Home | History | Annotate | Download | only in vl
      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_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 #include "util/u_upload_mgr.h"
     37 
     38 #include "tgsi/tgsi_ureg.h"
     39 
     40 #include "vl_csc.h"
     41 #include "vl_types.h"
     42 #include "vl_compositor.h"
     43 
     44 #define MIN_DIRTY (0)
     45 #define MAX_DIRTY (1 << 15)
     46 
     47 enum VS_OUTPUT
     48 {
     49    VS_O_VPOS = 0,
     50    VS_O_COLOR = 0,
     51    VS_O_VTEX = 0,
     52    VS_O_VTOP,
     53    VS_O_VBOTTOM,
     54 };
     55 
     56 static void *
     57 create_vert_shader(struct vl_compositor *c)
     58 {
     59    struct ureg_program *shader;
     60    struct ureg_src vpos, vtex, color;
     61    struct ureg_dst tmp;
     62    struct ureg_dst o_vpos, o_vtex, o_color;
     63    struct ureg_dst o_vtop, o_vbottom;
     64 
     65    shader = ureg_create(PIPE_SHADER_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    color = ureg_DECL_vs_input(shader, 2);
     72    tmp = ureg_DECL_temporary(shader);
     73    o_vpos = ureg_DECL_output(shader, TGSI_SEMANTIC_POSITION, VS_O_VPOS);
     74    o_color = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, VS_O_COLOR);
     75    o_vtex = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX);
     76    o_vtop = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTOP);
     77    o_vbottom = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VBOTTOM);
     78 
     79    /*
     80     * o_vpos = vpos
     81     * o_vtex = vtex
     82     * o_color = color
     83     */
     84    ureg_MOV(shader, o_vpos, vpos);
     85    ureg_MOV(shader, o_vtex, vtex);
     86    ureg_MOV(shader, o_color, color);
     87 
     88    /*
     89     * tmp.x = vtex.w / 2
     90     * tmp.y = vtex.w / 4
     91     *
     92     * o_vtop.x = vtex.x
     93     * o_vtop.y = vtex.y * tmp.x + 0.25f
     94     * o_vtop.z = vtex.y * tmp.y + 0.25f
     95     * o_vtop.w = 1 / tmp.x
     96     *
     97     * o_vbottom.x = vtex.x
     98     * o_vbottom.y = vtex.y * tmp.x - 0.25f
     99     * o_vbottom.z = vtex.y * tmp.y - 0.25f
    100     * o_vbottom.w = 1 / tmp.y
    101     */
    102    ureg_MUL(shader, ureg_writemask(tmp, TGSI_WRITEMASK_X),
    103             ureg_scalar(vtex, TGSI_SWIZZLE_W), ureg_imm1f(shader, 0.5f));
    104    ureg_MUL(shader, ureg_writemask(tmp, TGSI_WRITEMASK_Y),
    105             ureg_scalar(vtex, TGSI_SWIZZLE_W), ureg_imm1f(shader, 0.25f));
    106 
    107    ureg_MOV(shader, ureg_writemask(o_vtop, TGSI_WRITEMASK_X), vtex);
    108    ureg_MAD(shader, ureg_writemask(o_vtop, TGSI_WRITEMASK_Y), ureg_scalar(vtex, TGSI_SWIZZLE_Y),
    109             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_X), ureg_imm1f(shader, 0.25f));
    110    ureg_MAD(shader, ureg_writemask(o_vtop, TGSI_WRITEMASK_Z), ureg_scalar(vtex, TGSI_SWIZZLE_Y),
    111             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_Y), ureg_imm1f(shader, 0.25f));
    112    ureg_RCP(shader, ureg_writemask(o_vtop, TGSI_WRITEMASK_W),
    113             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_X));
    114 
    115    ureg_MOV(shader, ureg_writemask(o_vbottom, TGSI_WRITEMASK_X), vtex);
    116    ureg_MAD(shader, ureg_writemask(o_vbottom, TGSI_WRITEMASK_Y), ureg_scalar(vtex, TGSI_SWIZZLE_Y),
    117             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_X), ureg_imm1f(shader, -0.25f));
    118    ureg_MAD(shader, ureg_writemask(o_vbottom, TGSI_WRITEMASK_Z), ureg_scalar(vtex, TGSI_SWIZZLE_Y),
    119             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_Y), ureg_imm1f(shader, -0.25f));
    120    ureg_RCP(shader, ureg_writemask(o_vbottom, TGSI_WRITEMASK_W),
    121             ureg_scalar(ureg_src(tmp), TGSI_SWIZZLE_Y));
    122 
    123    ureg_END(shader);
    124 
    125    return ureg_create_shader_and_destroy(shader, c->pipe);
    126 }
    127 
    128 static void
    129 create_frag_shader_weave(struct ureg_program *shader, struct ureg_dst fragment)
    130 {
    131    struct ureg_src i_tc[2];
    132    struct ureg_src sampler[3];
    133    struct ureg_dst t_tc[2];
    134    struct ureg_dst t_texel[2];
    135    unsigned i, j;
    136 
    137    i_tc[0] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTOP, TGSI_INTERPOLATE_LINEAR);
    138    i_tc[1] = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VBOTTOM, TGSI_INTERPOLATE_LINEAR);
    139 
    140    for (i = 0; i < 3; ++i)
    141       sampler[i] = ureg_DECL_sampler(shader, i);
    142 
    143    for (i = 0; i < 2; ++i) {
    144       t_tc[i] = ureg_DECL_temporary(shader);
    145       t_texel[i] = ureg_DECL_temporary(shader);
    146    }
    147 
    148    /* calculate the texture offsets
    149     * t_tc.x = i_tc.x
    150     * t_tc.y = (round(i_tc.y - 0.5) + 0.5) / height * 2
    151     */
    152    for (i = 0; i < 2; ++i) {
    153       ureg_MOV(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_X), i_tc[i]);
    154       ureg_ADD(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_YZ),
    155                i_tc[i], ureg_imm1f(shader, -0.5f));
    156       ureg_ROUND(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_YZ), ureg_src(t_tc[i]));
    157       ureg_MOV(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_W),
    158                ureg_imm1f(shader, i ? 1.0f : 0.0f));
    159       ureg_ADD(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_YZ),
    160                ureg_src(t_tc[i]), ureg_imm1f(shader, 0.5f));
    161       ureg_MUL(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_Y),
    162                ureg_src(t_tc[i]), ureg_scalar(i_tc[0], TGSI_SWIZZLE_W));
    163       ureg_MUL(shader, ureg_writemask(t_tc[i], TGSI_WRITEMASK_Z),
    164                ureg_src(t_tc[i]), ureg_scalar(i_tc[1], TGSI_SWIZZLE_W));
    165    }
    166 
    167    /* fetch the texels
    168     * texel[0..1].x = tex(t_tc[0..1][0])
    169     * texel[0..1].y = tex(t_tc[0..1][1])
