src/gallium/auxiliary/vl/vl_compositor.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 Younes Manton.
   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_compiler.h"
  31 #include "pipe/p_context.h"
  32
  33 #include "util/u_memory.h"
  34 #include "util/u_draw.h"
  35 #include "util/u_surface.h"
  36
  37 #include "tgsi/tgsi_ureg.h"
  38
  39 #include "vl_csc.h"
  40 #include "vl_types.h"
  41 #include "vl_compositor.h"
  42
  43 #define MIN_DIRTY (0)
  44 #define MAX_DIRTY (1 << 15)
  45
  46 enum VS_OUTPUT
  47 {
  48    VS_O_VPOS,
  49    VS_O_VTEX,
  50    VS_O_VTOP,
  51    VS_O_VBOTTOM,
  52 };
  53
  54 typedef float csc_matrix[16];
  55
  56 static void *
  57 create_vert_shader(struct vl_compositor *c)
  58 {
  59    struct ureg_program *shader;
  60    struct ureg_src vpos, vtex;
  61    struct ureg_dst tmp;
  62    struct ureg_dst o_vpos, o_vtex;
  63    struct ureg_dst o_vtop, o_vbottom;
  64
  65    shader = ureg_create(TGSI_PROCESSOR_VERTEX);
  66    if (!shader)
  67       return false;
  68
  69    vpos = ureg_DECL_vs_input(shader, 0);
  70    vtex = ureg_DECL_vs_input(shader, 1);
  71    tmp = ureg_DECL_temporary(shader);
  72    o_vpos = ureg_DECL_output(shader, TGSI_SEMANTIC_POSITION, VS_O_VPOS);
  73    o_vtex = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX);
  74    o_vtop = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTOP);
  75    o_vbottom = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VBOTTOM);
  76
  77    /*
  78     * o_vpos = vpos
  79     * o_vtex = vtex
  80     */
  81    ureg_MOV(shader, o_vpos, vpos);
  82    ureg_MOV(shader, o_vtex, vtex);
  83
  84    ureg_MUL(shader, ureg_writemask(tmp, TGSI_WRITEMASK_X),
  85             ureg_scalar(vtex, TGSI_SWIZZLE_W), ureg_imm1f(shader, 0.5f));
  86    ureg_MUL(shader, ureg_writemask(tmp, TGSI_WRITEMASK_Y),
  87             ureg_scalar(vtex, TGSI_SWIZZLE_W), ureg_imm1f(shader, 0.25f));
  88
  89    ureg_MOV(shader, ureg_writemask(o_vtop, TGSI_WRITEMASK_X), vtex);
  90    ureg_MAD(shader, ureg_writemask(o_vtop, TGSI_WRITEMASK_Y), ureg_scalar(vtex, TGSI_SWIZZLE_Y),
  91             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_X), ureg_imm1f(shader, 0.25f));
  92    ureg_MAD(shader, ureg_writemask(o_vtop, TGSI_WRITEMASK_Z), ureg_scalar(vtex, TGSI_SWIZZLE_Y),
  93             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_Y), ureg_imm1f(shader, 0.25f));
  94    ureg_RCP(shader, ureg_writemask(o_vtop, TGSI_WRITEMASK_W),
  95             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_X));
  96
  97    ureg_MOV(shader, ureg_writemask(o_vbottom, TGSI_WRITEMASK_X), vtex);
  98    ureg_MAD(shader, ureg_writemask(o_vbottom, TGSI_WRITEMASK_Y), ureg_scalar(vtex, TGSI_SWIZZLE_Y),
  99             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_X), ureg_imm1f(shader, -0.25f));
 100    ureg_MAD(shader, ureg_writemask(o_vbottom, TGSI_WRITEMASK_Z), ureg_scalar(vtex, TGSI_SWIZZLE_Y),
 101             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_Y), ureg_imm1f(shader, -0.25f));
 102    ureg_RCP(shader, ureg_writemask(o_vbottom, TGSI_WRITEMASK_W),
 103             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_Y));
 104
 105    ureg_END(shader);
 106
 107    return ureg_create_shader_and_destroy(shader, c->pipe);
 108 }
 109
 110 static void *
 111 create_frag_shader_video_buffer(struct vl_compositor *c)
 112 {
 113    struct ureg_program *shader;
 114    struct ureg_src tc;
 115    struct ureg_src csc[3];
 116    struct ureg_src sampler[3];
 117    struct ureg_dst texel;
 118    struct ureg_dst fragment;
 119    unsigned i;
 120
 121    shader = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 122    if (!shader)
 123       return false;
 124
 125    tc = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, 1, TGSI_INTERPOLATE_LINEAR);
 126    for (i = 0; i < 3; ++i) {
 127       csc[i] = ureg_DECL_constant(shader, i);
 128       sampler[i] = ureg_DECL_sampler(shader, i);
 129    }
 130    texel = ureg_DECL_temporary(shader);
 131    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
 132
 133    /*
 134     * texel.xyz = tex(tc, sampler[i])
 135     * fragment = csc * texel
 136     */
 137    for (i = 0; i < 3; ++i)
 138       ureg_TEX(shader, ureg_writemask(texel, TGSI_WRITEMASK_X << i), TGSI_TEXTURE_3D, tc, sampler[i]);
 139
