2 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
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:
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
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
24 * Rob Clark <robclark@freedesktop.org>
27 #include "pipe/p_state.h"
28 #include "util/u_string.h"
29 #include "util/u_memory.h"
30 #include "util/u_inlines.h"
31 #include "util/u_format.h"
33 #include "freedreno_gmem.h"
34 #include "freedreno_context.h"
35 #include "freedreno_fence.h"
36 #include "freedreno_resource.h"
37 #include "freedreno_query_hw.h"
38 #include "freedreno_util.h"
41 * GMEM is the small (ie. 256KiB for a200, 512KiB for a220, etc) tile buffer
42 * inside the GPU. All rendering happens to GMEM. Larger render targets
43 * are split into tiles that are small enough for the color (and depth and/or
44 * stencil, if enabled) buffers to fit within GMEM. Before rendering a tile,
45 * if there was not a clear invalidating the previous tile contents, we need
46 * to restore the previous tiles contents (system mem -> GMEM), and after all
47 * the draw calls, before moving to the next tile, we need to save the tile
48 * contents (GMEM -> system mem).
50 * The code in this file handles dealing with GMEM and tiling.
52 * The structure of the ringbuffer ends up being:
54 * +--<---<-- IB ---<---+---<---+---<---<---<--+
57 * ------------------------------------------------------
58 * | clear/draw cmds | Tile0 | Tile1 | .... | TileN |
59 * ------------------------------------------------------
62 * address submitted in issueibcmds
64 * Where the per-tile section handles scissor setup, mem2gmem restore (if
65 * needed), IB to draw cmds earlier in the ringbuffer, and then gmem2mem
69 static uint32_t bin_width(struct fd_screen
*screen
)
71 if (is_a4xx(screen
) || is_a5xx(screen
) || is_a6xx(screen
))
79 total_size(uint8_t cbuf_cpp
[], uint8_t zsbuf_cpp
[2],
80 uint32_t bin_w
, uint32_t bin_h
, uint32_t gmem_align
,
81 struct fd_gmem_stateobj
*gmem
)
83 uint32_t total
= 0, i
;
85 for (i
= 0; i
< MAX_RENDER_TARGETS
; i
++) {
87 gmem
->cbuf_base
[i
] = align(total
, gmem_align
);
88 total
= gmem
->cbuf_base
[i
] + cbuf_cpp
[i
] * bin_w
* bin_h
;
93 gmem
->zsbuf_base
[0] = align(total
, gmem_align
);
94 total
= gmem
->zsbuf_base
[0] + zsbuf_cpp
[0] * bin_w
* bin_h
;
98 gmem
->zsbuf_base
[1] = align(total
, gmem_align
);
99 total
= gmem
->zsbuf_base
[1] + zsbuf_cpp
[1] * bin_w
* bin_h
;
106 calculate_tiles(struct fd_batch
*batch
)
108 struct fd_context
*ctx
= batch
->ctx
;
109 struct fd_screen
*screen
= ctx
->screen
;
110 struct fd_gmem_stateobj
*gmem
= &ctx
->gmem
;
111 struct pipe_scissor_state
*scissor
= &batch
->max_scissor
;
112 struct pipe_framebuffer_state
*pfb
= &batch
->framebuffer
;
113 const uint32_t gmem_alignw
= screen
->gmem_alignw
;
114 const uint32_t gmem_alignh
