[Mesa-dev] [PATCH 23/25] radeonsi: factor si_query_buffer logic out of si_query_hw

Timothy Arceri tarceri at itsqueeze.com
Wed Feb 6 00:27:13 UTC 2019


On 4/2/19 11:26 pm, Nicolai Hähnle wrote:
> On 01.02.19 05:25, Timothy Arceri wrote:
>> On 26/1/19 11:56 am, Marek Olšák wrote:
>>> Timothy, can you please test the attached fix?
>>
>> I'm having trouble compiling 32bit mesa on my machine at the moment so 
>> haven't been able to test Batman. But this commit also causes No Mans 
>> Sky to lock up my machine and the attached patch does not fix it.
> 
> Is there a trace or something else to easily reproduce it?

I've filed a bug report with all the info you should need to reproduce:

https://bugs.freedesktop.org/show_bug.cgi?id=109561

> 
> Cheers,
> Nicolai
> 
> 
> 
>>
>>>
>>> Thanks,
>>> Marek
>>>
>>> On Wed, Jan 2, 2019 at 10:58 PM Timothy Arceri <tarceri at itsqueeze.com 
>>> <mailto:tarceri at itsqueeze.com>> wrote:
>>>
>>>     This commit seems to cause bad stuttering in the Batman Arkham City
>>>     benchmark.
>>>
>>>     On 7/12/18 1:00 am, Nicolai Hähnle wrote:
>>>      > From: Nicolai Hähnle <nicolai.haehnle at amd.com
>>>     <mailto:nicolai.haehnle at amd.com>>
>>>      >
>>>      > This is a move towards using composition instead of 
>>> inheritance for
>>>      > different query types.
>>>      >
>>>      > This change weakens out-of-memory error reporting somewhat,
>>>     though this
>>>      > should be acceptable since we didn't consistently report such
>>>     errors in
>>>      > the first place.
>>>      > ---
>>>      >   src/gallium/drivers/radeonsi/si_perfcounter.c |   8 +-
>>>      >   src/gallium/drivers/radeonsi/si_query.c       | 177
>>>     +++++++++---------
>>>      >   src/gallium/drivers/radeonsi/si_query.h       |  17 +-
>>>      >   src/gallium/drivers/radeonsi/si_texture.c     |   7 +-
>>>      >   4 files changed, 99 insertions(+), 110 deletions(-)
>>>      >
>>>      > diff --git a/src/gallium/drivers/radeonsi/si_perfcounter.c
>>>     b/src/gallium/drivers/radeonsi/si_perfcounter.c
>>>      > index 0b3d8f89273..f0d10c054c4 100644
>>>      > --- a/src/gallium/drivers/radeonsi/si_perfcounter.c
>>>      > +++ b/src/gallium/drivers/radeonsi/si_perfcounter.c
>>>      > @@ -761,23 +761,22 @@ static void si_pc_query_destroy(struct
>>>     si_screen *sscreen,
>>>      >               struct si_query_group *group = query->groups;
>>>      >               query->groups = group->next;
>>>      >               FREE(group);
>>>      >       }
>>>      >
>>>      >       FREE(query->counters);
>>>      >
>>>      >       si_query_hw_destroy(sscreen, rquery);
>>>      >   }
>>>      >
>>>      > -static bool si_pc_query_prepare_buffer(struct si_screen *screen,
>>>      > -                                    struct si_query_hw *hwquery,
>>>      > -                                    struct r600_resource 
>>> *buffer)
>>>      > +static bool si_pc_query_prepare_buffer(struct si_context *ctx,
>>>      > +                                    struct si_query_buffer 
>>> *qbuf)
>>>      >   {
>>>      >       /* no-op */
>>>      >       return true;
>>>      >   }
>>>      >
>>>      >   static void si_pc_query_emit_start(struct si_context *sctx,
>>>      >                                  struct si_query_hw *hwquery,
>>>      >                                  struct r600_resource *buffer,
>>>     uint64_t va)
>>>      >   {
>>>      >       struct si_query_pc *query = (struct si_query_pc *)hwquery;
>>>      > @@ -1055,23 +1054,20 @@ struct pipe_query
>>>     *si_create_batch_query(struct pipe_context *ctx,
>>>      >               counter->base = group->result_base + j;
>>>      >               counter->stride = group->num_counters;
>>>      >
>>>      >               counter->qwords = 1;
>>>      >               if ((block->b->b->flags & SI_PC_BLOCK_SE) &&
>>>     group->se < 0)
>>>      >                       counter->qwords = screen->info.max_se;
>>>      >               if (group->instance < 0)
>>>      >                       counter->qwords *= block->num_instances;
>>>      >       }
>>>      >
>>>      > -     if (!si_query_hw_init(screen, &query->b))
>>>      > -             goto error;
>>>      > -
>>>      >       return (struct pipe_query *)query;
>>>      >
>>>      >   error:
>>>      >       si_pc_query_destroy(screen, &query->b.b);
