Update ops-related pbtxt files.

PiperOrigin-RevId: 173738765

tensorflow/core/ops/ops.pbtxt

@@ -7049,7 +7049,7 @@ op {
     }
   }
   summary: "Dequantize the \'input\' tensor into a float Tensor."
-  description: "[min_range, max_range] are scalar floats that specify the range for\nthe \'input\' data. The \'mode\' attribute controls exactly which calculations are\nused to convert the float values to their quantized equivalents.\n\nIn \'MIN_COMBINED\' mode, each value of the tensor will undergo the following:\n\n```\nif T == qint8, in[i] += (range(T) + 1)/ 2.0\nout[i] = min_range + (in[i]* (max_range - min_range) / range(T))\n```\nhere `range(T) = numeric_limits<T>::max() - numeric_limits<T>::min()`\n\n*MIN_COMBINED Mode Example*\n\nIf the input comes from a QuantizedRelu6, the output type is\nquint8 (range of 0-255) but the possible range of QuantizedRelu6 is\n0-6.  The min_range and max_range values are therefore 0.0 and 6.0.\nDequantize on quint8 will take each value, cast to float, and multiply\nby 6 / 255.\nNote that if quantizedtype is qint8, the operation will additionally add\neach value by 128 prior to casting.\n\nIf the mode is \'MIN_FIRST\', then this approach is used:\n\n```c++\nnumber_of_steps = 1 << (# of bits in T)\nrange_adjust = number_of_steps / (number_of_steps - 1)\nrange = (range_max - range_min) * range_adjust\nrange_scale = range / number_of_steps\nconst double offset_input = static_cast<double>(input) - lowest_quantized;\nresult = range_min + ((input - numeric_limits<T>::min()) * range_scale)\n```\n\n*SCALED mode Example*\n\n`SCALED` mode matches the quantization approach used in\n`QuantizeAndDequantize{V2|V3}`.\n\nIf the mode is `SCALED`, we do not use the full range of the output type,\nchoosing to elide the lowest possible value for symmetry (e.g., output range is\n-127 to 127, not -128 to 127 for signed 8 bit quantization), so that 0.0 maps to\n0.\n\nWe first find the range of values in our tensor. The\nrange we use is always centered on 0, so we find m such that\n```c++\n  m = max(abs(input_min), abs(input_max))\n```\n\nOur input tensor range is then `[-m, m]`.\n\nNext, we choose our fixed-point quantization buckets, `[min_fixed, max_fixed]`.\nIf T is signed, this is\n```\n  num_bits = sizeof(T) * 8\n  [min_fixed, max_fixed] =\n      [-(1 << (num_bits - 1) - 1), (1 << (num_bits - 1)) - 1]\n```\n\nOtherwise, if T is unsigned, the fixed-point range is\n```\n  [min_fixed, max_fixed] = [0, (1 << num_bits) - 1]\n```\n\nFrom this we compute our scaling factor, s:\n```c++\n  s = (2 * m) / (max_fixed - min_fixed)\n```\n\nNow we can dequantize the elements of our tensor:\n```c++\nresult = input * s\n```"
+  description: "[min_range, max_range] are scalar floats that specify the range for\nthe \'input\' data. The \'mode\' attribute controls exactly which calculations are\nused to convert the float values to their quantized equivalents.\n\nIn \'MIN_COMBINED\' mode, each value of the tensor will undergo the following:\n\n```\nif T == qint8, in[i] += (range(T) + 1)/ 2.0\nout[i] = min_range + (in[i]* (max_range - min_range) / range(T))\n```\nhere `range(T) = numeric_limits<T>::max() - numeric_limits<T>::min()`\n\n*MIN_COMBINED Mode Example*\n\nIf the input comes from a QuantizedRelu6, the output type is\nquint8 (range of 0-255) but the possible range of QuantizedRelu6 is\n0-6.  The min_range and max_range values are therefore 0.0 and 6.0.\nDequantize on quint8 will take each value, cast to float, and multiply\nby 6 / 255.