src/gallium/drivers/nvfx/nvfx_state_emit.c

   1 #include "nvfx_context.h"
   2 #include "nvfx_state.h"
   3 #include "nvfx_resource.h"
   4 #include "draw/draw_context.h"
   5
   6 static boolean
   7 nvfx_state_validate_common(struct nvfx_context *nvfx)
   8 {
   9         struct nouveau_channel* chan = nvfx->screen->base.channel;
  10         unsigned dirty;
  11         unsigned still_dirty = 0;
  12         int all_swizzled = -1;
  13         boolean flush_tex_cache = FALSE;
  14
  15         if(nvfx != nvfx->screen->cur_ctx)
  16         {
  17                 nvfx->dirty = ~0;
  18                 nvfx->hw_vtxelt_nr = 16;
  19                 nvfx->screen->cur_ctx = nvfx;
  20         }
  21
  22         /* These can trigger use the of 3D engine to copy temporaries.
  23          * That will recurse here and thus dirty all 3D state, so we need to this before anything else, and in a loop..
  24          * This converges to having clean temps, then binding both fragtexes and framebuffers.
  25          */
  26         while(nvfx->dirty & (NVFX_NEW_FB | NVFX_NEW_SAMPLER))
  27         {
  28                 if(nvfx->dirty & NVFX_NEW_SAMPLER)
  29                 {
  30                         nvfx->dirty &=~ NVFX_NEW_SAMPLER;
  31                         nvfx_fragtex_validate(nvfx);
  32
  33                         // TODO: only set this if really necessary
  34                         flush_tex_cache = TRUE;
  35                 }
  36
  37                 if(nvfx->dirty & NVFX_NEW_FB)
  38                 {
  39                         nvfx->dirty &=~ NVFX_NEW_FB;
  40                         all_swizzled = nvfx_framebuffer_prepare(nvfx);
  41
  42                         // TODO: make sure this doesn't happen, i.e. fbs have matching formats
  43                         assert(all_swizzled >= 0);
  44                 }
  45         }
  46
  47         dirty = nvfx->dirty;
  48
  49         if(nvfx->render_mode == HW)
  50         {
  51                 if(dirty & (NVFX_NEW_VERTPROG | NVFX_NEW_VERTCONST | NVFX_NEW_UCP))
  52                 {
  53                         if(!nvfx_vertprog_validate(nvfx))
  54                                 return FALSE;
  55                 }
  56
  57                 if(dirty & NVFX_NEW_ARRAYS)
  58                 {
  59                         if(!nvfx_vbo_validate(nvfx))
  60                                 return FALSE;
  61                 }
  62
  63                 if(dirty & NVFX_NEW_INDEX)
  64                 {
  65                         if(nvfx->use_index_buffer)
  66                                 nvfx_idxbuf_validate(nvfx);
  67                         else
  68                                 still_dirty = NVFX_NEW_INDEX;
  69                 }
  70         }
  71         else
  72         {
  73                 /* TODO: this looks a bit misdesigned */
  74                 if(dirty & (NVFX_NEW_VERTPROG | NVFX_NEW_UCP))
  75                         nvfx_vertprog_validate(nvfx);
  76
  77                 if(dirty & (NVFX_NEW_ARRAYS | NVFX_NEW_INDEX | NVFX_NEW_FRAGPROG))
  78                         nvfx_vtxfmt_validate(nvfx);
  79         }
  80
  81         if(dirty & NVFX_NEW_RAST)
  82                 sb_emit(chan, nvfx->rasterizer->sb, nvfx->rasterizer->sb_len);
  83
  84         if(dirty & NVFX_NEW_SCISSOR)
  85                 nvfx_state_scissor_validate(nvfx);
  86
  87         if(dirty & NVFX_NEW_STIPPLE)
  88                 nvfx_state_stipple_validate(nvfx);
  89
  90         if(dirty & (NVFX_NEW_FRAGPROG | NVFX_NEW_FRAGCONST | NVFX_NEW_VERTPROG | NVFX_NEW_SPRITE))
  91         {
  92                 nvfx_fragprog_validate(nvfx);
  93                 if(dirty & NVFX_NEW_FRAGPROG)
  94                         flush_tex_cache = TRUE; // TODO: do we need this?
