Merge pull request #38 from sifive/benchmarking

Benchmarking

15 files changed, 282 insertions, 32 deletions

diff --git a/Makefile b/Makefile
index 0515511..ffc2e77 100644
--- a/Makefile
+++ b/Makefile
@@ -64,7 +64,7 @@ openocd: $(openocd_dest)/bin/openocd
 
 $(toolchain_dest)/bin/$(target32)-gcc: $(toolchain_srcdir)
 	mkdir -p $(toolchain32_wrkdir)
-	cd $(toolchain32_wrkdir); $(toolchain_srcdir)/configure --prefix=$(toolchain_dest) --with-arch=rv32ima --with-abi=ilp32
+	cd $(toolchain32_wrkdir); $(toolchain_srcdir)/configure --prefix=$(toolchain_dest) --with-arch=rv32imac --with-abi=ilp32
 	$(MAKE) -C $(toolchain32_wrkdir)
 
 $(openocd_dest)/bin/openocd: $(openocd_srcdir)
diff --git a/README.md b/README.md
index 58664b0..d0dfc29 100644
--- a/README.md
+++ b/README.md
@@ -42,6 +42,53 @@ cd freedom-e-sdk
 make help
 ```
 
+### Benchmarking ###
+
+#### Dhrystone ####
+
+After setting up the software and debug toolchains, you can build and
+execute everyone's favorite benchmark as follows:
+
+- Compile the benchmark with the command `make software PROGRAM=dhrystone`.
+- Run on the HiFive1 board with the command `make upload PROGRAM=dhrystone`.
+  This will take a few minutes.  Sample output is provided below.
+- Compute DMIPS by dividing the Dhrystones per Second result by 1757, which
+  was the VAX 11/780's performance.  In the example below, 729927 / 1757 =
+  415 DMIPS.
+- Compute DMIPS/MHz by dividing by the clock rate: in the example below,
+  415 / 260 = 1.60 DMIPS/MHz.
+
+```
+core freq at 259830579 Hz
+
+Dhrystone Benchmark, Version 2.1 (Language: C)
+
+<snip>
+
+Microseconds for one run through Dhrystone: 1.3
+Dhrystones per Second:                      729927.0
+```
+
+#### CoreMark ####
+
+We cannot distribute the CoreMark benchmark, but following are instructions
+to download and run the benchmark on the HiFive1 board:
+
+- Download CoreMark from EEMBC's web site and extract the archive from
+  http://www.eembc.org/coremark/download.php.
+- Copy the following files from the extracted archive into the
+  `software/coremark` directory in this repository:
+  - `core_list_join.c`
+  - `core_main.c`
+  - `coremark.h`
+  - `core_matrix.c`
+  - `core_state.c`
+  - `core_util.c`
+- Compile the benchmark with the command `make software PROGRAM=coremark`.
+- Run on the HiFive1 board with the command `make upload PROGRAM=coremark`.
+- Divide the reported Iterations/Sec by the reported core frequency in MHz to
+  obtain a CoreMarks/MHz value.
+
 ### For More Information ###
 
 Documentation, Forums, and much more available at
diff --git a/bsp/env/entry.S b/bsp/env/entry.S
index cbf26eb..1f5de24 100644
--- a/bsp/env/entry.S
+++ b/bsp/env/entry.S
@@ -7,6 +7,7 @@
 #include "sifive/bits.h"
 
   .section      .text.entry	
+  .align 2
   .global trap_entry
 trap_entry:
   addi sp, sp, -32*REGBYTES
@@ -46,7 +47,7 @@ trap_entry:
   csrr a0, mcause
   csrr a1, mepc
   mv a2, sp
-  jal handle_trap
+  call handle_trap
   csrw mepc, a0
 
   # Remain in M-mode after mret
@@ -90,6 +91,7 @@ trap_entry:
 
 .weak handle_trap
 handle_trap:
-  j handle_trap
+1:
+  j 1b
 	
 #endif
diff --git a/bsp/env/freedom-e300-arty/init.c b/bsp/env/freedom-e300-arty/init.c
index c766e98..35b1104 100644
--- a/bsp/env/freedom-e300-arty/init.c
+++ b/bsp/env/freedom-e300-arty/init.c
@@ -9,11 +9,30 @@
 extern int main(int argc, char** argv);
