git.oblomov.eu Git - nouveau/blob - src/nv40_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  * Implements the filtering as described in
  69  * "Fast Third-Order Texture Filtering"
  70  * Sigg & Hardwiger in GPU Gems 2
  71  */
  72 #define TABLE_SIZE 512
  73 static void compute_filter_table(int8_t *t) {
  74         int i;
  75         float x;
  76         for(i=0;i<TABLE_SIZE;i++) {
  77                 x=(i+0.5)/TABLE_SIZE;
  78
  79                 float w0=filter_func(x+1.0);
  80                 float w1=filter_func(x);
  81                 float w2=filter_func(x-1.0);
  82                 float w3=filter_func(x-2.0);
  83
  84                 t[4*i+2]=f32tosb8(1.0+x-w1/(w0+w1));
  85                 t[4*i+1]=f32tosb8(1.0-x+w3/(w2+w3));
  86                 t[4*i+0]=f32tosb8(w0+w1);
  87                 t[4*i+3]=f32tosb8(0.0);
  88         }
  89 }
  90
  91 static struct nouveau_bo *table_mem = NULL;
  92 static void
  93 NV40_LoadFilterTable(ScrnInfoPtr pScrn)
  94 {
  95         NVPtr pNv = NVPTR(pScrn);
  96
  97         if (!table_mem) {
  98                 if (nouveau_bo_new(pNv->dev, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART,
  99                                 0, TABLE_SIZE*sizeof(float)*4, &table_mem)) {
 100                         xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 101                                 "Couldn't alloc filter table!\n");
 102                         return;
 103                 }
 104
 105                 if (nouveau_bo_map(table_mem, NOUVEAU_BO_RDWR)) {
 106                         xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 107                                    "Couldn't map filter table!\n");
 108                         return;
 109                 }
 110
 111                 int8_t *t=table_mem->map;
 112                 compute_filter_table(t);
 113
 114                 nouveau_bo_unmap(table_mem);
 115         }
 116 }
 117
 118 #define SWIZZLE(ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w)                                \
 119         (                                                                               \
 120         NV40TCL_TEX_SWIZZLE_S0_X_##ts0x | NV40TCL_TEX_SWIZZLE_S0_Y_##ts0y       |       \
 121         NV40TCL_TEX_SWIZZLE_S0_Z_##ts0z | NV40TCL_TEX_SWIZZLE_S0_W_##ts0w       |       \
 122         NV40TCL_TEX_SWIZZLE_S1_X_##ts1x | NV40TCL_TEX_SWIZZLE_S1_Y_##ts1y       |       \
 123         NV40TCL_TEX_SWIZZLE_S1_Z_##ts1z | NV40TCL_TEX_SWIZZLE_S1_W_##ts1w               \
 124         )
 125
 126 /*
 127  * Texture 0 : filter table
 128  * Texture 1 : Y data
 129  * Texture 2 : UV data
 130  */
 131 static Bool
 132 NV40VideoTexture(ScrnInfoPtr pScrn, struct nouveau_bo *src, int offset,
 133                  uint16_t width, uint16_t height, uint16_t src_pitch, int unit)
 134 {
 135         NVPtr pNv = NVPTR(pScrn);
 136         struct nouveau_channel *chan = pNv->chan;
 137         struct nouveau_grobj *curie = pNv->Nv3D;
 138
 139         uint32_t card_fmt = 0;
 140         uint32_t card_swz = 0;
 141
 142         switch(unit) {
 143         case 0:
 144                 card_fmt = NV40TCL_TEX_FORMAT_FORMAT_A8R8G8B8;
 145                 card_swz = SWIZZLE(S1, S1, S1, S1, X, Y, Z, W);
 146                 break;
 147         case 1:
 148                 card_fmt = NV40TCL_TEX_FORMAT_FORMAT_L8;
 149                 card_swz = SWIZZLE(S1, S1, S1, S1, X, X, X, X);
 150                 break;
 151         case 2:
 152                 card_fmt = NV40TCL_TEX_FORMAT_FORMAT_A8L8;
 153 #if X_BYTE_ORDER == X_BIG_ENDIAN
