tu: Implement utrace CS marker support

Adds support for emitting utrace markers into the CS, this allows
for useful debug information that can be decoded from a recorded
command stream.

Signed-off-by: Mark Collins <mark@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/18271>

src/freedreno/common/freedreno_pm4.h

@@ -39,6 +39,10 @@ extern "C" {
 #define CP_TYPE4_PKT 0x40000000
 #define CP_TYPE7_PKT 0x70000000
 
+#define CP_NOP_MESG 0x4D455347
+#define CP_NOP_BEGN 0x4245474E
+#define CP_NOP_END  0x454E4400
+
 /*
  * Helpers for pm4 pkt header building/parsing:
  */

src/freedreno/vulkan/tu_cs.c

@@ -435,3 +435,71 @@ tu_cs_reset(struct tu_cs *cs)
 
    cs->entry_count = 0;
 }
+
+void
+tu_cs_emit_debug_string(struct tu_cs *cs, const char *string, int len)
+{
+   assert(cs->mode == TU_CS_MODE_GROW);
+
+   /* max packet size is 0x3fff dwords */
+   len = MIN2(len, 0x3fff * 4);
+
+   tu_cs_emit_pkt7(cs, CP_NOP, align(len, 4) / 4);
+   const uint32_t *buf = (const uint32_t *) string;
+
+   tu_cs_emit_array(cs, buf, len / 4);
+   buf += len / 4;
+   len = len % 4;
+
+   /* copy remainder bytes without reading past end of input string */
+   if (len > 0) {
+      uint32_t w = 0;
+      memcpy(&w, buf, len);
+      tu_cs_emit(cs, w);
+   }
+}
+
+void
+tu_cs_emit_debug_magic_strv(struct tu_cs *cs,
+                            uint32_t magic,
+                            const char *fmt,
+                            va_list args)
+{
+   int fmt_len = vsnprintf(NULL, 0, fmt, args);
+   int len = 4 + fmt_len + 1;
+   char *string = (char *) malloc(len);
+
+   /* format: <magic><formatted string>\0 */
+   *(uint32_t *) string = magic;
+   vsnprintf(string + 4, fmt_len + 1, fmt, args);
+
+   tu_cs_emit_debug_string(cs, string, len);
+   free(string);
+}
+
+__attribute__((format(printf, 2, 3))) void
+tu_cs_emit_debug_msg(struct tu_cs *cs, const char *fmt, ...)
+{
+   va_list args;
+   va_start(args, fmt);
+   tu_cs_emit_debug_magic_strv(cs, CP_NOP_MESG, fmt, args);
+   va_end(args);
+}
+
+void
+tu_cs_trace_start(void *cs, const char *fmt, ...)
+{
+   va_list args;
+   va_start(args, fmt);
+   tu_cs_emit_debug_magic_strv((struct tu_cs *) cs, CP_NOP_BEGN, fmt, args);
+   va_end(args);
+}
+
+void
+tu_cs_trace_end(void *cs, const char *fmt, ...)
+{
+   va_list args;
+   va_start(args, fmt);
+   tu_cs_emit_debug_magic_strv((struct tu_cs *) cs, CP_NOP_END, fmt, args);
+   va_end(args);
+}

src/freedreno/vulkan/tu_cs.h

@@ -375,6 +375,40 @@ tu_cs_emit_call(struct tu_cs *cs, const struct tu_cs *target)
       tu_cs_emit_ib(cs, target->entries + i);
 }
 
+/**
+ * Emit a CP_NOP with a string tail into the command stream.
+ */
+void
+tu_cs_emit_debug_string(struct tu_cs *cs, const char *string, int len);
+
+void
+tu_cs_emit_debug_magic_strv(struct tu_cs *cs,
+                            uint32_t magic,
+                            const char *fmt,
+                            va_list args);
+
+__attribute__((format(printf, 2, 3))) void
+tu_cs_emit_debug_msg(struct tu_cs *cs, const char *fmt, ...);
+
+/**
+ * Emit a single message into the CS that denote the calling function and any
+ * optional printf-style parameters when utrace markers are enabled.
+ */
+#define TU_CS_DEBUG_MSG(CS, FORMAT_STRING, ...)                              \
+   do {                                                                      \
+      if (unlikely(u_trace_markers_enabled(&(CS)->device->trace_context)))   \
+         tu_cs_emit_debug_msg(CS, "%s(" FORMAT_STRING ")", __func__,         \
+                              ## __VA_ARGS__);                               \
+   } while (0)
+
+typedef struct tu_cs *tu_debug_scope;
+
+__attribute__((format(printf, 2, 3))) void
+tu_cs_trace_start(void *cs, const char *fmt, ...);
+
+__attribute__((format(printf, 2, 3))) void
+tu_cs_trace_end(void *cs, const char *fmt, ...);
+
 /* Helpers for bracketing a large sequence of commands of unknown size inside
  * a CP_COND_REG_EXEC packet.
  */

src/freedreno/vulkan/tu_tracepoints.py

@@ -45,7 +45,8 @@ tu_default_tps = []
 #
 
 def begin_end_tp(name, args=[], tp_struct=None, tp_print=None,
-                 tp_default_enabled=True):
+                 tp_default_enabled=True, marker_tp=True,
+                 queue_tp=True):
     global tu_default_tps
     if tp_default_enabled:
         tu_default_tps.append(name)
@@ -53,11 +54,13 @@ def begin_end_tp(name, args=[], tp_struct=None, tp_print=None,
                toggle_name=name,
                args=args,
                tp_struct=tp_struct,
-               tp_perfetto='tu_perfetto_start_{0}'.format(name),
-               tp_print=tp_print)
+               tp_perfetto='tu_perfetto_start_{0}'.format(name) if queue_tp else None,
+               tp_print=tp_print if queue_tp else None,
+               tp_markers='tu_cs_trace_start' if marker_tp else None)
     Tracepoint('end_{0}'.format(name),
                toggle_name=name,
-               tp_perfetto='tu_perfetto_end_{0}'.format(name))
+               tp_perfetto='tu_perfetto_end_{0}'.format(name),
+               tp_markers='tu_cs_trace_end' if marker_tp else None)
 
 begin_end_tp('cmd_buffer',
     args=[ArgStruct(type='const struct tu_cmd_buffer *', var='cmd')],