radeonsi: implement AMD_performance_monitor for CIK+

Expose most of the performance counter groups that are exposed by Catalyst.
Ideally, the driver will work with GPUPerfStudio at some point, but we are not
quite there yet. In any case, this is the reason for grouping multiple
instances of hardware blocks in the way it is implemented.

The counters can also be shown using the Gallium HUD. If one is interested to
see how work is distributed across multiple shader engines, one can set the
environment variable RADEON_PC_SEPARATE_SE=1 to obtain finer-grained performance
counter groups.

Part of the implementation is in radeon because an implementation for
older hardware would largely follow along the same lines, but exposing
a different set of blocks which are programmed slightly differently.

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

1 files changed, 121 insertions, 0 deletions

diff --git a/src/gallium/drivers/radeon/r600_query.h b/src/gallium/drivers/radeon/r600_query.h
index 0ea5707..64ac916 100644
--- a/src/gallium/drivers/radeon/r600_query.h
+++ b/src/gallium/drivers/radeon/r600_query.h
@@ -31,7 +31,11 @@
 #include "pipe/p_defines.h"
 #include "util/list.h"
 
+struct pipe_context;
+struct pipe_query;
+
 struct r600_common_context;
+struct r600_common_screen;
 struct r600_query;
 struct r600_query_hw;
 struct r600_resource;
@@ -133,4 +137,121 @@ boolean r600_query_hw_get_result(struct r600_common_context *rctx,
 				 boolean wait,
 				 union pipe_query_result *result);
 
+/* Performance counters */
+enum {
+	/* This block is part of the shader engine */
+	R600_PC_BLOCK_SE = (1 << 0),
+
+	/* Expose per-instance groups instead of summing all instances (within
+	 * an SE). */
+	R600_PC_BLOCK_INSTANCE_GROUPS = (1 << 1),
+
+	/* Expose per-SE groups instead of summing instances across SEs. */
+	R600_PC_BLOCK_SE_GROUPS = (1 << 2),
+
+	/* Shader block */
+	R600_PC_BLOCK_SHADER = (1 << 3),
+
+	/* Non-shader block with perfcounters windowed by shaders. */
+	R600_PC_BLOCK_SHADER_WINDOWED = (1 << 4),
+};
+
+/* Shader enable bits. Chosen to coincide with SQ_PERFCOUNTER_CTRL values */
+enum {
+	R600_PC_SHADER_PS = (1 << 0),
+	R600_PC_SHADER_VS = (1 << 1),
+	R600_PC_SHADER_GS = (1 << 2),
+	R600_PC_SHADER_ES = (1 << 3),
+	R600_PC_SHADER_HS = (1 << 4),
+	R600_PC_SHADER_LS = (1 << 5),
+	R600_PC_SHADER_CS = (1 << 6),
+
+	R600_PC_SHADER_ALL = R600_PC_SHADER_PS | R600_PC_SHADER_VS |
+			     R600_PC_SHADER_GS | R600_PC_SHADER_ES |
+			     R600_PC_SHADER_HS | R600_PC_SHADER_LS |
+			     R600_PC_SHADER_CS,
+
+	R600_PC_SHADER_WINDOWING = (1 << 31),
+};
+
+/* Describes a hardware block with performance counters. Multiple instances of
+ * each block, possibly per-SE, may exist on the chip. Depending on the block
+ * and on the user's configuration, we either
+ *  (a) expose every instance as a performance counter group,
+ *  (b) expose a single performance counter group that reports the sum over all
+ *      instances, or
+ *  (c) expose one performance counter group per instance, but summed over all
+ *      shader engines.
+ */
+struct r600_perfcounter_block {
+	const char *basename;
+	unsigned flags;
+	unsigned num_counters;
+	unsigned num_selectors;
+	unsigned num_instances;
+
+	unsigned num_groups;
+	char *group_names;
+	unsigned group_name_stride;
+
+	char *selector_names;
+	unsigned selector_name_stride;
+
+	void *data;
+};
+
+struct r600_perfcounters {
+	unsigned num_groups;
+	unsigned num_blocks;
+	struct r600_perfcounter_block *blocks;
+
+	unsigned num_start_cs_dwords;
+	unsigned num_stop_cs_dwords;
+	unsigned num_instance_cs_dwords;
+	unsigned num_shaders_cs_dwords;
+
+	void (*get_size)(struct r600_perfcounter_block *,
+			 unsigned count, unsigned *selectors,
+			 unsigned *num_select_dw, unsigned *num_read_dw);
+
+	void (*emit_instance)(struct r600_common_context *,
+			      int se, int instance);
+	void (*emit_shaders)(struct r600_common_context *, unsigned shaders);
+	void (*emit_select)(struct r600_common_context *,
+			    struct r600_perfcounter_block *,
+			    unsigned count, unsigned *selectors);
+	void (*emit_start)(struct r600_common_context *,
+			  struct r600_resource *buffer, uint64_t va);
+	void (*emit_stop)(struct r600_common_context *,
+			  struct r600_resource *buffer, uint64_t va);
+	void (*emit_read)(struct r600_common_context *,
+			  struct r600_perfcounter_block *,
+			  unsigned count, unsigned *selectors,
+			  struct r600_resource *buffer, uint64_t va);
+
+	void (*cleanup)(struct r600_common_screen *);
+
+	boolean separate_se;
+	boolean separate_instance;
+};
+
+struct pipe_query *r600_create_batch_query(struct pipe_context *ctx,
+					   unsigned num_queries,
+					   unsigned *query_types);
+
+int r600_get_perfcounter_info(struct r600_common_screen *,
+			      unsigned index,
+			      struct pipe_driver_query_info *info);
+int r600_get_perfcounter_group_info(struct r600_common_screen *,
+				    unsigned index,
+				    struct pipe_driver_query_group_info *info);
+
+boolean r600_perfcounters_init(struct r600_perfcounters *, unsigned num_blocks);
+boolean r600_perfcounters_add_block(struct r600_common_screen *,
+				    struct r600_perfcounters *,
+				    const char *name, unsigned flags,
+				    unsigned counters, unsigned selectors,
+				    unsigned instances, void *data);
+void r600_perfcounters_do_destroy(struct r600_perfcounters *);
+
 #endif /* R600_QUERY_H */