2 * Copyright © 2014 Broadcom
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
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
27 #include "pipe/p_defines.h"
28 #include "util/ralloc.h"
29 #include "util/u_inlines.h"
30 #include "util/u_memory.h"
31 #include "util/u_blitter.h"
32 #include "indices/u_primconvert.h"
33 #include "pipe/p_screen.h"
35 #include "vc4_screen.h"
36 #include "vc4_context.h"
37 #include "vc4_resource.h"
40 * Emits a no-op STORE_TILE_BUFFER_GENERAL.
42 * If we emit a PACKET_TILE_COORDINATES, it must be followed by a store of
43 * some sort before another load is triggered.
46 vc4_store_before_load(struct vc4_context
*vc4
, bool *coords_emitted
)
51 cl_u8(&vc4
->rcl
, VC4_PACKET_STORE_TILE_BUFFER_GENERAL
);
52 cl_u8(&vc4
->rcl
, VC4_LOADSTORE_TILE_BUFFER_NONE
);
53 cl_u8(&vc4
->rcl
, (VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR
|
54 VC4_STORE_TILE_BUFFER_DISABLE_ZS_CLEAR
|
55 VC4_STORE_TILE_BUFFER_DISABLE_VG_MASK_CLEAR
));
56 cl_u32(&vc4
->rcl
, 0); /* no address, since we're in None mode */
58 *coords_emitted
= false;
62 * Emits a PACKET_TILE_COORDINATES if one isn't already pending.
64 * The tile coordinates packet triggers a pending load if there is one, are
65 * used for clipping during rendering, and determine where loads/stores happen
66 * relative to their base address.
69 vc4_tile_coordinates(struct vc4_context
*vc4
, uint32_t x
, uint32_t y
,
75 cl_u8(&vc4
->rcl
, VC4_PACKET_TILE_COORDINATES
);
79 *coords_emitted
= true;
83 vc4_setup_rcl(struct vc4_context
*vc4
)
85 struct vc4_surface
*csurf
= vc4_surface(vc4
->framebuffer
.cbufs
[0]);
86 struct vc4_resource
*ctex
= csurf
? vc4_resource(csurf
->base
.texture
) : NULL
;
87 struct vc4_surface
*zsurf
= vc4_surface(vc4
->framebuffer
.zsbuf
);
88 struct vc4_resource
*ztex
= zsurf
? vc4_resource(zsurf
->base
.texture
) : NULL
;
91 vc4
->resolve
&= ~PIPE_CLEAR_COLOR0
;
93 vc4
->resolve
&= ~(PIPE_CLEAR_DEPTH
| PIPE_CLEAR_STENCIL
);
94 uint32_t resolve_uncleared
= vc4
->resolve
& ~vc4
->cleared
;
95 uint32_t width
= vc4
->framebuffer
.width
;
96 uint32_t height
= vc4
->framebuffer
.height
;
97 uint32_t xtiles
= align(width
, 64) / 64;
98 uint32_t ytiles
= align(height
, 64) / 64;
101 fprintf(stderr
, "RCL: resolve 0x%x clear 0x%x resolve uncleared 0x%x\n",
107 uint32_t reloc_size
= 9;
108 uint32_t clear_size
= 14;
109 uint32_t config_size
= 11 + reloc_size
;
110 uint32_t loadstore_size
= 7 + reloc_size
;
111 uint32_t tilecoords_size
= 3;
112 uint32_t branch_size
= 5 + reloc_size
;
113 uint32_t color_store_size
= 1;
114 cl_ensure_space(&vc4
->rcl
,
118 xtiles
* ytiles
* (loadstore_size
* 4 +
119 tilecoords_size
* 3 +
123 cl_u8(&vc4
->rcl
, VC4_PACKET_CLEAR_COLORS
);
124 cl_u32(&vc4
->rcl
, vc4
->clear_color
[0]);