 140    ureg_MOV(shader, ureg_writemask(texel, TGSI_WRITEMASK_W), ureg_imm1f(shader, 1.0f));
 141
 142    for (i = 0; i < 3; ++i)
 143       ureg_DP4(shader, ureg_writemask(fragment, TGSI_WRITEMASK_X << i), csc[i], ureg_src(texel));
 144
 145    ureg_MOV(shader, ureg_writemask(fragment, TGSI_WRITEMASK_W), ureg_imm1f(shader, 1.0f));
 146
 147    ureg_release_temporary(shader, texel);
 148    ureg_END(shader);
 149
 150    return ureg_create_shader_and_destroy(shader, c->pipe);
 151 }
 152
 153 static void *
 154 create_frag_shader_weave(struct vl_compositor *c)
 155 {
 156    struct ureg_program *shader;
 157    struct ureg_src i_tc[2];
 158    struct ureg_src csc[3];
 159    struct ureg_src sampler[3];
 160    struct ureg_dst t_tc[2];
 161    struct ureg_dst t_texel[2];
 162    struct ureg_dst o_fragment;
 163    unsigned i, j;
 164
 165    shader = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 166    if (!shader)
 167       return false;
 168
 169    i_tc[0] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTOP, TGSI_INTERPOLATE_LINEAR);
 170    i_tc[1] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VBOTTOM, TGSI_INTERPOLATE_LINEAR);
 171
 172    for (i = 0; i < 3; ++i) {
 173       csc[i] = ureg_DECL_constant(shader, i);
 174       sampler[i] = ureg_DECL_sampler(shader, i);
 175    }
 176
 177    for (i = 0; i < 2; ++i) {
 178       t_tc[i] = ureg_DECL_temporary(shader);
 179       t_texel[i] = ureg_DECL_temporary(shader);
 180    }
 181    o_fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
 182
 183    /* calculate the texture offsets
 184     * t_tc.x = i_tc.x
 185     * t_tc.y = (round(i_tc.y) + 0.5) / height * 2
 186     */
 187    for (i = 0; i < 2; ++i) {
 188       ureg_MOV(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_X), i_tc[i]);
 189       ureg_ROUND(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_YZ), i_tc[i]);
 190       ureg_MOV(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_W),
 191                ureg_imm1f(shader, i ? 0.75f : 0.25f));
 192       ureg_ADD(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_YZ),
 193                ureg_src(t_tc[i]), ureg_imm1f(shader, 0.5f));
 194       ureg_MUL(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_Y),
 195                ureg_src(t_tc[i]), ureg_scalar(i_tc[0], TGSI_SWIZZLE_W));
 196       ureg_MUL(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_Z),
 197                ureg_src(t_tc[i]), ureg_scalar(i_tc[1], TGSI_SWIZZLE_W));
 198    }
 199
 200    /* fetch the texels
 201     * texel[0..1].x = tex(t_tc[0..1][0])
 202     * texel[0..1].y = tex(t_tc[0..1][1])
 203     * texel[0..1].z = tex(t_tc[0..1][2])
 204     */
 205    for (i = 0; i < 2; ++i)
 206       for (j = 0; j < 3; ++j) {
 207          struct ureg_src src = ureg_swizzle(ureg_src(t_tc[i]),
 208             TGSI_SWIZZLE_X, j ? TGSI_SWIZZLE_Z : TGSI_SWIZZLE_Y, TGSI_SWIZZLE_W, TGSI_SWIZZLE_W);
 209
 210          ureg_TEX(shader, ureg_writemask(t_texel[i], TGSI_WRITEMASK_X << j),
 211                   TGSI_TEXTURE_3D, src, sampler[j]);
 212       }
 213
 214    /* calculate linear interpolation factor
 215     * factor = |round(i_tc.y) - i_tc.y| * 2
 216     */
 217    ureg_ROUND(shader, ureg_writemask(t_tc[0], TGSI_WRITEMASK_YZ), i_tc[0]);
 218    ureg_ADD(shader, ureg_writemask(t_tc[0], TGSI_WRITEMASK_YZ),
 219             ureg_src(t_tc[0]), ureg_negate(i_tc[0]));
 220    ureg_MUL(shader, ureg_writemask(t_tc[0], TGSI_WRITEMASK_XY),
 221             ureg_abs(ureg_src(t_tc[0])), ureg_imm1f(shader, 2.0f));
 222    ureg_LRP(shader, t_texel[0], ureg_swizzle(ureg_src(t_tc[0]),
 223             TGSI_SWIZZLE_Y, TGSI_SWIZZLE_Z, TGSI_SWIZZLE_Z, TGSI_SWIZZLE_Z),
 224             ureg_src(t_texel[1]), ureg_src(t_texel[0]));
 225
 226    /* and finally do colour space transformation
 227     * fragment = csc * texel
 228     */
 229    ureg_MOV(shader, ureg_writemask(t_texel[0], TGSI_WRITEMASK_W), ureg_imm1f(shader, 1.0f));
 230    for (i = 0; i < 3; ++i)
 231       ureg_DP4(shader, ureg_writemask(o_fragment, TGSI_WRITEMASK_X << i), csc[i], ureg_src(t_texel[0]));
 232
 233    ureg_MOV(shader, ureg_writemask(o_fragment, TGSI_WRITEMASK_W), ureg_imm1f(shader, 1.0f));
 234
 235    for (i = 0; i < 2; ++i) {
 236       ureg_release_temporary(shader, t_texel[i]);
 237       ureg_release_temporary(shader, t_tc[i]);
 238    }
 239
 240    ureg_END(shader);
 241
 242    return ureg_create_shader_and_destroy(shader, c->pipe);
 243 }
 244
 245 static void *
 246 create_frag_shader_palette(struct vl_compositor *c, bool include_cc)
 247 {
 248    struct ureg_program *shader;
 249    struct ureg_src csc[3];
 250    struct ureg_src tc;
 251    struct ureg_src sampler;
 252    struct ureg_src palette;
 253    struct ureg_dst texel;
 254    struct ureg_dst fragment;
 255    unsigned i;
 256
 257    shader = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 258    if (!shader)
 259       return false;
 260
 261    for (i = 0; include_cc && i < 3; ++i)
 262       csc[i] = ureg_DECL_constant(shader, i);
 263
 264    tc = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX, TGSI_INTERPOLATE_LINEAR);
 265    sampler = ureg_DECL_sampler(shader, 0);
 266    palette = ureg_DECL_sampler(shader, 1);
 267
 268    texel = ureg_DECL_temporary(shader);
 269    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
 270
 271    /*
 272     * texel = tex(tc, sampler)
 273     * fragment.xyz = tex(texel, palette) * csc
 274     * fragment.a = texel.a
 275     */
 276    ureg_TEX(shader, texel, TGSI_TEXTURE_2D, tc, sampler);
 277    ureg_MOV(shader, ureg_writemask(fragment, TGSI_WRITEMASK_W), ureg_src(texel));
 278
 279    if (include_cc) {
 280       ureg_TEX(shader, texel, TGSI_TEXTURE_1D, ureg_src(texel), palette);
 281       for (i = 0; i < 3; ++i)
 282          ureg_DP4(shader, ureg_writemask(fragment, TGSI_WRITEMASK_X << i), csc[i], ureg_src(texel));
 283    } else {
 284       ureg_TEX(shader, ureg_writemask(fragment, TGSI_WRITEMASK_XYZ),
 285                TGSI_TEXTURE_1D, ureg_src(texel), palette);
 286    }
 287
 288    ureg_release_temporary(shader, texel);
 289    ureg_END(shader);
 290
 291    return ureg_create_shader_and_destroy(shader, c->pipe);
 292 }
 293
 294 static void *
 295 create_frag_shader_rgba(struct vl_compositor *c)
 296 {
 297    struct ureg_program *shader;
 298    struct ureg_src tc;
 299    struct ureg_src sampler;
 300    struct ureg_dst fragment;
 301
 302    shader = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 303    if (!shader)
 304       return false;
 305
 306    tc = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX, TGSI_INTERPOLATE_LINEAR);
 307    sampler = ureg_DECL_sampler(shader, 0);
 308    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
 309
 310    /*
 311     * fragment = tex(tc, sampler)
 312     */
 313    ureg_TEX(shader, fragment, TGSI_TEXTURE_2D, tc, sampler);
 314    ureg_END(shader);
 315
 316    return ureg_create_shader_and_destroy(shader, c->pipe);
 317 }
 318
 319 static bool
 320 init_shaders(struct vl_compositor *c)
 321 {
 322    assert(c);
 323
 324    c->vs = create_vert_shader(c);
 325    if (!c->vs) {
 326       debug_printf("Unable to create vertex shader.\n");
 327       return false;
 328    }
 329
 330    c->fs_video_buffer = create_frag_shader_video_buffer(c);
 331    if (!c->fs_video_buffer) {
 332       debug_printf("Unable to create YCbCr-to-RGB fragment shader.\n");
 333       return false;
 334    }
 335
 336    c->fs_weave = create_frag_shader_weave(c);
 337    if (!c->fs_weave) {
 338       debug_printf("Unable to create YCbCr-to-RGB weave fragment shader.\n");
 339       return false;
 340    }
 341
 342    c->fs_palette.yuv = create_frag_shader_palette(c, true);
 343    if (!c->fs_palette.yuv) {
 344       debug_printf("Unable to create YUV-Palette-to-RGB fragment shader.\n");
 345       return false;
 346    }
 347
 348    c->fs_palette.rgb = create_frag_shader_palette(c, false);
 349    if (!c->fs_palette.rgb) {
 350       debug_printf("Unable to create RGB-Palette-to-RGB fragment shader.\n");
 351       return false;
 352    }
 353
 354    c->fs_rgba = create_frag_shader_rgba(c);
 355    if (!c->fs_rgba) {
 356       debug_printf("Unable to create RGB-to-RGB fragment shader.\n");
 357       return false;
 358    }
 359
 360    return true;
 361 }
 362
 363 static void cleanup_shaders(struct vl_compositor *c)
 364 {
 365    assert(c);
 366
 367    c->pipe->delete_vs_state(c->pipe, c->vs);
 368    c->pipe->delete_fs_state(c->pipe, c->fs_video_buffer);
 369    c->pipe->delete_fs_state(c->pipe, c->fs_weave);
 370    c->pipe->delete_fs_state(c->pipe, c->fs_palette.yuv);
 371    c->pipe->delete_fs_state(c->pipe, c->fs_palette.rgb);
 372    c->pipe->delete_fs_state(c->pipe, c->fs_rgba);
 373 }
 374
 375 static bool
 376 init_pipe_state(struct vl_compositor *c)
 377 {
 378    struct pipe_rasterizer_state rast;
 379    struct pipe_sampler_state sampler;
 380    struct pipe_blend_state blend;
 381    struct pipe_depth_stencil_alpha_state dsa;
 382    unsigned i;
 383
 384    assert(c);
 385
 386    c->fb_state.nr_cbufs = 1;
 387    c->fb_state.zsbuf = NULL;
 388
 389    memset(&sampler, 0, sizeof(sampler));
 390    sampler.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 391    sampler.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 392    sampler.wrap_r = PIPE_TEX_WRAP_REPEAT;
 393    sampler.min_img_filter = PIPE_TEX_FILTER_LINEAR;
 394    sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
 395    sampler.mag_img_filter = PIPE_TEX_FILTER_LINEAR;
 396    sampler.compare_mode = PIPE_TEX_COMPARE_NONE;
 397    sampler.compare_func = PIPE_FUNC_ALWAYS;
 398    sampler.normalized_coords = 1;
 399
 400    c->sampler_linear = c->pipe->create_sampler_state(c->pipe, &sampler);
 401
 402    sampler.min_img_filter = PIPE_TEX_FILTER_NEAREST;
 403    sampler.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
 404    c->sampler_nearest = c->pipe->create_sampler_state(c->pipe, &sampler);
 405
 406    memset(&blend, 0, sizeof blend);
 407    blend.independent_blend_enable = 0;
 408    blend.rt[0].blend_enable = 0;
 409    blend.logicop_enable = 0;
 410    blend.logicop_func = PIPE_LOGICOP_CLEAR;
 411    blend.rt[0].colormask = PIPE_MASK_RGBA;
 412    blend.dither = 0;
 413    c->blend_clear = c->pipe->create_blend_state(c->pipe, &blend);
 414
 415    blend.rt[0].blend_enable = 1;
 416    blend.rt[0].rgb_func = PIPE_BLEND_ADD;
 417    blend.rt[0].rgb_src_factor = PIPE_BLENDFACTOR_SRC_ALPHA;
 418    blend.rt[0].rgb_dst_factor = PIPE_BLENDFACTOR_INV_SRC_ALPHA;
 419    blend.rt[0].alpha_func = PIPE_BLEND_ADD;
 420    blend.rt[0].alpha_src_factor = PIPE_BLENDFACTOR_ONE;
 421    blend.rt[0].alpha_dst_factor = PIPE_BLENDFACTOR_ONE;
 422    c->blend_add = c->pipe->create_blend_state(c->pipe, &blend);
 423
 424    memset(&rast, 0, sizeof rast);
 425    rast.flatshade = 1;
 426    rast.front_ccw = 1;
 427    rast.cull_face = PIPE_FACE_NONE;
 428    rast.fill_back = PIPE_POLYGON_MODE_FILL;
 429    rast.fill_front = PIPE_POLYGON_MODE_FILL;
 430    rast.scissor = 1;
 431    rast.line_width = 1;
 432    rast.point_size_per_vertex = 1;
 433    rast.offset_units = 1;
 434    rast.offset_scale = 1;
 435    rast.gl_rasterization_rules = 1;
 436    rast.depth_clip = 1;
 437
 438    c->rast = c->pipe->create_rasterizer_state(c->pipe, &rast);
 439
 440    memset(&dsa, 0, sizeof dsa);
 441    dsa.depth.enabled = 0;
 442    dsa.depth.writemask = 0;
 443    dsa.depth.func = PIPE_FUNC_ALWAYS;
 444    for (i = 0; i < 2; ++i) {
 445       dsa.stencil[i].enabled = 0;
 446       dsa.stencil[i].func = PIPE_FUNC_ALWAYS;
 447       dsa.stencil[i].fail_op = PIPE_STENCIL_OP_KEEP;
 448       dsa.stencil[i].zpass_op = PIPE_STENCIL_OP_KEEP;
 449       dsa.stencil[i].zfail_op = PIPE_STENCIL_OP_KEEP;
 450       dsa.stencil[i].valuemask = 0;
 451       dsa.stencil[i].writemask = 0;
 452    }
 453    dsa.alpha.enabled = 0;
 454    dsa.alpha.func = PIPE_FUNC_ALWAYS;
 455    dsa.alpha.ref_value = 0;
 456    c->dsa = c->pipe->create_depth_stencil_alpha_state(c->pipe, &dsa);
 457    c->pipe->bind_depth_stencil_alpha_state(c->pipe, c->dsa);
 458
 459    return true;
 460 }
 461
 462 static void cleanup_pipe_state(struct vl_compositor *c)
 463 {
 464    assert(c);
 465
 466    /* Asserted in softpipe_delete_fs_state() for some reason */
 467    c->pipe->bind_vs_state(c->pipe, NULL);
 468    c->pipe->bind_fs_state(c->pipe, NULL);
 469
 470    c->pipe->delete_depth_stencil_alpha_state(c->pipe, c->dsa);
 471    c->pipe->delete_sampler_state(c->pipe, c->sampler_linear);
 472    c->pipe->delete_sampler_state(c->pipe, c->sampler_nearest);
 473    c->pipe->delete_blend_state(c->pipe, c->blend_clear);
 474    c->pipe->delete_blend_state(c->pipe, c->blend_add);
 475    c->pipe->delete_rasterizer_state(c->pipe, c->rast);
 476 }
 477
 478 static bool
 479 create_vertex_buffer(struct vl_compositor *c)
 480 {
 481    assert(c);
 482
 483    pipe_resource_reference(&c->vertex_buf.buffer, NULL);
 484    c->vertex_buf.buffer = pipe_buffer_create
 485    (
 486       c->pipe->screen,
 487       PIPE_BIND_VERTEX_BUFFER,
 488       PIPE_USAGE_STREAM,
 489       c->vertex_buf.stride * VL_COMPOSITOR_MAX_LAYERS * 4
 490    );
 491
 492    return c->vertex_buf.buffer != NULL;
 493 }
 494
 495 static bool
 496 init_buffers(struct vl_compositor *c)
 497 {
 498    struct pipe_vertex_element vertex_elems[2];
 499
 500    assert(c);
 501
 502    /*
 503     * Create our vertex buffer and vertex buffer elements
 504     */
 505    c->vertex_buf.stride = sizeof(struct vertex2f) + sizeof(struct vertex4f);
 506    c->vertex_buf.buffer_offset = 0;
 507    create_vertex_buffer(c);
 508
 509    vertex_elems[0].src_offset = 0;
 510    vertex_elems[0].instance_divisor = 0;
 511    vertex_elems[0].vertex_buffer_index = 0;
 512    vertex_elems[0].src_format = PIPE_FORMAT_R32G32_FLOAT;
 513    vertex_elems[1].src_offset = sizeof(struct vertex2f);
 514    vertex_elems[1].instance_divisor = 0;
 515    vertex_elems[1].vertex_buffer_index = 0;
 516    vertex_elems[1].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
 517    c->vertex_elems_state = c->pipe->create_vertex_elements_state(c->pipe, 2, vertex_elems);
 518
 519    return true;
 520 }
 521
 522 static void
 523 cleanup_buffers(struct vl_compositor *c)
 524 {
 525    assert(c);
 526
 527    c->pipe->delete_vertex_elements_state(c->pipe, c->vertex_elems_state);
 528    pipe_resource_reference(&c->vertex_buf.buffer, NULL);
 529 }
 530
 531 static INLINE struct u_rect
 532 default_rect(struct vl_compositor_layer *layer)
 533 {
 534    struct pipe_resource *res = layer->sampler_views[0]->texture;
 535    struct u_rect rect = { 0, res->width0, 0, res->height0 * res->depth0 };
 536    return rect;
 537 }
 538
 539 static INLINE struct vertex2f
 540 calc_topleft(struct vertex2f size, struct u_rect rect)
 541 {
 542    struct vertex2f res = { rect.x0 / size.x, rect.y0 / size.y };
 543    return res;
 544 }
 545
 546 static INLINE struct vertex2f
 547 calc_bottomright(struct vertex2f size, struct u_rect rect)
 548 {
 549    struct vertex2f res = { rect.x1 / size.x, rect.y1 / size.y };
 550    return res;
 551 }
 552
 553 static INLINE void
 554 calc_src_and_dst(struct vl_compositor_layer *layer, unsigned width, unsigned height,
 555                  struct u_rect src, struct u_rect dst)
 556 {
 557    struct vertex2f size =  { width, height };
 558
 559    layer->src.tl = calc_topleft(size, src);
 560    layer->src.br = calc_bottomright(size, src);
 561    layer->dst.tl = calc_topleft(size, dst);
 562    layer->dst.br = calc_bottomright(size, dst);
 563    layer->zw.x = 0.0f;
 564    layer->zw.y = size.y;
 565 }
 566
 567 static void
 568 gen_rect_verts(struct vertex2f *vb, struct vl_compositor_layer *layer)
 569 {
 570    assert(vb && layer);
 571
 572    vb[ 0].x = layer->dst.tl.x;
 573    vb[ 0].y = layer->dst.tl.y;
 574    vb[ 1].x = layer->src.tl.x;
 575    vb[ 1].y = layer->src.tl.y;
 576    vb[ 2] = layer->zw;
 577
 578    vb[ 3].x = layer->dst.br.x;
 579    vb[ 3].y = layer->dst.tl.y;
 580    vb[ 4].x = layer->src.br.x;
 581    vb[ 4].y = layer->src.tl.y;
 582    vb[ 5] = layer->zw;
 583
 584    vb[ 6].x = layer->dst.br.x;
 585    vb[ 6].y = layer->dst.br.y;
 586    vb[ 7].x = layer->src.br.x;
 587    vb[ 7].y = layer->src.br.y;
 588    vb[ 8] = layer->zw;
 589
 590    vb[ 9].x = layer->dst.tl.x;
 591    vb[ 9].y = layer->dst.br.y;
 592    vb[10].x = layer->src.tl.x;
 593    vb[10].y = layer->src.br.y;
 594    vb[11] = layer->zw;
 595 }
 596
 597 static INLINE struct u_rect
 598 calc_drawn_area(struct vl_compositor_state *s, struct vl_compositor_layer *layer)
 599 {
 600    struct u_rect result;
 601
 602    // scale
 603    result.x0 = layer->dst.tl.x * s->viewport.scale[0] + s->viewport.translate[0];
 604    result.y0 = layer->dst.tl.y * s->viewport.scale[1] + s->viewport.translate[1];
 605    result.x1 = layer->dst.br.x * s->viewport.scale[0] + s->viewport.translate[0];
 606    result.y1 = layer->dst.br.y * s->viewport.scale[1] + s->viewport.translate[1];
 607
 608    // and clip
 609    result.x0 = MAX2(result.x0, s->scissor.minx);
 610    result.y0 = MAX2(result.y0, s->scissor.miny);
 611    result.x1 = MIN2(result.x1, s->scissor.maxx);
 612    result.y1 = MIN2(result.y1, s->scissor.maxy);
 613    return result;
 614 }
 615
 616 static void
 617 gen_vertex_data(struct vl_compositor *c, struct vl_compositor_state *s, struct u_rect *dirty)
 618 {
 619    struct vertex2f *vb;
 620    struct pipe_transfer *buf_transfer;
 621    unsigned i;
 622
 623    assert(c);
 624
