refector: Rename last year experimental

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

1 files changed, 219 insertions, 0 deletions

diff --git a/expt_220809.cu b/expt_220809.cu
new file mode 100644
index 0000000..44d9992
--- /dev/null
+++ b/expt_220809.cu
@@ -0,0 +1,219 @@
+// Experimental content: test scale performance
+/*
+ * Sample output:
+scale: 2 time: 97.101822
+scale: 3 time: 80.391197
+scale: 4 time: 66.238525
+scale: 5 time: 70.442848
+scale: 6 time: 69.601219
+scale: 7 time: 69.934914
+scale: 8 time: 68.154366
+ */
+#include <algorithm>
+#include <cassert>
+#include <cstdio>
+#include <cstring>
+#include <iostream>
+#include <type_traits>
+#include <vector>
+
+constexpr size_t length = 2097152;
+std::vector<unsigned long long> population_vector, sample_vector;
+
+unsigned long long max_value = 0, min_value = 0xfffffffff;
+
+inline void store_into_vector(unsigned long long value) {
+  if (max_value < value) {
+    max_value = value;
+  }
+  if (min_value > value) {
+    min_value = value;
+  }
+  population_vector.push_back(value);
+}
+
+typedef unsigned long long key_type;
+
+typedef unsigned long long *key_type_ptr;
+
+constexpr long SAMPLE_LENGTH = 1024;
+
+constexpr long TEST_SIZE = 1024;
+
+__device__ key_type *cudaSampleItem, *cudaPopulationItem;
+__device__ bool *cdf_result;
+__device__ unsigned insert_value;
+#ifdef TEST_BOUNDS
+__device__ unsigned int index_max, index_min;
+#endif
+
+__device__ const key_type *cudaBinarySearch(key_type *start, key_type *end,
+                                            const key_type val) {
+  auto begin = start;
+  key_type *last_known_point = nullptr;
+  while (begin < end) {
+    auto mid = (end - begin) / 2;
+    auto mid_val = *(start + mid);
+    if (val == mid_val) {
+      return start + mid;
+    }
+    begin += ((mid_val - val) >> 63) & (mid + 1);
+    end -= (~((mid_val - val) >> 63)) & (mid + 1);
+    last_known_point = begin;
+  }
+  return last_known_point;
+}
+
+__device__ double sample_cdf(double x) {
+  auto it =
+      cudaBinarySearch(cudaSampleItem, cudaSampleItem + SAMPLE_LENGTH + 2, x);
+  if (it == cudaSampleItem + SAMPLE_LENGTH) {
+    return 1;
+  }
+  if (it == cudaSampleItem) {
+    return 0;
+  }
+  auto it_prev = it - 1;
+  return (double(it_prev - cudaSampleItem) +
+          (x - (double)*it_prev) / (double)(*it - *it_prev)) /
+         double(SAMPLE_LENGTH - 1);
+}
+
+__global__ void initStorage(unsigned scale, unsigned test_size) {
+  cudaFree(cdf_result);
+  cudaMalloc(&cdf_result, scale * test_size * sizeof(bool));
+  memset(cdf_result, 0, scale * test_size * sizeof(bool));
+  insert_value = 0;
+  // printf("init storage\n");
+#ifdef TEST_BOUNDS
+  index_max = 0;
+  index_min = 0x7fffffff;
+#endif
+}
+
+__global__ void init(key_type *sample_item, key_type *population_item) {
+  cudaSampleItem = sample_item;
+  cudaPopulationItem = population_item;
+  // cudaMalloc(&cudaSampleItem, (SAMPLE_LENGTH + 2) * sizeof(unsigned long
+  // long));
+  //  cudaMalloc(&Storage, TEST_SIZE * sizeof(key_type));
+  cdf_result = nullptr;
+}
+
+__global__ void kernel(unsigned long step, const double slice_size,
+                       const long split_size) {
+  for (int i = 0; i < step; i++) {
+    auto tid = step * gridDim.x * blockDim.x + blockIdx.x * blockDim.x +
