Merge pull request #6 from dingobits/darwin

Fix building issues on Darwin

CoreClockChecker/BoostClockChecker_arm.s

@@ -1,7 +1,12 @@
 .text
 .global clktsctest
 
+.global _clktsctest
+
+.balign 4
+
 /* x0 = iterations, return elapsed TSC in x0 */
+_clktsctest:
 clktsctest:
   sub sp, sp, #0x40
   stp x10, x11, [sp, #0x10]

GpuMemLatency/instruction_rate.c

@@ -214,6 +214,8 @@ float instruction_rate_test(cl_context context,
         float4_element_count, a_mem_obj, result_obj, A, result, opsPerIteration);
     fprintf(stderr, "INT8 G Multiplies/sec: %f\n", int8_mul_rate);
 
+    short checkExtensionSupport(const char *extension_name);
+
     if (checkExtensionSupport("cl_khr_fp64")) {
         fp64_instruction_rate_test(context, command_queue, thread_count, local_size, chase_iterations, float4_element_count,
             a_mem_obj, result_obj, A, result);
@@ -462,4 +464,4 @@ float fp16_instruction_rate_test(cl_context context,
     fprintf(stderr, "FP16 G FMAs/sec: %f : %f FP16 GFLOPs\n", gOpsPerSec, gOpsPerSec * 2);
 
     return gOpsPerSec;
-}
+}

MemoryBandwidth/MemoryBandwidth.c

@@ -18,9 +18,10 @@
 #include <sched.h>
 #include <math.h>
 #include <sys/mman.h>
-#include <sys/sysinfo.h>
 #include <errno.h>
+
 #ifdef NUMA
+#include <sys/sysinfo.h>
 #include <numa.h>
 #endif
 
@@ -39,7 +40,9 @@ typedef struct BandwidthTestThreadData {
     uint64_t start;
     float* arr;
     float bw; // written to by the thread
+    #ifdef NUMA
     cpu_set_t cpuset; // if numa set, will set affinity
+    #endif
 } BandwidthTestThreadData;
 
 float MeasureBw(uint64_t sizeKb, uint64_t iterations, uint64_t threads, int shared, int nopBytes, int coreNode, int memNode);

MemoryBandwidth/MemoryBandwidth_arm.s

@@ -18,6 +18,8 @@
 .global _flush_icache
 .global _readbankconflict
 
+.balign 4
+
 /* x0 = ptr to array (was rcx)
  * x1 = arr length (was rdx)
  * x2 = iterations (was r8)

MemoryLatency/MemoryLatency.c

@@ -17,6 +17,7 @@
 #include <numaif.h>
 #include <sys/sysinfo.h>
 #endif
+
 #include <errno.h>
 #include <sched.h>
 
@@ -204,7 +205,7 @@ int main(int argc, char* argv[]) {
         fprintf(stderr, "Usage: [-test <c/asm/tlb/mlp>] [-maxsizemb <max test size in MB>] [-iter <base iterations, default 100000000]\n");
     }
 
-#ifndef __MINGW32__
+#ifdef __linux__
     if (hugePages) {
        size_t hugePageSize = 1 << 21;
        size_t testSizeKb = singleSize ? singleSize : default_test_sizes[testSizeCount - 1];
@@ -595,10 +596,11 @@ float RunAsmTest(uint32_t size_kb, uint32_t iterations, uint32_t *preallocatedAr
     // Run test
     gettimeofday(&startTv, &startTz);
     #ifdef LONGPATTERN
-    if (longpattern) sum = longpatternlatencytest(scaled_iterations, A);
-    else 
-    #endif
+    if (longpattern)
+        sum = longpatternlatencytest(scaled_iterations, A);
+    else
         sum = latencytest(scaled_iterations, A);
+    #endif
     gettimeofday(&endTv, &endTz);
     uint64_t time_diff_ms = 1000 * (endTv.tv_sec - startTv.tv_sec) + ((endTv.tv_usec - startTv.tv_usec) / 1000);
     float latency = 1e6 * (float)time_diff_ms / (float)scaled_iterations;

MemoryLatency/MemoryLatency_arm.s

@@ -8,9 +8,20 @@
 .global stlftest128
 .global matchedstlftest
 
+.global _latencytest
+.global _longpatternlatencytest
+.global _preplatencyarr
+.global _stlftest
+.global _stlftest32
+.global _stlftest128
+.global _matchedstlftest
+
+.balign 4
+
 /* x0 = ptr to arr
    x1 = arr len
    convert values in array from array indexes to pointers */
+_preplatencyarr:
 preplatencyarr:
   sub sp, sp, #0x20
   stp x14, x15, [sp, #0x10]
@@ -30,6 +41,7 @@ preplatencyarr_loop:
 /* x0 = iteration count
    x1 = ptr to arr
    do pointer chasing for specified iteration count */
+_latencytest:
 latencytest:
   sub sp, sp, #0x20
   stp x14, x15, [sp, #0x10]
@@ -49,6 +61,7 @@ latencytest_loop:
    x1 = ptr to arr
    do pointer chasing with longer pattern, given different patterns
    within each cacheline */
+_longpatternlatencytest:
 longpatternlatencytest:
   sub sp, sp, #0x50
   stp x14, x15, [sp, #0x10]
@@ -86,6 +99,7 @@ longpatternlatencytest_loop_inc:
 
 /* x0 = iteration count
    x1 = ptr to arr. first 32-bit int = store offset, second = load offset */
+_stlftest:
 stlftest:
   sub sp, sp, #0x40
   stp x14, x15, [sp, #0x10]
@@ -114,6 +128,7 @@ stlftest_loop:
   add sp, sp, #0x40
   ret
 
+_stlftest32:
 stlftest32:
   sub sp, sp, #0x40
   stp x14, x15, [sp, #0x10]
@@ -144,6 +159,7 @@ stlftest32_loop:
 
 /* x0 = iteration count
    x1 = ptr to arr. first 32-bit int = store offset, second = load offset */
+_stlftest128:
 stlftest128:
   sub sp, sp, #0x40
   stp x14, x15, [sp, #0x10]
@@ -172,6 +188,7 @@ stlftest128_loop:
   add sp, sp, #0x40
   ret
 
+_matchedstlftest:
 matchedstlftest:
   sub sp, sp, #0x40
   stp x14, x15, [sp, #0x10]