[TIR] Refactor BF16Legalize

This PR refactors BF16Legalize to enable more f32 computations.
We also split the BF16Legalize into two steps.

- BF16ComputeLegalize changes all computation to f32 while keeping
  the external BF16 storages.
- BF16StorageLegalize changes all storage to u16.

Now BF16 kernels accept tvm.nd.array that are created as bfloat16 type.

Author: tqchen

include/tvm/tir/transform.h

@@ -337,11 +337,17 @@ TVM_DLL Pass CombineContextCall();
 TVM_DLL Pass NarrowDataType(int target_bits);
 
 /*!
- * \brief Legalize bf16 typed Ops. Add a cast to fp32
+ * \brief Legalize bf16 compute Ops. Add a cast to fp32
  *   before Ops, then add a cast back to bf16.
  * \return The pass.
  */
-TVM_DLL Pass BF16Legalize();
+TVM_DLL Pass BF16ComputeLegalize();
+
+/*!
+ * \brief Legalize bf16 storage types to u16.
+ * \return The pass.
+ */
+TVM_DLL Pass BF16StorageLegalize();
 
 /*!
  * \brief Rewrite the pointer content type of arguments,

include/tvm/topi/elemwise.h

@@ -310,11 +310,7 @@ inline Tensor cast(const Tensor& x, DataType type, std::string name = "T_cast",
 inline Tensor reinterpret(const Tensor& x, DataType type, std::string name = "tensor",
                           std::string tag = kElementWise) {
   return compute(
-      x->shape,
-      [&](const Array<Var>& i) {
-        return tvm::tir::Call(type, tvm::tir::builtin::reinterpret(), {x(i)});
-      },
-      name, tag);
+      x->shape, [&](const Array<Var>& i) { return reinterpret(type, x(i)); }, name, tag);
 }
 
 /*!

python/tvm/tir/transform/transform.py

@@ -286,59 +286,26 @@ def RemoveStoreUndef():
     return _ffi_api.RemoveStoreUndef()  # type: ignore
 
 
-def BF16Legalize():
-    """Legalize bf16 typed Ops.
-    Runs BF16Promote, BF16CastElimination and BF16TypeLowering
+def BF16ComputeLegalize():
+    """Legalize bf16 compute Ops.
 
     Returns
     -------
     fpass : tvm.transform.Pass
         The result pass
     """
-    return _ffi_api.BF16Legalize()  # type: ignore
+    return _ffi_api.BF16ComputeLegalize()  # type: ignore
 
 
-def BF16Promote():
-    """Promote bf16 to fp32. Add a cast to fp32
-    before Ops, then add a cast back to bf16.
+def BF16StorageLegalize():
+    """Legalize bf16 storage types to u16.
 
     Returns
     -------
     fpass : tvm.transform.Pass
         The result pass
     """
-    return _ffi_api.BF16Promote()  # type: ignore
-
-
-def BF16CastElimination():
-    """Eliminate verbose casting between fp32 and bf16
-    Checks if the AST has the pattern:
-    castto32(castto16(some_fp32_op(...)))
-    The verbose casting is generated by BF16Promote for multiple
-    bf16 Ops in a row. e.g.:
-    X[i] + Y[i] + T[i] =>
-    bf16((float32(bf16((float32(X[i]) + float32(Y[i])))) + float32(T[i])))
-    After this pass:
-    bf16(float32(X[i]) + float32(Y[i]) + float32(T[i]))
-
-    Returns
-    -------
-    fpass : tvm.transform.Pass
-        The result pass
-    """
-    return _ffi_api.BF16CastElimination()  # type: ignore
-
-
-def BF16TypeLowering():
-    """Replace all bf16 type with uint16. Also lower the casting
-    between fp32 and bf16
-
-    Returns
-    -------
-    fpass : tvm.transform.Pass
-        The result pass
-    """
-    return _ffi_api.BF16TypeLowering()  # type: ignore
+    return _ffi_api.BF16StorageLegalize()  # type: ignore
 
 
 def CommonSubexprElimTIR(enable_cse_tir: bool = True, identify_equiv_terms: bool = False):

src/driver/driver_api.cc

@@ -218,7 +218,7 @@ Array<tvm::transform::Pass> CreatePassList(bool disable_loop_partition) {
   pass_list.push_back(tir::transform::InjectSoftwarePipeline());
   pass_list.push_back(tir::transform::LowerOpaqueBlock());
   pass_list.push_back(tir::transform::FlattenBuffer());
-  pass_list.push_back(tir::transform::BF16Legalize());
+  pass_list.push_back(tir::transform::BF16ComputeLegalize());
   pass_list.push_back(tir::transform::NarrowDataType(32));
   pass_list.push_back(tir::transform::Simplify());
 
