Refactor and Fix Some Prombles on Autocast

1. Fix the wrong test cases
2. Refactor tests

ghstack-source-id: 9f3685af5dd509cbcd3c2dfd057de4cad9395f47
Pull Request resolved: #125118

aten/src/ATen/autocast_mode.cpp

@@ -321,19 +321,22 @@ TORCH_LIBRARY_IMPL(aten, Autocast, m) {
   KERNEL_CUDA(__VA_ARGS__, lower_precision_fp)
 
   AT_FORALL_LOWER_PRECISION_FP(_KERNEL_CUDA_LOW_PRECISION_FP)
+#undef _KERNEL_CUDA_LOW_PRECISION_FP
   KERNEL_CUDA(cudnn_convolution, lower_precision_fp)
   KERNEL_CUDA(cudnn_convolution_transpose, lower_precision_fp)
 
   // fp32
 #define _KERNEL_CUDA_FP32(...) KERNEL_CUDA(__VA_ARGS__, fp32)
 
   AT_FORALL_FP32(_KERNEL_CUDA_FP32)
+#undef _KERNEL_CUDA_FP32
 
   // fp32_set_opt_dtype
 #define _KERNEL_CUDA_FP32_SET_OPT_DTYPE(...) \
   KERNEL_CUDA(__VA_ARGS__, fp32_set_opt_dtype)
 
   AT_FORALL_FP32_SET_OPT_DTYPE(_KERNEL_CUDA_FP32_SET_OPT_DTYPE)
+#undef _KERNEL_CUDA_FP32_SET_OPT_DTYPE
   // commenting these out because they accept an explicit (not-optional) dtype, and we shouldn't try to flip that even
   // when autocasting.
   // KERNEL_CUDA(norm, ScalarOpt_dtype, fp32_set_opt_dtype)
@@ -350,9 +353,9 @@ TORCH_LIBRARY_IMPL(aten, Autocast, m) {
 #define _KERNEL_CUDA_PROMOTE(...) KERNEL_CUDA(__VA_ARGS__, promote)
 
   AT_FORALL_PROMOTE(_KERNEL_CUDA_PROMOTE)
+#undef _KERNEL_CUDA_PROMOTE
 
-  m.impl(TORCH_SELECTIVE_NAME("aten::binary_cross_entropy"),
-         TORCH_FN((&at::autocast::binary_cross_entropy_banned)));
+  KERNEL_FN("binary_cross_entropy", &at::autocast::binary_cross_entropy_banned)
 }
 
 TORCH_LIBRARY_IMPL(_, AutocastCPU, m) {
@@ -507,17 +510,20 @@ TORCH_LIBRARY_IMPL(aten, AutocastXPU, m) {
   KERNEL_XPU(__VA_ARGS__, lower_precision_fp)
 
   AT_FORALL_LOWER_PRECISION_FP(_KERNEL_XPU_LOW_PRECISION_FP)
+#undef _KERNEL_XPU_LOW_PRECISION_FP
 
   // fp32
 #define _KERNEL_XPU_FP32(...) KERNEL_XPU(__VA_ARGS__, fp32)
 
   AT_FORALL_FP32(_KERNEL_XPU_FP32)
+#undef _KERNEL_XPU_FP32
 
   // fp32_set_opt_dtype
 #define _KERNEL_XPU_FP32_SET_OPT_DTYPE(...) \
   KERNEL_XPU(__VA_ARGS__, fp32_set_opt_dtype)
 
   AT_FORALL_FP32_SET_OPT_DTYPE(_KERNEL_XPU_FP32_SET_OPT_DTYPE)
+#undef _KERNEL_XPU_FP32_SET_OPT_DTYPE
 
   // fp32_append_dtype
   // The fp32_append_dtype wrapper overrides implicit promotion behavior.
@@ -529,9 +535,9 @@ TORCH_LIBRARY_IMPL(aten, AutocastXPU, m) {
 #define _KERNEL_XPU_PROMOTE(...) KERNEL_XPU(__VA_ARGS__, promote)
 
   AT_FORALL_PROMOTE(_KERNEL_XPU_PROMOTE)
+#undef _KERNEL_XPU_PROMOTE
 
-  m.impl(TORCH_SELECTIVE_NAME("aten::binary_cross_entropy"),
-         TORCH_FN((&at::autocast::binary_cross_entropy_banned)));
+  KERNEL_FN("binary_cross_entropy", &at::autocast::binary_cross_entropy_banned)
 }
 
 } // namespace

aten/src/ATen/autocast_mode.h

@@ -623,6 +623,9 @@ copy pasted in from VariableTypeEverything.cpp with appropriate substitutions.
 #define _KERNEL_OVERLOAD_NARG(...) \
   C10_EXPAND_MSVC_WORKAROUND(_KERNEL_OVERLOAD_NARG_IMPL(__VA_ARGS__, 2, 1))
 
