add hard swish bench (pytorch#234)

Author: bwasti

benchmarks/tensorexpr/benchmark.py

@@ -12,6 +12,7 @@
 #import pooling
 #import conv
 #import matmul
+import swish
 
 
 def main():

benchmarks/tensorexpr/framework.py

@@ -8,6 +8,7 @@
 class BenchmarkBase(object):
     def __init__(self, mode, device):
         self.mode = mode
+        self.deterministic = False
         self.device = device
         if mode == 'both':
             self.requires_grad = True

benchmarks/tensorexpr/swish.py

@@ -0,0 +1,52 @@
+import framework
+import scipy.special
+import numpy as np
+import torch
+
+
+class SwishBench(framework.Benchmark):
+    def __init__(self, mode, device, M, N):
+        super().__init__(mode, device)
+        self.M = M
+        self.N = N
+        self.data = self.rand([M, N], device=device, requires_grad=self.requires_grad)
+        self.inputs = [self.data]
+        self.zeros = torch.zeros(M, N, device=device)
+        self.six = self.zeros + 6.0
+        self.three = self.zeros + 3.0
+        self.sixth = self.zeros + 1.0 / 6.0
+
+    def forward(self, inp):
+        y = inp * (torch.min(torch.relu(inp), self.six) + self.three) * self.sixth
+        return y
+
+    def reference(self):
+        return self.numpy(self.forward(self.data))
+
+    def config(self):
+        return [self.M, self.N]
+
+    @staticmethod
+    def module():
+        return "swish"
+
+    def memory_workload(self):
+        if self.mode == "fwd":
+            sol_count = 1 + 1
+            algorithmic_count = 3 + 1
+        else:
+            sol_count = (1 + 1) + (1 + 1)
+            algorithmic_count = (3 + 1) + (3 + 1)
+
+        buffer_size = self.M * self.N * 4
+        return {
+            "sol": buffer_size * sol_count,
+            "algorithmic": buffer_size * algorithmic_count,
+        }
+
+    @staticmethod
+    def default_configs():
+        return [[128, 1 << 16]]
+
+
+framework.register_benchmark_class(SwishBench)