• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2015 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice (including the next
13  * paragraph) shall be included in all copies or substantial portions of the
14  * Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #ifndef SI_QUERY_H
26 #define SI_QUERY_H
27 
28 #include "util/u_threaded_context.h"
29 
30 #include "ac_perfcounter.h"
31 
32 struct pipe_context;
33 struct pipe_query;
34 struct pipe_resource;
35 
36 struct si_screen;
37 struct si_context;
38 struct si_query;
39 struct si_query_buffer;
40 struct si_query_hw;
41 struct si_resource;
42 
43 #define SI_MAX_STREAMS 4
44 
45 enum
46 {
47    SI_QUERY_DRAW_CALLS = PIPE_QUERY_DRIVER_SPECIFIC,
48    SI_QUERY_DECOMPRESS_CALLS,
49    SI_QUERY_PRIM_RESTART_CALLS,
50    SI_QUERY_COMPUTE_CALLS,
51    SI_QUERY_CP_DMA_CALLS,
52    SI_QUERY_NUM_VS_FLUSHES,
53    SI_QUERY_NUM_PS_FLUSHES,
54    SI_QUERY_NUM_CS_FLUSHES,
55    SI_QUERY_NUM_CB_CACHE_FLUSHES,
56    SI_QUERY_NUM_DB_CACHE_FLUSHES,
57    SI_QUERY_NUM_L2_INVALIDATES,
58    SI_QUERY_NUM_L2_WRITEBACKS,
59    SI_QUERY_NUM_RESIDENT_HANDLES,
60    SI_QUERY_TC_OFFLOADED_SLOTS,
61    SI_QUERY_TC_DIRECT_SLOTS,
62    SI_QUERY_TC_NUM_SYNCS,
63    SI_QUERY_CS_THREAD_BUSY,
64    SI_QUERY_GALLIUM_THREAD_BUSY,
65    SI_QUERY_REQUESTED_VRAM,
66    SI_QUERY_REQUESTED_GTT,
67    SI_QUERY_MAPPED_VRAM,
68    SI_QUERY_MAPPED_GTT,
69    SI_QUERY_SLAB_WASTED_VRAM,
70    SI_QUERY_SLAB_WASTED_GTT,
71    SI_QUERY_BUFFER_WAIT_TIME,
72    SI_QUERY_NUM_MAPPED_BUFFERS,
73    SI_QUERY_NUM_GFX_IBS,
74    SI_QUERY_GFX_BO_LIST_SIZE,
75    SI_QUERY_GFX_IB_SIZE,
76    SI_QUERY_NUM_BYTES_MOVED,
77    SI_QUERY_NUM_EVICTIONS,
78    SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS,
79    SI_QUERY_VRAM_USAGE,
80    SI_QUERY_VRAM_VIS_USAGE,
81    SI_QUERY_GTT_USAGE,
82    SI_QUERY_GPU_TEMPERATURE,
83    SI_QUERY_CURRENT_GPU_SCLK,
84    SI_QUERY_CURRENT_GPU_MCLK,
85    SI_QUERY_GPU_LOAD,
86    SI_QUERY_GPU_SHADERS_BUSY,
87    SI_QUERY_GPU_TA_BUSY,
88    SI_QUERY_GPU_GDS_BUSY,
89    SI_QUERY_GPU_VGT_BUSY,
90    SI_QUERY_GPU_IA_BUSY,
91    SI_QUERY_GPU_SX_BUSY,
92    SI_QUERY_GPU_WD_BUSY,
93    SI_QUERY_GPU_BCI_BUSY,
94    SI_QUERY_GPU_SC_BUSY,
95    SI_QUERY_GPU_PA_BUSY,
96    SI_QUERY_GPU_DB_BUSY,
97    SI_QUERY_GPU_CP_BUSY,
98    SI_QUERY_GPU_CB_BUSY,
99    SI_QUERY_GPU_SDMA_BUSY,
100    SI_QUERY_GPU_PFP_BUSY,
101    SI_QUERY_GPU_MEQ_BUSY,
102    SI_QUERY_GPU_ME_BUSY,
103    SI_QUERY_GPU_SURF_SYNC_BUSY,
104    SI_QUERY_GPU_CP_DMA_BUSY,
105    SI_QUERY_GPU_SCRATCH_RAM_BUSY,
106    SI_QUERY_NUM_COMPILATIONS,
107    SI_QUERY_NUM_SHADERS_CREATED,
108    SI_QUERY_BACK_BUFFER_PS_DRAW_RATIO,
109    SI_QUERY_GPIN_ASIC_ID,
110    SI_QUERY_GPIN_NUM_SIMD,
111    SI_QUERY_GPIN_NUM_RB,
112    SI_QUERY_GPIN_NUM_SPI,
113    SI_QUERY_GPIN_NUM_SE,
114    SI_QUERY_LIVE_SHADER_CACHE_HITS,
115    SI_QUERY_LIVE_SHADER_CACHE_MISSES,
116    SI_QUERY_MEMORY_SHADER_CACHE_HITS,
117    SI_QUERY_MEMORY_SHADER_CACHE_MISSES,
118    SI_QUERY_DISK_SHADER_CACHE_HITS,
119    SI_QUERY_DISK_SHADER_CACHE_MISSES,
120 
121    SI_QUERY_FIRST_PERFCOUNTER = PIPE_QUERY_DRIVER_SPECIFIC + 100,
122 };
123 
124 enum
125 {
126    SI_QUERY_GROUP_GPIN = 0,
127    SI_NUM_SW_QUERY_GROUPS
128 };
129 
130 struct si_query_ops {
131    void (*destroy)(struct si_context *, struct si_query *);
132    bool (*begin)(struct si_context *, struct si_query *);
133    bool (*end)(struct si_context *, struct si_query *);
