Alert non-deterministic on kernels that use gpuAtomicAdd (#7582)

Signed-off-by: Edward Z. Yang <ezyang@meta.com>

torchvision/csrc/ops/cuda/deform_conv2d_kernel.cu

@@ -426,6 +426,8 @@ void compute_grad_input(
   // Checks if num_kernels or columns numel larger than 2 ** 31
   use_64bits_indexing |= num_kernels > (1 << 31);
 
+  at::globalContext().alertNotDeterministic("compute_grad_input");
+
   if (use_64bits_indexing) {
     AT_DISPATCH_FLOATING_TYPES_AND_HALF(
         columns.scalar_type(), "compute_grad_input", ([&] {

torchvision/csrc/ops/cuda/ps_roi_align_kernel.cu

@@ -412,6 +412,8 @@ at::Tensor ps_roi_align_backward_kernel(
 
   int channels_out = channels / (pooled_height * pooled_width);
 
+  at::globalContext().alertNotDeterministic("ps_roi_align_backward_kernel");
+
   auto grad_ = grad.contiguous(), rois_ = rois.contiguous();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
       grad.scalar_type(), "ps_roi_align_backward_kernel", [&] {

torchvision/csrc/ops/cuda/ps_roi_pool_kernel.cu

@@ -251,6 +251,8 @@ at::Tensor ps_roi_pool_backward_kernel(
 
   int channels_out = channels / (pooled_height * pooled_width);
 
+  at::globalContext().alertNotDeterministic("ps_roi_pool_backward_kernel");
+
   auto grad_ = grad.contiguous(), rois_ = rois.contiguous();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
       grad.scalar_type(), "ps_roi_pool_backward_kernel", [&] {

torchvision/csrc/ops/cuda/roi_align_kernel.cu

@@ -421,6 +421,8 @@ at::Tensor roi_align_backward_kernel(
   int h_stride = grad.stride(2);
   int w_stride = grad.stride(3);
 
+  at::globalContext().alertNotDeterministic("roi_align_backward_kernel");
+
   auto rois_ = rois.contiguous();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
       grad.scalar_type(), "roi_align_backward_kernel", [&] {

torchvision/csrc/ops/cuda/roi_pool_kernel.cu

@@ -232,6 +232,8 @@ at::Tensor roi_pool_backward_kernel(
   int h_stride = grad.stride(2);
   int w_stride = grad.stride(3);
 
+  at::globalContext().alertNotDeterministic("roi_pool_backward_kernel");
+
   auto argmax_ = argmax.contiguous(), rois_ = rois.contiguous();
   AT_DISPATCH_FLOATING_TYPES_AND_HALF(
       grad.scalar_type(), "roi_pool_backward_kernel", [&] {