src/gallium/drivers/panfrost/pan_mfbd.c

   1 /*
   2  * Copyright 2018-2019 Alyssa Rosenzweig
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  */
  24
  25 #include "pan_context.h"
  26 #include "pan_util.h"
  27 #include "pan_format.h"
  28
  29 #include "util/u_format.h"
  30
  31 static void
  32 panfrost_invert_swizzle(const unsigned char *in, unsigned char *out)
  33 {
  34         /* First, default to all zeroes to prevent uninitialized junk */
  35
  36         for (unsigned c = 0; c < 4; ++c)
  37                 out[c] = PIPE_SWIZZLE_0;
  38
  39         /* Now "do" what the swizzle says */
  40
  41         for (unsigned c = 0; c < 4; ++c) {
  42                 unsigned char i = in[c];
  43
  44                 /* Who cares? */
  45                 assert(PIPE_SWIZZLE_X == 0);
  46                 if (i > PIPE_SWIZZLE_W)
  47                         continue;
  48
  49                 /* Invert */
  50                 unsigned idx = i - PIPE_SWIZZLE_X;
  51                 out[idx] = PIPE_SWIZZLE_X + c;
  52         }
  53 }
  54
  55 static struct mali_rt_format
  56 panfrost_mfbd_format(struct pipe_surface *surf)
  57 {
  58         /* Explode details on the format */
  59
  60         const struct util_format_description *desc =
  61                 util_format_description(surf->format);
  62
  63         /* The swizzle for rendering is inverted from texturing */
  64
  65         unsigned char swizzle[4];
  66         panfrost_invert_swizzle(desc->swizzle, swizzle);
  67
  68         /* Fill in accordingly, defaulting to 8-bit UNORM */
  69
  70         struct mali_rt_format fmt = {
  71                 .unk1 = 0x4000000,
  72                 .unk2 = 0x1,
  73                 .nr_channels = MALI_POSITIVE(desc->nr_channels),
  74                 .unk3 = 0x4,
  75                 .flags = 0x8,
  76                 .swizzle = panfrost_translate_swizzle_4(swizzle),
  77                 .no_preload = true
  78         };
  79
  80         if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB)
  81                 fmt.flags |= MALI_MFBD_FORMAT_SRGB;
  82
  83         /* sRGB handled as a dedicated flag */
  84         enum pipe_format linearized = util_format_linear(surf->format);
  85
  86         /* If RGB, we're good to go */
  87         if (util_format_is_unorm8(desc))
  88                 return fmt;
  89
  90         /* Set flags for alternative formats */
  91
  92         switch (linearized) {
  93         case PIPE_FORMAT_B5G6R5_UNORM:
  94                 fmt.unk1 = 0x14000000;
  95                 fmt.nr_channels = MALI_POSITIVE(2);
  96                 fmt.unk3 |= 0x1;
  97                 break;
  98
  99         case PIPE_FORMAT_A4B4G4R4_UNORM:
 100         case PIPE_FORMAT_B4G4R4A4_UNORM:
 101                 fmt.unk1 = 0x10000000;
 102                 fmt.unk3 = 0x5;
 103                 fmt.nr_channels = MALI_POSITIVE(1);
 104                 break;
 105
 106         case PIPE_FORMAT_R10G10B10A2_UNORM:
 107         case PIPE_FORMAT_B10G10R10A2_UNORM:
 108         case PIPE_FORMAT_R10G10B10X2_UNORM:
 109         case PIPE_FORMAT_B10G10R10X2_UNORM:
 110                 fmt.unk1 = 0x08000000;
 111                 fmt.unk3 = 0x6;
 112                 fmt.nr_channels = MALI_POSITIVE(1);
 113                 break;
 114
 115         /* Generic 8-bit */
 116         case PIPE_FORMAT_R8_UINT:
 117         case PIPE_FORMAT_R8_SINT:
 118                 fmt.unk1 = 0x80000000;
 119                 fmt.unk3 = 0x0;
 120                 fmt.nr_channels = MALI_POSITIVE(1);
 121                 break;
 122
 123         /* Generic 32-bit */
