cconvey commented on code in PR #12340: URL: https://github.com/apache/tvm/pull/12340#discussion_r948050528
########## tests/python/contrib/test_hexagon/test_fixed_point_conversion.py: ########## @@ -0,0 +1,58 @@ +# 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. + +import math +import struct +import numpy as np +import tvm.topi.hexagon.utils as utils + +""" +Test float to fixed-point conversion. We do it by constructing a numpy array with the +wide range of floating-point values. These values are converted into the +fixed-point value using topi.hexagon.utils.get_fixed_point_value. Then, these values are +converted back into float using scale_factor provided by the function. These converted +floating point values are then compared against the original values and an assertion is +raised if they happened to be outside of the expected tolerance. +""" + + +class TestFixedPointConversion: + def test_fixed_point_conversion(self): + # Construct array with wide range of values + fp1 = np.random.uniform(0.00001, 0.0002, size=(10)) + fp2 = np.random.uniform(0.001, 0.02, size=(10)) + fp3 = np.random.uniform(1, 20, size=(10)) + fp4 = np.random.uniform(900, 1000, size=(10)) + fp5 = np.random.uniform(1e9, 1e10, size=(10)) + fp6 = np.random.uniform(2.44885652993e38, 2.54885652993e38, size=(1)) + fp7 = np.random.uniform(1.46711479073e-34, 1.76098837843e-34, size=(1)) + float_arr = np.concatenate((fp1, fp2, fp3, fp4, fp5, fp6, fp7)) + for flp in float_arr: + fxp, rsh = utils.get_fixed_point_value(flp, "int16") + # Compute scale_factor using rsh (rsh is log2 of the scale_factor). While doing this, + # we use IEEE-754 floating-point representation since rsh can be negative or positive. + + scale = ((rsh + 127) & 0xFF) << 23 # Add bias (127) and position it into exponent bits + scale_i = struct.pack("I", scale) # Pack it as integer + scale_f = struct.unpack("f", scale_i) # Unpack as float + + converted_flp = fxp / scale_f[0] Review Comment: Would it make sense to move this logic into new function in `utils`, e.g. `get_floating_point_value(fxp:int, rsh:int, dtype="float16") -> float` ? I'm just thinking that the two conversion functions probably belong in the same place, even if one is currently used only for testing. ########## tests/python/contrib/test_hexagon/test_fixed_point_conversion.py: ########## @@ -0,0 +1,58 @@ +# 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. + +import math +import struct +import numpy as np +import tvm.topi.hexagon.utils as utils + +""" +Test float to fixed-point conversion. We do it by constructing a numpy array with the +wide range of floating-point values. These values are converted into the +fixed-point value using topi.hexagon.utils.get_fixed_point_value. Then, these values are +converted back into float using scale_factor provided by the function. These converted +floating point values are then compared against the original values and an assertion is +raised if they happened to be outside of the expected tolerance. +""" + + +class TestFixedPointConversion: + def test_fixed_point_conversion(self): + # Construct array with wide range of values + fp1 = np.random.uniform(0.00001, 0.0002, size=(10)) + fp2 = np.random.uniform(0.001, 0.02, size=(10)) + fp3 = np.random.uniform(1, 20, size=(10)) + fp4 = np.random.uniform(900, 1000, size=(10)) + fp5 = np.random.uniform(1e9, 1e10, size=(10)) + fp6 = np.random.uniform(2.44885652993e38, 2.54885652993e38, size=(1)) + fp7 = np.random.uniform(1.46711479073e-34, 1.76098837843e-34, size=(1)) Review Comment: I'm wondering if random draws are worth the effort / complexity here... - If the goal is just simple, sanity-checking unit tests, then I'm not sure we really need randomness. Especially if they lead to test-failures that can't be reproduced for the sake of debugging, due to the randomization. - If the goal is to check corner cases, I would think that's better done using specifically chosen values, e.g. - extreme value / special values for floating-point numbers - floating point values that, by inspection of the conversion algorithm, are likely to be critical ########## tests/python/contrib/test_hexagon/test_fixed_point_conversion.py: ########## @@ -0,0 +1,58 @@ +# 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. + +import math +import struct +import numpy as np +import tvm.topi.hexagon.utils as utils + +""" +Test float to fixed-point conversion. We do it by constructing a numpy array with the +wide range of floating-point values. These values are converted into the +fixed-point value using topi.hexagon.utils.get_fixed_point_value. Then, these values are +converted back into float using scale_factor provided by the function. These converted +floating point values are then compared against the original values and an assertion is +raised if they happened to be outside of the expected tolerance. +""" + + +class TestFixedPointConversion: + def test_fixed_point_conversion(self): + # Construct array with wide range of values + fp1 = np.random.uniform(0.00001, 0.0002, size=(10)) + fp2 = np.random.uniform(0.001, 0.02, size=(10)) + fp3 = np.random.uniform(1, 20, size=(10)) + fp4 = np.random.uniform(900, 1000, size=(10)) + fp5 = np.random.uniform(1e9, 1e10, size=(10)) + fp6 = np.random.uniform(2.44885652993e38, 2.54885652993e38, size=(1)) + fp7 = np.random.uniform(1.46711479073e-34, 1.76098837843e-34, size=(1)) Review Comment: These numbers seem pretty specific. It would be nice to have a comment indicating what (if anything) they correspond to. ########## python/tvm/topi/hexagon/utils.py: ########## @@ -150,4 +157,126 @@ def get_layout_transform_fn(layout): return nc_2048_2d if layout == "nhwc-8h8w32c-2d": return nhwc_8h8w32c_2d + if layout == "n11c-2048c-2d": + return n11c_2048c_2d raise RuntimeError(f"Unexpected layout '{layout}'") + + +def get_fixed_point_value(flp: float, dtype: str = "int16"): + """ + Return fixed-point value and the corresponding log2 of the scale factor used to compute + this value. + + Parameters + ---------- + flp : float + Floating-point value to be converted + dtype : str + Type of the resulting fixed-point value. By default, it's set to "int16" + + Returns + ------- + fixed_point_value : int + Fixed-point value for the given floating-point value + exp_scale_factor : int + log2 of the scale factor + + Convert floating-point value into fixed-point number. This is done by + multiplying the value by a scaling factor and then rounding it to the nearest + integer value. + + As per IEEE-754 standard, a floating-point value can be represented as follows + [see: https://en.wikipedia.org/wiki/IEEE_754-1985]: + (-1)^S * M * 2^(E-Bias) + + Here, + * S is the signed bit (0 or 1). + * M is the mantissa. It's composed of an implicit 1 for the normalized floating-point + values or 0 for the denormalized values, and the fraction part. This ensures that + mantissa is always within [0, 2) range. Please note that this function doesn't + handle denormalized values. + * E is the exponent. + + In single precision, 23 bits are used to represent the fraction part of + the mantissa (and therefore, '23' shows up in one of the computations below) and + 8 bits are used for the exponent. Since exponent field needs to reperesent both + positive and negative values, a bias (127 for single precision) is added to the actual + value. Therefore, to compute the actual exponent, 127 must be subtracted from the stored + value. + + As mentioned above, to find the corresponding fixed-point number, we multiply the + value with a scaling factor and then round it to the nearest integer. The scaling factor + is chosen to be a power for 2 and it's the largest value that can be safely multiplied + to the floating-point value, without causing the resulting value to overflow the range + of the integer type used to represent the fixed-point value. + + So, if we assume the scaling factor to be 2^x, the resulting fixed-point value will be: + round((-1)^S * (M) * 2^(E-Bias) * 2^x) + + This can be simplified to: + round((-1)^S * M * 2^(E-Bias+x) + + Now, if 'int16' is used for fixed-point value, then it has to be >= -(2 * 2^14) + and <= (2 * 2^14) - 1. Since M (Mantissa) is always < 2, in order for the fixed-point value + to be within this range, 2^(E - Bias + x) must