[ETHOSN] Consolidate target string usage

python/tvm/driver/tvmc/composite_target.py

@@ -18,7 +18,6 @@
 Provides support to composite target on TVMC.
 """
 import logging
-import warnings
 
 # Make sure Vitis AI codegen is registered
 import tvm.contrib.target.vitis_ai  # pylint: disable=unused-import
@@ -72,11 +71,6 @@
         "config_key": "relay.ext.vitis_ai.options",
         "pass_pipeline": partition_for_vitis_ai,
     },
-    # Deprecated in favour of "ethos-n".
-    "ethos-n78": {
-        "config_key": "relay.ext.ethos-n.options",
-        "pass_pipeline": partition_for_ethosn,
-    },
 }
 
 
@@ -105,12 +99,6 @@ def get_codegen_by_target(name):
         requested target codegen information
     """
     try:
-        if name == "ethos-n78":
-            warnings.warn(
-                "Please use 'ethos-n' instead of the deprecated 'ethos-n78' target, "
-                "which will be removed in a later release of TVM.",
-                DeprecationWarning,
-            )
         return REGISTERED_CODEGEN[name]
     except KeyError:
         raise TVMCException("Composite target %s is not defined in TVMC." % name)

python/tvm/relay/op/contrib/ethosn.py

@@ -17,7 +17,6 @@
 # pylint: disable=invalid-name, unused-argument
 """Arm(R) Ethos(TM)-N NPU supported operators."""
 from enum import Enum
-import warnings
 from distutils.version import LooseVersion
 
 import tvm.ir
@@ -97,6 +96,8 @@ def is_inline_non_compute_intensive_partitions_enabled() -> bool:
     True if inlining should happen, False if not.
     """
     compiler_attrs = tvm.get_global_func("relay.ext.ethos-n.get_compiler_attrs")()
+    if not compiler_attrs:
+        return False
     return compiler_attrs.inline_non_compute_intensive_partitions
 
 
@@ -115,20 +116,6 @@ def partition_for_ethosn(mod, params=None, **opts):
     -------
     ret : annotated and partitioned module.
     """
-    opts = opts or {}
-    if "variant" not in opts:
-        raise ValueError("Please specify a variant in the target string, e.g. -variant=n78.")
-
-    # -variant=ethos-n78 deprecated in favour of -variant=n78
-    if opts["variant"].lower() == "ethos-n78":
-        warnings.warn(
-            "Please use '-variant=n78' instead of the deprecated "
-            "'-variant=ethos-n78', which will be removed in TVM v0.9.",
-            DeprecationWarning,
-        )
-    elif opts["variant"] != "n78":
-        raise ValueError("When targeting Ethos(TM)-N78, -variant=n78 should be set.")
-
     api_version = ethosn_api_version()
     supported_api_versions = ["3.1.0"]
     if all(api_version != LooseVersion(exp_ver) for exp_ver in supported_api_versions):

src/relay/backend/contrib/ethosn/codegen.cc

@@ -226,36 +226,27 @@ sl::TensorsAndId MakeOps(const sl::TensorAndId<sl::Operand>& op) {
   return ops;
 }
 
-String MakeVariant(Optional<EthosnCompilerConfig> configuration) {
-  String variant = configuration.value()->variant;
-  // Transform variant string to lowercase for comparison
-  std::string variant_string = variant.c_str();
-
-  // Checking deprecated variant format. Support for specifying
-  // the variant in this way only remains for backwards compatibility
-  // and will be removed in a later release of TVM.
-  std::string deprecated_variant_string = variant_string;
-  std::transform(deprecated_variant_string.begin(), deprecated_variant_string.end(),
-                 deprecated_variant_string.begin(), ::tolower);
-  if (variant_string == "n78" || deprecated_variant_string == "ethos-n78") {
-    String tops = configuration.value()->tops;
-    String ple_ratio = configuration.value()->ple_ratio;
-    variant = "Ethos-N78_" + tops + "TOPS_" + ple_ratio + "PLE_RATIO";
-  }
-  return variant;
+sl::EthosNVariant MakeVariant(EthosnCompilerConfig configuration) {
+  String variant = configuration->variant;
+  String tops = configuration->tops;
+  String ple_ratio = configuration->ple_ratio;
+
+  std::string capitalized_variant = variant;
+  std::transform(capitalized_variant.begin(), capitalized_variant.end(),
+                 capitalized_variant.begin(), ::toupper);
+  std::string sl_variant_string =
+      "Ethos-" + capitalized_variant + "_" + tops + "TOPS_" + ple_ratio + "PLE_RATIO";
+  return sl::EthosNVariantFromString(sl_variant_string.c_str());
 }
 
