Clean up non-exec_aten references to tensor types (#5254)

Summary:
Pull Request resolved: #5254

Most code should use the exec_aten:: types: only aten-mode-aware code should refer directly to e.g. torch::executor::Tensor.

ArrayRef is a little different because it's also a core type, but in that case it should be called executorch::runtime::ArrayRef.

Reviewed By: mergennachin

Differential Revision: D62476845

fbshipit-source-id: 869fb15ce342873697823425271955ca7ed4c14d

Author: dbort

backends/qualcomm/runtime/QnnExecuTorch.h

@@ -44,7 +44,7 @@ struct CustomMemTensorInfo {
   size_t tensor_bytes;
   uint32_t* shape;
   uint32_t rank;
-  torch::executor::ScalarType dtype;
+  exec_aten::ScalarType dtype;
 };
 
 /// Allocate specific tensors (usually graph inputs and outputs) on shared

backends/qualcomm/runtime/SharedBuffer.cpp

@@ -25,7 +25,7 @@ std::size_t std::hash<CustomMemTensorInfo>::operator()(
     hash_val ^= info.shape[i];
   }
   hash_val ^= std::hash<uint32_t>()(info.rank);
-  hash_val ^= std::hash<torch::executor::ScalarType>()(info.dtype);
+  hash_val ^= std::hash<exec_aten::ScalarType>()(info.dtype);
   return hash_val;
 }
 

backends/vulkan/test/utils/test_utils.cpp

@@ -8,7 +8,7 @@
 
 #include <executorch/backends/vulkan/test/utils/test_utils.h>
 
-#include <executorch/runtime/core/portable_type/half.h>
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
 
 #include <executorch/backends/vulkan/runtime/graph/ops/impl/utils/TensorUtils.h>
 
@@ -322,7 +322,7 @@ void record_reference_matmul(
   _(uint8_t, Byte)                \
   _(int8_t, Char)                 \
   _(int32_t, Int)                 \
-  _(torch::executor::Half, Half)  \
+  _(exec_aten::Half, Half)        \
   _(float, Float)                 \
   _(int8_t, QInt8)
 

backends/vulkan/test/vulkan_compute_api_test.cpp

@@ -11,7 +11,7 @@
 #include <utility>
 #include <vector>
 
-#include <executorch/runtime/core/portable_type/half.h>
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
 
 #include <executorch/backends/vulkan/runtime/api/api.h>
 
@@ -485,7 +485,7 @@ TEST_F(VulkanComputeAPITest, test_buffer_float16) {
   if (!context()->adapter_ptr()->has_full_float16_buffers_support()) {
     GTEST_SKIP();
   }
-  test_storage_buffer_type<torch::executor::Half, vkapi::kHalf>(16);
+  test_storage_buffer_type<exec_aten::Half, vkapi::kHalf>(16);
 }
 
 TEST_F(VulkanComputeAPITest, test_buffer_int8) {
@@ -567,7 +567,7 @@ TEST_F(VulkanComputeAPITest, buffer_tensor_sanity_check) {
             run_buffer_tensor_sanity_check<float>(a);
             break;
           case vkapi::kHalf:
-            run_buffer_tensor_sanity_check<torch::executor::Half>(a);
+            run_buffer_tensor_sanity_check<exec_aten::Half>(a);
             break;
           case vkapi::kChar:
             run_buffer_tensor_sanity_check<int8_t>(a);
@@ -2395,7 +2395,7 @@ TEST(VulkanToFromGPUShaderTest, round_trip_tests) {
 
   for (auto& sizes : to_test) {
     RUN_TESTS(float, vkapi::kFloat)
-    RUN_TESTS(torch::executor::Half, vkapi::kHalf)
+    RUN_TESTS(exec_aten::Half, vkapi::kHalf)
   }
 
   for (auto& sizes : to_test_int8) {

examples/models/flamingo/cross_attention/cross_attention_mask_test.cpp

@@ -11,9 +11,9 @@
 #include <gtest/gtest.h>
 
 using namespace ::testing;
-using torch::executor::ScalarType;
-using torch::executor::Tensor;
-using torch::executor::TensorImpl;
+using exec_aten::ScalarType;
+using exec_aten::Tensor;
+using exec_aten::TensorImpl;
 
 TEST(CrossAttentxnMaskTest, TestCrossAttentionMask) {
   std::vector<int> tokens = {

examples/qualcomm/oss_scripts/llama2/runner/runner.cpp

@@ -135,7 +135,7 @@ int32_t Runner::logitsToToken(const exec_aten::Tensor& logits_tensor) {
 
