Mesa (master): ilo: rework query support
Chia-I Wu
olv at kemper.freedesktop.org
Sun Sep 21 16:06:18 UTC 2014
Module: Mesa
Branch: master
Commit: 9c873816a8ba0fe19fe6d38839078c1b75957953
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=9c873816a8ba0fe19fe6d38839078c1b75957953
Author: Chia-I Wu <olvaffe at gmail.com>
Date: Sat Sep 20 13:23:01 2014 +0800
ilo: rework query support
This fixes some corner cases, but more importantly, the new code should be
easier to reason about.
Signed-off-by: Chia-I Wu <olvaffe at gmail.com>
---
src/gallium/drivers/ilo/ilo_3d.c | 497 +++++++++++-------------
src/gallium/drivers/ilo/ilo_3d.h | 9 +-
src/gallium/drivers/ilo/ilo_3d_pipeline.c | 12 +-
src/gallium/drivers/ilo/ilo_3d_pipeline.h | 12 +-
src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.c | 30 +-
src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.h | 9 +-
src/gallium/drivers/ilo/ilo_query.c | 172 ++++----
src/gallium/drivers/ilo/ilo_query.h | 18 +-
8 files changed, 357 insertions(+), 402 deletions(-)
diff --git a/src/gallium/drivers/ilo/ilo_3d.c b/src/gallium/drivers/ilo/ilo_3d.c
index 30686d6..0a746a1 100644
--- a/src/gallium/drivers/ilo/ilo_3d.c
+++ b/src/gallium/drivers/ilo/ilo_3d.c
@@ -38,246 +38,142 @@
#include "ilo_3d.h"
static void
-process_query_for_occlusion_counter(struct ilo_3d *hw3d,
- struct ilo_query *q)
+ilo_3d_own_render_ring(struct ilo_3d *hw3d)
{
- uint64_t *vals, depth_count = 0;
- int i;
-
- /* in pairs */
- assert(q->reg_read % 2 == 0);
-
- vals = intel_bo_map(q->bo, false);
- for (i = 1; i < q->reg_read; i += 2)
- depth_count += vals[i] - vals[i - 1];
- intel_bo_unmap(q->bo);
-
- /* accumulate so that the query can be resumed if wanted */
- q->data.u64 += depth_count;
- q->reg_read = 0;
+ ilo_cp_set_owner(hw3d->cp, INTEL_RING_RENDER, &hw3d->owner);
}
static uint64_t
-timestamp_to_ns(uint64_t timestamp)
+query_timestamp_to_ns(const struct ilo_3d *hw3d, uint64_t timestamp)
{
/* see ilo_get_timestamp() */
return (timestamp & 0xffffffff) * 80;
}
+/**
+ * Process the bo and accumulate the result. The bo is emptied.
+ */
static void
-process_query_for_timestamp(struct ilo_3d *hw3d, struct ilo_query *q)
-{
- uint64_t *vals, timestamp;
-
- assert(q->reg_read == 1);
-
- vals = intel_bo_map(q->bo, false);
- timestamp = vals[0];
- intel_bo_unmap(q->bo);
-
- q->data.u64 = timestamp_to_ns(timestamp);
- q->reg_read = 0;
-}
-
-static void
-process_query_for_time_elapsed(struct ilo_3d *hw3d, struct ilo_query *q)
-{
- uint64_t *vals, elapsed = 0;
- int i;
-
- /* in pairs */
- assert(q->reg_read % 2 == 0);
-
- vals = intel_bo_map(q->bo, false);
-
- for (i = 1; i < q->reg_read; i += 2)
- elapsed += vals[i] - vals[i - 1];
-
- intel_bo_unmap(q->bo);
-
- /* accumulate so that the query can be resumed if wanted */
- q->data.u64 += timestamp_to_ns(elapsed);
- q->reg_read = 0;
-}
-
-static void
-process_query_for_pipeline_statistics(struct ilo_3d *hw3d,
- struct ilo_query *q)
+query_process_bo(const struct ilo_3d *hw3d, struct ilo_query *q)
{
const uint64_t *vals;
+ uint64_t tmp;
int i;
- assert(q->reg_read % 22 == 0);
+ if (!q->used)
+ return;
vals = intel_bo_map(q->bo, false);
-
- for (i = 0; i < q->reg_read; i += 22) {
- struct pipe_query_data_pipeline_statistics *stats =
- &q->data.pipeline_statistics;
- const uint64_t *begin = vals + i;
- const uint64_t *end = begin + 11;
-
- stats->ia_vertices += end[0] - begin[0];
- stats->ia_primitives += end[1] - begin[1];
- stats->vs_invocations += end[2] - begin[2];
- stats->gs_invocations += end[3] - begin[3];
- stats->gs_primitives += end[4] - begin[4];
- stats->c_invocations += end[5] - begin[5];
- stats->c_primitives += end[6] - begin[6];
- stats->ps_invocations += end[7] - begin[7];
- stats->hs_invocations += end[8] - begin[8];
- stats->ds_invocations += end[9] - begin[9];
- stats->cs_invocations += end[10] - begin[10];
+ if (!vals) {
+ q->used = 0;
+ return;
}
- intel_bo_unmap(q->bo);
-
- q->reg_read = 0;
-}
-
-static void
-ilo_3d_resume_queries(struct ilo_3d *hw3d)
-{
- struct ilo_query *q;
+ switch (q->type) {
+ case PIPE_QUERY_OCCLUSION_COUNTER:
+ case PIPE_QUERY_TIME_ELAPSED:
+ assert(q->stride == sizeof(*vals) * 2);
- /* resume occlusion queries */
- LIST_FOR_EACH_ENTRY(q, &hw3d->occlusion_queries, list) {
- /* accumulate the result if the bo is alreay full */
- if (q->reg_read >= q->reg_total)
- process_query_for_occlusion_counter(hw3d, q);
+ tmp = 0;
+ for (i = 0; i < q->used; i++)
+ tmp += vals[2 * i + 1] - vals[2 * i];
- ilo_3d_pipeline_emit_write_depth_count(hw3d->pipeline,
- q->bo, q->reg_read++);
- }
+ if (q->type == PIPE_QUERY_TIME_ELAPSED)
+ tmp = query_timestamp_to_ns(hw3d, tmp);
- /* resume timer queries */
- LIST_FOR_EACH_ENTRY(q, &hw3d->time_elapsed_queries, list) {
- /* accumulate the result if the bo is alreay full */
- if (q->reg_read >= q->reg_total)
- process_query_for_time_elapsed(hw3d, q);
+ q->result.u64 += tmp;
+ break;
+ case PIPE_QUERY_TIMESTAMP:
+ assert(q->stride == sizeof(*vals));
- ilo_3d_pipeline_emit_write_timestamp(hw3d->pipeline,
- q->bo, q->reg_read++);
+ q->result.u64 = query_timestamp_to_ns(hw3d, vals[q->used - 1]);
+ break;
+ case PIPE_QUERY_PIPELINE_STATISTICS:
+ assert(q->stride == sizeof(*vals) * 22);
+
+ for (i = 0; i < q->used; i++) {
+ struct pipe_query_data_pipeline_statistics *stats =
+ &q->result.pipeline_statistics;
+ const uint64_t *begin = vals + 22 * i;
+ const uint64_t *end = begin + 11;
+
+ stats->ia_vertices += end[0] - begin[0];
+ stats->ia_primitives += end[1] - begin[1];
+ stats->vs_invocations += end[2] - begin[2];
+ stats->gs_invocations += end[3] - begin[3];
+ stats->gs_primitives += end[4] - begin[4];
+ stats->c_invocations += end[5] - begin[5];
+ stats->c_primitives += end[6] - begin[6];
+ stats->ps_invocations += end[7] - begin[7];
+ stats->hs_invocations += end[8] - begin[8];
+ stats->ds_invocations += end[9] - begin[9];
+ stats->cs_invocations += end[10] - begin[10];
+ }
+ break;
+ default:
+ break;
}
- /* resume pipeline statistics queries */
- LIST_FOR_EACH_ENTRY(q, &hw3d->pipeline_statistics_queries, list) {
- /* accumulate the result if the bo is alreay full */
- if (q->reg_read >= q->reg_total)
- process_query_for_pipeline_statistics(hw3d, q);
+ intel_bo_unmap(q->bo);
- ilo_3d_pipeline_emit_write_statistics(hw3d->pipeline,
- q->bo, q->reg_read);
- q->reg_read += 11;
- }
+ q->used = 0;
}
static void
-ilo_3d_pause_queries(struct ilo_3d *hw3d)
+query_begin_bo(struct ilo_3d *hw3d, struct ilo_query *q)
{
- struct ilo_query *q;
+ uint32_t offset;
- /* pause occlusion queries */
- LIST_FOR_EACH_ENTRY(q, &hw3d->occlusion_queries, list) {
- assert(q->reg_read < q->reg_total);
- ilo_3d_pipeline_emit_write_depth_count(hw3d->pipeline,
- q->bo, q->reg_read++);
- }
+ /* bo is full */
+ if (q->used >= q->count)
+ query_process_bo(hw3d, q);
- /* pause timer queries */
- LIST_FOR_EACH_ENTRY(q, &hw3d->time_elapsed_queries, list) {
- assert(q->reg_read < q->reg_total);
- ilo_3d_pipeline_emit_write_timestamp(hw3d->pipeline,
- q->bo, q->reg_read++);
- }
+ offset = q->stride * q->used;
- /* pause pipeline statistics queries */
- LIST_FOR_EACH_ENTRY(q, &hw3d->pipeline_statistics_queries, list) {
- assert(q->reg_read < q->reg_total);
- ilo_3d_pipeline_emit_write_statistics(hw3d->pipeline,
- q->bo, q->reg_read);
- q->reg_read += 11;
+ /* write the beginning value to the bo */
+ switch (q->type) {
+ case PIPE_QUERY_OCCLUSION_COUNTER:
+ ilo_3d_pipeline_emit_write_depth_count(hw3d->pipeline, q->bo, offset);
+ break;
+ case PIPE_QUERY_TIMESTAMP:
+ /* no-op */
+ break;
+ case PIPE_QUERY_TIME_ELAPSED:
+ ilo_3d_pipeline_emit_write_timestamp(hw3d->pipeline, q->bo, offset);
+ break;
+ case PIPE_QUERY_PIPELINE_STATISTICS:
+ ilo_3d_pipeline_emit_write_statistics(hw3d->pipeline, q->bo, offset);
+ break;
+ default:
+ assert(!"unknown query type");
+ break;
}
}
static void
-ilo_3d_own_render_ring(struct ilo_3d *hw3d)
-{
- ilo_cp_set_owner(hw3d->cp, INTEL_RING_RENDER, &hw3d->owner);
-}
-
-static void
-ilo_3d_reserve_for_query(struct ilo_3d *hw3d, struct ilo_query *q,
- enum ilo_3d_pipeline_action act)
+query_end_bo(struct ilo_3d *hw3d, struct ilo_query *q)
{
- q->reg_cmd_size = ilo_3d_pipeline_estimate_size(hw3d->pipeline, act, NULL);
+ uint32_t offset;
- /* XXX we should check the aperture size */
- if (ilo_cp_space(hw3d->cp) < q->reg_cmd_size * 2) {
- ilo_cp_submit(hw3d->cp, "out of space");
- assert(ilo_cp_space(hw3d->cp) >= q->reg_cmd_size * 2);
- }
+ assert(q->used < q->count);
- /* reserve space for pausing the query */
- hw3d->owner.reserve += q->reg_cmd_size;
-}
+ offset = q->stride * q->used;
+ if (q->in_pairs)
+ offset += q->stride >> 1;
-/**
- * Begin a query.
- */
-void
-ilo_3d_begin_query(struct ilo_context *ilo, struct ilo_query *q)
-{
- struct ilo_3d *hw3d = ilo->hw3d;
-
- ilo_3d_own_render_ring(hw3d);
+ q->used++;
+ /* write the ending value to the bo */
switch (q->type) {
case PIPE_QUERY_OCCLUSION_COUNTER:
- ilo_3d_reserve_for_query(hw3d, q, ILO_3D_PIPELINE_WRITE_DEPTH_COUNT);
- q->data.u64 = 0;
-
- if (ilo_query_alloc_bo(q, 2, -1, hw3d->cp->winsys)) {
- ilo_3d_pipeline_emit_write_depth_count(hw3d->pipeline,
- q->bo, q->reg_read++);
-
- list_add(&q->list, &hw3d->occlusion_queries);
- }
+ ilo_3d_pipeline_emit_write_depth_count(hw3d->pipeline, q->bo, offset);
break;
case PIPE_QUERY_TIMESTAMP:
- /* nop */
- break;
case PIPE_QUERY_TIME_ELAPSED:
- ilo_3d_reserve_for_query(hw3d, q, ILO_3D_PIPELINE_WRITE_TIMESTAMP);
- q->data.u64 = 0;
-
- if (ilo_query_alloc_bo(q, 2, -1, hw3d->cp->winsys)) {
- ilo_3d_pipeline_emit_write_timestamp(hw3d->pipeline,
- q->bo, q->reg_read++);
-
- list_add(&q->list, &hw3d->time_elapsed_queries);
- }
- break;
- case PIPE_QUERY_PRIMITIVES_GENERATED:
- q->data.u64 = 0;
- list_add(&q->list, &hw3d->prim_generated_queries);
- break;
- case PIPE_QUERY_PRIMITIVES_EMITTED:
- q->data.u64 = 0;
- list_add(&q->list, &hw3d->prim_emitted_queries);
+ ilo_3d_pipeline_emit_write_timestamp(hw3d->pipeline, q->bo, offset);
break;
case PIPE_QUERY_PIPELINE_STATISTICS:
- ilo_3d_reserve_for_query(hw3d, q, ILO_3D_PIPELINE_WRITE_STATISTICS);
- memset(&q->data.pipeline_statistics, 0,
- sizeof(q->data.pipeline_statistics));
-
- if (ilo_query_alloc_bo(q, 11 * 2, -1, hw3d->cp->winsys)) {
- ilo_3d_pipeline_emit_write_statistics(hw3d->pipeline,
- q->bo, q->reg_read);
- q->reg_read += 11;
-
- list_add(&q->list, &hw3d->pipeline_statistics_queries);
- }
+ ilo_3d_pipeline_emit_write_statistics(hw3d->pipeline, q->bo, offset);
break;
default:
assert(!"unknown query type");
@@ -285,93 +181,103 @@ ilo_3d_begin_query(struct ilo_context *ilo, struct ilo_query *q)
}
}
-/**
- * End a query.
- */
-void
-ilo_3d_end_query(struct ilo_context *ilo, struct ilo_query *q)
+bool
+ilo_3d_init_query(struct pipe_context *pipe, struct ilo_query *q)
{
- struct ilo_3d *hw3d = ilo->hw3d;
-
- ilo_3d_own_render_ring(hw3d);
+ struct ilo_context *ilo = ilo_context(pipe);
+ unsigned bo_size;
switch (q->type) {
case PIPE_QUERY_OCCLUSION_COUNTER:
- list_del(&q->list);
-
- assert(q->reg_read < q->reg_total);
- assert(hw3d->owner.reserve >= q->reg_cmd_size);
- hw3d->owner.reserve -= q->reg_cmd_size;
-
- ilo_3d_pipeline_emit_write_depth_count(hw3d->pipeline,
- q->bo, q->reg_read++);
+ q->cmd_len = ilo_3d_pipeline_estimate_size(ilo->hw3d->pipeline,
+ ILO_3D_PIPELINE_WRITE_DEPTH_COUNT, q);
+ q->stride = sizeof(uint64_t);
+ q->in_pairs = true;
break;
case PIPE_QUERY_TIMESTAMP:
- q->data.u64 = 0;
-
- if (ilo_query_alloc_bo(q, 1, 1, hw3d->cp->winsys)) {
- ilo_3d_pipeline_emit_write_timestamp(hw3d->pipeline,
- q->bo, q->reg_read++);
- }
+ q->cmd_len = ilo_3d_pipeline_estimate_size(ilo->hw3d->pipeline,
+ ILO_3D_PIPELINE_WRITE_TIMESTAMP, q);
+ q->stride = sizeof(uint64_t);
+ q->in_pairs = false;
break;
case PIPE_QUERY_TIME_ELAPSED:
- list_del(&q->list);
-
- assert(q->reg_read < q->reg_total);
- assert(hw3d->owner.reserve >= q->reg_cmd_size);
- hw3d->owner.reserve -= q->reg_cmd_size;
-
- ilo_3d_pipeline_emit_write_timestamp(hw3d->pipeline,
- q->bo, q->reg_read++);
+ q->cmd_len = ilo_3d_pipeline_estimate_size(ilo->hw3d->pipeline,
+ ILO_3D_PIPELINE_WRITE_TIMESTAMP, q);
+ q->stride = sizeof(uint64_t);
+ q->in_pairs = true;
+ break;
+ case PIPE_QUERY_PIPELINE_STATISTICS:
+ q->cmd_len = ilo_3d_pipeline_estimate_size(ilo->hw3d->pipeline,
+ ILO_3D_PIPELINE_WRITE_STATISTICS, q);
+ q->stride = sizeof(uint64_t) * 11;
+ q->in_pairs = true;
break;
case PIPE_QUERY_PRIMITIVES_GENERATED:
case PIPE_QUERY_PRIMITIVES_EMITTED:
- list_del(&q->list);
- break;
- case PIPE_QUERY_PIPELINE_STATISTICS:
- list_del(&q->list);
-
- assert(q->reg_read + 11 <= q->reg_total);
- assert(hw3d->owner.reserve >= q->reg_cmd_size);
- hw3d->owner.reserve -= q->reg_cmd_size;
-
- ilo_3d_pipeline_emit_write_statistics(hw3d->pipeline,
- q->bo, q->reg_read);
- q->reg_read += 11;
+ return true;
break;
default:
- assert(!"unknown query type");
+ return false;
break;
}
+
+ /* double cmd_len and stride if in pairs */
+ q->cmd_len <<= q->in_pairs;
+ q->stride <<= q->in_pairs;
+
+ bo_size = (q->stride > 4096) ? q->stride : 4096;
+ q->bo = intel_winsys_alloc_buffer(ilo->winsys, "query", bo_size, false);
+ if (!q->bo)
+ return false;
+
+ q->count = bo_size / q->stride;
+
+ return true;
}
-/**
- * Process the raw query data.
- */
void
-ilo_3d_process_query(struct ilo_context *ilo, struct ilo_query *q)
+ilo_3d_begin_query(struct pipe_context *pipe, struct ilo_query *q)
{
- struct ilo_3d *hw3d = ilo->hw3d;
+ struct ilo_3d *hw3d = ilo_context(pipe)->hw3d;
+
+ ilo_3d_own_render_ring(hw3d);
+
+ if (q->bo) {
+ /* need to submit first */
+ if (!ilo_builder_validate(&hw3d->cp->builder, 1, &q->bo) ||
+ ilo_cp_space(hw3d->cp) < q->cmd_len) {
+ ilo_cp_submit(hw3d->cp, "out of aperture or space");
+
+ assert(ilo_builder_validate(&hw3d->cp->builder, 1, &q->bo));
+ assert(ilo_cp_space(hw3d->cp) >= q->cmd_len);
+
+ ilo_3d_own_render_ring(hw3d);
+ }
+
+ /* reserve the space for ending/pausing the query */
+ hw3d->owner.reserve += q->cmd_len >> q->in_pairs;
+
+ query_begin_bo(hw3d, q);
+ }
switch (q->type) {
case PIPE_QUERY_OCCLUSION_COUNTER:
- if (q->bo)
- process_query_for_occlusion_counter(hw3d, q);
+ list_add(&q->list, &hw3d->occlusion_queries);
break;
case PIPE_QUERY_TIMESTAMP:
- if (q->bo)
- process_query_for_timestamp(hw3d, q);
+ /* no-op */
break;
case PIPE_QUERY_TIME_ELAPSED:
- if (q->bo)
- process_query_for_time_elapsed(hw3d, q);
+ list_add(&q->list, &hw3d->time_elapsed_queries);
break;
case PIPE_QUERY_PRIMITIVES_GENERATED:
+ list_add(&q->list, &hw3d->prim_generated_queries);
+ break;
case PIPE_QUERY_PRIMITIVES_EMITTED:
+ list_add(&q->list, &hw3d->prim_emitted_queries);
break;
case PIPE_QUERY_PIPELINE_STATISTICS:
- if (q->bo)
- process_query_for_pipeline_statistics(hw3d, q);
+ list_add(&q->list, &hw3d->pipeline_statistics_queries);
break;
default:
assert(!"unknown query type");
@@ -379,19 +285,34 @@ ilo_3d_process_query(struct ilo_context *ilo, struct ilo_query *q)
}
}
+void
+ilo_3d_end_query(struct pipe_context *pipe, struct ilo_query *q)
+{
+ struct ilo_3d *hw3d = ilo_context(pipe)->hw3d;
+
+ ilo_3d_own_render_ring(hw3d);
+
+ if (q->bo) {
+ /* reclaim the reserved space */
+ hw3d->owner.reserve -= q->cmd_len >> q->in_pairs;
+ assert(hw3d->owner.reserve >= 0);
+
+ query_end_bo(hw3d, q);
+ }
+
+ list_delinit(&q->list);
+}
+
/**
- * Hook for CP new-batch.
+ * Process the raw query data.
*/
void
-ilo_3d_cp_submitted(struct ilo_3d *hw3d)
+ilo_3d_process_query(struct pipe_context *pipe, struct ilo_query *q)
{
- /* invalidate the pipeline */
- ilo_3d_pipeline_invalidate(hw3d->pipeline,
- ILO_3D_PIPELINE_INVALIDATE_BATCH_BO |
- ILO_3D_PIPELINE_INVALIDATE_STATE_BO |
- ILO_3D_PIPELINE_INVALIDATE_KERNEL_BO);
+ struct ilo_3d *hw3d = ilo_context(pipe)->hw3d;
- hw3d->new_batch = true;
+ if (q->bo)
+ query_process_bo(hw3d, q);
}
static void
@@ -405,7 +326,35 @@ ilo_3d_own_cp(struct ilo_cp *cp, void *data)
assert(ilo_cp_space(hw3d->cp) >= hw3d->owner.reserve * 2);
}
- ilo_3d_resume_queries(hw3d);
+ while (true) {
+ struct ilo_builder_snapshot snapshot;
+ struct ilo_query *q;
+
+ ilo_builder_batch_snapshot(&hw3d->cp->builder, &snapshot);
+
+ /* resume occlusion queries */
+ LIST_FOR_EACH_ENTRY(q, &hw3d->occlusion_queries, list)
+ query_begin_bo(hw3d, q);
+
+ /* resume timer queries */
+ LIST_FOR_EACH_ENTRY(q, &hw3d->time_elapsed_queries, list)
+ query_begin_bo(hw3d, q);
+
+ /* resume pipeline statistics queries */
+ LIST_FOR_EACH_ENTRY(q, &hw3d->pipeline_statistics_queries, list)
+ query_begin_bo(hw3d, q);
+
+ if (!ilo_builder_validate(&hw3d->cp->builder, 0, NULL)) {
+ ilo_builder_batch_restore(&hw3d->cp->builder, &snapshot);
+
+ if (ilo_builder_batch_used(&hw3d->cp->builder)) {
+ ilo_cp_submit(hw3d->cp, "out of aperture");
+ continue;
+ }
+ }
+
+ break;
+ }
assert(ilo_cp_space(hw3d->cp) >= hw3d->owner.reserve);
}
@@ -414,8 +363,36 @@ static void
ilo_3d_release_cp(struct ilo_cp *cp, void *data)
{
struct ilo_3d *hw3d = data;
+ struct ilo_query *q;
+
+ assert(ilo_cp_space(hw3d->cp) >= hw3d->owner.reserve);
+
+ /* pause occlusion queries */
+ LIST_FOR_EACH_ENTRY(q, &hw3d->occlusion_queries, list)
+ query_end_bo(hw3d, q);
+
+ /* pause timer queries */
+ LIST_FOR_EACH_ENTRY(q, &hw3d->time_elapsed_queries, list)
+ query_end_bo(hw3d, q);
+
+ /* pause pipeline statistics queries */
+ LIST_FOR_EACH_ENTRY(q, &hw3d->pipeline_statistics_queries, list)
+ query_end_bo(hw3d, q);
+}
+
+/**
+ * Hook for CP new-batch.
+ */
+void
+ilo_3d_cp_submitted(struct ilo_3d *hw3d)
+{
+ /* invalidate the pipeline */
+ ilo_3d_pipeline_invalidate(hw3d->pipeline,
+ ILO_3D_PIPELINE_INVALIDATE_BATCH_BO |
+ ILO_3D_PIPELINE_INVALIDATE_STATE_BO |
+ ILO_3D_PIPELINE_INVALIDATE_KERNEL_BO);
- ilo_3d_pause_queries(hw3d);
+ hw3d->new_batch = true;
}
/**
@@ -512,10 +489,10 @@ update_prim_count(struct ilo_3d *hw3d, int generated, int emitted)
struct ilo_query *q;
LIST_FOR_EACH_ENTRY(q, &hw3d->prim_generated_queries, list)
- q->data.u64 += generated;
+ q->result.u64 += generated;
LIST_FOR_EACH_ENTRY(q, &hw3d->prim_emitted_queries, list)
- q->data.u64 += emitted;
+ q->result.u64 += emitted;
}
bool
diff --git a/src/gallium/drivers/ilo/ilo_3d.h b/src/gallium/drivers/ilo/ilo_3d.h
index ff3a59e..9053dcd 100644
--- a/src/gallium/drivers/ilo/ilo_3d.h
+++ b/src/gallium/drivers/ilo/ilo_3d.h
@@ -72,14 +72,17 @@ ilo_3d_cp_submitted(struct ilo_3d *hw3d);
void
ilo_3d_draw_rectlist(struct ilo_3d *hw3d, const struct ilo_blitter *blitter);
+bool
+ilo_3d_init_query(struct pipe_context *pipe, struct ilo_query *q);
+
void
-ilo_3d_begin_query(struct ilo_context *ilo, struct ilo_query *q);
+ilo_3d_begin_query(struct pipe_context *pipe, struct ilo_query *q);
void
-ilo_3d_end_query(struct ilo_context *ilo, struct ilo_query *q);
+ilo_3d_end_query(struct pipe_context *pipe, struct ilo_query *q);
void
-ilo_3d_process_query(struct ilo_context *ilo, struct ilo_query *q);
+ilo_3d_process_query(struct pipe_context *pipe, struct ilo_query *q);
bool
ilo_3d_pass_render_condition(struct ilo_context *ilo);
diff --git a/src/gallium/drivers/ilo/ilo_3d_pipeline.c b/src/gallium/drivers/ilo/ilo_3d_pipeline.c
index 37e3332..87c6934 100644
--- a/src/gallium/drivers/ilo/ilo_3d_pipeline.c
+++ b/src/gallium/drivers/ilo/ilo_3d_pipeline.c
@@ -237,10 +237,10 @@ ilo_3d_pipeline_emit_flush(struct ilo_3d_pipeline *p)
*/
void
ilo_3d_pipeline_emit_write_timestamp(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index)
+ struct intel_bo *bo, uint32_t offset)
{
handle_invalid_batch_bo(p, true);
- p->emit_write_timestamp(p, bo, index);
+ p->emit_write_timestamp(p, bo, offset);
}
/**
@@ -248,10 +248,10 @@ ilo_3d_pipeline_emit_write_timestamp(struct ilo_3d_pipeline *p,
*/
void
ilo_3d_pipeline_emit_write_depth_count(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index)
+ struct intel_bo *bo, uint32_t offset)
{
handle_invalid_batch_bo(p, true);
