src/gallium/auxiliary/vl/vl_bicubic_filter.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2016 Nayan Deshmukh.
   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 <stdio.h>
  29
  30 #include "pipe/p_context.h"
  31
  32 #include "tgsi/tgsi_ureg.h"
  33
  34 #include "util/u_draw.h"
  35 #include "util/u_memory.h"
  36 #include "util/u_math.h"
  37 #include "util/u_rect.h"
  38
  39 #include "vl_types.h"
  40 #include "vl_vertex_buffers.h"
  41 #include "vl_bicubic_filter.h"
  42
  43 enum VS_OUTPUT
  44 {
  45    VS_O_VPOS = 0,
  46    VS_O_VTEX = 0
  47 };
  48
  49 static void *
  50 create_vert_shader(struct vl_bicubic_filter *filter)
  51 {
  52    struct ureg_program *shader;
  53    struct ureg_src i_vpos;
  54    struct ureg_dst o_vpos, o_vtex;
  55
  56    shader = ureg_create(PIPE_SHADER_VERTEX);
  57    if (!shader)
  58       return NULL;
  59
  60    i_vpos = ureg_DECL_vs_input(shader, 0);
  61    o_vpos = ureg_DECL_output(shader, TGSI_SEMANTIC_POSITION, VS_O_VPOS);
  62    o_vtex = ureg_DECL_output(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX);
  63
  64    ureg_MOV(shader, o_vpos, i_vpos);
  65    ureg_MOV(shader, o_vtex, i_vpos);
  66
  67    ureg_END(shader);
  68
  69    return ureg_create_shader_and_destroy(shader, filter->pipe);
  70 }
  71
  72 static void
  73 create_frag_shader_cubic_interpolater(struct ureg_program *shader, struct ureg_src tex_a,
  74                                       struct ureg_src tex_b, struct ureg_src tex_c,
  75                                       struct ureg_src tex_d, struct ureg_src t,
  76                                       struct ureg_dst o_fragment)
  77 {
  78    struct ureg_dst temp[11];
  79    struct ureg_dst t_2;
  80    unsigned i;
  81
  82    for(i = 0; i < 11; ++i)
  83        temp[i] = ureg_DECL_temporary(shader);
  84    t_2 = ureg_DECL_temporary(shader);
  85
  86    /*
  87     * |temp[0]|   |  0  2  0  0 |  |tex_a|
  88     * |temp[1]| = | -1  0  1  0 |* |tex_b|
  89     * |temp[2]|   |  2 -5  4 -1 |  |tex_c|
  90     * |temp[3]|   | -1  3 -3  1 |  |tex_d|
  91     */
  92    ureg_MUL(shader, temp[0], tex_b, ureg_imm1f(shader, 2.0f));
  93
  94    ureg_MUL(shader, temp[1], tex_a, ureg_imm1f(shader, -1.0f));
  95    ureg_MAD(shader, temp[1], tex_c, ureg_imm1f(shader, 1.0f),
  96             ureg_src(temp[1]));
  97
  98    ureg_MUL(shader, temp[2], tex_a, ureg_imm1f(shader, 2.0f));
  99    ureg_MAD(shader, temp[2], tex_b, ureg_imm1f(shader, -5.0f),
 100             ureg_src(temp[2]));
 101    ureg_MAD(shader, temp[2], tex_c, ureg_imm1f(shader, 4.0f),
 102             ureg_src(temp[2]));
 103    ureg_MAD(shader, temp[2], tex_d, ureg_imm1f(shader, -1.0f),
 104              ureg_src(temp[2]));
 105
 106    ureg_MUL(shader, temp[3], tex_a, ureg_imm1f(shader, -1.0f));
 107    ureg_MAD(shader, temp[3], tex_b, ureg_imm1f(shader, 3.0f),
 108             ureg_src(temp[3]));
 109    ureg_MAD(shader, temp[3], tex_c, ureg_imm1f(shader, -3.0f),
 110             ureg_src(temp[3]));
 111    ureg_MAD(shader, temp[3], tex_d, ureg_imm1f(shader, 1.0f),
 112             ureg_src(temp[3]));
 113
 114    /*
 115     * t_2 = t*t
 116     * o_fragment = 0.5*|1  t  t^2  t^3|*|temp[0]|
 117     *                                   |temp[1]|
 118     *                                   |temp[2]|
 119     *                                   |temp[3]|
 120     */
 121
 122    ureg_MUL(shader, t_2, t, t);
 123    ureg_MUL(shader, temp[4], ureg_src(t_2), t);
 124
 125    ureg_MUL(shader, temp[4], ureg_src(temp[4]),
 126             ureg_src(temp[3]));
 127    ureg_MUL(shader, temp[5], ureg_src(t_2),
 128             ureg_src(temp[2]));
 129    ureg_MUL(shader, temp[6], t,
 130             ureg_src(temp[1]));
 131    ureg_MUL(shader, temp[7], ureg_imm1f(shader, 1.0f),
 132             ureg_src(temp[0]));
 133    ureg_ADD(shader, temp[8], ureg_src(temp[4]),
 134             ureg_src(temp[5]));
 135    ureg_ADD(shader, temp[9], ureg_src(temp[6]),
 136             ureg_src(temp[7]));
 137
 138    ureg_ADD(shader, temp[10], ureg_src(temp[8]),
 139             ureg_src(temp[9]));
 140    ureg_MUL(shader, o_fragment, ureg_src(temp[10]),
 141             ureg_imm1f(shader, 0.5f));
 142
 143
 144    for(i = 0; i < 11; ++i)
 145        ureg_release_temporary(shader, temp[i]);
 146    ureg_release_temporary(shader, t_2);
 147 }
 148
 149 static void *
 150 create_frag_shader(struct vl_bicubic_filter *filter, unsigned video_width,
 151                    unsigned video_height, struct vertex2f *offsets)
 152 {
 153    struct pipe_screen *screen = filter->pipe->screen;
 154    struct ureg_program *shader;
 155    struct ureg_src i_vtex, vtex;
 156    struct ureg_src sampler;
 157    struct ureg_src half_pixel;
 158    struct ureg_dst t_array[23];
 159    struct ureg_dst o_fragment;
 160    struct ureg_dst t;
 161    unsigned i;
 162
 163    if (screen->get_shader_param(
 164       screen, PIPE_SHADER_FRAGMENT, PIPE_SHADER_CAP_MAX_TEMPS) < 23) {
 165
 166       return NULL;
 167    }
 168
 169    shader = ureg_create(PIPE_SHADER_FRAGMENT);
 170    if (!shader) {
 171       return NULL;
 172    }
 173
 174    i_vtex = ureg_DECL_fs_input(shader, TGSI_SEMANTIC_GENERIC, VS_O_VTEX, TGSI_INTERPOLATE_LINEAR);
 175    sampler = ureg_DECL_sampler(shader, 0);
 176
 177    for (i = 0; i < 23; ++i)
 178       t_array[i] = ureg_DECL_temporary(shader);
 179    t = ureg_DECL_temporary(shader);
 180
 181    half_pixel = ureg_DECL_constant(shader, 0);
 182    o_fragment = ureg_DECL_output(shader, TGSI_SEMANTIC_COLOR, 0);
 183
 184    /*
 185     * temp = (i_vtex - (0.5/dst_size)) * i_size)
 186     * t = frac(temp)
 187     * vtex = floor(i_vtex)/i_size
 188     */
 189    ureg_SUB(shader, ureg_writemask(t_array[21], TGSI_WRITEMASK_XY),
 190             i_vtex, half_pixel);
 191    ureg_MUL(shader, ureg_writemask(t_array[22], TGSI_WRITEMASK_XY),
 192             ureg_src(t_array[21]), ureg_imm2f(shader, video_width, video_height));
 193    ureg_FRC(shader, ureg_writemask(t, TGSI_WRITEMASK_XY),
 194             ureg_src(t_array[22]));
 195
 196    ureg_FLR(shader, ureg_writemask(t_array[22], TGSI_WRITEMASK_XY),
 197             ureg_src(t_array[22]));
 198    ureg_DIV(shader, ureg_writemask(t_array[22], TGSI_WRITEMASK_XY),
 199             ureg_src(t_array[22]), ureg_imm2f(shader, video_width, video_height));
 200    ureg_ADD(shader, ureg_writemask(t_array[22], TGSI_WRITEMASK_XY),
 201             ureg_src(t_array[22]), half_pixel);
 202
 203    /*
 204     * t_array[0..*] = vtex + offset[0..*]
 205     * t_array[0..*] = tex(t_array[0..*], sampler)
 206     * t_array[16+i] = cubic_interpolate(t_array[4*i..4*i+3], t_x)
 207     * o_fragment = cubic_interpolate(t_array[16..19], t_y)
 208     */
 209    vtex = ureg_src(t_array[22]);
 210    for (i = 0; i < 16; ++i) {
 211         ureg_ADD(shader, ureg_writemask(t_array[i], TGSI_WRITEMASK_XY),
 212                   vtex, ureg_imm2f(shader, offsets[i].x, offsets[i].y));
 213         ureg_MOV(shader, ureg_writemask(t_array[i], TGSI_WRITEMASK_ZW),
 214                   ureg_imm1f(shader, 0.0f));
 215    }
 216
 217    for (i = 0; i < 16; ++i) {
 218       ureg_TEX(shader, t_array[i], TGSI_TEXTURE_2D, ureg_src(t_array[i]), sampler);
 219    }
 220
 221    for(i = 0; i < 4; ++i)
 222       create_frag_shader_cubic_interpolater(shader, ureg_src(t_array[4*i]),
 223               ureg_src(t_array[4*i+1]), ureg_src(t_array[4*i+2]), ureg_src(t_array[4*i+3]),
 224               ureg_scalar(ureg_src(t), TGSI_SWIZZLE_X), t_array[16+i]);
 225
 226    create_frag_shader_cubic_interpolater(shader, ureg_src(t_array[16]),
 227             ureg_src(t_array[17]), ureg_src(t_array[18]), ureg_src(t_array[19]),
 228             ureg_scalar(ureg_src(t), TGSI_SWIZZLE_Y), o_fragment);
 229
 230    for(i = 0; i < 23; ++i)
 231        ureg_release_temporary(shader, t_array[i]);
 232    ureg_release_temporary(shader, t);
 233
 234    ureg_END(shader);
 235
 236    return ureg_create_shader_and_destroy(shader, filter->pipe);
 237 }
 238
 239 bool
 240 vl_bicubic_filter_init(struct vl_bicubic_filter *filter, struct pipe_context *pipe,
 241                       unsigned width, unsigned height)
 242 {
 243    struct pipe_rasterizer_state rs_state;
 244    struct pipe_blend_state blend;
 245    struct vertex2f offsets[16];
 246    struct pipe_sampler_state sampler;
 247    struct pipe_vertex_element ve;
 248    unsigned i;
 249
 250    assert(filter && pipe);
 251    assert(width && height);
 252
 253    memset(filter, 0, sizeof(*filter));
 254    filter->pipe = pipe;
 255
 256    memset(&rs_state, 0, sizeof(rs_state));
 257    rs_state.half_pixel_center = true;
 258    rs_state.bottom_edge_rule = true;
 259    rs_state.depth_clip = 1;
 260    filter->rs_state = pipe->create_rasterizer_state(pipe, &rs_state);
 261    if (!filter->rs_state)
 262       goto error_rs_state;
 263
 264    memset(&blend, 0, sizeof blend);
 265    blend.rt[0].rgb_func = PIPE_BLEND_ADD;
 266    blend.rt[0].rgb_src_factor = PIPE_BLENDFACTOR_ONE;
 267    blend.rt[0].rgb_dst_factor = PIPE_BLENDFACTOR_ONE;
 268    blend.rt[0].alpha_func = PIPE_BLEND_ADD;
 269    blend.rt[0].alpha_src_factor = PIPE_BLENDFACTOR_ONE;
 270    blend.rt[0].alpha_dst_factor = PIPE_BLENDFACTOR_ONE;
 271    blend.logicop_func = PIPE_LOGICOP_CLEAR;
 272    blend.rt[0].colormask = PIPE_MASK_RGBA;
 273    filter->blend = pipe->create_blend_state(pipe, &blend);
 274    if (!filter->blend)
 275       goto error_blend;
 276
 277    memset(&sampler, 0, sizeof(sampler));
 278    sampler.wrap_s = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 279    sampler.wrap_t = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 280    sampler.wrap_r = PIPE_TEX_WRAP_CLAMP_TO_EDGE;
 281    sampler.min_img_filter = PIPE_TEX_FILTER_NEAREST;
 282    sampler.min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
 283    sampler.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