 extern void trap_entry();
 
-uint32_t get_cpu_freq()
+static unsigned long get_cpu_freq()
 {
   return 65000000;
 }
 
+unsigned long get_timer_freq()
+{
+  return get_cpu_freq();
+}
+
+uint64_t get_timer_value()
+{
+#if __riscv_xlen == 32
+  while (1) {
+    uint32_t hi = read_csr(mcycleh);
+    uint32_t lo = read_csr(mcycle);
+    if (hi == read_csr(mcycleh))
+      return ((uint64_t)hi << 32) | lo;
+  }
+#else
+  return read_csr(mcycle);
+#endif
+}
+
 static void uart_init(size_t baud_rate)
 {
   GPIO_REG(GPIO_IOF_SEL) &= ~IOF0_UART0_MASK;
@@ -59,8 +78,6 @@ void _init()
   printf("core freq at %d Hz\n", get_cpu_freq());
 
   write_csr(mtvec, &trap_entry);
-  
-  //  _exit(main(0, NULL));
 }
 
 
diff --git a/bsp/env/freedom-e300-arty/platform.h b/bsp/env/freedom-e300-arty/platform.h
index 1f62956..d5d6dda 100644
--- a/bsp/env/freedom-e300-arty/platform.h
+++ b/bsp/env/freedom-e300-arty/platform.h
@@ -119,4 +119,7 @@
 #define HAS_BOARD_BUTTONS
 #include "hifive1.h"
 
+unsigned long get_timer_freq(void);
+uint64_t get_timer_value(void);
+
 #endif /* _SIFIVE_PLATFORM_H */
diff --git a/bsp/env/freedom-e300-hifive1/init.c b/bsp/env/freedom-e300-hifive1/init.c
index c088079..71e1659 100644
--- a/bsp/env/freedom-e300-hifive1/init.c
+++ b/bsp/env/freedom-e300-hifive1/init.c
@@ -5,21 +5,43 @@
 #include "platform.h"
 #include "encoding.h"
 
-uint32_t cpu_freq = 0;
-
 extern int main(int argc, char** argv);
 extern void trap_entry();
 
-uint32_t mtime_lo(void)
+static unsigned long mtime_lo(void)
+{
+  return *(volatile unsigned long *)(CLINT_BASE_ADDR + CLINT_MTIME);
+}
+
+#ifdef __riscv32
+
+static uint32_t mtime_hi(void)
 {
-  return *(volatile uint32_t *)(CLINT_BASE_ADDR + CLINT_MTIME);
+  return *(volatile uint32_t *)(CLINT_BASE_ADDR + CLINT_MTIME + 4);
 }
 
-uint32_t mcycle_lo(void)
+uint64_t get_timer_value()
 {
-  uint32_t t;
-  asm volatile ("csrr %0, mcycle" : "=r" (t));
-  return t;
+  while (1) {
+    uint32_t hi = mtime_hi();
+    uint32_t lo = mtime_lo();
+    if (hi == mtime_hi())
+      return ((uint64_t)hi << 32) | lo;
+  }
+}
+
+#else /* __riscv32 */
+
+uint64_t get_timer_value()
+{
+  return mtime_lo();
+}
+
+#endif
+
+unsigned long get_timer_freq()
+{
+  return 32768;
 }
 
 static void use_hfrosc(int div, int trim)
@@ -116,21 +138,40 @@ static void use_default_clocks()
   use_hfrosc(4, 16);
 }
 
-void measure_cpu_freq(size_t n, size_t mtime_freq)
+static unsigned long __attribute__((noinline)) measure_cpu_freq(size_t n)
 {
-  uint32_t start_mtime = mtime_lo();
-  uint32_t start_mcycle = mcycle_lo();
+  unsigned long start_mtime, delta_mtime;
+  unsigned long mtime_freq = get_timer_freq();
+
+  // Don't start measuruing until we see an mtime tick
+  unsigned long tmp = mtime_lo();
+  do {
+    start_mtime = mtime_lo();
+  } while (start_mtime == tmp);
 
-  while (mtime_lo() - start_mtime < n) ;
+  unsigned long start_mcycle = read_csr(mcycle);
 
-  uint32_t end_mtime = mtime_lo();
-  uint32_t end_mcycle = mcycle_lo();
+  do {
+    delta_mtime = mtime_lo() - start_mtime;
+  } while (delta_mtime < n);
 
-  cpu_freq = (end_mcycle-start_mcycle)/n*mtime_freq;
+  unsigned long delta_mcycle = read_csr(mcycle) - start_mcycle;
+
+  return (delta_mcycle / delta_mtime) * mtime_freq
+         + ((delta_mcycle % delta_mtime) * mtime_freq) / delta_mtime;
 }
 
