turnip: Add debug option to print gmem load/store skip stats

TU_DEBUG=log_skip_gmem_ops would print stats about skipped
gmem/load every second.

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/15974>

src/freedreno/vulkan/tu_clear_blit.c

@@ -2808,6 +2808,22 @@ tu_begin_load_store_cond_exec(struct tu_cmd_buffer *cmd,
                               struct tu_cs *cs, bool load)
 {
    tu_cond_exec_start(cs, CP_COND_REG_EXEC_0_MODE(PRED_TEST));
+
+   if (!unlikely(cmd->device->physical_device->instance->debug_flags &
+                 TU_DEBUG_LOG_SKIP_GMEM_OPS))
+      return;
+
+   uint64_t result_iova;
+   if (load)
+      result_iova = global_iova(cmd, dbg_gmem_taken_loads);
+   else
+      result_iova = global_iova(cmd, dbg_gmem_taken_stores);
+
+   tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 7);
+   tu_cs_emit(cs, CP_MEM_TO_MEM_0_NEG_B);
+   tu_cs_emit_qw(cs, result_iova);
+   tu_cs_emit_qw(cs, result_iova);
+   tu_cs_emit_qw(cs, global_iova(cmd, dbg_one));
 }
 
 static void
@@ -2815,6 +2831,22 @@ tu_end_load_store_cond_exec(struct tu_cmd_buffer *cmd,
                             struct tu_cs *cs, bool load)
 {
    tu_cond_exec_end(cs);
+
+   if (!unlikely(cmd->device->physical_device->instance->debug_flags &
+                 TU_DEBUG_LOG_SKIP_GMEM_OPS))
+      return;
+
+   uint64_t result_iova;
+   if (load)
+      result_iova = global_iova(cmd, dbg_gmem_total_loads);
+   else
+      result_iova = global_iova(cmd, dbg_gmem_total_stores);
+
+   tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 7);
+   tu_cs_emit(cs, CP_MEM_TO_MEM_0_NEG_B);
+   tu_cs_emit_qw(cs, result_iova);
+   tu_cs_emit_qw(cs, result_iova);
+   tu_cs_emit_qw(cs, global_iova(cmd, dbg_one));
 }
 
 void

src/freedreno/vulkan/tu_device.c

@@ -340,6 +340,7 @@ static const struct debug_control tu_debug_options[] = {
    { "dontcare_as_load", TU_DEBUG_DONT_CARE_AS_LOAD },
    { "rast_order", TU_DEBUG_RAST_ORDER },
    { "unaligned_store", TU_DEBUG_UNALIGNED_STORE },
+   { "log_skip_gmem_ops", TU_DEBUG_LOG_SKIP_GMEM_OPS },
    { NULL, 0 }
 };
 
@@ -1810,6 +1811,11 @@ tu_CreateDevice(VkPhysicalDevice physicalDevice,
    struct tu6_global *global = device->global_bo->map;
    tu_init_clear_blit_shaders(device);
    global->predicate = 0;
+   global->dbg_one = (uint32_t)-1;
+   global->dbg_gmem_total_loads = 0;
+   global->dbg_gmem_taken_loads = 0;
+   global->dbg_gmem_total_stores = 0;
+   global->dbg_gmem_taken_stores = 0;
    tu6_pack_border_color(&global->bcolor_builtin[VK_BORDER_COLOR_FLOAT_TRANSPARENT_BLACK],
                          &(VkClearColorValue) {}, false);
    tu6_pack_border_color(&global->bcolor_builtin[VK_BORDER_COLOR_INT_TRANSPARENT_BLACK],

src/freedreno/vulkan/tu_drm.c

@@ -1143,6 +1143,11 @@ tu_queue_submit(struct vk_queue *vk_queue, struct vk_queue_submit *submit)
                               submit->perf_pass_index : ~0;
    struct tu_queue_submit submit_req;
 
+   if (unlikely(queue->device->physical_device->instance->debug_flags &
+                 TU_DEBUG_LOG_SKIP_GMEM_OPS)) {
+      tu_dbg_log_gmem_load_store_skips(queue->device);
+   }
+
    pthread_mutex_lock(&queue->device->submit_mutex);
 
    VkResult ret = tu_queue_submit_create_locked(queue, submit,

src/freedreno/vulkan/tu_kgsl.c

@@ -362,6 +362,11 @@ tu_QueueSubmit(VkQueue _queue,
    TU_FROM_HANDLE(tu_syncobj, fence, _fence);
    VkResult result = VK_SUCCESS;
 
