fix arange shape issue inconsistency across cpu and cuda (#18462)

Summary:
Fixes pytorch/pytorch#18363
Pull Request resolved: pytorch/pytorch#18462

Differential Revision: D14620263

Pulled By: soumith

fbshipit-source-id: 223524cdda2f5d55c2ca8d4cdcf6f7a05a6c15eb

Author: soumith

aten/src/ATen/native/RangeFactories.cpp

@@ -116,14 +116,29 @@ Tensor& arange_cpu_out(Tensor& result, Scalar start, Scalar end, Scalar step) {
     auto xend = end.to<accscalar_t>();
     auto xstep = step.to<accscalar_t>();
 
+    // we use double precision for (start - end) / step
+    // to compute size_d for consistency across devices.
+    // The problem with using accscalar_t is that accscalar_t might be float32 on gpu for a float32 scalar_t,
+    // but double on cpu for the same,
+    // and the effective output size starts differing on CPU vs GPU because of precision issues, which
+    // we dont want.
+    // the corner-case we do want to take into account is int64_t, which has higher precision than double
+    double size_d;
+    if (std::is_same<scalar_t, int64_t>::value) {
+      size_d = std::ceil(static_cast<double>(end.to<accscalar_t>() - start.to<accscalar_t>())
+			 / step.to<accscalar_t>());
+    } else {
+      size_d = std::ceil(static_cast<double>(end.to<double>() - start.to<double>())
+			 / step.to<double>());
+    }
+
     AT_CHECK(xstep > 0 || xstep < 0, "step must be nonzero");
     AT_CHECK(std::isfinite(static_cast<double>(xstart)) &&
              std::isfinite(static_cast<double>(xend)),
              "unsupported range: ", xstart, " -> ", xend);
     AT_CHECK(((xstep > 0) && (xend >= xstart)) || ((xstep < 0) && (xend <= xstart)),
              "upper bound and larger bound inconsistent with step sign");
 
-    double size_d = std::ceil(static_cast<double>(xend - xstart) / xstep);
     AT_CHECK(size_d >= 0 && size_d <= static_cast<double>(std::numeric_limits<int64_t>::max()),
              "invalid size, possible overflow?");
     int64_t size = static_cast<int64_t>(size_d);

aten/src/ATen/native/cuda/RangeFactories.cu

@@ -136,14 +136,29 @@ Tensor& arange_cuda_out(Tensor& result, Scalar start, Scalar end, Scalar step) {
     auto xend = end.to<accscalar_t>();
     auto xstep = step.to<accscalar_t>();
 
+    // we use double precision for (start - end) / step
+    // to compute size_d for consistency across devices.
+    // The problem with using accscalar_t is that accscalar_t might be float32 on gpu for a float32 scalar_t,
+    // but double on cpu for the same,
+    // and the effective output size starts differing on CPU vs GPU because of precision issues, which
+    // we dont want.
+    // the corner-case we do want to take into account is int64_t, which has higher precision than double
+    double size_d;
+    if (std::is_same<scalar_t, int64_t>::value) {
+      size_d = std::ceil(static_cast<double>(end.to<accscalar_t>() - start.to<accscalar_t>())
+			 / step.to<accscalar_t>());
+    } else {
+      size_d = std::ceil(static_cast<double>(end.to<double>() - start.to<double>())
+			 / step.to<double>());
+    }
+
     AT_CHECK(xstep > 0 || xstep < 0, "step must be nonzero");
     AT_CHECK(std::isfinite(static_cast<double>(xstart)) &&
              std::isfinite(static_cast<double>(xend)),
              "unsupported range: ", xstart, " -> ", xend);
     AT_CHECK(((xstep > 0) && (xend >= xstart)) || ((xstep < 0) && (xend <= xstart)),
              "upper bound and larger bound inconsistent with step sign");
 
-    double size_d = std::ceil(static_cast<double>(xend - xstart) / xstep);
     AT_CHECK(size_d >= 0 && size_d <= static_cast<double>(std::numeric_limits<int64_t>::max()),
              "invalid size, possible overflow?");
     int64_t size = static_cast<int64_t>(size_d);