src/gallium/auxiliary/vl/vl_mc.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 VMWARE 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_context.h"
  31
  32 #include "util/u_sampler.h"
  33 #include "util/u_draw.h"
  34
  35 #include "tgsi/tgsi_ureg.h"
  36
  37 #include "vl_defines.h"
  38 #include "vl_vertex_buffers.h"
  39 #include "vl_mc.h"
  40 #include "vl_idct.h"
  41
  42 enum VS_OUTPUT
  43 {
  44    VS_O_VPOS = 0,
  45    VS_O_VTOP = 0,
  46    VS_O_VBOTTOM,
  47
  48    VS_O_FLAGS = VS_O_VTOP,
  49    VS_O_VTEX = VS_O_VBOTTOM
  50 };
  51
  52 static struct ureg_dst
  53 calc_position(struct vl_mc *r, struct ureg_program *shader, struct ureg_src block_scale)
  54 {
  55    struct ureg_src vrect, vpos;
  56    struct ureg_dst t_vpos;
  57    struct ureg_dst o_vpos;
  58
  59    vrect = ureg_DECL_vs_input(shader, VS_I_RECT);
  60    vpos = ureg_DECL_vs_input(shader, VS_I_VPOS);
  61
  62    t_vpos = ureg_DECL_temporary(shader);
  63
  64    o_vpos = ureg_DECL_output(shader, TGSI_SEMANTIC_POSITION, VS_O_VPOS);
  65
  66    /*
  67     * block_scale = (VL_MACROBLOCK_WIDTH, VL_MACROBLOCK_HEIGHT) / (dst.width, dst.height)
  68     *
  69     * t_vpos = (vpos + vrect) * block_scale
  70     * o_vpos.xy = t_vpos
  71     * o_vpos.zw = vpos
  72     */
  73    ureg_ADD(shader, ureg_writemask(t_vpos, TGSI_WRITEMASK_XY), vpos, vrect);
  74    ureg_MUL(shader, ureg_writemask(t_vpos, TGSI_WRITEMASK_XY), ureg_src(t_vpos), block_scale);
  75    ureg_MOV(shader, ureg_writemask(o_vpos, TGSI_WRITEMASK_XY), ureg_src(t_vpos));
  76    ureg_MOV(shader, ureg_writemask(o_vpos, TGSI_WRITEMASK_ZW), ureg_imm1f(shader, 1.0f));
  77
  78    return t_vpos;
  79 }
  80
  81 static struct ureg_dst
  82 calc_line(struct pipe_screen *screen, struct ureg_program *shader)
  83 {
  84    struct ureg_dst tmp;
  85    struct ureg_src pos;
  86
  87    tmp = ureg_DECL_temporary(shader);
  88
  89    if (screen->get_param(screen, PIPE_CAP_TGSI_FS_POSITION_IS_SYSVAL))
  90       pos = ureg_DECL_system_value(shader, TGSI_SEMANTIC_POSITION, 0);
  91    else
  92       pos = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_POSITION, VS_O_VPOS,
  93                                TGSI_INTERPOLATE_LINEAR);
  94
  95    /*
  96     * tmp.y = fraction(pos.y / 2) >= 0.5 ? 1 : 0
  97     */
  98    ureg_MUL(shader, ureg_writemask(tmp, TGSI_WRITEMASK_Y), pos, ureg_imm1f(shader, 0.5f));
  99    ureg_FRC(shader, ureg_writemask(tmp, TGSI_WRITEMASK_Y), ureg_src(tmp));
 100    ureg_SGE(shader, ureg_writemask(tmp, TGSI_WRITEMASK_Y), ureg_src(tmp), ureg_imm1f(shader, 0.5f));
 101
 102    return tmp;
 103 }
 104
 105 static void *
 106 create_ref_vert_shader(struct vl_mc *r)
 107 {
 108    struct ureg_program *shader;
 109    struct ureg_src mv_scale;
 110    struct ureg_src vmv[2];
 111    struct ureg_dst t_vpos;
 112    struct ureg_dst o_vmv[2];
 113    unsigned i;
 114
 115    shader = ureg_create(TGSI_PROCESSOR_VERTEX);
 116    if (!shader)
 117       return NULL;
 118
 119    vmv[0] = ureg_DECL_vs_input(shader, VS_I_MV_TOP);
 120    vmv[1] = ureg_DECL_vs_input(shader, VS_I_MV_BOTTOM);
 121
 122    t_vpos = calc_position(r, shader, ureg_imm2f(shader,
 123       (float)VL_MACROBLOCK_WIDTH / r->buffer_width,
 124       (float)VL_MACROBLOCK_HEIGHT / r->buffer_height)
