radeonsi: move query suspend logic into the top-level si_query struct

Reviewed-by: Marek Olšák <marek.olsak@amd.com>

diff --git a/src/gallium/drivers/radeonsi/si_perfcounter.c b/src/gallium/drivers/radeonsi/si_perfcounter.c
index 69e149c76b618d448cec2bda4af71c619b72bbe8..0b3d8f892731624bfbd12a0e8912e15eabad7110 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_perfcounter.c
+++ b/src/gallium/drivers/radeonsi/si_perfcounter.c
@@ -868,7 +868,10 @@ static struct si_query_ops batch_query_ops = {
        .destroy = si_pc_query_destroy,
        .begin = si_query_hw_begin,
        .end = si_query_hw_end,
-       .get_result = si_query_hw_get_result
+       .get_result = si_query_hw_get_result,
+
+       .suspend = si_query_hw_suspend,
+       .resume = si_query_hw_resume,
 };
 
 static struct si_query_hw_ops batch_query_hw_ops = {
@@ -1001,8 +1004,8 @@ struct pipe_query *si_create_batch_query(struct pipe_context *ctx,
        }
 
        /* Compute result bases and CS size per group */
-       query->b.num_cs_dw_end = pc->num_stop_cs_dwords;
-       query->b.num_cs_dw_end += pc->num_instance_cs_dwords;
+       query->b.b.num_cs_dw_suspend = pc->num_stop_cs_dwords;
+       query->b.b.num_cs_dw_suspend += pc->num_instance_cs_dwords;
 
        i = 0;
        for (group = query->groups; group; group = group->next) {
@@ -1020,8 +1023,8 @@ struct pipe_query *si_create_batch_query(struct pipe_context *ctx,
                i += instances * group->num_counters;
 
                read_dw = 6 * group->num_counters;
-               query->b.num_cs_dw_end += instances * read_dw;
-               query->b.num_cs_dw_end += instances * pc->num_instance_cs_dwords;
+               query->b.b.num_cs_dw_suspend += instances * read_dw;
+               query->b.b.num_cs_dw_suspend += instances * pc->num_instance_cs_dwords;
        }
 
        if (query->shaders) {

diff --git a/src/gallium/drivers/radeonsi/si_query.c b/src/gallium/drivers/radeonsi/si_query.c
index aed3e1e80c1086bb0f896dc87ee51eaf4b2b965f..479a1bbf2c4372f47c2778abd45724615dacc53c 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_query.c
+++ b/src/gallium/drivers/radeonsi/si_query.c
@@ -34,6 +34,8 @@
 
 #define SI_MAX_STREAMS 4
 
+static struct si_query_ops query_hw_ops;
+
 struct si_hw_query_params {
        unsigned start_offset;
        unsigned end_offset;
@@ -607,14 +609,6 @@ static void si_query_hw_get_result_resource(struct si_context *sctx,
                                            struct pipe_resource *resource,
                                            unsigned offset);
 
-static struct si_query_ops query_hw_ops = {
-       .destroy = si_query_hw_destroy,
-       .begin = si_query_hw_begin,
-       .end = si_query_hw_end,
-       .get_result = si_query_hw_get_result,
-       .get_result_resource = si_query_hw_get_result_resource,
-};
-
 static void si_query_hw_do_emit_start(struct si_context *sctx,
                                      struct si_query_hw *query,
                                      struct r600_resource *buffer,
@@ -665,20 +659,19 @@ static struct pipe_query *si_query_hw_create(struct si_screen *sscreen,
        case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
                query->result_size = 16 * sscreen->info.num_render_backends;
                query->result_size += 16; /* for the fence + alignment */
-               query->num_cs_dw_end = 6 + si_cp_write_fence_dwords(sscreen);
+               query->b.num_cs_dw_suspend = 6 + si_cp_write_fence_dwords(sscreen);
                break;
        case SI_QUERY_TIME_ELAPSED_SDMA:
                /* GET_GLOBAL_TIMESTAMP only works if the offset is a multiple of 32. */
                query->result_size = 64;
-               query->num_cs_dw_end = 0;
                break;
        case PIPE_QUERY_TIME_ELAPSED:
                query->result_size = 24;
-               query->num_cs_dw_end = 8 + si_cp_write_fence_dwords(sscreen);
+               query->b.num_cs_dw_suspend = 8 + si_cp_write_fence_dwords(sscreen);
                break;
        case PIPE_QUERY_TIMESTAMP:
                query->result_size = 16;
-               query->num_cs_dw_end = 8 + si_cp_write_fence_dwords(sscreen);
+               query->b.num_cs_dw_suspend = 8 + si_cp_write_fence_dwords(sscreen);
                query->flags = SI_QUERY_HW_FLAG_NO_START;
                break;
        case PIPE_QUERY_PRIMITIVES_EMITTED:
@@ -687,19 +680,19 @@ static struct pipe_query *si_query_hw_create(struct si_screen *sscreen,
        case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
                /* NumPrimitivesWritten, PrimitiveStorageNeeded. */
                query->result_size = 32;
-               query->num_cs_dw_end = 6;
+               query->b.num_cs_dw_suspend = 6;
                query->stream = index;
                break;
        case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
                /* NumPrimitivesWritten, PrimitiveStorageNeeded. */
                query->result_size = 32 * SI_MAX_STREAMS;
-               query->num_cs_dw_end = 6 * SI_MAX_STREAMS;
+               query->b.num_cs_dw_suspend = 6 * SI_MAX_STREAMS;
                break;
        case PIPE_QUERY_PIPELINE_STATISTICS:
                /* 11 values on GCN. */
                query->result_size = 11 * 16;
                query->result_size += 8; /* for the fence + alignment */
-               query->num_cs_dw_end = 6 + si_cp_write_fence_dwords(sscreen);
+               query->b.num_cs_dw_suspend = 6 + si_cp_write_fence_dwords(sscreen);
                break;
        default:
                assert(0);
@@ -840,8 +833,6 @@ static void si_query_hw_emit_start(struct si_context *sctx,
        va = query->buffer.buf->gpu_address + query->buffer.results_end;
 
        query->ops->emit_start(sctx, query, query->buffer.buf, va);
-
-       sctx->num_cs_dw_queries_suspend += query->num_cs_dw_end;
 }
 
 static void si_query_hw_do_emit_stop(struct si_context *sctx,
@@ -935,9 +926,6 @@ static void si_query_hw_emit_stop(struct si_context *sctx,
 
        query->buffer.results_end += query->result_size;
 
-       if (!(query->flags & SI_QUERY_HW_FLAG_NO_START))
-               sctx->num_cs_dw_queries_suspend -= query->num_cs_dw_end;
-
        si_update_occlusion_query_state(sctx, query->b.type, -1);
        si_update_prims_generated_query_state(sctx, query->b.type, -1);
 }
@@ -1119,7 +1107,8 @@ bool si_query_hw_begin(struct si_context *sctx,
        if (!query->buffer.buf)
                return false;
 
-       LIST_ADDTAIL(&query->list, &sctx->active_queries);
+       LIST_ADDTAIL(&query->b.active_list, &sctx->active_queries);
+       sctx->num_cs_dw_queries_suspend += query->b.num_cs_dw_suspend;
        return true;
 }
 
@@ -1141,8 +1130,10 @@ bool si_query_hw_end(struct si_context *sctx,
 
        si_query_hw_emit_stop(sctx, query);
 
-       if (!(query->flags & SI_QUERY_HW_FLAG_NO_START))
-               LIST_DELINIT(&query->list);
+       if (!(query->flags & SI_QUERY_HW_FLAG_NO_START)) {
+               LIST_DELINIT(&query->b.active_list);
+               sctx->num_cs_dw_queries_suspend -= query->b.num_cs_dw_suspend;
+       }
 
        if (!query->buffer.buf)
                return false;
@@ -1349,6 +1340,27 @@ static void si_query_hw_add_result(struct si_screen *sscreen,
        }
 }
 
