src/gallium/drivers/nv50/nv50_transfer.c

   1
   2 #include "pipe/p_context.h"
   3 #include "pipe/p_inlines.h"
   4 #include "util/u_math.h"
   5
   6 #include "nv50_context.h"
   7
   8 struct nv50_transfer {
   9         struct pipe_transfer base;
  10         struct nouveau_bo *bo;
  11         unsigned level_offset;
  12         unsigned level_tiling;
  13         int level_pitch;
  14         int level_width;
  15         int level_height;
  16         int level_depth;
  17         int level_x;
  18         int level_y;
  19         unsigned nblocksx;
  20         unsigned nblocksy;
  21 };
  22
  23 static void
  24 nv50_transfer_rect_m2mf(struct pipe_screen *pscreen,
  25                         struct nouveau_bo *src_bo, unsigned src_offset,
  26                         int src_pitch, unsigned src_tile_mode,
  27                         int sx, int sy, int sw, int sh, int sd,
  28                         struct nouveau_bo *dst_bo, unsigned dst_offset,
  29                         int dst_pitch, unsigned dst_tile_mode,
  30                         int dx, int dy, int dw, int dh, int dd,
  31                         int cpp, int width, int height,
  32                         unsigned src_reloc, unsigned dst_reloc)
  33 {
  34         struct nv50_screen *screen = nv50_screen(pscreen);
  35         struct nouveau_channel *chan = screen->m2mf->channel;
  36         struct nouveau_grobj *m2mf = screen->m2mf;
  37
  38         src_reloc |= NOUVEAU_BO_RD;
  39         dst_reloc |= NOUVEAU_BO_WR;
  40
  41         WAIT_RING (chan, 14);
  42
  43         if (!src_bo->tile_flags) {
  44                 BEGIN_RING(chan, m2mf,
  45                         NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_IN, 1);
  46                 OUT_RING  (chan, 1);
  47                 BEGIN_RING(chan, m2mf,
  48                         NV50_MEMORY_TO_MEMORY_FORMAT_PITCH_IN, 1);
  49                 OUT_RING  (chan, src_pitch);
  50                 src_offset += (sy * src_pitch) + (sx * cpp);
  51         } else {
  52                 BEGIN_RING(chan, m2mf,
  53                         NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_IN, 6);
  54                 OUT_RING  (chan, 0);
  55                 OUT_RING  (chan, src_tile_mode << 4);
  56                 OUT_RING  (chan, sw * cpp);
  57                 OUT_RING  (chan, sh);
  58                 OUT_RING  (chan, sd);
  59                 OUT_RING  (chan, 0);
  60         }
  61
  62         if (!dst_bo->tile_flags) {
  63                 BEGIN_RING(chan, m2mf,
  64                         NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_OUT, 1);
  65                 OUT_RING  (chan, 1);
  66                 BEGIN_RING(chan, m2mf,
  67                         NV50_MEMORY_TO_MEMORY_FORMAT_PITCH_OUT, 1);
  68                 OUT_RING  (chan, dst_pitch);
  69                 dst_offset += (dy * dst_pitch) + (dx * cpp);
  70         } else {
  71                 BEGIN_RING(chan, m2mf,
  72                         NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_OUT, 6);
  73                 OUT_RING  (chan, 0);
  74                 OUT_RING  (chan, dst_tile_mode << 4);
  75                 OUT_RING  (chan, dw * cpp);
  76                 OUT_RING  (chan, dh);
  77                 OUT_RING  (chan, dd);
  78                 OUT_RING  (chan, 0);
  79         }
  80
  81         while (height) {
  82                 int line_count = height > 2047 ? 2047 : height;
  83
  84                 WAIT_RING (chan, 15);
  85                 BEGIN_RING(chan, m2mf,
  86                         NV50_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN_HIGH, 2);
  87                 OUT_RELOCh(chan, src_bo, src_offset, src_reloc);
  88                 OUT_RELOCh(chan, dst_bo, dst_offset, dst_reloc);
  89                 BEGIN_RING(chan, m2mf,
  90                         NV50_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 2);
  91                 OUT_RELOCl(chan, src_bo, src_offset, src_reloc);
  92                 OUT_RELOCl(chan, dst_bo, dst_offset, dst_reloc);
  93                 if (src_bo->tile_flags) {
  94                         BEGIN_RING(chan, m2mf,
  95                                 NV50_MEMORY_TO_MEMORY_FORMAT_TILING_POSITION_IN, 1);
  96                         OUT_RING  (chan, (sy << 16) | (sx * cpp));
  97                 } else {
  98                         src_offset += (line_count * src_pitch);
  99                 }
 100                 if (dst_bo->tile_flags) {
 101                         BEGIN_RING(chan, m2mf,
 102                                 NV50_MEMORY_TO_MEMORY_FORMAT_TILING_POSITION_OUT, 1);
 103                         OUT_RING  (chan, (dy << 16) | (dx * cpp));
 104                 } else {
 105                         dst_offset += (line_count * dst_pitch);
 106                 }
 107                 BEGIN_RING(chan, m2mf,
 108                         NV50_MEMORY_TO_MEMORY_FORMAT_LINE_LENGTH_IN, 4);
 109                 OUT_RING  (chan, width * cpp);
 110                 OUT_RING  (chan, line_count);
 111                 OUT_RING  (chan, 0x00000101);
 112                 OUT_RING  (chan, 0);
 113                 FIRE_RING (chan);
 114
 115                 height -= line_count;
 116                 sy += line_count;
 117                 dy += line_count;
 118         }
 119 }
 120
 121 static INLINE unsigned
 122 get_zslice_offset(unsigned tile_mode, unsigned z, unsigned pitch, unsigned ny)
 123 {
 124         unsigned tile_h = get_tile_height(tile_mode);
 125         unsigned tile_d = get_tile_depth(tile_mode);
 126
 127         /* pitch_2d == to next slice within this volume-tile */
 128         /* pitch_3d == to next slice in next 2D array of blocks */
 129         unsigned pitch_2d = tile_h * 64;
 130         unsigned pitch_3d = tile_d * align(ny, tile_h) * pitch;
 131
 132         return (z % tile_d) * pitch_2d + (z / tile_d) * pitch_3d;
 133 }
 134
 135 static struct pipe_transfer *
 136 nv50_transfer_new(struct pipe_screen *pscreen, struct pipe_texture *pt,
 137                   unsigned face, unsigned level, unsigned zslice,
 138                   enum pipe_transfer_usage usage,
 139                   unsigned x, unsigned y, unsigned w, unsigned h)
 140 {
 141         struct nouveau_device *dev = nouveau_screen(pscreen)->device;
 142         struct nv50_miptree *mt = nv50_miptree(pt);
 143         struct nv50_miptree_level *lvl = &mt->level[level];
 144         struct nv50_transfer *tx;
 145         unsigned nx, ny, image = 0;
 146         int ret;
 147
 148         if (pt->target == PIPE_TEXTURE_CUBE)
 149                 image = face;
 150
 151         tx = CALLOC_STRUCT(nv50_transfer);
 152         if (!tx)
 153                 return NULL;
 154
 155         pipe_texture_reference(&tx->base.texture, pt);
 156         tx->nblocksx = pf_get_nblocksx(pt->format, u_minify(pt->width0, level));
 157         tx->nblocksy = pf_get_nblocksy(pt->format, u_minify(pt->height0, level));
 158         tx->base.width = w;
 159         tx->base.height = h;
 160         tx->base.stride = tx->nblocksx * pf_get_blocksize(pt->format);
 161         tx->base.usage = usage;
 162
 163         tx->level_pitch = lvl->pitch;
 164         tx->level_width = u_minify(mt->base.base.width0, level);
 165         tx->level_height = u_minify(mt->base.base.height0, level);
 166         tx->level_depth = u_minify(mt->base.base.depth0, level);
 167         tx->level_offset = lvl->image_offset[image];
 168         tx->level_tiling = lvl->tile_mode;
 169         tx->level_x = pf_get_nblocksx(pt->format, x);
 170         tx->level_y = pf_get_nblocksy(pt->format, y);
 171         ret = nouveau_bo_new(dev, NOUVEAU_BO_GART | NOUVEAU_BO_MAP, 0,
 172                              tx->nblocksy * tx->base.stride, &tx->bo);
 173         if (ret) {
 174                 FREE(tx);
 175                 return NULL;
 176         }
 177
 178         if (pt->target == PIPE_TEXTURE_3D)
 179                 tx->level_offset += get_zslice_offset(lvl->tile_mode, zslice,
 180                                                       lvl->pitch,
 181                                                       tx->nblocksy);
 182
 183         if (usage & PIPE_TRANSFER_READ) {
 184                 nx = pf_get_nblocksx(pt->format, tx->base.width);
 185                 ny = pf_get_nblocksy(pt->format, tx->base.height);
 186
 187                 nv50_transfer_rect_m2mf(pscreen, mt->base.bo, tx->level_offset,
 188                                         tx->level_pitch, tx->level_tiling,
 189                                         x, y,
 190                                         tx->nblocksx, tx->nblocksy,
 191                                         tx->level_depth,
 192                                         tx->bo, 0,
 193                                         tx->base.stride, tx->bo->tile_mode,
 194                                         0, 0,
 195                                         tx->nblocksx, tx->nblocksy, 1,
 196                                         pf_get_blocksize(pt->format), nx, ny,
 197                                         NOUVEAU_BO_VRAM | NOUVEAU_BO_GART,
 198                                         NOUVEAU_BO_GART);
 199         }
 200
 201         return &tx->base;
 202 }
 203
 204 static void
 205 nv50_transfer_del(struct pipe_transfer *ptx)
 206 {
 207         struct nv50_transfer *tx = (struct nv50_transfer *)ptx;
 208         struct nv50_miptree *mt = nv50_miptree(ptx->texture);
 209         struct pipe_texture *pt = ptx->texture;
 210
 211         unsigned nx = pf_get_nblocksx(pt->format, tx->base.width);
 212         unsigned ny = pf_get_nblocksy(pt->format, tx->base.height);
 213
 214         if (ptx->usage & PIPE_TRANSFER_WRITE) {
 215                 struct pipe_screen *pscreen = pt->screen;
 216
 217                 nv50_transfer_rect_m2mf(pscreen, tx->bo, 0,
 218                                         tx->base.stride, tx->bo->tile_mode,
 219                                         0, 0,
 220                                         tx->nblocksx, tx->nblocksy, 1,
 221                                         mt->base.bo, tx->level_offset,
 222                                         tx->level_pitch, tx->level_tiling,
 223                                         tx->level_x, tx->level_y,
 224                                         tx->nblocksx, tx->nblocksy,
 225                                         tx->level_depth,
 226                                         pf_get_blocksize(pt->format), nx, ny,
 227                                         NOUVEAU_BO_GART, NOUVEAU_BO_VRAM |
 228                                         NOUVEAU_BO_GART);
 229         }
 230
 231         nouveau_bo_ref(NULL, &tx->bo);
 232         pipe_texture_reference(&ptx->texture, NULL);
 233         FREE(ptx);
 234 }
 235
 236 static void *
 237 nv50_transfer_map(struct pipe_screen *pscreen, struct pipe_transfer *ptx)
 238 {
 239         struct nv50_transfer *tx = (struct nv50_transfer *)ptx;
 240         unsigned flags = 0;
 241         int ret;
 242
 243         if (ptx->usage & PIPE_TRANSFER_WRITE)
 244                 flags |= NOUVEAU_BO_WR;
 245         if (ptx->usage & PIPE_TRANSFER_READ)
 246                 flags |= NOUVEAU_BO_RD;
 247
 248         ret = nouveau_bo_map(tx->bo, flags);
 249         if (ret)
 250                 return NULL;
 251         return tx->bo->map;
 252 }
 253
 254 static void
 255 nv50_transfer_unmap(struct pipe_screen *pscreen, struct pipe_transfer *ptx)
 256 {
 257         struct nv50_transfer *tx = (struct nv50_transfer *)ptx;
 258
 259         nouveau_bo_unmap(tx->bo);
 260 }
 261
 262 void
 263 nv50_transfer_init_screen_functions(struct pipe_screen *pscreen)
 264 {
 265         pscreen->get_tex_transfer = nv50_transfer_new;
 266         pscreen->tex_transfer_destroy = nv50_transfer_del;
 267         pscreen->transfer_map = nv50_transfer_map;
 268         pscreen->transfer_unmap = nv50_transfer_unmap;
 269 }
 270
 271 void
 272 nv50_upload_sifc(struct nv50_context *nv50,
 273                  struct nouveau_bo *bo, unsigned dst_offset, unsigned reloc,
 274                  unsigned dst_format, int dst_w, int dst_h, int dst_pitch,
 275                  void *src, unsigned src_format, int src_pitch,
 276                  int x, int y, int w, int h, int cpp)
 277 {
 278         struct nouveau_channel *chan = nv50->screen->base.channel;
 279         struct nouveau_grobj *eng2d = nv50->screen->eng2d;
 280         struct nouveau_grobj *tesla = nv50->screen->tesla;
 281         unsigned line_dwords = (w * cpp + 3) / 4;
 282
 283         reloc |= NOUVEAU_BO_WR;
 284
 285         WAIT_RING (chan, 32);
 286
 287         if (bo->tile_flags) {
 288                 BEGIN_RING(chan, eng2d, NV50_2D_DST_FORMAT, 5);
 289                 OUT_RING  (chan, dst_format);
 290                 OUT_RING  (chan, 0);
 291                 OUT_RING  (chan, bo->tile_mode << 4);
 292                 OUT_RING  (chan, 1);
 293                 OUT_RING  (chan, 0);
 294         } else {
 295                 BEGIN_RING(chan, eng2d, NV50_2D_DST_FORMAT, 2);
 296                 OUT_RING  (chan, dst_format);
 297                 OUT_RING  (chan, 1);
 298                 BEGIN_RING(chan, eng2d, NV50_2D_DST_PITCH, 1);
 299                 OUT_RING  (chan, dst_pitch);
 300         }
 301
 302         BEGIN_RING(chan, eng2d, NV50_2D_DST_WIDTH, 4);
 303         OUT_RING  (chan, dst_w);
 304         OUT_RING  (chan, dst_h);
 305         OUT_RELOCh(chan, bo, dst_offset, reloc);
 306         OUT_RELOCl(chan, bo, dst_offset, reloc);
 307
 308         /* NV50_2D_OPERATION_SRCCOPY assumed already set */
 309
 310         BEGIN_RING(chan, eng2d, NV50_2D_SIFC_UNK0800, 2);
 311         OUT_RING  (chan, 0);
 312         OUT_RING  (chan, src_format);
 313         BEGIN_RING(chan, eng2d, NV50_2D_SIFC_WIDTH, 10);
 314         OUT_RING  (chan, w);
 315         OUT_RING  (chan, h);
 316         OUT_RING  (chan, 0);
 317         OUT_RING  (chan, 1);
 318         OUT_RING  (chan, 0);
 319         OUT_RING  (chan, 1);
 320         OUT_RING  (chan, 0);
 321         OUT_RING  (chan, x);
 322         OUT_RING  (chan, 0);
 323         OUT_RING  (chan, y);
 324
 325         while (h--) {
 326                 const uint32_t *p = src;
 327                 unsigned count = line_dwords;
 328
 329                 while (count) {
 330                         unsigned nr = MIN2(count, 1792);
 331
 332                         if (chan->pushbuf->remaining <= nr) {
 333                                 FIRE_RING (chan);
 334
 335                                 BEGIN_RING(chan, eng2d,
 336                                            NV50_2D_DST_ADDRESS_HIGH, 2);
 337                                 OUT_RELOCh(chan, bo, dst_offset, reloc);
 338                                 OUT_RELOCl(chan, bo, dst_offset, reloc);
 339                         }
 340                         assert(chan->pushbuf->remaining > nr);
 341
 342                         BEGIN_RING(chan, eng2d,
 343                                    NV50_2D_SIFC_DATA | (2 << 29), nr);
 344                         OUT_RINGp (chan, p, nr);
 345
 346                         p += nr;
 347                         count -= nr;
 348                 }
 349
 350                 src += src_pitch;
 351         }
 352
 353         BEGIN_RING(chan, tesla, 0x1440, 1);
 354         OUT_RING  (chan, 0);
 355 }