static struct r600_resource *r600_new_query_buffer(struct r600_common_context *ctx,
struct r600_query_hw *query)
{
- unsigned buf_size = 4096;
+ unsigned buf_size = MAX2(query->result_size, 4096);
/* Queries are normally read by the CPU after
* being written by the gpu, hence staging is probably a good
pipe_buffer_create(ctx->b.screen, PIPE_BIND_CUSTOM,
PIPE_USAGE_STAGING, buf_size);
- if (query->ops->prepare_buffer)
+ if (query->flags & R600_QUERY_HW_FLAG_PREDICATE)
query->ops->prepare_buffer(ctx, query, buf);
return buf;
static void r600_query_hw_prepare_buffer(struct r600_common_context *ctx,
struct r600_query_hw *query,
struct r600_resource *buffer)
- {
- uint32_t *results;
-
- if (query->b.type == PIPE_QUERY_TIME_ELAPSED ||
- query->b.type == PIPE_QUERY_TIMESTAMP)
- return;
-
- results = r600_buffer_map_sync_with_rings(ctx, buffer,
- PIPE_TRANSFER_WRITE);
+{
+ /* Callers ensure that the buffer is currently unused by the GPU. */
+ uint32_t *results = ctx->ws->buffer_map(buffer->cs_buf, NULL,
+ PIPE_TRANSFER_WRITE |
+ PIPE_TRANSFER_UNSYNCHRONIZED);
memset(results, 0, buffer->b.b.width0);
}
}
-static boolean r600_query_hw_begin(struct r600_common_context *, struct r600_query *);
-static void r600_query_hw_end(struct r600_common_context *, struct r600_query *);
-static boolean r600_query_hw_get_result(struct r600_common_context *,
- struct r600_query *, boolean wait,
- union pipe_query_result *result);
-
static struct r600_query_ops query_hw_ops = {
.destroy = r600_query_hw_destroy,
.begin = r600_query_hw_begin,
struct r600_query_hw *query,
struct r600_resource *buffer,
uint64_t va);
+static void r600_query_hw_add_result(struct r600_common_context *ctx,
+ struct r600_query_hw *, void *buffer,
+ union pipe_query_result *result);
+static void r600_query_hw_clear_result(struct r600_query_hw *,
+ union pipe_query_result *);
static struct r600_query_hw_ops query_hw_default_hw_ops = {
.prepare_buffer = r600_query_hw_prepare_buffer,
.emit_start = r600_query_hw_do_emit_start,
.emit_stop = r600_query_hw_do_emit_stop,
+ .clear_result = r600_query_hw_clear_result,
+ .add_result = r600_query_hw_add_result,
};
+boolean r600_query_hw_init(struct r600_common_context *rctx,
+ struct r600_query_hw *query)
+{
+ query->buffer.buf = r600_new_query_buffer(rctx, query);
+ if (!query->buffer.buf)
+ return FALSE;
+
+ return TRUE;
+}
+
static struct pipe_query *r600_query_hw_create(struct r600_common_context *rctx,
unsigned query_type,
unsigned index)
case PIPE_QUERY_OCCLUSION_COUNTER:
case PIPE_QUERY_OCCLUSION_PREDICATE:
query->result_size = 16 * rctx->max_db;
- query->num_cs_dw = 6;
+ query->num_cs_dw_begin = 6;
+ query->num_cs_dw_end = 6;
+ query->flags |= R600_QUERY_HW_FLAG_PREDICATE;
break;
case PIPE_QUERY_TIME_ELAPSED:
query->result_size = 16;
- query->num_cs_dw = 8;
+ query->num_cs_dw_begin = 8;
+ query->num_cs_dw_end = 8;
query->flags = R600_QUERY_HW_FLAG_TIMER;
break;
case PIPE_QUERY_TIMESTAMP:
query->result_size = 8;
- query->num_cs_dw = 8;
+ query->num_cs_dw_end = 8;
query->flags = R600_QUERY_HW_FLAG_TIMER |
R600_QUERY_HW_FLAG_NO_START;
break;
case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
/* NumPrimitivesWritten, PrimitiveStorageNeeded. */
query->result_size = 32;
- query->num_cs_dw = 6;
+ query->num_cs_dw_begin = 6;
+ query->num_cs_dw_end = 6;
query->stream = index;
+ query->flags |= R600_QUERY_HW_FLAG_PREDICATE;
break;
case PIPE_QUERY_PIPELINE_STATISTICS:
/* 11 values on EG, 8 on R600. */
query->result_size = (rctx->chip_class >= EVERGREEN ? 11 : 8) * 16;
- query->num_cs_dw = 6;
+ query->num_cs_dw_begin = 6;
+ query->num_cs_dw_end = 6;
break;
default:
assert(0);
return NULL;
}
- query->buffer.buf = r600_new_query_buffer(rctx, query);
- if (!query->buffer.buf) {