 // Given an input token. Set up the inputs for the model and execute a single
 // step. Returning the logits tensor.
-Result<torch::executor::Tensor> Runner::run_model_step(
+Result<exec_aten::Tensor> Runner::run_model_step(
     int64_t input_token,
     TensorPtr& token,
     TensorPtr& start_pos,

examples/qualcomm/oss_scripts/llama2/runner/runner.h

@@ -120,7 +120,7 @@ class IoMemMgr {
     std::vector<uint32_t> shape;
     uint32_t rank;
     size_t element_size;
-    torch::executor::ScalarType dtype;
+    exec_aten::ScalarType dtype;
   };
 
   struct IoInfo {

extension/android/jni/jni_layer_constants.h

@@ -8,7 +8,7 @@
 
 #include <unordered_map>
 
-#include <executorch/runtime/core/portable_type/scalar_type.h>
+#include <executorch/runtime/core/exec_aten/exec_aten.h>
 
 namespace executorch::extension {
 
@@ -37,7 +37,7 @@ constexpr static int kTensorDTypeBits4x2 = 20;
 constexpr static int kTensorDTypeBits8 = 21;
 constexpr static int kTensorDTypeBits16 = 22;
 
-using torch::executor::ScalarType;
+using exec_aten::ScalarType;
 
 const std::unordered_map<ScalarType, int> scalar_type_to_java_dtype = {
     {ScalarType::Byte, kTensorDTypeUInt8},

extension/android/src/main/java/org/pytorch/executorch/DType.java

@@ -17,51 +17,51 @@ public enum DType {
   // NOTE: "jniCode" must be kept in sync with scalar_type.h.
   // NOTE: Never serialize "jniCode", because it can change between releases.
 
-  /** Code for dtype torch::executor::Byte */
+  /** Code for dtype ScalarType::Byte */
   UINT8(0),
-  /** Code for dtype torch::executor::Char */
+  /** Code for dtype ScalarType::Char */
   INT8(1),
-  /** Code for dtype torch::executor::Short */
+  /** Code for dtype ScalarType::Short */
   INT16(2),
-  /** Code for dtype torch::executor::Int */
+  /** Code for dtype ScalarType::Int */
   INT32(3),
-  /** Code for dtype torch::executor::Long */
+  /** Code for dtype ScalarType::Long */
   INT64(4),
-  /** Code for dtype torch::executor::Half */
+  /** Code for dtype ScalarType::Half */
   HALF(5),
-  /** Code for dtype torch::executor::Float */
+  /** Code for dtype ScalarType::Float */
   FLOAT(6),
-  /** Code for dtype torch::executor::Double */
+  /** Code for dtype ScalarType::Double */
   DOUBLE(7),
-  /** Code for dtype torch::executor::ComplexHalf */
+  /** Code for dtype ScalarType::ComplexHalf */
   COMPLEX_HALF(8),
-  /** Code for dtype torch::executor::ComplexFloat */
+  /** Code for dtype ScalarType::ComplexFloat */
   COMPLEX_FLOAT(9),
-  /** Code for dtype torch::executor::ComplexDouble */
+  /** Code for dtype ScalarType::ComplexDouble */
   COMPLEX_DOUBLE(10),
-  /** Code for dtype torch::executor::Bool */
+  /** Code for dtype ScalarType::Bool */
   BOOL(11),
-  /** Code for dtype torch::executor::QInt8 */
+  /** Code for dtype ScalarType::QInt8 */
   QINT8(12),
-  /** Code for dtype torch::executor::QUInt8 */
+  /** Code for dtype ScalarType::QUInt8 */
   QUINT8(13),
-  /** Code for dtype torch::executor::QInt32 */
+  /** Code for dtype ScalarType::QInt32 */
   QINT32(14),
-  /** Code for dtype torch::executor::BFloat16 */
+  /** Code for dtype ScalarType::BFloat16 */
   BFLOAT16(15),
-  /** Code for dtype torch::executor::QUInt4x2 */
+  /** Code for dtype ScalarType::QUInt4x2 */
   QINT4X2(16),
-  /** Code for dtype torch::executor::QUInt2x4 */
+  /** Code for dtype ScalarType::QUInt2x4 */
   QINT2X4(17),
-  /** Code for dtype torch::executor::Bits1x8 */
+  /** Code for dtype ScalarType::Bits1x8 */
   BITS1X8(18),
-  /** Code for dtype torch::executor::Bits2x4 */
+  /** Code for dtype ScalarType::Bits2x4 */
   BITS2X4(19),
-  /** Code for dtype torch::executor::Bits4x2 */
+  /** Code for dtype ScalarType::Bits4x2 */
   BITS4X2(20),
-  /** Code for dtype torch::executor::Bits8 */
+  /** Code for dtype ScalarType::Bits8 */
   BITS8(21),
-  /** Code for dtype torch::executor::Bits16 */
+  /** Code for dtype ScalarType::Bits16 */
   BITS16(22),
   ;
 

kernels/optimized/blas/CPUBlas.cpp

@@ -24,7 +24,8 @@ extern "C" void sgemm_(char *transa, char *transb, int *m, int *n, int *k, float
 namespace executorch {
 namespace cpublas {
 