 124         case PIPE_FORMAT_R11G11B10_FLOAT:
 125         case PIPE_FORMAT_R8G8B8A8_UINT:
 126         case PIPE_FORMAT_R8G8B8A8_SINT:
 127         case PIPE_FORMAT_R16G16_FLOAT:
 128         case PIPE_FORMAT_R16G16_UINT:
 129         case PIPE_FORMAT_R16G16_SINT:
 130         case PIPE_FORMAT_R32_FLOAT:
 131         case PIPE_FORMAT_R32_UINT:
 132         case PIPE_FORMAT_R32_SINT:
 133         case PIPE_FORMAT_R10G10B10A2_UINT:
 134                 fmt.unk1 = 0x88000000;
 135                 fmt.unk3 = 0x0;
 136                 fmt.nr_channels = MALI_POSITIVE(4);
 137                 break;
 138
 139         /* Generic 16-bit */
 140         case PIPE_FORMAT_R8G8_UINT:
 141         case PIPE_FORMAT_R8G8_SINT:
 142         case PIPE_FORMAT_R16_FLOAT:
 143         case PIPE_FORMAT_R16_UINT:
 144         case PIPE_FORMAT_R16_SINT:
 145         case PIPE_FORMAT_B5G5R5A1_UNORM:
 146                 fmt.unk1 = 0x84000000;
 147                 fmt.unk3 = 0x0;
 148                 fmt.nr_channels = MALI_POSITIVE(2);
 149                 break;
 150
 151         /* Generic 64-bit */
 152         case PIPE_FORMAT_R32G32_FLOAT:
 153         case PIPE_FORMAT_R32G32_SINT:
 154         case PIPE_FORMAT_R32G32_UINT:
 155         case PIPE_FORMAT_R16G16B16A16_FLOAT:
 156         case PIPE_FORMAT_R16G16B16A16_SINT:
 157         case PIPE_FORMAT_R16G16B16A16_UINT:
 158                 fmt.unk1 = 0x8c000000;
 159                 fmt.unk3 = 0x1;
 160                 fmt.nr_channels = MALI_POSITIVE(2);
 161                 break;
 162
 163         /* Generic 128-bit */
 164         case PIPE_FORMAT_R32G32B32A32_FLOAT:
 165         case PIPE_FORMAT_R32G32B32A32_SINT:
 166         case PIPE_FORMAT_R32G32B32A32_UINT:
 167                 fmt.unk1 = 0x90000000;
 168                 fmt.unk3 = 0x1;
 169                 fmt.nr_channels = MALI_POSITIVE(4);
 170                 break;
 171
 172         default:
 173                 unreachable("Invalid format rendering");
 174         }
 175
 176         return fmt;
 177 }
 178
 179
 180 static void
 181 panfrost_mfbd_clear(
 182         struct panfrost_job *job,
 183         struct bifrost_framebuffer *fb,
 184         struct bifrost_fb_extra *fbx,
 185         struct bifrost_render_target *rts,
 186         unsigned rt_count)
 187 {
 188         for (unsigned i = 0; i < rt_count; ++i) {
 189                 if (!(job->clear & (PIPE_CLEAR_COLOR0 << i)))
 190                         continue;
 191
 192                 rts[i].clear_color_1 = job->clear_color[i][0];
 193                 rts[i].clear_color_2 = job->clear_color[i][1];
 194                 rts[i].clear_color_3 = job->clear_color[i][2];
 195                 rts[i].clear_color_4 = job->clear_color[i][3];
 196         }
 197
 198         if (job->clear & PIPE_CLEAR_DEPTH) {
 199                 fb->clear_depth = job->clear_depth;
 200         }
 201
 202         if (job->clear & PIPE_CLEAR_STENCIL) {
 203                 fb->clear_stencil = job->clear_stencil;
 204         }
 205 }
 206
 207 static void
 208 panfrost_mfbd_set_cbuf(
 209         struct bifrost_render_target *rt,
 210         struct pipe_surface *surf)
 211 {
 212         struct panfrost_resource *rsrc = pan_resource(surf->texture);
 213
 214         unsigned level = surf->u.tex.level;
 215         unsigned first_layer = surf->u.tex.first_layer;
 216         assert(surf->u.tex.last_layer == first_layer);
 217         int stride = rsrc->slices[level].stride;
 218
 219         mali_ptr base = panfrost_get_texture_address(rsrc, level, first_layer);
 220
 221         rt->format = panfrost_mfbd_format(surf);
 222
 223         /* Now, we set the layout specific pieces */