>>>      >       return NULL;
>>>      >   }
>>>      >
>>>      >   static bool si_init_block_names(struct si_screen *screen,
>>>      >                               struct si_pc_block *block)
>>>      >   {
>>>      > diff --git a/src/gallium/drivers/radeonsi/si_query.c
>>>     b/src/gallium/drivers/radeonsi/si_query.c
>>>      > index 479a1bbf2c4..5b0fba0ed92 100644
>>>      > --- a/src/gallium/drivers/radeonsi/si_query.c
>>>      > +++ b/src/gallium/drivers/radeonsi/si_query.c
>>>      > @@ -514,86 +514,129 @@ static struct pipe_query
>>>     *si_query_sw_create(unsigned query_type)
>>>      >       query = CALLOC_STRUCT(si_query_sw);
>>>      >       if (!query)
>>>      >               return NULL;
>>>      >
>>>      >       query->b.type = query_type;
>>>      >       query->b.ops = &sw_query_ops;
>>>      >
>>>      >       return (struct pipe_query *)query;
>>>      >   }
>>>      >
>>>      > -void si_query_hw_destroy(struct si_screen *sscreen,
>>>      > -                      struct si_query *rquery)
>>>      > +void si_query_buffer_destroy(struct si_screen *sscreen, struct
>>>     si_query_buffer *buffer)
>>>      >   {
>>>      > -     struct si_query_hw *query = (struct si_query_hw *)rquery;
>>>      > -     struct si_query_buffer *prev = query->buffer.previous;
>>>      > +     struct si_query_buffer *prev = buffer->previous;
>>>      >
>>>      >       /* Release all query buffers. */
>>>      >       while (prev) {
>>>      >               struct si_query_buffer *qbuf = prev;
>>>      >               prev = prev->previous;
>>>      >               r600_resource_reference(&qbuf->buf, NULL);
>>>      >               FREE(qbuf);
>>>      >       }
>>>      >
>>>      > -     r600_resource_reference(&query->buffer.buf, NULL);
>>>      > -     r600_resource_reference(&query->workaround_buf, NULL);
>>>      > -     FREE(rquery);
>>>      > +     r600_resource_reference(&buffer->buf, NULL);
>>>      > +}
>>>      > +
>>>      > +void si_query_buffer_reset(struct si_context *sctx, struct
>>>     si_query_buffer *buffer)
>>>      > +{
>>>      > +     /* Discard all query buffers except for the oldest. */
>>>      > +     while (buffer->previous) {
>>>      > +             struct si_query_buffer *qbuf = buffer->previous;
>>>      > +             buffer->previous = qbuf->previous;
>>>      > +
>>>      > +             r600_resource_reference(&buffer->buf, NULL);
>>>      > +             buffer->buf = qbuf->buf; /* move ownership */
>>>      > +             FREE(qbuf);
>>>      > +     }
>>>      > +     buffer->results_end = 0;
>>>      > +
>>>      > +     /* Discard even the oldest buffer if it can't be mapped
>>>     without a stall. */
>>>      > +     if (buffer->buf &&
>>>      > +         (si_rings_is_buffer_referenced(sctx, buffer->buf->buf,
>>>     RADEON_USAGE_READWRITE) ||
>>>      > +          !sctx->ws->buffer_wait(buffer->buf->buf, 0,
>>>     RADEON_USAGE_READWRITE))) {
>>>      > +             r600_resource_reference(&buffer->buf, NULL);
>>>      > +     }
>>>      >   }
>>>      >
>>>      > -static struct r600_resource *si_new_query_buffer(struct
>>>     si_screen *sscreen,
>>>      > -                                              struct si_query_hw
>>>     *query)
>>>      > +bool si_query_buffer_alloc(struct si_context *sctx, struct
>>>     si_query_buffer *buffer,
>>>      > +                        bool (*prepare_buffer)(struct si_context
>>>     *, struct si_query_buffer*),
>>>      > +                        unsigned size)
>>>      >   {
>>>      > -     unsigned buf_size = MAX2(query->result_size,
>>>      > -                              sscreen->info.min_alloc_size);
>>>      > +     if (buffer->buf && buffer->results_end + size >=
>>>     buffer->buf->b.b.width0)
>>>      > +             return true;
>>>      > +
>>>      > +     if (buffer->buf) {
>>>      > +             struct si_query_buffer *qbuf =
>>>     MALLOC_STRUCT(si_query_buffer);
>>>      > +             memcpy(qbuf, buffer, sizeof(*qbuf));
>>>      > +             buffer->previous = qbuf;
>>>      > +     }
>>>      > +
>>>      > +     buffer->results_end = 0;
>>>      >
>>>      >       /* Queries are normally read by the CPU after
>>>      >        * being written by the gpu, hence staging is probably a 
>>> good
>>>      >        * usage pattern.
