ECE 408 sem project

kernels_req_5/local_softmax.cuh

@@ -0,0 +1,37 @@
+#ifndef __SOFTMAX_KERNEL_CUH__
+#define __SOFTMAX_KERNEL_CUH__
+
+#include <cuda_runtime.h>
+#include <math.h>
+#include <float.h>
+#include "../utils/cuda_utils.cuh"
+
+__global__ void local_softmax_forward_kernel(float* out, float inv_temperature, const float* inp, int N, int T) {
+    // Implement this
+    int x = blockDim.x * blockIdx.x + threadIdx.x;
+
+    if(x < N*T){
+        //row offset based on WIndow size
+        int rowoff = x * WINDOW;
+
+        float maxval = -FLT_MAX;
+        for (int i = 0; i < WINDOW; i++) {
+            maxval = fmaxf(maxval, inp[i + rowoff]);
+        }
+
+        float sumval = 0.0f;
+        //Traverse Window Length Again
+        for (int i = 0; i < WINDOW; i++) {
+            float ev = expf(inv_temperature * (inp[i + rowoff] - maxval));
+            sumval += ev;
+            out[i + rowoff] = ev;
+        }
+
+        float norm = 1.0f / sumval;
+        for (int i = 0; i < WINDOW; i++) {
+            out[i + rowoff] *= norm;
+        }
+    }
+}
+
+#endif // __SOFTMAX_KERNEL_CUH__