 125    );
 126
 127    o_vmv[0] = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTOP);
 128    o_vmv[1] = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VBOTTOM);
 129
 130    /*
 131     * mv_scale.xy = 0.5 / (dst.width, dst.height);
 132     * mv_scale.z = 1.0f / 4.0f
 133     * mv_scale.w = 1.0f / 255.0f
 134     *
 135     * // Apply motion vectors
 136     * o_vmv[0..1].xy = vmv[0..1] * mv_scale + t_vpos
 137     * o_vmv[0..1].zw = vmv[0..1] * mv_scale
 138     *
 139     */
 140
 141    mv_scale = ureg_imm4f(shader,
 142       0.5f / r->buffer_width,
 143       0.5f / r->buffer_height,
 144       1.0f / 4.0f,
 145       1.0f / PIPE_VIDEO_MV_WEIGHT_MAX);
 146
 147    for (i = 0; i < 2; ++i) {
 148       ureg_MAD(shader, ureg_writemask(o_vmv[i], TGSI_WRITEMASK_XY), mv_scale, vmv[i], ureg_src(t_vpos));
 149       ureg_MUL(shader, ureg_writemask(o_vmv[i], TGSI_WRITEMASK_ZW), mv_scale, vmv[i]);
 150    }
 151
 152    ureg_release_temporary(shader, t_vpos);
 153
 154    ureg_END(shader);
 155
 156    return ureg_create_shader_and_destroy(shader, r->pipe);
 157 }
 158
 159 static void *
 160 create_ref_frag_shader(struct vl_mc *r)
 161 {
 162    const float y_scale =
 163       r->buffer_height / 2 *
 164       r->macroblock_size / VL_MACROBLOCK_HEIGHT;
 165
 166    struct ureg_program *shader;
 167    struct ureg_src tc[2], sampler;
 168    struct ureg_dst ref, field;
 169    struct ureg_dst fragment;
 170    unsigned label;
 171
 172    shader = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 173    if (!shader)
 174       return NULL;
 175
 176    tc[0] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTOP, TGSI_INTERPOLATE_LINEAR);
 177    tc[1] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VBOTTOM, TGSI_INTERPOLATE_LINEAR);
 178
 179    sampler = ureg_DECL_sampler(shader, 0);
 180    ref = ureg_DECL_temporary(shader);
 181
 182    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
 183
 184    field = calc_line(r->pipe->screen, shader);
 185
 186    /*
 187     * ref = field.z ? tc[1] : tc[0]
 188     *
 189     * // Adjust tc acording to top/bottom field selection
 190     * if (|ref.z|) {
 191     *    ref.y *= y_scale
 192     *    ref.y = floor(ref.y)
 193     *    ref.y += ref.z
 194     *    ref.y /= y_scale
 195     * }
 196     * fragment.xyz = tex(ref, sampler[0])
 197     */
 198    ureg_CMP(shader, ureg_writemask(ref, TGSI_WRITEMASK_XYZ),
 199             ureg_negate(ureg_scalar(ureg_src(field), TGSI_SWIZZLE_Y)),
 200             tc[1], tc[0]);
 201    ureg_CMP(shader, ureg_writemask(fragment, TGSI_WRITEMASK_W),
 202             ureg_negate(ureg_scalar(ureg_src(field), TGSI_SWIZZLE_Y)),
 203             tc[1], tc[0]);
 204
 205    ureg_IF(shader, ureg_scalar(ureg_src(ref), TGSI_SWIZZLE_Z), &label);
 206
 207       ureg_MUL(shader, ureg_writemask(ref, TGSI_WRITEMASK_Y),
 208                ureg_src(ref), ureg_imm1f(shader, y_scale));
 209       ureg_FLR(shader, ureg_writemask(ref, TGSI_WRITEMASK_Y), ureg_src(ref));
 210       ureg_ADD(shader, ureg_writemask(ref, TGSI_WRITEMASK_Y),
 211                ureg_src(ref), ureg_scalar(ureg_src(ref), TGSI_SWIZZLE_Z));
 212       ureg_MUL(shader, ureg_writemask(ref, TGSI_WRITEMASK_Y),
