[Mesa-dev] [PATCH 2/2] r600g, radeonsi: add a driver query returning GPU load
Marek Olšák
maraeo at gmail.com
Mon Apr 27 07:46:28 PDT 2015
From: Marek Olšák <marek.olsak at amd.com>
---
src/gallium/drivers/radeon/Makefile.am | 4 +-
src/gallium/drivers/radeon/Makefile.sources | 1 +
src/gallium/drivers/radeon/r600_gpu_load.c | 141 ++++++++++++++++++++++
src/gallium/drivers/radeon/r600_pipe_common.c | 7 +-
src/gallium/drivers/radeon/r600_pipe_common.h | 13 ++
src/gallium/drivers/radeon/r600_query.c | 11 ++
src/gallium/winsys/radeon/drm/radeon_drm_winsys.c | 16 +++
src/gallium/winsys/radeon/drm/radeon_winsys.h | 3 +
8 files changed, 194 insertions(+), 2 deletions(-)
create mode 100644 src/gallium/drivers/radeon/r600_gpu_load.c
diff --git a/src/gallium/drivers/radeon/Makefile.am b/src/gallium/drivers/radeon/Makefile.am
index ef6b999..13d8976 100644
--- a/src/gallium/drivers/radeon/Makefile.am
+++ b/src/gallium/drivers/radeon/Makefile.am
@@ -4,7 +4,9 @@ include $(top_srcdir)/src/gallium/Automake.inc
AM_CFLAGS = \
$(GALLIUM_DRIVER_CFLAGS) \
- $(RADEON_CFLAGS)
+ $(RADEON_CFLAGS) \
+ -Wstrict-overflow=0
+# ^^ disable warnings about overflows (os_time_timeout)
noinst_LTLIBRARIES = libradeon.la
diff --git a/src/gallium/drivers/radeon/Makefile.sources b/src/gallium/drivers/radeon/Makefile.sources
index f2b70a1..469f6d1 100644
--- a/src/gallium/drivers/radeon/Makefile.sources
+++ b/src/gallium/drivers/radeon/Makefile.sources
@@ -3,6 +3,7 @@ C_SOURCES := \
r600_buffer_common.c \
r600_cs.h \
r600d_common.h \
+ r600_gpu_load.c \
r600_pipe_common.c \
r600_pipe_common.h \
r600_query.c \
diff --git a/src/gallium/drivers/radeon/r600_gpu_load.c b/src/gallium/drivers/radeon/r600_gpu_load.c
new file mode 100644
index 0000000..a653834
--- /dev/null
+++ b/src/gallium/drivers/radeon/r600_gpu_load.c
@@ -0,0 +1,141 @@
+/*
+ * Copyright 2015 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors: Marek Olšák <maraeo at gmail.com>
+ *
+ */
+
+/* The GPU load is measured as follows.
+ *
+ * There is a thread which samples the GRBM_STATUS register at a certain
+ * frequency and the "busy" or "idle" counter is incremented based on
+ * whether the GUI_ACTIVE bit is set or not.
+ *
+ * Then, the user can sample the counters twice and calculate the average
+ * GPU load between the two samples.
+ */
+
+#include "r600_pipe_common.h"
+#include "os/os_time.h"
+
+/* For good accuracy at 1000 fps or lower. This will be inaccurate for higher
+ * fps (there are too few samples per frame). */
+#define SAMPLES_PER_SEC 10000
+
+#define GRBM_STATUS 0x8010
+#define GUI_ACTIVE(x) (((x) >> 31) & 0x1)
+
+static bool r600_is_gpu_busy(struct r600_common_screen *rscreen)
+{
+ uint32_t value = 0;
+
+ rscreen->ws->read_registers(rscreen->ws, GRBM_STATUS, 1, &value);
+ return GUI_ACTIVE(value);
+}
+
+static PIPE_THREAD_ROUTINE(r600_gpu_load_thread, param)
+{
+ struct r600_common_screen *rscreen = (struct r600_common_screen*)param;
+ const int period_us = 1000000 / SAMPLES_PER_SEC;
+ int sleep_us = period_us;
+ int64_t cur_time, last_time = os_time_get();
+
+ while (!p_atomic_read(&rscreen->gpu_load_stop_thread)) {
+ if (sleep_us)
+ os_time_sleep(sleep_us);
+
+ /* Make sure we sleep the ideal amount of time to match
+ * the expected frequency. */
+ cur_time = os_time_get();
+
+ if (os_time_timeout(last_time, last_time + period_us,
+ cur_time))
+ sleep_us = MAX2(sleep_us - 1, 1);
+ else
+ sleep_us += 1;
+
+ /*printf("Hz: %.1f\n", 1000000.0 / (cur_time - last_time));*/
+ last_time = cur_time;
+
+ /* Update the counters. */
+ if (r600_is_gpu_busy(rscreen))
+ p_atomic_inc(&rscreen->gpu_load_counter_busy);
+ else
+ p_atomic_inc(&rscreen->gpu_load_counter_idle);
+ }
+ p_atomic_dec(&rscreen->gpu_load_stop_thread);
+ return 0;
+}
+
+void r600_gpu_load_kill_thread(struct r600_common_screen *rscreen)
+{
+ if (!rscreen->gpu_load_thread)
+ return;
+
+ p_atomic_inc(&rscreen->gpu_load_stop_thread);
+ pipe_thread_wait(rscreen->gpu_load_thread);
+ rscreen->gpu_load_thread = 0;
+}
+
+static uint64_t r600_gpu_load_read_counter(struct r600_common_screen *rscreen)
+{
+ /* Start the thread if needed. */
+ if (!rscreen->gpu_load_thread) {
+ pipe_mutex_lock(rscreen->gpu_load_mutex);
+ /* Check again inside the mutex. */
+ if (!rscreen->gpu_load_thread)
+ rscreen->gpu_load_thread =
+ pipe_thread_create(r600_gpu_load_thread, rscreen);
+ pipe_mutex_unlock(rscreen->gpu_load_mutex);
+ }
+
+ /* The busy counter is in the lower 32 bits.
+ * The idle counter is in the upper 32 bits. */
+ return p_atomic_read(&rscreen->gpu_load_counter_busy) |
+ ((uint64_t)p_atomic_read(&rscreen->gpu_load_counter_idle) << 32);
+}
+
+/**
+ * Just return the counters.
+ */
+uint64_t r600_gpu_load_begin(struct r600_common_screen *rscreen)
+{
+ return r600_gpu_load_read_counter(rscreen);
+}
+
+unsigned r600_gpu_load_end(struct r600_common_screen *rscreen, uint64_t begin)
+{
+ uint64_t end = r600_gpu_load_read_counter(rscreen);
+ unsigned busy = (end & 0xffffffff) - (begin & 0xffffffff);
+ unsigned idle = (end >> 32) - (begin >> 32);
+
+ /* Calculate the GPU load.
+ *
+ * If no counters have been incremented, return the current load.
+ * It's for the case when the load is queried faster than
+ * the counters are updated.
+ */
+ if (idle || busy)
+ return busy*100 / (busy + idle);
+ else
+ return r600_is_gpu_busy(rscreen) ? 100 : 0;
+}
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.c b/src/gallium/drivers/radeon/r600_pipe_common.c
index 97eed13..eddb37d 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.c
+++ b/src/gallium/drivers/radeon/r600_pipe_common.c
@@ -667,7 +667,8 @@ static int r600_get_driver_query_info(struct pipe_screen *screen,
{"GTT-usage", R600_QUERY_GTT_USAGE, rscreen->info.gart_size, TRUE},
{"temperature", R600_QUERY_GPU_TEMPERATURE, 100, FALSE},
{"shader-clock", R600_QUERY_CURRENT_GPU_SCLK, 0, FALSE},
- {"memory-clock", R600_QUERY_CURRENT_GPU_MCLK, 0, FALSE}
+ {"memory-clock", R600_QUERY_CURRENT_GPU_MCLK, 0, FALSE},
+ {"GPU-load", R600_QUERY_GPU_LOAD, 100, FALSE}
};
unsigned num_queries;
@@ -872,6 +873,7 @@ bool r600_common_screen_init(struct r600_common_screen *rscreen,
}
util_format_s3tc_init();
pipe_mutex_init(rscreen->aux_context_lock);
+ pipe_mutex_init(rscreen->gpu_load_mutex);
if (rscreen->info.drm_minor >= 28 && (rscreen->debug_flags & DBG_TRACE_CS)) {
rscreen->trace_bo = (struct r600_resource*)pipe_buffer_create(&rscreen->b,
@@ -915,6 +917,9 @@ bool r600_common_screen_init(struct r600_common_screen *rscreen,
void r600_destroy_common_screen(struct r600_common_screen *rscreen)
{
+ r600_gpu_load_kill_thread(rscreen);
+
+ pipe_mutex_destroy(rscreen->gpu_load_mutex);
pipe_mutex_destroy(rscreen->aux_context_lock);
rscreen->aux_context->destroy(rscreen->aux_context);
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h b/src/gallium/drivers/radeon/r600_pipe_common.h
index c23072c..faa6e0d 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -58,6 +58,7 @@
#define R600_QUERY_GPU_TEMPERATURE (PIPE_QUERY_DRIVER_SPECIFIC + 8)
#define R600_QUERY_CURRENT_GPU_SCLK (PIPE_QUERY_DRIVER_SPECIFIC + 9)
#define R600_QUERY_CURRENT_GPU_MCLK (PIPE_QUERY_DRIVER_SPECIFIC + 10)
+#define R600_QUERY_GPU_LOAD (PIPE_QUERY_DRIVER_SPECIFIC + 11)
#define R600_CONTEXT_STREAMOUT_FLUSH (1u << 0)
#define R600_CONTEXT_PRIVATE_FLAG (1u << 1)
@@ -283,6 +284,13 @@ struct r600_common_screen {
struct r600_resource *trace_bo;
uint32_t *trace_ptr;
unsigned cs_count;
+
+ /* GPU load thread. */
+ pipe_mutex gpu_load_mutex;
+ pipe_thread gpu_load_thread;
+ unsigned gpu_load_counter_busy;
+ unsigned gpu_load_counter_idle;
+ unsigned gpu_load_stop_thread; /* bool */
};
/* This encapsulates a state or an operation which can emitted into the GPU
@@ -478,6 +486,11 @@ struct pipe_resource *r600_resource_create_common(struct pipe_screen *screen,
const char *r600_get_llvm_processor_name(enum radeon_family family);
void r600_need_dma_space(struct r600_common_context *ctx, unsigned num_dw);
+/* r600_gpu_load.c */
+void r600_gpu_load_kill_thread(struct r600_common_screen *rscreen);
+uint64_t r600_gpu_load_begin(struct r600_common_screen *rscreen);
+unsigned r600_gpu_load_end(struct r600_common_screen *rscreen, uint64_t begin);
+
/* r600_query.c */
void r600_query_init(struct r600_common_context *rctx);
void r600_suspend_nontimer_queries(struct r600_common_context *ctx);
diff --git a/src/gallium/drivers/radeon/r600_query.c b/src/gallium/drivers/radeon/r600_query.c
index 758064a..1335087 100644
