src/gallium/drivers/nvfx/nvfx_state_fb.c

   1 #include "nvfx_context.h"
   2 #include "nvfx_resource.h"
   3 #include "util/u_format.h"
   4
   5 static inline boolean
   6 nvfx_surface_linear_renderable(struct pipe_surface* surf)
   7 {
   8         /* TODO: precompute this in nvfx_surface creation */
   9         return (surf->texture->flags & NVFX_RESOURCE_FLAG_LINEAR)
  10                 && !(surf->offset & 63)
  11                 && !(((struct nvfx_surface*)surf)->pitch & 63);
  12 }
  13
  14 static inline boolean
  15 nvfx_surface_swizzled_renderable(struct pipe_framebuffer_state* fb, struct pipe_surface* surf)
  16 {
  17         /* TODO: precompute this in nvfx_surface creation */
  18         return !((struct nvfx_miptree*)surf->texture)->linear_pitch
  19                 && (surf->texture->target != PIPE_TEXTURE_3D || u_minify(surf->texture->depth0, surf->level) <= 1)
  20                 && !(surf->offset & 127)
  21                 && (surf->width == fb->width)
  22                 && (surf->height == fb->height)
  23                 && !((struct nvfx_surface*)surf)->temp
  24                 && (surf->format == PIPE_FORMAT_B8G8R8A8_UNORM || surf->format == PIPE_FORMAT_B8G8R8X8_UNORM || surf->format == PIPE_FORMAT_B5G6R5_UNORM);
  25 }
  26
  27 static boolean
  28 nvfx_surface_get_render_target(struct pipe_surface* surf, int all_swizzled, struct nvfx_render_target* target)
  29 {
  30         struct nvfx_surface* ns = (struct nvfx_surface*)surf;
  31         if(!ns->temp)
  32         {
  33                 target->bo = ((struct nvfx_miptree*)surf->texture)->base.bo;
  34                 target->offset = surf->offset;
  35                 target->pitch = align(ns->pitch, 64);
  36                 assert(target->pitch);
  37                 return FALSE;
  38         }
  39         else
  40         {
  41                 target->offset = 0;
  42                 target->pitch = ns->temp->linear_pitch;
  43                 target->bo = ns->temp->base.bo;
  44                 assert(target->pitch);
  45                 return TRUE;
  46         }
  47 }
  48
  49 int
  50 nvfx_framebuffer_prepare(struct nvfx_context *nvfx)
  51 {
  52         struct pipe_framebuffer_state *fb = &nvfx->framebuffer;
  53         int i, color_format = 0, zeta_format = 0;
  54         int all_swizzled = 1;
  55
  56         if(!nvfx->is_nv4x)
  57                 assert(fb->nr_cbufs <= 2);
  58         else
  59                 assert(fb->nr_cbufs <= 4);
  60
  61         for (i = 0; i < fb->nr_cbufs; i++) {
  62                 if (color_format) {
  63                         if(color_format != fb->cbufs[i]->format)
  64                                 return -1;
  65                 } else
  66                         color_format = fb->cbufs[i]->format;
  67
  68                 if(!nvfx_surface_swizzled_renderable(fb, fb->cbufs[i]))
  69                         all_swizzled = 0;
  70         }
  71
  72         if (fb->zsbuf) {
  73                 /* TODO: return FALSE if we have a format not supporting a depth buffer (e.g. r8); currently those are not supported at all */
  74                 if(!nvfx_surface_swizzled_renderable(fb, fb->zsbuf))
  75                         all_swizzled = 0;
  76
  77                 if(all_swizzled && util_format_get_blocksize(color_format) != util_format_get_blocksize(zeta_format))
  78                         all_swizzled = 0;
  79         }
  80
  81         for (i = 0; i < fb->nr_cbufs; i++) {
  82                 if(!((struct nvfx_surface*)fb->cbufs[i])->temp && !all_swizzled && !nvfx_surface_linear_renderable(fb->cbufs[i]))
  83                         nvfx_surface_create_temp(&nvfx->pipe, fb->cbufs[i]);
  84         }
  85
  86         if(fb->zsbuf) {
  87                 if(!((struct nvfx_surface*)fb->zsbuf)->temp && !all_swizzled && !nvfx_surface_linear_renderable(fb->zsbuf))