  95         }
  96
  97         if(all_swizzled >= 0)
  98                 nvfx_framebuffer_validate(nvfx, all_swizzled);
  99
 100         if(dirty & NVFX_NEW_BLEND)
 101                 sb_emit(chan, nvfx->blend->sb, nvfx->blend->sb_len);
 102
 103         if(dirty & NVFX_NEW_BCOL)
 104                 nvfx_state_blend_colour_validate(nvfx);
 105
 106         if(dirty & NVFX_NEW_ZSA)
 107                 sb_emit(chan, nvfx->zsa->sb, nvfx->zsa->sb_len);
 108
 109         if(dirty & NVFX_NEW_SR)
 110                 nvfx_state_sr_validate(nvfx);
 111
 112 /* All these dependencies are wrong, but otherwise
 113    etracer, neverball, foobillard, glest totally misrender
 114    TODO: find the right fix
 115 */
 116         if(dirty & (NVFX_NEW_VIEWPORT | NVFX_NEW_RAST | NVFX_NEW_ZSA) || (all_swizzled >= 0))
 117         {
 118                 nvfx_state_viewport_validate(nvfx);
 119         }
 120
 121         if(dirty & NVFX_NEW_ZSA || (all_swizzled >= 0))
 122         {
 123                 WAIT_RING(chan, 3);
 124                 OUT_RING(chan, RING_3D(NV34TCL_DEPTH_WRITE_ENABLE, 2));
 125                 OUT_RING(chan, nvfx->framebuffer.zsbuf && nvfx->zsa->pipe.depth.writemask);
 126                 OUT_RING(chan, nvfx->framebuffer.zsbuf && nvfx->zsa->pipe.depth.enabled);
 127         }
 128
 129         if(flush_tex_cache)
 130         {
 131                 // TODO: what about nv30?
 132                 if(nvfx->is_nv4x)
 133                 {
 134                         WAIT_RING(chan, 4);
 135                         OUT_RING(chan, RING_3D(NV40TCL_TEX_CACHE_CTL, 1));
 136                         OUT_RING(chan, 2);
 137                         OUT_RING(chan, RING_3D(NV40TCL_TEX_CACHE_CTL, 1));
 138                         OUT_RING(chan, 1);
 139                 }
 140         }
 141
 142         nvfx->dirty = dirty & still_dirty;
 143
 144         unsigned render_temps = nvfx->state.render_temps;
 145         if(render_temps)
 146         {
 147                 for(int i = 0; i < nvfx->framebuffer.nr_cbufs; ++i)
 148                 {
 149                         if(render_temps & (1 << i))
 150                                 util_dirty_surface_set_dirty(nvfx_surface_get_dirty_surfaces(nvfx->framebuffer.cbufs[i]),
 151                                                 (struct util_dirty_surface*)nvfx->framebuffer.cbufs[i]);
 152                 }
 153
 154                 if(render_temps & 0x80)
 155                         util_dirty_surface_set_dirty(nvfx_surface_get_dirty_surfaces(nvfx->framebuffer.zsbuf),
 156                                         (struct util_dirty_surface*)nvfx->framebuffer.zsbuf);
 157         }
 158
 159         return TRUE;
 160 }
 161
 162 void
 163 nvfx_state_emit(struct nvfx_context *nvfx)
 164 {
 165         struct nouveau_channel* chan = nvfx->screen->base.channel;
 166         /* we need to ensure there is enough space to output relocations in one go */
 167         unsigned max_relocs = 0
 168               + 16 /* vertex buffers, incl. dma flag */
 169               + 2 /* index buffer plus format+dma flag */
 170               + 2 * 5 /* 4 cbufs + zsbuf, plus dma objects */
 171               + 2 * 16 /* fragment textures plus format+dma flag */
