src/gallium/drivers/nvfx/nvfx_fragtex.c

   1 #include "nvfx_context.h"
   2 #include "nvfx_resource.h"
   3 #include "nvfx_tex.h"
   4
   5 static void *
   6 nvfx_sampler_state_create(struct pipe_context *pipe,
   7                           const struct pipe_sampler_state *cso)
   8 {
   9         struct nvfx_context *nvfx = nvfx_context(pipe);
  10         struct nvfx_sampler_state *ps;
  11
  12         ps = MALLOC(sizeof(struct nvfx_sampler_state));
  13
  14         /* on nv30, we use this as an internal flag */
  15         ps->fmt = cso->normalized_coords ? 0 : NV40_3D_TEX_FORMAT_RECT;
  16         ps->en = 0;
  17         ps->filt = nvfx_tex_filter(cso) | 0x2000; /*voodoo*/
  18         ps->wrap = (nvfx_tex_wrap_mode(cso->wrap_s) << NV30_3D_TEX_WRAP_S__SHIFT) |
  19                     (nvfx_tex_wrap_mode(cso->wrap_t) << NV30_3D_TEX_WRAP_T__SHIFT) |
  20                     (nvfx_tex_wrap_mode(cso->wrap_r) << NV30_3D_TEX_WRAP_R__SHIFT);
  21         ps->compare = FALSE;
  22
  23         if(cso->compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE)
  24         {
  25                 ps->wrap |= nvfx_tex_wrap_compare_mode(cso->compare_func);
  26                 ps->compare = TRUE;
  27         }
  28         ps->bcol = nvfx_tex_border_color(cso->border_color.f);
  29
  30         if(nvfx->is_nv4x)
  31                 nv40_sampler_state_init(pipe, ps, cso);
  32         else
  33                 nv30_sampler_state_init(pipe, ps, cso);
  34
  35         return (void *)ps;
  36 }
  37
  38 static void
  39 nvfx_sampler_state_delete(struct pipe_context *pipe, void *hwcso)
  40 {
  41         FREE(hwcso);
  42 }
  43
  44 static void
  45 nvfx_sampler_state_bind(struct pipe_context *pipe, unsigned nr, void **sampler)
  46 {
  47         struct nvfx_context *nvfx = nvfx_context(pipe);
  48         unsigned unit;
  49
  50         for (unit = 0; unit < nr; unit++) {
  51                 nvfx->tex_sampler[unit] = sampler[unit];
  52                 nvfx->dirty_samplers |= (1 << unit);
  53         }
  54
  55         for (unit = nr; unit < nvfx->nr_samplers; unit++) {
  56                 nvfx->tex_sampler[unit] = NULL;
  57                 nvfx->dirty_samplers |= (1 << unit);
  58         }
  59
  60         nvfx->nr_samplers = nr;
  61         nvfx->dirty |= NVFX_NEW_SAMPLER;
  62 }
  63
  64 static struct pipe_sampler_view *
  65 nvfx_create_sampler_view(struct pipe_context *pipe,
  66                          struct pipe_resource *pt,
  67                          const struct pipe_sampler_view *templ)
  68 {
  69         struct nvfx_context *nvfx = nvfx_context(pipe);
  70         struct nvfx_sampler_view *sv = CALLOC_STRUCT(nvfx_sampler_view);
  71         struct nvfx_texture_format *tf = &nvfx_texture_formats[templ->format];
  72         unsigned txf;
  73
  74         if (!sv)
  75                 return NULL;
  76
  77         sv->base = *templ;
  78         sv->base.reference.count = 1;
  79         sv->base.texture = NULL;
  80         pipe_resource_reference(&sv->base.texture, pt);
  81         sv->base.context = pipe;
  82
  83         txf = NV30_3D_TEX_FORMAT_NO_BORDER;
  84
  85         switch (pt->target) {
  86         case PIPE_TEXTURE_CUBE:
  87                 txf |= NV30_3D_TEX_FORMAT_CUBIC;
  88                 /* fall-through */
  89         case PIPE_TEXTURE_2D:
  90         case PIPE_TEXTURE_RECT:
  91                 txf |= NV30_3D_TEX_FORMAT_DIMS_2D;
  92                 break;
  93         case PIPE_TEXTURE_3D:
  94                 txf |= NV30_3D_TEX_FORMAT_DIMS_3D;
  95                 break;
  96         case PIPE_TEXTURE_1D:
  97                 txf |= NV30_3D_TEX_FORMAT_DIMS_1D;
  98                 break;