 224
 225         if (rsrc->layout == PAN_LINEAR) {
 226                 rt->format.block = MALI_MFBD_BLOCK_LINEAR;
 227                 rt->framebuffer = base;
 228                 rt->framebuffer_stride = stride / 16;
 229         } else if (rsrc->layout == PAN_TILED) {
 230                 rt->format.block = MALI_MFBD_BLOCK_TILED;
 231                 rt->framebuffer = base;
 232                 rt->framebuffer_stride = stride;
 233         } else if (rsrc->layout == PAN_AFBC) {
 234                 rt->format.block = MALI_MFBD_BLOCK_AFBC;
 235
 236                 unsigned header_size = rsrc->slices[level].header_size;
 237
 238                 rt->framebuffer = base + header_size;
 239                 rt->afbc.metadata = base;
 240                 rt->afbc.stride = 0;
 241                 rt->afbc.unk = 0x30009;
 242
 243                 /* TODO: The blob sets this to something nonzero, but it's not
 244                  * clear what/how to calculate/if it matters */
 245                 rt->framebuffer_stride = 0;
 246         } else {
 247                 fprintf(stderr, "Invalid render layout (cbuf)");
 248                 assert(0);
 249         }
 250 }
 251
 252 /* Is a format encoded like Z24S8 and therefore compatible for render? */
 253
 254 static bool
 255 panfrost_is_z24s8_variant(enum pipe_format fmt)
 256 {
 257         switch (fmt) {
 258                 case PIPE_FORMAT_Z24_UNORM_S8_UINT:
 259                 case PIPE_FORMAT_Z24X8_UNORM:
 260                         return true;
 261                 default:
 262                         return false;
 263         }
 264 }
 265
 266 static void
 267 panfrost_mfbd_set_zsbuf(
 268         struct bifrost_framebuffer *fb,
 269         struct bifrost_fb_extra *fbx,
 270         struct pipe_surface *surf)
 271 {
 272         struct panfrost_resource *rsrc = pan_resource(surf->texture);
 273
 274         unsigned level = surf->u.tex.level;
 275         assert(surf->u.tex.first_layer == 0);
 276
 277         unsigned offset = rsrc->slices[level].offset;
 278
 279         if (rsrc->layout == PAN_AFBC) {
 280                 /* The only Z/S format we can compress is Z24S8 or variants
 281                  * thereof (handled by the state tracker) */
 282                 assert(panfrost_is_z24s8_variant(surf->format));
 283
 284                 mali_ptr base = rsrc->bo->gpu + offset;
 285                 unsigned header_size = rsrc->slices[level].header_size;
 286
 287                 fb->mfbd_flags |= MALI_MFBD_EXTRA;
 288
 289                 fbx->flags =
 290                         MALI_EXTRA_PRESENT |
 291                         MALI_EXTRA_AFBC |
 292                         MALI_EXTRA_AFBC_ZS |
 293                         MALI_EXTRA_ZS |
 294                         0x1; /* unknown */
 295
 296                 fbx->ds_afbc.depth_stencil = base + header_size;
 297                 fbx->ds_afbc.depth_stencil_afbc_metadata = base;
 298                 fbx->ds_afbc.depth_stencil_afbc_stride = 0;
 299
 300                 fbx->ds_afbc.zero1 = 0x10009;
 301                 fbx->ds_afbc.padding = 0x1000;
 302         } else if (rsrc->layout == PAN_LINEAR) {
 303                 /* TODO: Z32F(S8) support, which is always linear */
 304
 305                 int stride = rsrc->slices[level].stride;
 306
 307                 fb->mfbd_flags |= MALI_MFBD_EXTRA;
 308                 fbx->flags |= MALI_EXTRA_PRESENT | MALI_EXTRA_ZS;
 309
 310                 fbx->ds_linear.depth = rsrc->bo->gpu + offset;
 311                 fbx->ds_linear.depth_stride = stride;
 312
 313                 if (panfrost_is_z24s8_variant(surf->format)) {
 314                         fbx->flags |= 0x1;
 315                 } else if (surf->format == PIPE_FORMAT_Z32_UNORM) {
 316                         /* default flags (0 in bottom place) */
 317                 } else if (surf->format == PIPE_FORMAT_Z32_FLOAT) {
 318                         fbx->flags |= 0xA;
 319                         fb->mfbd_flags ^= 0x100;
 320                         fb->mfbd_flags |= 0x200;
 321                 } else if (surf->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT) {
 322                         fbx->flags |= 0x1000A;
 323                         fb->mfbd_flags ^= 0x100;
 324                         fb->mfbd_flags |= 0x201;
 325
 326                         struct panfrost_resource *stencil = rsrc->separate_stencil;
 327                         struct panfrost_slice stencil_slice = stencil->slices[level];
 328
 329                         fbx->ds_linear.stencil = stencil->bo->gpu + stencil_slice.offset;
 330                         fbx->ds_linear.stencil_stride = stencil_slice.stride;
 331                 }
 332
 333         } else {
 334                 assert(0);
 335         }
 336 }
 337
 338 /* Helper for sequential uploads used for MFBD */
 339
 340 #define UPLOAD(dest, offset, src, max) { \
 341         size_t sz = sizeof(*src); \
 342         memcpy(dest.cpu + offset, src, sz); \
 343         assert((offset + sz) <= max); \
 344         offset += sz; \
 345 }
 346
 347 static mali_ptr
 348 panfrost_mfbd_upload(
 349         struct panfrost_context *ctx,
 350         struct bifrost_framebuffer *fb,
 351         struct bifrost_fb_extra *fbx,
 352         struct bifrost_render_target *rts,
 353         unsigned rt_count)
 354 {
 355         off_t offset = 0;
 356
 357         /* There may be extra data stuck in the middle */
 358         bool has_extra = fb->mfbd_flags & MALI_MFBD_EXTRA;
 359
 360         /* Compute total size for transfer */
 361
 362         size_t total_sz =
 363                 sizeof(struct bifrost_framebuffer) +
 364                 (has_extra ? sizeof(struct bifrost_fb_extra) : 0) +
 365                 sizeof(struct bifrost_render_target) * 4;
 366
 367         struct panfrost_transfer m_f_trans =
 368                 panfrost_allocate_transient(ctx, total_sz);
 369
 370         /* Do the transfer */
 371
 372         UPLOAD(m_f_trans, offset, fb, total_sz);
 373
 374         if (has_extra)
 375                 UPLOAD(m_f_trans, offset, fbx, total_sz);
 376
 377         for (unsigned c = 0; c < 4; ++c) {
 378                 UPLOAD(m_f_trans, offset, &rts[c], total_sz);
 379         }
 380
 381         /* Return pointer suitable for the fragment section */
 382         unsigned tag =
 383                 MALI_MFBD |
 384                 (has_extra ? MALI_MFBD_TAG_EXTRA : 0) |
 385                 (MALI_POSITIVE(rt_count) << 2);
 386
 387         return m_f_trans.gpu | tag;
 388 }
 389
 390 #undef UPLOAD
 391
 392 /* Creates an MFBD for the FRAGMENT section of the bound framebuffer */
 393
 394 mali_ptr
 395 panfrost_mfbd_fragment(struct panfrost_context *ctx, bool has_draws)
 396 {
 397         struct panfrost_job *job = panfrost_get_job_for_fbo(ctx);
 398
 399         struct bifrost_framebuffer fb = panfrost_emit_mfbd(ctx, has_draws);
 400         struct bifrost_fb_extra fbx = {};
 401         struct bifrost_render_target rts[4] = {};
 402
 403         /* We always upload at least one dummy GL_NONE render target */
 404
 405         unsigned rt_descriptors =
 406                 MAX2(ctx->pipe_framebuffer.nr_cbufs, 1);
 407
 408         fb.rt_count_1 = MALI_POSITIVE(rt_descriptors);