 213                ureg_src(ref), ureg_imm1f(shader, 1.0f / y_scale));
 214
 215    ureg_fixup_label(shader, label, ureg_get_instruction_number(shader));
 216    ureg_ENDIF(shader);
 217
 218    ureg_TEX(shader, ureg_writemask(fragment, TGSI_WRITEMASK_XYZ), TGSI_TEXTURE_2D, ureg_src(ref), sampler);
 219
 220    ureg_release_temporary(shader, ref);
 221
 222    ureg_release_temporary(shader, field);
 223    ureg_END(shader);
 224
 225    return ureg_create_shader_and_destroy(shader, r->pipe);
 226 }
 227
 228 static void *
 229 create_ycbcr_vert_shader(struct vl_mc *r, vl_mc_ycbcr_vert_shader vs_callback, void *callback_priv)
 230 {
 231    struct ureg_program *shader;
 232
 233    struct ureg_src vrect, vpos;
 234    struct ureg_dst t_vpos, t_vtex;
 235    struct ureg_dst o_vpos, o_flags;
 236
 237    struct vertex2f scale = {
 238       (float)VL_BLOCK_WIDTH / r->buffer_width * VL_MACROBLOCK_WIDTH / r->macroblock_size,
 239       (float)VL_BLOCK_HEIGHT / r->buffer_height * VL_MACROBLOCK_HEIGHT / r->macroblock_size
 240    };
 241
 242    unsigned label;
 243
 244    shader = ureg_create(TGSI_PROCESSOR_VERTEX);
 245    if (!shader)
 246       return NULL;
 247
 248    vrect = ureg_DECL_vs_input(shader, VS_I_RECT);
 249    vpos = ureg_DECL_vs_input(shader, VS_I_VPOS);
 250
 251    t_vpos = calc_position(r, shader, ureg_imm2f(shader, scale.x, scale.y));
 252    t_vtex = ureg_DECL_temporary(shader);
 253
 254    o_vpos = ureg_DECL_output(shader, TGSI_SEMANTIC_POSITION, VS_O_VPOS);
 255    o_flags = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_FLAGS);
 256
 257    /*
 258     * o_vtex.xy = t_vpos
 259     * o_flags.z = intra * 0.5
 260     *
 261     * if(interlaced) {
 262     *    t_vtex.xy = vrect.y ? { 0, scale.y } : { -scale.y : 0 }
 263     *    t_vtex.z = vpos.y % 2
 264     *    t_vtex.y = t_vtex.z ? t_vtex.x : t_vtex.y
 265     *    o_vpos.y = t_vtex.y + t_vpos.y
 266     *
 267     *    o_flags.w = t_vtex.z ? 0 : 1
 268     * }
 269     *
 270     */
 271
 272    vs_callback(callback_priv, r, shader, VS_O_VTEX, t_vpos);
 273
 274    ureg_MUL(shader, ureg_writemask(o_flags, TGSI_WRITEMASK_Z),
 275             ureg_scalar(vpos, TGSI_SWIZZLE_Z), ureg_imm1f(shader, 0.5f));
 276    ureg_MOV(shader, ureg_writemask(o_flags, TGSI_WRITEMASK_W), ureg_imm1f(shader, -1.0f));
 277
 278    if (r->macroblock_size == VL_MACROBLOCK_HEIGHT) { //TODO
 279       ureg_IF(shader, ureg_scalar(vpos, TGSI_SWIZZLE_W), &label);
 280
 281          ureg_CMP(shader, ureg_writemask(t_vtex, TGSI_WRITEMASK_XY),
 282                   ureg_negate(ureg_scalar(vrect, TGSI_SWIZZLE_Y)),
 283                   ureg_imm2f(shader, 0.0f, scale.y),
 284                   ureg_imm2f(shader, -scale.y, 0.0f));
 285          ureg_MUL(shader, ureg_writemask(t_vtex, TGSI_WRITEMASK_Z),
 286                   ureg_scalar(vpos, TGSI_SWIZZLE_Y), ureg_imm1f(shader, 0.5f));
 287
 288          ureg_FRC(shader, ureg_writemask(t_vtex, TGSI_WRITEMASK_Z), ureg_src(t_vtex));
 289
 290          ureg_CMP(shader, ureg_writemask(t_vtex, TGSI_WRITEMASK_Y),
 291                   ureg_negate(ureg_scalar(ureg_src(t_vtex), TGSI_SWIZZLE_Z)),
 292                   ureg_scalar(ureg_src(t_vtex), TGSI_SWIZZLE_X),
 293                   ureg_scalar(ureg_src(t_vtex), TGSI_SWIZZLE_Y));