be <= 2^14 - 1. + And, if we ignore -1, (E - Bias + x) should be <= 14. Note: if mantissa gets too close to 2, + this will cause the resulting value to go out of range and require it to be saturated. + In the following implementation, we perform range check and adjust the scale to avoid + saturation. + For most cases, 2^x, where x = 14 - (E - Bias) or 14 - (E - 127) for single precision, is the + best scaling factor for 'int16' type that can be used to convert the floating-point value to + fixed-point with the least amount of precision loss. + + Additonal notes on various floating-point values: + ------------------------------------------------ + 1) Denormalized values: Can't be represented as fixed-point - causes assertion failure + 2) NaN and INF: assertion failure + """ + + def within_range(val, dtype): + if dtype == "int16": + return -32768 <= val <= 32767 + raise RuntimeError(f"Unsupported dtype, {dtype}'") + + # Make sure that 'flp' isn't NaN or infinity + if math.isnan(flp) or math.isinf(flp): + raise RuntimeError("Can not handle NaN or INF") Review Comment: Nitpick: Sometimes comments like this indicate a temporary limitation of the function, that could be addressed in a later version. But IIUC, the fixed-point format we're dealing with here is simply incapable of expressing those two concepts. It might be helpful to use an error message that's clearer about this. ########## python/tvm/topi/hexagon/utils.py: ########## @@ -150,4 +157,126 @@ def get_layout_transform_fn(layout): return nc_2048_2d if layout == "nhwc-8h8w32c-2d": return nhwc_8h8w32c_2d + if layout == "n11c-2048c-2d": + return n11c_2048c_2d raise RuntimeError(f"Unexpected layout '{layout}'") + + +def get_fixed_point_value(flp: float, dtype: str = "int16"): + """ + Return fixed-point value and the corresponding log2 of the scale factor used to compute + this value. + + Parameters + ---------- + flp : float + Floating-point value to be converted + dtype : str + Type of the resulting fixed-point value. By default, it's set to "int16" + + Returns + ------- + fixed_point_value : int + Fixed-point value for the given floating-point value + exp_scale_factor : int + log2 of the scale factor + + Convert floating-point value into fixed-point number. This is done by + multiplying the value by a scaling factor and then rounding it to the nearest + integer value. + + As per IEEE-754 standard, a floating-point value can be represented as follows + [see: https://en.wikipedia.org/wiki/IEEE_754-1985]: + (-1)^S * M * 2^(E-Bias) + + Here, + * S is the signed bit (0 or 1). + * M is the mantissa. It's composed of an implicit 1 for the normalized floating-point + values or 0 for the denormalized values, and the fraction part. This ensures that + mantissa is always within [0, 2) range. Please note that this function doesn't + handle denormalized values. + * E is the exponent. + + In single precision, 23 bits are used to represent the fraction part of + the mantissa (and therefore, '23' shows up in one of the computations below) and + 8 bits are used for the exponent. Since exponent field needs to reperesent both + positive and negative values, a bias (127 for single precision) is added to the actual + value. Therefore, to compute the actual exponent, 127 must be subtracted from the stored + value. + + As mentioned above, to find the corresponding fixed-point number, we multiply the + value with a scaling factor and then round it to the nearest integer. The scaling factor + is chosen to be a power for 2 and it's the largest value that can be safely multiplied + to the floating-point value, without causing the resulting value to overflow the range + of the integer type used to represent the fixed-point value. + + So, if we assume the scaling factor to be 2^x, the resulting fixed-point value will be: + round((-1)^S * (M) * 2^(E-Bias) * 2^x) + + This can be simplified to: + round((-1)^S * M * 2^(E-Bias+x) + + Now, if 'int16' is used for fixed-point value, then it has to be >= -(2 * 2^14) + and <= (2 * 2^14) - 1. Since M (Mantissa) is always < 2, in order for the fixed-point value + to be within this range, 