 NetworkWithIDs ConstructNetworkVisitor::Construct(const Function& func) {
   // Initialise everything
-  auto ctx = transform::PassContext::Current();
-  auto cfg = ctx->GetConfig<EthosnCompilerConfig>("relay.ext.ethos-n.options");
-  if (!cfg.defined()) {
-    cfg = AttrsWithDefaultValues<EthosnCompilerConfig>();
-  }
+  EthosnCompilerConfig cfg = GetCompilerAttrs();
+  sl::EthosNVariant variant = MakeVariant(cfg);
+
   NetworkWithIDs network_with_ids;
   network_ = sl::CreateNetwork(
-      sl::GetFwAndHwCapabilities(sl::EthosNVariantFromString(MakeVariant(cfg).c_str()),
-                                 static_cast<uint32_t>(std::stoul(cfg.value()->sram_size))));
+      sl::GetFwAndHwCapabilities(variant, static_cast<uint32_t>(std::stoul(cfg->sram_size))));
   network_with_ids.network = network_;
   operand_table_.clear();
 
@@ -744,28 +735,24 @@ runtime::ethosn::OrderedCompiledNetwork EthosnCompiler::CompileEthosnFunc(const
 }
 
 sl::CompilationOptions EthosnCompiler::CreateOptions() {
-  auto ctx = transform::PassContext::Current();
-  auto cfg = ctx->GetConfig<EthosnCompilerConfig>("relay.ext.ethos-n.options");
-  if (!cfg.defined()) {
-    cfg = AttrsWithDefaultValues<EthosnCompilerConfig>();
-  }
+  EthosnCompilerConfig cfg = GetCompilerAttrs();
 
   sl::CompilationOptions options;
-  options.m_Strategy0 = cfg.value()->strategy0;
-  options.m_Strategy1 = cfg.value()->strategy1;
-  options.m_Strategy3 = cfg.value()->strategy3;
-  options.m_Strategy4 = cfg.value()->strategy4;
-  options.m_Strategy6 = cfg.value()->strategy6;
-  options.m_Strategy7 = cfg.value()->strategy7;
-  options.m_DebugInfo.m_DumpRam = cfg.value()->dump_ram;
-  options.m_DebugInfo.m_InitialSramDump = cfg.value()->initial_sram_dump;
-  options.m_BlockConfig16x16 = cfg.value()->block_config_16x16;
-  options.m_BlockConfig32x8 = cfg.value()->block_config_32x8;
-  options.m_BlockConfig8x32 = cfg.value()->block_config_8x32;
-  options.m_BlockConfig8x8 = cfg.value()->block_config_8x8;
-  options.m_EnableIntermediateCompression = cfg.value()->enable_intermediate_compression;
-  options.m_DisableWinograd = cfg.value()->disable_winograd;
-  options.m_DebugInfo.m_DebugDir = cfg.value()->debug_dir;
+  options.m_Strategy0 = cfg->strategy0;
+  options.m_Strategy1 = cfg->strategy1;
+  options.m_Strategy3 = cfg->strategy3;
+  options.m_Strategy4 = cfg->strategy4;
+  options.m_Strategy6 = cfg->strategy6;
+  options.m_Strategy7 = cfg->strategy7;
+  options.m_DebugInfo.m_DumpRam = cfg->dump_ram;
+  options.m_DebugInfo.m_InitialSramDump = cfg->initial_sram_dump;
+  options.m_BlockConfig16x16 = cfg->block_config_16x16;
+  options.m_BlockConfig32x8 = cfg->block_config_32x8;
+  options.m_BlockConfig8x32 = cfg->block_config_8x32;
+  options.m_BlockConfig8x8 = cfg->block_config_8x8;
+  options.m_EnableIntermediateCompression = cfg->enable_intermediate_compression;
+  options.m_DisableWinograd = cfg->disable_winograd;
+  options.m_DebugInfo.m_DebugDir = cfg->debug_dir;
   return options;
 }
 
@@ -806,12 +793,10 @@ std::unique_ptr<sl::SupportQueries> EthosnCompiler::m_Queries;
 