134    bool (*get_result)(struct si_context *, struct si_query *, bool wait,
135                       union pipe_query_result *result);
136    void (*get_result_resource)(struct si_context *, struct si_query *,
137                                enum pipe_query_flags flags,
138                                enum pipe_query_value_type result_type, int index,
139                                struct pipe_resource *resource, unsigned offset);
140 
141    void (*suspend)(struct si_context *, struct si_query *);
142    void (*resume)(struct si_context *, struct si_query *);
143 };
144 
145 struct si_query {
146    struct threaded_query b;
147    const struct si_query_ops *ops;
148 
149    /* The PIPE_QUERY_xxx type of query */
150    unsigned type;
151 
152    /* The number of dwords for suspend. */
153    unsigned num_cs_dw_suspend;
154 
155    /* Linked list of queries that must be suspended at end of CS. */
156    struct list_head active_list;
157 };
158 
159 enum
160 {
161    SI_QUERY_HW_FLAG_NO_START = (1 << 0),
162    /* gap */
163    /* whether begin_query doesn't clear the result */
164    SI_QUERY_HW_FLAG_BEGIN_RESUMES = (1 << 2),
165    /* whether GS invocations and emitted primitives counters are emulated
166     * using atomic adds.
167     */
168    SI_QUERY_EMULATE_GS_COUNTERS = (1 << 3),
169 };
170 
171 struct si_query_hw_ops {
172    bool (*prepare_buffer)(struct si_context *, struct si_query_buffer *);
173    void (*emit_start)(struct si_context *, struct si_query_hw *, struct si_resource *buffer,
174                       uint64_t va);
175    void (*emit_stop)(struct si_context *, struct si_query_hw *, struct si_resource *buffer,
176                      uint64_t va);
177    void (*clear_result)(struct si_query_hw *, union pipe_query_result *);
178    void (*add_result)(struct si_screen *screen, struct si_query_hw *, void *buffer,
179                       union pipe_query_result *result);
180 };
181 
182 struct si_query_buffer {
183    /* The buffer where query results are stored. */
184    struct si_resource *buf;
185    /* If a query buffer is full, a new buffer is created and the old one
186     * is put in here. When we calculate the result, we sum up the samples
187     * from all buffers. */
188    struct si_query_buffer *previous;
189    /* Offset of the next free result after current query data */
190    unsigned results_end;
191    bool unprepared;
192 };
193 
194 void si_query_buffer_destroy(struct si_screen *sctx, struct si_query_buffer *buffer);
195 void si_query_buffer_reset(struct si_context *sctx, struct si_query_buffer *buffer);
196 bool si_query_buffer_alloc(struct si_context *sctx, struct si_query_buffer *buffer,
197                            bool (*prepare_buffer)(struct si_context *, struct si_query_buffer *),
198                            unsigned size);
199 
200 struct si_query_hw {
201    struct si_query b;
202    struct si_query_hw_ops *ops;
203    unsigned flags;
204 
205    /* The query buffer and how many results are in it. */
206    struct si_query_buffer buffer;
207    /* Size of the result in memory for both begin_query and end_query,
208     * this can be one or two numbers, or it could even be a size of a structure. */
209    unsigned result_size;
210    union {
211       /* For transform feedback: which stream the query is for */
212       unsigned stream;
213       /* For pipeline stats: which counter is active */
