/*
* Copyright 2015 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* 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@gmail.com>
- *
*/
/* The GPU load is measured as follows.
* GPU load between the two samples.
*/
-#include "r600_pipe_common.h"
+#include "radeonsi/si_pipe.h"
#include "r600_query.h"
-#include "os/os_time.h"
+#include "util/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 SRBM_STATUS2 0x0e4c
#define SDMA_BUSY(x) (((x) >> 5) & 0x1)
+#define CP_STAT 0x8680
+#define PFP_BUSY(x) (((x) >> 15) & 0x1)
+#define MEQ_BUSY(x) (((x) >> 16) & 0x1)
+#define ME_BUSY(x) (((x) >> 17) & 0x1)
+#define SURFACE_SYNC_BUSY(x) (((x) >> 21) & 0x1)
+#define DMA_BUSY(x) (((x) >> 22) & 0x1)
+#define SCRATCH_RAM_BUSY(x) (((x) >> 24) & 0x1)
+
+#define IDENTITY(x) x
+
#define UPDATE_COUNTER(field, mask) \
do { \
if (mask(value)) \
p_atomic_inc(&counters->named.field.idle); \
} while (0)
-static void r600_update_mmio_counters(struct r600_common_screen *rscreen,
- union r600_mmio_counters *counters)
+static void si_update_mmio_counters(struct si_screen *sscreen,
+ union si_mmio_counters *counters)
{
uint32_t value = 0;
+ bool gui_busy, sdma_busy = false;
/* GRBM_STATUS */
- rscreen->ws->read_registers(rscreen->ws, GRBM_STATUS, 1, &value);
+ sscreen->ws->read_registers(sscreen->ws, GRBM_STATUS, 1, &value);
UPDATE_COUNTER(ta, TA_BUSY);
UPDATE_COUNTER(gds, GDS_BUSY);
UPDATE_COUNTER(cp, CP_BUSY);
UPDATE_COUNTER(cb, CB_BUSY);
UPDATE_COUNTER(gui, GUI_ACTIVE);
+ gui_busy = GUI_ACTIVE(value);
- if (rscreen->chip_class >= EVERGREEN) {
+ if (sscreen->info.chip_class == CIK || sscreen->info.chip_class == VI) {
/* SRBM_STATUS2 */
- rscreen->ws->read_registers(rscreen->ws, SRBM_STATUS2, 1, &value);
+ sscreen->ws->read_registers(sscreen->ws, SRBM_STATUS2, 1, &value);
UPDATE_COUNTER(sdma, SDMA_BUSY);
+ sdma_busy = SDMA_BUSY(value);
}
+
+ if (sscreen->info.chip_class >= VI) {
+ /* CP_STAT */
+ sscreen->ws->read_registers(sscreen->ws, CP_STAT, 1, &value);
+
+ UPDATE_COUNTER(pfp, PFP_BUSY);
+ UPDATE_COUNTER(meq, MEQ_BUSY);
+ UPDATE_COUNTER(me, ME_BUSY);
+ UPDATE_COUNTER(surf_sync, SURFACE_SYNC_BUSY);
+ UPDATE_COUNTER(cp_dma, DMA_BUSY);
+ UPDATE_COUNTER(scratch_ram, SCRATCH_RAM_BUSY);
+ }
+
+ value = gui_busy || sdma_busy;
+ UPDATE_COUNTER(gpu, IDENTITY);
}
#undef UPDATE_COUNTER
-static PIPE_THREAD_ROUTINE(r600_gpu_load_thread, param)
+static int
+si_gpu_load_thread(void *param)
{
- struct r600_common_screen *rscreen = (struct r600_common_screen*)param;
+ struct si_screen *sscreen = (struct si_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)) {
+ while (!p_atomic_read(&sscreen->gpu_load_stop_thread)) {
if (sleep_us)
os_time_sleep(sleep_us);
last_time = cur_time;
/* Update the counters. */
- r600_update_mmio_counters(rscreen, &rscreen->mmio_counters);
+ si_update_mmio_counters(sscreen, &sscreen->mmio_counters);
}
- p_atomic_dec(&rscreen->gpu_load_stop_thread);
