fix: Fix fast_log

Signed-off-by: KunoiSayami <[email protected]>

2 files changed, 32 insertions, 6 deletions

diff --git a/expt_0516_2.cu b/expt_0516_2.cu
index f0eb69a..0661e47 100644
--- a/expt_0516_2.cu
+++ b/expt_0516_2.cu
@@ -1,4 +1,7 @@
 // Experimental content: Test CustomSort Calculation
+#define DISABLE_TEST_WARNING
+#include "sortlib.cuh"
+
 #include <algorithm>
 #include <cstdio>
 
@@ -50,12 +53,24 @@ long pow_for_sample(long n) {
 }
 
 int main(int argc, char const *argv[]) {
-  auto max = 30L;
+  auto max = 32L;
+  auto cuda_only = false;
+
   if (argc > 1) {
     max = strtol(argv[1], nullptr, 10);
   }
-  for (int i = 4; i < max; i++) {
+
+  if (argc > 2) {
+    cuda_only = strtol(argv[2], nullptr, 10);
+  }
+
+  for (int i = 4; i <= max && !cuda_only; i++) {
     CustomSort::testSelf(pow_for_sample(i));
     printf("%d pass\n", i);
   }
+
+  for (int i = 4; i <= max; i++) {
+    testCustomCalculationWithPrint<<<1, 1>>>(pow_for_sample(i));
+    cudaDeviceSynchronize();
+  }
 }
diff --git a/sortlib.cuh b/sortlib.cuh
index fa306b0..5d32e4b 100644
--- a/sortlib.cuh
+++ b/sortlib.cuh
@@ -17,10 +17,12 @@ public:
   const int MOVE_OFFSET;
   const unsigned int STEP_LIMIT;
 
-  // IEEE 754
   __device__ __host__ static size_t fast_log(size_t a) {
-    float t = a;
-    return (((*(int *)&t) >> 23) + 1) & 127;
+#ifdef __CUDA_ARCH__
+    return log2((double)a);
+#else
+    return std::log2(a);
+#endif
   }
 
   __device__ __host__ size_t calculate_index(size_t rank) const {
@@ -38,7 +40,11 @@ public:
     for (size_t i = 0; i < LENGTH; i++) {
       auto left = calculate_index(i);
       auto right = calculate_rank(left - 1);
-
+      if (i + 1 == LENGTH || i + 1 != right) {
+        printf("%lu %lu %lu\n", i, left, right);
+        printf("%lu %lu %lu\n", LENGTH, fast_log(i + 1),
+               (LENGTH + 1) >> fast_log(i + 1) >> 1);
+      }
       assert(i + 1 == right);
     }
   }
@@ -176,6 +182,11 @@ __global__ void testCustomCalculation(size_t length) {
   CustomSort(length, sizeof(long) * 8).testCalculation();
 }
 
+__global__ void testCustomCalculationWithPrint(size_t length) {
+  CustomSort(length, sizeof(long) * 8).testCalculation();
+  printf("%lu pass\n", length);
+}
+
 class FactorySort {
 
 public: