git.oblomov.eu Git - nouveau/blob - src/nv40_video_texture.c

   1 /*
   2  * Copyright 2007 Maarten Maathuis
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
  21  * DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #ifdef HAVE_CONFIG_H
  25 #include "config.h"
  26 #endif
  27
  28 #include "xf86.h"
  29 #include "xf86_OSproc.h"
  30 #include "xf86Resources.h"
  31 #include "compiler.h"
  32 #include "xf86PciInfo.h"
  33 #include "xf86Pci.h"
  34 #include "xf86fbman.h"
  35 #include "regionstr.h"
  36
  37 #include "xf86xv.h"
  38 #include <X11/extensions/Xv.h>
  39 #include "exa.h"
  40 #include "damage.h"
  41 #include "dixstruct.h"
  42 #include "fourcc.h"
  43
  44 #include "nv_include.h"
  45 #include "nv_dma.h"
  46
  47 #include "nv_shaders.h"
  48
  49 static nv_shader_t nv40_video = {
  50         .card_priv.NV30VP.vp_in_reg  = 0x00000309,
  51         .card_priv.NV30VP.vp_out_reg = 0x0000c001,
  52         .size = (3*4),
  53         .data = {
  54                 /* MOV result.position, vertex.position */
  55                 0x40041c6c, 0x0040000d, 0x8106c083, 0x6041ff80,
  56                 /* MOV result.texcoord[0], vertex.texcoord[0] */
  57                 0x401f9c6c, 0x0040080d, 0x8106c083, 0x6041ff9c,
  58                 /* MOV result.texcoord[1], vertex.texcoord[1] */
  59                 0x401f9c6c, 0x0040090d, 0x8106c083, 0x6041ffa1,
  60         }
  61 };
  62
  63 static nv_shader_t nv40_yv12 = {
  64         .card_priv.NV30FP.num_regs = 4,
  65         .size = (17*4),
  66         .data = {
  67                 /* INST 0.0, TEX R1 (TR0.xyzw), attrib.texcoord[0] */
  68                 0x17009e02, 0x1c9dc811, 0x0001c801, 0x0001c801,
  69                 /* INST 1.0, TEX R2 (TR0.xyzw), attrib.texcoord[1] */
  70                 0x1702be04, 0x1c9dc815, 0x0001c801, 0x0001c801,
  71                 /* INST 2.0, DP4R R3.x (TR0.xyzw), R1, { 0.00, 0.00, 0.00, 1.00 } */
  72                 0x06000206, 0x1c9dc804, 0x0001c802, 0x0001c801,
  73                 /* const */
  74                 0x00000000, 0x00000000, 0x00000000, 0x3f800000,
  75                 /* INST 3.0, DP4R R3.y (TR0.xyzw), R2, { 0.00, 1.00, 0.00, 0.00 } */
  76                 0x06000406, 0x1c9dc808, 0x0001c802, 0x0001c801,
  77                 /* const */
  78                 0x00000000, 0x3f800000, 0x00000000, 0x00000000,
  79                 /* INST 4.0, DP4R R3.z (TR0.xyzw), R2, { 1.00, 0.00, 0.00, 0.00 } */
  80                 0x06000806, 0x1c9dc808, 0x0001c802, 0x0001c801,
  81                 /* const */
  82                 0x3f800000, 0x00000000, 0x00000000, 0x00000000,
  83                 /* INST 5.0, ADDR R3 (TR0.xyzw), R3, { 0.00, -0.50, -0.50, 0.00 } */
  84                 0x03001e06, 0x1c9dc80c, 0x0001c802, 0x0001c801,
  85                 /* const */
  86                 0x00000000, 0xBF000000, 0xBF000000, 0x00000000,
  87                 /* INST 6.0, DP3R R0.x (TR0.xyzw), R3, { 1.00, 0.00, 1.4022, 0.00 } */
  88                 0x05000280, 0x1c9dc80c, 0x0001c802, 0x0001c801,
  89                 /* const */
  90                 0x3f800000, 0x00000000, 0x3FB37B4A, 0x00000000,
  91                 /* INST 7.0, DP3R R0.y (TR0.xyzw), R3, { 1.00, -0.3457, -0.7145, 0.00 } */
  92                 0x05000480, 0x1c9dc80c, 0x0001c802, 0x0001c801,
  93                 /* const */
  94                 0x3f800000, 0xBEB0FF97, 0xBF36E979, 0x00000000,
  95                 /* INST 8.0, DP3R R0.z (TR0.xyzw), R3, { 1.00, 1.7710, 0.00, 0.00 } */
