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 }; 166 167 struct si_query_hw_ops { 168 bool (*prepare_buffer)(struct si_context *, struct si_query_buffer *); 169 void (*emit_start)(struct si_context *, struct si_query_hw *, struct si_resource *buffer, 170 uint64_t va); 171 void (*emit_stop)(struct si_context *, struct si_query_hw *, struct si_resource *buffer, 172 uint64_t va); 173 void (*clear_result)(struct si_query_hw *, union pipe_query_result *); 174 void (*add_result)(struct si_screen *screen, struct si_query_hw *, void *buffer, 175 union pipe_query_result *result); 176 }; 177 178 struct si_query_buffer { 179 /* The buffer where query results are stored. */ 180 struct si_resource *buf; 181 /* If a query buffer is full, a new buffer is created and the old one 182 * is put in here. When we calculate the result, we sum up the samples 183 * from all buffers. */ 184 struct si_query_buffer *previous; 185 /* Offset of the next free result after current query data */ 186 unsigned results_end; 187 bool unprepared; 188 }; 189 190 void si_query_buffer_destroy(struct si_screen *sctx, struct si_query_buffer *buffer); 191 void si_query_buffer_reset(struct si_context *sctx, struct si_query_buffer *buffer); 192 bool si_query_buffer_alloc(struct si_context *sctx, struct si_query_buffer *buffer, 193 bool (*prepare_buffer)(struct si_context *, struct si_query_buffer *), 194 unsigned size); 195 196 struct si_query_hw { 197 struct si_query b; 198 struct si_query_hw_ops *ops; 199 unsigned flags; 200 201 /* The query buffer and how many results are in it. */ 202 struct si_query_buffer buffer; 203 /* Size of the result in memory for both begin_query and end_query, 204 * this can be one or two numbers, or it could even be a size of a structure. */ 205 unsigned result_size; 206 /* For transform feedback: which stream the query is for */ 207 unsigned stream; 208 209 /* Workaround via compute shader */ 210 struct si_resource *workaround_buf; 211 unsigned workaround_offset; 212 }; 213 214 void si_query_hw_destroy(struct si_context *sctx, struct si_query *squery); 215 bool si_query_hw_begin(struct si_context *sctx, struct si_query *squery); 216 bool si_query_hw_end(struct si_context *sctx, struct si_query *squery); 217 bool si_query_hw_get_result(struct si_context *sctx, struct si_query *squery, bool wait, 218 union pipe_query_result *result); 219 void si_query_hw_suspend(struct si_context *sctx, struct si_query *query); 220 void si_query_hw_resume(struct si_context *sctx, struct si_query *query); 221 222 /* Shader-based queries */ 223 224 /** 225 * The query buffer is written to by ESGS NGG shaders with statistics about 226 * generated and (streamout-)emitted primitives. 227 * 228 * The context maintains a ring of these query buffers, and queries simply 229 * point into the ring, allowing an arbitrary number of queries to be active 230 * without additional GPU cost. 231 */ 232 struct gfx10_sh_query_buffer { 233 struct list_head list; 234 struct si_resource *buf; 235 unsigned refcount; 236 237 /* Offset into the buffer in bytes; points at the first un-emitted entry. */ 238 unsigned head; 239 }; 240 241 /* Memory layout of the query buffer. Must be kept in sync with shaders 242 * (including QBO shaders) and should be aligned to cachelines. 243 * 244 * The somewhat awkward memory layout is for compatibility with the 245 * SET_PREDICATION packet, which also means that we're setting the high bit 246 * of all those values unconditionally. 247 */ 248 struct gfx10_sh_query_buffer_mem { 249 struct { 250 uint64_t generated_primitives_start_dummy; 251 uint64_t emitted_primitives_start_dummy; 252 uint64_t generated_primitives; 253 uint64_t emitted_primitives; 254 } stream[4]; 255 uint32_t fence; /* bottom-of-pipe fence: set to ~0 when draws have finished */ 256 uint32_t pad[31]; 257 }; 258 259 struct gfx10_sh_query { 260 struct si_query b; 261 262 struct gfx10_sh_query_buffer *first; 263 struct gfx10_sh_query_buffer *last; 264 unsigned first_begin; 265 unsigned last_end; 266 267 unsigned stream; 268 }; 269 270 struct pipe_query *gfx10_sh_query_create(struct si_screen *screen, enum pipe_query_type query_type, 271 unsigned index); 272 273 /* Performance counters */ 274 struct si_perfcounters { 275 struct ac_perfcounters base; 276 277 unsigned num_stop_cs_dwords; 278 unsigned num_instance_cs_dwords; 279 }; 280 281 struct pipe_query *si_create_batch_query(struct pipe_context *ctx, unsigned num_queries, 282 unsigned *query_types); 283 284 int si_get_perfcounter_info(struct si_screen *, unsigned index, 285 struct pipe_driver_query_info *info); 286 int si_get_perfcounter_group_info(struct si_screen *, unsigned index, 287 struct pipe_driver_query_group_info *info); 288 289 struct si_qbo_state { 290 struct pipe_constant_buffer saved_const0; 291 }; 292 293 #endif /* SI_QUERY_H */ 294