Mma op integration on ampere (#1440)

Author: shmsong

tools/build_variables.bzl

@@ -32,6 +32,7 @@ libtorch_nvfuser_runtime_sources = [
     "torch/csrc/jit/codegen/cuda/runtime/helpers.cu",
     "torch/csrc/jit/codegen/cuda/runtime/index_utils.cu",
     "torch/csrc/jit/codegen/cuda/runtime/tensorcore.cu",
+    "torch/csrc/jit/codegen/cuda/runtime/memory.cu",
     "torch/csrc/jit/codegen/cuda/runtime/random_numbers.cu",
     "torch/csrc/jit/codegen/cuda/runtime/tensor.cu",
     "torch/csrc/jit/codegen/cuda/runtime/tuple.cu",

torch/csrc/jit/codegen/cuda/codegen.cpp

@@ -477,6 +477,33 @@ class CudaKernelGenerator : private OptOutConstDispatch {
     TORCH_INTERNAL_ASSERT(false, "Unreachable");
   }
 
+  //! Utility for generating vectorized pointer access in ldsm and
+  //!  cpasync.
+  //! TODO: this access pattern as is could be merged with exisiting
+  //!  vectorization handling logic but this path will be updated in
+  //!  follow ups to optimize the generated assembly so keeping them
+  //!  separate path for now.
+  std::string genVectorPointer(Val* val, DataType dtype, int vec_size) {
+    std::stringstream ss;
+
+    ss << "reinterpret_cast<Array<" << dtype << "," << vec_size << ","
+       << vec_size << ">*>(&" << gen(val) << ")";
+
+    return ss.str();
+  }
+
+  void genLdMatrix(const LoadStoreOp* ldst, int vector_word_size) {
+    auto dtype = ldst->in()->getDataType().value();
+    indent() << "Turing::ldMatrix";
+    if (ldst->opType() == LoadStoreOpType::LdMatrixTranspose) {
+      code_ << "T";
+    }
+    code_ << " (";
+    code_ << "*" << genVectorPointer(ldst->out(), dtype, vector_word_size)
+          << ","
+          << "&" << gen(ldst->in()) << ");\n";
+  }
+
   void handle(const UnaryOp* uop) final {
     bool is_vector_op = false;
     size_t vector_word_size = 1;
@@ -918,7 +945,15 @@ class CudaKernelGenerator : private OptOutConstDispatch {
     if (init) {
       ss << "init";
     }
-    ss << toString(options.macro) << toString(options.operand_layout);
+    ss << toString(options.macro);
+
+    if (isVolta(options.macro)) {
+      ss << toString(options.operand_layout);
+    } else if (isTuring(options.macro) || isAmpere(options.macro)) {
+      // mma's in turing and ampere TN only, transpose is handled either
+      //  via ldmatrix for fp16 or explicitly for other types.
+      ss << "TN";
+    }
     // TODO: additional parameter could be removed by swizzling iterdomain
     auto acc_stride = mma->accStride();
     TORCH_INTERNAL_ASSERT(acc_stride > 0);
@@ -1123,6 +1158,49 @@ class CudaKernelGenerator : private OptOutConstDispatch {
     }
   }
 