+               threadIdx.x + i;
+
+    // printf("%d\n", tid);
+    auto index = (int)(sample_cdf(cudaPopulationItem[tid]) / slice_size);
+    if (cdf_result[index]) {
+      while (cdf_result[++index]) {
+        assert(index < split_size);
+      }
+    }
+    cdf_result[index] = true;
+    // atomicAdd(&insert_value, 1);
+
+#ifdef TEST_BOUNDS
+    while (true) {
+      unsigned tmp = index_min;
+      if (tmp < tid) {
+        break;
+      }
+      if (atomicCAS(&index_min, tmp, tid) == tmp) {
+        break;
+      }
+    }
+    while (true) {
+      unsigned tmp = index_max;
+      if (tmp > tid) {
+        break;
+      }
+      if (atomicCAS(&index_max, tmp, tid) == tmp) {
+        break;
+      }
+    }
+#endif
+  }
+}
+
+__global__ void print_function() {
+  // printf("%u\n", insert_value);
+#ifdef TEST_BOUNDS
+  printf("%u %u\n", index_min, index_max);
+#endif
+}
+
+int main(int argc, char const *argv[]) {
+  size_t test_size;
+  if (argc == 1) {
+    test_size = 1048576;
+  } else {
+    try {
+      test_size = std::stol(argv[1]);
+    } catch (...) {
+      return 1;
+    }
+  }
+
+  FILE *file = fopen("normal_distribution.txt", "r");
+  assert(file);
+  for (long long i; fscanf(file, "%lld ", &i) != EOF; store_into_vector(i))
+    ;
+  fclose(file);
+
+  assert(population_vector.size() == length);
+
+  sample_vector = std::vector<unsigned long long>(
+      population_vector.begin(), population_vector.begin() + SAMPLE_LENGTH);
+  sample_vector.push_back(min_value);
+  sample_vector.push_back(max_value);
+
+  std::sort(sample_vector.begin(), sample_vector.end());
+
+  key_type *cudaSample = nullptr, *cudaPopulation = nullptr;
+  cudaMalloc(&cudaSample, sizeof(key_type) * (SAMPLE_LENGTH + 2));
+  cudaMalloc(&cudaPopulation, sizeof(key_type) * TEST_SIZE);
+  cudaMemcpy(cudaSample, &sample_vector[0],
+             sizeof(key_type) * sample_vector.size(), cudaMemcpyHostToDevice);
+  cudaMemcpy(cudaPopulation, &population_vector[SAMPLE_LENGTH],
+             sizeof(key_type) * TEST_SIZE, cudaMemcpyHostToDevice);
+  // memcpy(sample_heap, sample_vector.cbegin(),sizeof(key_type) *(SAMPLE_LENGTH
+  // + 2));
+
+  init<<<1, 1>>>(cudaSample, cudaPopulation);
+  cudaDeviceSynchronize();
+
+  dim3 grid_dim = 16, block_dim = 64;
+
+  unsigned long step = test_size / (grid_dim.x * block_dim.x);
+  // printf("(0809)step: %lu\n", step);
+  assert(!(test_size % (grid_dim.x * block_dim.x)));
+
+  for (int scale = 2; scale <= 8; scale++) {
+    const long split_size = test_size * scale;
+    const auto slice_size = 1.0 / (double)split_size;
+    printf("scale: %i ", scale);
+    initStorage<<<1, 1>>>(scale, test_size);
+    cudaDeviceSynchronize();
+    cudaEvent_t start, stop;
+    cudaEventCreate(&start);
+    cudaEventCreate(&stop);
+    cudaEventRecord(start, nullptr);
+    kernel<<<grid_dim, block_dim>>>(step, slice_size, split_size);
+    cudaDeviceSynchronize();
+    cudaEventRecord(stop, nullptr);
+    cudaEventSynchronize(stop);
+    float time;
+    cudaEventElapsedTime(&time, start, stop);
+    cudaEventDestroy(start);
+    cudaEventDestroy(stop);
+
+    printf("time: %lf\n", time);
+    print_function<<<1, 1>>>();
+    cudaDeviceSynchronize();
+  }
+  return 0;
+}
\ No newline at end of file