 EthosnError EthosnCompiler::SupportedSetup() {
   if (m_Queries == nullptr) {
-    auto ctx = transform::PassContext::Current();
-    auto cfg = ctx->GetConfig<EthosnCompilerConfig>("relay.ext.ethos-n.options").defined()
-                   ? ctx->GetConfig<EthosnCompilerConfig>("relay.ext.ethos-n.options")
-                   : AttrsWithDefaultValues<EthosnCompilerConfig>();
-    m_Queries = std::make_unique<sl::SupportQueries>(sl::GetFwAndHwCapabilities(
-        sl::EthosNVariantFromString(MakeVariant(cfg).c_str()), std::stoul(cfg.value()->sram_size)));
+    EthosnCompilerConfig cfg = GetCompilerAttrs();
+    sl::EthosNVariant variant = MakeVariant(cfg);
+    m_Queries = std::make_unique<sl::SupportQueries>(
+        sl::GetFwAndHwCapabilities(variant, std::stoul(cfg->sram_size)));
     if (m_Queries == nullptr) {
       return EthosnError("Could not initialise Arm(R) Ethos(TM)-N compiler isSupported");
     }

src/relay/backend/contrib/ethosn/codegen_ethosn.h

@@ -296,13 +296,14 @@ class EthosnCompilerConfig : public Attrs {
 TVM_REGISTER_NODE_TYPE(EthosnCompilerConfigNode);
 TVM_REGISTER_PASS_CONFIG_OPTION("relay.ext.ethos-n.options", EthosnCompilerConfig);
 
-auto GetCompilerAttrs() {
+EthosnCompilerConfig GetCompilerAttrs() {
   auto ctx = transform::PassContext::Current();
-  auto cfg = ctx->GetConfig<EthosnCompilerConfig>("relay.ext.ethos-n.options");
+  Optional<EthosnCompilerConfig> cfg =
+      ctx->GetConfig<EthosnCompilerConfig>("relay.ext.ethos-n.options");
   if (!cfg.defined()) {
-    cfg = AttrsWithDefaultValues<EthosnCompilerConfig>();
+    return AttrsWithDefaultValues<EthosnCompilerConfig>();
   }
-  return cfg;
+  return cfg.value();
 }
 TVM_REGISTER_GLOBAL("relay.ext.ethos-n.get_compiler_attrs").set_body_typed(GetCompilerAttrs);
 

tests/python/contrib/test_ethosn/infrastructure.py

@@ -31,6 +31,7 @@
 from tvm import relay
 from tvm.contrib import utils, graph_executor, download
 from tvm.relay.op.contrib import partition_for_ethosn
+from tvm.driver.tvmc.target import parse_target
 
 from . import _infrastructure
 
@@ -143,7 +144,9 @@ def visit_call(self, call):
     return c.count
 
 
-def build(mod, params, npu=True, expected_host_ops=0, npu_partitions=1, optimize_partitions=True):
+def build(
+    mod, params, npu=True, expected_host_ops=0, npu_partitions=1, additional_config_args=None
+):
     """Build a network with or without Ethos-N offloading.
 
     Parameters
@@ -158,22 +161,18 @@ def build(mod, params, npu=True, expected_host_ops=0, npu_partitions=1, optimize
         The number of ops expected to remain on the host.
     npu_partitions : int, optional
         The number of Ethos-N partitions expected.
-    optimize_partitions : bool, optional
-        Disable the pass that optimizes NPU partitions post partitioning.
+    additional_config_args : dict, optional
+        Additional compiler config options for the NPU.
     """
     relay.backend.te_compiler.get().clear()
-    with tvm.transform.PassContext(
-        opt_level=3,
-        config={
-            "relay.ext.ethos-n.options": {
-                "variant": get_ethosn_variant(),
-                "inline_non_compute_intensive_partitions": optimize_partitions,
-            }
-        },
-    ):
+    if not additional_config_args:
+        additional_config_args = {}
+    npu_config = {**get_ethosn_device_options(), **additional_config_args}
+    print(npu_config)
+    with tvm.transform.PassContext(opt_level=3, config={"relay.ext.ethos-n.options": npu_config}):
         with tvm.target.Target("llvm"):
             if npu:
-                mod = partition_for_ethosn(mod, params, variant="n78")
+                mod = partition_for_ethosn(mod, params)
                 host_op_count = get_host_op_count(mod)
                 assert (
                     host_op_count == expected_host_ops
@@ -244,12 +243,12 @@ def build_and_run(
     npu=True,
     expected_host_ops=0,
     npu_partitions=1,
-    optimize_partitions=True,
+    additional_config_args=None,
 ):
     """
     Convenient wrapper for building and running a module on the NPU.
     """
-    lib = build(mod, params, npu, expected_host_ops, npu_partitions, optimize_partitions)
+    lib = build(mod, params, npu, expected_host_ops, npu_partitions, additional_config_args)
     return run(lib, inputs, outputs, npu)
 
 
@@ -285,7 +284,7 @@ def test_error(mod, params, err_msg):
 
     caught = None
     with tvm.transform.PassContext(
-        opt_level=3, config={"relay.ext.ethos-n.options": {"variant": get_ethosn_variant()}}
+        opt_level=3, config={"relay.ext.ethos-n.options": get_ethosn_device_options()}
     ):
         with tvm.target.Target("llvm"):
             try:
@@ -403,5 +402,9 @@ def get_same_padding(
     return (pad_top, pad_left, pad_bottom, pad_right)
 
 
-def get_ethosn_variant():
-    return os.getenv("ETHOSN_VARIANT_CONFIG", default="Ethos-N78_1TOPS_2PLE_RATIO")
+def get_ethosn_device_options():
+    """Determine the NPU configuration used for testing."""
+    default_target_string = "ethos-n -variant=n78 -tops=1 -ple_ratio=2"
+    target_string = os.getenv("ETHOSN_TEST_TARGET_CONFIG", default_target_string)
+    target = parse_target(target_string)
+    return target[0]["opts"]

tests/python/contrib/test_ethosn/test_addition.py

@@ -111,7 +111,16 @@ def test_addition(dtype, shape):
     model = _get_model(shape, shape, lhs_zp, lhs_sc, rhs_zp, rhs_sc, out_zp, out_sc, dtype)
     for npu in [False, True]:
         mod = tei.make_module(model, [])
-        outputs.append(tei.build_and_run(mod, inputs, 1, {}, npu=npu, optimize_partitions=False))
+        outputs.append(
+            tei.build_and_run(
+                mod,
+                inputs,
+                1,
+                {},
+                npu=npu,
+                additional_config_args={"inline_non_compute_intensive_partitions": False},
+            )
+        )
 
     tei.verify(outputs, dtype, 1)
 
@@ -227,7 +236,16 @@ def test_addition_to_reinterpret_quantize(lhs_shape, lhs_is_constant, rhs_shape,
     outputs = []
     for npu in [False, True]:
         mod = tei.make_module(model, {})
-        outputs.append(tei.build_and_run(mod, inputs, 1, {}, npu=npu, optimize_partitions=False))
+        outputs.append(
+            tei.build_and_run(
+                mod,
+                inputs,
+                1,
+                {},
+                npu=npu,
+                additional_config_args={"inline_non_compute_intensive_partitions": False},
+            )
+        )
     tei.verify(outputs, dtype, 1)
 
 

tests/python/contrib/test_ethosn/test_codegen.py

@@ -0,0 +1,52 @@
+# 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.
+
+"""NPU codegen tests"""
+
+import pytest
+import numpy as np
+
+import tvm
+from tvm import relay
+from tvm.testing import requires_ethosn
+
+from . import infrastructure as tei
+
+
+@requires_ethosn
+def test_compile_with_unsupported_variant():
+    """Test compilation with unsupported variant."""
+    dtype = "int8"
+    input_shape = (1, 2, 2, 2)
+
+    x = relay.var("x", shape=input_shape, dtype=dtype)
+    y = relay.reshape(x, newshape=(1, 1, 1, 8))
+    mod = tei.make_ethosn_partition(y)
+
+    additional_config_args = {
+        "variant": "foo",
+        "inline_non_compute_intensive_partitions": False,
+    }
+
+    inputs = {
+        "x": np.random.randint(
+            low=np.iinfo(dtype).min, high=np.iinfo(dtype).max, size=input_shape, dtype=dtype
+        )
+    }
+
+    with pytest.raises(tvm.TVMError, match=r"Unknown NPU type"):
+        tei.build_and_run(mod, inputs, 1, {}, True, additional_config_args=additional_config_args)

tests/python/contrib/test_ethosn/test_concatenate.py

@@ -76,7 +76,16 @@ def test_concatenate(dtype, shapes, axis):
     for npu in [False, True]:
         model = _get_model(shapes, dtype, axis)
         mod = tei.make_module(model, {})
-        outputs.append(tei.build_and_run(mod, inputs, 1, {}, npu=npu, optimize_partitions=False))
+        outputs.append(
+            tei.build_and_run(
+                mod,
+                inputs,
+                1,
+                {},
+                npu=npu,
+                additional_config_args={"inline_non_compute_intensive_partitions": False},
+            )
+        )
 
         tei.verify(outputs, dtype, 0)