-uint32_t get_cpu_freq()
+unsigned long get_cpu_freq()
 {
+  static uint32_t cpu_freq;
+
+  if (!cpu_freq) {
+    // warm up I$
+    measure_cpu_freq(1);
+    // measure for real
+    cpu_freq = measure_cpu_freq(10);
+  }
+
   return cpu_freq;
 }
 
@@ -178,7 +219,6 @@ void _init()
 {
   use_default_clocks();
   use_pll(0, 0, 1, 31, 1);
-  measure_cpu_freq(1000, 32768);
   uart_init(115200);
 
   printf("core freq at %d Hz\n", get_cpu_freq());
@@ -188,8 +228,6 @@ void _init()
     write_csr(mstatus, MSTATUS_FS); // allow FPU instructions without trapping
     write_csr(fcsr, 0); // initialize rounding mode, undefined at reset
   }
-
-  //_exit(main(0, NULL));
 }
 
 void _fini()
diff --git a/bsp/env/freedom-e300-hifive1/link.lds b/bsp/env/freedom-e300-hifive1/link.lds
index e224273..90e5c8f 100644
--- a/bsp/env/freedom-e300-hifive1/link.lds
+++ b/bsp/env/freedom-e300-hifive1/link.lds
@@ -26,6 +26,8 @@ SECTIONS
 
   .text           :
   {
+    *(.text.unlikely .text.unlikely.*)
+    *(.text.startup .text.startup.*)
     *(.text .text.*)
     *(.gnu.linkonce.t.*)
   } >flash AT>flash :flash
diff --git a/bsp/env/freedom-e300-hifive1/platform.h b/bsp/env/freedom-e300-hifive1/platform.h
index eca708e..63efc9e 100644
--- a/bsp/env/freedom-e300-hifive1/platform.h
+++ b/bsp/env/freedom-e300-hifive1/platform.h
@@ -126,6 +126,8 @@
 
 #include "hifive1.h"
 
-uint32_t get_cpu_freq();
+unsigned long get_cpu_freq(void);
+unsigned long get_timer_freq(void);
+uint64_t get_timer_value(void);
 
 #endif /* _SIFIVE_PLATFORM_H */
diff --git a/bsp/env/start.S b/bsp/env/start.S
index 77e223d..b526411 100644
--- a/bsp/env/start.S
+++ b/bsp/env/start.S
@@ -47,5 +47,8 @@ _start:
 1:
 #endif
 
+	/* argc = argv = 0 */
+	li a0, 0
+	li a1, 0
 	call main
 	tail exit
diff --git a/software/coremark/.gitignore b/software/coremark/.gitignore
new file mode 100644
index 0000000..6bd8438
--- /dev/null
+++ b/software/coremark/.gitignore
@@ -0,0 +1,9 @@
+/*.o
+/coremark
+/core_list_join.c
+/core_main.c
+/coremark.h
+/core_matrix.c
+/core_state.c
+/core_util.c
+/trans.c
diff --git a/software/coremark/Makefile b/software/coremark/Makefile
new file mode 100644
index 0000000..b57151e
--- /dev/null
+++ b/software/coremark/Makefile
@@ -0,0 +1,20 @@
+TARGET := coremark
+
+C_SRCS := \
+	core_list_join.c \
+	core_main.c \
+	core_matrix.c \
+	core_state.c \
+	core_util.c \
+	core_portme.c \
+
+HEADERS := \
+	coremark.h \
+	core_portme.h \
+
+CFLAGS := -O2 -fno-common -funroll-loops -finline-functions --param max-inline-insns-auto=20 -falign-functions=4 -falign-jumps=4 -falign-loops=4
+CFLAGS += -DFLAGS_STR=\""$(CFLAGS)"\"
+CFLAGS += -DITERATIONS=10000 -DPERFORMANCE_RUN=1
+
+BSP_BASE = ../../bsp
+include $(BSP_BASE)/env/common.mk
diff --git a/software/coremark/core_portme.c b/software/coremark/core_portme.c
new file mode 100644
index 0000000..cbf1396
--- /dev/null
+++ b/software/coremark/core_portme.c
@@ -0,0 +1,52 @@
+#include <stdio.h>
+#include <stdlib.h>
+#include "coremark.h"
+#include "platform.h"
+#include "encoding.h"
+
+#if VALIDATION_RUN
+	volatile ee_s32 seed1_volatile=0x3415;
+	volatile ee_s32 seed2_volatile=0x3415;
+	volatile ee_s32 seed3_volatile=0x66;
+#endif
+
+#if PERFORMANCE_RUN
+	volatile ee_s32 seed1_volatile=0x0;
+	volatile ee_s32 seed2_volatile=0x0;
+	volatile ee_s32 seed3_volatile=0x66;
+#endif
+
+#if PROFILE_RUN
