/* Max counters per HW block */
#define R600_QUERY_MAX_COUNTERS 16
-static const char * const r600_pc_shader_suffix[] = {
- "", "_PS", "_VS", "_GS", "_ES", "_HS", "_LS", "_CS"
-};
-
static struct r600_perfcounter_block *
lookup_counter(struct r600_perfcounters *pc, unsigned index,
unsigned *base_gid, unsigned *sub_index)
struct r600_pc_counter {
unsigned base;
- unsigned dwords;
- unsigned stride;
+ unsigned qwords;
+ unsigned stride; /* in uint64s */
};
+#define R600_PC_SHADERS_WINDOWING (1 << 31)
+
struct r600_query_pc {
struct r600_query_hw b;
pc->emit_read(ctx, block,
group->num_counters, group->selectors,
buffer, va);
- va += 4 * group->num_counters;
+ va += sizeof(uint64_t) * group->num_counters;
} while (group->instance < 0 && ++instance < block->num_instances);
} while (++se < se_end);
}
union pipe_query_result *result)
{
struct r600_query_pc *query = (struct r600_query_pc *)hwquery;
- uint32_t *results = buffer;
+ uint64_t *results = buffer;
unsigned i, j;
for (i = 0; i < query->num_counters; ++i) {
struct r600_pc_counter *counter = &query->counters[i];
- if (counter->base == ~0)
- continue;
-
- for (j = 0; j < counter->dwords; ++j) {
+ for (j = 0; j < counter->qwords; ++j) {
uint32_t value = results[counter->base + j * counter->stride];
- result->batch[i].u32 += value;
+ result->batch[i].u64 += value;
}
}
}
if (block->flags & R600_PC_BLOCK_SHADER) {
unsigned sub_gids = block->num_instances;
unsigned shader_id;
- unsigned shader_mask;
- unsigned query_shader_mask;
+ unsigned shaders;
+ unsigned query_shaders;
if (block->flags & R600_PC_BLOCK_SE_GROUPS)
sub_gids = sub_gids * screen->info.max_se;
shader_id = sub_gid / sub_gids;
sub_gid = sub_gid % sub_gids;
- if (shader_id == 0)
- shader_mask = R600_PC_SHADER_ALL;
- else
- shader_mask = 1 << (shader_id - 1);
+ shaders = screen->perfcounters->shader_type_bits[shader_id];
- query_shader_mask = query->shaders & R600_PC_SHADER_ALL;
- if (query_shader_mask && query_shader_mask != shader_mask) {
+ query_shaders = query->shaders & ~R600_PC_SHADERS_WINDOWING;
+ if (query_shaders && query_shaders != shaders) {
fprintf(stderr, "r600_perfcounter: incompatible shader groups\n");
FREE(group);
return NULL;
}
- query->shaders |= shader_mask;
+ query->shaders = shaders;
}
- if (block->flags & R600_PC_BLOCK_SHADER_WINDOWED) {
+ if (block->flags & R600_PC_BLOCK_SHADER_WINDOWED && !query->shaders) {
// A non-zero value in query->shaders ensures that the shader
// masking is reset unless the user explicitly requests one.