- p->emit_write_depth_count(p, bo, index);
+ p->emit_write_depth_count(p, bo, offset);
}
/**
@@ -259,10 +259,10 @@ ilo_3d_pipeline_emit_write_depth_count(struct ilo_3d_pipeline *p,
*/
void
ilo_3d_pipeline_emit_write_statistics(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index)
+ struct intel_bo *bo, uint32_t offset)
{
handle_invalid_batch_bo(p, true);
- p->emit_write_statistics(p, bo, index);
+ p->emit_write_statistics(p, bo, offset);
}
void
diff --git a/src/gallium/drivers/ilo/ilo_3d_pipeline.h b/src/gallium/drivers/ilo/ilo_3d_pipeline.h
index 0bf4ee2..5342b4b 100644
--- a/src/gallium/drivers/ilo/ilo_3d_pipeline.h
+++ b/src/gallium/drivers/ilo/ilo_3d_pipeline.h
@@ -81,13 +81,13 @@ struct ilo_3d_pipeline {
void (*emit_flush)(struct ilo_3d_pipeline *pipeline);
void (*emit_write_timestamp)(struct ilo_3d_pipeline *pipeline,
- struct intel_bo *bo, int index);
+ struct intel_bo *bo, uint32_t offset);
void (*emit_write_depth_count)(struct ilo_3d_pipeline *pipeline,
- struct intel_bo *bo, int index);
+ struct intel_bo *bo, uint32_t offset);
void (*emit_write_statistics)(struct ilo_3d_pipeline *pipeline,
- struct intel_bo *bo, int index);
+ struct intel_bo *bo, uint32_t offset);
void (*emit_rectlist)(struct ilo_3d_pipeline *pipeline,
const struct ilo_blitter *blitter);
@@ -176,15 +176,15 @@ ilo_3d_pipeline_emit_flush(struct ilo_3d_pipeline *p);
void
ilo_3d_pipeline_emit_write_timestamp(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index);
+ struct intel_bo *bo, uint32_t offset);
void
ilo_3d_pipeline_emit_write_depth_count(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index);
+ struct intel_bo *bo, uint32_t offset);
void
ilo_3d_pipeline_emit_write_statistics(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index);
+ struct intel_bo *bo, uint32_t offset);
void
ilo_3d_pipeline_emit_rectlist(struct ilo_3d_pipeline *p,
diff --git a/src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.c b/src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.c
index 5deb772..383e9fe 100644
--- a/src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.c
+++ b/src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.c
@@ -1466,20 +1466,21 @@ ilo_3d_pipeline_emit_flush_gen6(struct ilo_3d_pipeline *p)
void
ilo_3d_pipeline_emit_write_timestamp_gen6(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index)
+ struct intel_bo *bo,
+ uint32_t offset)
{
if (ilo_dev_gen(p->dev) == ILO_GEN(6))
gen6_wa_pipe_control_post_sync(p, true);
gen6_PIPE_CONTROL(p->builder,
GEN6_PIPE_CONTROL_WRITE_TIMESTAMP,
- bo, index * sizeof(uint64_t),
- true);
+ bo, offset, true);
}
void
ilo_3d_pipeline_emit_write_depth_count_gen6(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index)
+ struct intel_bo *bo,
+ uint32_t offset)
{
if (ilo_dev_gen(p->dev) == ILO_GEN(6))
gen6_wa_pipe_control_post_sync(p, false);
@@ -1487,15 +1488,15 @@ ilo_3d_pipeline_emit_write_depth_count_gen6(struct ilo_3d_pipeline *p,
gen6_PIPE_CONTROL(p->builder,
GEN6_PIPE_CONTROL_DEPTH_STALL |
GEN6_PIPE_CONTROL_WRITE_PS_DEPTH_COUNT,
- bo, index * sizeof(uint64_t),
- true);
+ bo, offset, true);
}
void
ilo_3d_pipeline_emit_write_statistics_gen6(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index)
+ struct intel_bo *bo,
+ uint32_t offset)
{
- uint32_t regs[] = {
+ const uint32_t regs[] = {
GEN6_REG_IA_VERTICES_COUNT,
GEN6_REG_IA_PRIMITIVES_COUNT,
GEN6_REG_VS_INVOCATION_COUNT,
@@ -1513,20 +1514,17 @@ ilo_3d_pipeline_emit_write_statistics_gen6(struct ilo_3d_pipeline *p,
p->emit_flush(p);
for (i = 0; i < Elements(regs); i++) {
- const uint32_t bo_offset = (index + i) * sizeof(uint64_t);
-
if (regs[i]) {
/* store lower 32 bits */
- gen6_MI_STORE_REGISTER_MEM(p->builder,
- bo, bo_offset, regs[i]);
+ gen6_MI_STORE_REGISTER_MEM(p->builder, bo, offset, regs[i]);
/* store higher 32 bits */
- gen6_MI_STORE_REGISTER_MEM(p->builder,
- bo, bo_offset + 4, regs[i] + 4);
+ gen6_MI_STORE_REGISTER_MEM(p->builder, bo, offset + 4, regs[i] + 4);
}
else {
- gen6_MI_STORE_DATA_IMM(p->builder,
- bo, bo_offset, 0, true);
+ gen6_MI_STORE_DATA_IMM(p->builder, bo, offset, 0, true);
}
+
+ offset += 8;
}
}
diff --git a/src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.h b/src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.h
index 59892cb..6121a06 100644
--- a/src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.h
+++ b/src/gallium/drivers/ilo/ilo_3d_pipeline_gen6.h
@@ -158,15 +158,18 @@ ilo_3d_pipeline_emit_flush_gen6(struct ilo_3d_pipeline *p);