 154                 card_swz = SWIZZLE(S1, S1, S1, S1, Z, W, X, Y); /* x = V, y = U */
 155 #else
 156                 card_swz = SWIZZLE(S1, S1, S1, S1, W, Z, Y, X); /* x = V, y = U */
 157 #endif
 158                 break;
 159         }
 160
 161         BEGIN_RING(chan, curie, NV40TCL_TEX_OFFSET(unit), 8);
 162         OUT_RELOCl(chan, src, offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
 163         if (unit==0) {
 164                 OUT_RELOCd(chan, pNv->FB, card_fmt |
 165                                  NV40TCL_TEX_FORMAT_DIMS_1D | 0x8000 |
 166                                  NV40TCL_TEX_FORMAT_NO_BORDER |
 167                                  (1 << NV40TCL_TEX_FORMAT_MIPMAP_COUNT_SHIFT),
 168                                  NOUVEAU_BO_VRAM | NOUVEAU_BO_RD,
 169                                  NV40TCL_TEX_FORMAT_DMA0, 0);
 170                 OUT_RING  (chan, NV40TCL_TEX_WRAP_S_REPEAT |
 171                                  NV40TCL_TEX_WRAP_T_CLAMP_TO_EDGE |
 172                                  NV40TCL_TEX_WRAP_R_CLAMP_TO_EDGE);
 173         } else {
 174                 OUT_RELOCd(chan, pNv->FB, card_fmt | NV40TCL_TEX_FORMAT_LINEAR |
 175                                  NV40TCL_TEX_FORMAT_RECT | 0x8000 |
 176                                  NV40TCL_TEX_FORMAT_DIMS_2D |
 177                                  NV40TCL_TEX_FORMAT_NO_BORDER |
 178                                  (1 << NV40TCL_TEX_FORMAT_MIPMAP_COUNT_SHIFT),
 179                                  NOUVEAU_BO_VRAM | NOUVEAU_BO_RD,
 180                                  NV40TCL_TEX_FORMAT_DMA0, 0);
 181                 OUT_RING  (chan, NV40TCL_TEX_WRAP_S_CLAMP_TO_EDGE |
 182                                  NV40TCL_TEX_WRAP_T_CLAMP_TO_EDGE |
 183                                  NV40TCL_TEX_WRAP_R_CLAMP_TO_EDGE);
 184         }
 185
 186         OUT_RING  (chan, NV40TCL_TEX_ENABLE_ENABLE);
 187         OUT_RING  (chan, card_swz);
 188         if (unit == 0)
 189                 OUT_RING  (chan, NV40TCL_TEX_FILTER_SIGNED_ALPHA |
 190                                  NV40TCL_TEX_FILTER_SIGNED_RED |
 191                                  NV40TCL_TEX_FILTER_SIGNED_GREEN |
 192                                  NV40TCL_TEX_FILTER_SIGNED_BLUE |
 193                                  NV40TCL_TEX_FILTER_MIN_LINEAR |
 194                                  NV40TCL_TEX_FILTER_MAG_LINEAR | 0x3fd6);
 195         else
 196                 OUT_RING  (chan, NV40TCL_TEX_FILTER_MIN_LINEAR |
 197                                  NV40TCL_TEX_FILTER_MAG_LINEAR | 0x3fd6);
 198         OUT_RING  (chan, (width << 16) | height);
 199         OUT_RING  (chan, 0); /* border ARGB */
 200
 201         BEGIN_RING(chan, curie, NV40TCL_TEX_SIZE1(unit), 1);
 202         OUT_RING  (chan, (1 << NV40TCL_TEX_SIZE1_DEPTH_SHIFT) |
 203                          (uint16_t) src_pitch);
 204
 205         return TRUE;
 206 }
 207
 208 Bool
 209 NV40GetSurfaceFormat(PixmapPtr ppix, int *fmt_ret)
 210 {
 211         switch (ppix->drawable.bitsPerPixel) {
 212         case 32:
 213                 *fmt_ret = NV40TCL_RT_FORMAT_COLOR_A8R8G8B8;
 214                 break;
 215         case 24:
 216                 *fmt_ret = NV40TCL_RT_FORMAT_COLOR_X8R8G8B8;
 217                 break;
 218         case 16:
 219                 *fmt_ret = NV40TCL_RT_FORMAT_COLOR_R5G6B5;
 220                 break;
 221         case 8:
 222                 *fmt_ret = NV40TCL_RT_FORMAT_COLOR_B8;
 223                 break;
 224         default:
 225                 return FALSE;
 226         }
 227
 228         return TRUE;