+   if (unlikely(queue->device->physical_device->instance->debug_flags &
+                 TU_DEBUG_LOG_SKIP_GMEM_OPS)) {
+      tu_dbg_log_gmem_load_store_skips(queue->device);
+   }
+
    uint32_t max_entry_count = 0;
    for (uint32_t i = 0; i < submitCount; ++i) {
       const VkSubmitInfo *submit = pSubmits + i;

src/freedreno/vulkan/tu_private.h

@@ -269,6 +269,7 @@ enum tu_debug_flags
    TU_DEBUG_RAST_ORDER = 1 << 14,
    TU_DEBUG_UNALIGNED_STORE = 1 << 15,
    TU_DEBUG_LAYOUT = 1 << 16,
+   TU_DEBUG_LOG_SKIP_GMEM_OPS = 1 << 15,
 };
 
 struct tu_instance
@@ -490,6 +491,12 @@ struct tu6_global
    /* To know when renderpass stats for autotune are valid */
    volatile uint32_t autotune_fence;
 
+   volatile uint32_t dbg_one;
+   volatile uint32_t dbg_gmem_total_loads;
+   volatile uint32_t dbg_gmem_taken_loads;
+   volatile uint32_t dbg_gmem_total_stores;
+   volatile uint32_t dbg_gmem_taken_stores;
+
    /* note: larger global bo will be used for customBorderColors */
    struct bcolor_entry bcolor_builtin[TU_BORDER_COLOR_BUILTIN], bcolor[];
 };

src/freedreno/vulkan/tu_util.c

@@ -31,6 +31,7 @@
 #include <string.h>
 
 #include "util/u_math.h"
+#include "util/timespec.h"
 #include "vk_enum_to_str.h"
 
 void PRINTFLIKE(3, 4)
@@ -216,3 +217,55 @@ tu_framebuffer_tiling_config(struct tu_framebuffer *fb,
    tu_tiling_config_update_pipe_layout(fb, device);
    tu_tiling_config_update_pipes(fb, device);
 }
+
+void
+tu_dbg_log_gmem_load_store_skips(struct tu_device *device)
+{
+   static uint32_t last_skipped_loads = 0;
+   static uint32_t last_skipped_stores = 0;
+   static uint32_t last_total_loads = 0;
+   static uint32_t last_total_stores = 0;
+   static struct timespec last_time = {};
+
+   pthread_mutex_lock(&device->submit_mutex);
+
+   struct timespec current_time;
+   clock_gettime(CLOCK_MONOTONIC, &current_time);
+
+   if (timespec_sub_to_nsec(&current_time, &last_time) > 1000 * 1000 * 1000) {
+      last_time = current_time;
+   } else {
+      pthread_mutex_unlock(&device->submit_mutex);
+      return;
+   }
+
+   struct tu6_global *global = device->global_bo->map;
+
+   uint32_t current_taken_loads = global->dbg_gmem_taken_loads;
+   uint32_t current_taken_stores = global->dbg_gmem_taken_stores;
+   uint32_t current_total_loads = global->dbg_gmem_total_loads;
+   uint32_t current_total_stores = global->dbg_gmem_total_stores;
+
+   uint32_t skipped_loads = current_total_loads - current_taken_loads;
+   uint32_t skipped_stores = current_total_stores - current_taken_stores;
+
+   uint32_t current_time_frame_skipped_loads = skipped_loads - last_skipped_loads;
+   uint32_t current_time_frame_skipped_stores = skipped_stores - last_skipped_stores;
+
+   uint32_t current_time_frame_total_loads = current_total_loads - last_total_loads;
+   uint32_t current_time_frame_total_stores = current_total_stores - last_total_stores;
+
+   mesa_logi("[GMEM] loads total: %u skipped: %.1f%%\n",
+         current_time_frame_total_loads,
+         current_time_frame_skipped_loads / (float) current_time_frame_total_loads * 100.f);
+   mesa_logi("[GMEM] stores total: %u skipped: %.1f%%\n",
+         current_time_frame_total_stores,
+         current_time_frame_skipped_stores / (float) current_time_frame_total_stores * 100.f);
+
+   last_skipped_loads = skipped_loads;
+   last_skipped_stores = skipped_stores;
+   last_total_loads = current_total_loads;
+   last_total_stores = current_total_stores;
+
+   pthread_mutex_unlock(&device->submit_mutex);
+}

src/freedreno/vulkan/tu_util.h

@@ -326,4 +326,7 @@ tu6_pack_border_color(struct bcolor_entry *bcolor, const VkClearColorValue *val,
 #undef PACK_F
 }
 
+void
+tu_dbg_log_gmem_load_store_skips(struct tu_device *device);
+
 #endif /* TU_UTIL_H */