src/gallium/drivers/nvfx/nvfx_draw.c

   1 #include "pipe/p_shader_tokens.h"
   2 #include "util/u_inlines.h"
   3 #include "tgsi/tgsi_ureg.h"
   4
   5 #include "util/u_pack_color.h"
   6
   7 #include "draw/draw_context.h"
   8 #include "draw/draw_vertex.h"
   9 #include "draw/draw_pipe.h"
  10
  11 #include "nvfx_context.h"
  12
  13 /* Simple, but crappy, swtnl path, hopefully we wont need to hit this very
  14  * often at all.  Uses "quadro style" vertex submission + a fixed vertex
  15  * layout to avoid the need to generate a vertex program or vtxfmt.
  16  */
  17
  18 struct nvfx_render_stage {
  19         struct draw_stage stage;
  20         struct nvfx_context *nvfx;
  21         unsigned prim;
  22 };
  23
  24 static INLINE struct nvfx_render_stage *
  25 nvfx_render_stage(struct draw_stage *stage)
  26 {
  27         return (struct nvfx_render_stage *)stage;
  28 }
  29
  30 static INLINE void
  31 nvfx_render_vertex(struct nvfx_context *nvfx, const struct vertex_header *v)
  32 {
  33         struct nvfx_screen *screen = nvfx->screen;
  34         struct nouveau_channel *chan = screen->base.channel;
  35         struct nouveau_grobj *eng3d = screen->eng3d;
  36         unsigned i;
  37
  38         for (i = 0; i < nvfx->swtnl.nr_attribs; i++) {
  39                 unsigned idx = nvfx->swtnl.draw[i];
  40                 unsigned hw = nvfx->swtnl.hw[i];
  41
  42                 switch (nvfx->swtnl.emit[i]) {
  43                 case EMIT_OMIT:
  44                         break;
  45                 case EMIT_1F:
  46                         BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_1F(hw), 1);
  47                         OUT_RING  (chan, fui(v->data[idx][0]));
  48                         break;
  49                 case EMIT_2F:
  50                         BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_2F_X(hw), 2);
  51                         OUT_RING  (chan, fui(v->data[idx][0]));
  52                         OUT_RING  (chan, fui(v->data[idx][1]));
  53                         break;
  54                 case EMIT_3F:
  55                         BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_3F_X(hw), 3);
  56                         OUT_RING  (chan, fui(v->data[idx][0]));
  57                         OUT_RING  (chan, fui(v->data[idx][1]));
  58                         OUT_RING  (chan, fui(v->data[idx][2]));
  59                         break;
  60                 case EMIT_4F:
  61                         BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_4F_X(hw), 4);
  62                         OUT_RING  (chan, fui(v->data[idx][0]));
  63                         OUT_RING  (chan, fui(v->data[idx][1]));
  64                         OUT_RING  (chan, fui(v->data[idx][2]));
  65                         OUT_RING  (chan, fui(v->data[idx][3]));
  66                         break;
  67                 case 0xff:
  68                         BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_4F_X(hw), 4);
  69                         OUT_RING  (chan, fui(v->data[idx][0] / v->data[idx][3]));
  70                         OUT_RING  (chan, fui(v->data[idx][1] / v->data[idx][3]));
  71                         OUT_RING  (chan, fui(v->data[idx][2] / v->data[idx][3]));
  72                         OUT_RING  (chan, fui(1.0f / v->data[idx][3]));
  73                         break;
  74                 case EMIT_4UB:
  75                         BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_4UB(hw), 1);
  76                         OUT_RING  (chan, pack_ub4(float_to_ubyte(v->data[idx][0]),
  77                                             float_to_ubyte(v->data[idx][1]),