  99         default:
 100                 assert(0);
 101         }
 102         sv->u.init_fmt = txf;
 103
 104         sv->swizzle = 0
 105                         | (tf->src[sv->base.swizzle_r] << NV30_3D_TEX_SWIZZLE_S0_Z__SHIFT)
 106                         | (tf->src[sv->base.swizzle_g] << NV30_3D_TEX_SWIZZLE_S0_Y__SHIFT)
 107                         | (tf->src[sv->base.swizzle_b] << NV30_3D_TEX_SWIZZLE_S0_X__SHIFT)
 108                         | (tf->src[sv->base.swizzle_a] << NV30_3D_TEX_SWIZZLE_S0_W__SHIFT)
 109                         | (tf->comp[sv->base.swizzle_r] << NV30_3D_TEX_SWIZZLE_S1_Z__SHIFT)
 110                         | (tf->comp[sv->base.swizzle_g] << NV30_3D_TEX_SWIZZLE_S1_Y__SHIFT)
 111                         | (tf->comp[sv->base.swizzle_b] << NV30_3D_TEX_SWIZZLE_S1_X__SHIFT)
 112                         | (tf->comp[sv->base.swizzle_a] << NV30_3D_TEX_SWIZZLE_S1_W__SHIFT);
 113
 114         sv->filt = tf->sign;
 115         sv->wrap = tf->wrap;
 116         sv->wrap_mask = ~0;
 117
 118         if (pt->target == PIPE_TEXTURE_CUBE)
 119         {
 120                 sv->offset = 0;
 121                 sv->npot_size = (pt->width0 << NV30_3D_TEX_NPOT_SIZE_W__SHIFT) | pt->height0;
 122         }
 123         else
 124         {
 125                 sv->offset = nvfx_subresource_offset(pt, 0, sv->base.u.tex.first_level, 0);
 126                 sv->npot_size = (u_minify(pt->width0, sv->base.u.tex.first_level) << NV30_3D_TEX_NPOT_SIZE_W__SHIFT) | u_minify(pt->height0, sv->base.u.tex.first_level);
 127
 128                 /* apparently, we need to ignore the t coordinate for 1D textures to fix piglit tex1d-2dborder */
 129                 if(pt->target == PIPE_TEXTURE_1D)
 130                 {
 131                         sv->wrap_mask &=~ NV30_3D_TEX_WRAP_T__MASK;
 132                         sv->wrap |= NV30_3D_TEX_WRAP_T_REPEAT;
 133                 }
 134         }
 135
 136         if(nvfx->is_nv4x)
 137                 nv40_sampler_view_init(pipe, sv);
 138         else
 139                 nv30_sampler_view_init(pipe, sv);
 140
 141         return &sv->base;
 142 }
 143
 144 static void
 145 nvfx_sampler_view_destroy(struct pipe_context *pipe,
 146                           struct pipe_sampler_view *view)
 147 {
 148         pipe_resource_reference(&view->texture, NULL);
 149         FREE(view);
 150 }
 151
 152 static void
 153 nvfx_set_fragment_sampler_views(struct pipe_context *pipe,
 154                                 unsigned nr,
 155                                 struct pipe_sampler_view **views)
 156 {
 157         struct nvfx_context *nvfx = nvfx_context(pipe);
 158         unsigned unit;
 159
 160         for (unit = 0; unit < nr; unit++) {
 161                 pipe_sampler_view_reference(&nvfx->fragment_sampler_views[unit],
 162                                             views[unit]);
 163                 nvfx->dirty_samplers |= (1 << unit);
 164         }
 165
 166         for (unit = nr; unit < nvfx->nr_textures; unit++) {
 167                 pipe_sampler_view_reference(&nvfx->fragment_sampler_views[unit],
 168                                             NULL);
 169                 nvfx->dirty_samplers |= (1 << unit);
 170         }
 171
 172         nvfx->nr_textures = nr;
 173         nvfx->dirty |= NVFX_NEW_SAMPLER;
 174 }
 175
 176 void
 177 nvfx_fragtex_validate(struct nvfx_context *nvfx)
 178 {
 179         struct nouveau_channel* chan = nvfx->screen->base.channel;
 180         struct nouveau_grobj *eng3d = nvfx->screen->eng3d;
 181         unsigned samplers, unit;
 182
 183         samplers = nvfx->dirty_samplers;
 184         if(!samplers)
 185                 return;
 186
