1 /* 2 * Copyright 2015 Advanced Micro Devices, Inc. 3 * 4 * SPDX-License-Identifier: MIT 5 */ 6 7 #ifndef SI_QUERY_H 8 #define SI_QUERY_H 9 10 #include "util/u_threaded_context.h" 11 12 #include "ac_perfcounter.h" 13 14 struct pipe_context; 15 struct pipe_query; 16 struct pipe_resource; 17 18 struct si_screen; 19 struct si_context; 20 struct si_query; 21 struct si_query_buffer; 22 struct si_query_hw; 23 struct si_resource; 24 25 #define SI_MAX_STREAMS 4 26 27 enum 28 { 29 SI_QUERY_DRAW_CALLS = PIPE_QUERY_DRIVER_SPECIFIC, 30 SI_QUERY_DECOMPRESS_CALLS, 31 SI_QUERY_COMPUTE_CALLS, 32 SI_QUERY_CP_DMA_CALLS, 33 SI_QUERY_NUM_VS_FLUSHES, 34 SI_QUERY_NUM_PS_FLUSHES, 35 SI_QUERY_NUM_CS_FLUSHES, 36 SI_QUERY_NUM_CB_CACHE_FLUSHES, 37 SI_QUERY_NUM_DB_CACHE_FLUSHES, 38 SI_QUERY_NUM_L2_INVALIDATES, 39 SI_QUERY_NUM_L2_WRITEBACKS, 40 SI_QUERY_NUM_RESIDENT_HANDLES, 41 SI_QUERY_TC_OFFLOADED_SLOTS, 42 SI_QUERY_TC_DIRECT_SLOTS, 43 SI_QUERY_TC_NUM_SYNCS, 44 SI_QUERY_CS_THREAD_BUSY, 45 SI_QUERY_GALLIUM_THREAD_BUSY, 46 SI_QUERY_REQUESTED_VRAM, 47 SI_QUERY_REQUESTED_GTT, 48 SI_QUERY_MAPPED_VRAM, 49 SI_QUERY_MAPPED_GTT, 50 SI_QUERY_SLAB_WASTED_VRAM, 51 SI_QUERY_SLAB_WASTED_GTT, 52 SI_QUERY_BUFFER_WAIT_TIME, 53 SI_QUERY_NUM_MAPPED_BUFFERS, 54 SI_QUERY_NUM_GFX_IBS, 55 SI_QUERY_GFX_BO_LIST_SIZE, 56 SI_QUERY_GFX_IB_SIZE, 57 SI_QUERY_NUM_BYTES_MOVED, 58 SI_QUERY_NUM_EVICTIONS, 59 SI_QUERY_NUM_VRAM_CPU_PAGE_FAULTS, 60 SI_QUERY_VRAM_USAGE, 61 SI_QUERY_VRAM_VIS_USAGE, 62 SI_QUERY_GTT_USAGE, 63 SI_QUERY_GPU_TEMPERATURE, 64 SI_QUERY_CURRENT_GPU_SCLK, 65 SI_QUERY_CURRENT_GPU_MCLK, 66 SI_QUERY_GPU_LOAD, 67 SI_QUERY_GPU_SHADERS_BUSY, 68 SI_QUERY_GPU_TA_BUSY, 69 SI_QUERY_GPU_GDS_BUSY, 70 SI_QUERY_GPU_VGT_BUSY, 71 SI_QUERY_GPU_IA_BUSY, 72 SI_QUERY_GPU_SX_BUSY, 73 SI_QUERY_GPU_WD_BUSY, 74 SI_QUERY_GPU_BCI_BUSY, 75 SI_QUERY_GPU_SC_BUSY, 76 SI_QUERY_GPU_PA_BUSY, 77 SI_QUERY_GPU_DB_BUSY, 78 SI_QUERY_GPU_CP_BUSY, 79 SI_QUERY_GPU_CB_BUSY, 80 SI_QUERY_GPU_SDMA_BUSY, 81 SI_QUERY_GPU_PFP_BUSY, 82 SI_QUERY_GPU_MEQ_BUSY, 83 SI_QUERY_GPU_ME_BUSY, 84 SI_QUERY_GPU_SURF_SYNC_BUSY, 85 SI_QUERY_GPU_CP_DMA_BUSY, 86 SI_QUERY_GPU_SCRATCH_RAM_BUSY, 87 SI_QUERY_NUM_COMPILATIONS, 88 SI_QUERY_NUM_SHADERS_CREATED, 89 SI_QUERY_BACK_BUFFER_PS_DRAW_RATIO, 90 SI_QUERY_GPIN_ASIC_ID, 91 SI_QUERY_GPIN_NUM_SIMD, 92 SI_QUERY_GPIN_NUM_RB, 93 SI_QUERY_GPIN_NUM_SPI, 94 SI_QUERY_GPIN_NUM_SE, 95 SI_QUERY_LIVE_SHADER_CACHE_HITS, 96 SI_QUERY_LIVE_SHADER_CACHE_MISSES, 97 SI_QUERY_MEMORY_SHADER_CACHE_HITS, 98 SI_QUERY_MEMORY_SHADER_CACHE_MISSES, 99 SI_QUERY_DISK_SHADER_CACHE_HITS, 100 SI_QUERY_DISK_SHADER_CACHE_MISSES, 101 102 SI_QUERY_FIRST_PERFCOUNTER = PIPE_QUERY_DRIVER_SPECIFIC + 100, 103 }; 104 105 enum 106 { 107 SI_QUERY_GROUP_GPIN = 0, 108 SI_NUM_SW_QUERY_GROUPS 109 }; 110 111 struct si_query_ops { 112 void (*destroy)(struct si_context *, struct si_query *); 113 bool (*begin)(struct si_context *, struct si_query *); 114 bool (*end)(struct si_context *, struct si_query *); 115 bool (*get_result)(struct si_context *, struct si_query *, bool wait, 116 union pipe_query_result *result); 117 void (*get_result_resource)(struct si_context *, struct si_query *, 118 enum pipe_query_flags flags, 119 enum pipe_query_value_type result_type, int index, 120 struct pipe_resource *resource, unsigned offset); 121 122 void (*suspend)(struct si_context *, struct si_query *); 123 void (*resume)(struct si_context *, struct si_query *); 124 }; 125 126 struct si_query { 127 struct threaded_query b; 128 const struct si_query_ops *ops; 129 130 /* The PIPE_QUERY_xxx type of query */ 131 unsigned type; 132 133 /* The number of dwords for suspend. */ 134 unsigned num_cs_dw_suspend; 135 136 /* Linked list of queries that must be suspended at end of CS. */ 137 struct list_head active_list; 138 }; 139 140 enum 141 { 142 SI_QUERY_HW_FLAG_NO_START = (1 << 0), 143 /* gap */ 144 /* whether begin_query doesn't clear the result */ 145 SI_QUERY_HW_FLAG_BEGIN_RESUMES = (1 << 2), 146 /* whether GS invocations and emitted primitives