  88                         nvfx_surface_create_temp(&nvfx->pipe, fb->zsbuf);
  89         }
  90
  91         return all_swizzled;
  92 }
  93
  94 void
  95 nvfx_framebuffer_validate(struct nvfx_context *nvfx, unsigned prepare_result)
  96 {
  97         struct pipe_framebuffer_state *fb = &nvfx->framebuffer;
  98         struct nouveau_channel *chan = nvfx->screen->base.channel;
  99         uint32_t rt_enable, rt_format;
 100         int i;
 101         unsigned rt_flags = NOUVEAU_BO_RDWR | NOUVEAU_BO_VRAM;
 102         unsigned w = fb->width;
 103         unsigned h = fb->height;
 104
 105         rt_enable = (NV30_3D_RT_ENABLE_COLOR0 << fb->nr_cbufs) - 1;
 106         if (rt_enable & (NV30_3D_RT_ENABLE_COLOR1 |
 107                          NV40_3D_RT_ENABLE_COLOR2 | NV40_3D_RT_ENABLE_COLOR3))
 108                 rt_enable |= NV30_3D_RT_ENABLE_MRT;
 109
 110         nvfx->state.render_temps = 0;
 111
 112         for (i = 0; i < fb->nr_cbufs; i++)
 113                 nvfx->state.render_temps |= nvfx_surface_get_render_target(fb->cbufs[i], prepare_result, &nvfx->hw_rt[i]) << i;
 114
 115         for(; i < 4; ++i)
 116                 nvfx->hw_rt[i].bo = 0;
 117
 118         if (fb->zsbuf) {
 119                 nvfx->state.render_temps |= nvfx_surface_get_render_target(fb->zsbuf, prepare_result, &nvfx->hw_zeta) << 7;
 120
 121                 assert(util_format_get_stride(fb->zsbuf->format, fb->width) <= nvfx->hw_zeta.pitch);
 122                 assert(nvfx->hw_zeta.offset + nvfx->hw_zeta.pitch * fb->height <= nvfx->hw_zeta.bo->size);
 123         }
 124
 125         if (prepare_result) {
 126                 assert(!(fb->width & (fb->width - 1)) && !(fb->height & (fb->height - 1)));
 127
 128                 rt_format = NV30_3D_RT_FORMAT_TYPE_SWIZZLED |
 129                         (util_logbase2(fb->width) << NV30_3D_RT_FORMAT_LOG2_WIDTH__SHIFT) |
 130                         (util_logbase2(fb->height) << NV30_3D_RT_FORMAT_LOG2_HEIGHT__SHIFT);
 131         } else
 132                 rt_format = NV30_3D_RT_FORMAT_TYPE_LINEAR;
 133
 134         if(fb->nr_cbufs > 0) {
 135                 switch (fb->cbufs[0]->format) {
 136                 case PIPE_FORMAT_B8G8R8X8_UNORM:
 137                         rt_format |= NV30_3D_RT_FORMAT_COLOR_X8R8G8B8;
 138                         break;
 139                 case PIPE_FORMAT_B8G8R8A8_UNORM:
 140                 case 0:
 141                         rt_format |= NV30_3D_RT_FORMAT_COLOR_A8R8G8B8;
 142                         break;
 143                 case PIPE_FORMAT_R8G8B8X8_UNORM:
 144                         rt_format |= NV30_3D_RT_FORMAT_COLOR_X8B8G8R8;
 145                         break;
 146                 case PIPE_FORMAT_R8G8B8A8_UNORM:
 147                         rt_format |= NV30_3D_RT_FORMAT_COLOR_A8B8G8R8;
 148                         break;
 149                 case PIPE_FORMAT_B5G6R5_UNORM:
 150                         rt_format |= NV30_3D_RT_FORMAT_COLOR_R5G6B5;
 151                         break;
 152                 case PIPE_FORMAT_R32G32B32A32_FLOAT:
 153                         rt_format |= NV30_3D_RT_FORMAT_COLOR_A32B32G32R32_FLOAT;
 154                         break;
 155                 case PIPE_FORMAT_R16G16B16A16_FLOAT:
 156                         rt_format |= NV30_3D_RT_FORMAT_COLOR_A16B16G16R16_FLOAT;
 157                         break;
 158                 default:
 159                         assert(0);
 160                 }
 161         } else if(fb->zsbuf && util_format_get_blocksize(fb->zsbuf->format) == 2)
 162                 rt_format |= NV30_3D_RT_FORMAT_COLOR_R5G6B5;
 163         else
 164                 rt_format |= NV30_3D_RT_FORMAT_COLOR_A8R8G8B8;
 165
 166         if(fb->zsbuf) {
 167                 switch (fb->zsbuf->format) {
 168                 case PIPE_FORMAT_Z16_UNORM:
 169                         rt_format |= NV30_3D_RT_FORMAT_ZETA_Z16;
 170                         break;
 171                 case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
 172                 case PIPE_FORMAT_X8Z24_UNORM:
 173                 case 0:
 174                         rt_format |= NV30_3D_RT_FORMAT_ZETA_Z24S8;
 175                         break;
 176                 default:
 177                         assert(0);
 178                 }
 179         } else if(fb->nr_cbufs && util_format_get_blocksize(fb->cbufs[0]->format) == 2)
 180                 rt_format |= NV30_3D_RT_FORMAT_ZETA_Z16;
 181         else
 182                 rt_format |= NV30_3D_RT_FORMAT_ZETA_Z24S8;
 183
 184         MARK_RING(chan, 42, 10);
 185
 186         if ((rt_enable & NV30_3D_RT_ENABLE_COLOR0) || fb->zsbuf) {
 187                 struct nvfx_render_target *rt0 = &nvfx->hw_rt[0];
 188                 uint32_t pitch;
 189
 190                 if(!(rt_enable & NV30_3D_RT_ENABLE_COLOR0))
 191                         rt0 = &nvfx->hw_zeta;
 192
 193                 pitch = rt0->pitch;
 194
 195                 if(!nvfx->is_nv4x)
 196                 {
 197                         if (nvfx->hw_zeta.bo)
 198                                 pitch |= (nvfx->hw_zeta.pitch << 16);
 199                         else
 200                                 pitch |= (pitch << 16);
 201                 }
 202
 203                 //printf("rendering to bo %p [%i] at offset %i with pitch %i\n", rt0->bo, rt0->bo->handle, rt0->offset, pitch);
 204
 205                 OUT_RING(chan, RING_3D(NV30_3D_DMA_COLOR0, 1));
 206                 OUT_RELOC(chan, rt0->bo, 0,
 207                               rt_flags | NOUVEAU_BO_OR,
 208                               chan->vram->handle, chan->gart->handle);
 209                 OUT_RING(chan, RING_3D(NV30_3D_COLOR0_PITCH, 2));
 210                 OUT_RING(chan, pitch);
 211                 OUT_RELOC(chan, rt0->bo,
 212                               rt0->offset, rt_flags | NOUVEAU_BO_LOW,
 213                               0, 0);
 214         }
 215
 216         if (rt_enable & NV30_3D_RT_ENABLE_COLOR1) {
 217                 OUT_RING(chan, RING_3D(NV30_3D_DMA_COLOR1, 1));
 218                 OUT_RELOC(chan, nvfx->hw_rt[1].bo, 0,
 219                               rt_flags | NOUVEAU_BO_OR,
 220                               chan->vram->handle, chan->gart->handle);
 221                 OUT_RING(chan, RING_3D(NV30_3D_COLOR1_OFFSET, 2));
 222                 OUT_RELOC(chan, nvfx->hw_rt[1].bo,
 223                                 nvfx->hw_rt[1].offset, rt_flags | NOUVEAU_BO_LOW,
 224                               0, 0);
 225                 OUT_RING(chan, nvfx->hw_rt[1].pitch);
 226         }
 227
 228         if(nvfx->is_nv4x)
 229         {
 230                 if (rt_enable & NV40_3D_RT_ENABLE_COLOR2) {
 231                         OUT_RING(chan, RING_3D(NV40_3D_DMA_COLOR2, 1));
 232                         OUT_RELOC(chan, nvfx->hw_rt[2].bo, 0,
 233                                       rt_flags | NOUVEAU_BO_OR,
 234                                       chan->vram->handle, chan->gart->handle);
 235                         OUT_RING(chan, RING_3D(NV40_3D_COLOR2_OFFSET, 1));
 236                         OUT_RELOC(chan, nvfx->hw_rt[2].bo,
 237                                       nvfx->hw_rt[2].offset, rt_flags | NOUVEAU_BO_LOW,
 238                                       0, 0);
 239                         OUT_RING(chan, RING_3D(NV40_3D_COLOR2_PITCH, 1));
 240                         OUT_RING(chan, nvfx->hw_rt[2].pitch);
 241                 }
 242
 243                 if (rt_enable & NV40_3D_RT_ENABLE_COLOR3) {
 244                         OUT_RING(chan, RING_3D(NV40_3D_DMA_COLOR3, 1));
 245                         OUT_RELOC(chan, nvfx->hw_rt[3].bo, 0,