    170     * texel[0..1].z = tex(t_tc[0..1][2])
    171     */
    172    for (i = 0; i < 2; ++i)
    173       for (j = 0; j < 3; ++j) {
    174          struct ureg_src src = ureg_swizzle(ureg_src(t_tc[i]),
    175             TGSI_SWIZZLE_X, j ? TGSI_SWIZZLE_Z : TGSI_SWIZZLE_Y, TGSI_SWIZZLE_W, TGSI_SWIZZLE_W);
    176 
    177          ureg_TEX(shader, ureg_writemask(t_texel[i], TGSI_WRITEMASK_X << j),
    178                   TGSI_TEXTURE_2D_ARRAY, src, sampler[j]);
    179       }
    180 
    181    /* calculate linear interpolation factor
    182     * factor = |round(i_tc.y) - i_tc.y| * 2
    183     */
    184    ureg_ROUND(shader, ureg_writemask(t_tc[0], TGSI_WRITEMASK_YZ), i_tc[0]);
    185    ureg_ADD(shader, ureg_writemask(t_tc[0], TGSI_WRITEMASK_YZ),
    186             ureg_src(t_tc[0]), ureg_negate(i_tc[0]));
    187    ureg_MUL(shader, ureg_writemask(t_tc[0], TGSI_WRITEMASK_YZ),
    188             ureg_abs(ureg_src(t_tc[0])), ureg_imm1f(shader, 2.0f));
    189    ureg_LRP(shader, fragment, ureg_swizzle(ureg_src(t_tc[0]),
    190             TGSI_SWIZZLE_Y, TGSI_SWIZZLE_Z, TGSI_SWIZZLE_Z, TGSI_SWIZZLE_Z),
    191             ureg_src(t_texel[0]), ureg_src(t_texel[1]));
    192 
    193    for (i = 0; i < 2; ++i) {
    194       ureg_release_temporary(shader, t_texel[i]);
    195       ureg_release_temporary(shader, t_tc[i]);
    196    }
    197 }
    198 
    199 static void
    200 create_frag_shader_csc(struct ureg_program *shader, struct ureg_dst texel,
    201 		       struct ureg_dst fragment)
    202 {
    203    struct ureg_src csc[3];
    204    struct ureg_src lumakey;
    205    struct ureg_dst temp[2];
    206    unsigned i;
    207 
    208    for (i = 0; i < 3; ++i)
    209       csc[i] = ureg_DECL_constant(shader, i);
    210 
    211    lumakey = ureg_DECL_constant(shader, 3);
    212 
    213    for (i = 0; i < 2; ++i)
    214       temp[i] = ureg_DECL_temporary(shader);
    215 
    216    ureg_MOV(shader, ureg_writemask(texel, TGSI_WRITEMASK_W),
    217 	    ureg_imm1f(shader, 1.0f));
    218 
    219    for (i = 0; i < 3; ++i)
    220       ureg_DP4(shader, ureg_writemask(fragment, TGSI_WRITEMASK_X << i), csc[i],
    221 	       ureg_src(texel));
    222 
    223    ureg_MOV(shader, ureg_writemask(temp[0], TGSI_WRITEMASK_W),
    224             ureg_scalar(ureg_src(texel), TGSI_SWIZZLE_Z));
    225    ureg_SLE(shader, ureg_writemask(temp[1],TGSI_WRITEMASK_W),
    226             ureg_src(temp[0]), ureg_scalar(lumakey, TGSI_SWIZZLE_X));
    227    ureg_SGT(shader, ureg_writemask(temp[0],TGSI_WRITEMASK_W),
    228             ureg_src(temp[0]), ureg_scalar(lumakey, TGSI_SWIZZLE_Y));
    229    ureg_MAX(shader, ureg_writemask(fragment, TGSI_WRITEMASK_W),
    230             ureg_src(temp[0]), ureg_src(temp[1]));
    231 
    232    for (i = 0; i < 2; ++i)
    233        ureg_release_temporary(shader, temp[i]);
    234 }
    235 
    236 static void *
    237 create_frag_shader_video_buffer(struct vl_compositor *c)
    238 {
    239    struct ureg_program *shader;
    240    struct ureg_src tc;
    241    struct ureg_src sampler[3];
    242    struct ureg_dst texel;
    243    struct ureg_dst fragment;
    244    unsigned i;
    245 
    246    shader = ureg_create(PIPE_SHADER_FRAGMENT);
    247    if (!shader)
    248       return false;
    249 
    250    tc = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX, TGSI_INTERPOLATE_LINEAR);
    251    for (i = 0; i < 3; ++i)
    252       sampler[i] = ureg_DECL_sampler(shader, i);
    253 
    254    texel = ureg_DECL_temporary(shader);
    255    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
    256 
    257    /*
    258     * texel.xyz = tex(tc, sampler[i])
    259     * fragment = csc * texel
    260     */
    261    for (i = 0; i < 3; ++i)
    262       ureg_TEX(shader, ureg_writemask(texel, TGSI_WRITEMASK_X << i), TGSI_TEXTURE_2D_ARRAY, tc, sampler[i]);
    263 
    264    create_frag_shader_csc(shader, texel, fragment);
    265 
    266    ureg_release_temporary(shader, texel);
    267    ureg_END(shader);
    268 
    269    return ureg_create_shader_and_destroy(shader, c->pipe);
    270 }
    271 
    272 static void *
    273 create_frag_shader_weave_rgb(struct vl_compositor *c)
    274 {
    275    struct ureg_program *shader;
    276    struct ureg_dst texel, fragment;
    277 
    278    shader = ureg_create(PIPE_SHADER_FRAGMENT);
    279    if (!shader)
    280       return false;
    281 
    282    texel = ureg_DECL_temporary(shader);
    283    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
    284 
    285    create_frag_shader_weave(shader, texel);
    286    create_frag_shader_csc(shader, texel, fragment);
    287 
    288    ureg_release_temporary(shader, texel);
    289 
    290    ureg_END(shader);
    291 
    292    return ureg_create_shader_and_destroy(shader, c->pipe);
    293 }
    294 
    295 static void *
    296 create_frag_shader_weave_yuv(struct vl_compositor *c, bool y)
    297 {
    298    struct ureg_program *shader;
    299    struct ureg_dst texel, fragment;
    300 
    301    shader = ureg_create(PIPE_SHADER_FRAGMENT);
    302    if (!shader)
    303       return false;
    304 
    305    texel = ureg_DECL_temporary(shader);
    306    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
    307 
    308    create_frag_shader_weave(shader, texel);
    309 
    310    if (y)
    311       ureg_MOV(shader, ureg_writemask(fragment, TGSI_WRITEMASK_X), ureg_src(texel));
    312    else
    313       ureg_MOV(shader, ureg_writemask(fragment, TGSI_WRITEMASK_XY),
    314                        ureg_swizzle(ureg_src(texel), TGSI_SWIZZLE_Y,
    315                                TGSI_SWIZZLE_Z, TGSI_SWIZZLE_W, TGSI_SWIZZLE_W));
    316 
    317    ureg_release_temporary(shader, texel);
    318 
    319    ureg_END(shader);
    320 
    321    return ureg_create_shader_and_destroy(shader, c->pipe);
    322 }
    323 
    324 static void *
    325 create_frag_shader_palette(struct vl_compositor *c, bool include_cc)
    326 {
    327    struct ureg_program *shader;