= screen
->gmem_alignh
;
115 const unsigned npipes
= screen
->num_vsc_pipes
;
116 const uint32_t gmem_size
= screen
->gmemsize_bytes
;
117 uint32_t minx
, miny
, width
, height
;
118 uint32_t nbins_x
= 1, nbins_y
= 1;
119 uint32_t bin_w
, bin_h
;
120 uint32_t gmem_align
= 0x4000;
121 uint32_t max_width
= bin_width(screen
);
122 uint8_t cbuf_cpp
[MAX_RENDER_TARGETS
] = {0}, zsbuf_cpp
[2] = {0};
123 uint32_t i
, j
, t
, xoff
, yoff
;
124 uint32_t tpp_x
, tpp_y
;
125 bool has_zs
= !!(batch
->gmem_reason
& (FD_GMEM_DEPTH_ENABLED
|
126 FD_GMEM_STENCIL_ENABLED
| FD_GMEM_CLEARS_DEPTH_STENCIL
));
130 struct fd_resource
*rsc
= fd_resource(pfb
->zsbuf
->texture
);
131 zsbuf_cpp
[0] = rsc
->cpp
;
133 zsbuf_cpp
[1] = rsc
->stencil
->cpp
;
135 /* we might have a zsbuf, but it isn't used */
136 batch
->restore
&= ~(FD_BUFFER_DEPTH
| FD_BUFFER_STENCIL
);
137 batch
->resolve
&= ~(FD_BUFFER_DEPTH
| FD_BUFFER_STENCIL
);
139 for (i
= 0; i
< pfb
->nr_cbufs
; i
++) {
141 cbuf_cpp
[i
] = util_format_get_blocksize(pfb
->cbufs
[i
]->format
);
144 /* if MSAA, color buffers are super-sampled in GMEM: */
145 cbuf_cpp
[i
] *= pfb
->samples
;
148 if (!memcmp(gmem
->zsbuf_cpp
, zsbuf_cpp
, sizeof(zsbuf_cpp
)) &&
149 !memcmp(gmem
->cbuf_cpp
, cbuf_cpp
, sizeof(cbuf_cpp
)) &&
150 !memcmp(&gmem
->scissor
, scissor
, sizeof(gmem
->scissor
))) {
151 /* everything is up-to-date */
155 if (fd_mesa_debug
& FD_DBG_NOSCIS
) {
159 height
= pfb
->height
;
161 /* round down to multiple of alignment: */
162 minx
= scissor
->minx
& ~(gmem_alignw
- 1);
163 miny
= scissor
->miny
& ~(gmem_alignh
- 1);
164 width
= scissor
->maxx
- minx
;
165 height
= scissor
->maxy
- miny
;
168 bin_w
= align(width
, gmem_alignw
);
169 bin_h
= align(height
, gmem_alignh
);
171 /* first, find a bin width that satisfies the maximum width
174 while (bin_w
> max_width
) {
176 bin_w
= align(width
/ nbins_x
, gmem_alignw
);
179 if (fd_mesa_debug
& FD_DBG_MSGS
) {
180 debug_printf("binning input: cbuf cpp:");
181 for (i
= 0; i
< pfb
->nr_cbufs
; i
++)
182 debug_printf(" %d", cbuf_cpp
[i
]);
183 debug_printf(", zsbuf cpp: %d; %dx%d\n",
184 zsbuf_cpp
[0], width
, height
);
187 if (is_a20x(screen
) && batch
->cleared
) {
188 /* under normal circumstances the requirement would be 4K
189 * but the fast clear path requires an alignment of 32K
194 /* then find a bin width/height that satisfies the memory
197 while (total_size(cbuf_cpp
, zsbuf_cpp
, bin_w
, bin_h
, gmem_align
, gmem
) >
201 bin_w
= align(width
/ nbins_x
, gmem_alignw
);
204 bin_h
= align(height
/ nbins_y
, gmem_alignh
);
208 DBG("using %d bins of size %dx%d", nbins_x
*nbins_y
, bin_w
, bin_h
);
210 gmem
->scissor
= *scissor
;
211 memcpy(gmem
->cbuf_cpp
, cbuf_cpp
, sizeof(cbuf_cpp
));
212 memcpy(gmem
->zsbuf_cpp
, zsbuf_cpp
, sizeof(zsbuf_cpp
));
215 gmem
->nbins_x
= nbins_x
;
216 gmem
->nbins_y
= nbins_y
;
220 gmem
->height
= height
;
223 * Assign tiles and pipes:
225 * At some point it might be worth playing with different
226 * strategies and seeing if that makes much impact on
230 #define div_round_up(v, a) (((v) + (a) - 1) / (a))
231 /* figure out number of tiles per pipe: */
232 if (is_a20x(ctx
->screen
)) {
233 /* for a20x we want to minimize the number of "pipes"
234 * binning data has 3 bits for x/y (8x8) but the edges are used to
235 * cull off-screen vertices with hw binning, so we have 6x6 pipes
241 while (div_round_up(nbins_y
, tpp_y
) > screen
->num_vsc_pipes
)
243 while ((div_round_up(nbins_y
, tpp_y
) *
244 div_round_up(nbins_x
, tpp_x
)) > screen
->num_vsc_pipes
)
251 /* configure pipes: */
253 for (i
= 0; i
< npipes
; i
++) {
254 struct fd_vsc_pipe
*pipe
= &ctx
->vsc_pipe
[i
];
256 if (xoff
>= nbins_x
) {
261 if (yoff
>= nbins_y
) {
267 pipe
->w
= MIN2(tpp_x
, nbins_x
- xoff
);
268 pipe
->h
= MIN2(tpp_y
, nbins_y
- yoff
);
273 /* number of pipes to use for a20x */
274 gmem
->num_vsc_pipes
= MAX2(1, i
);
276 for (; i
< npipes
; i
++) {
277 struct fd_vsc_pipe
*pipe
= &ctx
->vsc_pipe
[i
];
278 pipe
->x
= pipe
->y
= pipe
->w
= pipe
->h
= 0;
282 printf("%dx%d ... tpp=%dx%d\n", nbins_x
, nbins_y
, tpp_x
, tpp_y
);
283 for (i
= 0; i
< 8; i
++) {
284 struct fd_vsc_pipe
*pipe
= &ctx
->pipe
[i
];
285 printf("pipe[%d]: %ux%u @ %u,%u\n", i
,
286 pipe
->w
, pipe
->h
, pipe
->x
, pipe
->y
);
290 /* configure tiles: */
293 memset(tile_n
, 0, sizeof(tile_n
));
294 for (i
= 0; i
< nbins_y
; i
++) {
299 /* clip bin height: */
300 bh
= MIN2(bin_h
, miny
+ height
- yoff
);
302 for (j
= 0; j
< nbins_x
; j
++) {
303 struct fd_tile
*tile
= &ctx
->tile
[t
];
306 assert(t
< ARRAY_SIZE(ctx
->tile
));
309 p
= ((i
/ tpp_y
) * div_round_up(nbins_x
, tpp_x
)) + (j
/ tpp_x
);
310 assert(p
< gmem
->num_vsc_pipes
);
312 /* clip bin width: */
313 bw
= MIN2(bin_w
, minx
+ width
- xoff
);
314 tile
->n
= !is_a20x(ctx
->screen
) ? tile_n
[p
]++ :
315 ((i
% tpp_y
+ 1) << 3 | (j
% tpp_x
+ 1));
332 for (i
= 0; i
< nbins_y
; i
++) {
333 for (j
= 0; j
< nbins_x
; j
++) {
334 struct fd_tile
*tile
= &ctx
->tile
[t
++];
335 printf("|p:%u n:%u|", tile
->p
, tile
->n
);
343 render_tiles(struct fd_batch
*batch
)
345 struct fd_context
*ctx
= batch
->ctx
;
346 struct fd_gmem_stateobj
*gmem
= &ctx
->gmem
;
349 ctx
->emit_tile_init(batch
);
352 ctx
->stats
.batch_restore
++;
354 for (i
= 0; i
< (gmem
->nbins_x
* gmem
->nbins_y
); i
++) {
355 struct fd_tile
*tile
= &ctx
->tile
[i
];
357 DBG("bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
358 tile
->bin_h
, tile
->yoff
, tile
->bin_w
, tile
->xoff
);
360 ctx
->emit_tile_prep(batch
, tile
);
362 if (batch
->restore
) {
363 ctx
->emit_tile_mem2gmem(batch
, tile
);
366 ctx
->emit_tile_renderprep(batch
, tile
);
368 if (ctx
->query_prepare_tile
)
369 ctx
->query_prepare_tile(batch