+ if (!r600_query_hw_init(rctx, query)) {
FREE(query);
return NULL;
}
r600_update_occlusion_query_state(ctx, query->b.type, 1);
r600_update_prims_generated_query_state(ctx, query->b.type, 1);
- ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw * 2, TRUE);
+
+ ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw_begin + query->num_cs_dw_end,
+ TRUE);
/* Get a new query buffer if needed. */
if (query->buffer.results_end + query->result_size > query->buffer.buf->b.b.width0) {
query->ops->emit_start(ctx, query, query->buffer.buf, va);
if (query->flags & R600_QUERY_HW_FLAG_TIMER)
- ctx->num_cs_dw_timer_queries_suspend += query->num_cs_dw;
+ ctx->num_cs_dw_timer_queries_suspend += query->num_cs_dw_end;
else
- ctx->num_cs_dw_nontimer_queries_suspend += query->num_cs_dw;
-
+ ctx->num_cs_dw_nontimer_queries_suspend += query->num_cs_dw_end;
}
static void r600_query_hw_do_emit_stop(struct r600_common_context *ctx,
/* The queries which need begin already called this in begin_query. */
if (query->flags & R600_QUERY_HW_FLAG_NO_START) {
- ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw, FALSE);
+ ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw_end, FALSE);
}
/* emit end query */
if (!(query->flags & R600_QUERY_HW_FLAG_NO_START)) {
if (query->flags & R600_QUERY_HW_FLAG_TIMER)
- ctx->num_cs_dw_timer_queries_suspend -= query->num_cs_dw;
+ ctx->num_cs_dw_timer_queries_suspend -= query->num_cs_dw_end;
else
- ctx->num_cs_dw_nontimer_queries_suspend -= query->num_cs_dw;
+ ctx->num_cs_dw_nontimer_queries_suspend -= query->num_cs_dw_end;
}
r600_update_occlusion_query_state(ctx, query->b.type, -1);
return rquery->ops->begin(rctx, rquery);
}
-static boolean r600_query_hw_begin(struct r600_common_context *rctx,
- struct r600_query *rquery)
+static void r600_query_hw_reset_buffers(struct r600_common_context *rctx,
+ struct r600_query_hw *query)
{
- struct r600_query_hw *query = (struct r600_query_hw *)rquery;
struct r600_query_buffer *prev = query->buffer.previous;
- if (query->flags & R600_QUERY_HW_FLAG_NO_START) {
- assert(0);
- return false;
- }
-
/* Discard the old query buffers. */
while (prev) {
struct r600_query_buffer *qbuf = prev;
FREE(qbuf);
}
- /* Obtain a new buffer if the current one can't be mapped without a stall. */
- if (r600_rings_is_buffer_referenced(rctx, query->buffer.buf->cs_buf, RADEON_USAGE_READWRITE) ||
- !rctx->ws->buffer_wait(query->buffer.buf->buf, 0, RADEON_USAGE_READWRITE)) {
- pipe_resource_reference((struct pipe_resource**)&query->buffer.buf, NULL);
- query->buffer.buf = r600_new_query_buffer(rctx, query);
+ if (query->flags & R600_QUERY_HW_FLAG_PREDICATE) {
+ /* Obtain a new buffer if the current one can't be mapped without a stall. */
+ if (r600_rings_is_buffer_referenced(rctx, query->buffer.buf->cs_buf, RADEON_USAGE_READWRITE) ||
+ !rctx->ws->buffer_wait(query->buffer.buf->buf, 0, RADEON_USAGE_READWRITE)) {
+ pipe_resource_reference((struct pipe_resource**)&query->buffer.buf, NULL);
+ query->buffer.buf = r600_new_query_buffer(rctx, query);
+ } else {
+ query->ops->prepare_buffer(rctx, query, query->buffer.buf);
+ }
}
query->buffer.results_end = 0;
query->buffer.previous = NULL;
+}
+
+boolean r600_query_hw_begin(struct r600_common_context *rctx,
+ struct r600_query *rquery)
+{
+ struct r600_query_hw *query = (struct r600_query_hw *)rquery;
+
+ if (query->flags & R600_QUERY_HW_FLAG_NO_START) {
+ assert(0);
+ return false;
+ }
+
+ r600_query_hw_reset_buffers(rctx, query);
r600_query_hw_emit_start(rctx, query);
rquery->ops->end(rctx, rquery);
}
-static void r600_query_hw_end(struct r600_common_context *rctx,
+void r600_query_hw_end(struct r600_common_context *rctx,
struct r600_query *rquery)
{
struct r600_query_hw *query = (struct r600_query_hw *)rquery;
+ if (query->flags & R600_QUERY_HW_FLAG_NO_START)
+ r600_query_hw_reset_buffers(rctx, query);
+
r600_query_hw_emit_stop(rctx, query);
if (!(query->flags & R600_QUERY_HW_FLAG_NO_START))
LIST_DELINIT(&query->list);
}
-static unsigned r600_query_read_result(char *map, unsigned start_index, unsigned end_index,
+static unsigned r600_query_read_result(void *map, unsigned start_index, unsigned end_index,
bool test_status_bit)
{
uint32_t *current_result = (uint32_t*)map;
return 0;
}
-static boolean r600_get_query_buffer_result(struct r600_common_context *ctx,
- struct r600_query_hw *query,
- struct r600_query_buffer *qbuf,
- boolean wait,
- union pipe_query_result *result)
+static void r600_query_hw_add_result(struct r600_common_context *ctx,
+ struct r600_query_hw *query,
+ void *buffer,
+ union pipe_query_result *result)
{
- unsigned results_base = 0;
- char *map;
-
- map = r600_buffer_map_sync_with_rings(ctx, qbuf->buf,
- PIPE_TRANSFER_READ |
- (wait ? 0 : PIPE_TRANSFER_DONTBLOCK));
- if (!map)
- return FALSE;
-
- /* count all results across all data blocks */
switch (query->b.type) {
- case PIPE_QUERY_OCCLUSION_COUNTER:
- while (results_base != qbuf->results_end) {
+ case PIPE_QUERY_OCCLUSION_COUNTER: {
+ unsigned results_base = 0;
+ while (results_base != query->result_size) {
result->u64 +=
- r600_query_read_result(map + results_base, 0, 2, true);
+ r600_query_read_result(buffer + results_base, 0, 2, true);
results_base += 16;
}
break;
- case PIPE_QUERY_OCCLUSION_PREDICATE:
- while (results_base != qbuf->results_end) {
+ }
+ case PIPE_QUERY_OCCLUSION_PREDICATE: {
+ unsigned results_base = 0;
+ while (results_base != query->result_size) {
result->b = result->b ||
- r600_query_read_result(map + results_base, 0, 2, true) != 0;
+ r600_query_read_result(buffer + results_base, 0, 2, true) != 0;
results_base += 16;
}
break;
+ }
case PIPE_QUERY_TIME_ELAPSED:
- while (results_base != qbuf->results_end) {
- result->u64 +=
- r600_query_read_result(map + results_base, 0, 2, false);
- results_base += query->result_size;
- }
+ result->u64 += r600_query_read_result(buffer, 0, 2, false);
break;
case PIPE_QUERY_TIMESTAMP:
{
- uint32_t *current_result = (uint32_t*)map;
+ uint32_t *current_result = (uint32_t*)buffer;
result->u64 = (uint64_t)current_result[0] |
(uint64_t)current_result[1] << 32;
break;
* u64 PrimitiveStorageNeeded;
* }
* We only need NumPrimitivesWritten here. */
- while (results_base != qbuf->results_end) {
- result->u64 +=
- r600_query_read_result(map + results_base, 2, 6, true);
- results_base += query->result_size;
- }
+ result->u64 += r600_query_read_result(buffer, 2, 6, true);
break;
case PIPE_QUERY_PRIMITIVES_GENERATED:
/* Here we read PrimitiveStorageNeeded. */
- while (results_base != qbuf->results_end) {
- result->u64 +=
- r600_query_read_result(map + results_base, 0, 4, true);
- results_base += query->result_size;
- }
+ result->u64 += r600_query_read_result(buffer, 0, 4, true);
break;
case PIPE_QUERY_SO_STATISTICS:
- while (results_base != qbuf->results_end) {
- result->so_statistics.num_primitives_written +=
- r600_query_read_result(map + results_base, 2, 6, true);
- result->so_statistics.primitives_storage_needed +=
- r600_query_read_result(map + results_base, 0, 4, true);
- results_base += query->result_size;
- }
+ result->so_statistics.num_primitives_written +=
+ r600_query_read_result(buffer, 2, 6, true);
+ result->so_statistics.primitives_storage_needed +=
+ r600_query_read_result(buffer, 0, 4, true);
break;
case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
- while (results_base != qbuf->results_end) {
- result->b = result->b ||
- r600_query_read_result(map + results_base, 2, 6, true) !=
- r600_query_read_result(map + results_base, 0, 4, true);
