[PATCH 18/22] drm/amd/display: move idle pipe allocation logic into dcn specific layer
Tom Chung
chiahsuan.chung at amd.com
Wed Aug 2 06:51:28 UTC 2023
From: Wenjing Liu <wenjing.liu at amd.com>
[why]
generic dc resource file should not know what an optimal idle pipe is
because this is dcn hardware dependent.
[how]
We move the optimial pipe searching logic in dcn specific layer.
Reviewed-by: Jun Lei <jun.lei at amd.com>
Acked-by: Tom Chung <chiahsuan.chung at amd.com>
Signed-off-by: Wenjing Liu <wenjing.liu at amd.com>
---
.../gpu/drm/amd/display/dc/core/dc_resource.c | 184 +++++++-----------
.../drm/amd/display/dc/dcn32/dcn32_resource.c | 87 ++++++++-
drivers/gpu/drm/amd/display/dc/inc/resource.h | 19 +-
3 files changed, 174 insertions(+), 116 deletions(-)
diff --git a/drivers/gpu/drm/amd/display/dc/core/dc_resource.c b/drivers/gpu/drm/amd/display/dc/core/dc_resource.c
index 218fe2c401e1..103dfe82dc28 100644
--- a/drivers/gpu/drm/amd/display/dc/core/dc_resource.c
+++ b/drivers/gpu/drm/amd/display/dc/core/dc_resource.c
@@ -1623,139 +1623,101 @@ struct pipe_ctx *find_idle_secondary_pipe_legacy(
return secondary_pipe;
}
-/*
- * Find the most optimal idle pipe from res_ctx, which could be used as a
- * secondary dpp pipe for input opp head pipe.
- *
- * an idle pipe - a pipe in input res_ctx not yet used for any streams or
- * planes.
- * secondary dpp pipe - a pipe gets inserted to a head OPP pipe's MPC blending
- * tree. This is typical used for rendering MPO planes or additional offset
- * areas in MPCC combine.
- *
- * Hardware Transition Minimization Algorithm for Finding a Secondary DPP Pipe
- * -------------------------------------------------------------------------
- *
- * PROBLEM:
- *
- * 1. There is a hardware limitation that a secondary DPP pipe cannot be
- * transferred from one MPC blending tree to the other in a single frame.
- * Otherwise it could cause glitches on the screen.
- *
- * For instance, we cannot transition from state 1 to state 2 in one frame. This
- * is because PIPE1 is transferred from PIPE0's MPC blending tree over to
- * PIPE2's MPC blending tree, which is not supported by hardware.
- * To support this transition we need to first remove PIPE1 from PIPE0's MPC
- * blending tree in one frame and then insert PIPE1 to PIPE2's MPC blending tree
- * in the next frame. This is not optimal as it will delay the flip for two
- * frames.
- *
- * State 1:
- * PIPE0 -- secondary DPP pipe --> (PIPE1)
- * PIPE2 -- secondary DPP pipe --> NONE
- *
- * State 2:
- * PIPE0 -- secondary DPP pipe --> NONE
- * PIPE2 -- secondary DPP pipe --> (PIPE1)
- *
- * 2. We want to in general minimize the unnecessary changes in pipe topology.
- * If a pipe is already added in current blending tree and there are no changes
- * to plane topology, we don't want to swap it with another idle pipe
- * unnecessarily in every update. Powering up and down a pipe would require a
- * full update which delays the flip for 1 frame. If we use the original pipe
- * we don't have to toggle its power. So we can flip faster.
- */
-struct pipe_ctx *find_optimal_idle_pipe_as_secondary_dpp_pipe(
+int resource_find_idle_pipe_used_in_cur_mpc_blending_tree(
const struct resource_context *cur_res_ctx,
struct resource_context *new_res_ctx,
- const struct resource_pool *pool,
- const struct pipe_ctx *new_head)
+ const struct pipe_ctx *cur_opp_head)
{
- const struct pipe_ctx *cur_head, *cur_sec;
- struct pipe_ctx *new_sec;
- bool found = false;
- int i;
-
- cur_head = &cur_res_ctx->pipe_ctx[new_head->pipe_idx];
- cur_sec = cur_head->bottom_pipe;
+ const struct pipe_ctx *cur_sec_dpp = cur_opp_head->bottom_pipe;
+ struct pipe_ctx *new_sec_dpp;
+ int idle_pipe_idx = IDLE_PIPE_INDEX_NOT_FOUND;
- while (cur_sec) {
+ while (cur_sec_dpp) {
/* find an idle pipe used in current opp blend tree,
* this is to avoid MPO pipe switching to different opp blending
* tree
*/
- new_sec = &new_res_ctx->pipe_ctx[cur_sec->pipe_idx];
- if (new_sec->plane_state == NULL && new_sec->stream == NULL) {
- new_sec->pipe_idx = cur_sec->pipe_idx;
- found = true;
+ new_sec_dpp = &new_res_ctx->pipe_ctx[cur_sec_dpp->pipe_idx];
+ if (new_sec_dpp->plane_state == NULL &&
+ new_sec_dpp->stream == NULL) {
+ idle_pipe_idx = cur_sec_dpp->pipe_idx;
break;
}
- cur_sec = cur_sec->bottom_pipe;
+ cur_sec_dpp = cur_sec_dpp->bottom_pipe;
}
- /* Up until here if we have not found an idle secondary pipe, we will
- * need to wait for at least one frame to complete the transition
- * sequence.