--- a/src/gallium/drivers/radeon/r600_query.c
+++ b/src/gallium/drivers/radeon/r600_query.c
@@ -89,6 +89,7 @@ static struct r600_resource *r600_new_query_buffer(struct r600_common_context *c
case R600_QUERY_GPU_TEMPERATURE:
case R600_QUERY_CURRENT_GPU_SCLK:
case R600_QUERY_CURRENT_GPU_MCLK:
+ case R600_QUERY_GPU_LOAD:
return NULL;
}
@@ -388,6 +389,7 @@ static struct pipe_query *r600_create_query(struct pipe_context *ctx, unsigned q
case R600_QUERY_GPU_TEMPERATURE:
case R600_QUERY_CURRENT_GPU_SCLK:
case R600_QUERY_CURRENT_GPU_MCLK:
+ case R600_QUERY_GPU_LOAD:
skip_allocation = true;
break;
default:
@@ -459,6 +461,9 @@ static void r600_begin_query(struct pipe_context *ctx, struct pipe_query *query)
case R600_QUERY_NUM_BYTES_MOVED:
rquery->begin_result = rctx->ws->query_value(rctx->ws, RADEON_NUM_BYTES_MOVED);
return;
+ case R600_QUERY_GPU_LOAD:
+ rquery->begin_result = r600_gpu_load_begin(rctx->screen);
+ return;
}
/* Discard the old query buffers. */
@@ -531,6 +536,9 @@ static void r600_end_query(struct pipe_context *ctx, struct pipe_query *query)
case R600_QUERY_CURRENT_GPU_MCLK:
rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_CURRENT_MCLK) * 1000000;
return;
+ case R600_QUERY_GPU_LOAD:
+ rquery->end_result = r600_gpu_load_end(rctx->screen, rquery->begin_result);
+ return;
}
r600_emit_query_end(rctx, rquery);
@@ -593,6 +601,9 @@ static boolean r600_get_query_buffer_result(struct r600_common_context *ctx,
case R600_QUERY_CURRENT_GPU_MCLK:
result->u64 = query->end_result - query->begin_result;
return TRUE;
+ case R600_QUERY_GPU_LOAD:
+ result->u64 = query->end_result;
+ return TRUE;
}
map = r600_buffer_map_sync_with_rings(ctx, qbuf->buf,
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
index b3e3cb2..a6f847f 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
@@ -582,6 +582,21 @@ static uint64_t radeon_query_value(struct radeon_winsys *rws,
return 0;
}
+static void radeon_read_registers(struct radeon_winsys *rws,
+ unsigned reg_offset,
+ unsigned num_registers, uint32_t *out)
+{
+ struct radeon_drm_winsys *ws = (struct radeon_drm_winsys*)rws;
+ unsigned i;
+
+ for (i = 0; i < num_registers; i++) {
+ uint32_t reg = reg_offset + i*4;
+
+ radeon_get_drm_value(ws->fd, RADEON_INFO_READ_REG, "read-reg", ®);
+ out[i] = reg;
+ }
+}
+
static unsigned hash_fd(void *key)
{
int fd = pointer_to_intptr(key);
@@ -728,6 +743,7 @@ radeon_drm_winsys_create(int fd, radeon_screen_create_t screen_create)
ws->base.surface_init = radeon_drm_winsys_surface_init;
ws->base.surface_best = radeon_drm_winsys_surface_best;
ws->base.query_value = radeon_query_value;
+ ws->base.read_registers = radeon_read_registers;
radeon_bomgr_init_functions(ws);
radeon_drm_cs_init_functions(ws);
diff --git a/src/gallium/winsys/radeon/drm/radeon_winsys.h b/src/gallium/winsys/radeon/drm/radeon_winsys.h
index 996a201..ee0a904 100644
--- a/src/gallium/winsys/radeon/drm/radeon_winsys.h
+++ b/src/gallium/winsys/radeon/drm/radeon_winsys.h
@@ -589,6 +589,9 @@ struct radeon_winsys {
uint64_t (*query_value)(struct radeon_winsys *ws,
enum radeon_value_id value);
+
+ void (*read_registers)(struct radeon_winsys *ws, unsigned reg_offset,
+ unsigned num_registers, uint32_t *out);
};
--
2.1.0
More information about the mesa-dev
mailing list