void
ilo_3d_pipeline_emit_write_timestamp_gen6(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index);
+ struct intel_bo *bo,
+ uint32_t offset);
void
ilo_3d_pipeline_emit_write_depth_count_gen6(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index);
+ struct intel_bo *bo,
+ uint32_t offset);
void
ilo_3d_pipeline_emit_write_statistics_gen6(struct ilo_3d_pipeline *p,
- struct intel_bo *bo, int index);
+ struct intel_bo *bo,
+ uint32_t offset);
void
ilo_3d_pipeline_init_gen6(struct ilo_3d_pipeline *p);
diff --git a/src/gallium/drivers/ilo/ilo_query.c b/src/gallium/drivers/ilo/ilo_query.c
index 812e056..e0cd4ee 100644
--- a/src/gallium/drivers/ilo/ilo_query.c
+++ b/src/gallium/drivers/ilo/ilo_query.c
@@ -33,31 +33,30 @@
#include "ilo_query.h"
static const struct {
- const char *name;
-
- void (*begin)(struct ilo_context *ilo, struct ilo_query *q);
- void (*end)(struct ilo_context *ilo, struct ilo_query *q);
- void (*process)(struct ilo_context *ilo, struct ilo_query *q);
-} query_info[PIPE_QUERY_TYPES] = {
-#define INFO(prefix, desc) { \
- .name = desc, \
- .begin = prefix ## _begin_query, \
- .end = prefix ## _end_query, \
- .process = prefix ## _process_query, \
+ bool (*init)(struct pipe_context *pipe, struct ilo_query *q);
+ void (*begin)(struct pipe_context *pipe, struct ilo_query *q);
+ void (*end)(struct pipe_context *pipe, struct ilo_query *q);
+ void (*process)(struct pipe_context *pipe, struct ilo_query *q);
+} ilo_query_table[PIPE_QUERY_TYPES] = {
+#define INFO(prefix) { \
+ .init = prefix ## _init_query, \
+ .begin = prefix ## _begin_query, \
+ .end = prefix ## _end_query, \
+ .process = prefix ## _process_query, \
}
-#define INFOX(prefix, desc) { desc, NULL, NULL, NULL, }
-
- [PIPE_QUERY_OCCLUSION_COUNTER] = INFO(ilo_3d, "occlusion counter"),
- [PIPE_QUERY_OCCLUSION_PREDICATE] = INFOX(ilo_3d, "occlusion pred."),
- [PIPE_QUERY_TIMESTAMP] = INFO(ilo_3d, "timestamp"),
- [PIPE_QUERY_TIMESTAMP_DISJOINT] = INFOX(ilo_3d, "timestamp disjoint"),
- [PIPE_QUERY_TIME_ELAPSED] = INFO(ilo_3d, "time elapsed"),
- [PIPE_QUERY_PRIMITIVES_GENERATED] = INFO(ilo_3d, "primitives generated"),
- [PIPE_QUERY_PRIMITIVES_EMITTED] = INFO(ilo_3d, "primitives emitted"),
- [PIPE_QUERY_SO_STATISTICS] = INFOX(ilo_3d, "so statistics"),
- [PIPE_QUERY_SO_OVERFLOW_PREDICATE] = INFOX(ilo_3d, "so overflow pred."),
- [PIPE_QUERY_GPU_FINISHED] = INFOX(ilo_3d, "gpu finished"),
- [PIPE_QUERY_PIPELINE_STATISTICS] = INFO(ilo_3d, "pipeline statistics"),
+#define INFOX(prefix) { NULL, NULL, NULL, NULL, }
+
+ [PIPE_QUERY_OCCLUSION_COUNTER] = INFO(ilo_3d),
+ [PIPE_QUERY_OCCLUSION_PREDICATE] = INFOX(ilo_3d),
+ [PIPE_QUERY_TIMESTAMP] = INFO(ilo_3d),
+ [PIPE_QUERY_TIMESTAMP_DISJOINT] = INFOX(ilo_3d),
+ [PIPE_QUERY_TIME_ELAPSED] = INFO(ilo_3d),
+ [PIPE_QUERY_PRIMITIVES_GENERATED] = INFO(ilo_3d),
+ [PIPE_QUERY_PRIMITIVES_EMITTED] = INFO(ilo_3d),
+ [PIPE_QUERY_SO_STATISTICS] = INFOX(ilo_3d),
+ [PIPE_QUERY_SO_OVERFLOW_PREDICATE] = INFOX(ilo_3d),
+ [PIPE_QUERY_GPU_FINISHED] = INFOX(ilo_3d),
+ [PIPE_QUERY_PIPELINE_STATISTICS] = INFO(ilo_3d),
#undef INFO
#undef INFOX
@@ -91,8 +90,15 @@ ilo_create_query(struct pipe_context *pipe, unsigned query_type, unsigned index)
return NULL;
q->type = query_type;
+ q->index = index;
+
list_inithead(&q->list);
+ if (!ilo_query_table[q->type].init(pipe, q)) {
+ FREE(q);
+ return NULL;
+ }
+
return (struct pipe_query *) q;
}
@@ -110,62 +116,71 @@ ilo_destroy_query(struct pipe_context *pipe, struct pipe_query *query)
static void
ilo_begin_query(struct pipe_context *pipe, struct pipe_query *query)
{
- struct ilo_context *ilo = ilo_context(pipe);
struct ilo_query *q = ilo_query(query);
+ if (q->active)
+ return;
+
+ util_query_clear_result(&q->result, q->type);
+ q->used = 0;
q->active = true;
- query_info[q->type].begin(ilo, q);
+ ilo_query_table[q->type].begin(pipe, q);
}
static void
ilo_end_query(struct pipe_context *pipe, struct pipe_query *query)
{
- struct ilo_context *ilo = ilo_context(pipe);
struct ilo_query *q = ilo_query(query);
- query_info[q->type].end(ilo, q);
+ if (!q->active) {
+ /* require ilo_begin_query() first */
+ if (q->in_pairs)
+ return;
+
+ ilo_begin_query(pipe, query);
+ }
- /*
- * some queries such as timestamp query does not require a call to
- * begin_query() so q->active is always false
- */
q->active = false;
+
+ ilo_query_table[q->type].end(pipe, q);
}
/**
- * The type (union pipe_query_result) indicates only the size of the buffer.
