1/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
2
3/*
4 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the "Software"),
8 * to deal in the Software without restriction, including without limitation
9 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10 * and/or sell copies of the Software, and to permit persons to whom the
11 * Software is furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice (including the next
14 * paragraph) shall be included in all copies or substantial portions of the
15 * Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 * SOFTWARE.
24 *
25 * Authors:
26 *    Rob Clark <robclark@freedesktop.org>
27 */
28
29#include "pipe/p_state.h"
30#include "util/u_string.h"
31#include "util/u_memory.h"
32#include "util/u_inlines.h"
33#include "util/u_format.h"
34
35#include "freedreno_gmem.h"
36#include "freedreno_context.h"
37#include "freedreno_fence.h"
38#include "freedreno_resource.h"
39#include "freedreno_query_hw.h"
40#include "freedreno_util.h"
41
42/*
43 * GMEM is the small (ie. 256KiB for a200, 512KiB for a220, etc) tile buffer
44 * inside the GPU.  All rendering happens to GMEM.  Larger render targets
45 * are split into tiles that are small enough for the color (and depth and/or
46 * stencil, if enabled) buffers to fit within GMEM.  Before rendering a tile,
47 * if there was not a clear invalidating the previous tile contents, we need
48 * to restore the previous tiles contents (system mem -> GMEM), and after all
49 * the draw calls, before moving to the next tile, we need to save the tile
50 * contents (GMEM -> system mem).
51 *
52 * The code in this file handles dealing with GMEM and tiling.
53 *
54 * The structure of the ringbuffer ends up being:
55 *
56 *     +--<---<-- IB ---<---+---<---+---<---<---<--+
57 *     |                    |       |              |
58 *     v                    ^       ^              ^
59 *   ------------------------------------------------------
60 *     | clear/draw cmds | Tile0 | Tile1 | .... | TileN |
61 *   ------------------------------------------------------
62 *                       ^
63 *                       |
64 *                       address submitted in issueibcmds
65 *
66 * Where the per-tile section handles scissor setup, mem2gmem restore (if
67 * needed), IB to draw cmds earlier in the ringbuffer, and then gmem2mem
68 * resolve.
69 */
70
71static uint32_t bin_width(struct fd_screen *screen)
72{
73	if (is_a4xx(screen) || is_a5xx(screen))
74		return 1024;
75	if (is_a3xx(screen))
76		return 992;
77	return 512;
78}
79
80static uint32_t
81total_size(uint8_t cbuf_cpp[], uint8_t zsbuf_cpp[2],
82		   uint32_t bin_w, uint32_t bin_h, struct fd_gmem_stateobj *gmem)
83{
84	uint32_t total = 0, i;
85
86	for (i = 0; i < MAX_RENDER_TARGETS; i++) {
87		if (cbuf_cpp[i]) {
88			gmem->cbuf_base[i] = align(total, 0x4000);
89			total = gmem->cbuf_base[i] + cbuf_cpp[i] * bin_w * bin_h;
90		}
91	}
92
93	if (zsbuf_cpp[0]) {
94		gmem->zsbuf_base[0] = align(total, 0x4000);
95		total = gmem->zsbuf_base[0] + zsbuf_cpp[0] * bin_w * bin_h;
96	}
97
98	if (zsbuf_cpp[1]) {
99		gmem->zsbuf_base[1] = align(total, 0x4000);
100		total = gmem->zsbuf_base[1] + zsbuf_cpp[1] * bin_w * bin_h;
101	}
102
103	return total;
104}
105
106static void