+void si_query_hw_suspend(struct si_context *sctx, struct si_query *query)
+{
+       si_query_hw_emit_stop(sctx, (struct si_query_hw *)query);
+}
+
+void si_query_hw_resume(struct si_context *sctx, struct si_query *query)
+{
+       si_query_hw_emit_start(sctx, (struct si_query_hw *)query);
+}
+
+static struct si_query_ops query_hw_ops = {
+       .destroy = si_query_hw_destroy,
+       .begin = si_query_hw_begin,
+       .end = si_query_hw_end,
+       .get_result = si_query_hw_get_result,
+       .get_result_resource = si_query_hw_get_result_resource,
+
+       .suspend = si_query_hw_suspend,
+       .resume = si_query_hw_resume,
+};
+
 static boolean si_get_query_result(struct pipe_context *ctx,
                                   struct pipe_query *query, boolean wait,
                                   union pipe_query_result *result)
@@ -1644,26 +1656,21 @@ static void si_render_condition(struct pipe_context *ctx,
 
 void si_suspend_queries(struct si_context *sctx)
 {
-       struct si_query_hw *query;
+       struct si_query *query;
 
-       LIST_FOR_EACH_ENTRY(query, &sctx->active_queries, list) {
-               si_query_hw_emit_stop(sctx, query);
-       }
-       assert(sctx->num_cs_dw_queries_suspend == 0);
+       LIST_FOR_EACH_ENTRY(query, &sctx->active_queries, active_list)
+               query->ops->suspend(sctx, query);
 }
 
 void si_resume_queries(struct si_context *sctx)
 {
-       struct si_query_hw *query;
-
-       assert(sctx->num_cs_dw_queries_suspend == 0);
+       struct si_query *query;
 
        /* Check CS space here. Resuming must not be interrupted by flushes. */
        si_need_gfx_cs_space(sctx);
 
-       LIST_FOR_EACH_ENTRY(query, &sctx->active_queries, list) {
-               si_query_hw_emit_start(sctx, query);
-       }
+       LIST_FOR_EACH_ENTRY(query, &sctx->active_queries, active_list)
+               query->ops->resume(sctx, query);
 }
 
 #define XFULL(name_, query_type_, type_, result_type_, group_id_) \

diff --git a/src/gallium/drivers/radeonsi/si_query.h b/src/gallium/drivers/radeonsi/si_query.h
index 032946edf4d00bb14c3212479c8606f1570bc3fc..ebd965a004f42d9d30518142b9d05ef944c40c36 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_query.h
+++ b/src/gallium/drivers/radeonsi/si_query.h
@@ -133,14 +133,23 @@ struct si_query_ops {
                                    int index,
                                    struct pipe_resource *resource,
                                    unsigned offset);
+
+       void (*suspend)(struct si_context *, struct si_query *);
+       void (*resume)(struct si_context *, struct si_query *);
 };
 
 struct si_query {
        struct threaded_query b;
        struct si_query_ops *ops;
 
-       /* The type of query */
+       /* The PIPE_QUERY_xxx type of query */
        unsigned type;
+
+       /* The number of dwords for suspend. */
+       unsigned num_cs_dw_suspend;
+
+       /* Linked list of queries that must be suspended at end of CS. */
+       struct list_head active_list;
 };
 
 enum {
@@ -187,10 +196,6 @@ struct si_query_hw {
        /* Size of the result in memory for both begin_query and end_query,
         * this can be one or two numbers, or it could even be a size of a structure. */
        unsigned result_size;
-       /* The number of dwords for end_query. */
-       unsigned num_cs_dw_end;
-       /* Linked list of queries */
-       struct list_head list;
        /* For transform feedback: which stream the query is for */
        unsigned stream;
 
@@ -211,6 +216,9 @@ bool si_query_hw_get_result(struct si_context *sctx,
                            struct si_query *rquery,
                            bool wait,
                            union pipe_query_result *result);
+void si_query_hw_suspend(struct si_context *sctx, struct si_query *query);
+void si_query_hw_resume(struct si_context *sctx, struct si_query *query);
+
 
 /* Performance counters */
 struct si_perfcounters {