• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2019 Collabora, Ltd.
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  * Authors:
24  *   Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
25  */
26 
27 #include "util/u_math.h"
28 #include "util/macros.h"
29 #include "pan_encoder.h"
30 
31 /* Mali GPUs are tiled-mode renderers, rather than immediate-mode.
32  * Conceptually, the screen is divided into 16x16 tiles. Vertex shaders run.
33  * Then, a fixed-function hardware block (the tiler) consumes the gl_Position
34  * results. For each triangle specified, it marks each containing tile as
35  * containing that triangle. This set of "triangles per tile" form the "polygon
36  * list". Finally, the rasterization unit consumes the polygon list to invoke
37  * the fragment shader.
38  *
39  * In practice, it's a bit more complicated than this. On Midgard chips with an
40  * "advanced tiling unit" (all except T720/T820/T830), 16x16 is the logical
41  * tile size, but Midgard features "hierarchical tiling", where power-of-two
42  * multiples of the base tile size can be used: hierarchy level 0 (16x16),
43  * level 1 (32x32), level 2 (64x64), per public information about Midgard's
44  * tiling. In fact, tiling goes up to 4096x4096 (!), although in practice
45  * 128x128 is the largest usually used (though higher modes are enabled).  The
46  * idea behind hierarchical tiling is to use low tiling levels for small
47  * triangles and high levels for large triangles, to minimize memory bandwidth
48  * and repeated fragment shader invocations (the former issue inherent to
49  * immediate-mode rendering and the latter common in traditional tilers).
50  *
51  * The tiler itself works by reading varyings in and writing a polygon list
52  * out. Unfortunately (for us), both of these buffers are managed in main
53  * memory; although they ideally will be cached, it is the drivers'
54  * responsibility to allocate these buffers. Varying buffer allocation is
55  * handled elsewhere, as it is not tiler specific; the real issue is allocating
56  * the polygon list.
57  *
58  * This is hard, because from the driver's perspective, we have no information
59  * about what geometry will actually look like on screen; that information is
60  * only gained from running the vertex shader. (Theoretically, we could run the
61  * vertex shaders in software as a prepass, or in hardware with transform
62  * feedback as a prepass, but either idea is ludicrous on so many levels).
63  *
64  * Instead, Mali uses a bit of a hybrid approach, splitting the polygon list
65  * into three distinct pieces. First, the driver statically determines which
66  * tile hierarchy levels to use (more on that later). At this point, we know the
67  * framebuffer dimensions and all the possible tilings of the framebuffer, so
68  * we know exactly how many tiles exist across all hierarchy levels. The first
69  * piece of the polygon list is the header, which is exactly 8 bytes per tile,
70  * plus padding and a small 64-byte prologue. (If that doesn't remind you of
71  * AFBC, it should. See pan_afbc.c for some fun parallels). The next part is
72  * the polygon list body, which seems to contain 512 bytes per tile, again
73  * across every level of the hierarchy. These two parts form the polygon list
74  * buffer. This buffer has a statically determinable size, approximately equal
75  * to the # of tiles across all hierarchy levels * (8 bytes + 512 bytes), plus
76  * alignment / minimum restrictions / etc.
77  *
78  * The third piece is the easy one (for us): the tiler heap. In essence, the
79  * tiler heap is a gigantic slab that's as big as could possibly be necessary
80  * in the worst case imaginable. Just... a gigantic allocation that we give a
81  * start and end pointer to. What's the catch? The tiler heap is lazily
82  * allocated; that is, a huge amount of memory is _reserved_, but only a tiny
83  * bit is actually allocated upfront. The GPU just keeps using the
84  * unallocated-but-reserved portions as it goes along, generating page faults
85  * if it goes beyond the allocation, and then the kernel is instructed to
86  * expand the allocation on page fault (known in the vendor kernel as growable
87  * memory). This is quite a bit of bookkeeping of its own, but that task is
88  * pushed to kernel space and we can mostly ignore it here, just remembering to
89  * set the GROWABLE flag so the kernel actually uses this path rather than
90  * allocating a gigantic amount up front and burning a hole in RAM.