    328    struct ureg_src csc[3];
    329    struct ureg_src tc;
    330    struct ureg_src sampler;
    331    struct ureg_src palette;
    332    struct ureg_dst texel;
    333    struct ureg_dst fragment;
    334    unsigned i;
    335 
    336    shader = ureg_create(PIPE_SHADER_FRAGMENT);
    337    if (!shader)
    338       return false;
    339 
    340    for (i = 0; include_cc && i < 3; ++i)
    341       csc[i] = ureg_DECL_constant(shader, i);
    342 
    343    tc = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX, TGSI_INTERPOLATE_LINEAR);
    344    sampler = ureg_DECL_sampler(shader, 0);
    345    palette = ureg_DECL_sampler(shader, 1);
    346 
    347    texel = ureg_DECL_temporary(shader);
    348    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
    349 
    350    /*
    351     * texel = tex(tc, sampler)
    352     * fragment.xyz = tex(texel, palette) * csc
    353     * fragment.a = texel.a
    354     */
    355    ureg_TEX(shader, texel, TGSI_TEXTURE_2D, tc, sampler);
    356    ureg_MOV(shader, ureg_writemask(fragment, TGSI_WRITEMASK_W), ureg_src(texel));
    357 
    358    if (include_cc) {
    359       ureg_TEX(shader, texel, TGSI_TEXTURE_1D, ureg_src(texel), palette);
    360       for (i = 0; i < 3; ++i)
    361          ureg_DP4(shader, ureg_writemask(fragment, TGSI_WRITEMASK_X << i), csc[i], ureg_src(texel));
    362    } else {
    363       ureg_TEX(shader, ureg_writemask(fragment, TGSI_WRITEMASK_XYZ),
    364                TGSI_TEXTURE_1D, ureg_src(texel), palette);
    365    }
    366 
    367    ureg_release_temporary(shader, texel);
    368    ureg_END(shader);
    369 
    370    return ureg_create_shader_and_destroy(shader, c->pipe);
    371 }
    372 
    373 static void *
    374 create_frag_shader_rgba(struct vl_compositor *c)
    375 {
    376    struct ureg_program *shader;
    377    struct ureg_src tc, color, sampler;
    378    struct ureg_dst texel, fragment;
    379 
    380    shader = ureg_create(PIPE_SHADER_FRAGMENT);
    381    if (!shader)
    382       return false;
    383 
    384    tc = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX, TGSI_INTERPOLATE_LINEAR);
    385    color = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_COLOR, VS_O_COLOR, TGSI_INTERPOLATE_LINEAR);
    386    sampler = ureg_DECL_sampler(shader, 0);
    387    texel = ureg_DECL_temporary(shader);
    388    fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
    389 
    390    /*
    391     * fragment = tex(tc, sampler)
    392     */
    393    ureg_TEX(shader, texel, TGSI_TEXTURE_2D, tc, sampler);
    394    ureg_MUL(shader, fragment, ureg_src(texel), color);
    395    ureg_END(shader);
    396 
    397    return ureg_create_shader_and_destroy(shader, c->pipe);
    398 }
    399 
    400 static bool
    401 init_shaders(struct vl_compositor *c)
    402 {
    403    assert(c);
    404 
    405    c->vs = create_vert_shader(c);
    406    if (!c->vs) {
    407       debug_printf("Unable to create vertex shader.\n");
    408       return false;
    409    }
    410 
    411    c->fs_video_buffer = create_frag_shader_video_buffer(c);
    412    if (!c->fs_video_buffer) {
    413       debug_printf("Unable to create YCbCr-to-RGB fragment shader.\n");
    414       return false;
    415    }
    416 
    417    c->fs_weave_rgb = create_frag_shader_weave_rgb(c);
    418    if (!c->fs_weave_rgb) {
    419       debug_printf("Unable to create YCbCr-to-RGB weave fragment shader.\n");
    420       return false;
    421    }
    422 
    423    c->fs_weave_yuv.y = create_frag_shader_weave_yuv(c, true);
    424    c->fs_weave_yuv.uv = create_frag_shader_weave_yuv(c, false);
    425    if (!c->fs_weave_yuv.y || !c->fs_weave_yuv.uv) {
    426       debug_printf("Unable to create YCbCr i-to-YCbCr p weave fragment shader.\n");
    427       return false;
    428    }
    429 
    430    c->fs_palette.yuv = create_frag_shader_palette(c, true);
    431    if (!c->fs_palette.yuv) {
    432       debug_printf("Unable to create YUV-Palette-to-RGB fragment shader.\n");
    433       return false;
    434    }
    435 
    436    c->fs_palette.rgb = create_frag_shader_palette(c, false);
    437    if (!c->fs_palette.rgb) {
    438       debug_printf("Unable to create RGB-Palette-to-RGB fragment shader.\n");
    439       return false;
    440    }
    441 
    442    c->fs_rgba = create_frag_shader_rgba(c);
    443    if (!c->fs_rgba) {
    444       debug_printf("Unable to create RGB-to-RGB fragment shader.\n");
    445       return false;
    446    }
    447 
    448    return true;
    449 }
    450 
    451 static void cleanup_shaders(struct vl_compositor *c)
    452 {
    453    assert(c);
    454 
    455    c->pipe->delete_vs_state(c->pipe, c->vs);
    456    c->pipe->delete_fs_state(c->pipe, c->fs_video_buffer);
    457    c->pipe->delete_fs_state(c->pipe, c->fs_weave_rgb);
    458    c->pipe->delete_fs_state(c->pipe, c->fs_weave_yuv.y);
    459    c->pipe->delete_fs_state(c->pipe, c->fs_weave_yuv.uv);
    460    c->pipe->delete_fs_state(c->pipe, c->fs_palette.yuv);
    461    c->pipe->delete_fs_state(c->pipe, c->fs_palette.rgb);
    462    c->pipe->delete_fs_state(c->pipe, c->fs_rgba);
    463 }
    464 
    465 static bool
    466 init_pipe_state(struct vl_compositor *c)
    467 {
    468    struct pipe_rasterizer_state rast;
    469    struct pipe_sampler_state sampler;
    470    struct pipe_blend_state blend;
    471    struct pipe_depth_stencil_alpha_state dsa;
    472    unsigned i;
    473 
    474    assert(c);
    475 
    476    c->fb_state.nr_cbufs = 1;
    477    c->fb_state.zsbuf = NULL;
    478 
    479    memset(&sampler, 0, sizeof(sampler));
    480    sampler.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
    481    sampler.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
    482    sampler.wrap_r = PIPE_TEX_WRAP_REPEAT;
    483    sampler.min_img_filter = PIPE_TEX_FILTER_LINEAR;
    484    sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