-// using Half = exec_aten::Half;
+using exec_aten::BFloat16;
+using exec_aten::Half;
 
 #ifdef ET_BUILD_WITH_BLAS
 #ifdef ET_BUILD_FOR_APPLE

kernels/optimized/blas/CPUBlas.h

@@ -17,9 +17,6 @@
 namespace executorch {
 namespace cpublas {
 
-using BFloat16 = torch::executor::BFloat16;
-using Half = torch::executor::Half;
-
 enum class TransposeType {
   NoTranspose,
   Transpose,
@@ -100,20 +97,20 @@ void gemm(
 void gemm(
     TransposeType transa, TransposeType transb,
     int64_t m, int64_t n, int64_t k,
-    const Half alpha,
-    const Half *a, int64_t lda,
-    const Half *b, int64_t ldb,
-    const Half beta,
-    Half *c, int64_t ldc);
+    const exec_aten::Half alpha,
+    const exec_aten::Half *a, int64_t lda,
+    const exec_aten::Half *b, int64_t ldb,
+    const exec_aten::Half beta,
+    exec_aten::Half *c, int64_t ldc);
 
 void gemm(
     TransposeType transa, TransposeType transb,
     int64_t m, int64_t n, int64_t k,
-    const BFloat16 alpha,
-    const BFloat16 *a, int64_t lda,
-    const BFloat16 *b, int64_t ldb,
-    const BFloat16 beta,
-    BFloat16 *c, int64_t ldc);
+    const exec_aten::BFloat16 alpha,
+    const exec_aten::BFloat16 *a, int64_t lda,
+    const exec_aten::BFloat16 *b, int64_t ldb,
+    const exec_aten::BFloat16 beta,
+    exec_aten::BFloat16 *c, int64_t ldc);
 // clang-format on
 
 // clang-format off

kernels/optimized/cpu/op_exp.cpp

@@ -27,8 +27,8 @@ template <
     typename CTYPE_OUT,
     typename std::enable_if<
         std::is_same<CTYPE_IN, CTYPE_OUT>::value &&
-            !std::is_same<CTYPE_IN, torch::executor::Half>::value &&
-            !std::is_same<CTYPE_OUT, torch::executor::Half>::value,
+            !std::is_same<CTYPE_IN, exec_aten::Half>::value &&
+            !std::is_same<CTYPE_OUT, exec_aten::Half>::value,
         int>::type = 0>
 void exp_data(
     const CTYPE_IN* in_data,
@@ -47,8 +47,8 @@ template <
     typename CTYPE_OUT,
     typename std::enable_if<
         !std::is_same<CTYPE_IN, CTYPE_OUT>::value ||
-            std::is_same<CTYPE_IN, torch::executor::Half>::value ||
-            std::is_same<CTYPE_OUT, torch::executor::Half>::value,
+            std::is_same<CTYPE_IN, exec_aten::Half>::value ||
+            std::is_same<CTYPE_OUT, exec_aten::Half>::value,
         int>::type = 0>
 void exp_data(
     const CTYPE_IN* in_data,

kernels/portable/cpu/op_empty.cpp

@@ -26,7 +26,7 @@ using exec_aten::Tensor;
 Tensor& empty_out(
     KernelRuntimeContext& context,
     IntArrayRef size,
-    torch::executor::optional<torch::executor::MemoryFormat> memory_format,
+    exec_aten::optional<exec_aten::MemoryFormat> memory_format,
     Tensor& out) {
   (void)context;
 

kernels/portable/cpu/test/scalar_utils_test.cpp

@@ -16,7 +16,7 @@ struct promote_type_with_scalar_type_is_valid
           (std::is_same<T2, torch::executor::internal::B1>::value ||
            std::is_same<T2, torch::executor::internal::I8>::value ||
            std::is_same<T2, torch::executor::internal::F8>::value) &&
-              !std::is_same<T1, torch::executor::BFloat16>::value &&
+              !std::is_same<T1, exec_aten::BFloat16>::value &&
               !torch::executor::is_qint_type<T1>::value &&
               !torch::executor::is_bits_type<T1>::value> {};
 