 625    vb = pipe_buffer_map(c->pipe, c->vertex_buf.buffer,
 626                         PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE | PIPE_TRANSFER_DONTBLOCK,
 627                         &buf_transfer);
 628
 629    if (!vb) {
 630       // If buffer is still locked from last draw create a new one
 631       create_vertex_buffer(c);
 632       vb = pipe_buffer_map(c->pipe, c->vertex_buf.buffer,
 633                            PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE,
 634                            &buf_transfer);
 635    }
 636
 637    for (i = 0; i < VL_COMPOSITOR_MAX_LAYERS; i++) {
 638       if (s->used_layers & (1 << i)) {
 639          struct vl_compositor_layer *layer = &s->layers[i];
 640          gen_rect_verts(vb, layer);
 641          vb += 12;
 642
 643          if (dirty && layer->clearing) {
 644             struct u_rect drawn = calc_drawn_area(s, layer);
 645             if (
 646              dirty->x0 >= drawn.x0 &&
 647              dirty->y0 >= drawn.y0 &&
 648              dirty->x1 <= drawn.x1 &&
 649              dirty->y1 <= drawn.y1) {
 650
 651                // We clear the dirty area anyway, no need for clear_render_target
 652                dirty->x0 = dirty->y0 = MAX_DIRTY;
 653                dirty->x1 = dirty->y1 = MIN_DIRTY;
 654             }
 655          }
 656       }
 657    }
 658
 659    pipe_buffer_unmap(c->pipe, buf_transfer);
 660 }
 661
 662 static void
 663 draw_layers(struct vl_compositor *c, struct vl_compositor_state *s, struct u_rect *dirty)
 664 {
 665    unsigned vb_index, i;
 666
 667    assert(c);
 668
 669    for (i = 0, vb_index = 0; i < VL_COMPOSITOR_MAX_LAYERS; ++i) {
 670       if (s->used_layers & (1 << i)) {
 671          struct vl_compositor_layer *layer = &s->layers[i];
 672          struct pipe_sampler_view **samplers = &layer->sampler_views[0];
 673          unsigned num_sampler_views = !samplers[1] ? 1 : !samplers[2] ? 2 : 3;
 674          void *blend = layer->blend ? layer->blend : i ? c->blend_add : c->blend_clear;
 675
 676          c->pipe->bind_blend_state(c->pipe, blend);
 677          c->pipe->bind_fs_state(c->pipe, layer->fs);
 678          c->pipe->bind_fragment_sampler_states(c->pipe, num_sampler_views, layer->samplers);
 679          c->pipe->set_fragment_sampler_views(c->pipe, num_sampler_views, samplers);
 680          util_draw_arrays(c->pipe, PIPE_PRIM_QUADS, vb_index * 4, 4);
 681          vb_index++;
 682
 683          if (dirty) {
 684             // Remember the currently drawn area as dirty for the next draw command
 685             struct u_rect drawn = calc_drawn_area(s, layer);
 686             dirty->x0 = MIN2(drawn.x0, dirty->x0);
 687             dirty->y0 = MIN2(drawn.y0, dirty->y0);
 688             dirty->x1 = MAX2(drawn.x1, dirty->x1);
 689             dirty->y1 = MAX2(drawn.y1, dirty->y1);
 690          }
 691       }
 692    }
 693 }
 694
 695 void
 696 vl_compositor_reset_dirty_area(struct u_rect *dirty)
 697 {
 698    assert(dirty);
 699
 700    dirty->x0 = dirty->y0 = MIN_DIRTY;
 701    dirty->x1 = dirty->y1 = MAX_DIRTY;
 702 }
 703
 704 void
 705 vl_compositor_set_clear_color(struct vl_compositor_state *s, union pipe_color_union *color)
 706 {
 707    assert(s);
 708    assert(color);
 709
 710    s->clear_color = *color;
 711 }
 712
 713 void
 714 vl_compositor_get_clear_color(struct vl_compositor_state *s, union pipe_color_union *color)
 715 {
 716    assert(s);
 717    assert(color);
 718
 719    *color = s->clear_color;
 720 }
 721
 722 void
 723 vl_compositor_clear_layers(struct vl_compositor_state *s)
 724 {
 725    unsigned i, j;
 726
 727    assert(s);
 728
 729    s->used_layers = 0;
 730    for ( i = 0; i < VL_COMPOSITOR_MAX_LAYERS; ++i) {
 731       s->layers[i].clearing = i ? false : true;
 732       s->layers[i].blend = NULL;
 733       s->layers[i].fs = NULL;
 734       for ( j = 0; j < 3; j++)
 735          pipe_sampler_view_reference(&s->layers[i].sampler_views[j], NULL);
 736    }
 737 }
 738
 739 void
 740 vl_compositor_cleanup(struct vl_compositor *c)
 741 {
 742    assert(c);
 743
 744    cleanup_buffers(c);
 745    cleanup_shaders(c);
 746    cleanup_pipe_state(c);
 747 }
 748
 749 void
 750 vl_compositor_set_csc_matrix(struct vl_compositor_state *s, const float matrix[16])
 751 {
 752    struct pipe_transfer *buf_transfer;
 753
 754    assert(s);
 755
 756    memcpy
 757    (