  78                                             float_to_ubyte(v->data[idx][2]),
  79                                             float_to_ubyte(v->data[idx][3])));
  80                 case EMIT_4UB_BGRA:
  81                         BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_4UB(hw), 1);
  82                         OUT_RING  (chan, pack_ub4(float_to_ubyte(v->data[idx][2]),
  83                                             float_to_ubyte(v->data[idx][1]),
  84                                             float_to_ubyte(v->data[idx][0]),
  85                                             float_to_ubyte(v->data[idx][3])));
  86                         break;
  87                 default:
  88                         assert(0);
  89                         break;
  90                 }
  91         }
  92 }
  93
  94 static INLINE void
  95 nvfx_render_prim(struct draw_stage *stage, struct prim_header *prim,
  96                unsigned mode, unsigned count)
  97 {
  98         struct nvfx_render_stage *rs = nvfx_render_stage(stage);
  99         struct nvfx_context *nvfx = rs->nvfx;
 100
 101         struct nvfx_screen *screen = nvfx->screen;
 102         struct nouveau_channel *chan = screen->base.channel;
 103         struct nouveau_grobj *eng3d = screen->eng3d;
 104         unsigned i;
 105
 106         /* Ensure there's room for 4xfloat32 + potentially 3 begin/end */
 107         if (AVAIL_RING(chan) < ((count * 20) + 6)) {
 108                 if (rs->prim != NV34TCL_VERTEX_BEGIN_END_STOP) {
 109                         NOUVEAU_ERR("AIII, missed flush\n");
 110                         assert(0);
 111                 }
 112                 FIRE_RING(chan);
 113                 nvfx_state_emit(nvfx);
 114         }
 115
 116         /* Switch primitive modes if necessary */
 117         if (rs->prim != mode) {
 118                 if (rs->prim != NV34TCL_VERTEX_BEGIN_END_STOP) {
 119                         BEGIN_RING(chan, eng3d, NV34TCL_VERTEX_BEGIN_END, 1);
 120                         OUT_RING  (chan, NV34TCL_VERTEX_BEGIN_END_STOP);
 121                 }
 122
 123                 BEGIN_RING(chan, eng3d, NV34TCL_VERTEX_BEGIN_END, 1);
 124                 OUT_RING  (chan, mode);
 125                 rs->prim = mode;
 126         }
 127
 128         /* Emit vertex data */
 129         for (i = 0; i < count; i++)
 130                 nvfx_render_vertex(nvfx, prim->v[i]);
 131
 132         /* If it's likely we'll need to empty the push buffer soon, finish
 133          * off the primitive now.
 134          */
 135         if (AVAIL_RING(chan) < ((count * 20) + 6)) {
 136                 BEGIN_RING(chan, eng3d, NV34TCL_VERTEX_BEGIN_END, 1);
 137                 OUT_RING  (chan, NV34TCL_VERTEX_BEGIN_END_STOP);
 138                 rs->prim = NV34TCL_VERTEX_BEGIN_END_STOP;
 139         }
 140 }
 141
 142 static void
 143 nvfx_render_point(struct draw_stage *draw, struct prim_header *prim)
 144 {
 145         nvfx_render_prim(draw, prim, NV34TCL_VERTEX_BEGIN_END_POINTS, 1);
 146 }
 147
 148 static void
 149 nvfx_render_line(struct draw_stage *draw, struct prim_header *prim)
 150 {
 151         nvfx_render_prim(draw, prim, NV34TCL_VERTEX_BEGIN_END_LINES, 2);
 152 }
 153
 154 static void
 155 nvfx_render_tri(struct draw_stage *draw, struct prim_header *prim)
 156 {
 157         nvfx_render_prim(draw, prim, NV34TCL_VERTEX_BEGIN_END_TRIANGLES, 3);
 158 }
 159
 160 static void
 161 nvfx_render_flush(struct draw_stage *draw, unsigned flags)
 162 {
 163         struct nvfx_render_stage *rs = nvfx_render_stage(draw);
 164         struct nvfx_context *nvfx = rs->nvfx;
 165         struct nvfx_screen *screen = nvfx->screen;
 166         struct nouveau_channel *chan = screen->base.channel;