 229 }
 230
 231 void
 232 NV40StopTexturedVideo(ScrnInfoPtr pScrn, pointer data, Bool Exit)
 233 {
 234 }
 235
 236 #define VERTEX_OUT(sx,sy,dx,dy) do {                                           \
 237         BEGIN_RING(chan, curie, NV40TCL_VTX_ATTR_2F_X(8), 4);                  \
 238         OUT_RINGf (chan, (sx)); OUT_RINGf (chan, (sy));                        \
 239         OUT_RINGf (chan, (sx)/2.0); OUT_RINGf (chan, (sy)/2.0);                \
 240         BEGIN_RING(chan, curie, NV40TCL_VTX_ATTR_2I(0), 1);                    \
 241         OUT_RING  (chan, ((dy)<<16)|(dx));                                     \
 242 } while(0)
 243
 244 int
 245 NV40PutTextureImage(ScrnInfoPtr pScrn,
 246                     struct nouveau_bo *src, int src_offset, int src_offset2,
 247                     int id, int src_pitch, BoxPtr dstBox,
 248                     int x1, int y1, int x2, int y2,
 249                     uint16_t width, uint16_t height,
 250                     uint16_t src_w, uint16_t src_h,
 251                     uint16_t drw_w, uint16_t drw_h,
 252                     RegionPtr clipBoxes, PixmapPtr ppix,
 253                     NVPortPrivPtr pPriv)
 254 {
 255         NVPtr pNv = NVPTR(pScrn);
 256         struct nouveau_channel *chan = pNv->chan;
 257         struct nouveau_grobj *curie = pNv->Nv3D;
 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 (!NV40GetSurfaceFormat(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, curie, NV40TCL_BLEND_ENABLE, 1);
 284         OUT_RING  (chan, 0);
 285
 286         /* Setup surface */
 287         BEGIN_RING(chan, curie, NV40TCL_RT_FORMAT, 3);
 288         OUT_RING  (chan, NV40TCL_RT_FORMAT_TYPE_LINEAR |
 289                          NV40TCL_RT_FORMAT_ZETA_Z24S8 | dst_format);
 290         OUT_RING  (chan, exaGetPixmapPitch(ppix));
 291         OUT_PIXMAPl(chan, ppix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 292
 293         NV40_LoadFilterTable(pScrn);
 294
 295         NV40VideoTexture(pScrn, table_mem, 0, TABLE_SIZE, 1, 0 , 0);
 296         NV40VideoTexture(pScrn, src, src_offset, src_w, src_h, src_pitch, 1);
 297         /* We've got NV12 format, which means half width and half height texture of chroma channels. */
 298         NV40VideoTexture(pScrn, src, src_offset2, src_w/2, src_h/2, src_pitch, 2);
 299
 300         NV40_LoadVtxProg(pScrn, &nv40_vp_video);
 301         if (pPriv->bicubic)
 302                 NV40_LoadFragProg(pScrn, &nv40_fp_yv12_bicubic);
 303         else
 304                 NV40_LoadFragProg(pScrn, &nv30_fp_yv12_bilinear);
 305
 306         /* Appears to be some kind of cache flush, needed here at least
 307          * sometimes.. funky text rendering otherwise :)
 308          */
 309         BEGIN_RING(chan, curie, NV40TCL_TEX_CACHE_CTL, 1);
 310         OUT_RING  (chan, 2);
 311         BEGIN_RING(chan, curie, NV40TCL_TEX_CACHE_CTL, 1);
 312         OUT_RING  (chan, 1);
 313
 314         /* Just before rendering we wait for vblank in the non-composited case. */
 315         if (pPriv->SyncToVBlank && !redirected) {
 316                 uint8_t crtcs = nv_window_belongs_to_crtc(pScrn, dstBox->x1, dstBox->y1,
 317                         dstBox->x2 - dstBox->x1, dstBox->y2 - dstBox->y1);
 318
 319                 FIRE_RING (chan);
 320                 if (crtcs & 0x1)
 321                         NVWaitVSync(pScrn, 0);
 322                 else if (crtcs & 0x2)
 323                         NVWaitVSync(pScrn, 1);
 324         }
 325
 326         /* These are fixed point values in the 16.16 format. */
