src/gallium/drivers/nv50/nv50_transfer.c

   1
   2 #include "pipe/p_context.h"
   3 #include "pipe/p_inlines.h"
   4 #include "util/u_format.h"
   5 #include "util/u_math.h"
   6
   7 #include "nv50_context.h"
   8
   9 struct nv50_transfer {
  10         struct pipe_transfer base;
  11         struct nouveau_bo *bo;
  12         unsigned level_offset;
  13         unsigned level_tiling;
  14         int level_pitch;
  15         int level_width;
  16         int level_height;
  17         int level_depth;
  18         int level_x;
  19         int level_y;
  20         int level_z;
  21         unsigned nblocksx;
  22         unsigned nblocksy;
  23 };
  24
  25 static void
  26 nv50_transfer_rect_m2mf(struct pipe_screen *pscreen,
  27                         struct nouveau_bo *src_bo, unsigned src_offset,
  28                         int src_pitch, unsigned src_tile_mode,
  29                         int sx, int sy, int sz, int sw, int sh, int sd,
  30                         struct nouveau_bo *dst_bo, unsigned dst_offset,
  31                         int dst_pitch, unsigned dst_tile_mode,
  32                         int dx, int dy, int dz, int dw, int dh, int dd,
  33                         int cpp, int width, int height,
  34                         unsigned src_reloc, unsigned dst_reloc)
  35 {
  36         struct nv50_screen *screen = nv50_screen(pscreen);
  37         struct nouveau_channel *chan = screen->m2mf->channel;
  38         struct nouveau_grobj *m2mf = screen->m2mf;
  39
  40         src_reloc |= NOUVEAU_BO_RD;
  41         dst_reloc |= NOUVEAU_BO_WR;
  42
  43         WAIT_RING (chan, 14);
  44
  45         if (!src_bo->tile_flags) {
  46                 BEGIN_RING(chan, m2mf,
  47                         NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_IN, 1);
  48                 OUT_RING  (chan, 1);
  49                 BEGIN_RING(chan, m2mf,
  50                         NV04_MEMORY_TO_MEMORY_FORMAT_PITCH_IN, 1);
  51                 OUT_RING  (chan, src_pitch);
  52                 src_offset += (sy * src_pitch) + (sx * cpp);
  53         } else {
  54                 BEGIN_RING(chan, m2mf,
  55                         NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_IN, 6);
  56                 OUT_RING  (chan, 0);
  57                 OUT_RING  (chan, src_tile_mode << 4);
  58                 OUT_RING  (chan, sw * cpp);
  59                 OUT_RING  (chan, sh);
  60                 OUT_RING  (chan, sd);
  61                 OUT_RING  (chan, sz); /* copying only 1 zslice per call */
  62         }
  63
  64         if (!dst_bo->tile_flags) {
  65                 BEGIN_RING(chan, m2mf,
  66                         NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_OUT, 1);
  67                 OUT_RING  (chan, 1);
  68                 BEGIN_RING(chan, m2mf,
  69                         NV04_MEMORY_TO_MEMORY_FORMAT_PITCH_OUT, 1);
  70                 OUT_RING  (chan, dst_pitch);
  71                 dst_offset += (dy * dst_pitch) + (dx * cpp);
  72         } else {
  73                 BEGIN_RING(chan, m2mf,
  74                         NV50_MEMORY_TO_MEMORY_FORMAT_LINEAR_OUT, 6);
  75                 OUT_RING  (chan, 0);
  76                 OUT_RING  (chan, dst_tile_mode << 4);
  77                 OUT_RING  (chan, dw * cpp);
  78                 OUT_RING  (chan, dh);
  79                 OUT_RING  (chan, dd);
  80                 OUT_RING  (chan, dz); /* copying only 1 zslice per call */
  81         }
  82
  83         while (height) {
  84                 int line_count = height > 2047 ? 2047 : height;
  85
  86                 MARK_RING (chan, 15, 4); /* flush on lack of space or relocs */
  87                 BEGIN_RING(chan, m2mf,