counters are emulated 147 * using atomic adds. 148 */ 149 SI_QUERY_EMULATE_GS_COUNTERS = (1 << 3), 150 }; 151 152 struct si_query_buffer { 153 /* The buffer where query results are stored. */ 154 struct si_resource *buf; 155 /* If a query buffer is full, a new buffer is created and the old one 156 * is put in here. When we calculate the result, we sum up the samples 157 * from all buffers. */ 158 struct si_query_buffer *previous; 159 /* Offset of the next free result after current query data */ 160 unsigned results_end; 161 }; 162 163 void si_query_buffer_destroy(struct si_screen *sctx, struct si_query_buffer *buffer); 164 void si_query_buffer_reset(struct si_context *sctx, struct si_query_buffer *buffer); 165 bool si_query_buffer_alloc(struct si_context *sctx, struct si_query_buffer *buffer, 166 bool (*prepare_buffer)(struct si_context *, struct si_query_buffer *), 167 unsigned size); 168 169 struct si_query_hw { 170 struct si_query b; 171 unsigned flags; 172 173 /* The query buffer and how many results are in it. */ 174 struct si_query_buffer buffer; 175 /* Size of the result in memory for both begin_query and end_query, 176 * this can be one or two numbers, or it could even be a size of a structure. */ 177 unsigned result_size; 178 union { 179 /* For transform feedback: which stream the query is for */ 180 unsigned stream; 181 /* For pipeline stats: which counter is active */ 182 unsigned index; 183 }; 184 185 /* Workaround via compute shader */ 186 struct si_resource *workaround_buf; 187 unsigned workaround_offset; 188 }; 189 190 unsigned si_query_pipestat_end_dw_offset(struct si_screen *sscreen, 191 enum pipe_statistics_query_index index); 192 193 /* Shader-based queries */ 194 195 /** 196 * The query buffer is written to by ESGS NGG shaders with statistics about 197 * generated and (streamout-)emitted primitives. 198 * 199 * The context maintains a ring of these query buffers, and queries simply 200 * point into the ring, allowing an arbitrary number of queries to be active 201 * without additional GPU cost. 202 */ 203 struct gfx11_sh_query_buffer { 204 struct list_head list; 205 struct si_resource *buf; 206 unsigned refcount; 207 208 /* Offset into the buffer in bytes; points at the first un-emitted entry. */ 209 unsigned head; 210 }; 211 212 /* Memory layout of the query buffer. Must be kept in sync with shaders 213 * (including QBO shaders) and should be aligned to cachelines. 214 * 215 * The somewhat awkward memory layout is for compatibility with the 216 * SET_PREDICATION packet, which also means that we're setting the high bit 217 * of all those values unconditionally. 218 */ 219 struct gfx11_sh_query_buffer_mem { 220 struct { 221 uint64_t generated_primitives_start_dummy; 222 uint64_t emitted_primitives_start_dummy; 223 uint64_t generated_primitives; 224 uint64_t emitted_primitives; 225 } stream[4]; 226 uint32_t fence; /* bottom-of-pipe fence: set to ~0 when draws have finished */ 227 uint32_t pad[31]; 228 }; 229 230 struct gfx11_sh_query { 231 struct si_query b; 232 233 struct gfx11_sh_query_buffer *first; 234 struct gfx11_sh_query_buffer *last; 235 unsigned first_begin; 236 unsigned last_end; 237 238 unsigned stream; 239 }; 240 241 struct pipe_query *gfx11_sh_query_create(struct si_screen *screen, enum pipe_query_type query_type, 242 unsigned index); 243 244 /* Performance counters */ 245 struct si_perfcounters { 246 struct ac_perfcounters base; 247 248 unsigned num_stop_cs_dwords; 249 unsigned num_instance_cs_dwords; 250 }; 251 252 struct pipe_query *si_create_batch_query(struct pipe_context *ctx, unsigned num_queries, 253 unsigned *query_types); 254 255 int si_get_perfcounter_info(struct si_screen *, unsigned index, 256 struct pipe_driver_query_info *info); 257 int si_get_perfcounter_group_info(struct si_screen *, unsigned index, 258 struct pipe_driver_query_group_info *info); 259 260 struct si_qbo_state { 261 struct pipe_constant_buffer saved_const0; 262 }; 263 264 #endif /* SI_QUERY_H */ 265