llvmpipe: enable ARB_shader_group_vote.
[mesa.git] / src / gallium / drivers / llvmpipe / lp_query.c
index 1e2401fa84d62ddf950858b136bcff325c64621d..276fd3bab7843d173061a5a982f46482a2ea03fd 100644 (file)
@@ -1,6 +1,6 @@
 /**************************************************************************
  * 
- * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * Copyright 2007 VMware, Inc.
  * Copyright 2010 VMware, Inc.
  * All Rights Reserved.
  * 
 #include "draw/draw_context.h"
 #include "pipe/p_defines.h"
 #include "util/u_memory.h"
+#include "util/os_time.h"
 #include "lp_context.h"
 #include "lp_flush.h"
 #include "lp_fence.h"
 #include "lp_query.h"
+#include "lp_screen.h"
 #include "lp_state.h"
+#include "lp_rast.h"
 
 
 static struct llvmpipe_query *llvmpipe_query( struct pipe_query *p )
@@ -47,14 +50,19 @@ static struct llvmpipe_query *llvmpipe_query( struct pipe_query *p )
 
 static struct pipe_query *
 llvmpipe_create_query(struct pipe_context *pipe, 
-                      unsigned type)
+                      unsigned type,
+                      unsigned index)
 {
    struct llvmpipe_query *pq;
 
-   assert(type == PIPE_QUERY_OCCLUSION_COUNTER);
+   assert(type < PIPE_QUERY_TYPES);
 
    pq = CALLOC_STRUCT( llvmpipe_query );
 
+   if (pq) {
+      pq->type = type;
+   }
+
    return (struct pipe_query *) pq;
 }
 
@@ -81,44 +89,254 @@ llvmpipe_destroy_query(struct pipe_context *pipe, struct pipe_query *q)
 }
 
 
-static boolean
+static bool
 llvmpipe_get_query_result(struct pipe_context *pipe, 
                           struct pipe_query *q,
-                          boolean wait,
-                          void *vresult)
+                          bool wait,
+                          union pipe_query_result *vresult)
 {
+   struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
+   unsigned num_threads = MAX2(1, screen->num_threads);
    struct llvmpipe_query *pq = llvmpipe_query(q);
    uint64_t *result = (uint64_t *)vresult;
    int i;
 
-   if (!pq->fence) {
-      /* no fence because there was no scene, so results is zero */
-      *result = 0;
-      return TRUE;
-   }
+   if (pq->fence) {
+      /* only have a fence if there was a scene */
+      if (!lp_fence_signalled(pq->fence)) {
+         if (!lp_fence_issued(pq->fence))
+            llvmpipe_flush(pipe, NULL, __FUNCTION__);
 
-   if (!lp_fence_signalled(pq->fence)) {
-      if (!lp_fence_issued(pq->fence))
-         llvmpipe_flush(pipe, NULL, __FUNCTION__);
-         
-      if (!wait)
-         return FALSE;
+         if (!wait)
+            return false;
 
-      lp_fence_wait(pq->fence);
+         lp_fence_wait(pq->fence);
+      }
    }
 
    /* Sum the results from each of the threads:
     */
    *result = 0;
-   for (i = 0; i < LP_MAX_THREADS; i++) {
-      *result += pq->count[i];
+
+   switch (pq->type) {
+   case PIPE_QUERY_OCCLUSION_COUNTER:
+      for (i = 0; i < num_threads; i++) {
+         *result += pq->end[i];
+      }
+      break;
+   case PIPE_QUERY_OCCLUSION_PREDICATE:
+   case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
+      for (i = 0; i < num_threads; i++) {
+         /* safer (still not guaranteed) when there's an overflow */
+         vresult->b = vresult->b || pq->end[i];
+      }
+      break;
+   case PIPE_QUERY_TIMESTAMP:
+      for (i = 0; i < num_threads; i++) {
+         if (pq->end[i] > *result) {
+            *result = pq->end[i];
+         }
+      }
+      break;
+   case PIPE_QUERY_TIMESTAMP_DISJOINT: {
+      struct pipe_query_data_timestamp_disjoint *td =
+         (struct pipe_query_data_timestamp_disjoint *)vresult;
+      /* os_get_time_nano return nanoseconds */
+      td->frequency = UINT64_C(1000000000);
+      td->disjoint = false;
+   }
+      break;
+   case PIPE_QUERY_GPU_FINISHED:
+      vresult->b = true;
+      break;
+   case PIPE_QUERY_PRIMITIVES_GENERATED:
+      *result = pq->num_primitives_generated;
+      break;
+   case PIPE_QUERY_PRIMITIVES_EMITTED:
+      *result = pq->num_primitives_written;
+      break;
+   case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+   case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+      vresult->b = pq->num_primitives_generated > pq->num_primitives_written;
+      break;
+   case PIPE_QUERY_SO_STATISTICS: {
+      struct pipe_query_data_so_statistics *stats =
+         (struct pipe_query_data_so_statistics *)vresult;
+      stats->num_primitives_written = pq->num_primitives_written;
+      stats->primitives_storage_needed = pq->num_primitives_generated;
+   }
+      break;
+   case PIPE_QUERY_PIPELINE_STATISTICS: {
+      struct pipe_query_data_pipeline_statistics *stats =
+         (struct pipe_query_data_pipeline_statistics *)vresult;
+      /* only ps_invocations come from binned query */
+      for (i = 0; i < num_threads; i++) {
+         pq->stats.ps_invocations += pq->end[i];
+      }
+      pq->stats.ps_invocations *= LP_RASTER_BLOCK_SIZE * LP_RASTER_BLOCK_SIZE;
+      *stats = pq->stats;
+   }
+      break;
+   default:
+      assert(0);
+      break;
    }
 