91  *
92  * As far as determining which hierarchy levels to use, the simple answer is
93  * that right now, we don't. In the tiler configuration fields (consistent from
94  * the earliest Midgard's SFBD through the latest Bifrost traces we have),
95  * there is a hierarchy_mask field, controlling which levels (tile sizes) are
96  * enabled. Ideally, the hierarchical tiling dream -- mapping big polygons to
97  * big tiles and small polygons to small tiles -- would be realized here as
98  * well. As long as there are polygons at all needing tiling, we always have to
99  * have big tiles available, in case there are big polygons. But we don't
100  * necessarily need small tiles available. Ideally, when there are small
101  * polygons, small tiles are enabled (to avoid waste from putting small
102  * triangles in the big tiles); when there are not, small tiles are disabled to
103  * avoid enabling more levels than necessary, which potentially costs in memory
104  * bandwidth / power / tiler performance.
105  *
106  * Of course, the driver has to figure this out statically. When tile
107  * hiearchies are actually established, this occurs by the tiler in
108  * fixed-function hardware, after the vertex shaders have run and there is
109  * sufficient information to figure out the size of triangles. The driver has
110  * no such luxury, again barring insane hacks like additionally running the
111  * vertex shaders in software or in hardware via transform feedback. Thus, for
112  * the driver, we need a heuristic approach.
113  *
114  * There are lots of heuristics to guess triangle size statically you could
115  * imagine, but one approach shines as particularly simple-stupid: assume all
116  * on-screen triangles are equal size and spread equidistantly throughout the
117  * screen. Let's be clear, this is NOT A VALID ASSUMPTION. But if we roll with
118  * it, then we see:
119  *
120  *      Triangle Area   = (Screen Area / # of triangles)
121  *                      = (Width * Height) / (# of triangles)
122  *
123  * Or if you prefer, we can also make a third CRAZY assumption that we only draw
124  * right triangles with edges parallel/perpendicular to the sides of the screen
125  * with no overdraw, forming a triangle grid across the screen:
126  *
127  * |--w--|
128  *  _____   |
129  * | /| /|  |
130  * |/_|/_|  h
131  * | /| /|  |
132  * |/_|/_|  |
133  *
134  * Then you can use some middle school geometry and algebra to work out the
135  * triangle dimensions. I started working on this, but realised I didn't need
136  * to to make my point, but couldn't bare to erase that ASCII art. Anyway.
137  *
138  * POINT IS, by considering the ratio of screen area and triangle count, we can
139  * estimate the triangle size. For a small size, use small bins; for a large
140  * size, use large bins. Intuitively, this metric makes sense: when there are
141  * few triangles on a large screen, you're probably compositing a UI and
142  * therefore the triangles are large; when there are a lot of triangles on a
143  * small screen, you're probably rendering a 3D mesh and therefore the
144  * triangles are tiny. (Or better said -- there will be tiny triangles, even if
145  * there are also large triangles. There have to be unless you expect crazy
146  * overdraw. Generally, it's better to allow more small bin sizes than
147  * necessary than not allow enough.)
148  *
149  * From this heuristic (or whatever), we determine the minimum allowable tile
150  * size, and we use that to decide the hierarchy masking, selecting from the
151  * minimum "ideal" tile size to the maximum tile size (2048x2048 in practice).
152  *
153  * Once we have that mask and the framebuffer dimensions, we can compute the
154  * size of the statically-sized polygon list structures, allocate them, and go!
155  *
156  * -----
157  *
158  * On T720, T820, and T830, there is no support for hierarchical tiling.
159  * Instead, the hardware allows the driver to select the tile size dynamically
160  * on a per-framebuffer basis, including allowing rectangular/non-square tiles.
161  * Rules for tile size selection are as follows:
162  *
163  *  - Dimensions must be powers-of-two.
164  *  - The smallest tile is 16x16.
165  *  - The tile width/height is at most the framebuffer w/h (clamp up to 16 pix)
166  *  - There must be no more than 64 tiles in either dimension.
167  *
168  * Within these constraints, the driver is free to pick a tile size according
169  * to some heuristic, similar to units with an advanced tiling unit.
170  *
171  * To pick a size without any heuristics, we may satisfy the constraints by
172  * defaulting to 16x16 (a power-of-two). This fits the minimum. For the size
173  * constraint, consider:
174  *
175  *      # of tiles < 64
176  *      ceil (fb / tile) < 64
177  *      (fb / tile) <= (64 - 1)
178  *      tile <= fb / (64 - 1) <= next_power_of_two(fb / (64 - 1))
179  *
180  * Hence we clamp up to align_pot(fb / (64 - 1)).