 284    sampler.compare_mode = PIPE_TEX_COMPARE_NONE;
 285    sampler.compare_func = PIPE_FUNC_ALWAYS;
 286    sampler.normalized_coords = 1;
 287    filter->sampler = pipe->create_sampler_state(pipe, &sampler);
 288    if (!filter->sampler)
 289       goto error_sampler;
 290
 291    filter->quad = vl_vb_upload_quads(pipe);
 292    if(!filter->quad.buffer)
 293       goto error_quad;
 294
 295    memset(&ve, 0, sizeof(ve));
 296    ve.src_offset = 0;
 297    ve.instance_divisor = 0;
 298    ve.vertex_buffer_index = 0;
 299    ve.src_format = PIPE_FORMAT_R32G32_FLOAT;
 300    filter->ves = pipe->create_vertex_elements_state(pipe, 1, &ve);
 301    if (!filter->ves)
 302       goto error_ves;
 303
 304    offsets[0].x = -1.0f; offsets[0].y = -1.0f;
 305    offsets[1].x = 0.0f; offsets[1].y = -1.0f;
 306    offsets[2].x = 1.0f; offsets[2].y = -1.0f;
 307    offsets[3].x = 2.0f; offsets[3].y = -1.0f;
 308
 309    offsets[4].x = -1.0f; offsets[4].y = 0.0f;
 310    offsets[5].x = 0.0f; offsets[5].y = 0.0f;
 311    offsets[6].x = 1.0f; offsets[6].y = 0.0f;
 312    offsets[7].x = 2.0f; offsets[7].y = 0.0f;
 313
 314    offsets[8].x = -1.0f; offsets[8].y = 1.0f;
 315    offsets[9].x = 0.0f; offsets[9].y = 1.0f;
 316    offsets[10].x = 1.0f; offsets[10].y = 1.0f;
 317    offsets[11].x = 2.0f; offsets[11].y = 1.0f;
 318
 319    offsets[12].x = -1.0f; offsets[12].y = 2.0f;
 320    offsets[13].x = 0.0f; offsets[13].y = 2.0f;
 321    offsets[14].x = 1.0f; offsets[14].y = 2.0f;
 322    offsets[15].x = 2.0f; offsets[15].y = 2.0f;
 323
 324    for (i = 0; i < 16; ++i) {
 325       offsets[i].x /= width;
 326       offsets[i].y /= height;
 327    }
 328
 329    filter->vs = create_vert_shader(filter);
 330    if (!filter->vs)
 331       goto error_vs;
 332
 333    filter->fs = create_frag_shader(filter, width, height, offsets);
 334    if (!filter->fs)
 335       goto error_fs;
 336
 337    return true;
 338
 339 error_fs:
 340    pipe->delete_vs_state(pipe, filter->vs);
 341
 342 error_vs:
 343    pipe->delete_vertex_elements_state(pipe, filter->ves);
 344
 345 error_ves:
 346    pipe_resource_reference(&filter->quad.buffer, NULL);
 347
 348 error_quad:
 349    pipe->delete_sampler_state(pipe, filter->sampler);
 350
 351 error_sampler:
 352    pipe->delete_blend_state(pipe, filter->blend);
 353
 354 error_blend:
 355    pipe->delete_rasterizer_state(pipe, filter->rs_state);
 356
 357 error_rs_state:
 358    return false;
 359 }
 360
 361 void
 362 vl_bicubic_filter_cleanup(struct vl_bicubic_filter *filter)
 363 {
 364    assert(filter);
 365
 366    filter->pipe->delete_sampler_state(filter->pipe, filter->sampler);
 367    filter->pipe->delete_blend_state(filter->pipe, filter->blend);
 368    filter->pipe->delete_rasterizer_state(filter->pipe, filter->rs_state);
 369    filter->pipe->delete_vertex_elements_state(filter->pipe, filter->ves);
 370    pipe_resource_reference(&filter->quad.buffer, NULL);
 371
 372    filter->pipe->delete_vs_state(filter->pipe, filter->vs);
 373    filter->pipe->delete_fs_state(filter->pipe, filter->fs);
 374 }
 375
 376 void
 377 vl_bicubic_filter_render(struct vl_bicubic_filter *filter,
 378                         struct pipe_sampler_view *src,
 379                         struct pipe_surface *dst,