125 cl_u32(&vc4
->rcl
, vc4
->clear_color
[1]);
126 cl_u32(&vc4
->rcl
, vc4
->clear_depth
);
127 cl_u8(&vc4
->rcl
, vc4
->clear_stencil
);
129 /* The rendering mode config determines the pointer that's used for
130 * VC4_PACKET_STORE_MS_TILE_BUFFER address computations. The kernel
131 * could handle a no-relocation rendering mode config and deny those
132 * packets, but instead we just tell the kernel we're doing our color
133 * rendering to the Z buffer, and just don't emit any of those
136 struct vc4_surface
*render_surf
= csurf
? csurf
: zsurf
;
137 struct vc4_resource
*render_tex
= vc4_resource(render_surf
->base
.texture
);
138 cl_start_reloc(&vc4
->rcl
, 1);
139 cl_u8(&vc4
->rcl
, VC4_PACKET_TILE_RENDERING_MODE_CONFIG
);
140 cl_reloc(vc4
, &vc4
->rcl
, render_tex
->bo
, render_surf
->offset
);
141 cl_u16(&vc4
->rcl
, width
);
142 cl_u16(&vc4
->rcl
, height
);
143 cl_u16(&vc4
->rcl
, ((render_surf
->tiling
<<
144 VC4_RENDER_CONFIG_MEMORY_FORMAT_SHIFT
) |
145 (vc4_rt_format_is_565(render_surf
->base
.format
) ?
146 VC4_RENDER_CONFIG_FORMAT_BGR565
:
147 VC4_RENDER_CONFIG_FORMAT_RGBA8888
) |
148 VC4_RENDER_CONFIG_EARLY_Z_COVERAGE_DISABLE
));
150 /* The tile buffer normally gets cleared when the previous tile is
151 * stored. If the clear values changed between frames, then the tile
152 * buffer has stale clear values in it, so we have to do a store in
153 * None mode (no writes) so that we trigger the tile buffer clear.
155 * Excess clearing is only a performance cost, since per-tile contents
156 * will be loaded/stored in the loop below.
158 if (vc4
->cleared
& (PIPE_CLEAR_COLOR0
|
160 PIPE_CLEAR_STENCIL
)) {
161 cl_u8(&vc4
->rcl
, VC4_PACKET_TILE_COORDINATES
);
165 cl_u8(&vc4
->rcl
, VC4_PACKET_STORE_TILE_BUFFER_GENERAL
);
166 cl_u16(&vc4
->rcl
, VC4_LOADSTORE_TILE_BUFFER_NONE
);
167 cl_u32(&vc4
->rcl
, 0); /* no address, since we're in None mode */
170 uint32_t color_hindex
= ctex
? vc4_gem_hindex(vc4
, ctex
->bo
) : 0;
171 uint32_t depth_hindex
= ztex
? vc4_gem_hindex(vc4
, ztex
->bo
) : 0;
172 uint32_t tile_alloc_hindex
= vc4_gem_hindex(vc4
, vc4
->tile_alloc
);
174 for (int y
= 0; y
< ytiles
; y
++) {
175 for (int x
= 0; x
< xtiles
; x
++) {
176 bool end_of_frame
= (x
== xtiles
- 1 &&
178 bool coords_emitted
= false;
180 /* Note that the load doesn't actually occur until the
181 * tile coords packet is processed, and only one load
182 * may be outstanding at a time.
184 if (resolve_uncleared
& PIPE_CLEAR_COLOR
) {
185 vc4_store_before_load(vc4
, &coords_emitted
);
187 cl_start_reloc(&vc4
->rcl
, 1);
188 cl_u8(&vc4
->rcl
, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL
);
190 VC4_LOADSTORE_TILE_BUFFER_COLOR
|
192 VC4_LOADSTORE_TILE_BUFFER_FORMAT_SHIFT
));
194 vc4_rt_format_is_565(csurf
->base
.format
) ?
