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