| 1 |
/* |
| 2 |
* Copyright © 2018 Jonathan Marek <[email protected]> |
| 3 |
* SPDX-License-Identifier: MIT |
| 4 |
* |
| 5 |
* Authors: |
| 6 |
* Jonathan Marek <[email protected]> |
| 7 |
* Rob Clark <[email protected]> |
| 8 |
*/ |
| 9 |
|
| 10 |
/* NOTE: perfcntrs are 48-bits but we only have 32-bit accumulate (?) |
| 11 |
* so we work with 32-bits only. we accumulate start/stop separately, |
| 12 |
* which differs from a5xx but works with only accumulate (no add/neg) |
| 13 |
*/ |
| 14 |
|
| 15 |
#include "freedreno_query_acc.h" |
| 16 |
#include "freedreno_resource.h" |
| 17 |
|
| 18 |
#include "fd2_context.h" |
| 19 |
#include "fd2_query.h" |
| 20 |
|
| 21 |
struct PACKED fd2_query_sample { |
| 22 |
struct fd_acc_query_sample base; |
| 23 |
uint32_t start; |
| 24 |
uint32_t stop; |
| 25 |
}; |
| 26 |
FD_DEFINE_CAST(fd_acc_query_sample, fd2_query_sample); |
| 27 |
|
| 28 |
/* offset of a single field of an array of fd2_query_sample: */ |
| 29 |
#define query_sample_idx(aq, idx, field) \ |
| 30 |
fd_resource((aq)->prsc)->bo, \ |
| 31 |
(idx * sizeof(struct fd2_query_sample)) + \ |
| 32 |
offsetof(struct fd2_query_sample, field), \ |
| 33 |
0, 0 |
| 34 |
|
| 35 |
/* offset of a single field of fd2_query_sample: */ |
| 36 |
#define query_sample(aq, field) query_sample_idx(aq, 0, field) |
| 37 |
|
| 38 |
/* |
| 39 |
* Performance Counter (batch) queries: |
| 40 |
* |
| 41 |
* Only one of these is active at a time, per design of the gallium |
| 42 |
* batch_query API design. On perfcntr query tracks N query_types, |
| 43 |
* each of which has a 'fd_batch_query_entry' that maps it back to |
| 44 |
* the associated group and counter. |
| 45 |
*/ |
| 46 |
|
| 47 |
struct fd_batch_query_entry { |
| 48 |
uint8_t gid; /* group-id */ |
| 49 |
uint8_t cid; /* countable-id within the group */ |
| 50 |
}; |
| 51 |
|
| 52 |
struct fd_batch_query_data { |
| 53 |
struct fd_screen *screen; |
| 54 |
unsigned num_query_entries; |
| 55 |
struct fd_batch_query_entry query_entries[]; |
| 56 |
}; |
| 57 |
|
| 58 |
static void |
| 59 |
perfcntr_resume(struct fd_acc_query *aq, struct fd_batch *batch) assert_dt |
| 60 |
{ |
| 61 |
struct fd_batch_query_data *data = aq->query_data; |
| 62 |
struct fd_screen *screen = data->screen; |
| 63 |
struct fd_ringbuffer *ring = batch->draw; |
| 64 |
|
| 65 |
unsigned counters_per_group[screen->num_perfcntr_groups]; |
| 66 |
memset(counters_per_group, 0, sizeof(counters_per_group)); |
| 67 |
|
| 68 |
fd_wfi(batch, ring); |
| 69 |
|
| 70 |
/* configure performance counters for the requested queries: */ |
| 71 |
for (unsigned i = 0; i < data->num_query_entries; i++) { |
| 72 |
struct fd_batch_query_entry *entry = &data->query_entries[i]; |
| 73 |
const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid]; |
| 74 |
unsigned counter_idx = counters_per_group[entry->gid]++; |
| 75 |
|
| 76 |
assert(counter_idx < g->num_counters); |
| 77 |
|
| 78 |