107calculate_tiles(struct fd_batch *batch)
108{
109	struct fd_context *ctx = batch->ctx;
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 = ctx->screen->gmem_alignw;
114	const uint32_t gmem_alignh = ctx->screen->gmem_alignh;
115	const uint32_t gmem_size = ctx->screen->gmemsize_bytes;
116	uint32_t minx, miny, width, height;
117	uint32_t nbins_x = 1, nbins_y = 1;
118	uint32_t bin_w, bin_h;
119	uint32_t max_width = bin_width(ctx->screen);
120	uint8_t cbuf_cpp[MAX_RENDER_TARGETS] = {0}, zsbuf_cpp[2] = {0};
121	uint32_t i, j, t, xoff, yoff;
122	uint32_t tpp_x, tpp_y;
123	bool has_zs = !!(batch->resolve & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL));
124	int tile_n[ARRAY_SIZE(ctx->pipe)];
125
126	if (has_zs) {
127		struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
128		zsbuf_cpp[0] = rsc->cpp;
129		if (rsc->stencil)
130			zsbuf_cpp[1] = rsc->stencil->cpp;
131	}
132	for (i = 0; i < pfb->nr_cbufs; i++) {
133		if (pfb->cbufs[i])
134			cbuf_cpp[i] = util_format_get_blocksize(pfb->cbufs[i]->format);
135		else
136			cbuf_cpp[i] = 4;
137	}
138
139	if (!memcmp(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp)) &&
140		!memcmp(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp)) &&
141		!memcmp(&gmem->scissor, scissor, sizeof(gmem->scissor))) {
142		/* everything is up-to-date */
143		return;
144	}
145
146	if (fd_mesa_debug & FD_DBG_NOSCIS) {
147		minx = 0;
148		miny = 0;
149		width = pfb->width;
150		height = pfb->height;
151	} else {
152		/* round down to multiple of alignment: */
153		minx = scissor->minx & ~(gmem_alignw - 1);
154		miny = scissor->miny & ~(gmem_alignh - 1);
155		width = scissor->maxx - minx;
156		height = scissor->maxy - miny;
157	}
158
159	bin_w = align(width, gmem_alignw);
160	bin_h = align(height, gmem_alignh);
161
162	/* first, find a bin width that satisfies the maximum width
163	 * restrictions:
164	 */
165	while (bin_w > max_width) {
166		nbins_x++;
167		bin_w = align(width / nbins_x, gmem_alignw);
168	}
169
170	if (fd_mesa_debug & FD_DBG_MSGS) {
171		debug_printf("binning input: cbuf cpp:");
172		for (i = 0; i < pfb->nr_cbufs; i++)
173			debug_printf(" %d", cbuf_cpp[i]);
174		debug_printf(", zsbuf cpp: %d; %dx%d\n",
175				zsbuf_cpp[0], width, height);
176	}
177
178	/* then find a bin width/height that satisfies the memory
179	 * constraints:
180	 */
181	while (total_size(cbuf_cpp, zsbuf_cpp, bin_w, bin_h, gmem) > gmem_size) {
182		if (bin_w > bin_h) {
183			nbins_x++;
184			bin_w = align(width / nbins_x, gmem_alignw);
185		} else {
186			nbins_y++;
187			bin_h = align(height / nbins_y, gmem_alignh);
188		}
189	}
190
191	DBG("using %d bins of size %dx%d", nbins_x*nbins_y, bin_w, bin_h);
192
193	gmem->scissor = *scissor;
194	memcpy(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp));
195	memcpy(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp));
196	gmem->bin_h = bin_h;
197	gmem->bin_w = bin_w;
198	gmem->nbins_x = nbins_x;
199	gmem->nbins_y = nbins_y;
200	gmem->minx = minx;
201	gmem->miny = miny;
202	gmem->width = width;
203	gmem->height = height;
204
205	/*
206	 * Assign tiles and pipes:
207	 *
208	 * At some point it might be worth playing with different
209	 * strategies and seeing if that makes much impact on
210	 * performance.