>>>      >        */
>>>      > -     struct r600_resource *buf = r600_resource(
>>>      > -             pipe_buffer_create(&sscreen->b, 0,
>>>      > -                                PIPE_USAGE_STAGING, buf_size));
>>>      > -     if (!buf)
>>>      > -             return NULL;
>>>      > +     struct si_screen *screen = sctx->screen;
>>>      > +     unsigned buf_size = MAX2(size, 
>>> screen->info.min_alloc_size);
>>>      > +     buffer->buf = r600_resource(
>>>      > +             pipe_buffer_create(&screen->b, 0,
>>>     PIPE_USAGE_STAGING, buf_size));
>>>      > +     if (unlikely(!buffer->buf))
>>>      > +             return false;
>>>      >
>>>      > -     if (!query->ops->prepare_buffer(sscreen, query, buf)) {
>>>      > -             r600_resource_reference(&buf, NULL);
>>>      > -             return NULL;
>>>      > +     if (prepare_buffer) {
>>>      > +             if (unlikely(!prepare_buffer(sctx, buffer))) {
>>>      > +                     r600_resource_reference(&buffer->buf, 
>>> NULL);
>>>      > +                     return false;
>>>      > +             }
>>>      >       }
>>>      >
>>>      > -     return buf;
>>>      > +     return true;
>>>      >   }
>>>      >
>>>      > -static bool si_query_hw_prepare_buffer(struct si_screen 
>>> *sscreen,
>>>      > -                                    struct si_query_hw *query,
>>>      > -                                    struct r600_resource 
>>> *buffer)
>>>      > +
>>>      > +void si_query_hw_destroy(struct si_screen *sscreen,
>>>      > +                      struct si_query *rquery)
>>>      > +{
>>>      > +     struct si_query_hw *query = (struct si_query_hw *)rquery;
>>>      > +
>>>      > +     si_query_buffer_destroy(sscreen, &query->buffer);
>>>      > +     r600_resource_reference(&query->workaround_buf, NULL);
>>>      > +     FREE(rquery);
>>>      > +}
>>>      > +
>>>      > +static bool si_query_hw_prepare_buffer(struct si_context *sctx,
>>>      > +                                    struct si_query_buffer 
>>> *qbuf)
>>>      >   {
>>>      > -     /* Callers ensure that the buffer is currently unused by
>>>     the GPU. */
>>>      > -     uint32_t *results = sscreen->ws->buffer_map(buffer->buf, 
>>> NULL,
>>>      > +     static const struct si_query_hw si_query_hw_s;
>>>      > +     struct si_query_hw *query = container_of(qbuf,
>>>     &si_query_hw_s, buffer);
>>>      > +     struct si_screen *screen = sctx->screen;
>>>      > +
>>>      > +     /* The caller ensures that the buffer is currently unused
>>>     by the GPU. */
>>>      > +     uint32_t *results = screen->ws->buffer_map(qbuf->buf->buf,
>>>     NULL,
>>>      >     PIPE_TRANSFER_WRITE |
>>>      >     PIPE_TRANSFER_UNSYNCHRONIZED);
>>>      >       if (!results)
>>>      >               return false;
>>>      >
>>>      > -     memset(results, 0, buffer->b.b.width0);
>>>      > +     memset(results, 0, qbuf->buf->b.b.width0);
>>>      >
>>>      >       if (query->b.type == PIPE_QUERY_OCCLUSION_COUNTER ||
>>>      >           query->b.type == PIPE_QUERY_OCCLUSION_PREDICATE ||
>>>      >           query->b.type ==
>>>     PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
>>>      > -             unsigned max_rbs = 
>>> sscreen->info.num_render_backends;
>>>      > -             unsigned enabled_rb_mask =
>>>     sscreen->info.enabled_rb_mask;
>>>      > +             unsigned max_rbs = 
>>> screen->info.num_render_backends;
>>>      > +             unsigned enabled_rb_mask =
>>>     screen->info.enabled_rb_mask;
>>>      >               unsigned num_results;
>>>      >               unsigned i, j;
>>>      >
>>>      >               /* Set top bits for unused backends. */
>>>      > -             num_results = buffer->b.b.width0 / 
>>> query->result_size;
>>>      > +             num_results = qbuf->buf->b.b.width0 /
>>>     query->result_size;
>>>      >               for (j = 0; j < num_results; j++) {
>>>      >                       for (i = 0; i < max_rbs; i++) {
>>>      >                               if (!(enabled_rb_mask & (1<<i))) {
>>>      >                                       results[(i * 4)+1] =
>>>     0x80000000;
>>>      >                                       results[(i * 4)+3] =
>>>     0x80000000;
>>>      >                               }
>>>      >                       }
>>>      >                       results += 4 * max_rbs;
>>>      >               }
>>>      >       }
>>>      > @@ -624,30 +667,20 @@ static void si_query_hw_clear_result(struct
>>>     si_query_hw *,
>>>      >                                    union pipe_query_result *);
>>>      >
>>>      >   static struct si_query_hw_ops query_hw_default_hw_ops = {
>>>      >       .prepare_buffer = si_query_hw_prepare_buffer,
>>>      >       .emit_start = si_query_hw_do_emit_start,
>>>      >       .emit_stop = si_query_hw_do_emit_stop,
>>>      >       .clear_result = si_query_hw_clear_result,
>>>      >       .add_result = si_query_hw_add_result,
>>>      >   };
>>>      >
>>>      > -bool si_query_hw_init(struct si_screen *sscreen,
>>>      > -                   struct si_query_hw *query)
>>>      > -{
>>>      > -     query->buffer.buf = si_new_query_buffer(sscreen, query);
>>>      > -     if (!query->buffer.buf)
>>>      > -             return false;
>>>      > -
>>>      > -     return true;
>>>      > -}
>>>      > -
>>>      >   static struct pipe_query *si_query_hw_create(struct si_screen
>>>     *sscreen,
>>>      >                                            unsigned query_type,
>>>      >                                            unsigned index)
>>>      >   {
>>>      >       struct si_query_hw *query = CALLOC_STRUCT(si_query_hw);
>>>      >       if (!query)
>>>      >               return NULL;
>>>      >
>>>      >       query->b.type = query_type;
>>>      >       query->b.ops = &query_hw_ops;
>>>      > @@ -693,25 +726,20 @@ static struct pipe_query
>>>     *si_query_hw_create(struct si_screen *sscreen,
>>>      >               query->result_size = 11 * 16;
>>>      >               query->result_size += 8; /* for the fence +
>>>     alignment */
>>>      >               query->b.num_cs_dw_suspend = 6 +
>>>     si_cp_write_fence_dwords(sscreen);
>>>      >               break;
>>>      >       default:
>>>      >               assert(0);
>>>      >               FREE(query);
>>>      >               return NULL;
>>>      >       }
>>>      >
>>>      > -     if (!si_query_hw_init(sscreen, query)) {
>>>      > -             FREE(query);
>>>      > -             return NULL;
>>>      > -     }
>>>      > -
>>>      >       return (struct pipe_query *)query;
>>>      >   }
>>>      >
>>>      >   static void si_update_occlusion_query_state(struct si_context
>>>     *sctx,
>>>      >                                           unsigned type, int 
>>> diff)
>>>      >   {
>>>      >       if (type == PIPE_QUERY_OCCLUSION_COUNTER ||
>>>      >           type == PIPE_QUERY_OCCLUSION_PREDICATE ||
>>>      >           type == PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE) {
>>>      >               bool old_enable = sctx->num_occlusion_queries != 0;
>>>      > @@ -802,43 +830,31 @@ static void
>>>     si_query_hw_do_emit_start(struct si_context *sctx,
>>>      >       }
>>>      >       radeon_add_to_buffer_list(sctx, sctx->gfx_cs,
>>>     query->buffer.buf, RADEON_USAGE_WRITE,
>>>      >                                 RADEON_PRIO_QUERY);
>>>      >   }
>>>      >
>>>      >   static void si_query_hw_emit_start(struct si_context *sctx,
>>>      >                                  struct si_query_hw *query)
>>>      >   {
>>>      >       uint64_t va;
>>>      >
>>>      > -     if (!query->buffer.buf)
>>>      > -             return; // previous buffer allocation failure
>>>      > +     if (!si_query_buffer_alloc(sctx, &query->buffer,
>>>     query->ops->prepare_buffer,
>>>      > +                                query->result_size))