 294          ureg_ADD(shader, ureg_writemask(o_vpos, TGSI_WRITEMASK_Y),
 295                   ureg_src(t_vpos), ureg_src(t_vtex));
 296
 297          ureg_CMP(shader, ureg_writemask(o_flags, TGSI_WRITEMASK_W),
 298                   ureg_negate(ureg_scalar(ureg_src(t_vtex), TGSI_SWIZZLE_Z)),
 299                   ureg_imm1f(shader, 0.0f), ureg_imm1f(shader, 1.0f));
 300
 301       ureg_fixup_label(shader, label, ureg_get_instruction_number(shader));
 302       ureg_ENDIF(shader);
 303    }
 304
 305    ureg_release_temporary(shader, t_vtex);
 306    ureg_release_temporary(shader, t_vpos);
 307
 308    ureg_END(shader);
 309
 310    return ureg_create_shader_and_destroy(shader, r->pipe);
 311 }
 312
 313 static void *
 314 create_ycbcr_frag_shader(struct vl_mc *r, float scale, bool invert,
 315                          vl_mc_ycbcr_frag_shader fs_callback, void *callback_priv)
 316 {
 317    struct ureg_program *shader;
 318    struct ureg_src flags;
 319    struct ureg_dst tmp;
 320    struct ureg_dst fragment;
 321    unsigned label;
 322
 323    shader = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 324    if (!shader)
 325       return NULL;
 326
 327    flags = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_FLAGS, TGSI_INTERPOLATE_LINEAR);
 328
 329    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
 330
 331    tmp = calc_line(r->pipe->screen, shader);
 332
 333    /*
 334     * if (field == tc.w)
 335     *    kill();
 336     * else {
 337     *    fragment.xyz  = tex(tc, sampler) * scale + tc.z
 338     *    fragment.w = 1.0f
 339     * }
 340     */
 341
 342    ureg_SEQ(shader, ureg_writemask(tmp, TGSI_WRITEMASK_Y),
 343             ureg_scalar(flags, TGSI_SWIZZLE_W), ureg_src(tmp));
 344
 345    ureg_IF(shader, ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_Y), &label);
 346
 347       ureg_KILL(shader);
 348
 349    ureg_fixup_label(shader, label, ureg_get_instruction_number(shader));
 350    ureg_ELSE(shader, &label);
 351
 352       fs_callback(callback_priv, r, shader, VS_O_VTEX, tmp);
 353
 354       if (scale != 1.0f)
 355          ureg_MAD(shader, ureg_writemask(tmp, TGSI_WRITEMASK_XYZ),
 356                   ureg_src(tmp), ureg_imm1f(shader, scale),
 357                   ureg_scalar(flags, TGSI_SWIZZLE_Z));
 358       else
 359          ureg_ADD(shader, ureg_writemask(tmp, TGSI_WRITEMASK_XYZ),
 360                   ureg_src(tmp), ureg_scalar(flags, TGSI_SWIZZLE_Z));
 361
 362       ureg_MUL(shader, ureg_writemask(fragment, TGSI_WRITEMASK_XYZ), ureg_src(tmp), ureg_imm1f(shader, invert ? -1.0f : 1.0f));
 363       ureg_MOV(shader, ureg_writemask(fragment, TGSI_WRITEMASK_W), ureg_imm1f(shader, 1.0f));
 364
 365    ureg_fixup_label(shader, label, ureg_get_instruction_number(shader));
 366    ureg_ENDIF(shader);
 367
 368    ureg_release_temporary(shader, tmp);
 369
 370    ureg_END(shader);
 371
 372    return ureg_create_shader_and_destroy(shader, r->pipe);
 373 }
 374
 375 static bool
 376 init_pipe_state(struct vl_mc *r)
 377 {
 378    struct pipe_sampler_state sampler;
 379    struct pipe_blend_state blend;
 380    struct pipe_rasterizer_state rs_state;
 381    unsigned i;
 382
 383    assert(r);
 384
 385    memset(&sampler, 0, sizeof(sampler));