  88                         NV50_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN_HIGH, 2);
  89                 OUT_RELOCh(chan, src_bo, src_offset, src_reloc);
  90                 OUT_RELOCh(chan, dst_bo, dst_offset, dst_reloc);
  91                 BEGIN_RING(chan, m2mf,
  92                         NV04_MEMORY_TO_MEMORY_FORMAT_OFFSET_IN, 2);
  93                 OUT_RELOCl(chan, src_bo, src_offset, src_reloc);
  94                 OUT_RELOCl(chan, dst_bo, dst_offset, dst_reloc);
  95                 if (src_bo->tile_flags) {
  96                         BEGIN_RING(chan, m2mf,
  97                                 NV50_MEMORY_TO_MEMORY_FORMAT_TILING_POSITION_IN, 1);
  98                         OUT_RING  (chan, (sy << 16) | (sx * cpp));
  99                 } else {
 100                         src_offset += (line_count * src_pitch);
 101                 }
 102                 if (dst_bo->tile_flags) {
 103                         BEGIN_RING(chan, m2mf,
 104                                 NV50_MEMORY_TO_MEMORY_FORMAT_TILING_POSITION_OUT, 1);
 105                         OUT_RING  (chan, (dy << 16) | (dx * cpp));
 106                 } else {
 107                         dst_offset += (line_count * dst_pitch);
 108                 }
 109                 BEGIN_RING(chan, m2mf,
 110                         NV04_MEMORY_TO_MEMORY_FORMAT_LINE_LENGTH_IN, 4);
 111                 OUT_RING  (chan, width * cpp);
 112                 OUT_RING  (chan, line_count);
 113                 OUT_RING  (chan, 0x00000101);
 114                 OUT_RING  (chan, 0);
 115                 FIRE_RING (chan);
 116
 117                 height -= line_count;
 118                 sy += line_count;
 119                 dy += line_count;
 120         }
 121 }
 122
 123 static struct pipe_transfer *
 124 nv50_transfer_new(struct pipe_screen *pscreen, struct pipe_texture *pt,
 125                   unsigned face, unsigned level, unsigned zslice,
 126                   enum pipe_transfer_usage usage,
 127                   unsigned x, unsigned y, unsigned w, unsigned h)
 128 {
 129         struct nouveau_device *dev = nouveau_screen(pscreen)->device;
 130         struct nv50_miptree *mt = nv50_miptree(pt);
 131         struct nv50_miptree_level *lvl = &mt->level[level];
 132         struct nv50_transfer *tx;
 133         unsigned nx, ny, image = 0;
 134         int ret;
 135
 136         if (pt->target == PIPE_TEXTURE_CUBE)
 137                 image = face;
 138
 139         tx = CALLOC_STRUCT(nv50_transfer);
 140         if (!tx)
 141                 return NULL;
 142
 143         pipe_texture_reference(&tx->base.texture, pt);
 144         tx->nblocksx = util_format_get_nblocksx(pt->format, u_minify(pt->width0, level));
 145         tx->nblocksy = util_format_get_nblocksy(pt->format, u_minify(pt->height0, level));
 146         tx->base.width = w;
 147         tx->base.height = h;
 148         tx->base.stride = tx->nblocksx * util_format_get_blocksize(pt->format);
 149         tx->base.usage = usage;
 150
 151         tx->level_pitch = lvl->pitch;
 152         tx->level_width = u_minify(mt->base.base.width0, level);
 153         tx->level_height = u_minify(mt->base.base.height0, level);
 154         tx->level_depth = u_minify(mt->base.base.depth0, level);
 155         tx->level_offset = lvl->image_offset[image];
 156         tx->level_tiling = lvl->tile_mode;
 157         tx->level_z = zslice;
 158         tx->level_x = util_format_get_nblocksx(pt->format, x);
 159         tx->level_y = util_format_get_nblocksy(pt->format, y);
 160         ret = nouveau_bo_new(dev, NOUVEAU_BO_GART | NOUVEAU_BO_MAP, 0,
 161                              tx->nblocksy * tx->base.stride, &tx->bo);