 327         X1 = (float)(x1>>16)+(float)(x1&0xFFFF)/(float)0x10000;
 328         Y1 = (float)(y1>>16)+(float)(y1&0xFFFF)/(float)0x10000;
 329         X2 = (float)(x2>>16)+(float)(x2&0xFFFF)/(float)0x10000;
 330         Y2 = (float)(y2>>16)+(float)(y2&0xFFFF)/(float)0x10000;
 331
 332         BEGIN_RING(chan, curie, NV40TCL_BEGIN_END, 1);
 333         OUT_RING  (chan, NV40TCL_BEGIN_END_TRIANGLES);
 334
 335         while(nbox--) {
 336                 float tx1=X1+(float)(pbox->x1 - dstBox->x1)*(X2-X1)/(float)(drw_w);
 337                 float tx2=X1+(float)(pbox->x2 - dstBox->x1)*(src_w)/(float)(drw_w);
 338                 float ty1=Y1+(float)(pbox->y1 - dstBox->y1)*(Y2-Y1)/(float)(drw_h);
 339                 float ty2=Y1+(float)(pbox->y2 - dstBox->y1)*(src_h)/(float)(drw_h);
 340                 int sx1=pbox->x1;
 341                 int sx2=pbox->x2;
 342                 int sy1=pbox->y1;
 343                 int sy2=pbox->y2;
 344
 345                 BEGIN_RING(chan, curie, NV40TCL_SCISSOR_HORIZ, 2);
 346                 OUT_RING  (chan, (sx2 << 16) | 0);
 347                 OUT_RING  (chan, (sy2 << 16) | 0);
 348
 349                 VERTEX_OUT(tx1, ty1, sx1, sy1);
 350                 VERTEX_OUT(tx2+(tx2-tx1), ty1, sx2+(sx2-sx1), sy1);
 351                 VERTEX_OUT(tx1, ty2+(ty2-ty1), sx1, sy2+(sy2-sy1));
 352
 353                 pbox++;
 354         }
 355
 356         BEGIN_RING(chan, curie, NV40TCL_BEGIN_END, 1);
 357         OUT_RING  (chan, NV40TCL_BEGIN_END_STOP);
 358
 359         FIRE_RING (chan);
 360
 361         return Success;
 362 }
 363
 364 /**
 365  * NV40SetTexturePortAttribute
 366  * sets the attribute "attribute" of port "data" to value "value"
 367  * supported attributes:
 368  * Sync to vblank.
 369  *
 370  * @param pScrenInfo
 371  * @param attribute attribute to set
 372  * @param value value to which attribute is to be set
 373  * @param data port from which the attribute is to be set
 374  *
 375  * @return Success, if setting is successful
 376  * BadValue/BadMatch, if value/attribute are invalid
 377  */
 378 int
 379 NV40SetTexturePortAttribute(ScrnInfoPtr pScrn, Atom attribute,
 380                        INT32 value, pointer data)
 381 {
 382         NVPortPrivPtr pPriv = (NVPortPrivPtr)data;
 383         NVPtr           pNv = NVPTR(pScrn);
 384
 385         if ((attribute == xvSyncToVBlank) && pNv->WaitVSyncPossible) {
 386                 if ((value < 0) || (value > 1))
 387                         return BadValue;
 388                 pPriv->SyncToVBlank = value;
 389         } else
 390         if (attribute == xvSetDefaults) {
 391                 pPriv->SyncToVBlank = pNv->WaitVSyncPossible;
 392         } else
 393                 return BadMatch;
 394
 395         return Success;
 396 }
 397
 398 /**
 399  * NV40GetTexturePortAttribute
 400  * reads the value of attribute "attribute" from port "data" into INT32 "*value"
 401  * Sync to vblank.
 402  *
 403  * @param pScrn unused
 404  * @param attribute attribute to be read
 405  * @param value value of attribute will be stored here
 406  * @param data port from which attribute will be read
 407  * @return Success, if queried attribute exists
 408  */
 409 int
 410 NV40GetTexturePortAttribute(ScrnInfoPtr pScrn, Atom attribute,
 411                        INT32 *value, pointer data)
 412 {
 413         NVPortPrivPtr pPriv = (NVPortPrivPtr)data;
 414
 415         if(attribute == xvSyncToVBlank)
 416                 *value = (pPriv->SyncToVBlank) ? 1 : 0;
 417         else
 418                 return BadMatch;
 419
 420         return Success;
 421 }
 422