git.oblomov.eu Git - nouveau/blob - src/nv30_xv_tex.c

   1 /*
   2  * Copyright 2007-2008 Maarten Maathuis
   3  * Copyright 2008 Stephane Marchesin
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  22  * DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 #ifdef HAVE_CONFIG_H
  26 #include "config.h"
  27 #endif
  28
  29 #include "xf86xv.h"
  30 #include <X11/extensions/Xv.h>
  31 #include "exa.h"
  32 #include "damage.h"
  33 #include "dixstruct.h"
  34 #include "fourcc.h"
  35
  36 #include "nv_include.h"
  37 #include "nv_dma.h"
  38
  39 #include "nv30_shaders.h"
  40
  41 extern Atom xvSyncToVBlank, xvSetDefaults;
  42
  43 /*
  44  * The filtering function used for video scaling. We use a cubic filter as defined in
  45  * "Reconstruction Filters in Computer Graphics"
  46  * Mitchell & Netravali in SIGGRAPH '88
  47  */
  48 static float filter_func(float x)
  49 {
  50         const double B=0.75;
  51         const double C=(1.0-B)/2.0;
  52         double x1=fabs(x);
  53         double x2=fabs(x)*x1;
  54         double x3=fabs(x)*x2;
  55
  56         if (fabs(x)<1.0)
  57                 return ( (12.0-9.0*B-6.0*C)*x3+(-18.0+12.0*B+6.0*C)*x2+(6.0-2.0*B) )/6.0;
  58         else
  59                 return ( (-B-6.0*C)*x3+(6.0*B+30.0*C)*x2+(-12.0*B-48.0*C)*x1+(8.0*B+24.0*C) )/6.0;
  60 }
  61
  62 static int8_t f32tosb8(float v)
  63 {
  64         return (int8_t)(v*127.0);
  65 }
  66
  67 /*
  68  * 512 means 2048 bytes of VRAM
  69  */
  70 #define TABLE_SIZE 512
  71 static void compute_filter_table(int8_t *t) {
  72         int i;
  73         float x;
  74         for(i=0;i<TABLE_SIZE;i++) {
  75                 x=(i+0.5)/TABLE_SIZE;
  76
  77                 float w0=filter_func(x+1.0);
  78                 float w1=filter_func(x);
  79                 float w2=filter_func(x-1.0);
  80                 float w3=filter_func(x-2.0);
  81
  82                 t[4*i+2]=f32tosb8(1.0+x-w1/(w0+w1));
  83                 t[4*i+1]=f32tosb8(1.0-x+w3/(w2+w3));
  84                 t[4*i+0]=f32tosb8(w0+w1);
  85                 t[4*i+3]=f32tosb8(0.0);
  86         }
  87 }
  88
  89 static void
  90 NV30_LoadFilterTable(ScrnInfoPtr pScrn)
  91 {
  92         NVPtr pNv = NVPTR(pScrn);
  93
  94         if (!pNv->xv_filtertable_mem) {
  95                 if (nouveau_bo_new(pNv->dev, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART,
  96                                    0, TABLE_SIZE*sizeof(float)*4, &pNv->xv_filtertable_mem)) {
  97                         xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
  98                                 "Couldn't alloc filter table!\n");
  99                         return;
 100                 }
 101
 102                 if (nouveau_bo_map(pNv->xv_filtertable_mem, NOUVEAU_BO_RDWR)) {
 103                         xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 104                                    "Couldn't map filter table!\n");
 105                         return;
 106                 }
 107
 108                 int8_t *t=pNv->xv_filtertable_mem->map;
 109                 compute_filter_table(t);
 110
 111                 nouveau_bo_unmap(pNv->xv_filtertable_mem);
 112         }
 113 }
 114
 115 #define SWIZZLE(ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w)                        \
 116         (                                                                       \
 117         NV34TCL_TX_SWIZZLE_S0_X_##ts0x | NV34TCL_TX_SWIZZLE_S0_Y_##ts0y |       \