 167         struct nouveau_grobj *eng3d = screen->eng3d;
 168
 169         if (rs->prim != NV34TCL_VERTEX_BEGIN_END_STOP) {
 170                 BEGIN_RING(chan, eng3d, NV34TCL_VERTEX_BEGIN_END, 1);
 171                 OUT_RING  (chan, NV34TCL_VERTEX_BEGIN_END_STOP);
 172                 rs->prim = NV34TCL_VERTEX_BEGIN_END_STOP;
 173         }
 174 }
 175
 176 static void
 177 nvfx_render_reset_stipple_counter(struct draw_stage *draw)
 178 {
 179 }
 180
 181 static void
 182 nvfx_render_destroy(struct draw_stage *draw)
 183 {
 184         FREE(draw);
 185 }
 186
 187 static struct nvfx_vertex_program *
 188 nvfx_create_drawvp(struct nvfx_context *nvfx)
 189 {
 190         struct ureg_program *ureg;
 191         uint i;
 192
 193         ureg = ureg_create( TGSI_PROCESSOR_VERTEX );
 194         if (ureg == NULL)
 195                 return NULL;
 196
 197         ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_POSITION, 0), ureg_DECL_vs_input(ureg, 0));
 198         ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0), ureg_DECL_vs_input(ureg, 3));
 199         ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 1), ureg_DECL_vs_input(ureg, 4));
 200         ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_BCOLOR, 0), ureg_DECL_vs_input(ureg, 3));
 201         ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_BCOLOR, 1), ureg_DECL_vs_input(ureg, 4));
 202         ureg_MOV(ureg,
 203                    ureg_writemask(ureg_DECL_output(ureg, TGSI_SEMANTIC_FOG, 1), TGSI_WRITEMASK_X),
 204                    ureg_DECL_vs_input(ureg, 5));
 205         for (i = 0; i < 8; ++i)
 206                 ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_GENERIC, i), ureg_DECL_vs_input(ureg, 8 + i));
 207
 208         ureg_END( ureg );
 209
 210         return ureg_create_shader_and_destroy( ureg, &nvfx->pipe );
 211 }
 212
 213 struct draw_stage *
 214 nvfx_draw_render_stage(struct nvfx_context *nvfx)
 215 {
 216         struct nvfx_render_stage *render = CALLOC_STRUCT(nvfx_render_stage);
 217
 218         if (!nvfx->swtnl.vertprog)
 219                 nvfx->swtnl.vertprog = nvfx_create_drawvp(nvfx);
 220
 221         render->nvfx = nvfx;
 222         render->stage.draw = nvfx->draw;
 223         render->stage.point = nvfx_render_point;
 224         render->stage.line = nvfx_render_line;
 225         render->stage.tri = nvfx_render_tri;
 226         render->stage.flush = nvfx_render_flush;
 227         render->stage.reset_stipple_counter = nvfx_render_reset_stipple_counter;
 228         render->stage.destroy = nvfx_render_destroy;
 229
 230         return &render->stage;
 231 }
 232
 233 void
 234 nvfx_draw_elements_swtnl(struct pipe_context *pipe,
 235                          struct pipe_resource *idxbuf,
 236                          unsigned idxbuf_size, int idxbuf_bias,
 237                          unsigned mode, unsigned start, unsigned count)
 238 {
 239         struct nvfx_context *nvfx = nvfx_context(pipe);
 240         struct pipe_transfer *vb_transfer[PIPE_MAX_ATTRIBS];
 241         struct pipe_transfer *ib_transfer = NULL;
 242         struct pipe_transfer *cb_transfer = NULL;
 243         unsigned i;
 244         void *map;
 245
 246         if (!nvfx_state_validate_swtnl(nvfx))
 247                 return;
 248         nvfx_state_emit(nvfx);
 249
 250         for (i = 0; i < nvfx->vtxbuf_nr; i++) {
 251                 map = pipe_buffer_map(pipe, nvfx->vtxbuf[i].buffer,
 252                                       PIPE_TRANSFER_READ,
 253                                       &vb_transfer[i]);
 254                 draw_set_mapped_vertex_buffer(nvfx->draw, i, map);
 255         }
 256
 257         if (idxbuf) {