    485    sampler.mag_img_filter = PIPE_TEX_FILTER_LINEAR;
    486    sampler.compare_mode = PIPE_TEX_COMPARE_NONE;
    487    sampler.compare_func = PIPE_FUNC_ALWAYS;
    488    sampler.normalized_coords = 1;
    489 
    490    c->sampler_linear = c->pipe->create_sampler_state(c->pipe, &sampler);
    491 
    492    sampler.min_img_filter = PIPE_TEX_FILTER_NEAREST;
    493    sampler.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
    494    c->sampler_nearest = c->pipe->create_sampler_state(c->pipe, &sampler);
    495 
    496    memset(&blend, 0, sizeof blend);
    497    blend.independent_blend_enable = 0;
    498    blend.rt[0].blend_enable = 0;
    499    blend.logicop_enable = 0;
    500    blend.logicop_func = PIPE_LOGICOP_CLEAR;
    501    blend.rt[0].colormask = PIPE_MASK_RGBA;
    502    blend.dither = 0;
    503    c->blend_clear = c->pipe->create_blend_state(c->pipe, &blend);
    504 
    505    blend.rt[0].blend_enable = 1;
    506    blend.rt[0].rgb_func = PIPE_BLEND_ADD;
    507    blend.rt[0].rgb_src_factor = PIPE_BLENDFACTOR_SRC_ALPHA;
    508    blend.rt[0].rgb_dst_factor = PIPE_BLENDFACTOR_INV_SRC_ALPHA;
    509    blend.rt[0].alpha_func = PIPE_BLEND_ADD;
    510    blend.rt[0].alpha_src_factor = PIPE_BLENDFACTOR_ONE;
    511    blend.rt[0].alpha_dst_factor = PIPE_BLENDFACTOR_ONE;
    512    c->blend_add = c->pipe->create_blend_state(c->pipe, &blend);
    513 
    514    memset(&rast, 0, sizeof rast);
    515    rast.flatshade = 0;
    516    rast.front_ccw = 1;
    517    rast.cull_face = PIPE_FACE_NONE;
    518    rast.fill_back = PIPE_POLYGON_MODE_FILL;
    519    rast.fill_front = PIPE_POLYGON_MODE_FILL;
    520    rast.scissor = 1;
    521    rast.line_width = 1;
    522    rast.point_size_per_vertex = 1;
    523    rast.offset_units = 1;
    524    rast.offset_scale = 1;
    525    rast.half_pixel_center = 1;
    526    rast.bottom_edge_rule = 1;
    527    rast.depth_clip = 1;
    528 
    529    c->rast = c->pipe->create_rasterizer_state(c->pipe, &rast);
    530 
    531    memset(&dsa, 0, sizeof dsa);
    532    dsa.depth.enabled = 0;
    533    dsa.depth.writemask = 0;
    534    dsa.depth.func = PIPE_FUNC_ALWAYS;
    535    for (i = 0; i < 2; ++i) {
    536       dsa.stencil[i].enabled = 0;
    537       dsa.stencil[i].func = PIPE_FUNC_ALWAYS;
    538       dsa.stencil[i].fail_op = PIPE_STENCIL_OP_KEEP;
    539       dsa.stencil[i].zpass_op = PIPE_STENCIL_OP_KEEP;
    540       dsa.stencil[i].zfail_op = PIPE_STENCIL_OP_KEEP;
    541       dsa.stencil[i].valuemask = 0;
    542       dsa.stencil[i].writemask = 0;
    543    }
    544    dsa.alpha.enabled = 0;
    545    dsa.alpha.func = PIPE_FUNC_ALWAYS;
    546    dsa.alpha.ref_value = 0;
    547    c->dsa = c->pipe->create_depth_stencil_alpha_state(c->pipe, &dsa);
    548    c->pipe->bind_depth_stencil_alpha_state(c->pipe, c->dsa);
    549 
    550    return true;
    551 }
    552 
    553 static void cleanup_pipe_state(struct vl_compositor *c)
    554 {
    555    assert(c);
    556 
    557    /* Asserted in softpipe_delete_fs_state() for some reason */
    558    c->pipe->bind_vs_state(c->pipe, NULL);
    559    c->pipe->bind_fs_state(c->pipe, NULL);
    560 
    561    c->pipe->delete_depth_stencil_alpha_state(c->pipe, c->dsa);
    562    c->pipe->delete_sampler_state(c->pipe, c->sampler_linear);
    563    c->pipe->delete_sampler_state(c->pipe, c->sampler_nearest);
    564    c->pipe->delete_blend_state(c->pipe, c->blend_clear);
    565    c->pipe->delete_blend_state(c->pipe, c->blend_add);
    566    c->pipe->delete_rasterizer_state(c->pipe, c->rast);
    567 }
    568 
    569 static bool
    570 init_buffers(struct vl_compositor *c)
    571 {
    572    struct pipe_vertex_element vertex_elems[3];
    573 
    574    assert(c);
    575 
    576    /*
    577     * Create our vertex buffer and vertex buffer elements
    578     */
    579    c->vertex_buf.stride = sizeof(struct vertex2f) + sizeof(struct vertex4f) * 2;
    580    c->vertex_buf.buffer_offset = 0;
    581    c->vertex_buf.buffer = NULL;
    582 
    583    vertex_elems[0].src_offset = 0;
    584    vertex_elems[0].instance_divisor = 0;
    585    vertex_elems[0].vertex_buffer_index = 0;
    586    vertex_elems[0].src_format = PIPE_FORMAT_R32G32_FLOAT;
    587    vertex_elems[1].src_offset = sizeof(struct vertex2f);
    588    vertex_elems[1].instance_divisor = 0;
    589    vertex_elems[1].vertex_buffer_index = 0;
    590    vertex_elems[1].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
    591    vertex_elems[2].src_offset = sizeof(struct vertex2f) + sizeof(struct vertex4f);
    592    vertex_elems[2].instance_divisor = 0;
    593    vertex_elems[2].vertex_buffer_index = 0;
    594    vertex_elems[2].src_format = PIPE_FORMAT_R32G32B32A32_FLOAT;
    595    c->vertex_elems_state = c->pipe->create_vertex_elements_state(c->pipe, 3, vertex_elems);
    596 
    597    return true;
    598 }
    599 
    600 static void
    601 cleanup_buffers(struct vl_compositor *c)
    602 {
    603    assert(c);
    604 
    605    c->pipe->delete_vertex_elements_state(c->pipe, c->vertex_elems_state);
    606    pipe_resource_reference(&c->vertex_buf.buffer, NULL);
    607 }
    608 
    609 static inline struct u_rect
    610 default_rect(struct vl_compositor_layer *layer)
    611 {
    612    struct pipe_resource *res = layer->sampler_views[0]->texture;
    613    struct u_rect rect = { 0, res->width0, 0, res->height0 * res->array_size };
    614    return rect;
    615 }
    616 
    617 static inline struct vertex2f
    618 calc_topleft(struct vertex2f size, struct u_rect rect)
    619 {
    620    struct vertex2f res = { rect.x0 / size.x, rect.y0 / size.y };
    621    return res;
    622 }
    623 
    624 static inline struct vertex2f
    625 calc_bottomright(struct vertex2f size, struct u_rect rect)
    626 {
    627    struct vertex2f res = { rect.x1 / size.x, rect.y1 / size.y };
    628    return res;
    629 }
    630 
    631 static inline void
    632 calc_src_and_dst(struct vl_compositor_layer *layer, unsigned width, unsigned height,