+  void handle(const LoadStoreOp* ldst) {
+    // TODO:
+    //  Need to gradually merge the code path of this
+    //   with UnaryOp::Set for vectorization.
+    //  There is quite a bit of possible clean up.
+    bool vectorize_op = false;
+    size_t vector_word_size = 1;
+    auto ti = ldst->out()->as<kir::TensorIndex>();
+
+    // Check vectorization and set vector word size
+    for (auto id : ti->view()->domain()->domain()) {
+      if (!isParallelTypeVectorize(id->getParallelType())) {
+        continue;
+      }
+
+      ExpressionEvaluator expr_eval(id->fusion());
+      auto vector_size_optional = expr_eval.evaluate(id->extent());
+
+      TORCH_INTERNAL_ASSERT(
+          vector_size_optional.has_value(),
+          "Could not evaluate constant value bound to vectorized dim.");
+
+      TORCH_INTERNAL_ASSERT(
+          id->getParallelType() != ParallelType::MisalignedVectorize,
+          "LoadStoreOp: no support yet for mis-aligned vectorization");
+      vector_word_size = vector_size_optional.value();
+      vectorize_op = true;
+      break;
+    }
+
+    // Dispatch instruction generation:
+    switch (ldst->opType()) {
+      case LoadStoreOpType::LdMatrix:
+      case LoadStoreOpType::LdMatrixTranspose:
+        TORCH_INTERNAL_ASSERT(
+            vectorize_op, "LdMatrix: Vectorization required: ", ldst);
+        genLdMatrix(ldst, vector_word_size);
+        break;
+      default:
+        TORCH_INTERNAL_ASSERT(false, "LoadStoreOp: Unknown op type");
+    }
+  }
+
   void handle(const WelfordOp* wop) final {
     TORCH_INTERNAL_ASSERT(wop->out()->isA<kir::TensorIndex>());
 
@@ -2033,7 +2111,7 @@ class CudaKernelGenerator : private OptOutConstDispatch {
     }
   }
 
-  void handle(const kir::BlockSync*) final {
+  void handle(const kir::BlockSync* sync) final {
     // Use a custom synchronization method if enabled
     if (std::getenv("PYTORCH_NVFUSER_USE_BLOCK_SYNC_ATOMIC")) {
       indent() << "block_sync::sync();\n";

torch/csrc/jit/codegen/cuda/dispatch.cpp

@@ -110,6 +110,9 @@ void Expr::dispatch(T handler, Expr* expr) {
     case ExprType::WelfordOp:
       ptr(handler)->handle(expr->as<WelfordOp>());
       return;
+    case ExprType::LoadStoreOp:
+      ptr(handler)->handle(expr->as<LoadStoreOp>());
+      return;
     case ExprType::MmaOp:
       ptr(handler)->handle(expr->as<MmaOp>());
       return;
@@ -260,6 +263,9 @@ void Expr::constDispatch(T handler, const Expr* expr) {
     case ExprType::WelfordOp:
       ptr(handler)->handle(expr->as<WelfordOp>());
       return;
+    case ExprType::LoadStoreOp:
+      ptr(handler)->handle(expr->as<LoadStoreOp>());
+      return;
     case ExprType::MmaOp:
       ptr(handler)->handle(expr->as<MmaOp>());
       return;
@@ -418,6 +424,9 @@ void Expr::mutatorDispatch(T mutator, Expr* expr) {
     case ExprType::WelfordOp:
       ptr(mutator)->mutate(expr->as<WelfordOp>());
       return;
+    case ExprType::LoadStoreOp:
+      ptr(mutator)->mutate(expr->as<LoadStoreOp>());
+      return;
     case ExprType::MmaOp:
       ptr(mutator)->mutate(expr->as<MmaOp>());
       return;
@@ -641,6 +650,9 @@ void OptOutConstDispatch::handle(const GroupedReductionOp* stmt) {
 void OptOutConstDispatch::handle(const WelfordOp* stmt) {
   unhandled(stmt);
 }
+void OptOutConstDispatch::handle(const LoadStoreOp* stmt) {
+  unhandled(stmt);
+}
 void OptOutConstDispatch::handle(const MmaOp* stmt) {
   unhandled(stmt);
 }
@@ -761,6 +773,9 @@ void OptOutDispatch::handle(GroupedReductionOp* stmt) {
 void OptOutDispatch::handle(WelfordOp* stmt) {
   unhandled(stmt);
 }
+void OptOutDispatch::handle(LoadStoreOp* stmt) {
+  unhandled(stmt);
+}
 void OptOutDispatch::handle(MmaOp* stmt) {
   unhandled(stmt);
 }

