fix(exp): Add missing function in expt_0516_2

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

3 files changed, 24 insertions, 14 deletions

diff --git a/expt_0516.cu b/expt_0516.cu
index f141027..82755ac 100644
--- a/expt_0516.cu
+++ b/expt_0516.cu
@@ -260,12 +260,12 @@ int main(int argc, char const *argv[]) {
   // custom_sort.testCalculation();
   P_ERR("\rApply custom length and test calculation");
   applyCudaSampleLength<<<1, 1>>>(sample_length);
-  testCustomCalculation<<<1, 1>>>(sample_length, sizeof(key_type) * 8);
+  // testCustomCalculation<<<1, 1>>>(sample_length);
   cudaDeviceSynchronize();
   P_ERR("\rApply custom length and test calculation completed");
 
   // Can remove this function if pass
-  CustomSort::testSelf(sample_length, sizeof(key_type) * 8);
+  // CustomSort::testSelf(sample_length);
 
   initSample<<<1, 1>>>(cudaSample, cudaPopulation);
   cudaDeviceSynchronize();
diff --git a/expt_0516_2.cu b/expt_0516_2.cu
index 0661e47..860e146 100644
--- a/expt_0516_2.cu
+++ b/expt_0516_2.cu
@@ -11,37 +11,47 @@ class CustomSort {
 public:
   explicit CustomSort(size_t length) : LENGTH(length) {}
   const size_t LENGTH;
-  static size_t fast_log(size_t a) {
-    float t = a;
-    return (((*(int *)&t) >> 23) + 1) & 127;
+
+  __device__ __host__ static size_t fast_log(size_t a) {
+#ifdef __CUDA_ARCH__
+    return (size_t)log2((double)a);
+#else
+    return (size_t)std::log2(a);
+#endif
   }
 
-  size_t calculate_index(size_t rank) const {
+  __device__ __host__ size_t calculate_index(size_t rank) const {
     size_t bit_low = (LENGTH + 1) >> fast_log(++rank) >> 1;
     return (((rank << 1) | 1) * bit_low - LENGTH - 1);
   }
 
-  size_t calculate_rank(size_t index) const {
+  __device__ __host__ size_t calculate_rank(size_t index) const {
     index++;
     size_t low_bit = index & (-index);
     return ((LENGTH + index) / low_bit) >> 1;
   }
 
-  __attribute__((unused)) void testCalculation() const {
+  __device__ __host__ void testCalculation() const {
     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) {
+      /*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);
     }
   }
 
   static void testSelf(size_t length) { CustomSort(length).testCalculation(); }
 };
+
+__global__ void testCustomCalculationWithPrint(size_t length) {
+  CustomSort(length).testCalculation();
+  printf("%lu pass\n", length);
+}
+
 #endif
 
 long pow_for_sample(long n) {
diff --git a/sortlib.cuh b/sortlib.cuh
index 5d32e4b..5bacc11 100644
--- a/sortlib.cuh
+++ b/sortlib.cuh
@@ -19,9 +19,9 @@ public:
 
   __device__ __host__ static size_t fast_log(size_t a) {
 #ifdef __CUDA_ARCH__
-    return log2((double)a);
+    return (size_t)log2((double)a);
 #else
-    return std::log2(a);
+    return (size_t)std::log2(a);
 #endif
   }
 
@@ -40,11 +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) {
+      /*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);
     }
   }