- results_base += query->result_size;
- }
+ result->b = result->b ||
+ r600_query_read_result(buffer, 2, 6, true) !=
+ r600_query_read_result(buffer, 0, 4, true);
break;
case PIPE_QUERY_PIPELINE_STATISTICS:
if (ctx->chip_class >= EVERGREEN) {
- while (results_base != qbuf->results_end) {
- result->pipeline_statistics.ps_invocations +=
- r600_query_read_result(map + results_base, 0, 22, false);
- result->pipeline_statistics.c_primitives +=
- r600_query_read_result(map + results_base, 2, 24, false);
- result->pipeline_statistics.c_invocations +=
- r600_query_read_result(map + results_base, 4, 26, false);
- result->pipeline_statistics.vs_invocations +=
- r600_query_read_result(map + results_base, 6, 28, false);
- result->pipeline_statistics.gs_invocations +=
- r600_query_read_result(map + results_base, 8, 30, false);
- result->pipeline_statistics.gs_primitives +=
- r600_query_read_result(map + results_base, 10, 32, false);
- result->pipeline_statistics.ia_primitives +=
- r600_query_read_result(map + results_base, 12, 34, false);
- result->pipeline_statistics.ia_vertices +=
- r600_query_read_result(map + results_base, 14, 36, false);
- result->pipeline_statistics.hs_invocations +=
- r600_query_read_result(map + results_base, 16, 38, false);
- result->pipeline_statistics.ds_invocations +=
- r600_query_read_result(map + results_base, 18, 40, false);
- result->pipeline_statistics.cs_invocations +=
- r600_query_read_result(map + results_base, 20, 42, false);
- results_base += query->result_size;
- }
+ result->pipeline_statistics.ps_invocations +=
+ r600_query_read_result(buffer, 0, 22, false);
+ result->pipeline_statistics.c_primitives +=
+ r600_query_read_result(buffer, 2, 24, false);
+ result->pipeline_statistics.c_invocations +=
+ r600_query_read_result(buffer, 4, 26, false);
+ result->pipeline_statistics.vs_invocations +=
+ r600_query_read_result(buffer, 6, 28, false);
+ result->pipeline_statistics.gs_invocations +=
+ r600_query_read_result(buffer, 8, 30, false);
+ result->pipeline_statistics.gs_primitives +=
+ r600_query_read_result(buffer, 10, 32, false);
+ result->pipeline_statistics.ia_primitives +=
+ r600_query_read_result(buffer, 12, 34, false);
+ result->pipeline_statistics.ia_vertices +=
+ r600_query_read_result(buffer, 14, 36, false);
+ result->pipeline_statistics.hs_invocations +=
+ r600_query_read_result(buffer, 16, 38, false);
+ result->pipeline_statistics.ds_invocations +=
+ r600_query_read_result(buffer, 18, 40, false);
+ result->pipeline_statistics.cs_invocations +=
+ r600_query_read_result(buffer, 20, 42, false);
} else {
- while (results_base != qbuf->results_end) {
- result->pipeline_statistics.ps_invocations +=
- r600_query_read_result(map + results_base, 0, 16, false);
- result->pipeline_statistics.c_primitives +=
- r600_query_read_result(map + results_base, 2, 18, false);
- result->pipeline_statistics.c_invocations +=
- r600_query_read_result(map + results_base, 4, 20, false);
- result->pipeline_statistics.vs_invocations +=
- r600_query_read_result(map + results_base, 6, 22, false);
- result->pipeline_statistics.gs_invocations +=
- r600_query_read_result(map + results_base, 8, 24, false);
- result->pipeline_statistics.gs_primitives +=
- r600_query_read_result(map + results_base, 10, 26, false);
- result->pipeline_statistics.ia_primitives +=
- r600_query_read_result(map + results_base, 12, 28, false);
- result->pipeline_statistics.ia_vertices +=
- r600_query_read_result(map + results_base, 14, 30, false);
- results_base += query->result_size;
- }
+ result->pipeline_statistics.ps_invocations +=
+ r600_query_read_result(buffer, 0, 16, false);
+ result->pipeline_statistics.c_primitives +=
+ r600_query_read_result(buffer, 2, 18, false);