2^(E - Bias + x) must be <= 2^14 - 1. + And, if we ignore -1, (E - Bias + x) should be <= 14. Note: if mantissa gets too close to 2, + this will cause the resulting value to go out of range and require it to be saturated. + In the following implementation, we perform range check and adjust the scale to avoid + saturation. + For most cases, 2^x, where x = 14 - (E - Bias) or 14 - (E - 127) for single precision, is the + best scaling factor for 'int16' type that can be used to convert the floating-point value to + fixed-point with the least amount of precision loss. + + Additonal notes on various floating-point values: + ------------------------------------------------ + 1) Denormalized values: Can't be represented as fixed-point - causes assertion failure + 2) NaN and INF: assertion failure + """ + + def within_range(val, dtype): + if dtype == "int16": + return -32768 <= val <= 32767 + raise RuntimeError(f"Unsupported dtype, {dtype}'") + + # Make sure that 'flp' isn't NaN or infinity + if math.isnan(flp) or math.isinf(flp): + raise RuntimeError("Can not handle NaN or INF") + + flp_f = struct.pack("f", flp) + flp_i = struct.unpack("I", flp_f) + exp_stored_value = (flp_i[0] >> 23) & 0xFF + + if exp_stored_value == 0: + raise RuntimeError("Can not handle denormalized values") Review Comment: (This is somewhat redundant to a comment I left regarding the function's docstring, above.) It would be nice to have a comment regarding why denormalized values aren't handled. E.g.: - they're always indistinguishable from 0 in the resulting fixed-point representation, or - we don't need to support them yet, so we're just not dealing with them for now, or - (something else) ########## python/tvm/topi/hexagon/utils.py: ########## @@ -150,4 +157,126 @@ def get_layout_transform_fn(layout): return nc_2048_2d if layout == "nhwc-8h8w32c-2d": return nhwc_8h8w32c_2d + if layout == "n11c-2048c-2d": + return n11c_2048c_2d raise RuntimeError(f"Unexpected layout '{layout}'") + + +def get_fixed_point_value(flp: float, dtype: str = "int16"): Review Comment: ```suggestion def get_fixed_point_value(flp: float, dtype: str = "int16") -> Tuple[int, int]: ``` ########## python/tvm/topi/hexagon/qnn/avg_pool2d.py: ########## @@ -0,0 +1,205 @@ +# 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=invalid-name, unused-variable, unused-argument, too-many-locals + +""" Compute and schedule for quantized avg_pool2d op + +Please note the following assumptions made by the implementation: + +1) The input must be padded in advance to account for 'padding'. In addition, + both input and output must be padded as per the physical buffer layout. +2) The current implementation assumes 'count_include_pad' to be 'True'. It can be + modified to support 'False' case but the element count for the pooling window + must be pre-computed and provided as an input to reduce the run-time overhead. +3) 'padding' is ignored. It must be handled outside of the sliced op. +4) Please note that this implementation will not work if the output includes any + physical layout related padding as it can result into out-of-bound access + for the input. +""" + +from tvm import te +from tvm import tir +from ..utils import get_layout_transform_fn, get_fixed_point_value + + +def validate_out_shape(out_shape: list, in_shape: list, kernel: list, stride: list, dilation: list): + """Validate output shape""" + _, oh, ow, _ = out_shape + _, ih, iw, _ = in_shape + kh, kw = kernel + sh, sw = stride + dh, dw = dilation + if ih < (oh - 1) * sh + dh * (kh - 1) + 1: + raise RuntimeError("Output height is too large") + if iw < (ow - 1) * sw + dw * (kw - 1) + 1: + raise RuntimeError("Output width is too large") + + +def saturate(x: te.Tensor, dtype: str): + """Saturate value for the specified data type""" + return te.max(te.min_value(dtype), te.min(x, te.max_value(dtype))) Review Comment: When I looked at several of the Hexagon `.so` files produced by this PR's unit tests, I didn't see any indication that Hexagon's `saturate` or `:sat` instructions were being used. This isn't a critique of the PR; I'm just mentioning it as a point of interest for