src/gallium/drivers/radeonsi/si_query.h

   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 struct pipe_context;
  31 struct pipe_query;
  32 struct pipe_resource;
  33
  34 struct si_screen;
  35 struct si_context;
  36 struct si_query;
  37 struct si_query_hw;
  38 struct r600_resource;
  39
  40 enum {
  41         SI_QUERY_DRAW_CALLS = PIPE_QUERY_DRIVER_SPECIFIC,
  42         SI_QUERY_DECOMPRESS_CALLS,
  43         SI_QUERY_MRT_DRAW_CALLS,
  44         SI_QUERY_PRIM_RESTART_CALLS,
  45         SI_QUERY_SPILL_DRAW_CALLS,
  46         SI_QUERY_COMPUTE_CALLS,
  47         SI_QUERY_SPILL_COMPUTE_CALLS,
  48         SI_QUERY_DMA_CALLS,
  49         SI_QUERY_CP_DMA_CALLS,
  50         SI_QUERY_NUM_VS_FLUSHES,
  51         SI_QUERY_NUM_PS_FLUSHES,
  52         SI_QUERY_NUM_CS_FLUSHES,
  53         SI_QUERY_NUM_CB_CACHE_FLUSHES,
  54         SI_QUERY_NUM_DB_CACHE_FLUSHES,
  55         SI_QUERY_NUM_L2_INVALIDATES,
  56         SI_QUERY_NUM_L2_WRITEBACKS,
  57         SI_QUERY_NUM_RESIDENT_HANDLES,
  58         SI_QUERY_TC_OFFLOADED_SLOTS,
  59         SI_QUERY_TC_DIRECT_SLOTS,
  60         SI_QUERY_TC_NUM_SYNCS,
  61         SI_QUERY_CS_THREAD_BUSY,
  62         SI_QUERY_GALLIUM_THREAD_BUSY,
  63         SI_QUERY_REQUESTED_VRAM,
  64         SI_QUERY_REQUESTED_GTT,
  65         SI_QUERY_MAPPED_VRAM,
  66         SI_QUERY_MAPPED_GTT,
  67         SI_QUERY_BUFFER_WAIT_TIME,
  68         SI_QUERY_NUM_MAPPED_BUFFERS,
  69         SI_QUERY_NUM_GFX_IBS,
  70         SI_QUERY_NUM_SDMA_IBS,
  71         SI_QUERY_GFX_BO_LIST_SIZE,
  72         SI_QUERY_GFX_IB_SIZE,
  73         SI_QUERY_NUM_BYTES_MOVED,
  74         SI_QUERY_NUM_EVICTIONS,
  75         SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS,
  76         SI_QUERY_VRAM_USAGE,
  77         SI_QUERY_VRAM_VIS_USAGE,
  78         SI_QUERY_GTT_USAGE,
  79         SI_QUERY_GPU_TEMPERATURE,
  80         SI_QUERY_CURRENT_GPU_SCLK,
  81         SI_QUERY_CURRENT_GPU_MCLK,
  82         SI_QUERY_GPU_LOAD,
  83         SI_QUERY_GPU_SHADERS_BUSY,
  84         SI_QUERY_GPU_TA_BUSY,
  85         SI_QUERY_GPU_GDS_BUSY,
  86         SI_QUERY_GPU_VGT_BUSY,
  87         SI_QUERY_GPU_IA_BUSY,
  88         SI_QUERY_GPU_SX_BUSY,
  89         SI_QUERY_GPU_WD_BUSY,
  90         SI_QUERY_GPU_BCI_BUSY,
  91         SI_QUERY_GPU_SC_BUSY,
  92         SI_QUERY_GPU_PA_BUSY,
  93         SI_QUERY_GPU_DB_BUSY,
  94         SI_QUERY_GPU_CP_BUSY,
  95         SI_QUERY_GPU_CB_BUSY,
  96         SI_QUERY_GPU_SDMA_BUSY,
  97         SI_QUERY_GPU_PFP_BUSY,
  98         SI_QUERY_GPU_MEQ_BUSY,
  99         SI_QUERY_GPU_ME_BUSY,
 100         SI_QUERY_GPU_SURF_SYNC_BUSY,
 101         SI_QUERY_GPU_CP_DMA_BUSY,
 102         SI_QUERY_GPU_SCRATCH_RAM_BUSY,
 103         SI_QUERY_NUM_COMPILATIONS,
 104         SI_QUERY_NUM_SHADERS_CREATED,
 105         SI_QUERY_BACK_BUFFER_PS_DRAW_RATIO,
 106         SI_QUERY_NUM_SHADER_CACHE_HITS,
 107         SI_QUERY_GPIN_ASIC_ID,
 108         SI_QUERY_GPIN_NUM_SIMD,
 109         SI_QUERY_GPIN_NUM_RB,
 110         SI_QUERY_GPIN_NUM_SPI,
 111         SI_QUERY_GPIN_NUM_SE,
 112
 113         SI_QUERY_FIRST_PERFCOUNTER = PIPE_QUERY_DRIVER_SPECIFIC + 100,
 114 };
 115
 116 enum {
 117         SI_QUERY_GROUP_GPIN = 0,
 118         SI_NUM_SW_QUERY_GROUPS
 119 };
 120
 121 struct si_query_ops {
 122         void (*destroy)(struct si_screen *, struct si_query *);
 123         bool (*begin)(struct si_context *, struct si_query *);
 124         bool (*end)(struct si_context *, struct si_query *);
 125         bool (*get_result)(struct si_context *,
 126                            struct si_query *, bool wait,
 127                            union pipe_query_result *result);
 128         void (*get_result_resource)(struct si_context *,
 129                                     struct si_query *, bool wait,
 130                                     enum pipe_query_value_type result_type,
 131                                     int index,
 132                                     struct pipe_resource *resource,
 133                                     unsigned offset);
 134 };
 135
 136 struct si_query {
 137         struct threaded_query b;
 138         struct si_query_ops *ops;
 139
 140         /* The type of query */
 141         unsigned type;
 142 };
 143
 144 enum {
 145         SI_QUERY_HW_FLAG_NO_START = (1 << 0),
 146         /* gap */
 147         /* whether begin_query doesn't clear the result */
 148         SI_QUERY_HW_FLAG_BEGIN_RESUMES = (1 << 2),
 149 };
 150
 151 struct si_query_hw_ops {
 152         bool (*prepare_buffer)(struct si_screen *,
 153                                struct si_query_hw *,
 154                                struct r600_resource *);
 155         void (*emit_start)(struct si_context *,
 156                            struct si_query_hw *,
 157                            struct r600_resource *buffer, uint64_t va);
 158         void (*emit_stop)(struct si_context *,
 159                           struct si_query_hw *,
 160                           struct r600_resource *buffer, uint64_t va);
 161         void (*clear_result)(struct si_query_hw *, union pipe_query_result *);
 162         void (*add_result)(struct si_screen *screen,
 163                            struct si_query_hw *, void *buffer,
 164                            union pipe_query_result *result);
 165 };
 166
 167 struct si_query_buffer {
 168         /* The buffer where query results are stored. */
 169         struct r600_resource            *buf;
 170         /* Offset of the next free result after current query data */
 171         unsigned                        results_end;
 172         /* If a query buffer is full, a new buffer is created and the old one
 173          * is put in here. When we calculate the result, we sum up the samples
 174          * from all buffers. */
 175         struct si_query_buffer  *previous;
 176 };
 177
 178 struct si_query_hw {
 179         struct si_query b;
 180         struct si_query_hw_ops *ops;
 181         unsigned flags;
 182
 183         /* The query buffer and how many results are in it. */
 184         struct si_query_buffer buffer;
 185         /* Size of the result in memory for both begin_query and end_query,
 186          * this can be one or two numbers, or it could even be a size of a structure. */
 187         unsigned result_size;
 188         /* The number of dwords for end_query. */
 189         unsigned num_cs_dw_end;
 190         /* Linked list of queries */
 191         struct list_head list;
 192         /* For transform feedback: which stream the query is for */
 193         unsigned stream;
 194
 195         /* Workaround via compute shader */
 196         struct r600_resource *workaround_buf;
 197         unsigned workaround_offset;
 198 };
 199
 200 bool si_query_hw_init(struct si_screen *sscreen,
 201                       struct si_query_hw *query);
 202 void si_query_hw_destroy(struct si_screen *sscreen,
 203                          struct si_query *rquery);
 204 bool si_query_hw_begin(struct si_context *sctx,
 205                        struct si_query *rquery);
 206 bool si_query_hw_end(struct si_context *sctx,
 207                      struct si_query *rquery);
 208 bool si_query_hw_get_result(struct si_context *sctx,
 209                             struct si_query *rquery,
 210                             bool wait,
 211                             union pipe_query_result *result);
 212
 213 /* Performance counters */
 214 enum {
 215         /* This block is part of the shader engine */
 216         SI_PC_BLOCK_SE = (1 << 0),
 217
 218         /* Expose per-instance groups instead of summing all instances (within
 219          * an SE). */
 220         SI_PC_BLOCK_INSTANCE_GROUPS = (1 << 1),
 221
 222         /* Expose per-SE groups instead of summing instances across SEs. */
 223         SI_PC_BLOCK_SE_GROUPS = (1 << 2),
 224
 225         /* Shader block */
 226         SI_PC_BLOCK_SHADER = (1 << 3),
 227
 228         /* Non-shader block with perfcounters windowed by shaders. */
 229         SI_PC_BLOCK_SHADER_WINDOWED = (1 << 4),
 230 };
 231
 232 /* Describes a hardware block with performance counters. Multiple instances of
 233  * each block, possibly per-SE, may exist on the chip. Depending on the block
 234  * and on the user's configuration, we either
 235  *  (a) expose every instance as a performance counter group,
 236  *  (b) expose a single performance counter group that reports the sum over all
 237  *      instances, or
 238  *  (c) expose one performance counter group per instance, but summed over all
 239  *      shader engines.
 240  */
 241 struct si_perfcounter_block {
 242         const char *basename;
 243         unsigned flags;
 244         unsigned num_counters;
 245         unsigned num_selectors;
 246         unsigned num_instances;
 247
 248         unsigned num_groups;
 249         char *group_names;
 250         unsigned group_name_stride;
 251
 252         char *selector_names;
 253         unsigned selector_name_stride;
 254
 255         void *data;
 256 };
 257
 258 struct si_perfcounters {
 259         unsigned num_groups;
 260         unsigned num_blocks;
 261         struct si_perfcounter_block *blocks;
 262
 263         unsigned num_stop_cs_dwords;
 264         unsigned num_instance_cs_dwords;
 265
 266         unsigned num_shader_types;
 267         const char * const *shader_type_suffixes;
 268         const unsigned *shader_type_bits;
 269
 270         void (*emit_instance)(struct si_context *,
 271                               int se, int instance);
 272         void (*emit_shaders)(struct si_context *, unsigned shaders);
 273         void (*emit_select)(struct si_context *,
 274                             struct si_perfcounter_block *,
 275                             unsigned count, unsigned *selectors);
 276         void (*emit_start)(struct si_context *,
 277                           struct r600_resource *buffer, uint64_t va);
 278         void (*emit_stop)(struct si_context *,
 279                           struct r600_resource *buffer, uint64_t va);
 280         void (*emit_read)(struct si_context *,
 281                           struct si_perfcounter_block *,
 282                           unsigned count, unsigned *selectors,
 283                           struct r600_resource *buffer, uint64_t va);
 284
 285         void (*cleanup)(struct si_screen *);
 286
 287         bool separate_se;
 288         bool separate_instance;
 289 };
 290
 291 struct pipe_query *si_create_batch_query(struct pipe_context *ctx,
 292                                          unsigned num_queries,
 293                                          unsigned *query_types);
 294
 295 int si_get_perfcounter_info(struct si_screen *,
 296                             unsigned index,
 297                             struct pipe_driver_query_info *info);
 298 int si_get_perfcounter_group_info(struct si_screen *,
 299                                   unsigned index,
 300                                   struct pipe_driver_query_group_info *info);
 301
 302 bool si_perfcounters_init(struct si_perfcounters *, unsigned num_blocks);
 303 void si_perfcounters_add_block(struct si_screen *,
 304                                struct si_perfcounters *,
 305                                const char *name, unsigned flags,
 306                                unsigned counters, unsigned selectors,
 307                                unsigned instances, void *data);
 308 void si_perfcounters_do_destroy(struct si_perfcounters *);
 309 void si_query_hw_reset_buffers(struct si_context *sctx,
 310                                struct si_query_hw *query);
 311
 312 struct si_qbo_state {
 313         void *saved_compute;
 314         struct pipe_constant_buffer saved_const0;
 315         struct pipe_shader_buffer saved_ssbo[3];
 316 };
 317
 318 #endif /* SI_QUERY_H */