195 VC4_LOADSTORE_TILE_BUFFER_BGR565
:
196 VC4_LOADSTORE_TILE_BUFFER_RGBA8888
);
197 cl_reloc_hindex(&vc4
->rcl
, color_hindex
,
200 vc4_tile_coordinates(vc4
, x
, y
, &coords_emitted
);
203 if (resolve_uncleared
& (PIPE_CLEAR_DEPTH
| PIPE_CLEAR_STENCIL
)) {
204 vc4_store_before_load(vc4
, &coords_emitted
);
206 cl_start_reloc(&vc4
->rcl
, 1);
207 cl_u8(&vc4
->rcl
, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL
);
209 VC4_LOADSTORE_TILE_BUFFER_ZS
|
211 VC4_LOADSTORE_TILE_BUFFER_FORMAT_SHIFT
));
213 cl_reloc_hindex(&vc4
->rcl
, depth_hindex
,
216 vc4_tile_coordinates(vc4
, x
, y
, &coords_emitted
);
219 /* Clipping depends on tile coordinates having been
220 * emitted, so make sure it's happened even if
221 * everything was cleared to start.
223 vc4_tile_coordinates(vc4
, x
, y
, &coords_emitted
);
225 /* Wait for the binner before jumping to the first
228 if (x
== 0 && y
== 0)
229 cl_u8(&vc4
->rcl
, VC4_PACKET_WAIT_ON_SEMAPHORE
);
231 cl_start_reloc(&vc4
->rcl
, 1);
232 cl_u8(&vc4
->rcl
, VC4_PACKET_BRANCH_TO_SUB_LIST
);
233 cl_reloc_hindex(&vc4
->rcl
, tile_alloc_hindex
,
234 (y
* xtiles
+ x
) * 32);
236 if (vc4
->resolve
& (PIPE_CLEAR_DEPTH
| PIPE_CLEAR_STENCIL
)) {
237 vc4_tile_coordinates(vc4
, x
, y
, &coords_emitted
);
239 cl_start_reloc(&vc4
->rcl
, 1);
240 cl_u8(&vc4
->rcl
, VC4_PACKET_STORE_TILE_BUFFER_GENERAL
);
242 VC4_LOADSTORE_TILE_BUFFER_ZS
|
244 VC4_LOADSTORE_TILE_BUFFER_FORMAT_SHIFT
));
246 VC4_STORE_TILE_BUFFER_DISABLE_COLOR_CLEAR
);
247 cl_reloc_hindex(&vc4
->rcl
, depth_hindex
,
250 !(vc4
->resolve
& PIPE_CLEAR_COLOR0
)) ?
251 VC4_LOADSTORE_TILE_BUFFER_EOF
: 0));
253 coords_emitted
= false;
256 if (vc4
->resolve
& PIPE_CLEAR_COLOR0
) {
257 vc4_tile_coordinates(vc4
, x
, y
, &coords_emitted
);
260 VC4_PACKET_STORE_MS_TILE_BUFFER_AND_EOF
);
263 VC4_PACKET_STORE_MS_TILE_BUFFER
);
266 coords_emitted
= false;
269 /* One of the bits needs to have been set that would
270 * have triggered an EOF.