 246                                       rt_flags | NOUVEAU_BO_OR,
 247                                       chan->vram->handle, chan->gart->handle);
 248                         OUT_RING(chan, RING_3D(NV40_3D_COLOR3_OFFSET, 1));
 249                         OUT_RELOC(chan, nvfx->hw_rt[3].bo,
 250                                         nvfx->hw_rt[3].offset, rt_flags | NOUVEAU_BO_LOW,
 251                                       0, 0);
 252                         OUT_RING(chan, RING_3D(NV40_3D_COLOR3_PITCH, 1));
 253                         OUT_RING(chan, nvfx->hw_rt[3].pitch);
 254                 }
 255         }
 256
 257         if (fb->zsbuf) {
 258                 OUT_RING(chan, RING_3D(NV30_3D_DMA_ZETA, 1));
 259                 OUT_RELOC(chan, nvfx->hw_zeta.bo, 0,
 260                               rt_flags | NOUVEAU_BO_OR,
 261                               chan->vram->handle, chan->gart->handle);
 262                 OUT_RING(chan, RING_3D(NV30_3D_ZETA_OFFSET, 1));
 263                 /* TODO: reverse engineer LMA */
 264                 OUT_RELOC(chan, nvfx->hw_zeta.bo,
 265                              nvfx->hw_zeta.offset, rt_flags | NOUVEAU_BO_LOW, 0, 0);
 266                 if(nvfx->is_nv4x) {
 267                         OUT_RING(chan, RING_3D(NV40_3D_ZETA_PITCH, 1));
 268                         OUT_RING(chan, nvfx->hw_zeta.pitch);
 269                 }
 270         }
 271         else if(nvfx->is_nv4x) {
 272                 OUT_RING(chan, RING_3D(NV40_3D_ZETA_PITCH, 1));
 273                 OUT_RING(chan, 64);
 274         }
 275
 276         OUT_RING(chan, RING_3D(NV30_3D_RT_ENABLE, 1));
 277         OUT_RING(chan, rt_enable);
 278         OUT_RING(chan, RING_3D(NV30_3D_RT_HORIZ, 3));
 279         OUT_RING(chan, (w << 16) | 0);
 280         OUT_RING(chan, (h << 16) | 0);
 281         OUT_RING(chan, rt_format);
 282         OUT_RING(chan, RING_3D(NV30_3D_VIEWPORT_HORIZ, 2));
 283         OUT_RING(chan, (w << 16) | 0);
 284         OUT_RING(chan, (h << 16) | 0);
 285         OUT_RING(chan, RING_3D(NV30_3D_VIEWPORT_CLIP_HORIZ(0), 2));
 286         OUT_RING(chan, ((w - 1) << 16) | 0);
 287         OUT_RING(chan, ((h - 1) << 16) | 0);
 288
 289         if(!nvfx->is_nv4x) {
 290                 /* Wonder why this is needed, context should all be set to zero on init */
 291                 /* TODO: we can most likely remove this, after putting it in context init */
 292                 OUT_RING(chan, RING_3D(NV30_3D_VIEWPORT_TX_ORIGIN, 1));
 293                 OUT_RING(chan, 0);
 294         }
 295         nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAMEBUFFER;
 296 }
 297
 298 void
 299 nvfx_framebuffer_relocate(struct nvfx_context *nvfx)
 300 {
 301         struct nouveau_channel *chan = nvfx->screen->base.channel;
 302         unsigned rt_flags = NOUVEAU_BO_RDWR | NOUVEAU_BO_VRAM;
 303         rt_flags |= NOUVEAU_BO_DUMMY;
 304         MARK_RING(chan, 20, 20);
 305
 306 #define DO_(var, pfx, name) \
 307         if(var.bo) { \
 308                 OUT_RELOC(chan, var.bo, RING_3D(pfx##_3D_DMA_##name, 1), rt_flags, 0, 0); \
 309                 OUT_RELOC(chan, var.bo, 0, \
 310                         rt_flags | NOUVEAU_BO_OR, \
 311                         chan->vram->handle, chan->gart->handle); \
 312                 OUT_RELOC(chan, var.bo, RING_3D(pfx##_3D_##name##_OFFSET, 1), rt_flags, 0, 0); \
 313                 OUT_RELOC(chan, var.bo, \
 314                         var.offset, rt_flags | NOUVEAU_BO_LOW, \
 315                         0, 0); \
 316         }
 317
 318 #define DO(pfx, num) DO_(nvfx->hw_rt[num], pfx, COLOR##num)
 319         DO(NV30, 0);
 320         DO(NV30, 1);
 321         DO(NV40, 2);
 322         DO(NV40, 3);
 323
 324         DO_(nvfx->hw_zeta, NV30, ZETA);
 325         nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAMEBUFFER;
 326 }