  96                 0x05000880, 0x1c9dc80c, 0x0001c802, 0x0001c801,
  97                 /* const */
  98                 0x3f800000, 0x3FE2B021, 0x00000000, 0x00000000,
  99                 /* INST 9.0, MOVR R0.w (TR0.xyzw), R3.wwww + END */
 100                 0x01001081, 0x1c9dfe0c, 0x0001c801, 0x0001c801,
 101         }
 102 };
 103
 104 #define SWIZZLE(ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w)                                                        \
 105         (                                                                                                                                       \
 106         NV40TCL_TEX_SWIZZLE_S0_X_##ts0x | NV40TCL_TEX_SWIZZLE_S0_Y_##ts0y               |       \
 107         NV40TCL_TEX_SWIZZLE_S0_Z_##ts0z | NV40TCL_TEX_SWIZZLE_S0_W_##ts0w       |       \
 108         NV40TCL_TEX_SWIZZLE_S1_X_##ts1x | NV40TCL_TEX_SWIZZLE_S1_Y_##ts1y       |       \
 109         NV40TCL_TEX_SWIZZLE_S1_Z_##ts1z | NV40TCL_TEX_SWIZZLE_S1_W_##ts1w               \
 110         )
 111
 112 static Bool
 113 NV40VideoTexture(ScrnInfoPtr pScrn, int offset, uint16_t width, uint16_t height, uint16_t src_pitch, int unit)
 114 {
 115         NVPtr pNv = NVPTR(pScrn);
 116
 117         uint32_t card_fmt = 0;
 118         uint32_t card_swz = 0;
 119
 120         if (unit == 0) {
 121                 /* Pretend we've got a normal 8 bits format. */
 122                 card_fmt = NV40TCL_TEX_FORMAT_FORMAT_L8;
 123                 card_swz = SWIZZLE(ZERO, ZERO, ZERO, S1, X, X, X, X);
 124         } else {
 125                 /* Pretend we've got a normal 2x8 bits format. */
 126                 card_fmt = NV40TCL_TEX_FORMAT_FORMAT_A8L8;
 127                 card_swz = SWIZZLE(S1, S1, S1, S1, Y, X, W, Z); /* x = V, y = U */
 128         }
 129
 130         BEGIN_RING(Nv3D, NV40TCL_TEX_OFFSET(unit), 8);
 131         /* We get an obsolute offset, which needs to be corrected. */
 132         OUT_RELOCl(pNv->FB, (uint32_t)(offset - pNv->FB->offset), NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
 133         OUT_RELOCd(pNv->FB, card_fmt | NV40TCL_TEX_FORMAT_LINEAR |
 134                         NV40TCL_TEX_FORMAT_DIMS_2D | NV40TCL_TEX_FORMAT_NO_BORDER |
 135                         (0x8000) | (1 << NV40TCL_TEX_FORMAT_MIPMAP_COUNT_SHIFT),
 136                         NOUVEAU_BO_VRAM | NOUVEAU_BO_RD,
 137                         NV40TCL_TEX_FORMAT_DMA0, 0);
 138
 139         OUT_RING(NV40TCL_TEX_WRAP_S_CLAMP_TO_BORDER |
 140                         NV40TCL_TEX_WRAP_T_CLAMP_TO_BORDER |
 141                         NV40TCL_TEX_WRAP_R_CLAMP_TO_BORDER);
 142         OUT_RING(NV40TCL_TEX_ENABLE_ENABLE);
 143         OUT_RING(card_swz);
 144         if (unit == 0) {
 145                 OUT_RING(NV40TCL_TEX_FILTER_MIN_LINEAR |
 146                                 NV40TCL_TEX_FILTER_MAG_LINEAR |
 147                                 0x3fd6);
 148         } else { /* UV texture cannot be linearly filtered, because it's just offsets. */
 149                 OUT_RING(NV40TCL_TEX_FILTER_MIN_NEAREST |
 150                                 NV40TCL_TEX_FILTER_MAG_NEAREST |
 151                                 0x3fd6);
 152         }
 153         OUT_RING((width << 16) | height);
 154         OUT_RING(0); /* border ARGB */
 155         BEGIN_RING(Nv3D, NV40TCL_TEX_SIZE1(unit), 1);
 156         OUT_RING((1 << NV40TCL_TEX_SIZE1_DEPTH_SHIFT) |
 157                         (uint16_t) src_pitch);
 158
 159         return TRUE;
 160 }
 161
 162 Bool
 163 NV40GetSurfaceFormat(PixmapPtr pPix, int *fmt_ret)