272 assert(vc4
->resolve
& (PIPE_CLEAR_COLOR0
|
274 PIPE_CLEAR_STENCIL
));
275 /* Any coords emitted must also have been consumed by
278 assert(!coords_emitted
);
282 if (vc4
->resolve
& PIPE_CLEAR_COLOR0
)
285 if (vc4
->resolve
& (PIPE_CLEAR_DEPTH
| PIPE_CLEAR_STENCIL
))
290 vc4_flush(struct pipe_context
*pctx
)
292 struct vc4_context
*vc4
= vc4_context(pctx
);
294 if (!vc4
->needs_flush
)
297 /* Increment the semaphore indicating that binning is done and
298 * unblocking the render thread. Note that this doesn't act until the
301 cl_u8(&vc4
->bcl
, VC4_PACKET_INCREMENT_SEMAPHORE
);
302 /* The FLUSH caps all of our bin lists with a VC4_PACKET_RETURN. */
303 cl_u8(&vc4
->bcl
, VC4_PACKET_FLUSH
);
307 if (vc4_debug
& VC4_DEBUG_CL
) {
308 fprintf(stderr
, "BCL:\n");
309 vc4_dump_cl(vc4
->bcl
.base
, vc4
->bcl
.size
, false);
310 fprintf(stderr
, "RCL:\n");
311 vc4_dump_cl(vc4
->rcl
.base
, vc4
->rcl
.size
, true);
314 struct drm_vc4_submit_cl submit
;
315 memset(&submit
, 0, sizeof(submit
));
317 submit
.bo_handles
= vc4
->bo_handles
.base
;
318 submit
.bo_handle_count
= (vc4
->bo_handles
.next
-
319 vc4
->bo_handles
.base
) / 4;
320 submit
.bin_cl
= vc4
->bcl
.base
;
321 submit
.bin_cl_size
= vc4
->bcl
.next
- vc4
->bcl
.base
;
322 submit
.render_cl
= vc4
->rcl
.base
;
323 submit
.render_cl_size
= vc4
->rcl
.next
- vc4
->rcl
.base
;
324 submit
.shader_rec
= vc4
->shader_rec
.base
;
325 submit
.shader_rec_size
= vc4
->shader_rec
.next
- vc4
->shader_rec
.base
;
326 submit
.shader_rec_count
= vc4
->shader_rec_count
;
327 submit
.uniforms
= vc4
->uniforms
.base
;
328 submit
.uniforms_size
= vc4
->uniforms
.next
- vc4
->uniforms
.base
;
330 if (!(vc4_debug
& VC4_DEBUG_NORAST
)) {
333 #ifndef USE_VC4_SIMULATOR
334 ret
= drmIoctl(vc4
->fd
, DRM_IOCTL_VC4_SUBMIT_CL
, &submit
);
336 ret
= vc4_simulator_flush(vc4
, &submit
);
339 fprintf(stderr
, "VC4 submit failed\n");
344 vc4
->last_emit_seqno
= submit
.seqno
;
346 if (vc4_debug
& VC4_DEBUG_ALWAYS_SYNC
) {
347 if (!vc4_wait_seqno(vc4
->screen
, vc4
->last_emit_seqno
,
348 PIPE_TIMEOUT_INFINITE
)) {
349 fprintf(stderr
, "Wait failed.\n");
354 vc4_reset_cl(&vc4
->bcl
);
355 vc4_reset_cl(&vc4
->rcl
);
356 vc4_reset_cl(&vc4
->shader_rec
);
357 vc4_reset_cl(&vc4
->uniforms
);
358 vc4_reset_cl(&vc4
->bo_handles
);
359 struct vc4_bo
**referenced_bos
= vc4
->bo_pointers
.base
;
360 for (int i
= 0; i
< submit
.bo_handle_count
; i
++)
361 vc4_bo_unreference(&referenced_bos
[i
]);
362 vc4_reset_cl(&vc4
->bo_pointers
);
363 vc4
->shader_rec_count
= 0;
365 vc4
->needs_flush
= false;
366 vc4
->draw_call_queued
= false;
368 /* We have no hardware context saved between our draw calls, so we
369 * need to flag the next draw as needing all state emitted. Emitting
370 * all state at the start of our draws is also what ensures that we
371 * return to the state we need after a previous tile has finished.
379 vc4_pipe_flush(struct pipe_context
*pctx
, struct pipe_fence_handle
**fence
,
382 struct vc4_context
*vc4
= vc4_context(pctx
);
387 struct vc4_fence
*f
= vc4_fence_create(vc4
->screen
,
388 vc4
->last_emit_seqno
);
389 *fence
= (struct pipe_fence_handle
*)f
;
394 * Flushes the current command lists if they reference the given BO.
396 * This helps avoid flushing the command buffers when unnecessary.
