1 /**************************************************************************
3 * Copyright 2012 Marek Olšák <maraeo@gmail.com>
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL THE AUTHORS AND/OR THEIR SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
28 #include "util/u_cpu_detect.h"
29 #include "util/u_helpers.h"
30 #include "util/u_inlines.h"
31 #include "util/u_upload_mgr.h"
32 #include "util/u_thread.h"
33 #include "util/os_time.h"
37 * This function is used to copy an array of pipe_vertex_buffer structures,
38 * while properly referencing the pipe_vertex_buffer::buffer member.
40 * enabled_buffers is updated such that the bits corresponding to the indices
41 * of disabled buffers are set to 0 and the enabled ones are set to 1.
43 * \sa util_copy_framebuffer_state
45 void util_set_vertex_buffers_mask(struct pipe_vertex_buffer
*dst
,
46 uint32_t *enabled_buffers
,
47 const struct pipe_vertex_buffer
*src
,
48 unsigned start_slot
, unsigned count
)
55 *enabled_buffers
&= ~u_bit_consecutive(start_slot
, count
);
58 for (i
= 0; i
< count
; i
++) {
59 if (src
[i
].buffer
.resource
)
62 pipe_vertex_buffer_unreference(&dst
[i
]);
64 if (!src
[i
].is_user_buffer
)
65 pipe_resource_reference(&dst
[i
].buffer
.resource
, src
[i
].buffer
.resource
);
68 /* Copy over the other members of pipe_vertex_buffer. */
69 memcpy(dst
, src
, count
* sizeof(struct pipe_vertex_buffer
));
71 *enabled_buffers
|= bitmask
<< start_slot
;
74 /* Unreference the buffers. */
75 for (i
= 0; i
< count
; i
++)
76 pipe_vertex_buffer_unreference(&dst
[i
]);
81 * Same as util_set_vertex_buffers_mask, but it only returns the number
84 void util_set_vertex_buffers_count(struct pipe_vertex_buffer
*dst
,
86 const struct pipe_vertex_buffer
*src
,
87 unsigned start_slot
, unsigned count
)
90 uint32_t enabled_buffers
= 0;
92 for (i
= 0; i
< *dst_count
; i
++) {
93 if (dst
[i
].buffer
.resource
)
94 enabled_buffers
|= (1ull << i
);
97 util_set_vertex_buffers_mask(dst
, &enabled_buffers
, src
, start_slot
,
100 *dst_count
= util_last_bit(enabled_buffers
);
104 * This function is used to copy an array of pipe_shader_buffer structures,
105 * while properly referencing the pipe_shader_buffer::buffer member.
107 * \sa util_set_vertex_buffer_mask
109 void util_set_shader_buffers_mask(struct pipe_shader_buffer
*dst
,
110 uint32_t *enabled_buffers
,
111 const struct pipe_shader_buffer
*src
,
112 unsigned start_slot
, unsigned count
)
119 for (i
= 0; i
< count
; i
++) {
120 pipe_resource_reference(&dst
[i
].buffer
, src
[i
].buffer
);
123 *enabled_buffers
|= (1ull << (start_slot
+ i
));
125 *enabled_buffers
&= ~(1ull << (start_slot
+ i
));
128 /* Copy over the other members of pipe_shader_buffer. */
129 memcpy(dst
, src
, count
* sizeof(struct pipe_shader_buffer
));
132 /* Unreference the buffers. */
133 for (i
= 0; i
< count
; i
++)
134 pipe_resource_reference(&dst
[i
].buffer
, NULL
);
136 *enabled_buffers
&= ~(((1ull << count
) - 1) << start_slot
);
141 * Given a user index buffer, save the structure to "saved", and upload it.
144 util_upload_index_buffer(struct pipe_context
*pipe
,
145 const struct pipe_draw_info
*info
,
146 struct pipe_resource
**out_buffer
,
147 unsigned *out_offset
, unsigned alignment
)
149 unsigned start_offset
= info
->start
* info
->index_size
;
151 u_upload_data(pipe
->stream_uploader
, start_offset
,
152 info
->count
* info
->index_size
, alignment
,
153 (char*)info
->index
.user
+ start_offset
,
154 out_offset
, out_buffer
);
155 u_upload_unmap(pipe
->stream_uploader
);
156 *out_offset
-= start_offset
;
157 return *out_buffer
!= NULL
;
161 * Called by MakeCurrent. Used to notify the driver that the application
162 * thread may have been changed.