214       unsigned index;
215    };
216 
217    /* Workaround via compute shader */
218    struct si_resource *workaround_buf;
219    unsigned workaround_offset;
220 };
221 
222 unsigned si_query_pipestat_end_dw_offset(struct si_screen *sscreen,
223                                          enum pipe_statistics_query_index index);
224 void si_query_hw_destroy(struct si_context *sctx, struct si_query *squery);
225 bool si_query_hw_begin(struct si_context *sctx, struct si_query *squery);
226 bool si_query_hw_end(struct si_context *sctx, struct si_query *squery);
227 bool si_query_hw_get_result(struct si_context *sctx, struct si_query *squery, bool wait,
228                             union pipe_query_result *result);
229 void si_query_hw_suspend(struct si_context *sctx, struct si_query *query);
230 void si_query_hw_resume(struct si_context *sctx, struct si_query *query);
231 
232 /* Shader-based queries */
233 
234 /**
235  * The query buffer is written to by ESGS NGG shaders with statistics about
236  * generated and (streamout-)emitted primitives.
237  *
238  * The context maintains a ring of these query buffers, and queries simply
239  * point into the ring, allowing an arbitrary number of queries to be active
240  * without additional GPU cost.
241  */
242 struct gfx10_sh_query_buffer {
243    struct list_head list;
244    struct si_resource *buf;
245    unsigned refcount;
246 
247    /* Offset into the buffer in bytes; points at the first un-emitted entry. */
248    unsigned head;
249 };
250 
251 /* Memory layout of the query buffer. Must be kept in sync with shaders
252  * (including QBO shaders) and should be aligned to cachelines.
253  *
254  * The somewhat awkward memory layout is for compatibility with the
255  * SET_PREDICATION packet, which also means that we're setting the high bit
256  * of all those values unconditionally.
257  */
258 struct gfx10_sh_query_buffer_mem {
259    struct {
260       uint64_t generated_primitives_start_dummy;
261       uint64_t emitted_primitives_start_dummy;
262       uint64_t generated_primitives;
263       uint64_t emitted_primitives;
264    } stream[4];
265    uint32_t fence; /* bottom-of-pipe fence: set to ~0 when draws have finished */
266    uint32_t pad[31];
267 };
268 
269 struct gfx10_sh_query {
270    struct si_query b;
271 
272    struct gfx10_sh_query_buffer *first;
273    struct gfx10_sh_query_buffer *last;
274    unsigned first_begin;
275    unsigned last_end;
276 
277    unsigned stream;
278 };
279 
280 struct pipe_query *gfx10_sh_query_create(struct si_screen *screen, enum pipe_query_type query_type,
281                                          unsigned index);
282 
283 /* Performance counters */
284 struct si_perfcounters {
285    struct ac_perfcounters base;
286 
287    unsigned num_stop_cs_dwords;
288    unsigned num_instance_cs_dwords;
289 };
290 
291 struct pipe_query *si_create_batch_query(struct pipe_context *ctx, unsigned num_queries,
292                                          unsigned *query_types);
293 
294 int si_get_perfcounter_info(struct si_screen *, unsigned index,
295                             struct pipe_driver_query_info *info);
296 int si_get_perfcounter_group_info(struct si_screen *, unsigned index,
297                                   struct pipe_driver_query_group_info *info);
298 
299 struct si_qbo_state {
300    struct pipe_constant_buffer saved_const0;
301 };
302 
303 #endif /* SI_QUERY_H */
304