torch/csrc/jit/codegen/cuda/dispatch.h

@@ -74,6 +74,7 @@ class TernaryOp;
 class ReductionOp;
 class GroupedReductionOp;
 class WelfordOp;
+class LoadStoreOp;
 class MmaOp;
 class BroadcastOp;
 class TransposeOp;
@@ -136,6 +137,7 @@ class TORCH_CUDA_CU_API OptOutConstDispatch : public PolymorphicBase {
   virtual void handle(const ReductionOp* stmt);
   virtual void handle(const GroupedReductionOp* stmt);
   virtual void handle(const WelfordOp* stmt);
+  virtual void handle(const LoadStoreOp* stmt);
   virtual void handle(const MmaOp* stmt);
   virtual void handle(const BroadcastOp* stmt);
 
@@ -191,6 +193,7 @@ class TORCH_CUDA_CU_API OptOutDispatch : public PolymorphicBase {
   virtual void handle(ReductionOp* stmt);
   virtual void handle(GroupedReductionOp* stmt);
   virtual void handle(WelfordOp* stmt);
+  virtual void handle(LoadStoreOp* stmt);
   virtual void handle(MmaOp* stmt);
   virtual void handle(BroadcastOp* stmt);
 
@@ -287,6 +290,7 @@ class TORCH_CUDA_CU_API OptOutMutator : public PolymorphicBase {
   virtual void mutate(ReductionOp*);
   virtual void mutate(GroupedReductionOp*);
   virtual void mutate(WelfordOp*);
+  virtual void mutate(LoadStoreOp*);
   virtual void mutate(MmaOp*);
   virtual void mutate(BroadcastOp*);
 

torch/csrc/jit/codegen/cuda/executor_utils.cpp

@@ -28,6 +28,7 @@
 #include <nvfuser_resources/grid_sync.h>
 #include <nvfuser_resources/helpers.h>
 #include <nvfuser_resources/index_utils.h>
+#include <nvfuser_resources/memory.h>
 #include <nvfuser_resources/random_numbers.h>
 #include <nvfuser_resources/tensor.h>
 #include <nvfuser_resources/tensorcore.h>
@@ -98,6 +99,7 @@ std::string kernelPreamble() {
   ss << nvfuser_resources::welford_cu;
   ss << nvfuser_resources::warp_cu;
   ss << nvfuser_resources::tensorcore_cu;
+  ss << nvfuser_resources::memory_cu;
   ss << nvfuser_resources::fused_reduction_cu;
 
   // Random utilities

torch/csrc/jit/codegen/cuda/ir_builder.cpp

@@ -64,6 +64,7 @@ IR_BUILDER_INSTANTIATE(TernaryOp)
 IR_BUILDER_INSTANTIATE(ReductionOp)
 IR_BUILDER_INSTANTIATE(GroupedReductionOp)
 IR_BUILDER_INSTANTIATE(WelfordOp)
+IR_BUILDER_INSTANTIATE(LoadStoreOp)
 IR_BUILDER_INSTANTIATE(MmaOp)
 IR_BUILDER_INSTANTIATE(BroadcastOp)
 

torch/csrc/jit/codegen/cuda/ir_cloner.cpp

@@ -116,6 +116,10 @@ void IrCloner::handle(const WelfordOp* op) {
   clone_ = IrBuilder::clone(op, this);
 }
 
+void IrCloner::handle(const LoadStoreOp* op) {
+  clone_ = IrBuilder::clone(op, this);
+}
+
 void IrCloner::handle(const MmaOp* op) {
   clone_ = IrBuilder::clone(op, this);
 }

torch/csrc/jit/codegen/cuda/ir_cloner.h

@@ -75,6 +75,7 @@ class TORCH_CUDA_CU_API IrCloner : private OptInConstDispatch {
   void handle(const ReductionOp*) override;
   void handle(const GroupedReductionOp*) override;
   void handle(const WelfordOp*) override;
+  void handle(const LoadStoreOp*) override;
   void handle(const MmaOp*) override;
   void handle(const TransposeOp*) override;
   void handle(const ShiftOp*) override;

torch/csrc/jit/codegen/cuda/ir_interface_nodes.h

@@ -402,14 +402,22 @@ class TORCH_CUDA_CU_API TensorView : public Val {
       const std::vector<int>& axes,
       const std::vector<TensorView*>& tvs);
 