 118         NV34TCL_TX_SWIZZLE_S0_Z_##ts0z | NV34TCL_TX_SWIZZLE_S0_W_##ts0w |       \
 119         NV34TCL_TX_SWIZZLE_S1_X_##ts1x | NV34TCL_TX_SWIZZLE_S1_Y_##ts1y |       \
 120         NV34TCL_TX_SWIZZLE_S1_Z_##ts1z | NV34TCL_TX_SWIZZLE_S1_W_##ts1w         \
 121         )
 122
 123 /*
 124  * Texture 0 : filter table
 125  * Texture 1 : Y data
 126  * Texture 2 : UV data
 127  */
 128 static Bool
 129 NV30VideoTexture(ScrnInfoPtr pScrn, struct nouveau_bo *src, int offset,
 130                  uint16_t width, uint16_t height, uint16_t src_pitch, int unit)
 131 {
 132         NVPtr pNv = NVPTR(pScrn);
 133         struct nouveau_channel *chan = pNv->chan;
 134         struct nouveau_grobj *rankine = pNv->Nv3D;
 135
 136         uint32_t card_fmt = 0;
 137         uint32_t card_swz = 0;
 138
 139         switch(unit) {
 140                 case 0:
 141                 card_fmt = NV34TCL_TX_FORMAT_FORMAT_A8R8G8B8;
 142                 card_swz = SWIZZLE(S1, S1, S1, S1, X, Y, Z, W);
 143                 break;
 144                 case 1:
 145                 card_fmt = NV34TCL_TX_FORMAT_FORMAT_A8_RECT2;
 146                 card_swz = SWIZZLE(S1, S1, S1, S1, X, X, X, X);
 147                 break;
 148                 case 2:
 149                 card_fmt = NV34TCL_TX_FORMAT_FORMAT_L8A8_RECT;
 150 #if X_BYTE_ORDER == X_BIG_ENDIAN
 151                 card_swz = SWIZZLE(S1, S1, S1, S1, Z, W, X, Y); /* x = V, y = U */
 152 #else
 153                 card_swz = SWIZZLE(S1, S1, S1, S1, W, Z, Y, X); /* x = V, y = U */
 154 #endif
 155                 break;
 156         }
 157
 158         BEGIN_RING(chan, rankine, NV34TCL_TX_OFFSET(unit), 8);
 159         OUT_RELOCl(chan, src, offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
 160         if (unit==0) {
 161                 OUT_RELOCd(chan, pNv->FB, NV34TCL_TX_FORMAT_DIMS_1D | card_fmt |
 162                                  (1 << 16) |
 163                                  (log2i(width) <<
 164                                   NV34TCL_TX_FORMAT_BASE_SIZE_U_SHIFT) |
 165                                  (log2i(height) <<
 166                                   NV34TCL_TX_FORMAT_BASE_SIZE_V_SHIFT) |
 167                                  8 /* no idea */,
 168                                  NOUVEAU_BO_VRAM | NOUVEAU_BO_RD,
 169                                  NV34TCL_TX_FORMAT_DMA0, 0);
 170                 OUT_RING  (chan, NV34TCL_TX_WRAP_S_REPEAT |
 171                                  NV34TCL_TX_WRAP_T_CLAMP_TO_EDGE |
 172                                  NV34TCL_TX_WRAP_R_CLAMP_TO_EDGE);
 173         } else {
 174                 OUT_RELOCd(chan, pNv->FB, NV34TCL_TX_FORMAT_DIMS_2D | card_fmt |
 175                                  (1 << 16) |
 176                                  (log2i(width) <<
 177                                   NV34TCL_TX_FORMAT_BASE_SIZE_U_SHIFT) |
 178                                  (log2i(height) <<
 179                                   NV34TCL_TX_FORMAT_BASE_SIZE_V_SHIFT) |
 180                                  8 /* no idea */,
 181                                  NOUVEAU_BO_VRAM | NOUVEAU_BO_RD,
 182                                  NV34TCL_TX_FORMAT_DMA0, 0);
 183                 OUT_RING  (chan, NV34TCL_TX_WRAP_S_CLAMP_TO_EDGE |
 184                                  NV34TCL_TX_WRAP_T_CLAMP_TO_EDGE |
 185                                  NV34TCL_TX_WRAP_R_CLAMP_TO_EDGE);
 186         }
 187
 188         OUT_RING  (chan, NV34TCL_TX_ENABLE_ENABLE);