 380                         struct u_rect *dst_area,
 381                         struct u_rect *dst_clip)
 382 {
 383    struct pipe_viewport_state viewport;
 384    struct pipe_framebuffer_state fb_state;
 385    struct pipe_scissor_state scissor;
 386    union pipe_color_union clear_color;
 387    struct pipe_transfer *buf_transfer;
 388    struct pipe_resource *surface_size;
 389    assert(filter && src && dst);
 390
 391    if (dst_clip) {
 392       scissor.minx = dst_clip->x0;
 393       scissor.miny = dst_clip->y0;
 394       scissor.maxx = dst_clip->x1;
 395       scissor.maxy = dst_clip->y1;
 396    } else {
 397       scissor.minx = 0;
 398       scissor.miny = 0;
 399       scissor.maxx = dst->width;
 400       scissor.maxy = dst->height;
 401    }
 402
 403    clear_color.f[0] = clear_color.f[1] = 0.0f;
 404    clear_color.f[2] = clear_color.f[3] = 0.0f;
 405    surface_size = pipe_buffer_create
 406    (
 407       filter->pipe->screen,
 408       PIPE_BIND_CONSTANT_BUFFER,
 409       PIPE_USAGE_DEFAULT,
 410       2*sizeof(float)
 411    );
 412
 413
 414    memset(&viewport, 0, sizeof(viewport));
 415    if(dst_area){
 416       viewport.scale[0] = dst_area->x1 - dst_area->x0;
 417       viewport.scale[1] = dst_area->y1 - dst_area->y0;
 418       viewport.translate[0] = dst_area->x0;
 419       viewport.translate[1] = dst_area->y0;
 420    } else {
 421       viewport.scale[0] = dst->width;
 422       viewport.scale[1] = dst->height;
 423    }
 424    viewport.scale[2] = 1;
 425
 426    float *ptr = pipe_buffer_map(filter->pipe, surface_size,
 427                                PIPE_TRANSFER_WRITE | PIPE_TRANSFER_DISCARD_RANGE,
 428                                &buf_transfer);
 429
 430    ptr[0] = 0.5f/viewport.scale[0];
 431    ptr[1] = 0.5f/viewport.scale[1];
 432
 433    pipe_buffer_unmap(filter->pipe, buf_transfer);
 434
 435    memset(&fb_state, 0, sizeof(fb_state));
 436    fb_state.width = dst->width;
 437    fb_state.height = dst->height;
 438    fb_state.nr_cbufs = 1;
 439    fb_state.cbufs[0] = dst;
 440
 441    filter->pipe->set_scissor_states(filter->pipe, 0, 1, &scissor);
 442    filter->pipe->clear_render_target(filter->pipe, dst, &clear_color,
 443                                      0, 0, dst->width, dst->height, false);
 444    pipe_set_constant_buffer(filter->pipe, PIPE_SHADER_FRAGMENT, 0, surface_size);
 445    filter->pipe->bind_rasterizer_state(filter->pipe, filter->rs_state);
 446    filter->pipe->bind_blend_state(filter->pipe, filter->blend);
 447    filter->pipe->bind_sampler_states(filter->pipe, PIPE_SHADER_FRAGMENT,
 448                                      0, 1, &filter->sampler);
 449    filter->pipe->set_sampler_views(filter->pipe, PIPE_SHADER_FRAGMENT,
 450                                    0, 1, &src);
 451    filter->pipe->bind_vs_state(filter->pipe, filter->vs);
 452    filter->pipe->bind_fs_state(filter->pipe, filter->fs);
 453    filter->pipe->set_framebuffer_state(filter->pipe, &fb_state);
 454    filter->pipe->set_viewport_states(filter->pipe, 0, 1, &viewport);
 455    filter->pipe->set_vertex_buffers(filter->pipe, 0, 1, &filter->quad);
 456    filter->pipe->bind_vertex_elements_state(filter->pipe, filter->ves);
 457
 458    util_draw_arrays(filter->pipe, PIPE_PRIM_QUADS, 0, 4);
 459 }