-   return TRUE;
+   return true;
 }
 
-
 static void
+llvmpipe_get_query_result_resource(struct pipe_context *pipe,
+                                   struct pipe_query *q,
+                                   bool wait,
+                                   enum pipe_query_value_type result_type,
+                                   int index,
+                                   struct pipe_resource *resource,
+                                   unsigned offset)
+{
+   struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
+   unsigned num_threads = MAX2(1, screen->num_threads);
+   struct llvmpipe_query *pq = llvmpipe_query(q);
+   struct llvmpipe_resource *lpr = llvmpipe_resource(resource);
+   bool unflushed = false;
+   bool unsignalled = false;
+   if (pq->fence) {
+      /* only have a fence if there was a scene */
+      if (!lp_fence_signalled(pq->fence)) {
+         unsignalled = true;
+         if (!lp_fence_issued(pq->fence))
+            unflushed = true;
+      }
+   }
+
+
+   uint64_t value = 0;
+   if (index == -1)
+      if (unsignalled)
+         value = 0;
+      else
+         value = 1;
+   else {
+      unsigned i;
+
+      if (unflushed) {
+         llvmpipe_flush(pipe, NULL, __FUNCTION__);
+
+         if (!wait)
+            return;
+
+         lp_fence_wait(pq->fence);
+      }
+
+      switch (pq->type) {
+      case PIPE_QUERY_OCCLUSION_COUNTER:
+         for (i = 0; i < num_threads; i++) {
+            value += pq->end[i];
+         }
+         break;
+      case PIPE_QUERY_OCCLUSION_PREDICATE:
+      case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
+         for (i = 0; i < num_threads; i++) {
+            /* safer (still not guaranteed) when there's an overflow */
+            value = value || pq->end[i];
+         }
+         break;
+      case PIPE_QUERY_PRIMITIVES_GENERATED:
+         value = pq->num_primitives_generated;
+         break;
+      case PIPE_QUERY_PRIMITIVES_EMITTED:
+         value = pq->num_primitives_written;
+         break;
+      case PIPE_QUERY_TIMESTAMP:
+         for (i = 0; i < num_threads; i++) {
+            if (pq->end[i] > value) {
+               value = pq->end[i];
+            }
+         }
+         break;
+      case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+      case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+         value = !!(pq->num_primitives_generated > pq->num_primitives_written);
+         break;
+      case PIPE_QUERY_PIPELINE_STATISTICS:
+         switch ((enum pipe_statistics_query_index)index) {
+         case PIPE_STAT_QUERY_IA_VERTICES:
+            value = pq->stats.ia_vertices;
+            break;
+         case PIPE_STAT_QUERY_IA_PRIMITIVES:
+            value = pq->stats.ia_primitives;
+            break;
+         case PIPE_STAT_QUERY_VS_INVOCATIONS:
+            value = pq->stats.vs_invocations;
+            break;
+         case PIPE_STAT_QUERY_GS_INVOCATIONS:
+            value = pq->stats.gs_invocations;
+            break;
+         case PIPE_STAT_QUERY_GS_PRIMITIVES:
+            value = pq->stats.gs_primitives;
+            break;
+         case PIPE_STAT_QUERY_C_INVOCATIONS:
+            value = pq->stats.c_invocations;
+            break;
+         case PIPE_STAT_QUERY_C_PRIMITIVES:
+            value = pq->stats.c_primitives;
+            break;
+         case PIPE_STAT_QUERY_PS_INVOCATIONS:
+            value = 0;
+            for (i = 0; i < num_threads; i++) {
+               value += pq->end[i];
+            }
+            value *= LP_RASTER_BLOCK_SIZE * LP_RASTER_BLOCK_SIZE;
+            break;
+         case PIPE_STAT_QUERY_HS_INVOCATIONS:
+            value = pq->stats.hs_invocations;
+            break;
+         case PIPE_STAT_QUERY_DS_INVOCATIONS:
+            value = pq->stats.ds_invocations;
+            break;
+         case PIPE_STAT_QUERY_CS_INVOCATIONS:
+            value = pq->stats.cs_invocations;
+            break;
+         }
+         break;
+      default:
+         fprintf(stderr, "Unknown query type %d\n", pq->type);
+         break;
+      }
+   }
+
+   void *dst = (uint8_t *)lpr->data + offset;
+   switch (result_type) {
+   case PIPE_QUERY_TYPE_I32: {
+      int32_t *iptr = (int32_t *)dst;
+      if (value > 0x7fffffff)
+         *iptr = 0x7fffffff;
+      else
+         *iptr = (int32_t)value;
+      break;
+   }
+   case PIPE_QUERY_TYPE_U32: {
+      uint32_t *uptr = (uint32_t *)dst;
+      if (value > 0xffffffff)
+         *uptr = 0xffffffff;
+      else
+         *uptr = (uint32_t)value;
+      break;
+   }
+   case PIPE_QUERY_TYPE_I64: {
+      int64_t *iptr = (int64_t *)dst;
+      *iptr = (int64_t)value;
+      break;
+   }
+   case PIPE_QUERY_TYPE_U64: {
+      uint64_t *uptr = (uint64_t *)dst;
+      *uptr = (uint64_t)value;
+      break;
+   }
+   }
+}
+
+static bool
 llvmpipe_begin_query(struct pipe_context *pipe, struct pipe_query *q)
 {
    struct llvmpipe_context *llvmpipe = llvmpipe_context( pipe );
@@ -133,15 +351,50 @@ llvmpipe_begin_query(struct pipe_context *pipe, struct pipe_query *q)
    }
 