 172               + 2 * 4 /* vertex textures plus format+dma flag */
 173               + 1 /* fragprog incl dma flag */
 174               ;
 175         MARK_RING(chan, max_relocs * 2, max_relocs * 2);
 176         nvfx_state_relocate(nvfx);
 177 }
 178
 179 void
 180 nvfx_state_relocate(struct nvfx_context *nvfx)
 181 {
 182         nvfx_framebuffer_relocate(nvfx);
 183         nvfx_fragtex_relocate(nvfx);
 184         nvfx_fragprog_relocate(nvfx);
 185         if (nvfx->render_mode == HW)
 186         {
 187                 nvfx_vbo_relocate(nvfx);
 188                 if(nvfx->use_index_buffer)
 189                         nvfx_idxbuf_relocate(nvfx);
 190         }
 191 }
 192
 193 boolean
 194 nvfx_state_validate(struct nvfx_context *nvfx)
 195 {
 196         boolean was_sw = nvfx->fallback_swtnl ? TRUE : FALSE;
 197
 198         if (nvfx->render_mode != HW) {
 199                 /* Don't even bother trying to go back to hw if none
 200                  * of the states that caused swtnl previously have changed.
 201                  */
 202                 if ((nvfx->fallback_swtnl & nvfx->dirty)
 203                                 != nvfx->fallback_swtnl)
 204                         return FALSE;
 205
 206                 /* Attempt to go to hwtnl again */
 207                 nvfx->dirty |= (NVFX_NEW_VIEWPORT |
 208                                 NVFX_NEW_VERTPROG |
 209                                 NVFX_NEW_ARRAYS);
 210                 nvfx->render_mode = HW;
 211         }
 212
 213         if(!nvfx_state_validate_common(nvfx))
 214                 return FALSE;
 215
 216         if (was_sw)
 217                 NOUVEAU_ERR("swtnl->hw\n");
 218
 219         return TRUE;
 220 }
 221
 222 boolean
 223 nvfx_state_validate_swtnl(struct nvfx_context *nvfx)
 224 {
 225         struct draw_context *draw = nvfx->draw;
 226
 227         /* Setup for swtnl */
 228         if (nvfx->render_mode == HW) {
 229                 NOUVEAU_ERR("hw->swtnl 0x%08x\n", nvfx->fallback_swtnl);
 230                 nvfx->pipe.flush(&nvfx->pipe, 0, NULL);
 231                 nvfx->dirty |= (NVFX_NEW_VIEWPORT |
 232                                 NVFX_NEW_VERTPROG |
 233                                 NVFX_NEW_ARRAYS);
 234                 nvfx->render_mode = SWTNL;
 235         }
 236
 237         if (nvfx->draw_dirty & NVFX_NEW_VERTPROG)
 238                 draw_bind_vertex_shader(draw, nvfx->vertprog->draw);
 239
 240         if (nvfx->draw_dirty & NVFX_NEW_RAST)
 241            draw_set_rasterizer_state(draw, &nvfx->rasterizer->pipe,
 242                                      nvfx->rasterizer);
 243
 244         if (nvfx->draw_dirty & NVFX_NEW_UCP)
 245                 draw_set_clip_state(draw, &nvfx->clip);
 246
 247         if (nvfx->draw_dirty & NVFX_NEW_VIEWPORT)
 248                 draw_set_viewport_state(draw, &nvfx->viewport);
 249
 250         if (nvfx->draw_dirty & NVFX_NEW_ARRAYS) {
 251                 draw_set_vertex_buffers(draw, nvfx->vtxbuf_nr, nvfx->vtxbuf);
 252                 draw_set_vertex_elements(draw, nvfx->vtxelt->num_elements, nvfx->vtxelt->pipe);
 253         }
 254
 255         nvfx_state_validate_common(nvfx);
 256
 257         nvfx->draw_dirty = 0;
 258         return TRUE;
 259 }