OUT_PKT0(ring, g->counters[counter_idx].select_reg, 1); |
| 79 |
OUT_RING(ring, g->countables[entry->cid].selector); |
| 80 |
} |
| 81 |
|
| 82 |
memset(counters_per_group, 0, sizeof(counters_per_group)); |
| 83 |
|
| 84 |
/* and snapshot the start values */ |
| 85 |
for (unsigned i = 0; i < data->num_query_entries; i++) { |
| 86 |
struct fd_batch_query_entry *entry = &data->query_entries[i]; |
| 87 |
const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid]; |
| 88 |
unsigned counter_idx = counters_per_group[entry->gid]++; |
| 89 |
const struct fd_perfcntr_counter *counter = &g->counters[counter_idx]; |
| 90 |
|
| 91 |
OUT_PKT3(ring, CP_REG_TO_MEM, 2); |
| 92 |
OUT_RING(ring, counter->counter_reg_lo | CP_REG_TO_MEM_0_ACCUMULATE); |
| 93 |
OUT_RELOC(ring, query_sample_idx(aq, i, start)); |
| 94 |
} |
| 95 |
} |
| 96 |
|
| 97 |
static void |
| 98 |
perfcntr_pause(struct fd_acc_query *aq, struct fd_batch *batch) assert_dt |
| 99 |
{ |
| 100 |
struct fd_batch_query_data *data = aq->query_data; |
| 101 |
struct fd_screen *screen = data->screen; |
| 102 |
struct fd_ringbuffer *ring = batch->draw; |
| 103 |
|
| 104 |
unsigned counters_per_group[screen->num_perfcntr_groups]; |
| 105 |
memset(counters_per_group, 0, sizeof(counters_per_group)); |
| 106 |
|
| 107 |
fd_wfi(batch, ring); |
| 108 |
|
| 109 |
/* TODO do we need to bother to turn anything off? */ |
| 110 |
|
| 111 |
/* snapshot the end values: */ |
| 112 |
for (unsigned i = 0; i < data->num_query_entries; i++) { |
| 113 |
struct fd_batch_query_entry *entry = &data->query_entries[i]; |
| 114 |
const struct fd_perfcntr_group *g = &screen->perfcntr_groups[entry->gid]; |
| 115 |
unsigned counter_idx = counters_per_group[entry->gid]++; |
| 116 |
const struct fd_perfcntr_counter *counter = &g->counters[counter_idx]; |
| 117 |
|
| 118 |
OUT_PKT3(ring, CP_REG_TO_MEM, 2); |
| 119 |
OUT_RING(ring, counter->counter_reg_lo | CP_REG_TO_MEM_0_ACCUMULATE); |
| 120 |
OUT_RELOC(ring, query_sample_idx(aq, i, stop)); |
| 121 |
} |
| 122 |
} |
| 123 |
|
| 124 |
static void |
| 125 |
perfcntr_accumulate_result(struct fd_acc_query *aq, |
| 126 |
struct fd_acc_query_sample *s, |
| 127 |
union pipe_query_result *result) |
| 128 |
{ |
| 129 |
struct fd_batch_query_data *data = aq->query_data; |
| 130 |
struct fd2_query_sample *sp = fd2_query_sample(s); |
| 131 |
|
| 132 |
for (unsigned i = 0; i < data->num_query_entries; i++) |
| 133 |
result->batch[i].u64 = sp[i].stop - sp[i].start; |
| 134 |
} |
| 135 |
|
| 136 |
static const struct fd_acc_sample_provider perfcntr = { |
| 137 |
.query_type = FD_QUERY_FIRST_PERFCNTR, |
| 138 |
.always = true, |
| 139 |
.resume = perfcntr_resume, |
| 140 |
.pause = perfcntr_pause, |
| 141 |
.result = perfcntr_accumulate_result, |
| 142 |
}; |
| 143 |
|
| 144 |
static struct pipe_query * |
| 145 |
fd2_create_batch_query(struct pipe_context *pctx, unsigned num_queries, |
| 146 |
unsigned *query_types) |
| 147 |
{ |
| 148 |
struct fd_context *ctx = fd_context(pctx); |
