intel/perf: move get_query_data into gen_perf

This refactor moves several helper functions for get_query_data as
well:

 - accumulate_oa_reports
 - read_gt_frequency
 - get_pipeline_stats_data
 - get_oa_counter_data

Functions which are no longer referenced in brw_performance_query.c
have been removed.

Reviewed-by: Kenneth Graunke <[email protected]>

2 files changed, 2 insertions, 399 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_defines.h b/src/mesa/drivers/dri/i965/brw_defines.h
index a4ecc68ec0f..425f5534110 100644
--- a/src/mesa/drivers/dri/i965/brw_defines.h
+++ b/src/mesa/drivers/dri/i965/brw_defines.h
@@ -1645,18 +1645,6 @@ enum brw_pixel_shader_coverage_mask_mode {
 #define CS_DEBUG_MODE2                     0x20d8 /* Gen9+ */
 # define CSDBG2_CONSTANT_BUFFER_ADDRESS_OFFSET_DISABLE (1 << 4)
 
-#define GEN7_RPSTAT1                       0xA01C
-#define  GEN7_RPSTAT1_CURR_GT_FREQ_SHIFT   7
-#define  GEN7_RPSTAT1_CURR_GT_FREQ_MASK    INTEL_MASK(13, 7)
-#define  GEN7_RPSTAT1_PREV_GT_FREQ_SHIFT   0
-#define  GEN7_RPSTAT1_PREV_GT_FREQ_MASK    INTEL_MASK(6, 0)
-
-#define GEN9_RPSTAT0                       0xA01C
-#define  GEN9_RPSTAT0_CURR_GT_FREQ_SHIFT   23
-#define  GEN9_RPSTAT0_CURR_GT_FREQ_MASK    INTEL_MASK(31, 23)
-#define  GEN9_RPSTAT0_PREV_GT_FREQ_SHIFT   0
-#define  GEN9_RPSTAT0_PREV_GT_FREQ_MASK    INTEL_MASK(8, 0)
-
 #define SLICE_COMMON_ECO_CHICKEN1          0x731c /* Gen9+ */
 # define GLK_SCEC_BARRIER_MODE_GPGPU       (0 << 7)
 # define GLK_SCEC_BARRIER_MODE_3D_HULL     (1 << 7)
diff --git a/src/mesa/drivers/dri/i965/brw_performance_query.c b/src/mesa/drivers/dri/i965/brw_performance_query.c
index 45298e79112..dc4a833f563 100644
--- a/src/mesa/drivers/dri/i965/brw_performance_query.c
+++ b/src/mesa/drivers/dri/i965/brw_performance_query.c
@@ -244,246 +244,12 @@ brw_get_perf_counter_info(struct gl_context *ctx,
    *raw_max = counter->raw_max;
 }
 
-/**
- * Remove a query from the global list of unaccumulated queries once
- * after successfully accumulating the OA reports associated with the
- * query in accumulate_oa_reports() or when discarding unwanted query
- * results.
- */
-static void
-drop_from_unaccumulated_query_list(struct brw_context *brw,
-                                   struct gen_perf_query_object *obj)
-{
-   struct gen_perf_context *perf_ctx = &brw->perf_ctx;
-   for (int i = 0; i < perf_ctx->unaccumulated_elements; i++) {
-      if (perf_ctx->unaccumulated[i] == obj) {
-         int last_elt = --perf_ctx->unaccumulated_elements;
-
-         if (i == last_elt)
-            perf_ctx->unaccumulated[i] = NULL;
-         else {
-            perf_ctx->unaccumulated[i] =
-               perf_ctx->unaccumulated[last_elt];
-         }
-
-         break;
-      }
-   }
-
-   /* Drop our samples_head reference so that associated periodic
-    * sample data buffers can potentially be reaped if they aren't
-    * referenced by any other queries...
-    */
-
-   struct oa_sample_buf *buf =
-      exec_node_data(struct oa_sample_buf, obj->oa.samples_head, link);
-
-   assert(buf->refcount > 0);
-   buf->refcount--;
-
-   obj->oa.samples_head = NULL;
-
-   gen_perf_reap_old_sample_buffers(&brw->perf_ctx);
-}
-
-/* In general if we see anything spurious while accumulating results,
- * we don't try and continue accumulating the current query, hoping
- * for the best, we scrap anything outstanding, and then hope for the
- * best with new queries.
- */
-static void
-discard_all_queries(struct brw_context *brw)
-{
-   struct gen_perf_context *perf_ctx = &brw->perf_ctx;
-   while (perf_ctx->unaccumulated_elements) {
-      struct gen_perf_query_object *obj = perf_ctx->unaccumulated[0];
-
-      obj->oa.results_accumulated = true;
-      drop_from_unaccumulated_query_list(brw, perf_ctx->unaccumulated[0]);
-
-      gen_perf_dec_n_users(perf_ctx);
-   }
-}
-
 enum OaReadStatus {
    OA_READ_STATUS_ERROR,
    OA_READ_STATUS_UNFINISHED,
    OA_READ_STATUS_FINISHED,
 };
 
-/**
- * Accumulate raw OA counter values based on deltas between pairs of
- * OA reports.
- *
- * Accumulation starts from the first report captured via
- * MI_REPORT_PERF_COUNT (MI_RPC) by brw_begin_perf_query() until the
- * last MI_RPC report requested by brw_end_perf_query(). Between these
- * two reports there may also some number of periodically sampled OA
- * reports collected via the i915 perf interface - depending on the
- * duration of the query.
- *
- * These periodic snapshots help to ensure we handle counter overflow
- * correctly by being frequent enough to ensure we don't miss multiple
- * overflows of a counter between snapshots. For Gen8+ the i915 perf
- * snapshots provide the extra context-switch reports that let us
- * subtract out the progress of counters associated with other
- * contexts running on the system.
- */
-static void
-accumulate_oa_reports(struct brw_context *brw,
-                      struct brw_perf_query_object *brw_query)
-{
-   const struct gen_device_info *devinfo = &brw->screen->devinfo;
-   struct gen_perf_query_object *obj = brw_query->query;
-   struct gen_perf_context *perf_ctx = &brw->perf_ctx;
-   uint32_t *start;
-   uint32_t *last;
-   uint32_t *end;
-   struct exec_node *first_samples_node;
-   bool in_ctx = true;
-   int out_duration = 0;
-
-   assert(brw_query->base.Ready);
-   assert(obj->oa.map != NULL);
-
-   start = last = obj->oa.map;
-   end = obj->oa.map + MI_RPC_BO_END_OFFSET_BYTES;
-
-   if (start[0] != obj->oa.begin_report_id) {
-      DBG("Spurious start report id=%"PRIu32"\n", start[0]);
-      goto error;
-   }
-   if (end[0] != (obj->oa.begin_report_id + 1)) {
-      DBG("Spurious end report id=%"PRIu32"\n", end[0]);
-      goto error;
-   }
-
-   /* See if we have any periodic reports to accumulate too... */
-
-   /* N.B. The oa.samples_head was set when the query began and
-    * pointed to the tail of the perf_ctx->sample_buffers list at
-    * the time the query started. Since the buffer existed before the
-    * first MI_REPORT_PERF_COUNT command was emitted we therefore know
-    * that no data in this particular node's buffer can possibly be
-    * associated with the query - so skip ahead one...
-    */
-   first_samples_node = obj->oa.samples_head->next;
-
-   foreach_list_typed_from(struct oa_sample_buf, buf, link,
-                           &brw->perf_ctx.sample_buffers,
-                           first_samples_node)
-   {
-      int offset = 0;
-
-      while (offset < buf->len) {
-         const struct drm_i915_perf_record_header *header =
-            (const struct drm_i915_perf_record_header *)(buf->buf + offset);
-
-         assert(header->size != 0);
-         assert(header->size <= buf->len);
-
-         offset += header->size;
-
-         switch (header->type) {
-         case DRM_I915_PERF_RECORD_SAMPLE: {
-            uint32_t *report = (uint32_t *)(header + 1);
-            bool add = true;
-
-            /* Ignore reports that come before the start marker.
-             * (Note: takes care to allow overflow of 32bit timestamps)
-             */
-            if (gen_device_info_timebase_scale(devinfo,
-                                               report[1] - start[1]) > 5000000000) {
-               continue;
-            }
-
-            /* Ignore reports that come after the end marker.
-             * (Note: takes care to allow overflow of 32bit timestamps)
-             */
-            if (gen_device_info_timebase_scale(devinfo,
-                                               report[1] - end[1]) <= 5000000000) {
-               goto end;
-            }
-
-            /* For Gen8+ since the counters continue while other
-             * contexts are running we need to discount any unrelated
-             * deltas. The hardware automatically generates a report
-             * on context switch which gives us a new reference point
-             * to continuing adding deltas from.
-             *
-             * For Haswell we can rely on the HW to stop the progress
-             * of OA counters while any other context is acctive.
-             */
-            if (devinfo->gen >= 8) {
-               if (in_ctx && report[2] != obj->oa.result.hw_id) {
-                  DBG("i915 perf: Switch AWAY (observed by ID change)\n");
-                  in_ctx = false;
-                  out_duration = 0;
-               } else if (in_ctx == false && report[2] == obj->oa.result.hw_id) {
-                  DBG("i915 perf: Switch TO\n");
-                  in_ctx = true;
-
-                  /* From experimentation in IGT, we found that the OA unit
-                   * might label some report as "idle" (using an invalid
-                   * context ID), right after a report for a given context.
-                   * Deltas generated by those reports actually belong to the
-                   * previous context, even though they're not labelled as
-                   * such.
-                   *
-                   * We didn't *really* Switch AWAY in the case that we e.g.
-                   * saw a single periodic report while idle...
-                   */
-                  if (out_duration >= 1)
-                     add = false;
-               } else if (in_ctx) {
-                  assert(report[2] == obj->oa.result.hw_id);
-                  DBG("i915 perf: Continuation IN\n");
-               } else {
-                  assert(report[2] != obj->oa.result.hw_id);
-                  DBG("i915 perf: Continuation OUT\n");
-                  add = false;
-                  out_duration++;
-               }
-            }
-
-            if (add) {
-               gen_perf_query_result_accumulate(&obj->oa.result, obj->queryinfo,
-                                                last, report);
-            }
-
-            last = report;
-
-            break;
-         }
-
-         case DRM_I915_PERF_RECORD_OA_BUFFER_LOST:
-             DBG("i915 perf: OA error: all reports lost\n");
-             goto error;
-         case DRM_I915_PERF_RECORD_OA_REPORT_LOST:
-             DBG("i915 perf: OA report lost\n");
-             break;
-         }
-      }
-   }
-
-end:
-
-   gen_perf_query_result_accumulate(&obj->oa.result, obj->queryinfo,
-                                    last, end);
-
-   DBG("Marking %d accumulated - results gathered\n", brw_query->base.Id);
-
-   obj->oa.results_accumulated = true;
-   drop_from_unaccumulated_query_list(brw, obj);
-   gen_perf_dec_n_users(perf_ctx);
-
-   return;
-
-error:
-
-   discard_all_queries(brw);
-}
-
 /******************************************************************************/
 
 static void
@@ -574,123 +340,6 @@ brw_is_perf_query_ready(struct gl_context *ctx,
    return gen_perf_is_query_ready(&brw->perf_ctx, obj, &brw->batch);
 }
 
-static void
-read_slice_unslice_frequencies(struct brw_context *brw,
-                               struct gen_perf_query_object *obj)
-{
-   const struct gen_device_info *devinfo = &brw->screen->devinfo;
-   uint32_t *begin_report = obj->oa.map, *end_report = obj->oa.map + MI_RPC_BO_END_OFFSET_BYTES;
-
-   gen_perf_query_result_read_frequencies(&obj->oa.result,
-                                          devinfo, begin_report, end_report);
-}
-
-static void
-read_gt_frequency(struct brw_context *brw,
-                  struct gen_perf_query_object *obj)
-{
-   const struct gen_device_info *devinfo = &brw->screen->devinfo;
-   uint32_t start = *((uint32_t *)(obj->oa.map + MI_FREQ_START_OFFSET_BYTES)),
-      end = *((uint32_t *)(obj->oa.map + MI_FREQ_END_OFFSET_BYTES));
-
-   switch (devinfo->gen) {
-   case 7:
-   case 8:
-      obj->oa.gt_frequency[0] = GET_FIELD(start, GEN7_RPSTAT1_CURR_GT_FREQ) * 50ULL;
-      obj->oa.gt_frequency[1] = GET_FIELD(end, GEN7_RPSTAT1_CURR_GT_FREQ) * 50ULL;
-      break;
-   case 9:
-   case 10:
-   case 11:
-      obj->oa.gt_frequency[0] = GET_FIELD(start, GEN9_RPSTAT0_CURR_GT_FREQ) * 50ULL / 3ULL;
-      obj->oa.gt_frequency[1] = GET_FIELD(end, GEN9_RPSTAT0_CURR_GT_FREQ) * 50ULL / 3ULL;
-      break;
-   default:
-      unreachable("unexpected gen");
-   }
-
-   /* Put the numbers into Hz. */
-   obj->oa.gt_frequency[0] *= 1000000ULL;
-   obj->oa.gt_frequency[1] *= 1000000ULL;
-}
-
-static int
-get_oa_counter_data(struct brw_context *brw,
-                    struct gen_perf_query_object *obj,
-                    size_t data_size,
-                    uint8_t *data)
-{
-   struct gen_perf_config *perf = brw->perf_ctx.perf;
-   const struct gen_perf_query_info *query = obj->queryinfo;
-   int n_counters = query->n_counters;
-   int written = 0;
-
-   for (int i = 0; i < n_counters; i++) {
-      const struct gen_perf_query_counter *counter = &query->counters[i];
-      uint64_t *out_uint64;
-      float *out_float;
-      size_t counter_size = gen_perf_query_counter_get_size(counter);
-
-      if (counter_size) {
-         switch (counter->data_type) {
-         case GEN_PERF_COUNTER_DATA_TYPE_UINT64:
-            out_uint64 = (uint64_t *)(data + counter->offset);
-            *out_uint64 =
-               counter->oa_counter_read_uint64(perf, query,
-                                               obj->oa.result.accumulator);
-            break;
-         case GEN_PERF_COUNTER_DATA_TYPE_FLOAT:
-            out_float = (float *)(data + counter->offset);
-            *out_float =
-               counter->oa_counter_read_float(perf, query,
-                                              obj->oa.result.accumulator);
-            break;
-         default:
-            /* So far we aren't using uint32, double or bool32... */
-            unreachable("unexpected counter data type");
-         }
-         written = counter->offset + counter_size;