164 * The function pins the current thread and driver threads to a group of
165 * CPU cores that share the same L3 cache. This is needed for good multi-
166 * threading performance on AMD Zen CPUs.
168 * \param upper_thread thread in the state tracker that also needs to be
172 util_pin_driver_threads_to_random_L3(struct pipe_context
*ctx
,
173 thrd_t
*upper_thread
)
175 /* If pinning has no effect, don't do anything. */
176 if (util_cpu_caps
.nr_cpus
== util_cpu_caps
.cores_per_L3
)
179 unsigned num_L3_caches
= util_cpu_caps
.nr_cpus
/
180 util_cpu_caps
.cores_per_L3
;
182 /* Get a semi-random number. */
183 int64_t t
= os_time_get_nano();
184 unsigned cache
= (t
^ (t
>> 8) ^ (t
>> 16)) % num_L3_caches
;
186 /* Tell the driver to pin its threads to the selected L3 cache. */
187 if (ctx
->set_context_param
) {
188 ctx
->set_context_param(ctx
, PIPE_CONTEXT_PARAM_PIN_THREADS_TO_L3_CACHE
,
192 /* Do the same for the upper level thread if there is any (e.g. glthread) */
194 util_pin_thread_to_L3(*upper_thread
, cache
, util_cpu_caps
.cores_per_L3
);
197 /* This is a helper for hardware bring-up. Don't remove. */
199 util_begin_pipestat_query(struct pipe_context
*ctx
)
201 struct pipe_query
*q
=
202 ctx
->create_query(ctx
, PIPE_QUERY_PIPELINE_STATISTICS
, 0);
206 ctx
->begin_query(ctx
, q
);
210 /* This is a helper for hardware bring-up. Don't remove. */
212 util_end_pipestat_query(struct pipe_context
*ctx
, struct pipe_query
*q
,
215 static unsigned counter
;
216 struct pipe_query_data_pipeline_statistics stats
;
218 ctx
->end_query(ctx
, q
);
219 ctx
->get_query_result(ctx
, q
, true, (void*)&stats
);
220 ctx
->destroy_query(ctx
, q
);
224 " ia_vertices = %"PRIu64
"\n"
225 " ia_primitives = %"PRIu64
"\n"
226 " vs_invocations = %"PRIu64
"\n"
227 " gs_invocations = %"PRIu64
"\n"
228 " gs_primitives = %"PRIu64
"\n"
229 " c_invocations = %"PRIu64
"\n"
230 " c_primitives = %"PRIu64
"\n"
231 " ps_invocations = %"PRIu64
"\n"
232 " hs_invocations = %"PRIu64
"\n"
233 " ds_invocations = %"PRIu64
"\n"
234 " cs_invocations = %"PRIu64
"\n",
235 (unsigned)p_atomic_inc_return(&counter
),
238 stats
.vs_invocations
,
239 stats
.gs_invocations
,
243 stats
.ps_invocations
,
244 stats
.hs_invocations
,
245 stats
.ds_invocations
,
246 stats
.cs_invocations
);
249 /* This is a helper for hardware bring-up. Don't remove. */
251 util_wait_for_idle(struct pipe_context
*ctx
)
253 struct pipe_fence_handle
*fence
= NULL
;
255 ctx
->flush(ctx
, &fence
, 0);
256 ctx
->screen
->fence_finish(ctx
->screen
, NULL
, fence
, PIPE_TIMEOUT_INFINITE
);
260 util_throttle_init(struct util_throttle
*t
, uint64_t max_mem_usage
)
262 t
->max_mem_usage
= max_mem_usage
;
266 util_throttle_deinit(struct pipe_screen
*screen
, struct util_throttle
*t
)
268 for (unsigned i
= 0; i
< ARRAY_SIZE(t
->ring
); i
++)
269 screen
->fence_reference(screen
, &t
->ring
[i
].fence
, NULL
);
273 util_get_throttle_total_memory_usage(struct util_throttle
*t
)
275 uint64_t total_usage
= 0;
277 for (unsigned i
= 0; i
< ARRAY_SIZE(t
->ring
); i
++)
278 total_usage
+= t
->ring
[i
].mem_usage
;
282 static void util_dump_throttle_ring(struct util_throttle
*t
)
284 printf("Throttle:\n");
285 for (unsigned i
= 0; i
< ARRAY_SIZE(t
->ring
); i
++) {
286 printf(" ring[%u]: fence = %s, mem_usage = %"PRIu64
"%s%s\n",
287 i
, t
->ring
[i
].fence
? "yes" : " no",
288 t
->ring
[i
].mem_usage
,
289 t
->flush_index
== i
? " [flush]" : "",
290 t
->wait_index
== i
? " [wait]" : "");
295 * Notify util_throttle that the next operation allocates memory.