 189         OUT_RING  (chan, (src_pitch << NV34TCL_TX_SWIZZLE_RECT_PITCH_SHIFT) |
 190                          card_swz);
 191         if (unit==0)
 192                 OUT_RING  (chan, NV34TCL_TX_FILTER_SIGNED_ALPHA |
 193                                  NV34TCL_TX_FILTER_SIGNED_RED |
 194                                  NV34TCL_TX_FILTER_SIGNED_GREEN |
 195                                  NV34TCL_TX_FILTER_SIGNED_BLUE |
 196                                  NV34TCL_TX_FILTER_MINIFY_LINEAR |
 197                                  NV34TCL_TX_FILTER_MAGNIFY_LINEAR | 0x2000);
 198         else
 199                 OUT_RING  (chan, NV34TCL_TX_FILTER_MINIFY_LINEAR |
 200                                  NV34TCL_TX_FILTER_MAGNIFY_LINEAR | 0x2000);
 201         OUT_RING  (chan, (width << NV34TCL_TX_NPOT_SIZE_W_SHIFT) | height);
 202         OUT_RING  (chan, 0); /* border ARGB */
 203
 204         return TRUE;
 205 }
 206
 207 Bool
 208 NV30GetSurfaceFormat(PixmapPtr ppix, int *fmt_ret)
 209 {
 210         switch (ppix->drawable.bitsPerPixel) {
 211                 case 32:
 212                         *fmt_ret = NV34TCL_RT_FORMAT_COLOR_A8R8G8B8;
 213                         break;
 214                 case 24:
 215                         *fmt_ret = NV34TCL_RT_FORMAT_COLOR_X8R8G8B8;
 216                         break;
 217                 case 16:
 218                         *fmt_ret = NV34TCL_RT_FORMAT_COLOR_R5G6B5;
 219                         break;
 220                 case 8:
 221                         *fmt_ret = NV34TCL_RT_FORMAT_COLOR_B8;
 222                         break;
 223                 default:
 224                         return FALSE;
 225         }
 226
 227         return TRUE;
 228 }
 229
 230 void
 231 NV30StopTexturedVideo(ScrnInfoPtr pScrn, pointer data, Bool Exit)
 232 {
 233 }
 234
 235 #define VERTEX_OUT(sx,sy,dx,dy) do {                                           \
 236         BEGIN_RING(chan, rankine, NV34TCL_VTX_ATTR_2F_X(8), 4);                \
 237         OUT_RINGf (chan, (sx)); OUT_RINGf (chan, (sy));                        \
 238         OUT_RINGf (chan, (sx)/2.0); OUT_RINGf (chan, (sy)/2.0);                \
 239         BEGIN_RING(chan, rankine, NV34TCL_VTX_ATTR_2I(0), 1);                  \
 240         OUT_RING  (chan, ((dy)<<16)|(dx));                                     \
 241 } while(0)
 242
 243 int
 244 NV30PutTextureImage(ScrnInfoPtr pScrn, struct nouveau_bo *src, int src_offset,
 245                     int src_offset2, int id, int src_pitch, BoxPtr dstBox,
 246                     int x1, int y1, int x2, int y2,
 247                     uint16_t width, uint16_t height,
 248                     uint16_t src_w, uint16_t src_h,
 249                     uint16_t drw_w, uint16_t drw_h,
 250                     RegionPtr clipBoxes,
 251                     PixmapPtr ppix,
 252                     NVPortPrivPtr pPriv)
 253 {
 254         NVPtr pNv = NVPTR(pScrn);
 255         struct nouveau_channel *chan = pNv->chan;
 256         struct nouveau_grobj *rankine = pNv->Nv3D;
 257         struct nouveau_pixmap *rt = nouveau_pixmap(ppix);
 258         Bool redirected = FALSE;
 259         float X1, X2, Y1, Y2;
 260         BoxPtr pbox;
 261         int nbox;
 262         int dst_format = 0;
 263
 264         if (drw_w > 4096 || drw_h > 4096) {
 265                 xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 266                         "XV: Draw size too large.\n");
