[TIR] Add VNNI dot product intrinsic for TIR

python/tvm/script/tir/special_stmt.py

@@ -25,10 +25,9 @@
 
 import tvm.tir
 from tvm.runtime import Object, String
-from tvm import te
 from tvm.target import Target
 from tvm.ir import Span
-from tvm.tir import IntImm, IterVar
+from tvm.tir import IntImm, IterVar, Var
 
 from .node import BufferSlice
 from .utils import buffer_slice_to_region
@@ -800,7 +799,7 @@ def var(dtype, span):
                 self.context.report_error(
                     f"VarDef expected assign to only one var, but got {names}", span
                 )
-            v = te.var(names[0], dtype, span=span)
+            v = Var(names[0], dtype, span=span)
             self.context.update_symbol(v.name, v, self.node)
 
         super().__init__(var, def_symbol=True)
@@ -821,7 +820,7 @@ def buffer_var(dtype, storage_scope, span):
                     f"VarDef expected assign to only one var, but got {names}", span
                 )
             ptr_type = tvm.ir.PointerType(tvm.ir.PrimType(dtype), storage_scope)
-            v = te.var(names[0], ptr_type, span=span)
+            v = Var(names[0], ptr_type, span=span)
             self.context.update_symbol(v.name, v, self.node)
 
         super().__init__(buffer_var, def_symbol=True)
@@ -841,7 +840,7 @@ def env_thread(env_name, span):
                 self.context.report_error(
                     f"VarDef expected assign to only one var, but got {names}", span
                 )
-            v = te.var(names[0], span=span)
+            v = Var(names[0], span=span)
             self.context.func_var_env_dict[v] = env_name
             self.context.update_symbol(v.name, v, self.node)
 

python/tvm/tir/tensor_intrin/__init__.py

@@ -0,0 +1,19 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=unused-import
+"""Intrinsics for tensorization."""
+from . import vnni

python/tvm/tir/tensor_intrin/vnni.py

@@ -0,0 +1,69 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+from .. import TensorIntrin
+from tvm.script import tir as T
+
+
+@T.prim_func
+def dot_product_desc(a: T.handle, b: T.handle, c: T.handle) -> None:
+    A = T.match_buffer(a, (4,), "uint8", offset_factor=1)
+    B = T.match_buffer(b, (16, 4), "int8", offset_factor=1)
+    C = T.match_buffer(c, (16,), "int32", offset_factor=1)
+
+    with T.block("root"):
+        T.reads(C[0:16], A[0:4], B[0:16, 0:4])
+        T.writes(C[0:16])
+        for i in T.serial(0, 16):
+            with T.init():
+                C[i] = T.int32(0)
+            for k in T.serial(0, 4):
+                with T.block("update"):
+                    vi, vk = T.axis.remap("SR", [i, k])
+                    C[vi] = C[vi] + T.cast(A[vk], "int32") * T.cast(B[vi, vk], "int32")
+
+
+@T.prim_func
+def dot_product_intrin(a: T.handle, b: T.handle, c: T.handle) -> None:
+    A = T.match_buffer(a, (4,), "uint8", offset_factor=1)
+    B = T.match_buffer(b, (16, 4), "int8", offset_factor=1)
+    C = T.match_buffer(c, (16,), "int32", offset_factor=1)
+
+    with T.block("root"):
+        T.reads(C[0:16], A[0:4], B[0:16, 0:4])
+        T.writes(C[0:16])
+
+        A_u8x4 = A.vload([0], "uint8x4")
+        A_i32 = T.reinterpret(A_u8x4, dtype="int32")
+
+        B_i8x64 = B.vload([0, 0], dtype="int8x64")
+        B_i32x16 = T.reinterpret(B_i8x64, dtype="int32x16")
+
+        C[
+            T.ramp(T.int32(0), 1, 16)
+        ] += T.call_llvm_pure_intrin(  # Note: this is an update +=
+            T.llvm_lookup_intrinsic_id("llvm.x86.avx512.vpdpbusd.512"),
+            T.uint32(0),
+            T.int32x16(0),
+            T.broadcast(A_i32, 16),
+            B_i32x16,
+            dtype="int32x16",
+        )
+
+
+TensorIntrin.register(
+    "dot_16x1x16_uint8_int8_int32_cascadelake", dot_product_desc, dot_product_intrin
+)