future work. ########## python/tvm/topi/hexagon/utils.py: ########## @@ -150,4 +157,126 @@ def get_layout_transform_fn(layout): return nc_2048_2d if layout == "nhwc-8h8w32c-2d": return nhwc_8h8w32c_2d + if layout == "n11c-2048c-2d": + return n11c_2048c_2d raise RuntimeError(f"Unexpected layout '{layout}'") + + +def get_fixed_point_value(flp: float, dtype: str = "int16"): + """ + Return fixed-point value and the corresponding log2 of the scale factor used to compute + this value. + + Parameters + ---------- + flp : float + Floating-point value to be converted + dtype : str + Type of the resulting fixed-point value. By default, it's set to "int16" + + Returns + ------- + fixed_point_value : int + Fixed-point value for the given floating-point value + exp_scale_factor : int + log2 of the scale factor + + Convert floating-point value into fixed-point number. This is done by + multiplying the value by a scaling factor and then rounding it to the nearest + integer value. + + As per IEEE-754 standard, a floating-point value can be represented as follows + [see: https://en.wikipedia.org/wiki/IEEE_754-1985]: + (-1)^S * M * 2^(E-Bias) + + Here, + * S is the signed bit (0 or 1). + * M is the mantissa. It's composed of an implicit 1 for the normalized floating-point + values or 0 for the denormalized values, and the fraction part. This ensures that + mantissa is always within [0, 2) range. Please note that this function doesn't + handle denormalized values. + * E is the exponent. + + In single precision, 23 bits are used to represent the fraction part of + the mantissa (and therefore, '23' shows up in one of the computations below) and + 8 bits are used for the exponent. Since exponent field needs to reperesent both + positive and negative values, a bias (127 for single precision) is added to the actual + value. Therefore, to compute the actual exponent, 127 must be subtracted from the stored + value. + + As mentioned above, to find the corresponding fixed-point number, we multiply the + value with a scaling factor and then round it to the nearest integer. The scaling factor + is chosen to be a power for 2 and it's the largest value that can be safely multiplied + to the floating-point value, without causing the resulting value to overflow the range + of the integer type used to represent the fixed-point value. + + So, if we assume the scaling factor to be 2^x, the resulting fixed-point value will be: + round((-1)^S * (M) * 2^(E-Bias) * 2^x) + + This can be simplified to: + round((-1)^S * M * 2^(E-Bias+x) + + Now, if 'int16' is used for fixed-point value, then it has to be >= -(2 * 2^14) + and <= (2 * 2^14) - 1. Since M (Mantissa) is always < 2, in order for the fixed-point value + to be within this range, 2^(E - Bias + x) must be <= 2^14 - 1. + And, if we ignore -1, (E - Bias + x) should be <= 14. Note: if mantissa gets too close to 2, + this will cause the resulting value to go out of range and require it to be saturated. + In the following implementation, we perform range check and adjust the scale to avoid + saturation. + For most cases, 2^x, where x = 14 - (E - Bias) or 14 - (E - 127) for single precision, is the + best scaling factor for 'int16' type that can be used to convert the floating-point value to + fixed-point with the least amount of precision loss. + + Additonal notes on various floating-point values: + ------------------------------------------------ + 1) Denormalized values: Can't be represented as fixed-point - causes assertion failure Review Comment: I'm confused by the claim that denormal values can't be expressed as fixed-point. My understanding is that IEEE-754 denormalized values are simply a special way of encoding values that are much closer to 0 than normalized float16 values can express. I don't understand why that's fundamentally inexpressable as fixed-point. Are we assuming some additional unstated limitations on our fixedpoint representation? E.g., the range of values that we're willing to let `rsh` take on? -- 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-unsubscr...@tvm.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org