399 vc4_cl_references_bo(struct pipe_context
*pctx
, struct vc4_bo
*bo
)
401 struct vc4_context
*vc4
= vc4_context(pctx
);
403 if (!vc4
->needs_flush
)
406 /* Walk all the referenced BOs in the drawing command list to see if
409 struct vc4_bo
**referenced_bos
= vc4
->bo_pointers
.base
;
410 for (int i
= 0; i
< (vc4
->bo_handles
.next
-
411 vc4
->bo_handles
.base
) / 4; i
++) {
412 if (referenced_bos
[i
] == bo
) {
417 /* Also check for the Z/color buffers, since the references to those
418 * are only added immediately before submit.
420 struct vc4_surface
*csurf
= vc4_surface(vc4
->framebuffer
.cbufs
[0]);
422 struct vc4_resource
*ctex
= vc4_resource(csurf
->base
.texture
);
423 if (ctex
->bo
== bo
) {
428 struct vc4_surface
*zsurf
= vc4_surface(vc4
->framebuffer
.zsbuf
);
430 struct vc4_resource
*ztex
=
431 vc4_resource(zsurf
->base
.texture
);
432 if (ztex
->bo
== bo
) {
441 vc4_context_destroy(struct pipe_context
*pctx
)
443 struct vc4_context
*vc4
= vc4_context(pctx
);
446 util_blitter_destroy(vc4
->blitter
);
448 if (vc4
->primconvert
)
449 util_primconvert_destroy(vc4
->primconvert
);
451 util_slab_destroy(&vc4
->transfer_pool
);
453 pipe_surface_reference(&vc4
->framebuffer
.cbufs
[0], NULL
);
454 pipe_surface_reference(&vc4
->framebuffer
.zsbuf
, NULL
);
455 vc4_bo_unreference(&vc4
->tile_alloc
);
456 vc4_bo_unreference(&vc4
->tile_state
);
458 vc4_program_fini(pctx
);
463 struct pipe_context
*
464 vc4_context_create(struct pipe_screen
*pscreen
, void *priv
)
466 struct vc4_screen
*screen
= vc4_screen(pscreen
);
467 struct vc4_context
*vc4
;
469 /* Prevent dumping of the shaders built during context setup. */
470 uint32_t saved_shaderdb_flag
= vc4_debug
& VC4_DEBUG_SHADERDB
;
471 vc4_debug
&= ~VC4_DEBUG_SHADERDB
;
473 vc4
= rzalloc(NULL
, struct vc4_context
);
476 struct pipe_context
*pctx
= &vc4
->base
;
478 vc4
->screen
= screen
;
480 pctx
->screen
= pscreen
;
482 pctx
->destroy
= vc4_context_destroy
;
483 pctx
->flush
= vc4_pipe_flush
;
486 vc4_state_init(pctx
);
487 vc4_program_init(pctx
);
488 vc4_query_init(pctx
);
489 vc4_resource_context_init(pctx
);
491 vc4_init_cl(vc4
, &vc4
->bcl
);
492 vc4_init_cl(vc4
, &vc4
->rcl
);
493 vc4_init_cl(vc4
, &vc4
->shader_rec
);
494 vc4_init_cl(vc4
, &vc4
->uniforms
);
495 vc4_init_cl(vc4
, &vc4
->bo_handles
);
496 vc4_init_cl(vc4
, &vc4
->bo_pointers
);
499 vc4
->fd
= screen
->fd
;
501 util_slab_create(&vc4
->transfer_pool
, sizeof(struct vc4_transfer
),
502 16, UTIL_SLAB_SINGLETHREADED
);
503 vc4
->blitter
= util_blitter_create(pctx
);
507 vc4
->primconvert
= util_primconvert_create(pctx
,
508 (1 << PIPE_PRIM_QUADS
) - 1);
509 if (!vc4
->primconvert
)
512 vc4_debug
|= saved_shaderdb_flag
;