+ result->pipeline_statistics.c_invocations +=
+ r600_query_read_result(buffer, 4, 20, false);
+ result->pipeline_statistics.vs_invocations +=
+ r600_query_read_result(buffer, 6, 22, false);
+ result->pipeline_statistics.gs_invocations +=
+ r600_query_read_result(buffer, 8, 24, false);
+ result->pipeline_statistics.gs_primitives +=
+ r600_query_read_result(buffer, 10, 26, false);
+ result->pipeline_statistics.ia_primitives +=
+ r600_query_read_result(buffer, 12, 28, false);
+ result->pipeline_statistics.ia_vertices +=
+ r600_query_read_result(buffer, 14, 30, false);
}
#if 0 /* for testing */
printf("Pipeline stats: IA verts=%llu, IA prims=%llu, VS=%llu, HS=%llu, "
default:
assert(0);
}
-
- return TRUE;
}
static boolean r600_get_query_result(struct pipe_context *ctx,
return rquery->ops->get_result(rctx, rquery, wait, result);
}
-static boolean r600_query_hw_get_result(struct r600_common_context *rctx,
- struct r600_query *rquery,
- boolean wait, union pipe_query_result *result)
+static void r600_query_hw_clear_result(struct r600_query_hw *query,
+ union pipe_query_result *result)
+{
+ util_query_clear_result(result, query->b.type);
+}
+
+boolean r600_query_hw_get_result(struct r600_common_context *rctx,
+ struct r600_query *rquery,
+ boolean wait, union pipe_query_result *result)
{
struct r600_query_hw *query = (struct r600_query_hw *)rquery;
struct r600_query_buffer *qbuf;
- util_query_clear_result(result, rquery->type);
+ query->ops->clear_result(query, result);
for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
- if (!r600_get_query_buffer_result(rctx, query, qbuf, wait, result)) {
+ unsigned results_base = 0;
+ void *map;
+
+ map = r600_buffer_map_sync_with_rings(rctx, qbuf->buf,
+ PIPE_TRANSFER_READ |
+ (wait ? 0 : PIPE_TRANSFER_DONTBLOCK));
+ if (!map)
return FALSE;
+
+ while (results_base != qbuf->results_end) {
+ query->ops->add_result(rctx, query, map + results_base,
+ result);
+ results_base += query->result_size;
}
}
LIST_FOR_EACH_ENTRY(query, query_list, list) {
/* begin + end */
- num_dw += query->num_cs_dw * 2;
+ num_dw += query->num_cs_dw_begin + query->num_cs_dw_end;
/* Workaround for the fact that
* num_cs_dw_nontimer_queries_suspend is incremented for every
* resumed query, which raises the bar in need_cs_space for
* queries about to be resumed.
*/
- num_dw += query->num_cs_dw;
+ num_dw += query->num_cs_dw_end;
}
/* primitives generated query */
num_dw += ctx->streamout.enable_atom.num_dw;
struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
unsigned num_queries = r600_get_num_queries(rscreen);
- if (!info)
- return num_queries;
+ if (!info) {
+ unsigned num_perfcounters =
+ r600_get_perfcounter_info(rscreen, 0, NULL);
+
+ return num_queries + num_perfcounters;
+ }
if (index >= num_queries)
- return 0;
+ return r600_get_perfcounter_info(rscreen, index - num_queries, info);
*info = r600_driver_query_list[index];
return 1;
}
+static int r600_get_driver_query_group_info(struct pipe_screen *screen,
+ unsigned index,
+ struct pipe_driver_query_group_info *info)
+{
+ struct r600_common_screen *rscreen = (struct r600_common_screen *)screen;
+
+ return r600_get_perfcounter_group_info(rscreen, index, info);
+}
+
void r600_query_init(struct r600_common_context *rctx)
{
rctx->b.create_query = r600_create_query;
+ rctx->b.create_batch_query = r600_create_batch_query;
rctx->b.destroy_query = r600_destroy_query;
rctx->b.begin_query = r600_begin_query;
rctx->b.end_query = r600_end_query;
void r600_init_screen_query_functions(struct r600_common_screen *rscreen)
{
rscreen->b.get_driver_query_info = r600_get_driver_query_info;
+ rscreen->b.get_driver_query_group_info = r600_get_driver_query_group_info;
}
projects / mesa.git / blobdiff