-  // Create a TensorView before the original tensor. A common use case is to
-  // write results into shared memory or registers before moving to global
-  // memory. Analogous to TVM Cache_Write
-  TensorView* cacheBefore();
+  //! Create a TensorView before the original tensor. A common use case is to
+  //! write results into shared memory or registers before moving to global
+  //! memory. Analogous to TVM Cache_Write
+  //!
+  //! @param cache_op: memory operator to use for the inserted op between
+  //!   the the data tensor and the cache tensor
+  TensorView* cacheBefore(
+      c10::optional<LoadStoreOpType> cache_op = c10::nullopt);
 
-  // Create a TensorView after the original tensor. A common use case is to
-  // read tensor into shared memory or registers. Analogous to TVM Cache_Read
-  TensorView* cacheAfter();
+  //! Create a TensorView after the original tensor. A common use case is to
+  //! read tensor into shared memory or registers. Analogous to TVM Cache_Read
+  //!
+  //! @param cache_op: memory operator to use for the inserted op between
+  //!   the the data tensor and the cache tensor
+  TensorView* cacheAfter(
+      c10::optional<LoadStoreOpType> cache_op = c10::nullopt);
 
   // For a fusion output with other uses, we want to avoid writing to global
   // memory and then reading the output again. We write to global memory
@@ -438,17 +446,6 @@ class TORCH_CUDA_CU_API TensorView : public Val {
     return is_double_buffered_;
   }
 
-  //! Fill in mma options in scheduling time.
-  //!  Each mma op in Fusion IR must be configured once before lowering.
-  //!  Mma options are configuration parameters used in lowering to mma
-  //!  instrinsics, mainly the type of mma macro to use and input data layout
-  //!  etc.
-  //!
-  //! TODO: This step will very likely be removed in a follow up PR. All of
-  //!  the options configured here could actually be inferred from fusion IR
-  //!  once we are feature complete.
-  void configureMma(MmaOptions options);
-
   //! Transforms the innermost iterdomains according to the given mma swizzle,
   //!  this should be used on the tvs that are either inputs/outputs of an
   //!  MmaOp, or any tv's that are involved in prolog/epilog fusions and need to

torch/csrc/jit/codegen/cuda/ir_internal_nodes.h

@@ -611,6 +611,36 @@ class TORCH_CUDA_CU_API ViewOp : public Expr {
   TensorView* const in_ = nullptr;
 };
 
+//! This operator explicitly models data movement between
+//!   state spaces on GPU. Currently the modeled state spaces include
+//!   global memory, shared memory and register.
+//!
+//! The main usage of this op is to facilitate generation of hardware
+//!   accelerated memory ops, i.e. ldmatrix, cp.async and more to come.
+class TORCH_CUDA_CU_API LoadStoreOp : public Expr {
+ public:
+  LoadStoreOp(IrBuilderPasskey, LoadStoreOpType op_type, Val* out, Val* in);
+
+  LoadStoreOp(const LoadStoreOp* src, IrCloner* ir_cloner);
+
+  Val* out() const {
+    return out_;
+  }
+
+  Val* in() const {
+    return in_;
+  }
+
+  LoadStoreOpType opType() const {
+    return load_store_type_;
+  }
+
+ private:
+  LoadStoreOpType load_store_type_ = LoadStoreOpType::LdMatrix;
+  Val* const out_ = nullptr;
+  Val* const in_ = nullptr;
+};
+
 // Friends for direct access to split
 class TensorDomain;
 class ReplayTransformations;