@@ -605,6 +605,7 @@ transform::Sequential MixedModulePassManager(IRModule mixed_mod, Target target)
   } else {
     mixed_pass_list.push_back(tir::transform::MakePackedAPI());
   }
+  mixed_pass_list.push_back(tir::transform::BF16StorageLegalize());
   mixed_pass_list.push_back(tir::transform::SplitHostDevice());
 
   return transform::Sequential(mixed_pass_list);

src/meta_schedule/postproc/disallow_async_strided_mem_copy.cc

@@ -138,7 +138,7 @@ class DisallowAsyncStridedMemCopyNode : public PostprocNode {
           pass_list.push_back(tir::transform::InjectSoftwarePipeline());
           pass_list.push_back(tir::transform::LowerOpaqueBlock());
           pass_list.push_back(tir::transform::FlattenBuffer());
-          pass_list.push_back(tir::transform::BF16Legalize());
+          pass_list.push_back(tir::transform::BF16ComputeLegalize());
           pass_list.push_back(tir::transform::NarrowDataType(32));
           pass_list.push_back(tir::transform::Simplify());
           pass_list.push_back(tir::transform::InjectVirtualThread());

src/meta_schedule/postproc/verify_gpu_code.cc

@@ -169,7 +169,7 @@ class VerifyGPUCodeNode : public PostprocNode {
           pass_list.push_back(tir::transform::InjectSoftwarePipeline());
           pass_list.push_back(tir::transform::LowerOpaqueBlock());
           pass_list.push_back(tir::transform::FlattenBuffer());
-          pass_list.push_back(tir::transform::BF16Legalize());
+          pass_list.push_back(tir::transform::BF16ComputeLegalize());
           pass_list.push_back(tir::transform::NarrowDataType(32));
           pass_list.push_back(tir::transform::Simplify());
           // Phase 2

src/target/codegen.cc

@@ -46,7 +46,6 @@ runtime::Module Build(IRModule mod, Target target) {
           .value()) {
     mod = tir::transform::SkipAssert()(mod);
   }
-
   auto target_attr_map = tvm::TargetKind::GetAttrMap<FTVMTIRToRuntime>("TIRToRuntime");
   if (target_attr_map.count(target->kind)) {
     return target_attr_map[target->kind](mod, target);

src/target/llvm/codegen_llvm.cc

@@ -828,6 +828,10 @@ void CodeGenLLVM::CreateSerialFor(llvm::Value* begin, llvm::Value* end, llvm::Va
 llvm::Value* CodeGenLLVM::CreateCast(DataType from, DataType to, llvm::Value* value) {
   llvm::Type* target = DTypeToLLVMType(to);
   if (value->getType() == target) return value;
+  // TODO(tvm-team): consider add native support
+  ICHECK(!from.is_bfloat16()) << "BF16 needs to be storaged lowered first";
+  ICHECK(!to.is_bfloat16()) << "BF16 needs to be storaged lowered first";
+
   if (to.is_handle()) {
     return builder_->CreateBitCast(value, target);
   } else if (to.is_uint() && to.bits() == 1) {

src/target/llvm/llvm_module.cc

@@ -325,7 +325,6 @@ void LLVMModuleNode::Init(const IRModule& mod, const Target& target) {
   if (tm->getTargetTriple().isOSDarwin()) {
     module_->addModuleFlag(llvm::Module::Override, "Dwarf Version", 2);
   }
-
   std::string verify_errors_storage;
   llvm::raw_string_ostream verify_errors(verify_errors_storage);
   LOG_IF(FATAL, llvm::verifyModule(*module_, &verify_errors))

src/tir/op/op.cc

@@ -324,6 +324,8 @@ PrimExpr cast(const DataType& t, PrimExpr value, Span span) {
 // reinterpret
 PrimExpr reinterpret(const DataType& t, PrimExpr value, Span span) {
   if (value.dtype() == t) return value;
+  ICHECK(value.dtype().bits() * value.dtype().lanes() == t.bits() * t.lanes())
+      << "Bitcast requires size match " << t << " vs " << value.dtype();
   return tir::Call(t, tir::builtin::reinterpret(), {value}, span);
 }