>>>      > +             return;
>>>      >
>>>      >       si_update_occlusion_query_state(sctx, query->b.type, 1);
>>>      >       si_update_prims_generated_query_state(sctx, 
>>> query->b.type, 1);
>>>      >
>>>      >       if (query->b.type != SI_QUERY_TIME_ELAPSED_SDMA)
>>>      >               si_need_gfx_cs_space(sctx);
>>>      >
>>>      > -     /* Get a new query buffer if needed. */
>>>      > -     if (query->buffer.results_end + query->result_size >
>>>     query->buffer.buf->b.b.width0) {
>>>      > -             struct si_query_buffer *qbuf =
>>>     MALLOC_STRUCT(si_query_buffer);
>>>      > -             *qbuf = query->buffer;
>>>      > -             query->buffer.results_end = 0;
>>>      > -             query->buffer.previous = qbuf;
>>>      > -             query->buffer.buf =
>>>     si_new_query_buffer(sctx->screen, query);
>>>      > -             if (!query->buffer.buf)
>>>      > -                     return;
>>>      > -     }
>>>      > -
>>>      > -     /* emit begin query */
>>>      >       va = query->buffer.buf->gpu_address +
>>>     query->buffer.results_end;
>>>      > -
>>>      >       query->ops->emit_start(sctx, query, query->buffer.buf, va);
>>>      >   }
>>>      >
>>>      >   static void si_query_hw_do_emit_stop(struct si_context *sctx,
>>>      >                                      struct si_query_hw *query,
>>>      >                                      struct r600_resource 
>>> *buffer,
>>>      >                                      uint64_t va)
>>>      >   {
>>>      >       struct radeon_cmdbuf *cs = sctx->gfx_cs;
>>>      >       uint64_t fence_va = 0;
>>>      > @@ -905,26 +921,30 @@ static void si_query_hw_do_emit_stop(struct
>>>     si_context *sctx,
>>>      >                                 query->buffer.buf, fence_va,
>>>     0x80000000,
>>>      >                                 query->b.type);
>>>      >       }
>>>      >   }
>>>      >
>>>      >   static void si_query_hw_emit_stop(struct si_context *sctx,
>>>      >                                 struct si_query_hw *query)
>>>      >   {
>>>      >       uint64_t va;
>>>      >
>>>      > -     if (!query->buffer.buf)
>>>      > -             return; // previous buffer allocation failure
>>>      > -
>>>      >       /* The queries which need begin already called this in
>>>     begin_query. */
>>>      > -     if (query->flags & SI_QUERY_HW_FLAG_NO_START)
>>>      > +     if (query->flags & SI_QUERY_HW_FLAG_NO_START) {
>>>      >               si_need_gfx_cs_space(sctx);
>>>      > +             if (!si_query_buffer_alloc(sctx, &query->buffer,
>>>     query->ops->prepare_buffer,
>>>      > +                                        query->result_size))
>>>      > +                     return;
>>>      > +     }
>>>      > +
>>>      > +     if (!query->buffer.buf)
>>>      > +             return; // previous buffer allocation failure
>>>      >
>>>      >       /* emit end query */
>>>      >       va = query->buffer.buf->gpu_address +
>>>     query->buffer.results_end;
>>>      >
>>>      >       query->ops->emit_stop(sctx, query, query->buffer.buf, va);
>>>      >
>>>      >       query->buffer.results_end += query->result_size;
>>>      >
>>>      >       si_update_occlusion_query_state(sctx, query->b.type, -1);
>>>      >       si_update_prims_generated_query_state(sctx, 
>>> query->b.type, -1);
>>>      > @@ -1054,59 +1074,32 @@ static void si_destroy_query(struct
>>>     pipe_context *ctx, struct pipe_query *query)
>>>      >
>>>      >   static boolean si_begin_query(struct pipe_context *ctx,
>>>      >                                   struct pipe_query *query)
>>>      >   {
>>>      >       struct si_context *sctx = (struct si_context *)ctx;
>>>      >       struct si_query *rquery = (struct si_query *)query;
>>>      >
>>>      >       return rquery->ops->begin(sctx, rquery);
>>>      >   }
>>>      >
>>>      > -void si_query_hw_reset_buffers(struct si_context *sctx,
>>>      > -                            struct si_query_hw *query)
>>>      > -{
>>>      > -     struct si_query_buffer *prev = query->buffer.previous;
>>>      > -
>>>      > -     /* Discard the old query buffers. */
>>>      > -     while (prev) {
>>>      > -             struct si_query_buffer *qbuf = prev;
>>>      > -             prev = prev->previous;
>>>      > -             r600_resource_reference(&qbuf->buf, NULL);
>>>      > -             FREE(qbuf);
>>>      > -     }
>>>      > -
>>>      > -     query->buffer.results_end = 0;
>>>      > -     query->buffer.previous = NULL;
>>>      > -
>>>      > -     /* Obtain a new buffer if the current one can't be mapped
>>>     without a stall. */
>>>      > -     if (si_rings_is_buffer_referenced(sctx,
>>>     query->buffer.buf->buf, RADEON_USAGE_READWRITE) ||
>>>      > -         !sctx->ws->buffer_wait(query->buffer.buf->buf, 0,
>>>     RADEON_USAGE_READWRITE)) {
>>>      > -             r600_resource_reference(&query->buffer.buf, NULL);
>>>      > -             query->buffer.buf =
>>>     si_new_query_buffer(sctx->screen, query);
>>>      > -     } else {
>>>      > -             if (!query->ops->prepare_buffer(sctx->screen,
>>>     query, query->buffer.buf))
>>>      > -                     r600_resource_reference(&query->buffer.buf,
>>>     NULL);
>>>      > -     }
>>>      > -}
>>>      > -
>>>      >   bool si_query_hw_begin(struct si_context *sctx,
>>>      >                      struct si_query *rquery)
>>>      >   {
>>>      >       struct si_query_hw *query = (struct si_query_hw *)rquery;
>>>      >
>>>      >       if (query->flags & SI_QUERY_HW_FLAG_NO_START) {
>>>      >               assert(0);
>>>      >               return false;
>>>      >       }
>>>      >
>>>      >       if (!(query->flags & SI_QUERY_HW_FLAG_BEGIN_RESUMES))
>>>      > -             si_query_hw_reset_buffers(sctx, query);
>>>      > +             si_query_buffer_reset(sctx, &query->buffer);
>>>      >
>>>      >       r600_resource_reference(&query->workaround_buf, NULL);
>>>      >
>>>      >       si_query_hw_emit_start(sctx, query);
>>>      >       if (!query->buffer.buf)
>>>      >               return false;
>>>      >
>>>      >       LIST_ADDTAIL(&query->b.active_list, &sctx->active_queries);
>>>      >       sctx->num_cs_dw_queries_suspend += 
>>> query->b.num_cs_dw_suspend;
>>>      >       return true;
>>>      > @@ -1119,21 +1112,21 @@ static bool si_end_query(struct
>>>     pipe_context *ctx, struct pipe_query *query)
>>>      >
>>>      >       return rquery->ops->end(sctx, rquery);
>>>      >   }
>>>      >
>>>      >   bool si_query_hw_end(struct si_context *sctx,
>>>      >                    struct si_query *rquery)
>>>      >   {
>>>      >       struct si_query_hw *query = (struct si_query_hw *)rquery;
>>>      >
>>>      >       if (query->flags & SI_QUERY_HW_FLAG_NO_START)
>>>      > -             si_query_hw_reset_buffers(sctx, query);
>>>      > +             si_query_buffer_reset(sctx, &query->buffer);
>>>      >
>>>      >       si_query_hw_emit_stop(sctx, query);
>>>      >
>>>      >       if (!(query->flags & SI_QUERY_HW_FLAG_NO_START)) {
>>>      >               LIST_DELINIT(&query->b.active_list);
>>>      >               sctx->num_cs_dw_queries_suspend -=
>>>     query->b.num_cs_dw_suspend;
>>>      >       }
>>>      >
>>>      >       if (!query->buffer.buf)
>>>      >               return false;
>>>      > diff --git a/src/gallium/drivers/radeonsi/si_query.h
>>>     b/src/gallium/drivers/radeonsi/si_query.h
>>>      > index ebd965a004f..63af760a271 100644
>>>      > --- a/src/gallium/drivers/radeonsi/si_query.h
>>>      > +++ b/src/gallium/drivers/radeonsi/si_query.h
>>>      > @@ -27,20 +27,21 @@
>>>      >
>>>      >   #include "util/u_threaded_context.h"
>>>      >
>>>      >   struct pipe_context;
>>>      >   struct pipe_query;
>>>      >   struct pipe_resource;
>>>      >
>>>      >   struct si_screen;