    633                  struct u_rect src, struct u_rect dst)
    634 {
    635    struct vertex2f size =  { width, height };
    636 
    637    layer->src.tl = calc_topleft(size, src);
    638    layer->src.br = calc_bottomright(size, src);
    639    layer->dst.tl = calc_topleft(size, dst);
    640    layer->dst.br = calc_bottomright(size, dst);
    641    layer->zw.x = 0.0f;
    642    layer->zw.y = size.y;
    643 }
    644 
    645 static void
    646 gen_rect_verts(struct vertex2f *vb, struct vl_compositor_layer *layer)
    647 {
    648    struct vertex2f tl, tr, br, bl;
    649 
    650    assert(vb && layer);
    651 
    652    switch (layer->rotate) {
    653    default:
    654    case VL_COMPOSITOR_ROTATE_0:
    655       tl = layer->dst.tl;
    656       tr.x = layer->dst.br.x;
    657       tr.y = layer->dst.tl.y;
    658       br = layer->dst.br;
    659       bl.x = layer->dst.tl.x;
    660       bl.y = layer->dst.br.y;
    661       break;
    662    case VL_COMPOSITOR_ROTATE_90:
    663       tl.x = layer->dst.br.x;
    664       tl.y = layer->dst.tl.y;
    665       tr = layer->dst.br;
    666       br.x = layer->dst.tl.x;
    667       br.y = layer->dst.br.y;
    668       bl = layer->dst.tl;
    669       break;
    670    case VL_COMPOSITOR_ROTATE_180:
    671       tl = layer->dst.br;
    672       tr.x = layer->dst.tl.x;
    673       tr.y = layer->dst.br.y;
    674       br = layer->dst.tl;
    675       bl.x = layer->dst.br.x;
    676       bl.y = layer->dst.tl.y;
    677       break;
    678    case VL_COMPOSITOR_ROTATE_270:
    679       tl.x = layer->dst.tl.x;
    680       tl.y = layer->dst.br.y;
    681       tr = layer->dst.tl;
    682       br.x = layer->dst.br.x;
    683       br.y = layer->dst.tl.y;
    684       bl = layer->dst.br;
    685       break;
    686    }
    687 
    688    vb[ 0].x = tl.x;
    689    vb[ 0].y = tl.y;
    690    vb[ 1].x = layer->src.tl.x;
    691    vb[ 1].y = layer->src.tl.y;
    692    vb[ 2] = layer->zw;
    693    vb[ 3].x = layer->colors[0].x;
    694    vb[ 3].y = layer->colors[0].y;
    695    vb[ 4].x = layer->colors[0].z;
    696    vb[ 4].y = layer->colors[0].w;
    697 
    698    vb[ 5].x = tr.x;
    699    vb[ 5].y = tr.y;
    700    vb[ 6].x = layer->src.br.x;
    701    vb[ 6].y = layer->src.tl.y;
    702    vb[ 7] = layer->zw;
    703    vb[ 8].x = layer->colors[1].x;
    704    vb[ 8].y = layer->colors[1].y;
    705    vb[ 9].x = layer->colors[1].z;
    706    vb[ 9].y = layer->colors[1].w;
    707 
    708    vb[10].x = br.x;
    709    vb[10].y = br.y;
    710    vb[11].x = layer->src.br.x;
    711    vb[11].y = layer->src.br.y;
    712    vb[12] = layer->zw;
    713    vb[13].x = layer->colors[2].x;
    714    vb[13].y = layer->colors[2].y;
    715    vb[14].x = layer->colors[2].z;
    716    vb[14].y = layer->colors[2].w;
    717 
    718    vb[15].x = bl.x;
    719    vb[15].y = bl.y;
    720    vb[16].x = layer->src.tl.x;
    721    vb[16].y = layer->src.br.y;
    722    vb[17] = layer->zw;
    723    vb[18].x = layer->colors[3].x;
    724    vb[18].y = layer->colors[3].y;
    725    vb[19].x = layer->colors[3].z;
    726    vb[19].y = layer->colors[3].w;
    727 }
    728 
    729 static inline struct u_rect
    730 calc_drawn_area(struct vl_compositor_state *s, struct vl_compositor_layer *layer)
    731 {
    732    struct vertex2f tl, br;
    733    struct u_rect result;
    734 
    735    assert(s && layer);
    736 
    737    // rotate
    738    switch (layer->rotate) {
    739    default:
    740    case VL_COMPOSITOR_ROTATE_0:
    741       tl = layer->dst.tl;
    742       br = layer->dst.br;
    743       break;
    744    case VL_COMPOSITOR_ROTATE_90:
    745       tl.x = layer->dst.br.x;
    746       tl.y = layer->dst.tl.y;
    747       br.x = layer->dst.tl.x;
    748       br.y = layer->dst.br.y;
    749       break;
    750    case VL_COMPOSITOR_ROTATE_180:
    751       tl = layer->dst.br;
    752       br = layer->dst.tl;
    753       break;
    754    case VL_COMPOSITOR_ROTATE_270:
    755       tl.x = layer->dst.tl.x;
    756       tl.y = layer->dst.br.y;
    757       br.x = layer->dst.br.x;
    758       br.y = layer->dst.tl.y;
    759       break;
    760    }
    761 
    762    // scale
    763    result.x0 = tl.x * layer->viewport.scale[0] + layer->viewport.translate[0];
    764    result.y0 = tl.y * layer->viewport.scale[1] + layer->viewport.translate[1];
    765    result.x1 = br.x * layer->viewport.scale[0] + layer->viewport.translate[0];
    766    result.y1 = br.y * layer->viewport.scale[1] + layer->viewport.translate[1];
    767 
    768    // and clip
    769    result.x0 = MAX2(result.x0, s->scissor.minx);
    770    result.y0 = MAX2(result.y0, s->scissor.miny);
    771    result.x1 = MIN2(result.x1, s->scissor.maxx);
    772    result.y1 = MIN2(result.y1, s->scissor.maxy);
    773    return result;
    774 }
    775 
    776 static void
    777 gen_vertex_data(struct vl_compositor *c, struct vl_compositor_state *s, struct u_rect *dirty)
    778 {
    779    struct vertex2f *vb;
    780    unsigned i;
    781 
    782    assert(c);
    783 
    784    /* Allocate new memory for vertices. */
    785    u_upload_alloc(c->upload, 0,
    786                   c->vertex_buf.stride * VL_COMPOSITOR_MAX_LAYERS * 4, /* size */
    787                   4, /* alignment */
    788                   &c->vertex_buf.buffer_offset, &c->vertex_buf.buffer,
    789                   (void**)&vb);
    790 
    791    for (i = 0; i < VL_COMPOSITOR_MAX_LAYERS; i++) {
    792       if (s->used_layers & (1 << i)) {
    793          struct vl_compositor_layer *layer = &s->layers[i];
    794          gen_rect_verts(vb, layer);
    795          vb += 20;
    796 
    797          if (!layer->viewport_valid) {
    798             layer->viewport.scale[0] = c->fb_state.width;
    799             layer->viewport.scale[1] = c->fb_state.height;
    800             layer->viewport.translate[0] = 0;
    801             layer->viewport.translate[1] = 0;