211	 */
212
213#define div_round_up(v, a)  (((v) + (a) - 1) / (a))
214	/* figure out number of tiles per pipe: */
215	tpp_x = tpp_y = 1;
216	while (div_round_up(nbins_y, tpp_y) > 8)
217		tpp_y += 2;
218	while ((div_round_up(nbins_y, tpp_y) *
219			div_round_up(nbins_x, tpp_x)) > 8)
220		tpp_x += 1;
221
222	/* configure pipes: */
223	xoff = yoff = 0;
224	for (i = 0; i < ARRAY_SIZE(ctx->pipe); i++) {
225		struct fd_vsc_pipe *pipe = &ctx->pipe[i];
226
227		if (xoff >= nbins_x) {
228			xoff = 0;
229			yoff += tpp_y;
230		}
231
232		if (yoff >= nbins_y) {
233			break;
234		}
235
236		pipe->x = xoff;
237		pipe->y = yoff;
238		pipe->w = MIN2(tpp_x, nbins_x - xoff);
239		pipe->h = MIN2(tpp_y, nbins_y - yoff);
240
241		xoff += tpp_x;
242	}
243
244	for (; i < ARRAY_SIZE(ctx->pipe); i++) {
245		struct fd_vsc_pipe *pipe = &ctx->pipe[i];
246		pipe->x = pipe->y = pipe->w = pipe->h = 0;
247	}
248
249#if 0 /* debug */
250	printf("%dx%d ... tpp=%dx%d\n", nbins_x, nbins_y, tpp_x, tpp_y);
251	for (i = 0; i < 8; i++) {
252		struct fd_vsc_pipe *pipe = &ctx->pipe[i];
253		printf("pipe[%d]: %ux%u @ %u,%u\n", i,
254				pipe->w, pipe->h, pipe->x, pipe->y);
255	}
256#endif
257
258	/* configure tiles: */
259	t = 0;
260	yoff = miny;
261	memset(tile_n, 0, sizeof(tile_n));
262	for (i = 0; i < nbins_y; i++) {
263		uint32_t bw, bh;
264
265		xoff = minx;
266
267		/* clip bin height: */
268		bh = MIN2(bin_h, miny + height - yoff);
269
270		for (j = 0; j < nbins_x; j++) {
271			struct fd_tile *tile = &ctx->tile[t];
272			uint32_t p;
273
274			assert(t < ARRAY_SIZE(ctx->tile));
275
276			/* pipe number: */
277			p = ((i / tpp_y) * div_round_up(nbins_x, tpp_x)) + (j / tpp_x);
278
279			/* clip bin width: */
280			bw = MIN2(bin_w, minx + width - xoff);
281
282			tile->n = tile_n[p]++;
283			tile->p = p;
284			tile->bin_w = bw;
285			tile->bin_h = bh;
286			tile->xoff = xoff;
287			tile->yoff = yoff;
288
289			t++;
290
291			xoff += bw;
292		}
293
294		yoff += bh;
295	}
296
297#if 0 /* debug */
298	t = 0;
299	for (i = 0; i < nbins_y; i++) {
300		for (j = 0; j < nbins_x; j++) {
301			struct fd_tile *tile = &ctx->tile[t++];
302			printf("|p:%u n:%u|", tile->p, tile->n);
303		}
304		printf("\n");
305	}
306#endif
307}
308
309static void
310render_tiles(struct fd_batch *batch)
311{
312	struct fd_context *ctx = batch->ctx;
313	struct fd_gmem_stateobj *gmem = &ctx->gmem;
314	int i;
315
316	ctx->emit_tile_init(batch);
317
318	if (batch->restore)
319		ctx->stats.batch_restore++;
320
321	for (i = 0; i < (gmem->nbins_x * gmem->nbins_y); i++) {
322		struct fd_tile *tile = &ctx->tile[i];
323
324		DBG("bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
325			tile->bin_h, tile->yoff, tile->bin_w, tile->xoff);
326
327		ctx->emit_tile_prep(batch, tile);
328
329		if (batch->restore) {
330			ctx->emit_tile_mem2gmem(batch, tile);
331		}
332
333		ctx->emit_tile_renderprep(batch, tile);
334
335		fd_hw_query_prepare_tile(batch, i, batch->gmem);
336
337		/* emit IB to drawcmds: */
338		ctx->emit_ib(batch->gmem, batch->draw);
339		fd_reset_wfi(batch);
340
341		/* emit gmem2mem to transfer tile back to system memory: */
342		ctx->emit_tile_gmem2mem(batch, tile);
343	}
344
345	if (ctx->emit_tile_fini)
346		ctx->emit_tile_fini(batch);
347}
348
349static void
350render_sysmem(struct fd_batch *batch)
351{
352	struct fd_context *ctx = batch->ctx;
353
354	ctx->emit_sysmem_prep(batch);
355
356	fd_hw_query_prepare_tile(batch, 0, batch->gmem);
357
358	/* emit IB to drawcmds: */
359	ctx->emit_ib(batch->gmem, batch->draw);
360	fd_reset_wfi(batch);
361
362	if (ctx->emit_sysmem_fini)
363		ctx->emit_sysmem_fini(batch);
364}
365
366static void
367flush_ring(struct fd_batch *batch)
368{
369	struct fd_context *ctx = batch->ctx;
370	int out_fence_fd = -1;