181 
182  * Extending to use a selection heuristic left for future work.
183  *
184  * Once the tile size (w, h) is chosen, we compute the hierarchy "mask":
185  *
186  *      hierarchy_mask = (log2(h / 16) << 6) | log2(w / 16)
187  *
188  * Of course with no hierarchical tiling, this is not a mask; it's just a field
189  * specifying the tile size. But I digress.
190  *
191  * We also compute the polgon list sizes (with framebuffer size W, H) as:
192  *
193  *      full_size = 0x200 + 0x200 * ceil(W / w) * ceil(H / h)
194  *      offset = 8 * ceil(W / w) * ceil(H / h)
195  *
196  * It further appears necessary to round down offset to the nearest 0x200.
197  * Possibly we would also round down full_size to the nearest 0x200 but
198  * full_size/0x200 = (1 + ceil(W / w) * ceil(H / h)) is an integer so there's
199  * nothing to do.
200  */
201 
202 /* Hierarchical tiling spans from 16x16 to 4096x4096 tiles */
203 
204 #define MIN_TILE_SIZE 16
205 #define MAX_TILE_SIZE 4096
206 
207 /* Constants as shifts for easier power-of-two iteration */
208 
209 #define MIN_TILE_SHIFT util_logbase2(MIN_TILE_SIZE)
210 #define MAX_TILE_SHIFT util_logbase2(MAX_TILE_SIZE)
211 
212 /* The hierarchy has a 64-byte prologue */
213 #define PROLOGUE_SIZE 0x40
214 
215 /* For each tile (across all hierarchy levels), there is 8 bytes of header */
216 #define HEADER_BYTES_PER_TILE 0x8
217 
218 /* Likewise, each tile per level has 512 bytes of body */
219 #define FULL_BYTES_PER_TILE 0x200
220 
221 /* If the width-x-height framebuffer is divided into tile_size-x-tile_size
222  * tiles, how many tiles are there? Rounding up in each direction. For the
223  * special case of tile_size=16, this aligns with the usual Midgard count.
224  * tile_size must be a power-of-two. Not really repeat code from AFBC/checksum,
225  * because those care about the stride (not just the overall count) and only at
226  * a a fixed-tile size (not any of a number of power-of-twos) */
227 
228 static unsigned
pan_tile_count(unsigned width,unsigned height,unsigned tile_width,unsigned tile_height)229 pan_tile_count(unsigned width, unsigned height, unsigned tile_width, unsigned tile_height)
230 {
231         unsigned aligned_width = ALIGN_POT(width, tile_width);
232         unsigned aligned_height = ALIGN_POT(height, tile_height);
233 
234         unsigned tile_count_x = aligned_width / tile_width;
235         unsigned tile_count_y = aligned_height / tile_height;
236 
237         return tile_count_x * tile_count_y;
238 }
239 
240 /* For `masked_count` of the smallest tile sizes masked out, computes how the
241  * size of the polygon list header. We iterate the tile sizes (16x16 through
242  * 2048x2048). For each tile size, we figure out how many tiles there are at
243  * this hierarchy level and therefore many bytes this level is, leaving us with
244  * a byte count for each level. We then just sum up the byte counts across the
245  * levels to find a byte count for all levels. */
246 
247 static unsigned
panfrost_hierarchy_size(unsigned width,unsigned height,unsigned mask,unsigned bytes_per_tile)248 panfrost_hierarchy_size(
249                 unsigned width,
250                 unsigned height,
251                 unsigned mask,
252                 unsigned bytes_per_tile)
253 {
254         unsigned size = PROLOGUE_SIZE;
255 
256         /* Iterate hierarchy levels */
257 
258         for (unsigned b = 0; b < (MAX_TILE_SHIFT - MIN_TILE_SHIFT); ++b) {
259                 /* Check if this level is enabled */
260                 if (!(mask & (1 << b)))
261                         continue;
262 
263                 /* Shift from a level to a tile size */
264                 unsigned tile_size = (1 << b) * MIN_TILE_SIZE;
265 
266                 unsigned tile_count = pan_tile_count(width, height, tile_size, tile_size);
267                 unsigned level_count = bytes_per_tile * tile_count;
268 
269                 size += level_count;
270         }
271 
272         /* This size will be used as an offset, so ensure it's aligned */
273         return ALIGN_POT(size, 0x200);
274 }
275 
276 /* Implement the formula:
277  *
278  *      0x200 + bytes_per_tile * ceil(W / w) * ceil(H / h)
279  *
280  * rounding down the answer to the nearest 0x200. This is used to compute both
281  * header and body sizes for GPUs without hierarchical tiling. Essentially,
282  * computing a single hierarchy level, since there isn't any hierarchy!