\nNote that if quantizedtype is qint8, the operation will additionally add\neach value by 128 prior to casting.\n\nIf the mode is \'MIN_FIRST\', then this approach is used:\n\n```c++\nnum_discrete_values = 1 << (# of bits in T)\nrange_adjust = num_discrete_values / (num_discrete_values - 1)\nrange = (range_max - range_min) * range_adjust\nrange_scale = range / num_discrete_values\nconst double offset_input = static_cast<double>(input) - lowest_quantized;\nresult = range_min + ((input - numeric_limits<T>::min()) * range_scale)\n```\n\n*SCALED mode Example*\n\n`SCALED` mode matches the quantization approach used in\n`QuantizeAndDequantize{V2|V3}`.\n\nIf the mode is `SCALED`, we do not use the full range of the output type,\nchoosing to elide the lowest possible value for symmetry (e.g., output range is\n-127 to 127, not -128 to 127 for signed 8 bit quantization), so that 0.0 maps to\n0.\n\nWe first find the range of values in our tensor. The\nrange we use is always centered on 0, so we find m such that\n```c++\n  m = max(abs(input_min), abs(input_max))\n```\n\nOur input tensor range is then `[-m, m]`.\n\nNext, we choose our fixed-point quantization buckets, `[min_fixed, max_fixed]`.\nIf T is signed, this is\n```\n  num_bits = sizeof(T) * 8\n  [min_fixed, max_fixed] =\n      [-(1 << (num_bits - 1) - 1), (1 << (num_bits - 1)) - 1]\n```\n\nOtherwise, if T is unsigned, the fixed-point range is\n```\n  [min_fixed, max_fixed] = [0, (1 << num_bits) - 1]\n```\n\nFrom this we compute our scaling factor, s:\n```c++\n  s = (2 * m) / (max_fixed - min_fixed)\n```\n\nNow we can dequantize the elements of our tensor:\n```c++\nresult = input * s\n```"
 }
 op {
   name: "DeserializeIterator"
@@ -17500,7 +17500,7 @@ op {
     }
   }
   summary: "Quantize the \'input\' tensor of type float to \'output\' tensor of type \'T\'."
-  description: "[min_range, max_range] are scalar floats that specify the range for\nthe \'input\' data. The \'mode\' attribute controls exactly which calculations are\nused to convert the float values to their quantized equivalents.  The\n\'round_mode\' attribute controls which rounding tie-breaking algorithm is used\nwhen rounding float values to their quantized equivalents.\n\nIn \'MIN_COMBINED\' mode, each value of the tensor will undergo the following:\n\n```\nout[i] = (in[i] - min_range) * range(T) / (max_range - min_range)\nif T == qint8, out[i] -= (range(T) + 1) / 2.0\n```\nhere `range(T) = numeric_limits<T>::max() - numeric_limits<T>::min()`\n\n*MIN_COMBINED Mode Example*\n\nAssume the input is type float and has a possible range of [0.0, 6.0] and the\noutput type is quint8 ([0, 255]). The min_range and max_range values should be\nspecified as 0.0 and 6.0. Quantizing from float to quint8 will multiply each\nvalue of the input by 255/6 and cast to quint8.\n\nIf the output type was qint8 ([-128, 127]), the operation will additionally\nsubtract each value by 128 prior to casting, so that the range of values aligns\nwith the range of qint8.\n\nIf the mode is \'MIN_FIRST\', then this approach is used:\n\n```\nnumber_of_steps = 1 << (# of bits in T)\nrange_adjust = number_of_steps / (number_of_steps - 1)\nrange = (range_max - range_min) * range_adjust\nrange_scale = number_of_steps / range\nquantized = round(input * range_scale) - round(range_min * range_scale) +\n  numeric_limits<T>::min()\nquantized = max(quantized, numeric_limits<T>::min())\nquantized = min(quantized, numeric_limits<T>::max())\n```\n\nThe biggest difference between this and MIN_COMBINED is that the minimum range\nis rounded first, before it\'s subtracted from the rounded value. With\nMIN_COMBINED, a small bias is introduced where repeated iterations of quantizing\nand dequantizing will introduce a larger and larger error.