kernels/portable/cpu/util/math_util.h

@@ -96,9 +96,8 @@ INT_T max_override(INT_T a, INT_T b) {
 
 template <
     typename T,
-    typename std::enable_if<
-        std::is_same<T, torch::executor::Half>::value,
-        bool>::type = true>
+    typename std::enable_if<std::is_same<T, exec_aten::Half>::value, bool>::
+        type = true>
 T min_override(T a, T b) {
   const auto float_a = static_cast<float>(a);
   if (std::isnan(float_a)) {
@@ -117,9 +116,8 @@ T min_override(T a, T b) {
 
 template <
     typename T,
-    typename std::enable_if<
-        std::is_same<T, torch::executor::Half>::value,
-        bool>::type = true>
+    typename std::enable_if<std::is_same<T, exec_aten::Half>::value, bool>::
+        type = true>
 T max_override(T a, T b) {
   const auto float_a = static_cast<float>(a);
   if (std::isnan(float_a)) {

kernels/portable/cpu/util/test/broadcast_test.cpp

@@ -20,8 +20,8 @@
 using namespace ::testing;
 using exec_aten::ScalarType;
 using exec_aten::Tensor;
-using torch::executor::ArrayRef;
-using torch::executor::testing::TensorFactory;
+using executorch::runtime::ArrayRef;
+using executorch::runtime::testing::TensorFactory;
 
 TEST(BroadcastUtilTest, BroadcastTensor) {
   TensorFactory<ScalarType::Int> tf;
@@ -112,17 +112,17 @@ TEST(BroadcastUtilTest, GetBroadcastTargetSize) {
   Tensor a = tf.zeros({2, 1});
   Tensor b = tf.zeros({5, 1, 2});
 
-  get_broadcast_target_size(
+  executorch::runtime::Error err = get_broadcast_target_size(
       a,
       b,
       expected_output_size,
       torch::executor::kTensorDimensionLimit,
       &expected_output_dim);
+  EXPECT_EQ(err, torch::executor::Error::Ok);
 
   EXPECT_TRUE(
-      torch::executor::ArrayRef<Tensor::SizesType>(
-          expected_output_size, expected_output_dim)
-          .equals(torch::executor::ArrayRef<Tensor::SizesType>({5, 2, 2})));
+      ArrayRef<Tensor::SizesType>(expected_output_size, expected_output_dim)
+          .equals(ArrayRef<Tensor::SizesType>({5, 2, 2})));
 }
 
 size_t linearize_indexes(size_t* indexes, size_t indexes_len, const Tensor& t) {

kernels/portable/test/op_mul_test.cpp

@@ -36,19 +36,18 @@ class OpMulOutKernelTest : public OperatorTest {
 TEST_F(OpMulOutKernelTest, UnhandledDtypeDies) {
   // mul_out() doesn't handle QInt8.
   // TensorFactory cannot be used with ScalarType::QInt8 since
-  // torch::executor::qint8 does not have a default constructor. It must be
+  // exec_aten::qint8 does not have a default constructor. It must be
   // initialized with an explicit value. So, we need to manually create the
   // underlying data without default construction and then the tensors from that
   // data via TensorImpl.
 
   std::vector<SizesType> sizes = {2, 2};
 
-  std::vector<torch::executor::qint8> a_data{};
-  std::generate_n(std::back_inserter(a_data), 4, []() {
-    return torch::executor::qint8{0};
-  });
-  std::vector<torch::executor::qint8> b_data(a_data);
-  std::vector<torch::executor::qint8> out_data(a_data);
+  std::vector<exec_aten::qint8> a_data{};
+  std::generate_n(
+      std::back_inserter(a_data), 4, []() { return exec_aten::qint8{0}; });
+  std::vector<exec_aten::qint8> b_data(a_data);
+  std::vector<exec_aten::qint8> out_data(a_data);
 
   auto a_impl = torch::executor::TensorImpl(
       ScalarType::QInt8, 2, sizes.data(), a_data.data());

kernels/prim_ops/register_prim_ops.cpp

@@ -12,7 +12,6 @@
 #include <executorch/runtime/kernel/kernel_includes.h>
 #include <executorch/runtime/kernel/operator_registry.h>
 
-using KernelArrayRef = ::torch::executor::ArrayRef<::torch::executor::Kernel>;
 using torch::executor::function::et_copy_index;
 
 namespace torch {
@@ -294,13 +293,14 @@ static Kernel prim_ops[] = {
 
 };
 
-static KernelArrayRef kernel_array_ref(
+executorch::runtime::Span<const executorch::runtime::Kernel> kernel_span(
     prim_ops,
     prim_ops + sizeof(prim_ops) / sizeof(Kernel));
 
 // Return value not used. Keep the static variable assignment to register
 // operators in static initialization time.
-static auto success_with_kernel_reg = register_kernels(kernel_array_ref);
+auto success_with_kernel_reg =
+    executorch::runtime::register_kernels(kernel_span);
 
 } // namespace
 } // namespace function