[Mesa-dev] [PATCH 8/9] i965: Implement ARB_query_buffer_object for HSW+

[Kristian Høgsberg]

On Wed, Apr 27, 2016 at 12:13 PM, Kenneth Graunke <kenneth at whitecape.org> wrote:
> On Thursday, April 21, 2016 10:18:48 PM PDT Jordan Justen wrote:
> [snip]
>> diff --git a/src/mesa/drivers/dri/i965/brw_queryobj.c b/src/mesa/drivers/
> dri/i965/brw_queryobj.c
>> index a8e5aba..81ee5ea 100644
>> --- a/src/mesa/drivers/dri/i965/brw_queryobj.c
>> +++ b/src/mesa/drivers/dri/i965/brw_queryobj.c
>> @@ -462,7 +462,7 @@ brw_emit_query_end(struct brw_context *brw)
>>   * current GPU time.  This is unlike GL_TIME_ELAPSED, which measures the
>>   * time while the query is active.
>>   */
>> -static void
>> +void
>>  brw_query_counter(struct gl_context *ctx, struct gl_query_object *q)
>>  {
>>     struct brw_context *brw = brw_context(ctx);
>> @@ -507,12 +507,42 @@ brw_get_timestamp(struct gl_context *ctx)
>>     return result;
>>  }
>>
>
> /**
>  * Is this type of query written by PIPE_CONTROL?
>  */
>
>> +bool
>> +brw_is_query_pipelined(struct brw_query_object *query)
>> +{
>> +   switch (query->Base.Target) {
>> +   case GL_TIMESTAMP:
>> +   case GL_TIME_ELAPSED:
>> +   case GL_ANY_SAMPLES_PASSED:
>> +   case GL_ANY_SAMPLES_PASSED_CONSERVATIVE:
>> +   case GL_SAMPLES_PASSED_ARB:
>> +      return true;
>> +
>> +   case GL_PRIMITIVES_GENERATED:
>> +   case GL_TRANSFORM_FEEDBACK_PRIMITIVES_WRITTEN:
>> +   case GL_VERTICES_SUBMITTED_ARB:
>> +   case GL_PRIMITIVES_SUBMITTED_ARB:
>> +   case GL_VERTEX_SHADER_INVOCATIONS_ARB:
>> +   case GL_GEOMETRY_SHADER_INVOCATIONS:
>> +   case GL_GEOMETRY_SHADER_PRIMITIVES_EMITTED_ARB:
>> +   case GL_FRAGMENT_SHADER_INVOCATIONS_ARB:
>> +   case GL_CLIPPING_INPUT_PRIMITIVES_ARB:
>> +   case GL_CLIPPING_OUTPUT_PRIMITIVES_ARB:
>> +   case GL_COMPUTE_SHADER_INVOCATIONS_ARB:
>> +   case GL_TESS_CONTROL_SHADER_PATCHES_ARB:
>> +   case GL_TESS_EVALUATION_SHADER_INVOCATIONS_ARB:
>> +      return false;
>> +
>> +   default:
>> +      unreachable("Unrecognized query target in is_query_pipelined()");
>> +   }
>> +}
>> +
>>  /* Initialize query object functions used on all generations. */
>>  void brw_init_common_queryobj_functions(struct dd_function_table
> *functions)
>>  {
>>     functions->NewQueryObject = brw_new_query_object;
>>     functions->DeleteQuery = brw_delete_query;
>> -   functions->QueryCounter = brw_query_counter;
>>     functions->GetTimestamp = brw_get_timestamp;
>>  }
>>
>> @@ -523,4 +553,5 @@ void gen4_init_queryobj_functions(struct
> dd_function_table *functions)
>>     functions->EndQuery = brw_end_query;
>>     functions->CheckQuery = brw_check_query;
>>     functions->WaitQuery = brw_wait_query;
>> +   functions->QueryCounter = brw_query_counter;
>>  }
>> diff --git a/src/mesa/drivers/dri/i965/gen6_queryobj.c b/src/mesa/drivers/
> dri/i965/gen6_queryobj.c
>> index 960ccfd..f95c9fc 100644
>> --- a/src/mesa/drivers/dri/i965/gen6_queryobj.c
>> +++ b/src/mesa/drivers/dri/i965/gen6_queryobj.c
>> @@ -37,8 +37,25 @@
>>  #include "brw_defines.h"
>>  #include "brw_state.h"
>>  #include "intel_batchbuffer.h"
>> +#include "intel_buffer_objects.h"
>>  #include "intel_reg.h"
>>
>> +static inline void
>> +set_query_availability(struct brw_context *brw, struct brw_query_object
> *query,
>> +                       bool available)
>> +{
>> +   /* For ARB_query_buffer_object we write the query availability for
>> +    * pipelined results
>> +    */
>
> I could use some more comments here.  How about:
>
>    /* For platforms that support ARB_query_buffer_object, we write the
>     * query availability for "pipelined" queries.
>     *
>     * Most counter snapshots are written by the command streamer, by
>     * doing a CS stall and then MI_STORE_REGISTER_MEM.  For these
>     * counters, the CS stall guarantees that the results will be
>     * available when subsequent GPU commands run.  So we don't need to

Maybe say CS commands instead of GPU commands?

>     * do any additional tracking.
>     *
>     * Other counters (occlusion queries and timestamp) are written by
>     * PIPE_CONTROL, without a CS stall.  This means that we can't be
>     * sure whether the writes have landed yet or not.  Performing a
>     * PIPE_CONTROL with an immediate write will synchronize with
>     * those earlier writes, so we write 1 when the value has landed.
>     */
>
>> +   if (brw->ctx.Extensions.ARB_query_buffer_object &&
>> +       brw_is_query_pipelined(query)) {
>> +      brw_emit_pipe_control_write(brw,
>> +                                  PIPE_CONTROL_WRITE_IMMEDIATE,
>> +                                  query->bo, 2 * sizeof(uint64_t),
>> +                                  available, 0);
>> +   }
>> +}
>> +
>>  static void
>>  write_primitives_generated(struct brw_context *brw,
>>                             drm_intel_bo *query_bo, int stream, int idx)
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>