From 90c90747b88b2eea8ca60be9b0b68f9d0d57215d Mon Sep 17 00:00:00 2001
From: Gabriel Wang <gabriel.wang@arm.com>
Date: Wed, 28 Feb 2024 20:02:30 +0000
Subject: [PATCH] add __cpu_perf__

---
 perf_counter.h   |  6 ++++--
 perfc_port_pmu.h | 37 +++++++++++++++++++++++++++++++++++++
 2 files changed, 41 insertions(+), 2 deletions(-)

diff --git a/perf_counter.h b/perf_counter.h
index 6904ec0..a3801b1 100644
--- a/perf_counter.h
+++ b/perf_counter.h
@@ -380,14 +380,14 @@ __asm(".global __ensure_systick_wrapper\n\t");
     E.g.
     \code
         while (1) {
-            __cpu_time__(100) {
+            __cpu_usage__(100) {
                 delay_us(5000);
             }
             delay_us(5000);
         }
     \endcode
  */
-#define __cpu_time__(__CNT, ...)                                                \
+#define __cpu_usage__(__CNT, ...)                                                \
     static int64_t SAFE_NAME(s_lTimestamp) = 0, SAFE_NAME(s_lTotal) = 0;        \
     static uint32_t s_wLoopCounter = (__CNT);                                   \
     using(float __usage__ = 0, ({                                               \
@@ -412,6 +412,8 @@ __asm(".global __ensure_systick_wrapper\n\t");
     ({SAFE_NAME(s_lTotal) += stop_task_cycle_counter();                         \
     s_wLoopCounter--;}))
 
+#define __cpu_time__    __cpu_usage__
+
 /*!
  * \addtogroup gBasicTimerService 1.2 Timer Service
  * \ingroup gBasic
diff --git a/perfc_port_pmu.h b/perfc_port_pmu.h
index fbb47d1..0ce7b68 100644
--- a/perfc_port_pmu.h
+++ b/perfc_port_pmu.h
@@ -23,6 +23,43 @@
 
 /*============================ MACROS ========================================*/
 /*============================ MACROFIED FUNCTIONS ===========================*/
+#define __cpu_perf__(__str, ...)                                                \
+    using(                                                                      \
+        struct {                                                                \
+            uint64_t dwNoInstr;                                                 \
+            int64_t lCycles;                                                    \
+            uint32_t wCalib;                                                    \
+            float fCPI;                                                         \
+        } __PERF_INFO__ = {0},                                                  \
+        ({                                                                      \
+            __PERF_INFO__.dwNoInstr = perfc_pmu_get_instruction_count();        \
+            __PERF_INFO__.wCalib = perfc_pmu_get_instruction_count()            \
+                                - __PERF_INFO__.dwNoInstr;                      \
+            __PERF_INFO__.dwNoInstr = perfc_pmu_get_instruction_count();        \
+        }),                                                                     \
+        ({                                                                      \
+            __PERF_INFO__.dwNoInstr = perfc_pmu_get_instruction_count()         \
+                              - __PERF_INFO__.dwNoInstr                         \
+                              - __PERF_INFO__.wCalib;                           \
+            __PERF_INFO__.fCPI = (float)(    (double)__PERF_INFO__.lCycles      \
+                                       /    (double)__PERF_INFO__.dwNoInstr);   \
+            if (__PLOOC_VA_NUM_ARGS(__VA_ARGS__) == 0) {                        \
+                printf( "\r\n"                                                  \
+                        "[Report for " __str "]\r\n"                            \
+                        "-----------------------------------------\r\n"         \
+                        "Instruction executed: %lld\r\n"                        \
+                        "Cycle Used: %lld\r\n"                                  \
+                        "Cycles per Instructions: %3.3f \r\n",                  \
+                        __PERF_INFO__.dwNoInstr,                                \
+                        __PERF_INFO__.lCycles,                                  \
+                        __PERF_INFO__.fCPI);                                    \
+             } else {                                                           \
+                __VA_ARGS__                                                     \
+             }                                                                  \
+        }))                                                                     \
+        __cycleof__("", { __PERF_INFO__.lCycles = __cycle_count__; })
+
+
 /*============================ TYPES =========================================*/
 typedef uint32_t perfc_global_interrupt_status_t;