    802          }
    803 
    804          if (dirty && layer->clearing) {
    805             struct u_rect drawn = calc_drawn_area(s, layer);
    806             if (
    807              dirty->x0 >= drawn.x0 &&
    808              dirty->y0 >= drawn.y0 &&
    809              dirty->x1 <= drawn.x1 &&
    810              dirty->y1 <= drawn.y1) {
    811 
    812                // We clear the dirty area anyway, no need for clear_render_target
    813                dirty->x0 = dirty->y0 = MAX_DIRTY;
    814                dirty->x1 = dirty->y1 = MIN_DIRTY;
    815             }
    816          }
    817       }
    818    }
    819 
    820    u_upload_unmap(c->upload);
    821 }
    822 
    823 static void
    824 draw_layers(struct vl_compositor *c, struct vl_compositor_state *s, struct u_rect *dirty)
    825 {
    826    unsigned vb_index, i;
    827 
    828    assert(c);
    829 
    830    for (i = 0, vb_index = 0; i < VL_COMPOSITOR_MAX_LAYERS; ++i) {
    831       if (s->used_layers & (1 << i)) {
    832          struct vl_compositor_layer *layer = &s->layers[i];
    833          struct pipe_sampler_view **samplers = &layer->sampler_views[0];
    834          unsigned num_sampler_views = !samplers[1] ? 1 : !samplers[2] ? 2 : 3;
    835          void *blend = layer->blend ? layer->blend : i ? c->blend_add : c->blend_clear;
    836 
    837          c->pipe->bind_blend_state(c->pipe, blend);
    838          c->pipe->set_viewport_states(c->pipe, 0, 1, &layer->viewport);
    839          c->pipe->bind_fs_state(c->pipe, layer->fs);
    840          c->pipe->bind_sampler_states(c->pipe, PIPE_SHADER_FRAGMENT, 0,
    841                                       num_sampler_views, layer->samplers);
    842          c->pipe->set_sampler_views(c->pipe, PIPE_SHADER_FRAGMENT, 0,
    843                                     num_sampler_views, samplers);
    844 
    845          util_draw_arrays(c->pipe, PIPE_PRIM_QUADS, vb_index * 4, 4);
    846          vb_index++;
    847 
    848          if (dirty) {
    849             // Remember the currently drawn area as dirty for the next draw command
    850             struct u_rect drawn = calc_drawn_area(s, layer);
    851             dirty->x0 = MIN2(drawn.x0, dirty->x0);
    852             dirty->y0 = MIN2(drawn.y0, dirty->y0);
    853             dirty->x1 = MAX2(drawn.x1, dirty->x1);
    854             dirty->y1 = MAX2(drawn.y1, dirty->y1);
    855          }
    856       }
    857    }
    858 }
    859 
    860 void
    861 vl_compositor_reset_dirty_area(struct u_rect *dirty)
    862 {
    863    assert(dirty);
    864 
    865    dirty->x0 = dirty->y0 = MIN_DIRTY;
    866    dirty->x1 = dirty->y1 = MAX_DIRTY;
    867 }
    868 
    869 void
    870 vl_compositor_set_clear_color(struct vl_compositor_state *s, union pipe_color_union *color)
    871 {
    872    assert(s);
    873    assert(color);
    874 
    875    s->clear_color = *color;
    876 }
    877 
    878 void
    879 vl_compositor_get_clear_color(struct vl_compositor_state *s, union pipe_color_union *color)
    880 {
    881    assert(s);
    882    assert(color);
    883 
    884    *color = s->clear_color;
    885 }
    886 
    887 void
    888 vl_compositor_clear_layers(struct vl_compositor_state *s)
    889 {
    890    unsigned i, j;
    891 
    892    assert(s);
    893 
    894    s->used_layers = 0;
    895    for ( i = 0; i < VL_COMPOSITOR_MAX_LAYERS; ++i) {
    896       struct vertex4f v_one = { 1.0f, 1.0f, 1.0f, 1.0f };
    897       s->layers[i].clearing = i ? false : true;
    898       s->layers[i].blend = NULL;
    899       s->layers[i].fs = NULL;
    900       s->layers[i].viewport.scale[2] = 1;
    901       s->layers[i].viewport.translate[2] = 0;
    902       s->layers[i].rotate = VL_COMPOSITOR_ROTATE_0;
    903 
    904       for ( j = 0; j < 3; j++)
    905          pipe_sampler_view_reference(&s->layers[i].sampler_views[j], NULL);
    906       for ( j = 0; j < 4; ++j)
    907          s->layers[i].colors[j] = v_one;
    908    }
    909 }
    910 
    911 void
    912 vl_compositor_cleanup(struct vl_compositor *c)
    913 {
    914    assert(c);
    915 
    916    u_upload_destroy(c->upload);
    917    cleanup_buffers(c);
    918    cleanup_shaders(c);
    919    cleanup_pipe_state(c);
    920 }
    921 
    922 bool
    923 vl_compositor_set_csc_matrix(struct vl_compositor_state *s,
    924                              vl_csc_matrix const *matrix,
    925                              float luma_min, float luma_max)
    926 {
    927    struct pipe_transfer *buf_transfer;
    928 
    929    assert(s);
    930 
    931    float *ptr = pipe_buffer_map(s->pipe, s->csc_matrix,
    932                                PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE,
    933                                &buf_transfer);
    934 
    935    if (!ptr)
    936       return false;
    937 
    938    memcpy(ptr, matrix, sizeof(vl_csc_matrix));
    939 
    940    ptr += sizeof(vl_csc_matrix)/sizeof(float);
    941    ptr[0] = luma_min;
    942    ptr[1] = luma_max;
    943 
    944    pipe_buffer_unmap(s->pipe, buf_transfer);
    945 
    946    return true;
    947 }
    948 
    949 void
    950 vl_compositor_set_dst_clip(struct vl_compositor_state *s, struct u_rect *dst_clip)
    951 {
    952    assert(s);
    953 
    954    s->scissor_valid = dst_clip != NULL;
    955    if (dst_clip) {
    956       s->scissor.minx = dst_clip->x0;
    957       s->scissor.miny = dst_clip->y0;
    958       s->scissor.maxx = dst_clip->x1;
    959       s->scissor.maxy = dst_clip->y1;
    960    }
    961 }
    962 
    963 void
    964 vl_compositor_set_layer_blend(struct vl_compositor_state *s,
    965                               unsigned layer, void *blend,
    966                               bool is_clearing)
    967 {
    968    assert(s && blend);
    969 
    970    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
    971 
    972    s->layers[layer].clearing = is_clearing;
    973    s->layers[layer].blend = blend;