- */
- if (!found) {
- /* find a free pipe not used in current res ctx, this is to
- * avoid tearing down other pipe's topology
- */
- for (i = 0; i < pool->pipe_count; i++) {
- cur_sec = &cur_res_ctx->pipe_ctx[i];
- new_sec = &new_res_ctx->pipe_ctx[i];
-
- if (cur_sec->plane_state == NULL &&
- cur_sec->stream == NULL &&
- new_sec->plane_state == NULL &&
- new_sec->stream == NULL) {
- new_sec->pipe_idx = i;
- found = true;
- break;
- }
+ return idle_pipe_idx;
+}
+
+int recource_find_idle_pipe_not_used_in_cur_res_ctx(
+ const struct resource_context *cur_res_ctx,
+ struct resource_context *new_res_ctx,
+ const struct resource_pool *pool)
+{
+ int idle_pipe_idx = IDLE_PIPE_INDEX_NOT_FOUND;
+ const struct pipe_ctx *new_sec_dpp, *cur_sec_dpp;
+ int i;
+
+ for (i = 0; i < pool->pipe_count; i++) {
+ cur_sec_dpp = &cur_res_ctx->pipe_ctx[i];
+ new_sec_dpp = &new_res_ctx->pipe_ctx[i];
+
+ if (cur_sec_dpp->plane_state == NULL &&
+ cur_sec_dpp->stream == NULL &&
+ new_sec_dpp->plane_state == NULL &&
+ new_sec_dpp->stream == NULL) {
+ idle_pipe_idx = i;
+ break;
}
}
- /* Up until here if we have not found an idle secondary pipe, we will
- * need to wait for at least two frames to complete the transition
- * sequence. It really doesn't matter which pipe we decide take from
- * current enabled pipes. It won't save our frame time when we swap only
- * one pipe or more pipes.
- */
- if (!found) {
- /* find a free pipe by taking away a secondary dpp pipe from an
- * MPCC combine in current context
- */
- for (i = 0; i < pool->pipe_count; i++) {
- cur_sec = &cur_res_ctx->pipe_ctx[i];
- new_sec = &new_res_ctx->pipe_ctx[i];
-
- if (cur_sec->plane_state &&
- cur_sec->bottom_pipe &&
- cur_sec->bottom_pipe->plane_state == cur_sec->plane_state &&
- new_sec->plane_state == NULL &&
- new_sec->stream == NULL) {
- found = true;
- new_sec->pipe_idx = i;
- break;
- }
+ return idle_pipe_idx;
+}
+
+int resource_find_idle_pipe_used_as_cur_sec_dpp_in_mpcc_combine(
+ const struct resource_context *cur_res_ctx,
+ struct resource_context *new_res_ctx,
+ const struct resource_pool *pool)
+{
+ int idle_pipe_idx = IDLE_PIPE_INDEX_NOT_FOUND;
+ const struct pipe_ctx *new_sec_dpp, *cur_sec_dpp;
+ int i;
+
+ for (i = 0; i < pool->pipe_count; i++) {
+ cur_sec_dpp = &cur_res_ctx->pipe_ctx[i];
+ new_sec_dpp = &new_res_ctx->pipe_ctx[i];
+
+ if (cur_sec_dpp->plane_state &&
+ cur_sec_dpp->top_pipe &&
+ cur_sec_dpp->top_pipe->plane_state == cur_sec_dpp->plane_state &&
+ new_sec_dpp->plane_state == NULL &&
+ new_sec_dpp->stream == NULL) {
+ idle_pipe_idx = i;
+ break;
}
}
- if (!found) {
- /* find any pipe not used by new state */
- for (i = 0; i < pool->pipe_count; i++) {
- new_sec = &new_res_ctx->pipe_ctx[i];
+ return idle_pipe_idx;
+}
- if (new_sec->plane_state == NULL) {
- found = true;
- new_sec->pipe_idx = i;
- break;
- }
+int resource_find_any_idle_pipe(struct resource_context *new_res_ctx,
+ const struct resource_pool *pool)
+{
+ int idle_pipe_idx = IDLE_PIPE_INDEX_NOT_FOUND;
+ const struct pipe_ctx *new_sec_dpp;
+ int i;
+
+ for (i = 0; i < pool->pipe_count; i++) {
+ new_sec_dpp = &new_res_ctx->pipe_ctx[i];
+