+	volatile ee_s32 seed1_volatile=0x8;
+	volatile ee_s32 seed2_volatile=0x8;
+	volatile ee_s32 seed3_volatile=0x8;
+#endif
+
+volatile ee_s32 seed4_volatile=ITERATIONS;
+volatile ee_s32 seed5_volatile=0;
+
+static CORE_TICKS t0, t1;
+
+void start_time(void)
+{
+  t0 = get_timer_value();
+}
+
+void stop_time(void)
+{
+  t1 = get_timer_value();
+}
+
+CORE_TICKS get_time(void)
+{
+  return t1 - t0;
+}
+
+secs_ret time_in_secs(CORE_TICKS ticks)
+{
+  // scale timer down to avoid uint64_t -> double conversion in RV32
+  int scale = 256;
+  uint32_t delta = ticks / scale;
+  uint32_t freq = get_timer_freq() / scale;
+  return delta / (double)freq;
+}
diff --git a/software/coremark/core_portme.h b/software/coremark/core_portme.h
new file mode 100644
index 0000000..82298d6
--- /dev/null
+++ b/software/coremark/core_portme.h
@@ -0,0 +1,58 @@
+#ifndef FESDK_CORE_PORTME_H
+#define FESDK_CORE_PORTME_H
+
+#include <stdint.h>
+#include <stddef.h>
+
+#define HAS_FLOAT 1
+#define HAS_TIME_H 1
+#define USE_CLOCK 1
+#define HAS_STDIO 1
+#define HAS_PRINTF 1
+#define SEED_METHOD SEED_VOLATILE
+#define CORE_TICKS uint64_t
+#define ee_u8 uint8_t
+#define ee_u16 uint16_t
+#define ee_u32 uint32_t
+#define ee_s16 int16_t
+#define ee_s32 int32_t
+#define ee_ptr_int uintptr_t
+#define ee_size_t size_t
+#define COMPILER_FLAGS FLAGS_STR
+
+#define align_mem(x) (void *)(((ee_ptr_int)(x) + sizeof(ee_u32) - 1) & -sizeof(ee_u32))
+
+#ifdef __GNUC__
+# define COMPILER_VERSION "GCC"__VERSION__
+#else
+# error
+#endif
+
+#define MEM_METHOD MEM_STACK
+#define MEM_LOCATION "STACK"
+
+#define MAIN_HAS_NOARGC 0
+#define MAIN_HAS_NORETURN 0
+
+#define MULTITHREAD 1
+#define USE_PTHREAD 0
+#define USE_FORK 0
+#define USE_SOCKET 0
+
+#define default_num_contexts MULTITHREAD
+
+typedef int core_portable;
+static void portable_init(core_portable *p, int *argc, char *argv[]) {}
+static void portable_fini(core_portable *p) {}
+
+#if !defined(PROFILE_RUN) && !defined(PERFORMANCE_RUN) && !defined(VALIDATION_RUN)
+#if (TOTAL_DATA_SIZE==1200)
+#define PROFILE_RUN 1
+#elif (TOTAL_DATA_SIZE==2000)
+#define PERFORMANCE_RUN 1
+#else
+#define VALIDATION_RUN 1
+#endif
+#endif
+
+#endif
diff --git a/software/dhrystone/Makefile b/software/dhrystone/Makefile
index a55b1ec..78a7b23 100644
--- a/software/dhrystone/Makefile
+++ b/software/dhrystone/Makefile
@@ -5,7 +5,7 @@ C_SRCS := dhry_stubs.c dhry_printf.c
 HEADERS := dhry.h
 
 DHRY_SRCS := dhry_1.c dhry_2.c
-DHRY_CFLAGS := -O2 -DTIME -fno-inline -fno-builtin-printf -Wno-implicit
+DHRY_CFLAGS := -O2 -DTIME -fno-inline -fno-builtin-printf -Wno-implicit -march=rv32ima
 
 XLEN ?= 32
 CFLAGS := -Os -fno-common
diff --git a/software/dhrystone/dhry_stubs.c b/software/dhrystone/dhry_stubs.c
index d3bd14c..0616f86 100644
--- a/software/dhrystone/dhry_stubs.c
+++ b/software/dhrystone/dhry_stubs.c
@@ -3,16 +3,13 @@
 /* The functions in this file are only meant to support Dhrystone on an
  * embedded RV32 system and are obviously incorrect in general. */
 
-// return the cycle counter as though it were the current time
 long time(void)
 {
-  long t;
-  asm volatile ("csrr %0, mcycle" : "=r" (t));
-  return t / (get_cpu_freq() / 1000);
+  return get_timer_value() / get_timer_freq();
 }
 
 // set the number of dhrystone iterations
 void __wrap_scanf(const char* fmt, int* n)
 {
-  *n = 1500000;
+  *n = 100000000;
 }