-      }
-   }
-
-   return written;
-}
-
-static int
-get_pipeline_stats_data(struct brw_context *brw,
-                        struct gen_perf_query_object *obj,
-                        size_t data_size,
-                        uint8_t *data)
-
-{
-   const struct gen_perf_query_info *query = obj->queryinfo;
-   struct gen_perf_context *perf_ctx = &brw->perf_ctx;
-   struct gen_perf_config *perf_cfg = perf_ctx->perf;
-   int n_counters = obj->queryinfo->n_counters;
-   uint8_t *p = data;
-
-   uint64_t *start = perf_cfg->vtbl.bo_map(perf_ctx->ctx, obj->pipeline_stats.bo, MAP_READ);
-   uint64_t *end = start + (STATS_BO_END_OFFSET_BYTES / sizeof(uint64_t));
-
-   for (int i = 0; i < n_counters; i++) {
-      const struct gen_perf_query_counter *counter = &query->counters[i];
-      uint64_t value = end[i] - start[i];
-
-      if (counter->pipeline_stat.numerator !=
-          counter->pipeline_stat.denominator) {
-         value *= counter->pipeline_stat.numerator;
-         value /= counter->pipeline_stat.denominator;
-      }
-
-      *((uint64_t *)p) = value;
-      p += 8;
-   }
-
-   perf_cfg->vtbl.bo_unmap(obj->pipeline_stats.bo);
-
-   return p - data;
-}
-
 /**
  * Driver hook for glGetPerfQueryDataINTEL().
  */
@@ -704,7 +353,6 @@ brw_get_perf_query_data(struct gl_context *ctx,
    struct brw_context *brw = brw_context(ctx);
    struct brw_perf_query_object *brw_query = brw_perf_query(o);
    struct gen_perf_query_object *obj = brw_query->query;
-   int written = 0;
 
    assert(brw_is_perf_query_ready(ctx, o));
 
@@ -718,41 +366,8 @@ brw_get_perf_query_data(struct gl_context *ctx,
     */
    assert(o->Ready);
 
-   switch (obj->queryinfo->kind) {
-   case GEN_PERF_QUERY_TYPE_OA:
-   case GEN_PERF_QUERY_TYPE_RAW:
-      if (!obj->oa.results_accumulated) {
-         read_gt_frequency(brw, obj);
-         read_slice_unslice_frequencies(brw, obj);
-         accumulate_oa_reports(brw, brw_query);
-         assert(obj->oa.results_accumulated);
-
-         brw->perf_ctx.perf->vtbl.bo_unmap(obj->oa.bo);
-         obj->oa.map = NULL;
-      }
-      if (obj->queryinfo->kind == GEN_PERF_QUERY_TYPE_OA) {
-         written = get_oa_counter_data(brw, obj, data_size, (uint8_t *)data);
-      } else {
-         const struct gen_device_info *devinfo = &brw->screen->devinfo;
-
-         written = gen_perf_query_result_write_mdapi((uint8_t *)data, data_size,
-                                                     devinfo, &obj->oa.result,
-                                                     obj->oa.gt_frequency[0],
-                                                     obj->oa.gt_frequency[1]);
-      }
-      break;
-
-   case GEN_PERF_QUERY_TYPE_PIPELINE:
-      written = get_pipeline_stats_data(brw, obj, data_size, (uint8_t *)data);
-      break;
-
-   default:
-      unreachable("Unknown query type");
-      break;
-   }
-
-   if (bytes_written)
-      *bytes_written = written;
+   gen_perf_get_query_data(&brw->perf_ctx, obj,
+                           data_size, data, bytes_written);
 }
 
 static struct gl_perf_query_object *