+ if (new_sec_dpp->plane_state == NULL &&
+ new_sec_dpp->stream == NULL) {
+ idle_pipe_idx = i;
+ break;
}
}
- return found ? new_sec : NULL;
+ return idle_pipe_idx;
}
/* TODO: Unify the pipe naming convention:
diff --git a/drivers/gpu/drm/amd/display/dc/dcn32/dcn32_resource.c b/drivers/gpu/drm/amd/display/dc/dcn32/dcn32_resource.c
index fd6511da2903..2615a89d580b 100644
--- a/drivers/gpu/drm/amd/display/dc/dcn32/dcn32_resource.c
+++ b/drivers/gpu/drm/amd/display/dc/dcn32/dcn32_resource.c
@@ -2485,18 +2485,100 @@ struct resource_pool *dcn32_create_resource_pool(
return NULL;
}
+/*
+ * Find the most optimal idle pipe from res_ctx, which could be used as a
+ * secondary dpp pipe for input opp head pipe.
+ *
+ * an idle pipe - a pipe in input res_ctx not yet used for any streams or
+ * planes.
+ * secondary dpp pipe - a pipe gets inserted to a head OPP pipe's MPC blending
+ * tree. This is typical used for rendering MPO planes or additional offset
+ * areas in MPCC combine.
+ *
+ * Hardware Transition Minimization Algorithm for Finding a Secondary DPP Pipe
+ * -------------------------------------------------------------------------
+ *
+ * PROBLEM:
+ *
+ * 1. There is a hardware limitation that a secondary DPP pipe cannot be
+ * transferred from one MPC blending tree to the other in a single frame.
+ * Otherwise it could cause glitches on the screen.
+ *
+ * For instance, we cannot transition from state 1 to state 2 in one frame. This
+ * is because PIPE1 is transferred from PIPE0's MPC blending tree over to
+ * PIPE2's MPC blending tree, which is not supported by hardware.
+ * To support this transition we need to first remove PIPE1 from PIPE0's MPC
+ * blending tree in one frame and then insert PIPE1 to PIPE2's MPC blending tree
+ * in the next frame. This is not optimal as it will delay the flip for two
+ * frames.
+ *
+ * State 1:
+ * PIPE0 -- secondary DPP pipe --> (PIPE1)
+ * PIPE2 -- secondary DPP pipe --> NONE
+ *
+ * State 2:
+ * PIPE0 -- secondary DPP pipe --> NONE
+ * PIPE2 -- secondary DPP pipe --> (PIPE1)
+ *
+ * 2. We want to in general minimize the unnecessary changes in pipe topology.
+ * If a pipe is already added in current blending tree and there are no changes
+ * to plane topology, we don't want to swap it with another idle pipe
+ * unnecessarily in every update. Powering up and down a pipe would require a
+ * full update which delays the flip for 1 frame. If we use the original pipe
+ * we don't have to toggle its power. So we can flip faster.
+ */
+static int find_optimal_idle_pipe_as_secondary_dpp_pipe(
+ const struct resource_context *cur_res_ctx,
+ struct resource_context *new_res_ctx,
+ const struct resource_pool *pool,
+ const struct pipe_ctx *new_opp_head)
+{
+ const struct pipe_ctx *cur_opp_head;
+ int idle_pipe_idx;
+
+ cur_opp_head = &cur_res_ctx->pipe_ctx[new_opp_head->pipe_idx];
+ idle_pipe_idx = resource_find_idle_pipe_used_in_cur_mpc_blending_tree(
+ cur_res_ctx, new_res_ctx, cur_opp_head);
+
+ /* Up until here if we have not found an idle secondary pipe, we will
+ * need to wait for at least one frame to complete the transition
+ * sequence.