296 * util_throttle tracks memory usage and waits for fences until its tracked
297 * memory usage decreases.
300 * util_throttle_memory_usage(..., w*h*d*Bpp);
301 * TexSubImage(..., w, h, d, ...);
303 * This means that TexSubImage can't allocate more memory its maximum limit
304 * set during initialization.
307 util_throttle_memory_usage(struct pipe_context
*pipe
,
308 struct util_throttle
*t
, uint64_t memory_size
)
310 (void)util_dump_throttle_ring
; /* silence warning */
312 if (!t
->max_mem_usage
)
315 struct pipe_screen
*screen
= pipe
->screen
;
316 struct pipe_fence_handle
**fence
= NULL
;
317 unsigned ring_size
= ARRAY_SIZE(t
->ring
);
318 uint64_t total
= util_get_throttle_total_memory_usage(t
);
320 /* If there is not enough memory, walk the list of fences and find
321 * the latest one that we need to wait for.
323 while (t
->wait_index
!= t
->flush_index
&&
324 total
&& total
+ memory_size
> t
->max_mem_usage
) {
325 assert(t
->ring
[t
->wait_index
].fence
);
327 /* Release an older fence if we need to wait for a newer one. */
329 screen
->fence_reference(screen
, fence
, NULL
);
331 fence
= &t
->ring
[t
->wait_index
].fence
;
332 t
->ring
[t
->wait_index
].mem_usage
= 0;
333 t
->wait_index
= (t
->wait_index
+ 1) % ring_size
;
335 total
= util_get_throttle_total_memory_usage(t
);
338 /* Wait for the fence to decrease memory usage. */
340 screen
->fence_finish(screen
, pipe
, *fence
, PIPE_TIMEOUT_INFINITE
);
341 screen
->fence_reference(screen
, fence
, NULL
);
344 /* Flush and get a fence if we've exhausted memory usage for the current
347 if (t
->ring
[t
->flush_index
].mem_usage
&&
348 t
->ring
[t
->flush_index
].mem_usage
+ memory_size
>
349 t
->max_mem_usage
/ (ring_size
/ 2)) {
350 struct pipe_fence_handle
**fence
=
351 &t
->ring
[t
->flush_index
].fence
;
353 /* Expect that the current flush slot doesn't have a fence yet. */
356 pipe
->flush(pipe
, fence
, PIPE_FLUSH_ASYNC
);
357 t
->flush_index
= (t
->flush_index
+ 1) % ring_size
;
359 /* Vacate the next slot if it's occupied. This should be rare. */
360 if (t
->flush_index
== t
->wait_index
) {
361 struct pipe_fence_handle
**fence
=
362 &t
->ring
[t
->wait_index
].fence
;
364 t
->ring
[t
->wait_index
].mem_usage
= 0;
365 t
->wait_index
= (t
->wait_index
+ 1) % ring_size
;
368 screen
->fence_finish(screen
, pipe
, *fence
, PIPE_TIMEOUT_INFINITE
);
369 screen
->fence_reference(screen
, fence
, NULL
);
372 assert(!t
->ring
[t
->flush_index
].mem_usage
);
373 assert(!t
->ring
[t
->flush_index
].fence
);
376 t
->ring
[t
->flush_index
].mem_usage
+= memory_size
;