[Mesa-dev] [PATCH 3/4] gallium/radeon: add a HUD query for monitoring the CS thread activity

Marek Olšák maraeo at gmail.com
Mon Feb 13 15:57:46 UTC 2017


On Mon, Feb 13, 2017 at 4:37 PM, Nicolai Hähnle <nhaehnle at gmail.com> wrote:
> On 11.02.2017 20:58, Marek Olšák wrote:
>>
>> From: Marek Olšák <marek.olsak at amd.com>
>>
>> ---
>>  src/gallium/drivers/radeon/r600_query.c           | 25
>> ++++++++++++++++++++++-
>>  src/gallium/drivers/radeon/r600_query.h           |  1 +
>>  src/gallium/drivers/radeon/radeon_winsys.h        |  1 +
>>  src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c     |  2 ++
>>  src/gallium/winsys/radeon/drm/radeon_drm_winsys.c |  2 ++
>>  5 files changed, 30 insertions(+), 1 deletion(-)
>>
>> diff --git a/src/gallium/drivers/radeon/r600_query.c
>> b/src/gallium/drivers/radeon/r600_query.c
>> index 05741d3..8009416 100644
>> --- a/src/gallium/drivers/radeon/r600_query.c
>> +++ b/src/gallium/drivers/radeon/r600_query.c
>> @@ -19,37 +19,41 @@
>>   * THE AUTHOR(S) AND/OR THEIR SUPPLIERS 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.
>>   */
>>
>>  #include "r600_query.h"
>>  #include "r600_cs.h"
>>  #include "util/u_memory.h"
>>  #include "util/u_upload_mgr.h"
>> -
>> +#include "os/os_time.h"
>>  #include "tgsi/tgsi_text.h"
>>
>>  struct r600_hw_query_params {
>>         unsigned start_offset;
>>         unsigned end_offset;
>>         unsigned fence_offset;
>>         unsigned pair_stride;
>>         unsigned pair_count;
>>  };
>>
>>  /* Queries without buffer handling or suspend/resume. */
>>  struct r600_query_sw {
>>         struct r600_query b;
>>
>>         uint64_t begin_result;
>>         uint64_t end_result;
>> +
>> +       uint64_t begin_time;
>> +       uint64_t end_time;
>> +
>>         /* Fence for GPU_FINISHED. */
>>         struct pipe_fence_handle *fence;
>>  };
>>
>>  static void r600_query_sw_destroy(struct r600_common_context *rctx,
>>                                   struct r600_query *rquery)
>>  {
>>         struct pipe_screen *screen = rctx->b.screen;
>>         struct r600_query_sw *query = (struct r600_query_sw *)rquery;
>>
>> @@ -69,28 +73,30 @@ static enum radeon_value_id
>> winsys_id_from_type(unsigned type)
>>         case R600_QUERY_NUM_GFX_IBS: return RADEON_NUM_GFX_IBS;
>>         case R600_QUERY_NUM_SDMA_IBS: return RADEON_NUM_SDMA_IBS;
>>         case R600_QUERY_NUM_BYTES_MOVED: return RADEON_NUM_BYTES_MOVED;
>>         case R600_QUERY_NUM_EVICTIONS: return RADEON_NUM_EVICTIONS;
>>         case R600_QUERY_VRAM_USAGE: return RADEON_VRAM_USAGE;
>>         case R600_QUERY_VRAM_VIS_USAGE: return RADEON_VRAM_VIS_USAGE;
>>         case R600_QUERY_GTT_USAGE: return RADEON_GTT_USAGE;
>>         case R600_QUERY_GPU_TEMPERATURE: return RADEON_GPU_TEMPERATURE;
>>         case R600_QUERY_CURRENT_GPU_SCLK: return RADEON_CURRENT_SCLK;
>>         case R600_QUERY_CURRENT_GPU_MCLK: return RADEON_CURRENT_MCLK;
>> +       case R600_QUERY_CS_THREAD_BUSY: return RADEON_CS_THREAD_TIME;
>>         default: unreachable("query type does not correspond to winsys
>> id");
>>         }
>>  }
>>
>>  static bool r600_query_sw_begin(struct r600_common_context *rctx,
>>                                 struct r600_query *rquery)
>>  {
>>         struct r600_query_sw *query = (struct r600_query_sw *)rquery;
>> +       enum radeon_value_id ws_id;
>>
>>         switch(query->b.type) {
>>         case PIPE_QUERY_TIMESTAMP_DISJOINT:
>>         case PIPE_QUERY_GPU_FINISHED:
>>                 break;
>>         case R600_QUERY_DRAW_CALLS:
>>                 query->begin_result = rctx->num_draw_calls;
>>                 break;
>>         case R600_QUERY_SPILL_DRAW_CALLS:
>>                 query->begin_result = rctx->num_spill_draw_calls;
>> @@ -139,22 +145,28 @@ static bool r600_query_sw_begin(struct
>> r600_common_context *rctx,
>>         case R600_QUERY_NUM_MAPPED_BUFFERS:
>>                 query->begin_result = 0;
>>                 break;
>>         case R600_QUERY_BUFFER_WAIT_TIME:
>>         case R600_QUERY_NUM_GFX_IBS:
>>         case R600_QUERY_NUM_SDMA_IBS:
>>         case R600_QUERY_NUM_BYTES_MOVED:
>>         case R600_QUERY_NUM_EVICTIONS: {
>>                 enum radeon_value_id ws_id =
>> winsys_id_from_type(query->b.type);
>>                 query->begin_result = rctx->ws->query_value(rctx->ws,
>> ws_id);
>> +               query->begin_time = os_time_get_nano();
>>                 break;
>>         }
>> +       case R600_QUERY_CS_THREAD_BUSY:
>> +               ws_id = winsys_id_from_type(query->b.type);
>> +               query->begin_result = rctx->ws->query_value(rctx->ws,
>> ws_id);
>> +               query->begin_time = os_time_get_nano();
>> +               break;
>>         case R600_QUERY_GPU_LOAD:
>>         case R600_QUERY_GPU_SHADERS_BUSY:
>>         case R600_QUERY_GPU_TA_BUSY:
>>         case R600_QUERY_GPU_GDS_BUSY:
>>         case R600_QUERY_GPU_VGT_BUSY:
>>         case R600_QUERY_GPU_IA_BUSY:
>>         case R600_QUERY_GPU_SX_BUSY:
>>         case R600_QUERY_GPU_WD_BUSY:
>>         case R600_QUERY_GPU_BCI_BUSY:
>>         case R600_QUERY_GPU_SC_BUSY:
>> @@ -193,20 +205,21 @@ static bool r600_query_sw_begin(struct
>> r600_common_context *rctx,
>>                 unreachable("r600_query_sw_begin: bad query type");
>>         }
>>
>>         return true;
>>  }
>>
>>  static bool r600_query_sw_end(struct r600_common_context *rctx,
>>                               struct r600_query *rquery)
>>  {
>>         struct r600_query_sw *query = (struct r600_query_sw *)rquery;
>> +       enum radeon_value_id ws_id;
>>
>>         switch(query->b.type) {
>>         case PIPE_QUERY_TIMESTAMP_DISJOINT:
>>                 break;
>>         case PIPE_QUERY_GPU_FINISHED:
>>                 rctx->b.flush(&rctx->b, &query->fence,
>> PIPE_FLUSH_DEFERRED);
>>                 break;
>>         case R600_QUERY_DRAW_CALLS:
>>                 query->end_result = rctx->num_draw_calls;
>>                 break;
>> @@ -256,20 +269,25 @@ static bool r600_query_sw_end(struct
>> r600_common_context *rctx,
>>         case R600_QUERY_BUFFER_WAIT_TIME:
>>         case R600_QUERY_NUM_MAPPED_BUFFERS:
>>         case R600_QUERY_NUM_GFX_IBS:
>>         case R600_QUERY_NUM_SDMA_IBS:
>>         case R600_QUERY_NUM_BYTES_MOVED:
>>         case R600_QUERY_NUM_EVICTIONS: {
>>                 enum radeon_value_id ws_id =
>> winsys_id_from_type(query->b.type);
>>                 query->end_result = rctx->ws->query_value(rctx->ws,
>> ws_id);
>>                 break;
>>         }
>> +       case R600_QUERY_CS_THREAD_BUSY:
>> +               ws_id = winsys_id_from_type(query->b.type);
>> +               query->end_result = rctx->ws->query_value(rctx->ws,
>> ws_id);
>> +               query->end_time = os_time_get_nano();
>> +               break;
>>         case R600_QUERY_GPU_LOAD:
>>         case R600_QUERY_GPU_SHADERS_BUSY:
>>         case R600_QUERY_GPU_TA_BUSY:
>>         case R600_QUERY_GPU_GDS_BUSY:
>>         case R600_QUERY_GPU_VGT_BUSY:
>>         case R600_QUERY_GPU_IA_BUSY:
>>         case R600_QUERY_GPU_SX_BUSY:
>>         case R600_QUERY_GPU_WD_BUSY:
>>         case R600_QUERY_GPU_BCI_BUSY:
>>         case R600_QUERY_GPU_SC_BUSY:
>> @@ -330,20 +348,24 @@ static bool r600_query_sw_get_result(struct
>> r600_common_context *rctx,
>>                         (uint64_t)rctx->screen->info.clock_crystal_freq *
>> 1000;
>>                 result->timestamp_disjoint.disjoint = false;
>>                 return true;
>>         case PIPE_QUERY_GPU_FINISHED: {
>>                 struct pipe_screen *screen = rctx->b.screen;
>>                 result->b = screen->fence_finish(screen, &rctx->b,
>> query->fence,
>>                                                  wait ?
>> PIPE_TIMEOUT_INFINITE : 0);
>>                 return result->b;
>>         }
>>
>> +       case R600_QUERY_CS_THREAD_BUSY:
>> +               result->u64 = (query->end_result - query->begin_result) *
>> 100 /
>> +                             (query->end_time - query->begin_time);
>> +               return true;
>>         case R600_QUERY_GPIN_ASIC_ID:
>>                 result->u32 = 0;
>>                 return true;
>>         case R600_QUERY_GPIN_NUM_SIMD:
>>                 result->u32 = rctx->screen->info.num_good_compute_units;
>>                 return true;
>>         case R600_QUERY_GPIN_NUM_RB:
>>                 result->u32 = rctx->screen->info.num_render_backends;
>>                 return true;
>>         case R600_QUERY_GPIN_NUM_SPI:
>> @@ -1735,20 +1757,21 @@ static struct pipe_driver_query_info
>> r600_driver_query_list[] = {
>>         X("compute-calls",              COMPUTE_CALLS,          UINT64,
>> AVERAGE),
>>         X("spill-compute-calls",        SPILL_COMPUTE_CALLS,    UINT64,
>> AVERAGE),
>>         X("dma-calls",                  DMA_CALLS,              UINT64,
>> AVERAGE),
>>         X("cp-dma-calls",               CP_DMA_CALLS,           UINT64,
>> AVERAGE),
>>         X("num-vs-flushes",             NUM_VS_FLUSHES,         UINT64,
>> AVERAGE),
>>         X("num-ps-flushes",             NUM_PS_FLUSHES,         UINT64,
>> AVERAGE),
>>         X("num-cs-flushes",             NUM_CS_FLUSHES,         UINT64,
>> AVERAGE),
>>         X("num-fb-cache-flushes",       NUM_FB_CACHE_FLUSHES,   UINT64,
>> AVERAGE),
>>         X("num-L2-invalidates",         NUM_L2_INVALIDATES,     UINT64,
>> AVERAGE),
>>         X("num-L2-writebacks",          NUM_L2_WRITEBACKS,      UINT64,
>> AVERAGE),
>> +       X("CS-thread-busy",             CS_THREAD_BUSY,         UINT64,
>> AVERAGE),
>
>
> This should logically be PERCENTAGE instead of UINT64.
>
> I don't think the HUD handles that correctly today (according to p_defines,
> PERCENTAGE should be a float), but then we should either fix the HUD or nuke
> PERCENTAGE.

PERCENTAGE is used by nouveau for AMD_performance_monitor. We don't
have any other queries using PERCENTAGE. Certainly the HUD doesn't
handle it correctly, but that doesn't bother me much.

Marek


More information about the mesa-dev mailing list