polish log_softmax precision

python/jittor/__init__.py

@@ -9,7 +9,7 @@
 # file 'LICENSE.txt', which is part of this source code package.
 # ***************************************************************
 
-__version__ = '1.3.4.5'
+__version__ = '1.3.4.6'
 from jittor_utils import lock
 with lock.lock_scope():
     ori_int = int

python/jittor/nn.py

@@ -496,14 +496,12 @@ def softmax(x, dim=None, log=False):
     import jittor.other.code_softmax as code_softmax
     if code_softmax.can_softmax_v1(x, dim):
         return code_softmax.softmax_v1(x, log)
-    if dim is None:
+    if dim is None: dim = ()
-        x = (x - x.max()).exp()
+    if log:
-        ret = x / x.sum()
+        a = x-x.max(dim, keepdims=True)
-    else:
+        return a - a.exp().sum(dim, keepdims=True).log()
-        x = (x-x.max(dim, keepdims=True)).exp()
+    x = (x-x.max(dim, keepdims=True)).exp()
-        ret = x / x.sum(dim, keepdims=True)
+    return x / x.sum(dim, keepdims=True)
-    if log: return ret.log()
-    return ret
 jt.Var.softmax = softmax
 
 def log_softmax(x,dim=None):

python/jittor/other/code_softmax.py

@@ -39,6 +39,7 @@ def softmax_v1(a, log=False):
 ''', cuda_src=f'''
 __global__ void kernel(in0_type* x, out0_type* y, int len) {{
     typedef cub::BlockReduce<float, {tnum}> BlockReduce;
+    constexpr int need_log = {int(log)};
     __shared__ typename BlockReduce::TempStorage temp_storage;
 
     int id = blockIdx.x * len;
@@ -62,8 +63,13 @@ __global__ void kernel(in0_type* x, out0_type* y, int len) {{
     {for_loop}
         #pragma unroll
         for (int j=0; j<{ILP}; j++) {{
-            v[i][j] = expf(float(v[i][j]) - vmax);
+            if (need_log) {{
-            v1 += float(v[i][j]);
+                v[i][j] = float(v[i][j]) - vmax;
+                v1 += expf(float(v[i][j]));
+            }} else {{
+                v[i][j] = expf(float(v[i][j]) - vmax);
+                v1 += float(v[i][j]);
+            }}
         }}
 
     tmp = BlockReduce(temp_storage).Sum(v1);
@@ -74,11 +80,12 @@ __global__ void kernel(in0_type* x, out0_type* y, int len) {{
 
     {for_loop}
         #pragma unroll
-        for (int j=0; j<{ILP}; j++)
+        for (int j=0; j<{ILP}; j++) {{
-            v[i][j] = {
+            if (need_log)
-                "@expand_op(log,@in0_type,float(v[i][j])/vsum)" if log
+                v[i][j] = v[i][j] - @expand_op(log,@in0_type,vsum);
-                else "float(v[i][j])/vsum"
+            else
-            };
+                v[i][j] = float(v[i][j])/vsum;
+        }}
     {for_loop}
         vload<sizeof(in0_type)*{ILP}>(&y[id+(i*{tnum}+threadIdx.x)*{ILP}], v[i]);
 }}

python/jittor/test/test_misc_op.py

@@ -265,6 +265,19 @@ class TestOther(unittest.TestCase):
         z2 = np.arctan2(y.data, x.data)
         np.testing.assert_allclose(z.data, z2)
 
+    def test_softmax_precision(self):
+        # jt.flags.use_cuda = 1
+        a = -jt.array([1.0,2.0,1e5])
+        b = a.log_softmax(0)
+        assert b.isfinite().all().item()
+        print("test_softmax_precision cpu ok")
+        if not jt.has_cuda: return
+        jt.flags.use_cuda = 1
+        a = -jt.array([1.0,2.0,1e5])
+        b = a.log_softmax(0)
+        assert b.isfinite().all().item()
+        print("test_softmax_precision gpu ok")
+
     def test_code_softmax(self):
         if not jt.has_cuda: return