- * Callers expect the result to be "serialized".
+ * Serialize the result. The size of \p buf is
+ * sizeof(union pipe_query_result).
*/
static void
-serialize_query_data(unsigned type, const union pipe_query_result *data,
- void *buf)
+query_serialize(const struct ilo_query *q, void *buf)
{
- switch (type) {
+ switch (q->type) {
case PIPE_QUERY_OCCLUSION_COUNTER:
case PIPE_QUERY_TIMESTAMP:
case PIPE_QUERY_TIME_ELAPSED:
case PIPE_QUERY_PRIMITIVES_GENERATED:
case PIPE_QUERY_PRIMITIVES_EMITTED:
{
- uint64_t *r = buf;
- r[0] = data->u64;
+ uint64_t *dst = buf;
+ dst[0] = q->result.u64;
}
break;
case PIPE_QUERY_PIPELINE_STATISTICS:
{
- uint64_t *r = buf;
- r[0] = data->pipeline_statistics.ia_vertices;
- r[1] = data->pipeline_statistics.ia_primitives;
- r[2] = data->pipeline_statistics.vs_invocations;
- r[3] = data->pipeline_statistics.gs_invocations;
- r[4] = data->pipeline_statistics.gs_primitives;
- r[5] = data->pipeline_statistics.c_invocations;
- r[6] = data->pipeline_statistics.c_primitives;
- r[7] = data->pipeline_statistics.ps_invocations;
- r[8] = data->pipeline_statistics.hs_invocations;
- r[9] = data->pipeline_statistics.ds_invocations;
- r[10] = data->pipeline_statistics.cs_invocations;
+ const struct pipe_query_data_pipeline_statistics *stats =
+ &q->result.pipeline_statistics;
+ uint64_t *dst = buf;
+
+ dst[0] = stats->ia_vertices;
+ dst[1] = stats->ia_primitives;
+ dst[2] = stats->vs_invocations;
+ dst[3] = stats->gs_invocations;
+ dst[4] = stats->gs_primitives;
+ dst[5] = stats->c_invocations;
+ dst[6] = stats->c_primitives;
+ dst[7] = stats->ps_invocations;
+ dst[8] = stats->hs_invocations;
+ dst[9] = stats->ds_invocations;
+ dst[10] = stats->cs_invocations;
}
break;
default:
@@ -178,71 +193,30 @@ static boolean
ilo_get_query_result(struct pipe_context *pipe, struct pipe_query *query,
boolean wait, union pipe_query_result *result)
{
- struct ilo_context *ilo = ilo_context(pipe);
struct ilo_query *q = ilo_query(query);
if (q->active)
return false;
if (q->bo) {
- if (ilo_builder_has_reloc(&ilo->cp->builder, q->bo))
- ilo_cp_submit(ilo->cp, "syncing for queries");
+ struct ilo_cp *cp = ilo_context(pipe)->cp;
+
+ if (ilo_builder_has_reloc(&cp->builder, q->bo))
+ ilo_cp_submit(cp, "syncing for queries");
if (!wait && intel_bo_is_busy(q->bo))
return false;
-
- query_info[q->type].process(ilo, q);
}
+ ilo_query_table[q->type].process(pipe, q);
+
if (result)
- serialize_query_data(q->type, &q->data, (void *) result);
+ query_serialize(q, (void *) result);
return true;
}
/**
- * Allocate a query bo for reading hardware statistics.
- *
- * \param reg_count specifies how many registers need to be read.
- * \param repeat_count specifies how many times the registers are read. If
- * zero or negative, a 4KB bo is allocated.
- */
-bool
-ilo_query_alloc_bo(struct ilo_query *q, int reg_count, int repeat_count,
- struct intel_winsys *winsys)
-{
- const char *name;
- int reg_total;
-
- name = query_info[q->type].name;
-
- reg_total = reg_count * repeat_count;
- if (reg_total <= 0)
- reg_total = 4096 / sizeof(uint64_t);
-
- /* (re-)allocate the bo */
- if (q->reg_total < reg_total) {
- /* registers are 64-bit */
- const int size = reg_total * sizeof(uint64_t);
-
- if (q->bo)
- intel_bo_unreference(q->bo);
-
- q->bo = intel_winsys_alloc_buffer(winsys,
- name, size, false);
- q->reg_total = (q->bo) ? reg_total : 0;
- }
-
- /* avoid partial reads */
- if (reg_count)
- q->reg_total -= q->reg_total % reg_count;
-
- q->reg_read = 0;
-
- return (q->bo != NULL);
-}
-
-/**
* Initialize query-related functions.
*/
void
diff --git a/src/gallium/drivers/ilo/ilo_query.h b/src/gallium/drivers/ilo/ilo_query.h
index 7fedb2f..e451e4e 100644
--- a/src/gallium/drivers/ilo/ilo_query.h
+++ b/src/gallium/drivers/ilo/ilo_query.h
@@ -39,24 +39,24 @@ struct ilo_context;
*/
struct ilo_query {
unsigned type;
- bool active;
+ unsigned index;
struct list_head list;
- /* storage for the collected data */
- union pipe_query_result data;
+ bool active;
/* for queries that need to read hardware statistics */
+ int cmd_len; /* in dwords, as expected by ilo_cp */
+ bool in_pairs;
struct intel_bo *bo;
- int reg_read, reg_total;
- int reg_cmd_size; /* in dwords, as expected by ilo_cp */
+ int stride, count;
+ int used;
+
+ /* storage for the collected data */
+ union pipe_query_result result;
};
void
ilo_init_query_functions(struct ilo_context *ilo);
-bool
-ilo_query_alloc_bo(struct ilo_query *q, int reg_count, int repeat_count,
- struct intel_winsys *winsys);
-
#endif /* ILO_QUERY_H */
More information about the mesa-commit
mailing list