, i
, batch
->gmem
);
371 /* emit IB to drawcmds: */
372 ctx
->emit_ib(batch
->gmem
, batch
->draw
);
375 /* emit gmem2mem to transfer tile back to system memory: */
376 ctx
->emit_tile_gmem2mem(batch
, tile
);
379 if (ctx
->emit_tile_fini
)
380 ctx
->emit_tile_fini(batch
);
384 render_sysmem(struct fd_batch
*batch
)
386 struct fd_context
*ctx
= batch
->ctx
;
388 ctx
->emit_sysmem_prep(batch
);
390 if (ctx
->query_prepare_tile
)
391 ctx
->query_prepare_tile(batch
, 0, batch
->gmem
);
393 /* emit IB to drawcmds: */
394 ctx
->emit_ib(batch
->gmem
, batch
->draw
);
397 if (ctx
->emit_sysmem_fini
)
398 ctx
->emit_sysmem_fini(batch
);
402 flush_ring(struct fd_batch
*batch
)
405 int out_fence_fd
= -1;
407 fd_submit_flush(batch
->submit
, batch
->in_fence_fd
,
408 batch
->needs_out_fence_fd
? &out_fence_fd
: NULL
,
411 fd_fence_populate(batch
->fence
, timestamp
, out_fence_fd
);
415 fd_gmem_render_tiles(struct fd_batch
*batch
)
417 struct fd_context
*ctx
= batch
->ctx
;
418 struct pipe_framebuffer_state
*pfb
= &batch
->framebuffer
;
421 if (ctx
->emit_sysmem_prep
&& !batch
->nondraw
) {
422 if (batch
->cleared
|| batch
->gmem_reason
||
423 ((batch
->num_draws
> 5) && !batch
->blit
) ||
424 (pfb
->samples
> 1)) {
425 DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u, samples=%u",
426 batch
->cleared
, batch
->gmem_reason
, batch
->num_draws
,
428 } else if (!(fd_mesa_debug
& FD_DBG_NOBYPASS
)) {
432 /* For ARB_framebuffer_no_attachments: */
433 if ((pfb
->nr_cbufs
== 0) && !pfb
->zsbuf
) {
440 ctx
->stats
.batch_total
++;
442 if (batch
->nondraw
) {
443 DBG("%p: rendering non-draw", batch
);
444 ctx
->stats
.batch_nondraw
++;
446 DBG("%p: rendering sysmem %ux%u (%s/%s), num_draws=%u",
447 batch
, pfb
->width
, pfb
->height
,
448 util_format_short_name(pipe_surface_format(pfb
->cbufs
[0])),
449 util_format_short_name(pipe_surface_format(pfb
->zsbuf
)),
451 if (ctx
->query_prepare
)
452 ctx
->query_prepare(batch
, 1);
453 render_sysmem(batch
);
454 ctx
->stats
.batch_sysmem
++;
456 struct fd_gmem_stateobj
*gmem
= &ctx
->gmem
;
457 calculate_tiles(batch
);
458 DBG("%p: rendering %dx%d tiles %ux%u (%s/%s)",
459 batch
, pfb
->width
, pfb
->height
, gmem
->nbins_x
, gmem
->nbins_y
,
460 util_format_short_name(pipe_surface_format(pfb
->cbufs
[0])),
461 util_format_short_name(pipe_surface_format(pfb
->zsbuf
)));
462 if (ctx
->query_prepare
)
463 ctx
->query_prepare(batch
, gmem
->nbins_x
* gmem
->nbins_y
);
465 ctx
->stats
.batch_gmem
++;
471 /* When deciding whether a tile needs mem2gmem, we need to take into
472 * account the scissor rect(s) that were cleared. To simplify we only
473 * consider the last scissor rect for each buffer, since the common
474 * case would be a single clear.
477 fd_gmem_needs_restore(struct fd_batch
*batch
, struct fd_tile
*tile
,
480 if (!(batch
->restore
& buffers
))