    974 }
    975 
    976 void
    977 vl_compositor_set_layer_dst_area(struct vl_compositor_state *s,
    978                                  unsigned layer, struct u_rect *dst_area)
    979 {
    980    assert(s);
    981 
    982    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
    983 
    984    s->layers[layer].viewport_valid = dst_area != NULL;
    985    if (dst_area) {
    986       s->layers[layer].viewport.scale[0] = dst_area->x1 - dst_area->x0;
    987       s->layers[layer].viewport.scale[1] = dst_area->y1 - dst_area->y0;
    988       s->layers[layer].viewport.translate[0] = dst_area->x0;
    989       s->layers[layer].viewport.translate[1] = dst_area->y0;
    990    }
    991 }
    992 
    993 void
    994 vl_compositor_set_buffer_layer(struct vl_compositor_state *s,
    995                                struct vl_compositor *c,
    996                                unsigned layer,
    997                                struct pipe_video_buffer *buffer,
    998                                struct u_rect *src_rect,
    999                                struct u_rect *dst_rect,
   1000                                enum vl_compositor_deinterlace deinterlace)
   1001 {
   1002    struct pipe_sampler_view **sampler_views;
   1003    unsigned i;
   1004 
   1005    assert(s && c && buffer);
   1006 
   1007    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
   1008 
   1009    s->used_layers |= 1 << layer;
   1010    sampler_views = buffer->get_sampler_view_components(buffer);
   1011    for (i = 0; i < 3; ++i) {
   1012       s->layers[layer].samplers[i] = c->sampler_linear;
   1013       pipe_sampler_view_reference(&s->layers[layer].sampler_views[i], sampler_views[i]);
   1014    }
   1015 
   1016    calc_src_and_dst(&s->layers[layer], buffer->width, buffer->height,
   1017                     src_rect ? *src_rect : default_rect(&s->layers[layer]),
   1018                     dst_rect ? *dst_rect : default_rect(&s->layers[layer]));
   1019 
   1020    if (buffer->interlaced) {
   1021       float half_a_line = 0.5f / s->layers[layer].zw.y;
   1022       switch(deinterlace) {
   1023       case VL_COMPOSITOR_WEAVE:
   1024          s->layers[layer].fs = c->fs_weave_rgb;
   1025          break;
   1026 
   1027       case VL_COMPOSITOR_BOB_TOP:
   1028          s->layers[layer].zw.x = 0.0f;
   1029          s->layers[layer].src.tl.y += half_a_line;
   1030          s->layers[layer].src.br.y += half_a_line;
   1031          s->layers[layer].fs = c->fs_video_buffer;
   1032          break;
   1033 
   1034       case VL_COMPOSITOR_BOB_BOTTOM:
   1035          s->layers[layer].zw.x = 1.0f;
   1036          s->layers[layer].src.tl.y -= half_a_line;
   1037          s->layers[layer].src.br.y -= half_a_line;
   1038          s->layers[layer].fs = c->fs_video_buffer;
   1039          break;
   1040       }
   1041 
   1042    } else
   1043       s->layers[layer].fs = c->fs_video_buffer;
   1044 }
   1045 
   1046 void
   1047 vl_compositor_set_palette_layer(struct vl_compositor_state *s,
   1048                                 struct vl_compositor *c,
   1049                                 unsigned layer,
   1050                                 struct pipe_sampler_view *indexes,
   1051                                 struct pipe_sampler_view *palette,
   1052                                 struct u_rect *src_rect,
   1053                                 struct u_rect *dst_rect,
   1054                                 bool include_color_conversion)
   1055 {
   1056    assert(s && c && indexes && palette);
   1057 
   1058    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
   1059 
   1060    s->used_layers |= 1 << layer;
   1061 
   1062    s->layers[layer].fs = include_color_conversion ?
   1063       c->fs_palette.yuv : c->fs_palette.rgb;
   1064 
   1065    s->layers[layer].samplers[0] = c->sampler_linear;
   1066    s->layers[layer].samplers[1] = c->sampler_nearest;
   1067    s->layers[layer].samplers[2] = NULL;
   1068    pipe_sampler_view_reference(&s->layers[layer].sampler_views[0], indexes);
   1069    pipe_sampler_view_reference(&s->layers[layer].sampler_views[1], palette);
   1070    pipe_sampler_view_reference(&s->layers[layer].sampler_views[2], NULL);
   1071    calc_src_and_dst(&s->layers[layer], indexes->texture->width0, indexes->texture->height0,
   1072                     src_rect ? *src_rect : default_rect(&s->layers[layer]),
   1073                     dst_rect ? *dst_rect : default_rect(&s->layers[layer]));
   1074 }
   1075 
   1076 void
   1077 vl_compositor_set_rgba_layer(struct vl_compositor_state *s,
   1078                              struct vl_compositor *c,
   1079                              unsigned layer,
   1080                              struct pipe_sampler_view *rgba,
   1081                              struct u_rect *src_rect,
   1082                              struct u_rect *dst_rect,
   1083                              struct vertex4f *colors)
   1084 {
   1085    unsigned i;
   1086 
   1087    assert(s && c && rgba);
   1088 
   1089    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
   1090 
   1091    s->used_layers |= 1 << layer;
   1092    s->layers[layer].fs = c->fs_rgba;
   1093    s->layers[layer].samplers[0] = c->sampler_linear;
   1094    s->layers[layer].samplers[1] = NULL;
   1095    s->layers[layer].samplers[2] = NULL;
   1096    pipe_sampler_view_reference(&s->layers[layer].sampler_views[0], rgba);
   1097    pipe_sampler_view_reference(&s->layers[layer].sampler_views[1], NULL);
   1098    pipe_sampler_view_reference(&s->layers[layer].sampler_views[2], NULL);
   1099    calc_src_and_dst(&s->layers[layer], rgba->texture->width0, rgba->texture->height0,
   1100                     src_rect ? *src_rect : default_rect(&s->layers[layer]),
   1101                     dst_rect ? *dst_rect : default_rect(&s->layers[layer]));
   1102 
   1103    if (colors)
   1104       for (i = 0; i < 4; ++i)
   1105          s->layers[layer].colors[i] = colors[i];
   1106 }
   1107 