 164 {
 165         switch (pPix->drawable.bitsPerPixel) {
 166                 case 32:
 167                         *fmt_ret = NV40TCL_RT_FORMAT_COLOR_A8R8G8B8;
 168                         break;
 169                 case 24:
 170                         *fmt_ret = NV40TCL_RT_FORMAT_COLOR_X8R8G8B8;
 171                         break;
 172                 case 16:
 173                         *fmt_ret = NV40TCL_RT_FORMAT_COLOR_R5G6B5;
 174                         break;
 175                 case 8:
 176                         *fmt_ret = NV40TCL_RT_FORMAT_COLOR_B8;
 177                         break;
 178                 default:
 179                         return FALSE;
 180         }
 181
 182         return TRUE;
 183 }
 184
 185 #ifndef ExaOffscreenMarkUsed
 186 extern void ExaOffscreenMarkUsed(PixmapPtr);
 187 #endif
 188 #ifndef exaGetDrawablePixmap
 189 extern PixmapPtr exaGetDrawablePixmap(DrawablePtr);
 190 #endif
 191 #ifndef exaPixmapIsOffscreen
 192 extern Bool exaPixmapIsOffscreen(PixmapPtr p);
 193 #endif
 194 /* To support EXA 2.0, 2.1 has this in the header */
 195 #ifndef exaMoveInPixmap
 196 extern void exaMoveInPixmap(PixmapPtr pPixmap);
 197 #endif
 198
 199 #define SF(bf) (NV40TCL_BLEND_FUNC_SRC_RGB_##bf |                              \
 200                 NV40TCL_BLEND_FUNC_SRC_ALPHA_##bf)
 201 #define DF(bf) (NV40TCL_BLEND_FUNC_DST_RGB_##bf |                              \
 202                 NV40TCL_BLEND_FUNC_DST_ALPHA_##bf)
 203
 204 #define VERTEX_OUT(sx,sy,dx,dy) do {                                        \
 205         BEGIN_RING(Nv3D, NV40TCL_VTX_ATTR_2F_X(8), 4);                         \
 206         OUT_RINGf ((sx)); OUT_RINGf ((sy));                                    \
 207         OUT_RINGf ((sx)); OUT_RINGf ((sy));                                    \
 208         BEGIN_RING(Nv3D, NV40TCL_VTX_ATTR_2I(0), 1);                           \
 209         OUT_RING  (((dy)<<16)|(dx));                                           \
 210 } while(0)
 211
 212 void NV40PutTextureImage(ScrnInfoPtr pScrn, int src_offset,
 213                 int src_offset2, int id,
 214                 int src_pitch, BoxPtr dstBox,
 215                 int x1, int y1, int x2, int y2,
 216                 uint16_t width, uint16_t height,
 217                 uint16_t src_w, uint16_t src_h,
 218                 uint16_t drw_w, uint16_t drw_h,
 219                 RegionPtr clipBoxes,
 220                 DrawablePtr pDraw)
 221 {
 222         /* Remove some warnings. */
 223         /* This has to be done better at some point. */
 224         (void)nv40_vp_exa_render;
 225         (void)nv30_fp_pass_col0;
 226         (void)nv30_fp_pass_tex0;
 227         (void)nv30_fp_composite_mask;
 228         (void)nv30_fp_composite_mask_sa_ca;
 229         (void)nv30_fp_composite_mask_ca;
 230
 231         NVPtr pNv = NVPTR(pScrn);
 232         float X1, X2, X3, X4, Y1, Y2, Y3, Y4;
 233         ScreenPtr pScreen = pScrn->pScreen;
 234         PixmapPtr pPix = exaGetDrawablePixmap(pDraw);
 235         int dst_format = 0;
 236         if (!NV40GetSurfaceFormat(pPix, &dst_format)) {
 237                 ErrorF("No surface format, bad.\n");
 238         }
 239
 240         /* Try to get the dest drawable into vram */
 241         if (!exaPixmapIsOffscreen(pPix)) {
 242                 exaMoveInPixmap(pPix);
 243                 ExaOffscreenMarkUsed(pPix);
 244         }
 245
 246         /* If we failed, draw directly onto the screen pixmap.
 247          * Not sure if this is the best approach, maybe failing
 248          * with BadAlloc would be better?
 249          */
 250         if (!exaPixmapIsOffscreen(pPix)) {
 251                 xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
 252                         "XV: couldn't move dst surface into vram\n");
 253                 pPix = pScreen->GetScreenPixmap(pScreen);
 254         }
 255
 256 #ifdef COMPOSITE
 257         /* Adjust coordinates if drawing to an offscreen pixmap */
 258         if (pPix->screen_x || pPix->screen_y) {
 259                 REGION_TRANSLATE(pScrn->pScreen, clipBoxes,
 260                                                         -pPix->screen_x,
 261                                                         -pPix->screen_y);
 262                 dstBox->x1 -= pPix->screen_x;
 263                 dstBox->x2 -= pPix->screen_x;
 264                 dstBox->y1 -= pPix->screen_y;
 265                 dstBox->y2 -= pPix->screen_y;
 266         }
 267
 268         DamageDamageRegion((DrawablePtr)pPix, clipBoxes);
 269 #endif
 270
 271         /* Disable blending */
 272         BEGIN_RING(Nv3D, NV40TCL_BLEND_ENABLE, 1);
 273         OUT_RING(0);
 274
 275         /* Setup surface */
 276         BEGIN_RING(Nv3D, NV40TCL_RT_FORMAT, 3);
 277         OUT_RING  (NV40TCL_RT_FORMAT_TYPE_LINEAR |
 278                         NV40TCL_RT_FORMAT_ZETA_Z24S8 |
 279                         dst_format);
 280         OUT_RING  (exaGetPixmapPitch(pPix));
 281         OUT_PIXMAPl(pPix, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
 282
 283         NV40VideoTexture(pScrn, src_offset, width, height, src_pitch, 0);
 284         /* We've got NV12 format, which means half width and half height texture of chroma channels. */
 285         NV40VideoTexture(pScrn, src_offset2, width/2, height/2, src_pitch, 1);
 286
 287         NV40_LoadVtxProg(pScrn, &nv40_video);
 288         NV40_LoadFragProg(pScrn, &nv40_yv12);
 289
 290         /* Appears to be some kind of cache flush, needed here at least
 291          * sometimes.. funky text rendering otherwise :)
 292          */
 293         BEGIN_RING(Nv3D, NV40TCL_TEX_CACHE_CTL, 1);
 294         OUT_RING  (2);
 295         BEGIN_RING(Nv3D, NV40TCL_TEX_CACHE_CTL, 1);
 296         OUT_RING  (1);
 297
 298         BEGIN_RING(Nv3D, NV40TCL_BEGIN_END, 1);
 299         OUT_RING  (NV40TCL_BEGIN_END_QUADS);
 300
 301         /* These are fixed point values in the 16.16 format. */
 302         x1 >>= 16;
 303         x2 >>= 16;
 304         y1 >>= 16;
 305         y2 >>= 16;
 306
 307         X1 = (float)x1/(float)src_w;
 308         Y1 = (float)y1/(float)src_h;
 309         X2 = (float)x2/(float)src_w;
 310         Y2 = (float)y1/(float)src_h;
 311         X3 = (float)x2/(float)src_w;
 312         Y3 = (float)y2/(float)src_h;
 313         X4 = (float)x1/(float)src_w;
 314         Y4 = (float)y2/(float)src_h;
 315
 316         /* Submit the appropriate vertices. */
 317         /* This submits the same vertices for the Y and the UV texture. */
 318         VERTEX_OUT(X1, Y1, dstBox->x1, dstBox->y1);
 319         VERTEX_OUT(X2, Y2, dstBox->x2, dstBox->y1);
 320         VERTEX_OUT(X3, Y3, dstBox->x2, dstBox->y2);
 321         VERTEX_OUT(X4, Y4, dstBox->x1, dstBox->y2);
 322
 323         FIRE_RING();
 324
 325         BEGIN_RING(Nv3D, NV40TCL_BEGIN_END, 1);
 326         OUT_RING  (NV40TCL_BEGIN_END_STOP);
 327 }