 187         while (samplers) {
 188                 unit = ffs(samplers) - 1;
 189                 samplers &= ~(1 << unit);
 190
 191                 if(nvfx->fragment_sampler_views[unit] && nvfx->tex_sampler[unit]) {
 192                         util_dirty_surfaces_use_for_sampling(&nvfx->pipe,
 193                                         &((struct nvfx_miptree*)nvfx->fragment_sampler_views[unit]->texture)->dirty_surfaces,
 194                                         nvfx_surface_flush);
 195
 196                         if(!nvfx->is_nv4x)
 197                                 nv30_fragtex_set(nvfx, unit);
 198                         else
 199                                 nv40_fragtex_set(nvfx, unit);
 200                 } else {
 201                         /* this is OK for nv40 too */
 202                         BEGIN_RING(chan, eng3d, NV30_3D_TEX_ENABLE(unit), 1);
 203                         OUT_RING(chan, 0);
 204                         nvfx->hw_samplers &= ~(1 << unit);
 205                 }
 206         }
 207         nvfx->dirty_samplers = 0;
 208         nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAGTEX;
 209 }
 210
 211 void
 212 nvfx_fragtex_relocate(struct nvfx_context *nvfx)
 213 {
 214         struct nouveau_channel* chan = nvfx->screen->base.channel;
 215         unsigned samplers, unit;
 216         unsigned tex_flags = NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD;
 217
 218         samplers = nvfx->hw_samplers;
 219         while (samplers) {
 220                 struct nvfx_miptree* mt;
 221                 struct nouveau_bo *bo;
 222
 223                 unit = ffs(samplers) - 1;
 224                 samplers &= ~(1 << unit);
 225
 226                 mt = (struct nvfx_miptree*)nvfx->fragment_sampler_views[unit]->texture;
 227                 bo = mt->base.bo;
 228
 229                 MARK_RING(chan, 3, 3);
 230                 OUT_RELOC(chan, bo, RING_3D(NV30_3D_TEX_OFFSET(unit), 2), tex_flags | NOUVEAU_BO_DUMMY, 0, 0);
 231                 OUT_RELOC(chan, bo, 0, tex_flags | NOUVEAU_BO_LOW | NOUVEAU_BO_DUMMY, 0, 0);
 232                 OUT_RELOC(chan, bo, nvfx->hw_txf[unit], tex_flags | NOUVEAU_BO_OR | NOUVEAU_BO_DUMMY,
 233                                 NV30_3D_TEX_FORMAT_DMA0, NV30_3D_TEX_FORMAT_DMA1);
 234         }
 235         nvfx->relocs_needed &=~ NVFX_RELOCATE_FRAGTEX;
 236 }
 237
 238 void
 239 nvfx_init_sampling_functions(struct nvfx_context *nvfx)
 240 {
 241         nvfx->pipe.create_sampler_state = nvfx_sampler_state_create;
 242         nvfx->pipe.bind_fragment_sampler_states = nvfx_sampler_state_bind;
 243         nvfx->pipe.delete_sampler_state = nvfx_sampler_state_delete;
 244         nvfx->pipe.set_fragment_sampler_views = nvfx_set_fragment_sampler_views;
 245         nvfx->pipe.create_sampler_view = nvfx_create_sampler_view;
 246         nvfx->pipe.sampler_view_destroy = nvfx_sampler_view_destroy;
 247 }
 248
 249 #define NV30_3D_TEX_FORMAT_FORMAT_DXT1_RECT NV30_3D_TEX_FORMAT_FORMAT_DXT1
 250 #define NV30_3D_TEX_FORMAT_FORMAT_DXT3_RECT NV30_3D_TEX_FORMAT_FORMAT_DXT3
 251 #define NV30_3D_TEX_FORMAT_FORMAT_DXT5_RECT NV30_3D_TEX_FORMAT_FORMAT_DXT5
 252
 253 #define NV40_3D_TEX_FORMAT_FORMAT_HILO16 NV40_3D_TEX_FORMAT_FORMAT_A16L16
 254
 255 #define NV30_3D_TEX_FORMAT_FORMAT_RGBA16F 0x00004a00
 256 #define NV30_3D_TEX_FORMAT_FORMAT_RGBA16F_RECT NV30_3D_TEX_FORMAT_FORMAT_RGBA16F
 257 #define NV30_3D_TEX_FORMAT_FORMAT_RGBA32F 0x00004b00
 258 #define NV30_3D_TEX_FORMAT_FORMAT_RGBA32F_RECT NV30_3D_TEX_FORMAT_FORMAT_RGBA32F
 259 #define NV30_3D_TEX_FORMAT_FORMAT_R32F 0x00004c00
 260 #define NV30_3D_TEX_FORMAT_FORMAT_R32F_RECT NV30_3D_TEX_FORMAT_FORMAT_R32F
 261
 262 // TODO: guess!