- query->shaders |= R600_PC_SHADER_WINDOWING;
+ query->shaders = R600_PC_SHADERS_WINDOWING;
}
if (block->flags & R600_PC_BLOCK_SE_GROUPS) {
query->b.b.ops = &batch_query_ops;
query->b.ops = &batch_query_hw_ops;
- query->b.flags = R600_QUERY_HW_FLAG_TIMER;
query->num_counters = num_queries;
instances *= block->num_instances;
group->result_base = i;
- query->b.result_size += 4 * instances * group->num_counters;
+ query->b.result_size += sizeof(uint64_t) * instances * group->num_counters;
i += instances * group->num_counters;
pc->get_size(block, group->num_counters, group->selectors,
}
if (query->shaders) {
- if ((query->shaders & R600_PC_SHADER_ALL) == 0)
- query->shaders |= R600_PC_SHADER_ALL;
+ if (query->shaders == R600_PC_SHADERS_WINDOWING)
+ query->shaders = 0xffffffff;
query->b.num_cs_dw_begin += pc->num_shaders_cs_dwords;
}
counter->base = group->result_base + j;
counter->stride = group->num_counters;
- counter->dwords = 1;
+ counter->qwords = 1;
if ((block->flags & R600_PC_BLOCK_SE) && group->se < 0)
- counter->dwords = screen->info.max_se;
+ counter->qwords = screen->info.max_se;
if (group->instance < 0)
- counter->dwords *= block->num_instances;
+ counter->qwords *= block->num_instances;
}
if (!r600_query_hw_init(rctx, &query->b))
return NULL;
}
+static boolean r600_init_block_names(struct r600_common_screen *screen,
+ struct r600_perfcounter_block *block)
+{
+ unsigned i, j, k;
+ unsigned groups_shader = 1, groups_se = 1, groups_instance = 1;
+ unsigned namelen;
+ char *groupname;
+ char *p;
+
+ if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS)
+ groups_instance = block->num_instances;
+ if (block->flags & R600_PC_BLOCK_SE_GROUPS)
+ groups_se = screen->info.max_se;
+ if (block->flags & R600_PC_BLOCK_SHADER)
+ groups_shader = screen->perfcounters->num_shader_types;
+
+ namelen = strlen(block->basename);
+ block->group_name_stride = namelen + 1;
+ if (block->flags & R600_PC_BLOCK_SHADER)
+ block->group_name_stride += 3;
+ if (block->flags & R600_PC_BLOCK_SE_GROUPS) {
+ assert(groups_se <= 10);
+ block->group_name_stride += 1;
+
+ if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS)
+ block->group_name_stride += 1;
+ }
+ if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS) {
+ assert(groups_instance <= 100);
+ block->group_name_stride += 2;
+ }
+
+ block->group_names = MALLOC(block->num_groups * block->group_name_stride);
+ if (!block->group_names)
+ return FALSE;
+
+ groupname = block->group_names;
+ for (i = 0; i < groups_shader; ++i) {
+ const char *shader_suffix = screen->perfcounters->shader_type_suffixes[i];
+ unsigned shaderlen = strlen(shader_suffix);
+ for (j = 0; j < groups_se; ++j) {
+ for (k = 0; k < groups_instance; ++k) {
+ strcpy(groupname, block->basename);
+ p = groupname + namelen;
+
+ if (block->flags & R600_PC_BLOCK_SHADER) {
+ strcpy(p, shader_suffix);
+ p += shaderlen;
+ }
+
+ if (block->flags & R600_PC_BLOCK_SE_GROUPS) {
+ p += sprintf(p, "%d", j);
+ if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS)
+ *p++ = '_';
+ }
+
+ if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS)
+ p += sprintf(p, "%d", k);
+
+ groupname += block->group_name_stride;
+ }
+ }
+ }
+
+ assert(block->num_selectors <= 1000);
+ block->selector_name_stride = block->group_name_stride + 4;
+ block->selector_names = MALLOC(block->num_groups * block->num_selectors *
+ block->selector_name_stride);
+ if (!block->selector_names)
+ return FALSE;
+
+ groupname = block->group_names;
+ p = block->selector_names;
+ for (i = 0; i < block->num_groups; ++i) {
+ for (j = 0; j < block->num_selectors; ++j) {
+ sprintf(p, "%s_%03d", groupname, j);
+ p += block->selector_name_stride;
+ }
+ groupname += block->group_name_stride;
+ }
+
+ return TRUE;
+}
+
int r600_get_perfcounter_info(struct r600_common_screen *screen,
unsigned index,
struct pipe_driver_query_info *info)
if (!block)
return 0;
+ if (!block->selector_names) {
+ if (!r600_init_block_names(screen, block))
+ return 0;
+ }
info->name = block->selector_names + sub * block->selector_name_stride;
info->query_type = R600_QUERY_FIRST_PERFCOUNTER + index;
info->max_value.u64 = 0;
- info->type = PIPE_DRIVER_QUERY_TYPE_UINT;
+ info->type = PIPE_DRIVER_QUERY_TYPE_UINT64;
info->result_type = PIPE_DRIVER_QUERY_RESULT_TYPE_CUMULATIVE;
info->group_id = base_gid + sub / block->num_selectors;
info->flags = PIPE_DRIVER_QUERY_FLAG_BATCH;
+ if (sub > 0 && sub + 1 < block->num_selectors * block->num_groups)
+ info->flags |= PIPE_DRIVER_QUERY_FLAG_DONT_LIST;
return 1;
}
block = lookup_group(pc, &index);
if (!block)
return 0;
+
+ if (!block->group_names) {
+ if (!r600_init_block_names(screen, block))
+ return 0;
+ }
info->name = block->group_names + index * block->group_name_stride;
info->num_queries = block->num_selectors;
info->max_active_queries = block->num_counters;
return TRUE;
}
-boolean r600_perfcounters_add_block(struct r600_common_screen *rscreen,
- struct r600_perfcounters *pc,
- const char *name, unsigned flags,
- unsigned counters, unsigned selectors,
- unsigned instances, void *data)
+void r600_perfcounters_add_block(struct r600_common_screen *rscreen,
+ struct r600_perfcounters *pc,
+ const char *name, unsigned flags,
+ unsigned counters, unsigned selectors,
+ unsigned instances, void *data)
{
struct r600_perfcounter_block *block = &pc->blocks[pc->num_blocks];
- unsigned i, j, k;
- unsigned groups_shader = 1, groups_se = 1, groups_instance = 1;
- unsigned namelen;
- char *groupname;
- char *p;
assert(counters <= R600_QUERY_MAX_COUNTERS);
block->flags |= R600_PC_BLOCK_INSTANCE_GROUPS;
if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS) {
- groups_instance = block->num_instances;
- block->num_groups = groups_instance;
+ block->num_groups = block->num_instances;
} else {
block->num_groups = 1;
}
- if (block->flags & R600_PC_BLOCK_SE_GROUPS) {
- groups_se = rscreen->info.max_se;
- block->num_groups *= groups_se;
- }
-
- if (block->flags & R600_PC_BLOCK_SHADER) {
- groups_shader = ARRAY_SIZE(r600_pc_shader_suffix);
- block->num_groups *= groups_shader;
- }
-
- namelen = strlen(name);
- block->group_name_stride = namelen + 1;
+ if (block->flags & R600_PC_BLOCK_SE_GROUPS)
+ block->num_groups *= rscreen->info.max_se;
if (block->flags & R600_PC_BLOCK_SHADER)
- block->group_name_stride += 3;
- if (block->flags & R600_PC_BLOCK_SE_GROUPS) {
- assert(groups_se <= 10);
- block->group_name_stride += 1;
-
- if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS)
- block->group_name_stride += 1;
- }
- if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS) {
- assert(groups_instance <= 100);
- block->group_name_stride += 2;
- }
-
- block->group_names = MALLOC(block->num_groups * block->group_name_stride);
- if (!block->group_names)
- goto error;
-
- groupname = block->group_names;
- for (i = 0; i < groups_shader; ++i) {
- unsigned shaderlen = strlen(r600_pc_shader_suffix[i]);
- for (j = 0; j < groups_se; ++j) {
- for (k = 0; k < groups_instance; ++k) {
- strcpy(groupname, name);
- p = groupname + namelen;
-
- if (block->flags & R600_PC_BLOCK_SHADER) {
- strcpy(p, r600_pc_shader_suffix[i]);
- p += shaderlen;
- }
-
- if (block->flags & R600_PC_BLOCK_SE_GROUPS) {
- p += sprintf(p, "%d", j);
- if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS)
- *p++ = '_';
- }
-
- if (block->flags & R600_PC_BLOCK_INSTANCE_GROUPS)
- p += sprintf(p, "%d", k);
-
- groupname += block->group_name_stride;
- }
- }
- }
-
- assert(selectors <= 1000);
- block->selector_name_stride = block->group_name_stride + 4;
- block->selector_names = MALLOC(block->num_groups * selectors *
- block->selector_name_stride);
- if (!block->selector_names)
- goto error_groupnames;
-
- groupname = block->group_names;
- p = block->selector_names;
- for (i = 0; i < block->num_groups; ++i) {
- for (j = 0; j < selectors; ++j) {
- sprintf(p, "%s_%03d", groupname, j);
- p += block->selector_name_stride;
- }
- groupname += block->group_name_stride;
- }
+ block->num_groups *= pc->num_shader_types;
++pc->num_blocks;
pc->num_groups += block->num_groups;
-
- return TRUE;
-
-error_groupnames:
- FREE(block->group_names);
-error:
- return FALSE;
}
void r600_perfcounters_do_destroy(struct r600_perfcounters *pc)
projects / mesa.git / blobdiff