+ p_atomic_dec(&sscreen->gpu_load_stop_thread);
return 0;
}
-void r600_gpu_load_kill_thread(struct r600_common_screen *rscreen)
+void si_gpu_load_kill_thread(struct si_screen *sscreen)
{
- if (!rscreen->gpu_load_thread)
+ if (!sscreen->gpu_load_thread)
return;
- p_atomic_inc(&rscreen->gpu_load_stop_thread);
- pipe_thread_wait(rscreen->gpu_load_thread);
- rscreen->gpu_load_thread = 0;
+ p_atomic_inc(&sscreen->gpu_load_stop_thread);
+ thrd_join(sscreen->gpu_load_thread, NULL);
+ sscreen->gpu_load_thread = 0;
}
-static uint64_t r600_read_mmio_counter(struct r600_common_screen *rscreen,
- unsigned busy_index)
+static uint64_t si_read_mmio_counter(struct si_screen *sscreen,
+ unsigned busy_index)
{
/* Start the thread if needed. */
- if (!rscreen->gpu_load_thread) {
- pipe_mutex_lock(rscreen->gpu_load_mutex);
+ if (!sscreen->gpu_load_thread) {
+ mtx_lock(&sscreen->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);
+ if (!sscreen->gpu_load_thread)
+ sscreen->gpu_load_thread =
+ u_thread_create(si_gpu_load_thread, sscreen);
+ mtx_unlock(&sscreen->gpu_load_mutex);
}
- unsigned busy = p_atomic_read(&rscreen->mmio_counters.array[busy_index]);
- unsigned idle = p_atomic_read(&rscreen->mmio_counters.array[busy_index + 1]);
+ unsigned busy = p_atomic_read(&sscreen->mmio_counters.array[busy_index]);
+ unsigned idle = p_atomic_read(&sscreen->mmio_counters.array[busy_index + 1]);
return busy | ((uint64_t)idle << 32);
}
-static unsigned r600_end_mmio_counter(struct r600_common_screen *rscreen,
- uint64_t begin, unsigned busy_index)
+static unsigned si_end_mmio_counter(struct si_screen *sscreen,
+ uint64_t begin, unsigned busy_index)
{
- uint64_t end = r600_read_mmio_counter(rscreen, busy_index);
+ uint64_t end = si_read_mmio_counter(sscreen, busy_index);
unsigned busy = (end & 0xffffffff) - (begin & 0xffffffff);
unsigned idle = (end >> 32) - (begin >> 32);
if (idle || busy) {
return busy*100 / (busy + idle);
} else {
- union r600_mmio_counters counters;
+ union si_mmio_counters counters;
memset(&counters, 0, sizeof(counters));
- r600_update_mmio_counters(rscreen, &counters);
+ si_update_mmio_counters(sscreen, &counters);
return counters.array[busy_index] ? 100 : 0;
}
}
#define BUSY_INDEX(rscreen, field) (&rscreen->mmio_counters.named.field.busy - \
rscreen->mmio_counters.array)
-static unsigned busy_index_from_type(struct r600_common_screen *rscreen,
+static unsigned busy_index_from_type(struct si_screen *sscreen,
unsigned type)
{
switch (type) {
- case R600_QUERY_GPU_LOAD:
- return BUSY_INDEX(rscreen, gui);
- case R600_QUERY_GPU_SHADERS_BUSY:
- return BUSY_INDEX(rscreen, spi);
- case R600_QUERY_GPU_TA_BUSY:
- return BUSY_INDEX(rscreen, ta);
- case R600_QUERY_GPU_GDS_BUSY:
- return BUSY_INDEX(rscreen, gds);
- case R600_QUERY_GPU_VGT_BUSY:
- return BUSY_INDEX(rscreen, vgt);
- case R600_QUERY_GPU_IA_BUSY:
- return BUSY_INDEX(rscreen, ia);
- case R600_QUERY_GPU_SX_BUSY:
- return BUSY_INDEX(rscreen, sx);
- case R600_QUERY_GPU_WD_BUSY:
- return BUSY_INDEX(rscreen, wd);
- case R600_QUERY_GPU_BCI_BUSY:
- return BUSY_INDEX(rscreen, bci);
- case R600_QUERY_GPU_SC_BUSY:
- return BUSY_INDEX(rscreen, sc);
- case R600_QUERY_GPU_PA_BUSY:
- return BUSY_INDEX(rscreen, pa);
- case R600_QUERY_GPU_DB_BUSY:
- return BUSY_INDEX(rscreen, db);
- case R600_QUERY_GPU_CP_BUSY:
- return BUSY_INDEX(rscreen, cp);
- case R600_QUERY_GPU_CB_BUSY:
- return BUSY_INDEX(rscreen, cb);
- case R600_QUERY_GPU_SDMA_BUSY:
- return BUSY_INDEX(rscreen, sdma);
+ case SI_QUERY_GPU_LOAD:
+ return BUSY_INDEX(sscreen, gpu);
+ case SI_QUERY_GPU_SHADERS_BUSY:
+ return BUSY_INDEX(sscreen, spi);
+ case SI_QUERY_GPU_TA_BUSY:
+ return BUSY_INDEX(sscreen, ta);
+ case SI_QUERY_GPU_GDS_BUSY:
+ return BUSY_INDEX(sscreen, gds);
+ case SI_QUERY_GPU_VGT_BUSY:
+ return BUSY_INDEX(sscreen, vgt);
+ case SI_QUERY_GPU_IA_BUSY:
+ return BUSY_INDEX(sscreen, ia);
+ case SI_QUERY_GPU_SX_BUSY:
+ return BUSY_INDEX(sscreen, sx);
+ case SI_QUERY_GPU_WD_BUSY:
+ return BUSY_INDEX(sscreen, wd);
+ case SI_QUERY_GPU_BCI_BUSY:
+ return BUSY_INDEX(sscreen, bci);
+ case SI_QUERY_GPU_SC_BUSY:
+ return BUSY_INDEX(sscreen, sc);
+ case SI_QUERY_GPU_PA_BUSY:
+ return BUSY_INDEX(sscreen, pa);
+ case SI_QUERY_GPU_DB_BUSY:
+ return BUSY_INDEX(sscreen, db);
+ case SI_QUERY_GPU_CP_BUSY:
+ return BUSY_INDEX(sscreen, cp);
+ case SI_QUERY_GPU_CB_BUSY:
+ return BUSY_INDEX(sscreen, cb);
+ case SI_QUERY_GPU_SDMA_BUSY:
+ return BUSY_INDEX(sscreen, sdma);
+ case SI_QUERY_GPU_PFP_BUSY:
+ return BUSY_INDEX(sscreen, pfp);
+ case SI_QUERY_GPU_MEQ_BUSY:
+ return BUSY_INDEX(sscreen, meq);
+ case SI_QUERY_GPU_ME_BUSY:
+ return BUSY_INDEX(sscreen, me);
+ case SI_QUERY_GPU_SURF_SYNC_BUSY:
+ return BUSY_INDEX(sscreen, surf_sync);
+ case SI_QUERY_GPU_CP_DMA_BUSY:
+ return BUSY_INDEX(sscreen, cp_dma);
+ case SI_QUERY_GPU_SCRATCH_RAM_BUSY:
+ return BUSY_INDEX(sscreen, scratch_ram);
default:
unreachable("invalid query type");
}
}
-uint64_t r600_begin_counter(struct r600_common_screen *rscreen, unsigned type)
+uint64_t si_begin_counter(struct si_screen *sscreen, unsigned type)
{
- unsigned busy_index = busy_index_from_type(rscreen, type);
- return r600_read_mmio_counter(rscreen, busy_index);
+ unsigned busy_index = busy_index_from_type(sscreen, type);
+ return si_read_mmio_counter(sscreen, busy_index);
}
-unsigned r600_end_counter(struct r600_common_screen *rscreen, unsigned type,
- uint64_t begin)
+unsigned si_end_counter(struct si_screen *sscreen, unsigned type,
+ uint64_t begin)
{
- unsigned busy_index = busy_index_from_type(rscreen, type);
- return r600_end_mmio_counter(rscreen, begin, busy_index);
+ unsigned busy_index = busy_index_from_type(sscreen, type);
+ return si_end_mmio_counter(sscreen, begin, busy_index);
}
projects / mesa.git / blobdiff