| 149 |
struct fd_screen *screen = ctx->screen; |
| 150 |
struct fd_query *q; |
| 151 |
struct fd_acc_query *aq; |
| 152 |
struct fd_batch_query_data *data; |
| 153 |
|
| 154 |
data = CALLOC_VARIANT_LENGTH_STRUCT( |
| 155 |
fd_batch_query_data, num_queries * sizeof(data->query_entries[0])); |
| 156 |
|
| 157 |
data->screen = screen; |
| 158 |
data->num_query_entries = num_queries; |
| 159 |
|
| 160 |
/* validate the requested query_types and ensure we don't try |
| 161 |
* to request more query_types of a given group than we have |
| 162 |
* counters: |
| 163 |
*/ |
| 164 |
unsigned counters_per_group[screen->num_perfcntr_groups]; |
| 165 |
memset(counters_per_group, 0, sizeof(counters_per_group)); |
| 166 |
|
| 167 |
for (unsigned i = 0; i < num_queries; i++) { |
| 168 |
unsigned idx = query_types[i] - FD_QUERY_FIRST_PERFCNTR; |
| 169 |
|
| 170 |
/* verify valid query_type, ie. is it actually a perfcntr? */ |
| 171 |
if ((query_types[i] < FD_QUERY_FIRST_PERFCNTR) || |
| 172 |
(idx >= screen->num_perfcntr_queries)) { |
| 173 |
mesa_loge("invalid batch query query_type: %u", query_types[i]); |
| 174 |
goto error; |
| 175 |
} |
| 176 |
|
| 177 |
struct fd_batch_query_entry *entry = &data->query_entries[i]; |
| 178 |
struct pipe_driver_query_info *pq = &screen->perfcntr_queries[idx]; |
| 179 |
|
| 180 |
entry->gid = pq->group_id; |
| 181 |
|
| 182 |
/* the perfcntr_queries[] table flattens all the countables |
| 183 |
* for each group in series, ie: |
| 184 |
* |
| 185 |
* (G0,C0), .., (G0,Cn), (G1,C0), .., (G1,Cm), ... |
| 186 |
* |
| 187 |
* So to find the countable index just step back through the |
| 188 |
* table to find the first entry with the same group-id. |
| 189 |
*/ |
| 190 |
while (pq > screen->perfcntr_queries) { |
| 191 |
pq--; |
| 192 |
if (pq->group_id == entry->gid) |
| 193 |
entry->cid++; |
| 194 |
} |
| 195 |
|
| 196 |
if (counters_per_group[entry->gid] >= |
| 197 |
screen->perfcntr_groups[entry->gid].num_counters) { |
| 198 |
mesa_loge("too many counters for group %u", entry->gid); |
| 199 |
goto error; |
| 200 |
} |
| 201 |
|
| 202 |
counters_per_group[entry->gid]++; |
| 203 |
} |
| 204 |
|
| 205 |
q = fd_acc_create_query2(ctx, 0, 0, &perfcntr); |
| 206 |
aq = fd_acc_query(q); |
| 207 |
|
| 208 |
/* sample buffer size is based on # of queries: */ |
| 209 |
aq->size = num_queries * sizeof(struct fd2_query_sample); |
| 210 |
aq->query_data = data; |
| 211 |
|
| 212 |
return (struct pipe_query *)q; |
| 213 |
|
| 214 |
error: |
| 215 |
free(data); |
| 216 |
return NULL; |
| 217 |
} |
| 218 |
|
| 219 |
void |
| 220 |
fd2_query_context_init(struct pipe_context *pctx) disable_thread_safety_analysis |
| 221 |
{ |
| 222 |
struct fd_context *ctx = fd_context(pctx); |
| 223 |
|
| 224 |
ctx->create_query = fd_acc_create_query; |
| 225 |
ctx->query_update_batch = fd_acc_query_update_batch; |
| 226 |
|
| 227 |
pctx->create_batch_query = fd2_create_batch_query; |
| 228 |
} |