 409         fb.rt_count_2 = rt_descriptors;
 410         fb.mfbd_flags = 0x100;
 411
 412         /* TODO: MRT clear */
 413         panfrost_mfbd_clear(job, &fb, &fbx, rts, fb.rt_count_2);
 414
 415
 416         /* Upload either the render target or a dummy GL_NONE target */
 417
 418         for (int cb = 0; cb < rt_descriptors; ++cb) {
 419                 struct pipe_surface *surf = ctx->pipe_framebuffer.cbufs[cb];
 420
 421                 if (surf) {
 422                         panfrost_mfbd_set_cbuf(&rts[cb], surf);
 423
 424                         /* What is this? Looks like some extension of the bpp
 425                          * field. Maybe it establishes how much internal
 426                          * tilebuffer space is reserved? */
 427
 428                         unsigned bpp = util_format_get_blocksize(surf->format);
 429                         fb.rt_count_2 = MAX2(fb.rt_count_2, ALIGN_POT(bpp, 4) / 4);
 430                 } else {
 431                         struct mali_rt_format null_rt = {
 432                                 .unk1 = 0x4000000,
 433                                 .no_preload = true
 434                         };
 435
 436                         rts[cb].format = null_rt;
 437                         rts[cb].framebuffer = 0;
 438                         rts[cb].framebuffer_stride = 0;
 439                 }
 440
 441                 /* TODO: Break out the field */
 442                 rts[cb].format.unk1 |= (cb * 0x400);
 443         }
 444
 445         if (ctx->pipe_framebuffer.zsbuf) {
 446                 panfrost_mfbd_set_zsbuf(&fb, &fbx, ctx->pipe_framebuffer.zsbuf);
 447         }
 448
 449         /* When scanning out, the depth buffer is immediately invalidated, so
 450          * we don't need to waste bandwidth writing it out. This can improve
 451          * performance substantially (Z32_UNORM 1080p @ 60fps is 475 MB/s of
 452          * memory bandwidth!).
 453          *
 454          * The exception is ReadPixels, but this is not supported on GLES so we
 455          * can safely ignore it. */
 456
 457         if (panfrost_is_scanout(ctx)) {
 458                 job->requirements &= ~PAN_REQ_DEPTH_WRITE;
 459         }
 460
 461         /* Actualize the requirements */
 462
 463         if (job->requirements & PAN_REQ_MSAA) {
 464                 rts[0].format.flags |= MALI_MFBD_FORMAT_MSAA;
 465
 466                 /* XXX */
 467                 fb.unk1 |= (1 << 4) | (1 << 1);
 468                 fb.rt_count_2 = 4;
 469         }
 470
 471         if (job->requirements & PAN_REQ_DEPTH_WRITE)
 472                 fb.mfbd_flags |= MALI_MFBD_DEPTH_WRITE;
 473
 474         /* Checksumming only works with a single render target */
 475
 476         if (ctx->pipe_framebuffer.nr_cbufs == 1) {
 477                 struct pipe_surface *surf = ctx->pipe_framebuffer.cbufs[0];
 478                 struct panfrost_resource *rsrc = pan_resource(surf->texture);
 479                 struct panfrost_bo *bo = rsrc->bo;
 480
 481                 if (rsrc->checksummed) {
 482                         unsigned level = surf->u.tex.level;
 483                         struct panfrost_slice *slice = &rsrc->slices[level];
 484
 485                         fb.mfbd_flags |= MALI_MFBD_EXTRA;
 486                         fbx.flags |= MALI_EXTRA_PRESENT;
 487                         fbx.checksum_stride = slice->checksum_stride;
 488                         fbx.checksum = bo->gpu + slice->checksum_offset;
 489                 }
 490         }
 491
 492         return panfrost_mfbd_upload(ctx, &fb, &fbx, rts, rt_descriptors);
 493 }