 758       pipe_buffer_map(s->pipe, s->csc_matrix,
 759                       PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE,
 760                       &buf_transfer),
 761       matrix,
 762       sizeof(csc_matrix)
 763    );
 764
 765    pipe_buffer_unmap(s->pipe, buf_transfer);
 766 }
 767
 768 void
 769 vl_compositor_set_dst_area(struct vl_compositor_state *s, struct u_rect *dst_area)
 770 {
 771    assert(s);
 772
 773    s->viewport_valid = dst_area != NULL;
 774    if (dst_area) {
 775       s->viewport.scale[0] = dst_area->x1 - dst_area->x0;
 776       s->viewport.scale[1] = dst_area->y1 - dst_area->y0;
 777       s->viewport.translate[0] = dst_area->x0;
 778       s->viewport.translate[1] = dst_area->y0;
 779    }
 780 }
 781
 782 void
 783 vl_compositor_set_dst_clip(struct vl_compositor_state *s, struct u_rect *dst_clip)
 784 {
 785    assert(s);
 786
 787    s->scissor_valid = dst_clip != NULL;
 788    if (dst_clip) {
 789       s->scissor.minx = dst_clip->x0;
 790       s->scissor.miny = dst_clip->y0;
 791       s->scissor.maxx = dst_clip->x1;
 792       s->scissor.maxy = dst_clip->y1;
 793    }
 794 }
 795
 796 void
 797 vl_compositor_set_layer_blend(struct vl_compositor_state *s,
 798                               unsigned layer, void *blend,
 799                               bool is_clearing)
 800 {
 801    assert(s && blend);
 802
 803    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
 804
 805    s->layers[layer].clearing = is_clearing;
 806    s->layers[layer].blend = blend;
 807 }
 808
 809 void
 810 vl_compositor_set_buffer_layer(struct vl_compositor_state *s,
 811                                struct vl_compositor *c,
 812                                unsigned layer,
 813                                struct pipe_video_buffer *buffer,
 814                                struct u_rect *src_rect,
 815                                struct u_rect *dst_rect,
 816                                enum vl_compositor_deinterlace deinterlace)
 817 {
 818    struct pipe_sampler_view **sampler_views;
 819    unsigned i;
 820
 821    assert(s && c && buffer);
 822
 823    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
 824
 825    s->used_layers |= 1 << layer;
 826    sampler_views = buffer->get_sampler_view_components(buffer);
 827    for (i = 0; i < 3; ++i) {
 828       s->layers[layer].samplers[i] = c->sampler_linear;
 829       pipe_sampler_view_reference(&s->layers[layer].sampler_views[i], sampler_views[i]);
 830    }
 831
 832    calc_src_and_dst(&s->layers[layer], buffer->width, buffer->height,
 833                     src_rect ? *src_rect : default_rect(&s->layers[layer]),
 834                     dst_rect ? *dst_rect : default_rect(&s->layers[layer]));
 835
 836    if (buffer->interlaced) {
 837       float half_a_line = 0.5f / s->layers[layer].zw.y;
 838       switch(deinterlace) {
 839       case VL_COMPOSITOR_WEAVE:
 840          s->layers[layer].fs = c->fs_weave;
 841          break;
 842
 843       case VL_COMPOSITOR_BOB_TOP:
 844          s->layers[layer].zw.x = 0.25f;
 845          s->layers[layer].src.tl.y += half_a_line;
 846          s->layers[layer].src.br.y += half_a_line;
 847          s->layers[layer].fs = c->fs_video_buffer;
 848          break;
 849
 850       case VL_COMPOSITOR_BOB_BOTTOM:
 851          s->layers[layer].zw.x = 0.75f;
 852          s->layers[layer].src.tl.y -= half_a_line;
 853          s->layers[layer].src.br.y -= half_a_line;
 854          s->layers[layer].fs = c->fs_video_buffer;
 855          break;
 856       }
 857
 858    } else
 859       s->layers[layer].fs = c->fs_video_buffer;
 860 }
 861
 862 void
 863 vl_compositor_set_palette_layer(struct vl_compositor_state *s,
 864                                 struct vl_compositor *c,
 865                                 unsigned layer,
 866                                 struct pipe_sampler_view *indexes,
 867                                 struct pipe_sampler_view *palette,
 868                                 struct u_rect *src_rect,
 869                                 struct u_rect *dst_rect,
 870                                 bool include_color_conversion)
 871 {
 872    assert(s && c && indexes && palette);
 873
 874    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
 875
 876    s->used_layers |= 1 << layer;
 877
 878    s->layers[layer].fs = include_color_conversion ?