>>>      >   struct si_context;
>>>      >   struct si_query;
>>>      > +struct si_query_buffer;
>>>      >   struct si_query_hw;
>>>      >   struct r600_resource;
>>>      >
>>>      >   enum {
>>>      >       SI_QUERY_DRAW_CALLS = PIPE_QUERY_DRIVER_SPECIFIC,
>>>      >       SI_QUERY_DECOMPRESS_CALLS,
>>>      >       SI_QUERY_MRT_DRAW_CALLS,
>>>      >       SI_QUERY_PRIM_RESTART_CALLS,
>>>      >       SI_QUERY_SPILL_DRAW_CALLS,
>>>      >       SI_QUERY_COMPUTE_CALLS,
>>>      > @@ -153,23 +154,21 @@ struct si_query {
>>>      >   };
>>>      >
>>>      >   enum {
>>>      >       SI_QUERY_HW_FLAG_NO_START = (1 << 0),
>>>      >       /* gap */
>>>      >       /* whether begin_query doesn't clear the result */
>>>      >       SI_QUERY_HW_FLAG_BEGIN_RESUMES = (1 << 2),
>>>      >   };
>>>      >
>>>      >   struct si_query_hw_ops {
>>>      > -     bool (*prepare_buffer)(struct si_screen *,
>>>      > -                            struct si_query_hw *,
>>>      > -                            struct r600_resource *);
>>>      > +     bool (*prepare_buffer)(struct si_context *, struct
>>>     si_query_buffer *);
>>>      >       void (*emit_start)(struct si_context *,
>>>      >                          struct si_query_hw *,
>>>      >                          struct r600_resource *buffer, 
>>> uint64_t va);
>>>      >       void (*emit_stop)(struct si_context *,
>>>      >                         struct si_query_hw *,
>>>      >                         struct r600_resource *buffer, uint64_t 
>>> va);
>>>      >       void (*clear_result)(struct si_query_hw *, union
>>>     pipe_query_result *);
>>>      >       void (*add_result)(struct si_screen *screen,
>>>      >                          struct si_query_hw *, void *buffer,
>>>      >                          union pipe_query_result *result);
>>>      > @@ -179,40 +178,45 @@ struct si_query_buffer {
>>>      >       /* The buffer where query results are stored. */
>>>      >       struct r600_resource            *buf;
>>>      >       /* Offset of the next free result after current query 
>>> data */
>>>      >       unsigned                        results_end;
>>>      >       /* If a query buffer is full, a new buffer is created and
>>>     the old one
>>>      >        * is put in here. When we calculate the result, we sum up
>>>     the samples
>>>      >        * from all buffers. */
>>>      >       struct si_query_buffer  *previous;
>>>      >   };
>>>      >
>>>      > +void si_query_buffer_destroy(struct si_screen *sctx, struct
>>>     si_query_buffer *buffer);
>>>      > +void si_query_buffer_reset(struct si_context *sctx, struct
>>>     si_query_buffer *buffer);
>>>      > +bool si_query_buffer_alloc(struct si_context *sctx, struct
>>>     si_query_buffer *buffer,
>>>      > +                        bool (*prepare_buffer)(struct si_context
>>>     *, struct si_query_buffer*),
>>>      > +                        unsigned size);
>>>      > +
>>>      > +
>>>      >   struct si_query_hw {
>>>      >       struct si_query b;
>>>      >       struct si_query_hw_ops *ops;
>>>      >       unsigned flags;
>>>      >
>>>      >       /* The query buffer and how many results are in it. */
>>>      >       struct si_query_buffer buffer;
>>>      >       /* Size of the result in memory for both begin_query and
>>>     end_query,
>>>      >        * this can be one or two numbers, or it could even be a
>>>     size of a structure. */
>>>      >       unsigned result_size;
>>>      >       /* For transform feedback: which stream the query is for */
>>>      >       unsigned stream;
>>>      >
>>>      >       /* Workaround via compute shader */
>>>      >       struct r600_resource *workaround_buf;
>>>      >       unsigned workaround_offset;
>>>      >   };
>>>      >
>>>      > -bool si_query_hw_init(struct si_screen *sscreen,
>>>      > -                   struct si_query_hw *query);
>>>      >   void si_query_hw_destroy(struct si_screen *sscreen,