 263 #define NV40_3D_TEX_FORMAT_FORMAT_R32F 0x00001c00
 264
 265 #define SRGB 0x00700000
 266
 267 #define __(m,tf,tfc,ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w,sign,wrap) \
 268 [PIPE_FORMAT_##m] = { \
 269   {NV30_3D_TEX_FORMAT_FORMAT_##tf, \
 270   NV30_3D_TEX_FORMAT_FORMAT_##tfc, \
 271   NV30_3D_TEX_FORMAT_FORMAT_##tf##_RECT, \
 272   NV30_3D_TEX_FORMAT_FORMAT_##tfc##_RECT, \
 273   NV40_3D_TEX_FORMAT_FORMAT_##tf, \
 274   NV40_3D_TEX_FORMAT_FORMAT_##tfc}, \
 275   sign, wrap, \
 276   {ts0z, ts0y, ts0x, ts0w, 0, 1}, {ts1z, ts1y, ts1x, ts1w, 0, 0} \
 277 }
 278
 279 #define _(m,tf,ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w,sign, wrap) \
 280         __(m,tf,tf,ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w,sign, wrap)
 281
 282 /* Depth formats works by reading the depth value most significant 8/16 bits.
 283  * We are losing precision, but nVidia loses even more by using A8R8G8B8 instead of HILO16
 284  * There is no 32-bit integer texture support, so other things are infeasible.
 285  *
 286  * TODO: is it possible to read 16 bits for Z16? A16 doesn't seem to work, either due to normalization or endianness issues
 287  */
 288
 289 #define T 2
 290
 291 #define X 3
 292 #define Y 2
 293 #define Z 1
 294 #define W 0
 295
 296 #define SNORM ((NV30_3D_TEX_FILTER_SIGNED_RED) | (NV30_3D_TEX_FILTER_SIGNED_GREEN) | (NV30_3D_TEX_FILTER_SIGNED_BLUE) | (NV30_3D_TEX_FILTER_SIGNED_ALPHA))
 297 #define UNORM 0
 298
 299 struct nvfx_texture_format
 300 nvfx_texture_formats[PIPE_FORMAT_COUNT] = {
 301         [0 ... PIPE_FORMAT_COUNT - 1] = {{-1, -1, -1, -1, -1, -1}},
 302         _(B8G8R8X8_UNORM,       A8R8G8B8,       T, T, T, 1, X, Y, Z, W, UNORM, 0),
 303         _(B8G8R8X8_SRGB,        A8R8G8B8,       T, T, T, 1, X, Y, Z, W, UNORM, SRGB),
 304         _(B8G8R8A8_UNORM,       A8R8G8B8,       T, T, T, T, X, Y, Z, W, UNORM, 0),
 305         _(B8G8R8A8_SRGB,        A8R8G8B8,       T, T, T, T, X, Y, Z, W, UNORM, SRGB),
 306
 307         _(R8G8B8A8_UNORM,       A8R8G8B8,       T, T, T, T, Z, Y, X, W, UNORM, 0),
 308         _(R8G8B8A8_SRGB,        A8R8G8B8,       T, T, T, T, Z, Y, X, W, UNORM, SRGB),
 309         _(R8G8B8X8_UNORM,       A8R8G8B8,       T, T, T, 1, Z, Y, X, W, UNORM, 0),
 310
 311         _(A8R8G8B8_UNORM,       A8R8G8B8,       T, T, T, T, W, Z, Y, X, UNORM, 0),
 312         _(A8R8G8B8_SRGB,        A8R8G8B8,       T, T, T, T, W, Z, Y, X, UNORM, SRGB),
 313         _(A8B8G8R8_UNORM,       A8R8G8B8,       T, T, T, T, W, X, Y, Z, UNORM, 0),
 314         _(A8B8G8R8_SRGB,        A8R8G8B8,       T, T, T, T, W, X, Y, Z, UNORM, SRGB),
 315         _(X8R8G8B8_UNORM,       A8R8G8B8,       T, T, T, 1, W, Z, Y, X, UNORM, 0),
 316         _(X8R8G8B8_SRGB,        A8R8G8B8,       T, T, T, 1, W, Z, Y, X, UNORM, SRGB),
 317
 318         _(B5G5R5A1_UNORM,       A1R5G5B5,       T, T, T, T, X, Y, Z, W, UNORM, 0),