+ */
+ if (idle_pipe_idx == IDLE_PIPE_INDEX_NOT_FOUND)
+ idle_pipe_idx = recource_find_idle_pipe_not_used_in_cur_res_ctx(
+ cur_res_ctx, new_res_ctx, pool);
+
+ /* Up until here if we have not found an idle secondary pipe, we will
+ * need to wait for at least two frames to complete the transition
+ * sequence. It really doesn't matter which pipe we decide take from
+ * current enabled pipes. It won't save our frame time when we swap only
+ * one pipe or more pipes.
+ */
+ if (idle_pipe_idx == IDLE_PIPE_INDEX_NOT_FOUND)
+ idle_pipe_idx = resource_find_idle_pipe_used_as_cur_sec_dpp_in_mpcc_combine(
+ cur_res_ctx, new_res_ctx, pool);
+
+ if (idle_pipe_idx == IDLE_PIPE_INDEX_NOT_FOUND)
+ idle_pipe_idx = resource_find_any_idle_pipe(new_res_ctx, pool);
+
+ return idle_pipe_idx;
+}
+
struct pipe_ctx *dcn32_acquire_idle_pipe_for_layer(
const struct dc_state *cur_ctx,
struct dc_state *new_ctx,
const struct resource_pool *pool,
const struct pipe_ctx *opp_head_pipe)
{
- struct pipe_ctx *idle_pipe =
+ int idle_pipe_idx =
find_optimal_idle_pipe_as_secondary_dpp_pipe(
&cur_ctx->res_ctx, &new_ctx->res_ctx,
pool, opp_head_pipe);
+ struct pipe_ctx *idle_pipe;
- if (idle_pipe) {
+ if (idle_pipe_idx >= 0) {
+ idle_pipe = &new_ctx->res_ctx.pipe_ctx[idle_pipe_idx];
+ idle_pipe->pipe_idx = idle_pipe_idx;
idle_pipe->stream = opp_head_pipe->stream;
idle_pipe->stream_res.tg = opp_head_pipe->stream_res.tg;
idle_pipe->stream_res.opp = opp_head_pipe->stream_res.opp;
@@ -2508,6 +2590,7 @@ struct pipe_ctx *dcn32_acquire_idle_pipe_for_layer(
pool->dpps[idle_pipe->pipe_idx]->inst;
} else {
ASSERT(opp_head_pipe);
+ idle_pipe = NULL;
}
return idle_pipe;
diff --git a/drivers/gpu/drm/amd/display/dc/inc/resource.h b/drivers/gpu/drm/amd/display/dc/inc/resource.h
index 6ac6680711ff..c518ee8b1a03 100644
--- a/drivers/gpu/drm/amd/display/dc/inc/resource.h
+++ b/drivers/gpu/drm/amd/display/dc/inc/resource.h
@@ -37,6 +37,7 @@
#define IS_PIPE_SYNCD_VALID(pipe) ((((pipe)->pipe_idx_syncd) & 0x80)?1:0)
#define GET_PIPE_SYNCD_FROM_PIPE(pipe) ((pipe)->pipe_idx_syncd & 0x7F)
#define SET_PIPE_SYNCD_TO_PIPE(pipe, pipe_syncd) ((pipe)->pipe_idx_syncd = (0x80 | pipe_syncd))
+#define IDLE_PIPE_INDEX_NOT_FOUND -1
enum dce_version resource_parse_asic_id(
struct hw_asic_id asic_id);
@@ -158,11 +159,23 @@ struct pipe_ctx *find_idle_secondary_pipe_legacy(
const struct resource_pool *pool,
const struct pipe_ctx *primary_pipe);
-struct pipe_ctx *find_optimal_idle_pipe_as_secondary_dpp_pipe(
+int resource_find_idle_pipe_used_in_cur_mpc_blending_tree(
const struct resource_context *cur_res_ctx,
struct resource_context *new_res_ctx,
- const struct resource_pool *pool,
- const struct pipe_ctx *new_pri);
+ const struct pipe_ctx *cur_opp_head);
+
+int recource_find_idle_pipe_not_used_in_cur_res_ctx(
+ const struct resource_context *cur_res_ctx,
+ struct resource_context *new_res_ctx,
+ const struct resource_pool *pool);
+
+int resource_find_idle_pipe_used_as_cur_sec_dpp_in_mpcc_combine(
+ const struct resource_context *cur_res_ctx,
+ struct resource_context *new_res_ctx,
+ const struct resource_pool *pool);
+
+int resource_find_any_idle_pipe(struct resource_context *new_res_ctx,
+ const struct resource_pool *pool);
bool resource_validate_attach_surfaces(
const struct dc_validation_set set[],
--
2.25.1
More information about the amd-gfx
mailing list