 267                 return BadAlloc;
 268         }
 269
 270         if (!NV30GetSurfaceFormat(ppix, &dst_format)) {
 271                 ErrorF("No surface format, bad.\n");
 272         }
 273
 274 #ifdef COMPOSITE
 275         if (!NVExaPixmapIsOnscreen(ppix))
 276                 redirected = TRUE;
 277 #endif
 278
 279         pbox = REGION_RECTS(clipBoxes);
 280         nbox = REGION_NUM_RECTS(clipBoxes);
 281
 282         /* Disable blending */
 283         BEGIN_RING(chan, rankine, NV34TCL_BLEND_FUNC_ENABLE, 1);
 284         OUT_RING  (chan, 0);
 285
 286         /* Setup surface */
 287         BEGIN_RING(chan, rankine, NV34TCL_RT_FORMAT, 3);
 288         OUT_RING  (chan, NV34TCL_RT_FORMAT_TYPE_LINEAR |
 289                          NV34TCL_RT_FORMAT_ZETA_Z24S8 |
 290                          dst_format);
 291         OUT_RING  (chan, (exaGetPixmapPitch(ppix) << 16) |
 292                           exaGetPixmapPitch(ppix));
 293         OUT_RELOCl(chan, rt->bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 294
 295         if (pNv->NVArch == 0x30) {
 296                 int x = 0;
 297                 int y = 0;
 298                 int w = ppix->drawable.x + ppix->drawable.width;
 299                 int h = ppix->drawable.y + ppix->drawable.height;
 300
 301                 BEGIN_RING(chan, rankine, NV34TCL_VIEWPORT_HORIZ, 2);
 302                 OUT_RING  (chan, (w<<16)|x);
 303                 OUT_RING  (chan, (h<<16)|y);
 304                 BEGIN_RING(chan, rankine, NV34TCL_VIEWPORT_CLIP_HORIZ(0), 2);
 305                 OUT_RING  (chan, (w-1+x)<<16);
 306                 OUT_RING  (chan, (h-1+y)<<16);
 307                 BEGIN_RING(chan, rankine, NV34TCL_VIEWPORT_TX_ORIGIN, 1);
 308                 OUT_RING  (chan, (y<<16)|x);
 309         }
 310
 311         NV30_LoadFilterTable(pScrn);
 312
 313         BEGIN_RING(chan, rankine, NV34TCL_TX_UNITS_ENABLE, 1);
 314         OUT_RING  (chan, NV34TCL_TX_UNITS_ENABLE_TX0 |
 315                          NV34TCL_TX_UNITS_ENABLE_TX1);
 316
 317         NV30VideoTexture(pScrn, pNv->xv_filtertable_mem, 0, TABLE_SIZE, 1, 0 , 0);
 318         NV30VideoTexture(pScrn, src, src_offset, src_w, src_h, src_pitch, 1);
 319         /* We've got NV12 format, which means half width and half height texture of chroma channels. */
 320         NV30VideoTexture(pScrn, src, src_offset2, src_w/2, src_h/2, src_pitch, 2);
 321
 322         BEGIN_RING(chan, rankine, NV34TCL_TX_ENABLE(3), 1);
 323         OUT_RING  (chan, 0x0);
 324
 325         if (pPriv->bicubic)
 326                 NV30_LoadFragProg(pScrn, &nv30_fp_yv12_bicubic);
 327         else
 328                 NV30_LoadFragProg(pScrn, &nv30_fp_yv12_bilinear);
 329
 330         /* Just before rendering we wait for vblank in the non-composited case. */
 331         if (pPriv->SyncToVBlank && !redirected) {
 332                 uint8_t crtcs = nv_window_belongs_to_crtc(pScrn, dstBox->x1, dstBox->y1,
 333                         dstBox->x2 - dstBox->x1, dstBox->y2 - dstBox->y1);
 334
 335                 FIRE_RING (chan);
 336                 if (crtcs & 0x1)
 337                         NVWaitVSync(pScrn, 0);
 338                 else if (crtcs & 0x2)
 339                         NVWaitVSync(pScrn, 1);
 340         }
 341
 342         /* These are fixed point values in the 16.16 format. */
 343         X1 = (float)(x1>>16)+(float)(x1&0xFFFF)/(float)0x10000;
 344         Y1 = (float)(y1>>16)+(float)(y1&0xFFFF)/(float)0x10000;