371
372	fd_ringbuffer_flush2(batch->gmem, batch->in_fence_fd,
373			batch->needs_out_fence_fd ? &out_fence_fd : NULL);
374
375	fd_fence_ref(&ctx->screen->base, &ctx->last_fence, NULL);
376	ctx->last_fence = fd_fence_create(ctx,
377			fd_ringbuffer_timestamp(batch->gmem), out_fence_fd);
378}
379
380void
381fd_gmem_render_tiles(struct fd_batch *batch)
382{
383	struct fd_context *ctx = batch->ctx;
384	struct pipe_framebuffer_state *pfb = &batch->framebuffer;
385	bool sysmem = false;
386
387	if (ctx->emit_sysmem_prep) {
388		if (batch->cleared || batch->gmem_reason || (batch->num_draws > 5)) {
389			DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u",
390				batch->cleared, batch->gmem_reason, batch->num_draws);
391		} else if (!(fd_mesa_debug & FD_DBG_NOBYPASS)) {
392			sysmem = true;
393		}
394	}
395
396	fd_reset_wfi(batch);
397
398	ctx->stats.batch_total++;
399
400	if (sysmem) {
401		DBG("%p: rendering sysmem %ux%u (%s/%s)",
402			batch, pfb->width, pfb->height,
403			util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
404			util_format_short_name(pipe_surface_format(pfb->zsbuf)));
405		fd_hw_query_prepare(batch, 1);
406		render_sysmem(batch);
407		ctx->stats.batch_sysmem++;
408	} else {
409		struct fd_gmem_stateobj *gmem = &ctx->gmem;
410		calculate_tiles(batch);
411		DBG("%p: rendering %dx%d tiles %ux%u (%s/%s)",
412			batch, pfb->width, pfb->height, gmem->nbins_x, gmem->nbins_y,
413			util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
414			util_format_short_name(pipe_surface_format(pfb->zsbuf)));
415		fd_hw_query_prepare(batch, gmem->nbins_x * gmem->nbins_y);
416		render_tiles(batch);
417		ctx->stats.batch_gmem++;
418	}
419
420	flush_ring(batch);
421}
422
423/* special case for when we need to create a fence but have no rendering
424 * to flush.. just emit a no-op string-marker packet.
425 */
426void
427fd_gmem_render_noop(struct fd_batch *batch)
428{
429	struct fd_context *ctx = batch->ctx;
430	struct pipe_context *pctx = &ctx->base;
431
432	pctx->emit_string_marker(pctx, "noop", 4);
433	/* emit IB to drawcmds (which contain the string marker): */
434	ctx->emit_ib(batch->gmem, batch->draw);
435	flush_ring(batch);
436}
437
438/* tile needs restore if it isn't completely contained within the
439 * cleared scissor:
440 */
441static bool
442skip_restore(struct pipe_scissor_state *scissor, struct fd_tile *tile)
443{
444	unsigned minx = tile->xoff;
445	unsigned maxx = tile->xoff + tile->bin_w;
446	unsigned miny = tile->yoff;
447	unsigned maxy = tile->yoff + tile->bin_h;
448	return (minx >= scissor->minx) && (maxx <= scissor->maxx) &&
449			(miny >= scissor->miny) && (maxy <= scissor->maxy);
450}
451
452/* When deciding whether a tile needs mem2gmem, we need to take into
453 * account the scissor rect(s) that were cleared.  To simplify we only
454 * consider the last scissor rect for each buffer, since the common
455 * case would be a single clear.
456 */
457bool
458fd_gmem_needs_restore(struct fd_batch *batch, struct fd_tile *tile,
459		uint32_t buffers)
460{
461	if (!(batch->restore & buffers))
462		return false;
463
464	/* if buffers partially cleared, then slow-path to figure out
465	 * if this particular tile needs restoring:
466	 */
467	if ((buffers & FD_BUFFER_COLOR) &&
468			(batch->partial_cleared & FD_BUFFER_COLOR) &&
469			skip_restore(&batch->cleared_scissor.color, tile))
470		return false;
471	if ((buffers & FD_BUFFER_DEPTH) &&
472			(batch->partial_cleared & FD_BUFFER_DEPTH) &&
473			skip_restore(&batch->cleared_scissor.depth, tile))
474		return false;
475	if ((buffers & FD_BUFFER_STENCIL) &&
476			(batch->partial_cleared & FD_BUFFER_STENCIL) &&
477			skip_restore(&batch->cleared_scissor.stencil, tile))
478		return false;
479
480	return true;
481}
482