Ready for PR - Fixing float

Author: Daniel Holanda Noronha

qtorch/quant/quant_cuda/float_kernel.cu

@@ -10,12 +10,6 @@ __global__ void float_kernel_stochastic(float* __restrict__ a,
                                         int exp_bits) {
   int index = blockIdx.x * blockDim.x + threadIdx.x;
   if (index < size) {
-    //unsigned int old_num = FLOAT_TO_BITS(&a[index]);
-    //unsigned int rand_prob = (unsigned int) r[index];
-    //unsigned int quantize = round_bitwise_stochastic(old_num, rand_prob, man_bits);
-    //quantize = clip_exponent(exp_bits, man_bits, old_num, quantize);
-    //float quantize_float = BITS_TO_FLOAT(&quantize);
-    //o[index] = quantize_float;
     unsigned int rand_prob = (unsigned int) r[index];
     unsigned int target,quantize_bits;
     target = FLOAT_TO_BITS(&a[index]);
@@ -50,12 +44,6 @@ __global__ void float_kernel_nearest(float* __restrict__ a,
                                      int exp_bits) {
   int index = blockIdx.x * blockDim.x + threadIdx.x;
   if (index < size) {
-
-    //unsigned int old_num = FLOAT_TO_BITS(&a[index]);
-    //unsigned int quantize = round_bitwise_nearest(old_num, man_bits);
-    //quantize = clip_exponent(exp_bits, man_bits, old_num, quantize);
-    //float quantize_float = BITS_TO_FLOAT(&quantize);
-    //o[index] = quantize_float;
     unsigned int target,quantize_bits;
     target = FLOAT_TO_BITS(&a[index]);
     float quantized;

test/test_clamp.py

@@ -15,8 +15,7 @@ class TestStochastic(unittest.TestCase):
 
     def test_fixed(self):
         """test fixed point clamping"""
-        #for d in ["cpu", "cuda"]:
-        for d in ["cpu"]:
+        for d in ["cpu", "cuda"]:
             for r in ["stochastic", "nearest"]:
                 wl = 5
                 fl = 4
@@ -37,8 +36,7 @@ def test_float(self):
         formats = [(2,2),(2,3),(3,2)]
 
         for exp, man in formats:
-            #for d in ["cpu", "cuda"]:
-            for d in ["cpu"]:
+            for d in ["cpu", "cuda"]:
                 for r in ["stochastic", "nearest"]:
                     a_max = 2 ** (2 ** (exp - 1)) * (1 - 2 ** (-man - 1))
                     a_min = 2 ** (-(2 ** (exp - 1)) + 1)