 162         if (ret) {
 163                 FREE(tx);
 164                 return NULL;
 165         }
 166
 167         if (usage & PIPE_TRANSFER_READ) {
 168                 nx = util_format_get_nblocksx(pt->format, tx->base.width);
 169                 ny = util_format_get_nblocksy(pt->format, tx->base.height);
 170
 171                 nv50_transfer_rect_m2mf(pscreen, mt->base.bo, tx->level_offset,
 172                                         tx->level_pitch, tx->level_tiling,
 173                                         x, y, zslice,
 174                                         tx->nblocksx, tx->nblocksy,
 175                                         tx->level_depth,
 176                                         tx->bo, 0,
 177                                         tx->base.stride, tx->bo->tile_mode,
 178                                         0, 0, 0,
 179                                         tx->nblocksx, tx->nblocksy, 1,
 180                                         util_format_get_blocksize(pt->format), nx, ny,
 181                                         NOUVEAU_BO_VRAM | NOUVEAU_BO_GART,
 182                                         NOUVEAU_BO_GART);
 183         }
 184
 185         return &tx->base;
 186 }
 187
 188 static void
 189 nv50_transfer_del(struct pipe_transfer *ptx)
 190 {
 191         struct nv50_transfer *tx = (struct nv50_transfer *)ptx;
 192         struct nv50_miptree *mt = nv50_miptree(ptx->texture);
 193         struct pipe_texture *pt = ptx->texture;
 194
 195         unsigned nx = util_format_get_nblocksx(pt->format, tx->base.width);
 196         unsigned ny = util_format_get_nblocksy(pt->format, tx->base.height);
 197
 198         if (ptx->usage & PIPE_TRANSFER_WRITE) {
 199                 struct pipe_screen *pscreen = pt->screen;
 200
 201                 nv50_transfer_rect_m2mf(pscreen, tx->bo, 0,
 202                                         tx->base.stride, tx->bo->tile_mode,
 203                                         0, 0, 0,
 204                                         tx->nblocksx, tx->nblocksy, 1,
 205                                         mt->base.bo, tx->level_offset,
 206                                         tx->level_pitch, tx->level_tiling,
 207                                         tx->level_x, tx->level_y, tx->level_z,
 208                                         tx->nblocksx, tx->nblocksy,
 209                                         tx->level_depth,
 210                                         util_format_get_blocksize(pt->format), nx, ny,
 211                                         NOUVEAU_BO_GART, NOUVEAU_BO_VRAM |
 212                                         NOUVEAU_BO_GART);
 213         }
 214
 215         nouveau_bo_ref(NULL, &tx->bo);
 216         pipe_texture_reference(&ptx->texture, NULL);
 217         FREE(ptx);
 218 }
 219
 220 static void *
 221 nv50_transfer_map(struct pipe_screen *pscreen, struct pipe_transfer *ptx)
 222 {
 223         struct nv50_transfer *tx = (struct nv50_transfer *)ptx;
 224         unsigned flags = 0;
 225         int ret;
 226
 227         if (ptx->usage & PIPE_TRANSFER_WRITE)
 228                 flags |= NOUVEAU_BO_WR;
 229         if (ptx->usage & PIPE_TRANSFER_READ)
 230                 flags |= NOUVEAU_BO_RD;
 231
 232         ret = nouveau_bo_map(tx->bo, flags);
 233         if (ret)
 234                 return NULL;
 235         return tx->bo->map;
 236 }
 237
 238 static void
 239 nv50_transfer_unmap(struct pipe_screen *pscreen, struct pipe_transfer *ptx)
 240 {
 241         struct nv50_transfer *tx = (struct nv50_transfer *)ptx;
 242
 243         nouveau_bo_unmap(tx->bo);
 244 }
 245
 246 void
 247 nv50_transfer_init_screen_functions(struct pipe_screen *pscreen)
 248 {
