    if (!res)
 366       goto error_resource;
 367
 368    buf_transfer = pipe->get_transfer
 369    (
 370       pipe, res,
 371       0, PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD,
 372       &rect
 373    );
 374    if (!buf_transfer)
 375       goto error_transfer;
 376
 377    pitch = buf_transfer->stride / sizeof(float);
 378
 379    f = pipe->transfer_map(pipe, buf_transfer);
 380    if (!f)
 381       goto error_map;
 382
 383    for (i = 0; i < blocks_per_line; ++i)
 384       for (y = 0; y < BLOCK_HEIGHT; ++y)
 385          for (x = 0; x < BLOCK_WIDTH; ++x) {
 386             float addr = patched_layout[x + y * BLOCK_WIDTH] +
 387                i * BLOCK_WIDTH * BLOCK_HEIGHT;
 388
 389             addr /= total_size;
 390
 391             f[i * BLOCK_WIDTH + y * pitch + x] = addr;
 392          }
 393
 394    pipe->transfer_unmap(pipe, buf_transfer);
 395    pipe->transfer_destroy(pipe, buf_transfer);
 396
 397    memset(&sv_tmpl, 0, sizeof(sv_tmpl));
 398    u_sampler_view_default_template(&sv_tmpl, res, res->format);
 399    sv = pipe->create_sampler_view(pipe, res, &sv_tmpl);
 400    pipe_resource_reference(&res, NULL);
 401    if (!sv)
 402       goto error_map;
 403
 404    return sv;
 405
 406 error_map:
 407    pipe->transfer_destroy(pipe, buf_transfer);
 408
 409 error_transfer:
 410    pipe_resource_reference(&res, NULL);
 411
 412 error_resource:
 413    return NULL;
 414 }
 415
 416 #if 0
 417 // TODO
 418 struct pipe_sampler_view *
 419 vl_zscan_normal(struct pipe_context *pipe, unsigned blocks_per_line);
 420
 421 struct pipe_sampler_view *
 422 vl_zscan_alternate(struct pipe_context *pipe, unsigned blocks_per_line);
 423 #endif
 424
 425 bool
 426 vl_zscan_init(struct vl_zscan *zscan, struct pipe_context *pipe,
 427               unsigned buffer_width, unsigned buffer_height,
 428               unsigned blocks_per_line, unsigned blocks_total,
 429               unsigned num_channels)
 430 {
 431    assert(zscan && pipe);
 432
 433    zscan->pipe = pipe;
 434    zscan->buffer_width = buffer_width;
 435    zscan->buffer_height = buffer_height;
 436    zscan->num_channels = num_channels;
 437    zscan->blocks_per_line = blocks_per_line;
 438    zscan->blocks_total = blocks_total;
 439
 440    if(!init_shaders(zscan))
 441       return false;
 442
 443    if(!init_state(zscan)) {
 444       cleanup_shaders(zscan);
 445       return false;
 446    }
 447
 448    return true;
 449 }
 450
 451 void
 452 vl_zscan_cleanup(struct vl_zscan *zscan)
 453 {
 454    assert(zscan);
 455
 456    cleanup_shaders(zscan);
 457    cleanup_state(zscan);
 458 }
 459
 460 #if 0
 461 // TODO
 462 void
 463 vl_zscan_upload_quant(struct vl_zscan *zscan, ...);
 464 #endif
 465
 466 bool
 467 vl_zscan_init_buffer(struct vl_zscan *zscan, struct vl_zscan_buffer *buffer,
 468                      struct pipe_sampler_view *src, struct pipe_surface *dst)
 469 {
 470    assert(zscan && buffer);
 471
 472    memset(buffer, 0, sizeof(struct vl_zscan_buffer));
 473
 474    buffer->zscan = zscan;
 475
 476    pipe_sampler_view_reference(&buffer->src, src);
 477
 478    buffer->viewport.scale[0] = dst->width;
 479    buffer->viewport.scale[1] = dst->height;
 480    buffer->viewport.scale[2] = 1;
 481    buffer->viewport.scale[3] = 1;
 482    buffer->viewport.translate[0] = 0;
 483    buffer->viewport.translate[1] = 0;
 484    buffer->viewport.translate[2] = 0;
 485    buffer->viewport.translate[3] = 0;
 486
 487    buffer->fb_state.width = dst->width;
 488    buffer->fb_state.height = dst->height;
 489    buffer->fb_state.nr_cbufs = 1;
 490    pipe_surface_reference(&buffer->fb_state.cbufs[0], dst);
 491
 492    return true;
 493 }
 494
 495 void
 496 vl_zscan_cleanup_buffer(struct vl_zscan_buffer *buffer)
 497 {
 498    assert(buffer);
 499
 500    pipe_sampler_view_reference(&buffer->src, NULL);
 501    pipe_sampler_view_reference(&buffer->layout, NULL);
 502    pipe_sampler_view_reference(&buffer->quant, NULL);
 503    pipe_surface_reference(&buffer->fb_state.cbufs[0], NULL);
 504 }
 505
 506 void
 507 vl_zscan_set_layout(struct vl_zscan_buffer *buffer, struct pipe_sampler_view *layout)
 508 {
 509    assert(buffer);
 510    assert(layout);
 511
 512    pipe_sampler_view_reference(&buffer->layout, layout);
 513 }
 514
 515 void
 516 vl_zscan_render(struct vl_zscan_buffer *buffer, unsigned num_instances)
 517 {
 518    struct vl_zscan *zscan;
 519
 520    assert(buffer);
 521
 522    zscan = buffer->zscan;
 523
 524    zscan->pipe->bind_rasterizer_state(zscan->pipe, zscan->rs_state);
 525    zscan->pipe->bind_blend_state(zscan->pipe, zscan->blend);
 526    zscan->pipe->bind_fragment_sampler_states(zscan->pipe, 2, zscan->samplers);
 527    zscan->pipe->set_framebuffer_state(zscan->pipe, &buffer->fb_state);
 528    zscan->pipe->set_viewport_state(zscan->pipe, &buffer->viewport);
 529    zscan->pipe->set_fragment_sampler_views(zscan->pipe, 2, &buffer->src);
 530    zscan->pipe->bind_vs_state(zscan->pipe, zscan->vs);
 531    zscan->pipe->bind_fs_state(zscan->pipe, zscan->fs);
 532    util_draw_arrays_instanced(zscan->pipe, PIPE_PRIM_QUADS, 0, 4, 0, num_instances);
 533 }