 386    sampler.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 387    sampler.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 388    sampler.wrap_r = PIPE_TEX_WRAP_CLAMP_TO_BORDER;
 389    sampler.min_img_filter = PIPE_TEX_FILTER_LINEAR;
 390    sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
 391    sampler.mag_img_filter = PIPE_TEX_FILTER_LINEAR;
 392    sampler.compare_mode = PIPE_TEX_COMPARE_NONE;
 393    sampler.compare_func = PIPE_FUNC_ALWAYS;
 394    sampler.normalized_coords = 1;
 395    r->sampler_ref = r->pipe->create_sampler_state(r->pipe, &sampler);
 396    if (!r->sampler_ref)
 397       goto error_sampler_ref;
 398
 399    for (i = 0; i < VL_MC_NUM_BLENDERS; ++i) {
 400       memset(&blend, 0, sizeof blend);
 401       blend.independent_blend_enable = 0;
 402       blend.rt[0].blend_enable = 1;
 403       blend.rt[0].rgb_func = PIPE_BLEND_ADD;
 404       blend.rt[0].rgb_src_factor = PIPE_BLENDFACTOR_SRC_ALPHA;
 405       blend.rt[0].rgb_dst_factor = PIPE_BLENDFACTOR_ZERO;
 406       blend.rt[0].alpha_func = PIPE_BLEND_ADD;
 407       blend.rt[0].alpha_src_factor = PIPE_BLENDFACTOR_SRC_ALPHA;
 408       blend.rt[0].alpha_dst_factor = PIPE_BLENDFACTOR_ZERO;
 409       blend.logicop_enable = 0;
 410       blend.logicop_func = PIPE_LOGICOP_CLEAR;
 411       blend.rt[0].colormask = i;
 412       blend.dither = 0;
 413       r->blend_clear[i] = r->pipe->create_blend_state(r->pipe, &blend);
 414       if (!r->blend_clear[i])
 415          goto error_blend;
 416
 417       blend.rt[0].rgb_dst_factor = PIPE_BLENDFACTOR_ONE;
 418       blend.rt[0].alpha_dst_factor = PIPE_BLENDFACTOR_ONE;
 419       r->blend_add[i] = r->pipe->create_blend_state(r->pipe, &blend);
 420       if (!r->blend_add[i])
 421          goto error_blend;
 422
 423       blend.rt[0].rgb_func = PIPE_BLEND_REVERSE_SUBTRACT;
 424       blend.rt[0].alpha_dst_factor = PIPE_BLEND_REVERSE_SUBTRACT;
 425       r->blend_sub[i] = r->pipe->create_blend_state(r->pipe, &blend);
 426       if (!r->blend_sub[i])
 427          goto error_blend;
 428    }
 429
 430    memset(&rs_state, 0, sizeof(rs_state));
 431    /*rs_state.sprite_coord_enable */
 432    rs_state.sprite_coord_mode = PIPE_SPRITE_COORD_UPPER_LEFT;
 433    rs_state.point_quad_rasterization = true;
 434    rs_state.point_size = VL_BLOCK_WIDTH;
 435    rs_state.half_pixel_center = true;
 436    rs_state.bottom_edge_rule = true;
 437    rs_state.depth_clip = 1;
 438    r->rs_state = r->pipe->create_rasterizer_state(r->pipe, &rs_state);
 439    if (!r->rs_state)
 440       goto error_rs_state;
 441
 442    return true;
 443
 444 error_rs_state:
 445 error_blend:
 446    for (i = 0; i < VL_MC_NUM_BLENDERS; ++i) {
 447       if (r->blend_sub[i])
 448          r->pipe->delete_blend_state(r->pipe, r->blend_sub[i]);
 449
 450       if (r->blend_add[i])
 451          r->pipe->delete_blend_state(r->pipe, r->blend_add[i]);
 452
 453       if (r->blend_clear[i])
 454          r->pipe->delete_blend_state(r->pipe, r->blend_clear[i]);
 455    }
 456
 457    r->pipe->delete_sampler_state(r->pipe, r->sampler_ref);
 458
 459 error_sampler_ref:
 460    return false;
 461 }
 462
 463 static void
 464 cleanup_pipe_state(struct vl_mc *r)
 465 {
 466    unsigned i;
 467
 468    assert(r);
 469