 249         pscreen->get_tex_transfer = nv50_transfer_new;
 250         pscreen->tex_transfer_destroy = nv50_transfer_del;
 251         pscreen->transfer_map = nv50_transfer_map;
 252         pscreen->transfer_unmap = nv50_transfer_unmap;
 253 }
 254
 255 void
 256 nv50_upload_sifc(struct nv50_context *nv50,
 257                  struct nouveau_bo *bo, unsigned dst_offset, unsigned reloc,
 258                  unsigned dst_format, int dst_w, int dst_h, int dst_pitch,
 259                  void *src, unsigned src_format, int src_pitch,
 260                  int x, int y, int w, int h, int cpp)
 261 {
 262         struct nouveau_channel *chan = nv50->screen->base.channel;
 263         struct nouveau_grobj *eng2d = nv50->screen->eng2d;
 264         struct nouveau_grobj *tesla = nv50->screen->tesla;
 265         unsigned line_dwords = (w * cpp + 3) / 4;
 266
 267         reloc |= NOUVEAU_BO_WR;
 268
 269         MARK_RING (chan, 32, 2); /* flush on lack of space or relocs */
 270
 271         if (bo->tile_flags) {
 272                 BEGIN_RING(chan, eng2d, NV50_2D_DST_FORMAT, 5);
 273                 OUT_RING  (chan, dst_format);
 274                 OUT_RING  (chan, 0);
 275                 OUT_RING  (chan, bo->tile_mode << 4);
 276                 OUT_RING  (chan, 1);
 277                 OUT_RING  (chan, 0);
 278         } else {
 279                 BEGIN_RING(chan, eng2d, NV50_2D_DST_FORMAT, 2);
 280                 OUT_RING  (chan, dst_format);
 281                 OUT_RING  (chan, 1);
 282                 BEGIN_RING(chan, eng2d, NV50_2D_DST_PITCH, 1);
 283                 OUT_RING  (chan, dst_pitch);
 284         }
 285
 286         BEGIN_RING(chan, eng2d, NV50_2D_DST_WIDTH, 4);
 287         OUT_RING  (chan, dst_w);
 288         OUT_RING  (chan, dst_h);
 289         OUT_RELOCh(chan, bo, dst_offset, reloc);
 290         OUT_RELOCl(chan, bo, dst_offset, reloc);
 291
 292         /* NV50_2D_OPERATION_SRCCOPY assumed already set */
 293
 294         BEGIN_RING(chan, eng2d, NV50_2D_SIFC_BITMAP_ENABLE, 2);
 295         OUT_RING  (chan, 0);
 296         OUT_RING  (chan, src_format);
 297         BEGIN_RING(chan, eng2d, NV50_2D_SIFC_WIDTH, 10);
 298         OUT_RING  (chan, w);
 299         OUT_RING  (chan, h);
 300         OUT_RING  (chan, 0);
 301         OUT_RING  (chan, 1);
 302         OUT_RING  (chan, 0);
 303         OUT_RING  (chan, 1);
 304         OUT_RING  (chan, 0);
 305         OUT_RING  (chan, x);
 306         OUT_RING  (chan, 0);
 307         OUT_RING  (chan, y);
 308
 309         while (h--) {
 310                 const uint32_t *p = src;
 311                 unsigned count = line_dwords;
 312
 313                 while (count) {
 314                         unsigned nr = MIN2(count, 1792);
 315
 316                         if (chan->pushbuf->remaining <= nr) {
 317                                 FIRE_RING (chan);
 318
 319                                 BEGIN_RING(chan, eng2d,
 320                                            NV50_2D_DST_ADDRESS_HIGH, 2);
 321                                 OUT_RELOCh(chan, bo, dst_offset, reloc);
 322                                 OUT_RELOCl(chan, bo, dst_offset, reloc);
 323                         }
 324                         assert(chan->pushbuf->remaining > nr);
 325
 326                         BEGIN_RING(chan, eng2d,
 327                                    NV50_2D_SIFC_DATA | (2 << 29), nr);
 328                         OUT_RINGp (chan, p, nr);
 329
 330                         p += nr;
 331                         count -= nr;
 332                 }
 333
 334                 src += src_pitch;
 335         }
 336
 337         BEGIN_RING(chan, tesla, NV50TCL_CODE_CB_FLUSH, 1);
 338         OUT_RING  (chan, 0);
 339 }