Specify the operand data type constraints of operation #283

huningxin · 2022-08-26T06:08:09Z

The current spec doesn't specify the operand type constraints of an operation. However, some operations, e.g., softmax should only support float32 operand type according to the survey of frameworks and native ML APIs in the following table. The lack of the operand type constraints specification would lead to implementation issue, such as Chromium CL 3856752.

Framework / API	supported data types
TensorFlow tf.nn.softmax	half, float32, float64
ONNX Softmax	tensor(float16), tensor(float), tensor(double), tensor(bfloat16)
NNAPI ANEURALNETWORKS_SOFTMAX	FLOAT32, FLOAT16
DirectML DML_ACTIVATION_SOFTMAX_OPERATOR_DESC	FLOAT32, FLOAT16
BNNS ActivationFunction	float

Thanks @wacky6 for pointing this out.

The text was updated successfully, but these errors were encountered:

huningxin · 2022-09-21T06:39:12Z

There are other operations should constrain the operand to floating-point types, e.g., batchNormalization, elu, hardSigmoid, hardSwish, instanceNormalization, leakyRelu, linear, sigmoid, softplus, softsign, tanh.

huningxin · 2023-04-11T04:48:14Z

Also for prelu op, it should also only accept floating-point types. It was raised by @wacky6 in the Chromium CL review. Thanks again Jiewei!

huningxin · 2023-04-19T03:21:02Z

For some element-wise unary ops, ceil and floor should only accept floating-point types. abs and neg should only accept signed types, including signed integer and floating-point types. This feedback is from Chromium CL review. Thanks @wacky6 and @miaobin!

Additionally, cos, exp, log, sin and tan should accept floating-point types.

huningxin · 2023-10-25T06:56:18Z

(Feedback raised by @wacky6 from Chromium CL review)

For reduction ops, reduceL2, reduceLogSum, reduceLogSumExp and reduceMean should only accept floating-point types.

/cc @lisa0314 @fdwr

fdwr · 2023-10-25T09:36:07Z

reduceL2, reduceLogSum, reduceLogSumExp and reduceMean should only accept floating-point types.

That's consistent with DML's data type support too:

REDUCE:
  ARGMIN, ARGMAX:
    featureLevel: 4.1
    InputDataType:  uint8, uint16, uint32, uint64, int8, int16, int32, int64, float16, float32
    OutputDataType: uint32, uint64, int32, int64
    MinRank: 1
    MaxRank: 8
  AVERAGE, L2, LOG_SUM, LOG_SUM_EXP:
    featureLevel: 3.0
    InputDataType:  float16, float32
    OutputDataType: float16, float32
    MinRank: 1
    MaxRank: 8
  L1, SUM_SQUARE:
    featureLevel: 5.0
    InputDataType:  uint8, uint16, uint32, uint64, int8, int16, int32, int64, float16, float32
    OutputDataType: uint8, uint16, uint32, uint64, int8, int16, int32, int64, float16, float32
    MinRank: 1
    MaxRank: 8
  MIN, MaxRank:
    featureLevel: 5.0
    InputDataType:  uint8, uint16, uint32, uint64, int8, int16, int32, int64, float16, float32
    OutputDataType: uint8, uint16, uint32, uint64, int8, int16, int32, int64, float16, float32
    MinRank: 1
    MaxRank: 8
  MULTIPLY, SUM:
    featureLevel: 5.0
    InputDataType:  uint8, uint16, uint32, uint64, int8, int16, int32, int64, float16, float32
    OutputDataType: uint8, uint16, uint32, uint64, int8, int16, int32, int64, float16, float32
    MinRank: 1
    MaxRank: 8

wacky6 · 2023-10-26T03:42:52Z

Just curious, what's the expected behavior for integer overflows for MULTIPLY and SUM?

I guess float overflow should just become Infinity.

fdwr · 2023-10-26T23:50:01Z

@wacky6

Just curious, what's the expected behavior for integer overflows for MULTIPLY and SUM?

I recall surveying a number of libraries a while back for integers, and they all did two's complement wrap rather than saturate. That is what DML does. I presume XNNPack too?

e.g.

import numpy

x = numpy.array(200, dtype=numpy.uint8)
y = numpy.add(x, x)
print("value:", y)
print("shape:", y.shape)

# Prints:
# value: 144
# shape: ()

I guess float overflow should just become Infinity.

👍

huningxin · 2023-11-20T07:41:21Z

[WIP] Summary of the operand data type constraints for current WebNN operations.

argMin/argMax

input: all supported data types
output: int64

batchNormalization

input: float32, float16
mean: same as input
variance: same as input
scale: same as input
bias: same as input
output: same as input

clamp

operand: all supported data types
output: same as input

concat

inputs: all supported data types
output: same as input

conv2d

input: float32, float16
filter: same as input
bias: same as input
output: same as input

convTranspose2d

input: float32, float16
filter: same as input
bias: same as input
output: same as input

Element-wise binary operations

add

a: all supported data types
b: same as a
output: same as a

sub

a: all supported data types
b: same as a
output: same as a

mul

a: all supported data types
b: same as a
output: same as a

div

a: all supported data types
b: same as a
output: same as a

min

a: all supported data types
b: same as a
output: same as a

min

a: all supported data types
b: same as a
output: same as a

pow

a: all supported data types
b: same as a
output: same as a

Element-wise unary operations

abs

input: float32, float16, int32, int8
output: same as input

ceil

input: float32, float16
output: same as input

cos

input: float32, float16
output: same as input

exp

input: float32, float16
output: same as input

floor

input: float32, float16
output: same as input

log

input: float32, float16
output: same as input

neg

input: float32, float16, int32, int8
output: same as input

sin

input: float32, float16
output: same as input

tan

input: float32, float16
output: same as input

elu

input: float32, float16
output: same as input

expand

input: all supported data types
output: same as input

gather

input: all supported data types
indices: uint32, int64
output: same as input

gelu

input: float32, float16
output: same as input

gemm

a: float32, float16
b: same as a
output: same as a

gru

input: float32, float16
weight: same as input
recurrentWeight: same as input
bias: same as input
recurrentBias: same as input
initialHiddenState: same as input
outputs: same as input

gruCell

input: float32, float16
weight: same as input
recurrentWeight: same as input
hiddenState: same as input
bias: same as input
recurrentBias: same as input
output: same as input

hardSigmoid

input: float32, float16
output: same as input

hardSwish

input: float32, float16
output: same as input

instanceNormalization

input: float32, float16
scale: same as input
bias: same as input
output: same as input

layerNormalization

input: float32, float16
scale: same as input
bias: same as input
output: same as input

leakyRelu

input: float32, float16
output: same as input

linear

input: float32, float16
output: same as input

lstm

input: float32, float16
weight: same as input
recurrentWeight: same as input
bias: same as input
recurrentBias: same as input
peepholeWeight: same as input
initialHiddenState: same as input
initialCellState: same as input
output: same as input

lstmCell

input: float32, float16
weight: same as input
recurrentWeight: same as input
hiddenState: same as input
cellState: same as input
bias: same as input
recurrentBias: same as input
peepholeWeight: same as input
outputs: same as input

matmul

a: float32, float16
b: same as a
output: same as a

pad

input: all supported data types
output: same as input

Pooling operations

averagePool2d

input: float32, float16
output: same as input

l2Pool2d

input: float32, float16
output: same as input

maxPool2d

input: all supported data types
output: same as input

prelu

input: float32, float16, int32, int8
output: same as input

Reduction operations

reduceMax

input: all supported data types
output: same as input

reduceMean

input: float32, float16
output: same as input

reduceMin

input: all supported data types
output: same as input

reduceProduct

input: float32, float16, int32, uint32, int64, uint64
output: same as input

reduceSum

input: float32, float16, int32, uint32, int64, uint64
output: same as input

reduceSumSquare

input: float32, float16, int32, uint32, int64, uint64
output: same as input

relu

input: float32, float16, int32, int8
output: same as input

resample2d

input: float32, float16
output: same as input

reshape

input: all supported data types
output: same as input

sigmoid

input: float32, float16
output: same as input

slice

input: all supported data types
output: same as input

softmax

input: float32, float16
output: same as input

softplus

input: float32, float16
output: same as input

softsign

input: float32, float16
output: same as input

split

input: all supported data types
outputs: same as input

input: float32, float16
output: same as input

transpose

input: all supported data types
output: same as input

triangular

input: all supported data types
output: same as input

where

condition: uint8
input: all supported data types
other: same as input
output: same as input

huningxin · 2023-11-23T03:17:29Z

@fdwr , if I read the DML doc correctly, L1, SUM_SQUARE, MULTIPLY, SUM reduce functions only support 32 and 64 bit integers.

https://learn.microsoft.com/en-us/windows/win32/api/directml/ns-directml-dml_reduce_operator_desc

fdwr · 2023-11-23T03:29:04Z

@fdwr , if I read the DML doc correctly, L1, SUM_SQUARE, MULTIPLY, SUM reduce functions only support 32 and 64 bit integers.

https://learn.microsoft.com/en-us/windows/win32/api/directml/ns-directml-dml_reduce_operator_desc

@huningxin : Correct.

Summary of the operand data type constraints for current WebNN operations

Is this the intersection of DML with XNNPack? (because FL3+ DML ABS supports int16 too, and FL4.1+ supports int8/int16/int32/int64).

inexorabletash · 2024-04-09T21:01:40Z

@huningxin in the list above, when an op is listed like this:

batchNormalization
input: float32, float16
mean: same as input

Should that "same as input" statement be interpreted as:
(1) mean's data type has the same restriction as input (i.e. float32 or float16); or
(2) the given mean operand's data type must exactly match the given input operantd's data type

I assume the latter (based on other examples) but wanted to confirm.

fdwr · 2024-04-09T22:25:59Z

Should that "same as input" statement be interpreted as: (1) mean's data type has the same restriction as input (i.e. float32 or float16); or (2) the given_mean_ operand's data type must exactly match the given input operantd's data type

@inexorabletash Yep, latter. If input is float32, then mean must also be float32.

inexorabletash · 2024-04-09T23:52:26Z

I have a local change for this. Lots of copy/paste. I'm wondering if we want to be table-driven; I'll bring that up in the eventual PR.

The table is missing:

softmax but it kicked off the issue as "float32" and "float16" only
softplus, softsign, tanh are mentioned in a comment as floating-point only
gelu
layerNormalization
gru, gruCell
lstm, lstmCell
argmin/argmax (presumably anything, like most of the reduction ops)
expand, gather, slice, transpose, triangular (presumably anything)
where (already in the spec, condition must be uint8, other other must match input)

wacky6 · 2024-04-10T00:48:58Z

Should mixed precision be allowed when the op involves accumulation?

For example, would this be acceptable for conv2d (likewise, matmul, reduceSum, ...)?

conv2d(
   /* input */ fp16,
  /* weight */ fp16,
    /* bias */ fp16,
) => fp32

I think requiring input, weight and bias to be the same type is reasonable (caller shouldn't add/multiple matrices of different types.

We don't need to block on mixed precision topics (because it's relaxing the constraints). A well documented type constraint table will be a big spec improvement. 😄

My read of DML doc is that conv(fp16) -> fp32 isn't permitted.

Does DML use a fp32 accumulator internally, then casts the result to fp16? Or is it fp16 accumulation all the way (might saturate the range and yield Infinity / NaN) ?

@fdwr

fdwr · 2024-04-10T01:04:55Z

My read of DML doc is that conv(fp16) -> fp32 isn't permitted.

@wacky6 You are correct - DML almost always requires the input and output data types to be the same (some notable exceptions are cast and quantize/dequantize operators).

Does DML use a fp32 accumulator internally, then casts the result to fp16? Or is it fp16 accumulation all the way (might saturate the range and yield Infinity / NaN)?

It depends on the op (reduction needs more intermediate precision than simple addition) and which flags are passed (e.g. DML_EXECUTION_FLAG_ALLOW_HALF_PRECISION_COMPUTATION which can be faster on some devices, at the cost of precision).

caller shouldn't add/multiple matrices of different types

I generally agree, as mixing any multiple different input types would explode the test matrix and increase backend complexity.

huningxin · 2024-04-10T12:42:41Z

@inexorabletash

The table is missing:

Added the following ops into the table.

softmax but it kicked off the issue as "float32" and "float16" only

+1

softplus, softsign, tanh are mentioned in a comment as floating-point only

+1

gelu

floating-point only

layerNormalization

floating-point only

gru, gruCell

floating-point only

lstm, lstmCell

floating-point only

argmin/argmax (presumably anything, like most of the reduction ops)

argmin/argmax's output is int64

expand, gather, slice, transpose, triangular (presumably anything)

gather's indices is uint32 or int64

where (already in the spec, condition must be uint8, other other must match input)

+1

philloooo · 2024-04-11T18:31:49Z

Hi,
I've compared it with what CoreML support, the main difference is:

For almost all the ops that support all types, CoreML support fp32, fp16, int32.
Argmax/argmin output is int32
Scalar parameters, like gather indices, conv2d groups type is int32.
For floating points computation, NPU only supports fp16. CPU & GPU supports both fp32 and fp16.
Overal Model input/output: fp16, fp32, int32. So if the last op of the model generates int8, it errors out.

If WebNN declares wider type set, I think we would need some way to feature detect(#463).

Introduce constraints for input operands, either directly (e.g. input's dataType can only be "float32" or "float16") or indirectly (e.g. weight's dataType must be the same as input's). Fixes webmachinelearning#283

* Specify the operand data type constraints of operations Introduce constraints for input operands, either directly (e.g. input's dataType can only be "float32" or "float16") or indirectly (e.g. weight's dataType must be the same as input's). Fixes #283 * gruCell: bundle hiddenState in with other type validations * Identity should accept all types * Add reduceMean restriction * Update gemm to check c data type too --------- Co-authored-by: Dwayne Robinson <dwayner@microsoft.com>

fdwr · 2024-05-02T06:00:30Z

For almost all the ops that support all types, CoreML support fp32, fp64, int32.

@philloooo The float64 type is interesting because in contrast, DML doesn't support float64 for any math operators (only for bitwise data movement operators like transposing and gather and bitwise and/or, plus cast at the edges of partitions) but DML does support int64, whereas CoreML supports float64 but doesn't support int64 for any operators. 🤹

If WebNN declares wider type set, I think we would need some way to feature detect(#463).

Agreed. Differences across implementations will be inevitable (similarly, WebGPU doesn't support all GPUTextureFormats, like astc-4x4-unorm across all GPU's), but if they can be testable up-front, that avoids wastefully loading a model only to fail later, and if any differences occur broadly (e.g. entire data types missing), that's easier to deal with than sporadic differences (certain operators missing) because you could easily generate two separate models with different data types, but it would be tougher to generate multiple models depending on various permutations of spotty operator support. The MLContext is probably the right place to ask that before you even make a graphBuilder. e.g.

interface MLContext {
    Promise<MLComputeResult> compute(MLGraph graph, MLNamedArrayBufferViews inputs, MLNamedArrayBufferViews outputs);

+   boolean isTypeSupported(MLOperandDataType dataType);
};

philloooo · 2024-05-02T16:11:56Z

oops @fdwr sorry that was a typo, it's fp16 not fp64. Just edited. For almost all the ops that support all types, CoreML support fp32, fp16, int32.

inexorabletash · 2024-05-06T23:54:30Z

I know it's bad form to comment on a closed issue, but... the table lists this for ReLU and PReLU:

input: float32, float16, int32, int8

@philloooo points out that most other activations support only floats, and #283 (comment) says that prelu should only accept floats?

@huningxin - is this intentional?

huningxin · 2024-05-07T06:58:22Z

@inexorabletash

@huningxin - is this intentional?

Yes. The intention is having them to accept the signed values.

#283 (comment) says that prelu should only accept floats?

In Chromium CL review, we'd like to prototype prelu by staring from floating point data point in sake of simplicity.

huningxin · 2024-05-27T05:57:05Z

@fdwr , if I read the DML doc correctly, L1, SUM_SQUARE, MULTIPLY, SUM reduce functions only support 32 and 64 bit integers.

https://learn.microsoft.com/en-us/windows/win32/api/directml/ns-directml-dml_reduce_operator_desc

Added missing int64 and uint64 support of reduceL1, reduceSum, reduceProduct and reduceSumSquare into above table.

`reduceL1`, `reduceProduct`, `reduceSum` and `reduceSumSquare` support already 32-bit integers. 64-bit integers should also be supported. Fix webmachinelearning#283, webmachinelearning#694

`reduceL1`, `reduceProduct`, `reduceSum` and `reduceSumSquare` already support 32-bit integers. 64-bit integers should also be supported. Fix webmachinelearning#283, webmachinelearning#694

wacky6 mentioned this issue Aug 26, 2022

Need for type casting? #284

Closed

huningxin changed the title ~~Softmax should only support input of floating-point types~~ Specify the operand type constraints of operation Sep 21, 2022

anssiko added the question label Mar 3, 2023

huningxin changed the title ~~Specify the operand type constraints of operation~~ Specify the operand data type constraints of operation Jan 31, 2024

huningxin mentioned this issue Feb 2, 2024

Process: Add documentation for labels, current and proposed #533

Merged

3 tasks

huningxin mentioned this issue Feb 21, 2024

Synchronously validate input operands/activations #572

Closed

inexorabletash self-assigned this Apr 10, 2024

inexorabletash mentioned this issue Apr 17, 2024

Update 2024-04-18-wg-agenda.md webmachinelearning/meetings#22

Merged

inexorabletash mentioned this issue Apr 18, 2024

Specify the operand data type constraints of operations #646

Merged

a-sully mentioned this issue Apr 18, 2024

Introduce MLNumber for specifying numeric inputs of any type #647

Merged

fdwr closed this as completed in #646 Apr 26, 2024

philloooo mentioned this issue May 15, 2024

Allow checking whether operators/types are supported for a backend before creating a graph #463

Open

huningxin mentioned this issue May 27, 2024

Bugfix: Add missing 64-bit integers support for some reduction operators #695

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Specify the operand data type constraints of operation #283

Specify the operand data type constraints of operation #283

huningxin commented Aug 26, 2022 •

edited

Loading

huningxin commented Sep 21, 2022

huningxin commented Apr 11, 2023 •

edited

Loading

huningxin commented Apr 19, 2023 •

edited

Loading

huningxin commented Oct 25, 2023

fdwr commented Oct 25, 2023

wacky6 commented Oct 26, 2023

fdwr commented Oct 26, 2023

huningxin commented Nov 20, 2023 •

edited

Loading

huningxin commented Nov 23, 2023

fdwr commented Nov 23, 2023

inexorabletash commented Apr 9, 2024 •

edited

Loading

fdwr commented Apr 9, 2024 •

edited

Loading

inexorabletash commented Apr 9, 2024

wacky6 commented Apr 10, 2024

fdwr commented Apr 10, 2024

huningxin commented Apr 10, 2024

philloooo commented Apr 11, 2024 •

edited

Loading

fdwr commented May 2, 2024 •

edited

Loading

philloooo commented May 2, 2024

inexorabletash commented May 6, 2024

huningxin commented May 7, 2024

huningxin commented May 27, 2024

Specify the operand data type constraints of operation #283

Specify the operand data type constraints of operation #283

Comments

huningxin commented Aug 26, 2022 • edited Loading

huningxin commented Sep 21, 2022

huningxin commented Apr 11, 2023 • edited Loading

huningxin commented Apr 19, 2023 • edited Loading

huningxin commented Oct 25, 2023

fdwr commented Oct 25, 2023

wacky6 commented Oct 26, 2023

fdwr commented Oct 26, 2023

huningxin commented Nov 20, 2023 • edited Loading

[WIP] Summary of the operand data type constraints for current WebNN operations.

Element-wise binary operations

Element-wise unary operations

Pooling operations

Reduction operations

huningxin commented Nov 23, 2023

fdwr commented Nov 23, 2023

inexorabletash commented Apr 9, 2024 • edited Loading

fdwr commented Apr 9, 2024 • edited Loading

inexorabletash commented Apr 9, 2024

wacky6 commented Apr 10, 2024

fdwr commented Apr 10, 2024

huningxin commented Apr 10, 2024

philloooo commented Apr 11, 2024 • edited Loading

fdwr commented May 2, 2024 • edited Loading

philloooo commented May 2, 2024

inexorabletash commented May 6, 2024

huningxin commented May 7, 2024

huningxin commented May 27, 2024

huningxin commented Aug 26, 2022 •

edited

Loading

huningxin commented Apr 11, 2023 •

edited

Loading

huningxin commented Apr 19, 2023 •

edited

Loading

huningxin commented Nov 20, 2023 •

edited

Loading

inexorabletash commented Apr 9, 2024 •

edited

Loading

fdwr commented Apr 9, 2024 •

edited

Loading

philloooo commented Apr 11, 2024 •

edited

Loading

fdwr commented May 2, 2024 •

edited

Loading