 470    r->pipe->delete_sampler_state(r->pipe, r->sampler_ref);
 471    for (i = 0; i < VL_MC_NUM_BLENDERS; ++i) {
 472       r->pipe->delete_blend_state(r->pipe, r->blend_clear[i]);
 473       r->pipe->delete_blend_state(r->pipe, r->blend_add[i]);
 474       r->pipe->delete_blend_state(r->pipe, r->blend_sub[i]);
 475    }
 476    r->pipe->delete_rasterizer_state(r->pipe, r->rs_state);
 477 }
 478
 479 bool
 480 vl_mc_init(struct vl_mc *renderer, struct pipe_context *pipe,
 481            unsigned buffer_width, unsigned buffer_height,
 482            unsigned macroblock_size, float scale,
 483            vl_mc_ycbcr_vert_shader vs_callback,
 484            vl_mc_ycbcr_frag_shader fs_callback,
 485            void *callback_priv)
 486 {
 487    assert(renderer);
 488    assert(pipe);
 489
 490    memset(renderer, 0, sizeof(struct vl_mc));
 491
 492    renderer->pipe = pipe;
 493    renderer->buffer_width = buffer_width;
 494    renderer->buffer_height = buffer_height;
 495    renderer->macroblock_size = macroblock_size;
 496
 497    if (!init_pipe_state(renderer))
 498       goto error_pipe_state;
 499
 500    renderer->vs_ref = create_ref_vert_shader(renderer);
 501    if (!renderer->vs_ref)
 502       goto error_vs_ref;
 503
 504    renderer->vs_ycbcr = create_ycbcr_vert_shader(renderer, vs_callback, callback_priv);
 505    if (!renderer->vs_ycbcr)
 506       goto error_vs_ycbcr;
 507
 508    renderer->fs_ref = create_ref_frag_shader(renderer);
 509    if (!renderer->fs_ref)
 510       goto error_fs_ref;
 511
 512    renderer->fs_ycbcr = create_ycbcr_frag_shader(renderer, scale, false, fs_callback, callback_priv);
 513    if (!renderer->fs_ycbcr)
 514       goto error_fs_ycbcr;
 515
 516    renderer->fs_ycbcr_sub = create_ycbcr_frag_shader(renderer, scale, true, fs_callback, callback_priv);
 517    if (!renderer->fs_ycbcr_sub)
 518       goto error_fs_ycbcr_sub;
 519
 520    return true;
 521
 522 error_fs_ycbcr_sub:
 523    renderer->pipe->delete_fs_state(renderer->pipe, renderer->fs_ycbcr);
 524
 525 error_fs_ycbcr:
 526    renderer->pipe->delete_fs_state(renderer->pipe, renderer->fs_ref);
 527
 528 error_fs_ref:
 529    renderer->pipe->delete_vs_state(renderer->pipe, renderer->vs_ycbcr);
 530
 531 error_vs_ycbcr:
 532    renderer->pipe->delete_vs_state(renderer->pipe, renderer->vs_ref);
 533
 534 error_vs_ref:
 535    cleanup_pipe_state(renderer);
 536
 537 error_pipe_state:
 538    return false;
 539 }
 540
 541 void
 542 vl_mc_cleanup(struct vl_mc *renderer)
 543 {
 544    assert(renderer);
 545
 546    cleanup_pipe_state(renderer);
 547
 548    renderer->pipe->delete_vs_state(renderer->pipe, renderer->vs_ref);
 549    renderer->pipe->delete_vs_state(renderer->pipe, renderer->vs_ycbcr);
 550    renderer->pipe->delete_fs_state(renderer->pipe, renderer->fs_ref);
 551    renderer->pipe->delete_fs_state(renderer->pipe, renderer->fs_ycbcr);
 552    renderer->pipe->delete_fs_state(renderer->pipe, renderer->fs_ycbcr_sub);
 553 }
 554
 555 bool
 556 vl_mc_init_buffer(struct vl_mc *renderer, struct vl_mc_buffer *buffer)
 557 {
 558    assert(renderer && buffer);
 559
 560    buffer->viewport.scale[2] = 1;
 561    buffer->viewport.translate[0] = 0;
 562    buffer->viewport.translate[1] = 0;
 563    buffer->viewport.translate[2] = 0;
 564
 565    buffer->fb_state.nr_cbufs = 1;
 566    buffer->fb_state.zsbuf = NULL;
 567
 568    return true;
 569 }
 570
 571 void
 572 vl_mc_cleanup_buffer(struct vl_mc_buffer *buffer)
 573 {
 574    assert(buffer);
 575 }
 576
 577 void
 578 vl_mc_set_surface(struct vl_mc_buffer *buffer, struct pipe_surface *surface)
 579 {
 580    assert(buffer && surface);
 581
 582    buffer->surface_cleared = false;
 583
 584    buffer->viewport.scale[0] = surface->width;
 585    buffer->viewport.scale[1] = surface->height;
 586
 587    buffer->fb_state.width = surface->width;
 588    buffer->fb_state.height = surface->height;
 589    buffer->fb_state.cbufs[0] = surface;
 590 }
 591
 592 static void
 593 prepare_pipe_4_rendering(struct vl_mc *renderer, struct vl_mc_buffer *buffer, unsigned mask)
 594 {
 595    assert(buffer);
 596
 597    renderer->pipe->bind_rasterizer_state(renderer->pipe, renderer->rs_state);
 598
 599    if (buffer->surface_cleared)
 600       renderer->pipe->bind_blend_state(renderer->pipe, renderer->blend_add[mask]);
 601    else
 602       renderer->pipe->bind_blend_state(renderer->pipe, renderer->blend_clear[mask]);
 603
 604    renderer->pipe->set_framebuffer_state(renderer->pipe, &buffer->fb_state);
 605    renderer->pipe->set_viewport_states(renderer->pipe, 0, 1, &buffer->viewport);
 606 }
 607
 608 void
 609 vl_mc_render_ref(struct vl_mc *renderer, struct vl_mc_buffer *buffer, struct pipe_sampler_view *ref)
 610 {
 611    assert(buffer && ref);
 612
 613    prepare_pipe_4_rendering(renderer, buffer, PIPE_MASK_R | PIPE_MASK_G | PIPE_MASK_B);
 614
 615    renderer->pipe->bind_vs_state(renderer->pipe, renderer->vs_ref);
 616    renderer->pipe->bind_fs_state(renderer->pipe, renderer->fs_ref);
 617
 618    renderer->pipe->set_sampler_views(renderer->pipe, PIPE_SHADER_FRAGMENT,
 619                                      0, 1, &ref);
 620    renderer->pipe->bind_sampler_states(renderer->pipe, PIPE_SHADER_FRAGMENT,
 621                                        0, 1, &renderer->sampler_ref);
 622
 623    util_draw_arrays_instanced(renderer->pipe, PIPE_PRIM_QUADS, 0, 4, 0,
 624                               renderer->buffer_width / VL_MACROBLOCK_WIDTH *
 625                               renderer->buffer_height / VL_MACROBLOCK_HEIGHT);
 626
 627    buffer->surface_cleared = true;
 628 }
 629
 630 void
 631 vl_mc_render_ycbcr(struct vl_mc *renderer, struct vl_mc_buffer *buffer, unsigned component, unsigned num_instances)
 632 {
 633    unsigned mask = 1 << component;
 634
 635    assert(buffer);
 636
 637    if (num_instances == 0)
 638       return;
 639
 640    prepare_pipe_4_rendering(renderer, buffer, mask);
 641
 642    renderer->pipe->bind_vs_state(renderer->pipe, renderer->vs_ycbcr);
 643    renderer->pipe->bind_fs_state(renderer->pipe, renderer->fs_ycbcr);
 644
 645    util_draw_arrays_instanced(renderer->pipe, PIPE_PRIM_QUADS, 0, 4, 0, num_instances);
 646
 647    if (buffer->surface_cleared) {
 648       renderer->pipe->bind_blend_state(renderer->pipe, renderer->blend_sub[mask]);
 649       renderer->pipe->bind_fs_state(renderer->pipe, renderer->fs_ycbcr_sub);
 650       util_draw_arrays_instanced(renderer->pipe, PIPE_PRIM_QUADS, 0, 4, 0, num_instances);
 651    }
 652 }