tp2

TP2_reduction/Reduce.cu

@@ -37,13 +37,16 @@ __global__ void reduce_kernel( int n, const int *in_buffer, int *out_buffer, con
 {
   // Allocate shared memory inside the block.
   extern __shared__ int s_mem[];
-float my_sum=0;
+  int my_sum=0;
   // The range of data to work with.
   int2 range = block_ranges[blockIdx.x];
 
   // Compute the sum of my elements.
-  
   // TODO: fill-in that section of the code
+  for(int i = range.x + threadIdx.x; i < range.y; i+=blockDim.x)
+  {
+    my_sum += in_buffer[i];
+  }
 
   // Copy my sum in shared memory.
   s_mem[threadIdx.x] = my_sum;
@@ -54,6 +57,16 @@ float my_sum=0;
   // Compute the sum inside the block.
   
   // TODO: fill-in that section of the code
+  
+  for(int e = 1; 1<<e <= blockDim.x; e++)
+  {
+    int a = threadIdx.x+(1<<(e-1));
+    if(threadIdx.x%(1<<e) == 0 && a < blockDim.x)
+    {
+      s_mem[threadIdx.x] += s_mem[a];
+    }
+    __syncthreads();
+  }
 
   // The first thread of the block stores its result.
   if( threadIdx.x == 0 )
@@ -124,7 +137,7 @@ __global__ void reduce_kernel_optimized( int n, const int *in_buffer, int *out_b
 {
   // The number of warps in the block.
   const int NUM_WARPS = BLOCK_DIM / WARP_SIZE;
-float my_sum=0;
+  int my_sum=0;
   // Allocate shared memory inside the block.
   __shared__ volatile int s_mem[BLOCK_DIM];
 
@@ -138,12 +151,16 @@ float my_sum=0;
   // Compute the sum of my elements.
   
   // TODO: fill-in that section of the code
+  for(int i = range.x + threadIdx.x; i < range.y; i+=blockDim.x)
+  {
+    my_sum += in_buffer[i];
+  }
 
   // Copy my sum in shared memory.
   s_mem[threadIdx.x] = my_sum;
 
   // Compute the sum inside each warp.
- 
+
   // TODO: fill-in that section of the code
   
   // Each warp leader stores the result for the warp.