 319         _(B5G5R5X1_UNORM,       A1R5G5B5,       T, T, T, 1, X, Y, Z, W, UNORM, 0),
 320
 321         _(B4G4R4A4_UNORM,       A4R4G4B4,       T, T, T, T, X, Y, Z, W, UNORM, 0),
 322         _(B4G4R4X4_UNORM,       A4R4G4B4,       T, T, T, 1, X, Y, Z, W, UNORM, 0),
 323
 324         _(B5G6R5_UNORM,         R5G6B5,         T, T, T, 1, X, Y, Z, W, UNORM, 0),
 325
 326         _(R8_UNORM,             L8,             T, 0, 0, 1, X, X, X, X, UNORM, 0),
 327         _(R8_SNORM,             L8,             T, 0, 0, 1, X, X, X, X, SNORM, 0),
 328         _(L8_UNORM,             L8,             T, T, T, 1, X, X, X, X, UNORM, 0),
 329         _(L8_SRGB,              L8,             T, T, T, 1, X, X, X, X, UNORM, SRGB),
 330         _(A8_UNORM,             L8,             0, 0, 0, T, X, X, X, X, UNORM, 0),
 331         _(I8_UNORM,             L8,             T, T, T, T, X, X, X, X, UNORM, 0),
 332
 333         _(R8G8_UNORM,           A8L8,           T, T, T, T, X, X, X, W, UNORM, 0),
 334         _(R8G8_SNORM,           A8L8,           T, T, T, T, X, X, X, W, SNORM, 0),
 335         _(L8A8_UNORM,           A8L8,           T, T, T, T, X, X, X, W, UNORM, 0),
 336         _(L8A8_SRGB,            A8L8,           T, T, T, T, X, X, X, W, UNORM, SRGB),
 337
 338         _(DXT1_RGB,             DXT1,           T, T, T, 1, X, Y, Z, W, UNORM, 0),
 339         _(DXT1_SRGB,            DXT1,           T, T, T, 1, X, Y, Z, W, UNORM, SRGB),
 340         _(DXT1_RGBA,            DXT1,           T, T, T, T, X, Y, Z, W, UNORM, 0),
 341         _(DXT1_SRGBA,           DXT1,           T, T, T, T, X, Y, Z, W, UNORM, SRGB),
 342         _(DXT3_RGBA,            DXT3,           T, T, T, T, X, Y, Z, W, UNORM, 0),
 343         _(DXT3_SRGBA,           DXT3,           T, T, T, T, X, Y, Z, W, UNORM, SRGB),
 344         _(DXT5_RGBA,            DXT5,           T, T, T, T, X, Y, Z, W, UNORM, 0),
 345         _(DXT5_SRGBA,           DXT5,           T, T, T, T, X, Y, Z, W, UNORM, SRGB),
 346
 347         __(Z16_UNORM,           A8L8, Z16,      T, T, T, 1, W, W, W, W, UNORM, 0),
 348         __(S8_USCALED_Z24_UNORM,HILO16,Z24,     T, T, T, 1, W, W, W, W, UNORM, 0),
 349         __(X8Z24_UNORM,         HILO16,Z24,     T, T, T, 1, W, W, W, W, UNORM, 0),
 350
 351         _(R16_UNORM,            A16,            T, 0, 0, 1, X, X, X, X, UNORM, 0),
 352         _(R16_SNORM,            A16,            T, 0, 0, 1, X, X, X, X, SNORM, 0),
 353         _(R16G16_UNORM,         HILO16,         T, T, 0, 1, X, Y, X, X, UNORM, 0),
 354         _(R16G16_SNORM,         HILO16,         T, T, 0, 1, X, Y, X, X, SNORM, 0),
 355
 356         _(R16G16B16A16_FLOAT,           RGBA16F,        T, T, T, T, X, Y, Z, W, UNORM, 0),
 357         _(R32G32B32A32_FLOAT,           RGBA32F,        T, T, T, T, X, Y, Z, W, UNORM, 0),
 358         _(R32_FLOAT,            R32F,   T, 0, 0, 1, X, X, X, X, UNORM, 0)
 359 };