   1108 void
   1109 vl_compositor_set_layer_rotation(struct vl_compositor_state *s,
   1110                                  unsigned layer,
   1111                                  enum vl_compositor_rotation rotate)
   1112 {
   1113    assert(s);
   1114    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
   1115    s->layers[layer].rotate = rotate;
   1116 }
   1117 
   1118 void
   1119 vl_compositor_set_yuv_layer(struct vl_compositor_state *s,
   1120                             struct vl_compositor *c,
   1121                             unsigned layer,
   1122                             struct pipe_video_buffer *buffer,
   1123                             struct u_rect *src_rect,
   1124                             struct u_rect *dst_rect,
   1125                             bool y)
   1126 {
   1127    struct pipe_sampler_view **sampler_views;
   1128    unsigned i;
   1129 
   1130    assert(s && c && buffer);
   1131 
   1132    assert(layer < VL_COMPOSITOR_MAX_LAYERS);
   1133 
   1134    s->used_layers |= 1 << layer;
   1135    sampler_views = buffer->get_sampler_view_components(buffer);
   1136    for (i = 0; i < 3; ++i) {
   1137       s->layers[layer].samplers[i] = c->sampler_linear;
   1138       pipe_sampler_view_reference(&s->layers[layer].sampler_views[i], sampler_views[i]);
   1139    }
   1140 
   1141    calc_src_and_dst(&s->layers[layer], buffer->width, buffer->height,
   1142                     src_rect ? *src_rect : default_rect(&s->layers[layer]),
   1143                     dst_rect ? *dst_rect : default_rect(&s->layers[layer]));
   1144 
   1145    s->layers[layer].fs = (y) ? c->fs_weave_yuv.y : c->fs_weave_yuv.uv;
   1146 }
   1147 
   1148 void
   1149 vl_compositor_render(struct vl_compositor_state *s,
   1150                      struct vl_compositor       *c,
   1151                      struct pipe_surface        *dst_surface,
   1152                      struct u_rect              *dirty_area,
   1153                      bool                        clear_dirty)
   1154 {
   1155    assert(c);
   1156    assert(dst_surface);
   1157 
   1158    c->fb_state.width = dst_surface->width;
   1159    c->fb_state.height = dst_surface->height;
   1160    c->fb_state.cbufs[0] = dst_surface;
   1161 
   1162    if (!s->scissor_valid) {
   1163       s->scissor.minx = 0;
   1164       s->scissor.miny = 0;
   1165       s->scissor.maxx = dst_surface->width;
   1166       s->scissor.maxy = dst_surface->height;
   1167    }
   1168    c->pipe->set_scissor_states(c->pipe, 0, 1, &s->scissor);
   1169 
   1170    gen_vertex_data(c, s, dirty_area);
   1171 
   1172    if (clear_dirty && dirty_area &&
   1173        (dirty_area->x0 < dirty_area->x1 || dirty_area->y0 < dirty_area->y1)) {
   1174 
   1175       c->pipe->clear_render_target(c->pipe, dst_surface, &s->clear_color,
   1176                                    0, 0, dst_surface->width, dst_surface->height, false);
   1177       dirty_area->x0 = dirty_area->y0 = MAX_DIRTY;
   1178       dirty_area->x1 = dirty_area->y1 = MIN_DIRTY;
   1179    }
   1180 
   1181    c->pipe->set_framebuffer_state(c->pipe, &c->fb_state);
   1182    c->pipe->bind_vs_state(c->pipe, c->vs);
   1183    c->pipe->set_vertex_buffers(c->pipe, 0, 1, &c->vertex_buf);
   1184    c->pipe->bind_vertex_elements_state(c->pipe, c->vertex_elems_state);
   1185    pipe_set_constant_buffer(c->pipe, PIPE_SHADER_FRAGMENT, 0, s->csc_matrix);
   1186    c->pipe->bind_rasterizer_state(c->pipe, c->rast);
   1187 
   1188    draw_layers(c, s, dirty_area);
   1189 }
   1190 
   1191 bool
   1192 vl_compositor_init(struct vl_compositor *c, struct pipe_context *pipe)
   1193 {
   1194    assert(c);
   1195 
   1196    memset(c, 0, sizeof(*c));
   1197 
   1198    c->pipe = pipe;
   1199 
   1200    c->upload = u_upload_create(pipe, 128 * 1024, PIPE_BIND_VERTEX_BUFFER,
   1201                                PIPE_USAGE_STREAM);
   1202 
   1203    if (!c->upload)
   1204       return false;
   1205 
   1206    if (!init_pipe_state(c)) {
   1207       u_upload_destroy(c->upload);
   1208       return false;
   1209    }
   1210 
   1211    if (!init_shaders(c)) {
   1212       u_upload_destroy(c->upload);
   1213       cleanup_pipe_state(c);
   1214       return false;
   1215    }
   1216 
   1217    if (!init_buffers(c)) {
   1218       u_upload_destroy(c->upload);
   1219       cleanup_shaders(c);
   1220       cleanup_pipe_state(c);
   1221       return false;
   1222    }
   1223 
   1224    return true;
   1225 }
   1226 
   1227 bool
   1228 vl_compositor_init_state(struct vl_compositor_state *s, struct pipe_context *pipe)
   1229 {
   1230    vl_csc_matrix csc_matrix;
   1231 
   1232    assert(s);
   1233 
   1234    memset(s, 0, sizeof(*s));
   1235 
   1236    s->pipe = pipe;
   1237 
   1238    s->clear_color.f[0] = s->clear_color.f[1] = 0.0f;
   1239    s->clear_color.f[2] = s->clear_color.f[3] = 0.0f;
   1240 
   1241    /*
   1242     * Create our fragment shader's constant buffer
   1243     * Const buffer contains the color conversion matrix and bias vectors
   1244     */
   1245    /* XXX: Create with IMMUTABLE/STATIC... although it does change every once in a long while... */
   1246    s->csc_matrix = pipe_buffer_create
   1247    (
   1248       pipe->screen,
   1249       PIPE_BIND_CONSTANT_BUFFER,
   1250       PIPE_USAGE_DEFAULT,
   1251       sizeof(csc_matrix) + 2*sizeof(float)
   1252    );
   1253 
   1254    if (!s->csc_matrix)
   1255       return false;
   1256 
   1257    vl_compositor_clear_layers(s);
   1258 
   1259    vl_csc_get_matrix(VL_CSC_COLOR_STANDARD_IDENTITY, NULL, true, &csc_matrix);
   1260    if (!vl_compositor_set_csc_matrix(s, (const vl_csc_matrix *)&csc_matrix, 1.0f, 0.0f))
   1261       return false;
   1262 
   1263    return true;
   1264 }
   1265 
   1266 void
   1267 vl_compositor_cleanup_state(struct vl_compositor_state *s)
   1268 {
   1269    assert(s);
   1270 
   1271    vl_compositor_clear_layers(s);
   1272    pipe_resource_reference(&s->csc_matrix, NULL);
   1273 }
   1274