[ETHOSN] Remove support for older versions of the driver stack

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

@@ -18,6 +18,7 @@
 """Arm(R) Ethos(TM)-N NPU supported operators."""
 from enum import Enum
 import warnings
+from distutils.version import LooseVersion
 
 import tvm.ir
 from tvm.relay import transform
@@ -46,8 +47,16 @@ def ethosn_available():
     return Available.SW_AND_HW if hw else Available.SW_ONLY
 
 
-def ethosn_api_version():
-    """Returns the version of the driver stack api that is being used."""
+def ethosn_api_version() -> str:
+    """
+    Returns the semantic version of the driver stack api that is
+    being used.
+
+    Returns
+    -------
+    str
+        Semantic version string (e.g. 3.0.1).
+    """
     return tvm.get_global_func("relay.ethos-n.api.version")()
 
 
@@ -80,6 +89,14 @@ def partition_for_ethosn(mod, params=None, **opts):
     elif opts["variant"] != "n78":
         raise ValueError("When targeting Ethos(TM)-N78, -variant=n78 should be set.")
 
+    api_version = ethosn_api_version()
+    expected_api_version = "3.0.1"
+    if api_version != LooseVersion(expected_api_version):
+        raise ValueError(
+            f"Driver stack version {api_version} is unsupported. "
+            f"Please use version {expected_api_version}."
+        )
+
     if params:
         mod["main"] = bind_params_by_name(mod["main"], params)
 
@@ -298,7 +315,7 @@ def split(expr):
     """Check if a split is supported by Ethos-N."""
     if not ethosn_available():
         return False
-    if ethosn_api_version() == 2205:
+    if ethosn_api_version() >= LooseVersion("3.0.1"):
         return False
     if not support.split(expr):
         return False

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

@@ -22,6 +22,7 @@
  * \brief The Relay -> Arm(R) Ethos(TM)-N command stream compiler.
  */
 #include <tvm/relay/expr_functor.h>
+#include <tvm/runtime/container/string.h>
 #include <tvm/runtime/module.h>
 
 #include "codegen_ethosn.h"
@@ -965,8 +966,8 @@ TVM_REGISTER_GLOBAL("relay.ethos-n.query").set_body([](tvm::TVMArgs args, tvm::T
 #endif
 });
 
-TVM_REGISTER_GLOBAL("relay.ethos-n.api.version").set_body_typed([]() -> int {
-  return _ETHOSN_API_VERSION_;
+TVM_REGISTER_GLOBAL("relay.ethos-n.api.version").set_body_typed([]() -> String {
+  return sl::GetLibraryVersion().ToString();
 });
 
 }  // namespace ethosn

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

@@ -46,7 +46,6 @@
 #include "../../../../runtime/contrib/ethosn/ethosn_runtime.h"
 #include "../codegen_c/codegen_c.h"
 #include "ethosn_api.h"
-#include "ethosn_api_version.h"
 #include "ethosn_support_library/Support.hpp"
 #include "ethosn_support_library/SupportQueries.hpp"
 

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

@@ -36,7 +36,6 @@
 #include <utility>
 #include <vector>
 
-#include "ethosn_api_version.h"
 #include "ethosn_support_library/Support.hpp"
 #include "ethosn_support_library/SupportQueries.hpp"
 #include "tvm/relay/qnn/attrs.h"

src/runtime/contrib/ethosn/ethosn_device.cc

@@ -80,16 +80,9 @@ void CopyOutput(dl::Buffer* source_buffers[], std::vector<DLTensor*>* outputs) {
     dl::Buffer* source_buffer = source_buffers[0];
     T* dest_pointer = static_cast<T*>(tensor->data);
     size_t size = source_buffer->GetSize();
-#if _ETHOSN_API_VERSION_ < 2111
-    uint8_t* source_buffer_data = source_buffer->GetMappedBuffer();
-#else
     uint8_t* source_buffer_data = source_buffer->Map();
-#endif
     std::copy_backward(source_buffer_data, source_buffer_data + size, dest_pointer + size);
-#if _ETHOSN_API_VERSION_ >= 2111
     source_buffer->Unmap();
-#else
-#endif
     source_buffers++;
   }
 }
@@ -107,17 +100,10 @@ void CreateBuffers(std::vector<std::shared_ptr<dl::Buffer> >* fm,
   }
 }
 
-#if _ETHOSN_API_VERSION_ <= 2102
-bool Inference(tvm::runtime::TVMArgs args, sl::CompiledNetwork* network,
-               const std::vector<uint32_t>& input_order, const std::vector<uint32_t>& output_order,
-               const std::vector<uint32_t>& input_sizes,
-               const std::vector<uint32_t>& output_sizes) {
-#else
 bool Inference(tvm::runtime::TVMArgs args, dl::Network* npu,
                const std::vector<uint32_t>& input_order, const std::vector<uint32_t>& output_order,
                const std::vector<uint32_t>& input_sizes,
                const std::vector<uint32_t>& output_sizes) {
-#endif
   // Unpack parameters
   size_t n_inputs = input_order.size();
   size_t n_outputs = output_order.size();
@@ -149,10 +135,6 @@ bool Inference(tvm::runtime::TVMArgs args, dl::Network* npu,
     ofm_raw[i] = ofm[i].get();
   }
 
-#if _ETHOSN_API_VERSION_ <= 2102
-  auto npu = std::make_unique<dl::Network>(*network);
-#endif
-
   // Execute the inference.
   std::unique_ptr<dl::Inference> result(
       npu->ScheduleInference(ifm_raw, sizeof(ifm_raw) / sizeof(ifm_raw[0]), ofm_raw,
@@ -164,20 +146,13 @@ bool Inference(tvm::runtime::TVMArgs args, dl::Network* npu,
         dl::Buffer** ofms = &ofm_raw[0];
         for (DLTensor* tensor : outputs) {
           dl::Buffer* source_buffer = (*ofms++);
-#if _ETHOSN_API_VERSION_ < 2111
-          uint8_t* source_buffer_data = source_buffer->GetMappedBuffer();
-#else
           uint8_t* source_buffer_data = source_buffer->Map();
-#endif
           uint8_t* dest_pointer = static_cast<uint8_t*>(tensor->data);
           if (source_buffer_data != dest_pointer) {
             CopyOutput<uint8_t>(ofm_raw, &outputs);
             break;
           }
-#if _ETHOSN_API_VERSION_ >= 2111
           source_buffer->Unmap();
-#else
-#endif
         }
         break;
       }
@@ -229,17 +204,10 @@ TVM_REGISTER_GLOBAL("relay.ethos-n.test.infra.inference_result")
     });
 
 // Allow the ethos-n support code to be tested without a device
-#if _ETHOSN_API_VERSION_ <= 2102
-bool Inference(tvm::runtime::TVMArgs args, sl::CompiledNetwork* network,
-               const std::vector<uint32_t>& input_order, const std::vector<uint32_t>& output_order,
-               const std::vector<uint32_t>& input_sizes,
-               const std::vector<uint32_t>& output_sizes) {
-#else
 bool Inference(tvm::runtime::TVMArgs args, dl::Network* /* npu */,
                const std::vector<uint32_t>& input_order, const std::vector<uint32_t>& output_order,
                const std::vector<uint32_t>& input_sizes,
                const std::vector<uint32_t>& output_sizes) {
-#endif
   std::vector<DLTensor*> outputs;
   for (int argc = input_order.size(); argc < args.size(); argc++) {
     outputs.push_back(args[argc]);

src/runtime/contrib/ethosn/ethosn_device.h

@@ -28,7 +28,6 @@
 
 #include <vector>
 
-#include "../../../relay/backend/contrib/ethosn/ethosn_api_version.h"
 #include "ethosn_runtime.h"
 
 namespace tvm {
@@ -39,15 +38,9 @@ namespace dl = ::ethosn::driver_library;
 
 using tvm::runtime::TVMArgs;
 
-#if _ETHOSN_API_VERSION_ <= 2102
-bool Inference(TVMArgs args, sl::CompiledNetwork* npu, const std::vector<uint32_t>& input_order,
-               const std::vector<uint32_t>& output_order, const std::vector<uint32_t>& input_sizes,
-               const std::vector<uint32_t>& output_sizes);
-#else
 bool Inference(tvm::runtime::TVMArgs args, dl::Network* npu,
                const std::vector<uint32_t>& input_order, const std::vector<uint32_t>& output_order,
                const std::vector<uint32_t>& input_sizes, const std::vector<uint32_t>& output_sizes);
-#endif
 
 }  // namespace ethosn
 }  // namespace runtime

src/runtime/contrib/ethosn/ethosn_runtime.cc

@@ -53,11 +53,9 @@ EthosnModule::EthosnModule(std::vector<OrderedCompiledNetwork>* cmms) {
     if (it.compiled_cmm != nullptr) {
       network_map_[it.name].compiled_cmm = std::move(it.compiled_cmm);
     }
-#if _ETHOSN_API_VERSION_ > 2102
     if (it.runtime_cmm != nullptr) {
       network_map_[it.name].runtime_cmm = std::move(it.runtime_cmm);
     }
-#endif
     network_map_[it.name].inputs = it.inputs;
     network_map_[it.name].outputs = it.outputs;
     network_map_[it.name].input_sizes = it.input_sizes;
@@ -69,15 +67,9 @@ PackedFunc EthosnModule::GetFunction(const std::string& name,
                                      const ObjectPtr<Object>& sptr_to_self) {
   if (network_map_.find(name) != network_map_.end()) {
     return PackedFunc([sptr_to_self, this, name](TVMArgs args, TVMRetValue* rv) {
-#if _ETHOSN_API_VERSION_ <= 2102
-      *rv = Inference(args, network_map_[name].compiled_cmm.get(), network_map_[name].inputs,
-                      network_map_[name].outputs, network_map_[name].input_sizes,
-                      network_map_[name].output_sizes);
-#else
       *rv = Inference(args, network_map_[name].runtime_cmm.get(), network_map_[name].inputs,
                       network_map_[name].outputs, network_map_[name].input_sizes,
                       network_map_[name].output_sizes);
-#endif
     });
   } else {
     return PackedFunc();
@@ -119,14 +111,10 @@ Module EthosnModule::LoadFromBinary(void* strm) {
     // Read the serialized command stream
     stream->Read(&cmm);
     std::istringstream cmm_strm(cmm);
-#if _ETHOSN_API_VERSION_ <= 2102
-    compiled.compiled_cmm = sl::DeserializeCompiledNetwork(cmm_strm);
-#else
 #if defined ETHOSN_HW
     // If hardware unavaiable use the mock inference functionality. If hardware is
     // avaiable, deserialize the compiled graph.
     compiled.runtime_cmm = std::make_unique<dl::Network>(cmm.c_str(), cmm.size());
-#endif
 #endif
     // Read the number of inputs
     stream->Read<uint64_t>(&input_size);

src/runtime/contrib/ethosn/ethosn_runtime.h

@@ -33,7 +33,6 @@
 #include <unordered_map>
 #include <vector>
 
-#include "../../../relay/backend/contrib/ethosn/ethosn_api_version.h"
 #include "ethosn_driver_library/Network.hpp"
 #include "ethosn_support_library/Support.hpp"
 
@@ -46,9 +45,7 @@ namespace dl = ::ethosn::driver_library;
 
 struct OrderedCompiledNetwork {
   std::unique_ptr<sl::CompiledNetwork> compiled_cmm;
-#if _ETHOSN_API_VERSION_ > 2102
   std::unique_ptr<dl::Network> runtime_cmm;
-#endif
   std::string name;
   std::vector<uint32_t> inputs;
   std::vector<uint32_t> outputs;

tests/python/contrib/test_ethosn/infrastructure.py

@@ -270,33 +270,6 @@ def test_error(mod, params, err_msg):
     assert err_msg in caught, caught
 
 
-def get_overall_scale_range_expected_error_message():
-    """
-    Get the expected error message when the overall scale is out of range.
-    Different versions of the driver stack support different scale ranges,
-    so creating a unified utility to obtain the expected message.
-    """
-    if get_ethosn_api_version() >= 2205:
-        lb = "2^-32"
-    elif get_ethosn_api_version() > 2102:
-        lb = "2.328306e-10"
-    else:
-        lb = "0"
-
-    if get_ethosn_api_version() >= 2205:
-        ub = "65536"
-    else:
-        ub = "1"
-
-    lb_bracket = "(" if get_ethosn_api_version() >= 2205 else "["
-    ub_bracket = ")"
-
-    return (
-        "Overall scale (of the input * weights / output) should be in the range "
-        f"{lb_bracket}{lb}, {ub}{ub_bracket}"
-    )
-
-
 def get_conv2d(var, shape, dtype):
     """Standard convolution to test activation functions"""
 
@@ -359,10 +332,5 @@ def get_conv2d_qnn_params(
     return output_zp, output_sc
 
 
-def get_ethosn_api_version():
-    gf = tvm.get_global_func("relay.ethos-n.api.version", True)
-    return gf() if gf else None
-
-
 def get_ethosn_variant():
     return os.getenv("ETHOSN_VARIANT_CONFIG", default="Ethos-N78_1TOPS_2PLE_RATIO")

tests/python/contrib/test_ethosn/test_conv2d.py

@@ -226,7 +226,7 @@ def test_conv2d_failure():
             "uint8",
             8,
             "HWIO",
-            tei.get_overall_scale_range_expected_error_message(),
+            "Overall scale (of the input * weights / output) should be in the range (2^-32, 65536)",
         ),
         (
             (1, 4, 4, 4),

tests/python/contrib/test_ethosn/test_fullyconnected.py

@@ -127,7 +127,7 @@ def test_fullyconnected_failure():
             0,
             1,
             "uint8",
-            tei.get_overall_scale_range_expected_error_message(),
+            "Overall scale (of the input * weights / output) should be in the range (2^-32, 65536)",
         ),
         (
             (1, 1, 1, 64),