 345         X2 = (float)(x2>>16)+(float)(x2&0xFFFF)/(float)0x10000;
 346         Y2 = (float)(y2>>16)+(float)(y2&0xFFFF)/(float)0x10000;
 347
 348         BEGIN_RING(chan, rankine, NV34TCL_VERTEX_BEGIN_END, 1);
 349         OUT_RING  (chan, NV34TCL_VERTEX_BEGIN_END_TRIANGLES);
 350
 351         while(nbox--) {
 352                 float tx1=X1+(float)(pbox->x1 - dstBox->x1)*(X2-X1)/(float)(drw_w);
 353                 float tx2=X1+(float)(pbox->x2 - dstBox->x1)*(src_w)/(float)(drw_w);
 354                 float ty1=Y1+(float)(pbox->y1 - dstBox->y1)*(Y2-Y1)/(float)(drw_h);
 355                 float ty2=Y1+(float)(pbox->y2 - dstBox->y1)*(src_h)/(float)(drw_h);
 356                 int sx1=pbox->x1;
 357                 int sx2=pbox->x2;
 358                 int sy1=pbox->y1;
 359                 int sy2=pbox->y2;
 360
 361                 BEGIN_RING(chan, rankine, NV34TCL_SCISSOR_HORIZ, 2);
 362                 OUT_RING  (chan, (sx2 << 16) | 0);
 363                 OUT_RING  (chan, (sy2 << 16) | 0);
 364
 365                 VERTEX_OUT(tx1, ty1, sx1, sy1);
 366                 VERTEX_OUT(tx2+(tx2-tx1), ty1, sx2+(sx2-sx1), sy1);
 367                 VERTEX_OUT(tx1, ty2+(ty2-ty1), sx1, sy2+(sy2-sy1));
 368
 369                 pbox++;
 370         }
 371
 372         BEGIN_RING(chan, rankine, NV34TCL_VERTEX_BEGIN_END, 1);
 373         OUT_RING  (chan, NV34TCL_VERTEX_BEGIN_END_STOP);
 374
 375         FIRE_RING (chan);
 376
 377         return Success;
 378 }
 379
 380 /**
 381  * NV30SetTexturePortAttribute
 382  * sets the attribute "attribute" of port "data" to value "value"
 383  * supported attributes:
 384  * Sync to vblank.
 385  *
 386  * @param pScrenInfo
 387  * @param attribute attribute to set
 388  * @param value value to which attribute is to be set
 389  * @param data port from which the attribute is to be set
 390  *
 391  * @return Success, if setting is successful
 392  * BadValue/BadMatch, if value/attribute are invalid
 393  */
 394 int
 395 NV30SetTexturePortAttribute(ScrnInfoPtr pScrn, Atom attribute,
 396                        INT32 value, pointer data)
 397 {
 398         NVPortPrivPtr pPriv = (NVPortPrivPtr)data;
 399         NVPtr           pNv = NVPTR(pScrn);
 400
 401         if ((attribute == xvSyncToVBlank) && pNv->WaitVSyncPossible) {
 402                 if ((value < 0) || (value > 1))
 403                         return BadValue;
 404                 pPriv->SyncToVBlank = value;
 405         } else
 406         if (attribute == xvSetDefaults) {
 407                 pPriv->SyncToVBlank = pNv->WaitVSyncPossible;
 408         } else
 409                 return BadMatch;
 410
 411         return Success;
 412 }
 413
 414 /**
 415  * NV30GetTexturePortAttribute
 416  * reads the value of attribute "attribute" from port "data" into INT32 "*value"
 417  * Sync to vblank.
 418  *
 419  * @param pScrn unused
 420  * @param attribute attribute to be read
 421  * @param value value of attribute will be stored here
 422  * @param data port from which attribute will be read
 423  * @return Success, if queried attribute exists
 424  */
 425 int
 426 NV30GetTexturePortAttribute(ScrnInfoPtr pScrn, Atom attribute,
 427                        INT32 *value, pointer data)
 428 {
 429         NVPortPrivPtr pPriv = (NVPortPrivPtr)data;
 430
 431         if(attribute == xvSyncToVBlank)
 432                 *value = (pPriv->SyncToVBlank) ? 1 : 0;
 433         else
 434                 return BadMatch;
 435
 436         return Success;
 437 }
 438