>>>      >                        struct si_query *rquery);
>>>      >   bool si_query_hw_begin(struct si_context *sctx,
>>>      >                      struct si_query *rquery);
>>>      >   bool si_query_hw_end(struct si_context *sctx,
>>>      >                    struct si_query *rquery);
>>>      >   bool si_query_hw_get_result(struct si_context *sctx,
>>>      >                           struct si_query *rquery,
>>>      >                           bool wait,
>>>      >                           union pipe_query_result *result);
>>>      > @@ -237,20 +241,17 @@ struct pipe_query
>>>     *si_create_batch_query(struct pipe_context *ctx,
>>>      >                                        unsigned num_queries,
>>>      >                                        unsigned *query_types);
>>>      >
>>>      >   int si_get_perfcounter_info(struct si_screen *,
>>>      >                           unsigned index,
>>>      >                           struct pipe_driver_query_info *info);
>>>      >   int si_get_perfcounter_group_info(struct si_screen *,
>>>      >                                 unsigned index,
>>>      >                                 struct
>>>     pipe_driver_query_group_info *info);
>>>      >
>>>      > -void si_query_hw_reset_buffers(struct si_context *sctx,
>>>      > -                            struct si_query_hw *query);
>>>      > -
>>>      >   struct si_qbo_state {
>>>      >       void *saved_compute;
>>>      >       struct pipe_constant_buffer saved_const0;
>>>      >       struct pipe_shader_buffer saved_ssbo[3];
>>>      >   };
>>>      >
>>>      >   #endif /* SI_QUERY_H */
>>>      > diff --git a/src/gallium/drivers/radeonsi/si_texture.c
>>>     b/src/gallium/drivers/radeonsi/si_texture.c
>>>      > index ac1a0aa6097..9df12e0f5bd 100644
>>>      > --- a/src/gallium/drivers/radeonsi/si_texture.c
>>>      > +++ b/src/gallium/drivers/radeonsi/si_texture.c
>>>      > @@ -2276,25 +2276,24 @@ void
>>>     vi_separate_dcc_process_and_reset_stats(struct pipe_context *ctx,
>>>      >       struct si_context *sctx = (struct si_context*)ctx;
>>>      >       struct pipe_query *tmp;
>>>      >       unsigned i = vi_get_context_dcc_stats_index(sctx, tex);
>>>      >       bool query_active = sctx->dcc_stats[i].query_active;
>>>      >       bool disable = false;
>>>      >
>>>      >       if (sctx->dcc_stats[i].ps_stats[2]) {
>>>      >               union pipe_query_result result;
>>>      >
>>>      >               /* Read the results. */
>>>      > -             ctx->get_query_result(ctx,
>>>     sctx->dcc_stats[i].ps_stats[2],
>>>      > +             struct pipe_query *query =
>>>     sctx->dcc_stats[i].ps_stats[2];
>>>      > +             ctx->get_query_result(ctx, query,
>>>      >                                     true, &result);
>>>      > -             si_query_hw_reset_buffers(sctx,
>>>      > -                                       (struct si_query_hw*)
>>>      > -       sctx->dcc_stats[i].ps_stats[2]);
>>>      > +             si_query_buffer_reset(sctx, &((struct
>>>     si_query_hw*)query)->buffer);
>>>      >
>>>      >               /* Compute the approximate number of fullscreen
>>>     draws. */
>>>      >               tex->ps_draw_ratio =
>>>      >                       result.pipeline_statistics.ps_invocations /
>>>      >                       (tex->buffer.b.b.width0 *
>>>     tex->buffer.b.b.height0);
>>>      >               sctx->last_tex_ps_draw_ratio = tex->ps_draw_ratio;
>>>      >
>>>      >               disable = tex->dcc_separate_buffer &&
>>>      >                         !vi_should_enable_separate_dcc(tex);
>>>      >       }
>>>      >
>>>     _______________________________________________
>>>     mesa-dev mailing list
>>>     mesa-dev at lists.freedesktop.org 
>>> <mailto:mesa-dev at lists.freedesktop.org>
>>>     https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>>>
> 


More information about the mesa-dev mailing list