283  */
284 
285 static unsigned
panfrost_flat_size(unsigned width,unsigned height,unsigned dim,unsigned bytes_per_tile)286 panfrost_flat_size(unsigned width, unsigned height, unsigned dim, unsigned bytes_per_tile)
287 {
288         /* First, extract the tile dimensions */
289 
290         unsigned tw = (1 << (dim & 0b111)) * 8;
291         unsigned th = (1 << ((dim & (0b111 << 6)) >> 6)) * 8;
292 
293         /* tile_count is ceil(W/w) * ceil(H/h) */
294         unsigned raw = pan_tile_count(width, height, tw, th) * bytes_per_tile;
295 
296         /* Round down and add offset */
297         return 0x200 + ((raw / 0x200) * 0x200);
298 }
299 
300 /* Given a hierarchy mask and a framebuffer size, compute the header size */
301 
302 unsigned
panfrost_tiler_header_size(unsigned width,unsigned height,unsigned mask,bool hierarchy)303 panfrost_tiler_header_size(unsigned width, unsigned height, unsigned mask, bool hierarchy)
304 {
305         if (hierarchy)
306                 return panfrost_hierarchy_size(width, height, mask, HEADER_BYTES_PER_TILE);
307         else
308                 return panfrost_flat_size(width, height, mask, HEADER_BYTES_PER_TILE);
309 }
310 
311 /* The combined header/body is sized similarly (but it is significantly
312  * larger), except that it can be empty when the tiler disabled, rather than
313  * getting clamped to a minimum size.
314  */
315 
316 unsigned
panfrost_tiler_full_size(unsigned width,unsigned height,unsigned mask,bool hierarchy)317 panfrost_tiler_full_size(unsigned width, unsigned height, unsigned mask, bool hierarchy)
318 {
319         if (hierarchy)
320                 return panfrost_hierarchy_size(width, height, mask, FULL_BYTES_PER_TILE);
321         else
322                 return panfrost_flat_size(width, height, mask, FULL_BYTES_PER_TILE);
323 }
324 
325 /* On GPUs without hierarchical tiling, we choose a tile size directly and
326  * stuff it into the field otherwise known as hierarchy mask (not a mask). */
327 
328 static unsigned
panfrost_choose_tile_size(unsigned width,unsigned height,unsigned vertex_count)329 panfrost_choose_tile_size(
330         unsigned width, unsigned height, unsigned vertex_count)
331 {
332         /* Figure out the ideal tile size. Eventually a heuristic should be
333          * used for this */
334 
335         unsigned best_w = 16;
336         unsigned best_h = 16;
337 
338         /* Clamp so there are less than 64 tiles in each direction */
339 
340         best_w = MAX2(best_w, util_next_power_of_two(width / 63));
341         best_h = MAX2(best_h, util_next_power_of_two(height / 63));
342 
343         /* We have our ideal tile size, so encode */
344 
345         unsigned exp_w = util_logbase2(best_w / 16);
346         unsigned exp_h = util_logbase2(best_h / 16);
347 
348         return exp_w | (exp_h << 6);
349 }
350 
351 /* In the future, a heuristic to choose a tiler hierarchy mask would go here.
352  * At the moment, we just default to 0xFF, which enables all possible hierarchy
353  * levels. Overall this yields good performance but presumably incurs a cost in
354  * memory bandwidth / power consumption / etc, at least on smaller scenes that
355  * don't really need all the smaller levels enabled */
356 
357 unsigned
panfrost_choose_hierarchy_mask(unsigned width,unsigned height,unsigned vertex_count,bool hierarchy)358 panfrost_choose_hierarchy_mask(
359         unsigned width, unsigned height,
360         unsigned vertex_count, bool hierarchy)
361 {
362         /* If there is no geometry, we don't bother enabling anything */
363 
364         if (!vertex_count)
365                 return 0x00;
366 
367         if (!hierarchy)
368                 return panfrost_choose_tile_size(width, height, vertex_count);
369 
370         /* Otherwise, default everything on. TODO: Proper tests */
371 
372         return 0xFF;
373 }
374