 258                 map = pipe_buffer_map(pipe, idxbuf,
 259                                       PIPE_TRANSFER_READ,
 260                                       &ib_transfer);
 261                 draw_set_mapped_element_buffer(nvfx->draw, idxbuf_size, idxbuf_bias, map);
 262         } else {
 263                 draw_set_mapped_element_buffer(nvfx->draw, 0, 0, NULL);
 264         }
 265
 266         if (nvfx->constbuf[PIPE_SHADER_VERTEX]) {
 267                 const unsigned nr = nvfx->constbuf_nr[PIPE_SHADER_VERTEX];
 268
 269                 map = pipe_buffer_map(pipe,
 270                                       nvfx->constbuf[PIPE_SHADER_VERTEX],
 271                                       PIPE_TRANSFER_READ,
 272                                       &cb_transfer);
 273                 draw_set_mapped_constant_buffer(nvfx->draw, PIPE_SHADER_VERTEX, 0,
 274                                                 map, nr);
 275         }
 276
 277         draw_arrays(nvfx->draw, mode, start, count);
 278
 279         for (i = 0; i < nvfx->vtxbuf_nr; i++)
 280                 pipe_buffer_unmap(pipe, nvfx->vtxbuf[i].buffer, vb_transfer[i]);
 281
 282         if (idxbuf)
 283                 pipe_buffer_unmap(pipe, idxbuf, ib_transfer);
 284
 285         if (nvfx->constbuf[PIPE_SHADER_VERTEX])
 286                 pipe_buffer_unmap(pipe, nvfx->constbuf[PIPE_SHADER_VERTEX],
 287                                   cb_transfer);
 288
 289         draw_flush(nvfx->draw);
 290         pipe->flush(pipe, 0, NULL);
 291 }
 292
 293 static INLINE void
 294 emit_attrib(struct nvfx_context *nvfx, unsigned hw, unsigned emit,
 295             unsigned semantic, unsigned index)
 296 {
 297         unsigned draw_out = draw_find_shader_output(nvfx->draw, semantic, index);
 298         unsigned a = nvfx->swtnl.nr_attribs++;
 299
 300         nvfx->swtnl.hw[a] = hw;
 301         nvfx->swtnl.emit[a] = emit;
 302         nvfx->swtnl.draw[a] = draw_out;
 303 }
 304
 305 void
 306 nvfx_vtxfmt_validate(struct nvfx_context *nvfx)
 307 {
 308         struct nvfx_fragment_program *fp = nvfx->fragprog;
 309         unsigned colour = 0, texcoords = 0, fog = 0, i;
 310
 311         /* Determine needed fragprog inputs */
 312         for (i = 0; i < fp->info.num_inputs; i++) {
 313                 switch (fp->info.input_semantic_name[i]) {
 314                 case TGSI_SEMANTIC_POSITION:
 315                         break;
 316                 case TGSI_SEMANTIC_COLOR:
 317                         colour |= (1 << fp->info.input_semantic_index[i]);
 318                         break;
 319                 case TGSI_SEMANTIC_GENERIC:
 320                         texcoords |= (1 << fp->info.input_semantic_index[i]);
 321                         break;
 322                 case TGSI_SEMANTIC_FOG:
 323                         fog = 1;
 324                         break;
 325                 default:
 326                         assert(0);
 327                 }
 328         }
 329
 330         nvfx->swtnl.nr_attribs = 0;
 331
 332         /* Map draw vtxprog output to hw attribute IDs */
 333         for (i = 0; i < 2; i++) {
 334                 if (!(colour & (1 << i)))
 335                         continue;
 336                 emit_attrib(nvfx, 3 + i, EMIT_4F, TGSI_SEMANTIC_COLOR, i);
 337         }
 338
 339         for (i = 0; i < 8; i++) {
 340                 if (!(texcoords & (1 << i)))
 341                         continue;
 342                 emit_attrib(nvfx, 8 + i, EMIT_4F, TGSI_SEMANTIC_GENERIC, i);
 343         }
 344
 345         if (fog) {
 346                 emit_attrib(nvfx, 5, EMIT_1F, TGSI_SEMANTIC_FOG, 0);
 347         }
 348
 349         emit_attrib(nvfx, 0, 0xff, TGSI_SEMANTIC_POSITION, 0);
 350 }