+#define KERNEL_FN(OP, Function) \
+  m.impl(TORCH_SELECTIVE_NAME("aten::" OP), TORCH_FN(Function));
+
 // Common cases where registration signature matches redispatch signature
 // (that's why SIGNATURE is repeated in the WrapFunction instantiation)
 #define KERNEL1(DISPATCHKEY, OP, POLICY)      \

c10/test/core/DispatchKeySet_test.cpp

@@ -9,6 +9,20 @@
 
 using namespace c10;
 
+static bool isRealDispatchKey(DispatchKey k) {
+  if (k == DispatchKey::EndOfFunctionalityKeys ||
+      k == DispatchKey::StartOfDenseBackends ||
+      k == DispatchKey::StartOfQuantizedBackends ||
+      k == DispatchKey::StartOfSparseBackends ||
+      k == DispatchKey::StartOfSparseCsrBackends ||
+      k == DispatchKey::StartOfNestedTensorBackends ||
+      k == DispatchKey::StartOfAutogradFunctionalityBackends) {
+    return false;
+  }
+
+  return true;
+}
+
 // This test exists not to be comprehensive, but to more clearly show
 // what the semantics of DispatchKeySet are.
 TEST(DispatchKeySet, ShowSemantics) {
@@ -179,10 +193,7 @@ TEST(DispatchKeySet, SingletonPerBackendFunctionalityKeys) {
        i++) {
     auto tid = static_cast<DispatchKey>(i);
     // Skip these because they aren't real keys.
-    if (tid == DispatchKey::StartOfDenseBackends ||
-        tid == DispatchKey::StartOfSparseBackends ||
-        tid == DispatchKey::StartOfQuantizedBackends ||
-        tid == DispatchKey::StartOfAutogradFunctionalityBackends) {
+    if (isRealDispatchKey(tid)) {
       continue;
     }
     DispatchKeySet sing(tid);
@@ -221,20 +232,9 @@ TEST(DispatchKeySet, DoubletonPerBackend) {
       auto tid2 = static_cast<DispatchKey>(j);
 
       // Skip these because they aren't real keys.
-      if (tid1 == DispatchKey::StartOfDenseBackends ||
-          tid1 == DispatchKey::StartOfSparseBackends ||
-          tid1 == DispatchKey::StartOfSparseCsrBackends ||
-          tid1 == DispatchKey::StartOfQuantizedBackends ||
-          tid1 == DispatchKey::StartOfNestedTensorBackends ||
-          tid1 == DispatchKey::StartOfAutogradFunctionalityBackends)
-        continue;
-      if (tid2 == DispatchKey::StartOfDenseBackends ||
-          tid2 == DispatchKey::StartOfSparseBackends ||
-          tid2 == DispatchKey::StartOfSparseCsrBackends ||
-          tid2 == DispatchKey::StartOfQuantizedBackends ||
-          tid2 == DispatchKey::StartOfNestedTensorBackends ||
-          tid2 == DispatchKey::StartOfAutogradFunctionalityBackends)
+      if (!isRealDispatchKey(tid1) || !isRealDispatchKey(tid2)) {
         continue;
+      }
 
       auto backend1 = toBackendComponent(tid1);
       auto backend2 = toBackendComponent(tid2);
@@ -421,14 +421,9 @@ TEST(DispatchKeySet, TestFunctionalityDispatchKeyToString) {
     auto k = static_cast<DispatchKey>(i);
     // These synthetic keys never actually get used and don't need
     // to be printed
-    if (k == DispatchKey::EndOfFunctionalityKeys ||
-        k == DispatchKey::StartOfDenseBackends ||
-        k == DispatchKey::StartOfQuantizedBackends ||
-        k == DispatchKey::StartOfSparseBackends ||
-        k == DispatchKey::StartOfSparseCsrBackends ||
-        k == DispatchKey::StartOfNestedTensorBackends ||
-        k == DispatchKey::StartOfAutogradFunctionalityBackends)
+    if (!isRealDispatchKey(k)) {
       continue;
+    }
     auto res = std::string(toString(k));
     ASSERT_TRUE(res.find("Unknown") == std::string::npos)
         << i << " (before is " << toString(static_cast<DispatchKey>(i - 1))

test/test_ops.py

@@ -2350,10 +2350,10 @@ def test_refs_are_in_decomp_table(self, op):
     "narrow",  # Fails only for one overload with DataDependentOutputException (hence skip).
 )
 
-fake_autocast_device_skips = defaultdict(dict)
-
 # TODO: investigate/fix
+fake_autocast_device_skips = defaultdict(set)
 fake_autocast_device_skips["cpu"] = {"linalg.pinv"}
+fake_autocast_device_skips["cuda"] = {"linalg.pinv", "pinverse"}
 
 
 dynamic_output_op_tests = (
@@ -2594,6 +2594,8 @@ def test_fake(self, device, dtype, op):
 
     @ops(op_db, dtypes=OpDTypes.any_one)
     def test_fake_autocast(self, device, dtype, op):
+        # remove the index from the device, first
+        device = device.split(":")[0]
         if op.name in fake_autocast_device_skips[device]:
             self.skipTest("Skip failing test")
         context = (