Add support for benchmarking on mps3

Author: bremoran

nix/m55-an547-arm-none-eabi/default.nix

@@ -8,15 +8,14 @@
 
 stdenvNoCC.mkDerivation {
   pname = "mldsa-native-m55-an547";
-  version = "main-2025-10-02";
-
+  version = "main-2025-10-13";
 
   # Fetch platform files from pqmx (envs/m55-an547)
   src = fetchFromGitHub {
-    owner = "slothy-optimizer";
+    owner = "bremoran";
     repo = "pqmx";
-    rev = "4ed493d3cf2af62a08fd9fe36c3472a0dc50ad9f";
-    hash = "sha256-jLIqwknjRwcoDeEAETlMhRqZQ5a3QGCDZX9DENelGeQ=";
+    rev = "3b4177cb76a614ed1887bcc2f90e571469d6fe2f";
+    hash = "sha256-27ur5JYtZbwQ0fnrp/0Cj/R60TaZwouimFgOJGPjpMk=";
   };
 
   dontBuild = true;

scripts/tests

@@ -961,7 +961,7 @@ def cli():
         "-c",
         "--cycles",
         help="Method for counting clock cycles. PMU requires (user-space) access to the Arm Performance Monitor Unit (PMU). PERF requires a kernel with perf support. MAC works on some Apple platforms, at least Apple M1.",
-        choices=["NO", "PMU", "PERF", "MAC"],
+        choices=["NO", "PMU", "PERF", "MAC", "SYSTICK"],
         type=str.upper,
         required=True,
     )

test/baremetal/platform/m55-an547/platform.mk

@@ -59,3 +59,5 @@ LDFLAGS += \
 # Extra sources to be included in test binaries
 EXTRA_SOURCES = $(wildcard $(M55_AN547_PATH)/*.c)
 EXEC_WRAPPER := $(realpath $(PLATFORM_PATH)/exec_wrapper.py)
+
+CYCLES ?= SYSTICK

test/bench_mldsa.c

@@ -56,7 +56,7 @@ static void print_percentile_legend(void)
   printf("%21s", "percentile");
   for (i = 0; i < sizeof(percentiles) / sizeof(percentiles[0]); i++)
   {
-    printf("%7d", percentiles[i]);
+    printf("%9d", percentiles[i]);
   }
   printf("\n");
 }
@@ -67,7 +67,7 @@ static void print_percentiles(const char *txt, uint64_t cyc[NTESTS])
   printf("%10s percentiles:", txt);
   for (i = 0; i < sizeof(percentiles) / sizeof(percentiles[0]); i++)
   {
-    printf("%7" PRIu64, (cyc)[NTESTS * percentiles[i] / 100] / NITERATIONS);
+    printf("%9" PRIu64, (cyc)[NTESTS * percentiles[i] / 100] / NITERATIONS);
   }
   printf("\n");
 }

test/hal/hal.c

@@ -94,6 +94,41 @@ uint64_t get_cyclecounter(void)
   __asm __volatile("mrs    %[retval], pmccntr_el0\n" : [retval] "=r"(retval));
   return retval;
 }
+#elif defined(__ARM_ARCH_8_1M_MAIN__) || defined(__ARM_FEATURE_MVE)
+#include <ARMCM55.h>
+#include <system_ARMCM55.h>
+#include <uart.h>
+#include "pmu_armv8.h"
+
+volatile uint64_t v8m_pmu_cycles;
+void DebugMon_Handler(void) {
+  uint32_t ovsclr;
+  ARM_PMU_CNTR_Disable(PMU_CNTENSET_CCNTR_ENABLE_Msk);
+  v8m_pmu_cycles += ARM_PMU_Get_CCNTR();
+  ovsclr = ARM_PMU_Get_CNTR_OVS();
+  ARM_PMU_Set_CNTR_OVS(ovsclr);
+  ARM_PMU_CYCCNT_Reset();
+  ARM_PMU_CNTR_Enable(PMU_CNTENSET_CCNTR_ENABLE_Msk);
+}
+void enable_cyclecounter(void)
+{
+  CoreDebug->DEMCR |= CoreDebug_DEMCR_TRCENA_Msk | CoreDebug_DEMCR_MON_EN_Msk;
+  ARM_PMU_Enable();
+
+  v8m_pmu_cycles = 0;
+  ARM_PMU_CYCCNT_Reset();
+}
+void disable_cyclecounter(void)
+{
+  ARM_PMU_Disable();
+  CoreDebug->DEMCR &= (~CoreDebug_DEMCR_TRCENA_Msk) & (~CoreDebug_DEMCR_MON_EN_Msk);
+}
+uint64_t get_cyclecounter(void)
+{
+  v8m_pmu_cycles += ARM_PMU_Get_CCNTR();
+  ARM_PMU_CYCCNT_Reset();
+  return v8m_pmu_cycles;
+}
 
 #else
 #error PMU_CYCLES option only supported on x86_64 and AArch64
@@ -305,6 +340,56 @@ uint64_t get_cyclecounter(void)
   return g_counters[2];
 }
 
+#elif defined(SYSTICK_CYCLES)
+#include <ARMCM55.h>
+#include <system_ARMCM55.h>
+static volatile uint32_t systick_overflows;
+void SysTick_Handler(void)
+{
+  uint32_t cnt;
+  uint32_t fail;
+  do {
+    __asm__ volatile (
+      "ldrex %[val], [%[guard]]\n"
+      "addw %[val], %[val], #1\n"
+      "strex %[fail], %[val], [%[guard]]\n"
+      :[fail] "=r" (fail), [val] "=&r" (cnt)
+      :[guard] "r" (&systick_overflows) : "memory");
+  } while (fail);
+}
+
+void enable_cyclecounter(void) {
+  uint32_t cnt;
+  uint32_t fail;
+  do {
+    __asm__ volatile ("ldrex.w %[val], [%[guard]]"
+      :[val] "=&r" (cnt)
+      :[guard] "r" (&systick_overflows));
+    cnt = 0;
+    __asm__ volatile ("strex.w %[fail], %[val], [%[guard]]"
+      :[fail] "=r" (fail), [val] "=&r" (cnt)
+      :[guard] "r" (&systick_overflows) : "memory");
+  } while (fail);
+  SysTick_Config(0xFFFFFFu);
+}
+
+void disable_cyclecounter(void) { return; }
+
+uint64_t get_cyclecounter(void) {
+  uint32_t val;
+  uint32_t cnt;
+  uint32_t fail;
+  do {
+    __asm__ volatile (
+      "ldrex %[cnt], [%[guard]]\n"
+      "ldr %[val], [%[systv]]\n"
+      "strex %[fail], %[cnt], [%[guard]]\n"
+      :[fail] "=r" (fail), [cnt] "=&r" (cnt), [val] "=r" (val)
+      :[guard] "r" (&systick_overflows), [systv] "r" (&(SysTick->VAL)): "memory");
+  } while (fail);
+  return ((uint64_t)cnt + 1) * 16777216llu - val;
+}
+
 #else
 
 void enable_cyclecounter(void) { return; }

test/mk/config.mk

@@ -74,6 +74,10 @@ ifeq ($(CYCLES),MAC)
 	CFLAGS += -DMAC_CYCLES
 endif
 
+ifeq ($(CYCLES),SYSTICK)
+	CFLAGS += -DSYSTICK_CYCLES
+endif
+
 ##############################
 # Include retained variables #
 ##############################