intel/perf: repurpose INTEL_DEBUG=no-oaconfig

[mesa.git] / src / intel / perf / gen_perf.c

diff --git a/src/intel/perf/gen_perf.c b/src/intel/perf/gen_perf.c
index 2dc63c5226f38638a146ae309a6c4f0149dcdeb7..6fd66a1d2dd3dc1bbe2777424c2977d1ad1428e8 100644 (file)
--- a/src/intel/perf/gen_perf.c
+++ b/src/intel/perf/gen_perf.c
@@ -80,6 +80,9 @@ get_sysfs_dev_dir(struct gen_perf_config *perf, int fd)
 
    perf->sysfs_dev_dir[0] = '\0';
 
+   if (unlikely(INTEL_DEBUG & DEBUG_NO_OACONFIG))
+      return true;
+
    if (fstat(fd, &sb)) {
       DBG("Failed to stat DRM fd\n");
       return false;
@@ -232,6 +235,16 @@ enumerate_sysfs_metrics(struct gen_perf_config *perf,
    closedir(metricsdir);
 }
 
+static void
+add_all_metrics(struct gen_perf_config *perf,
+                const struct gen_device_info *devinfo)
+{
+   hash_table_foreach(perf->oa_metrics_table, entry) {
+      const struct gen_perf_query_info *query = entry->data;
+      register_oa_config(perf, devinfo, query, 0);
+   }
+}
+
 static bool
 kernel_has_dynamic_config_support(struct gen_perf_config *perf, int fd)
 {
@@ -394,11 +407,16 @@ init_oa_sys_vars(struct gen_perf_config *perf, const struct gen_device_info *dev
 {
    uint64_t min_freq_mhz = 0, max_freq_mhz = 0;
 
-   if (!read_sysfs_drm_device_file_uint64(perf, "gt_min_freq_mhz", &min_freq_mhz))
-      return false;
+   if (likely(!(INTEL_DEBUG & DEBUG_NO_OACONFIG))) {
+      if (!read_sysfs_drm_device_file_uint64(perf, "gt_min_freq_mhz", &min_freq_mhz))
+         return false;
 
-   if (!read_sysfs_drm_device_file_uint64(perf,  "gt_max_freq_mhz", &max_freq_mhz))
-      return false;
+      if (!read_sysfs_drm_device_file_uint64(perf,  "gt_max_freq_mhz", &max_freq_mhz))
+         return false;
+   } else {
+      min_freq_mhz = 300;
+      max_freq_mhz = 1000;
+   }
 
    memset(&perf->sys_vars, 0, sizeof(perf->sys_vars));
    perf->sys_vars.gt_min_freq = min_freq_mhz * 1000000;
@@ -679,11 +697,14 @@ load_oa_metrics(struct gen_perf_config *perf, int fd,
     */
    oa_register(perf);
 
-   if (likely((INTEL_DEBUG & DEBUG_NO_OACONFIG) == 0) &&
-       kernel_has_dynamic_config_support(perf, fd))
-      init_oa_configs(perf, fd, devinfo);
-   else
-      enumerate_sysfs_metrics(perf, devinfo);
+   if (likely(!(INTEL_DEBUG & DEBUG_NO_OACONFIG))) {
+      if (kernel_has_dynamic_config_support(perf, fd))
+         init_oa_configs(perf, fd, devinfo);
+      else
+         enumerate_sysfs_metrics(perf, devinfo);
+   } else {
+      add_all_metrics(perf, devinfo);
+   }
 
    build_unique_counter_list(perf);
 
@@ -939,7 +960,7 @@ gen_perf_query_result_accumulate(struct gen_perf_query_result *result,
                                  const uint32_t *start,
                                  const uint32_t *end)
 {
-   int i, idx = 0;
+   int i;
 
    if (result->hw_id == OA_REPORT_INVALID_CTX_ID &&
        start[2] != OA_REPORT_INVALID_CTX_ID)
@@ -950,27 +971,43 @@ gen_perf_query_result_accumulate(struct gen_perf_query_result *result,
 
    switch (query->oa_format) {
    case I915_OA_FORMAT_A32u40_A4u32_B8_C8:
-      accumulate_uint32(start + 1, end + 1, result->accumulator + idx++); /* timestamp */
-      accumulate_uint32(start + 3, end + 3, result->accumulator + idx++); /* clock */
+      accumulate_uint32(start + 1, end + 1,
+                        result->accumulator + query->gpu_time_offset); /* timestamp */
+      accumulate_uint32(start + 3, end + 3,
+                        result->accumulator + query->gpu_clock_offset); /* clock */
 
       /* 32x 40bit A counters... */
-      for (i = 0; i < 32; i++)
-         accumulate_uint40(i, start, end, result->accumulator + idx++);
+      for (i = 0; i < 32; i++) {
+         accumulate_uint40(i, start, end,
+                           result->accumulator + query->a_offset + i);
+      }
 
       /* 4x 32bit A counters... */
-      for (i = 0; i < 4; i++)
-         accumulate_uint32(start + 36 + i, end + 36 + i, result->accumulator + idx++);
+      for (i = 0; i < 4; i++) {
+         accumulate_uint32(start + 36 + i, end + 36 + i,
+                           result->accumulator + query->a_offset + 32 + i);
+      }
+
+      /* 8x 32bit B counters */
+      for (i = 0; i < 8; i++) {
+         accumulate_uint32(start + 48 + i, end + 48 + i,
+                           result->accumulator + query->b_offset + i);
+      }
 
-      /* 8x 32bit B counters + 8x 32bit C counters... */
-      for (i = 0; i < 16; i++)
-         accumulate_uint32(start + 48 + i, end + 48 + i, result->accumulator + idx++);
+      /* 8x 32bit C counters... */
+      for (i = 0; i < 8; i++) {
+         accumulate_uint32(start + 56 + i, end + 56 + i,
+                           result->accumulator + query->c_offset + i);
+      }
       break;
 
    case I915_OA_FORMAT_A45_B8_C8:
       accumulate_uint32(start + 1, end + 1, result->accumulator); /* timestamp */
 
-      for (i = 0; i < 61; i++)
-         accumulate_uint32(start + 3 + i, end + 3 + i, result->accumulator + 1 + i);
+      for (i = 0; i < 61; i++) {
+         accumulate_uint32(start + 3 + i, end + 3 + i,
+                           result->accumulator + query->a_offset + i);
+      }
       break;
 
    default: