You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2021/12/01 16:27:52 UTC
[GitHub] [tvm] lhutton1 opened a new pull request #9626: [microNPU] Support different constant datatypes
lhutton1 opened a new pull request #9626:
URL: https://github.com/apache/tvm/pull/9626
Currently only uint8 datatype is supported for constants, as this is all that was necessary until now. This PR allows different datatypes to be used for constants, including different datatypes within the same graph.
A workaround was previously added for Mean legalization, this has also been removed and replaced with the expected datatype of the constant.
Note: this PR is dependent on #9576 so contains the contents of that PR also.
cc @manupa-arm @ekalda @mbaret @NicolaLancellotti @dchauhan-arm
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] lhutton1 commented on a change in pull request #9626: [microNPU] Support different constant datatypes
Posted by GitBox <gi...@apache.org>.
lhutton1 commented on a change in pull request #9626:
URL: https://github.com/apache/tvm/pull/9626#discussion_r761254958
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -1156,6 +1156,7 @@ def transform_module(
mod = LegalizeReshape()(mod)
mod = LegalizeStridedSlice()(mod)
mod = LegalizeNoOps()(mod)
+ print(mod)
Review comment:
Thanks! Will wait till CI is fixed :)
##########
File path: python/tvm/relay/backend/contrib/ethosu/tir_to_cs_translator.py
##########
@@ -277,27 +276,24 @@ def classify_io(buffer):
raise ValueError(f"Unused IO : {buffer} in tir module.")
scratch_size = 0
- constant_tensor = None
+ constant_hex_data = []
+ total_constant_len = 0
buffer_addresses = dict()
for _buffer, info in buffer_info.items():
if info.values is not None:
- assert np.dtype(info.dtype) == np.uint8
assert info.btype == BufferType.constant
assert len(info.shape) == 1
- if constant_tensor is None:
- buffer_addresses[_buffer] = (0, info.btype)
- assert info.values.dtype == np.uint8
- size_in_bytes = info.values.size
- # Every memory address the NPU access have to be 16 byte aligned
- size_in_bytes = util.round_up(size_in_bytes, 16)
- constant_tensor = np.resize(info.values, size_in_bytes)
- else:
- buffer_addresses[_buffer] = (constant_tensor.size, info.btype)
- assert info.values.dtype == np.uint8
- size_in_bytes = info.values.size
- # Every memory address the NPU access have to be 16 byte aligned
- size_in_bytes = util.round_up(size_in_bytes, 16)
- constant_tensor = np.append(constant_tensor, np.resize(info.values, size_in_bytes))
+ buffer_addresses[_buffer] = (
+ (total_constant_len, info.btype) if constant_hex_data else (0, info.btype)
+ )
+ dtype_bytes = np.iinfo(np.dtype(info.dtype)).bits // 8
+ size_in_bytes = dtype_bytes * np.prod(list(info.shape))
+ # Every memory address the NPU access have to be 16 byte aligned
+ size_in_bytes = util.round_up(size_in_bytes, 16)
+ constant_tensor = np.resize(info.values, size_in_bytes // dtype_bytes)
+ constant_tensor = constant_tensor.tobytes().hex()
+ constant_hex_data.append(constant_tensor)
+ total_constant_len += len(constant_tensor) // 2
Review comment:
Each byte is represented by two hex values
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -1040,7 +1040,7 @@ def callback(
n = int(filter_height * filter_width)
eps = 1 / (256 * (n + 1)) if n % 2 == 0 else 0
- scalar_tensor = relay.const(np.ones([1, 1, 1, 1], dtype="uint8"), dtype="uint8")
+ scalar_tensor = relay.const(np.ones([1, 1, 1, 1], dtype="int16"), dtype="int16")
Review comment:
Simply because the output of the depthwise convolution that comes before is `int16`, this is just to match the datatype of the other input
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] ekalda commented on a change in pull request #9626: [microNPU] Support different constant datatypes
Posted by GitBox <gi...@apache.org>.
ekalda commented on a change in pull request #9626:
URL: https://github.com/apache/tvm/pull/9626#discussion_r761199497
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -1156,6 +1156,7 @@ def transform_module(
mod = LegalizeReshape()(mod)
mod = LegalizeStridedSlice()(mod)
mod = LegalizeNoOps()(mod)
+ print(mod)
Review comment:
Nit: remove the print
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -1040,7 +1040,7 @@ def callback(
n = int(filter_height * filter_width)
eps = 1 / (256 * (n + 1)) if n % 2 == 0 else 0
- scalar_tensor = relay.const(np.ones([1, 1, 1, 1], dtype="uint8"), dtype="uint8")
+ scalar_tensor = relay.const(np.ones([1, 1, 1, 1], dtype="int16"), dtype="int16")
Review comment:
For enlightenment, why `int16` (as opposed to e.g. `int32`)?
##########
File path: python/tvm/relay/backend/contrib/ethosu/tir_to_cs_translator.py
##########
@@ -277,27 +276,24 @@ def classify_io(buffer):
raise ValueError(f"Unused IO : {buffer} in tir module.")
scratch_size = 0
- constant_tensor = None
+ constant_hex_data = []
+ total_constant_len = 0
buffer_addresses = dict()
for _buffer, info in buffer_info.items():
if info.values is not None:
- assert np.dtype(info.dtype) == np.uint8
assert info.btype == BufferType.constant
assert len(info.shape) == 1
- if constant_tensor is None:
- buffer_addresses[_buffer] = (0, info.btype)
- assert info.values.dtype == np.uint8
- size_in_bytes = info.values.size
- # Every memory address the NPU access have to be 16 byte aligned
- size_in_bytes = util.round_up(size_in_bytes, 16)
- constant_tensor = np.resize(info.values, size_in_bytes)
- else:
- buffer_addresses[_buffer] = (constant_tensor.size, info.btype)
- assert info.values.dtype == np.uint8
- size_in_bytes = info.values.size
- # Every memory address the NPU access have to be 16 byte aligned
- size_in_bytes = util.round_up(size_in_bytes, 16)
- constant_tensor = np.append(constant_tensor, np.resize(info.values, size_in_bytes))
+ buffer_addresses[_buffer] = (
+ (total_constant_len, info.btype) if constant_hex_data else (0, info.btype)
+ )
+ dtype_bytes = np.iinfo(np.dtype(info.dtype)).bits // 8
+ size_in_bytes = dtype_bytes * np.prod(list(info.shape))
+ # Every memory address the NPU access have to be 16 byte aligned
+ size_in_bytes = util.round_up(size_in_bytes, 16)
+ constant_tensor = np.resize(info.values, size_in_bytes // dtype_bytes)
+ constant_tensor = constant_tensor.tobytes().hex()
+ constant_hex_data.append(constant_tensor)
+ total_constant_len += len(constant_tensor) // 2
Review comment:
For enlightenment, why division by 2?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] lhutton1 commented on a change in pull request #9626: [microNPU] Support different constant datatypes
Posted by GitBox <gi...@apache.org>.
lhutton1 commented on a change in pull request #9626:
URL: https://github.com/apache/tvm/pull/9626#discussion_r762966862
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -189,11 +176,11 @@ bool EthosuBinaryElementwiseRel(const Array<Type>& types, int num_inputs, const
return false;
}
if (ofm_dtype != DataType::UInt(8) && ofm_dtype != DataType::Int(8) &&
- ofm_dtype != DataType::Int(32)) {
+ ofm_dtype != DataType::Int(16) && ofm_dtype != DataType::Int(32)) {
reporter->GetDiagCtx().EmitFatal(
Diagnostic::Error(reporter->GetSpan())
<< "Invalid operator: expected ethosu_binary_elementwise " << operator_type
- << " type(uint8) or type(int8) or type(int32) for ofm but was " << ofm_dtype);
+ << " type(uint8), type(int8), type(int16) or type(int32) for ofm but was " << ofm_dtype);
return false;
}
} else if (operator_type == "MIN" || operator_type == "MAX") {
Review comment:
Given that adding `int16` to MIN and MAX is not essential for this change, I think it would be best to add as a separate patch WDYT?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9626: [microNPU] Support different constant datatypes
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9626:
URL: https://github.com/apache/tvm/pull/9626#discussion_r762934909
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -189,11 +176,11 @@ bool EthosuBinaryElementwiseRel(const Array<Type>& types, int num_inputs, const
return false;
}
if (ofm_dtype != DataType::UInt(8) && ofm_dtype != DataType::Int(8) &&
- ofm_dtype != DataType::Int(32)) {
+ ofm_dtype != DataType::Int(16) && ofm_dtype != DataType::Int(32)) {
reporter->GetDiagCtx().EmitFatal(
Diagnostic::Error(reporter->GetSpan())
<< "Invalid operator: expected ethosu_binary_elementwise " << operator_type
- << " type(uint8) or type(int8) or type(int32) for ofm but was " << ofm_dtype);
+ << " type(uint8), type(int8), type(int16) or type(int32) for ofm but was " << ofm_dtype);
return false;
}
} else if (operator_type == "MIN" || operator_type == "MAX") {
Review comment:
For MIN and MAX, ifm_dtype can be int16 too.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] manupa-arm commented on pull request #9626: [microNPU] Support different constant datatypes
Posted by GitBox <gi...@apache.org>.
manupa-arm commented on pull request #9626:
URL: https://github.com/apache/tvm/pull/9626#issuecomment-987850790
This is merged! Thanks @ekalda @lhutton1 @NicolaLancellotti.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] manupa-arm merged pull request #9626: [microNPU] Support different constant datatypes
Posted by GitBox <gi...@apache.org>.
manupa-arm merged pull request #9626:
URL: https://github.com/apache/tvm/pull/9626
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org