 
-   memset(pq->count, 0, sizeof(pq->count));
+   memset(pq->start, 0, sizeof(pq->start));
+   memset(pq->end, 0, sizeof(pq->end));
    lp_setup_begin_query(llvmpipe->setup, pq);
 
-   llvmpipe->active_query_count++;
-   llvmpipe->dirty |= LP_NEW_QUERY;
+   switch (pq->type) {
+   case PIPE_QUERY_PRIMITIVES_EMITTED:
+      pq->num_primitives_written = llvmpipe->so_stats.num_primitives_written;
+      break;
+   case PIPE_QUERY_PRIMITIVES_GENERATED:
+      pq->num_primitives_generated = llvmpipe->so_stats.primitives_storage_needed;
+      llvmpipe->active_primgen_queries++;
+      break;
+   case PIPE_QUERY_SO_STATISTICS:
+      pq->num_primitives_written = llvmpipe->so_stats.num_primitives_written;
+      pq->num_primitives_generated = llvmpipe->so_stats.primitives_storage_needed;
+      break;
+   case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+   case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+      pq->num_primitives_written = llvmpipe->so_stats.num_primitives_written;
+      pq->num_primitives_generated = llvmpipe->so_stats.primitives_storage_needed;
+      break;
+   case PIPE_QUERY_PIPELINE_STATISTICS:
+      /* reset our cache */
+      if (llvmpipe->active_statistics_queries == 0) {
+         memset(&llvmpipe->pipeline_statistics, 0,
+                sizeof(llvmpipe->pipeline_statistics));
+      }
+      memcpy(&pq->stats, &llvmpipe->pipeline_statistics, sizeof(pq->stats));
+      llvmpipe->active_statistics_queries++;
+      break;
+   case PIPE_QUERY_OCCLUSION_COUNTER:
+   case PIPE_QUERY_OCCLUSION_PREDICATE:
+   case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
+      llvmpipe->active_occlusion_queries++;
+      llvmpipe->dirty |= LP_NEW_OCCLUSION_QUERY;
+      break;
+   default:
+      break;
+   }
+   return true;
 }
 
 
-static void
+static bool
 llvmpipe_end_query(struct pipe_context *pipe, struct pipe_query *q)
 {
    struct llvmpipe_context *llvmpipe = llvmpipe_context( pipe );
@@ -149,11 +402,95 @@ llvmpipe_end_query(struct pipe_context *pipe, struct pipe_query *q)
 
    lp_setup_end_query(llvmpipe->setup, pq);
 
-   assert(llvmpipe->active_query_count);
-   llvmpipe->active_query_count--;
-   llvmpipe->dirty |= LP_NEW_QUERY;
+   switch (pq->type) {
+
+   case PIPE_QUERY_PRIMITIVES_EMITTED:
+      pq->num_primitives_written =
+         llvmpipe->so_stats.num_primitives_written - pq->num_primitives_written;
+      break;
+   case PIPE_QUERY_PRIMITIVES_GENERATED:
+      assert(llvmpipe->active_primgen_queries);
+      llvmpipe->active_primgen_queries--;
+      pq->num_primitives_generated =
+         llvmpipe->so_stats.primitives_storage_needed - pq->num_primitives_generated;
+      break;
+   case PIPE_QUERY_SO_STATISTICS:
+      pq->num_primitives_written =
+         llvmpipe->so_stats.num_primitives_written - pq->num_primitives_written;
+      pq->num_primitives_generated =
+         llvmpipe->so_stats.primitives_storage_needed - pq->num_primitives_generated;
+      break;
+   case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
+   case PIPE_QUERY_SO_OVERFLOW_ANY_PREDICATE:
+      pq->num_primitives_written =
+         llvmpipe->so_stats.num_primitives_written - pq->num_primitives_written;
+      pq->num_primitives_generated =
+         llvmpipe->so_stats.primitives_storage_needed - pq->num_primitives_generated;
+      break;
+   case PIPE_QUERY_PIPELINE_STATISTICS:
+      pq->stats.ia_vertices =
+         llvmpipe->pipeline_statistics.ia_vertices - pq->stats.ia_vertices;
+      pq->stats.ia_primitives =
+         llvmpipe->pipeline_statistics.ia_primitives - pq->stats.ia_primitives;
+      pq->stats.vs_invocations =
+         llvmpipe->pipeline_statistics.vs_invocations - pq->stats.vs_invocations;
+      pq->stats.gs_invocations =
+         llvmpipe->pipeline_statistics.gs_invocations - pq->stats.gs_invocations;
+      pq->stats.gs_primitives =
+         llvmpipe->pipeline_statistics.gs_primitives - pq->stats.gs_primitives;
+      pq->stats.c_invocations =
+         llvmpipe->pipeline_statistics.c_invocations - pq->stats.c_invocations;
+      pq->stats.c_primitives =
+         llvmpipe->pipeline_statistics.c_primitives - pq->stats.c_primitives;
+      pq->stats.ps_invocations =
+         llvmpipe->pipeline_statistics.ps_invocations - pq->stats.ps_invocations;
+      pq->stats.cs_invocations =
+         llvmpipe->pipeline_statistics.cs_invocations - pq->stats.cs_invocations;
+      llvmpipe->active_statistics_queries--;
+      break;
+   case PIPE_QUERY_OCCLUSION_COUNTER:
+   case PIPE_QUERY_OCCLUSION_PREDICATE:
+   case PIPE_QUERY_OCCLUSION_PREDICATE_CONSERVATIVE:
+      assert(llvmpipe->active_occlusion_queries);
+      llvmpipe->active_occlusion_queries--;
+      llvmpipe->dirty |= LP_NEW_OCCLUSION_QUERY;
+      break;
+   default:
+      break;
+   }
+
+   return true;
+}
+
+boolean
+llvmpipe_check_render_cond(struct llvmpipe_context *lp)
+{
+   struct pipe_context *pipe = &lp->pipe;
+   boolean b, wait;
+   uint64_t result;
+
+   if (!lp->render_cond_query)
+      return TRUE; /* no query predicate, draw normally */
+
+   wait = (lp->render_cond_mode == PIPE_RENDER_COND_WAIT ||
+           lp->render_cond_mode == PIPE_RENDER_COND_BY_REGION_WAIT);
+
+   b = pipe->get_query_result(pipe, lp->render_cond_query, wait, (void*)&result);
+   if (b)
+      return ((!result) == lp->render_cond_cond);
+   else
+      return TRUE;
 }
 
+static void
+llvmpipe_set_active_query_state(struct pipe_context *pipe, bool enable)
+{
+   struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
+
+   llvmpipe->queries_disabled = !enable;
+   /* for OQs we need to regenerate the fragment shader */
+   llvmpipe->dirty |= LP_NEW_OCCLUSION_QUERY;
+}
 
 void llvmpipe_init_query_funcs(struct llvmpipe_context *llvmpipe )
 {
@@ -162,6 +499,8 @@ void llvmpipe_init_query_funcs(struct llvmpipe_context *llvmpipe )
    llvmpipe->pipe.begin_query = llvmpipe_begin_query;
    llvmpipe->pipe.end_query = llvmpipe_end_query;
    llvmpipe->pipe.get_query_result = llvmpipe_get_query_result;
+   llvmpipe->pipe.get_query_result_resource = llvmpipe_get_query_result_resource;
+   llvmpipe->pipe.set_active_query_state = llvmpipe_set_active_query_state;
 }