 879       c->fs_palette.yuv : c->fs_palette.rgb;
 880
 881    s->layers[layer].samplers[0] = c->sampler_linear;
 882    s->layers[layer].samplers[1] = c->sampler_nearest;
 883    s->layers[layer].samplers[2] = NULL;
 884    pipe_sampler_view_reference(&s->layers[layer].sampler_views[0], indexes);
 885    pipe_sampler_view_reference(&s->layers[layer].sampler_views[1], palette);
 886    pipe_sampler_view_reference(&s->layers[layer].sampler_views[2], NULL);
 887    calc_src_and_dst(&s->layers[layer], indexes->texture->width0, indexes->texture->height0,
 888                     src_rect ? *src_rect : default_rect(&s->layers[layer]),
 889                     dst_rect ? *dst_rect : default_rect(&s->layers[layer]));
 890 }
 891
 892 void
 893 vl_compositor_set_rgba_layer(struct vl_compositor_state *s,
 894                              struct vl_compositor *c,
 895                              unsigned layer,
 896                              struct pipe_sampler_view *rgba,
 897                              struct u_rect *src_rect,
 898                              struct u_rect *dst_rect)
 899 {
 900    assert(s && c && rgba);
 901
 902    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
 903
 904    s->used_layers |= 1 << layer;
 905    s->layers[layer].fs = c->fs_rgba;
 906    s->layers[layer].samplers[0] = c->sampler_linear;
 907    s->layers[layer].samplers[1] = NULL;
 908    s->layers[layer].samplers[2] = NULL;
 909    pipe_sampler_view_reference(&s->layers[layer].sampler_views[0], rgba);
 910    pipe_sampler_view_reference(&s->layers[layer].sampler_views[1], NULL);
 911    pipe_sampler_view_reference(&s->layers[layer].sampler_views[2], NULL);
 912    calc_src_and_dst(&s->layers[layer], rgba->texture->width0, rgba->texture->height0,
 913                     src_rect ? *src_rect : default_rect(&s->layers[layer]),
 914                     dst_rect ? *dst_rect : default_rect(&s->layers[layer]));
 915 }
 916
 917 void
 918 vl_compositor_render(struct vl_compositor_state *s,
 919                      struct vl_compositor       *c,
 920                      struct pipe_surface        *dst_surface,
 921                      struct u_rect              *dirty_area)
 922 {
 923    assert(c);
 924    assert(dst_surface);
 925
 926    c->fb_state.width = dst_surface->width;
 927    c->fb_state.height = dst_surface->height;
 928    c->fb_state.cbufs[0] = dst_surface;
 929
 930    if (!s->viewport_valid) {
 931       s->viewport.scale[0] = dst_surface->width;
 932       s->viewport.scale[1] = dst_surface->height;
 933       s->viewport.translate[0] = 0;
 934       s->viewport.translate[1] = 0;
 935    }
 936
 937    if (!s->scissor_valid) {
 938       s->scissor.minx = 0;
 939       s->scissor.miny = 0;
 940       s->scissor.maxx = dst_surface->width;
 941       s->scissor.maxy = dst_surface->height;
 942    }
 943
 944    gen_vertex_data(c, s, dirty_area);
 945
 946    if (dirty_area && (dirty_area->x0 < dirty_area->x1 ||
 947                       dirty_area->y0 < dirty_area->y1)) {
 948
 949       c->pipe->clear_render_target(c->pipe, dst_surface, &s->clear_color,
 950                                    0, 0, dst_surface->width, dst_surface->height);
 951       dirty_area->x0 = dirty_area->y0 = MAX_DIRTY;
 952       dirty_area->x1 = dirty_area->y1 = MIN_DIRTY;
 953    }
 954
 955    c->pipe->set_scissor_state(c->pipe, &s->scissor);
 956    c->pipe->set_framebuffer_state(c->pipe, &c->fb_state);
 957    c->pipe->set_viewport_state(c->pipe, &s->viewport);
 958    c->pipe->bind_vs_state(c->pipe, c->vs);
 959    c->pipe->set_vertex_buffers(c->pipe, 1, &c->vertex_buf);
 960    c->pipe->bind_vertex_elements_state(c->pipe, c->vertex_elems_state);
 961    c->pipe->set_constant_buffer(c->pipe, PIPE_SHADER_FRAGMENT, 0, s->csc_matrix);
 962    c->pipe->bind_rasterizer_state(c->pipe, c->rast);
 963
 964    draw_layers(c, s, dirty_area);
 965 }
 966
 967 bool
 968 vl_compositor_init(struct vl_compositor *c, struct pipe_context *pipe)
 969 {
 970    assert(c);
 971
 972    memset(c, 0, sizeof(*c));
 973
 974    c->pipe = pipe;
 975
 976    if (!init_pipe_state(c))
 977       return false;
 978
 979    if (!init_shaders(c)) {
 980       cleanup_pipe_state(c);
 981       return false;
 982    }
 983
 984    if (!init_buffers(c)) {
 985       cleanup_shaders(c);
 986       cleanup_pipe_state(c);
 987       return false;
 988    }
 989
 990    return true;
 991 }
 992
 993 bool
 994 vl_compositor_init_state(struct vl_compositor_state *s, struct pipe_context *pipe)
 995 {
 996    csc_matrix csc_matrix;
 997
 998    assert(s);
 999
1000    memset(s, 0, sizeof(*s));
1001
1002    s->pipe = pipe;
1003
1004    s->viewport.scale[2] = 1;
1005    s->viewport.scale[3] = 1;
1006    s->viewport.translate[2] = 0;
1007    s->viewport.translate[3] = 0;
1008
1009    s->clear_color.f[0] = s->clear_color.f[1] = 0.0f;
1010    s->clear_color.f[2] = s->clear_color.f[3] = 0.0f;
1011
1012    /*
1013     * Create our fragment shader's constant buffer
1014     * Const buffer contains the color conversion matrix and bias vectors
1015     */
1016    /* XXX: Create with IMMUTABLE/STATIC... although it does change every once in a long while... */
1017    s->csc_matrix = pipe_buffer_create
1018    (
1019       pipe->screen,
1020       PIPE_BIND_CONSTANT_BUFFER,
1021       PIPE_USAGE_STATIC,
1022       sizeof(csc_matrix)
1023    );
1024
1025    vl_compositor_clear_layers(s);
1026
1027    vl_csc_get_matrix(VL_CSC_COLOR_STANDARD_IDENTITY, NULL, true, csc_matrix);
1028    vl_compositor_set_csc_matrix(s, csc_matrix);
1029
1030    return true;
1031 }
1032
1033 void
1034 vl_compositor_cleanup_state(struct vl_compositor_state *s)
1035 {
1036    assert(s);
1037
1038    vl_compositor_clear_layers(s);
1039    pipe_resource_reference(&s->csc_matrix, NULL);
1040 }