\n\n*SCALED mode Example*\n\n`SCALED` mode matches the quantization approach used in\n`QuantizeAndDequantize{V2|V3}`.\n\nIf the mode is `SCALED`, we do not use the full range of the output type,\nchoosing to elide the lowest possible value for symmetry (e.g., output range is\n-127 to 127, not -128 to 127 for signed 8 bit quantization), so that 0.0 maps to\n0.\n\nWe first find the range of values in our tensor. The\nrange we use is always centered on 0, so we find m such that\n```c++\n  m = max(abs(input_min), abs(input_max))\n```\n\nOur input tensor range is then `[-m, m]`.\n\nNext, we choose our fixed-point quantization buckets, `[min_fixed, max_fixed]`.\nIf T is signed, this is\n```\n  num_bits = sizeof(T) * 8\n  [min_fixed, max_fixed] =\n      [-(1 << (num_bits - 1) - 1), (1 << (num_bits - 1)) - 1]\n```\n\nOtherwise, if T is unsigned, the fixed-point range is\n```\n  [min_fixed, max_fixed] = [0, (1 << num_bits) - 1]\n```\n\nFrom this we compute our scaling factor, s:\n```c++\n  s = (max_fixed - min_fixed) / (2 * m)\n```\n\nNow we can quantize the elements of our tensor:\n```c++\nresult = round(input * s)\n```\n\nOne thing to watch out for is that the operator may choose to adjust the\nrequested minimum and maximum values slightly during the quantization process,\nso you should always use the output ports as the range for further calculations.\nFor example, if the requested minimum and maximum values are close to equal,\nthey will be separated by a small epsilon value to prevent ill-formed quantized\nbuffers from being created. Otherwise, you can end up with buffers where all the\nquantized values map to the same float value, which causes problems for\noperations that have to perform further calculations on them."
+  description: "[min_range, max_range] are scalar floats that specify the range for\nthe \'input\' data. The \'mode\' attribute controls exactly which calculations are\nused to convert the float values to their quantized equivalents.  The\n\'round_mode\' attribute controls which rounding tie-breaking algorithm is used\nwhen rounding float values to their quantized equivalents.\n\nIn \'MIN_COMBINED\' mode, each value of the tensor will undergo the following:\n\n```\nout[i] = (in[i] - min_range) * range(T) / (max_range - min_range)\nif T == qint8, out[i] -= (range(T) + 1) / 2.0\n```\nhere `range(T) = numeric_limits<T>::max() - numeric_limits<T>::min()`\n\n*MIN_COMBINED Mode Example*\n\nAssume the input is type float and has a possible range of [0.0, 6.0] and the\noutput type is quint8 ([0, 255]). The min_range and max_range values should be\nspecified as 0.0 and 6.0. Quantizing from float to quint8 will multiply each\nvalue of the input by 255/6 and cast to quint8.\n\nIf the output type was qint8 ([-128, 127]), the operation will additionally\nsubtract each value by 128 prior to casting, so that the range of values aligns\nwith the range of qint8.\n\nIf the mode is \'MIN_FIRST\', then this approach is used:\n\n```\nnum_discrete_values = 1 << (# of bits in T)\nrange_adjust = num_discrete_values / (num_discrete_values - 1)\nrange = (range_max - range_min) * range_adjust\nrange_scale = num_discrete_values / range\nquantized = round(input * range_scale) - round(range_min * range_scale) +\n  numeric_limits<T>::min()\nquantized = max(quantized, numeric_limits<T>::min())\nquantized = min(quantized, numeric_limits<T>::max())\n```\n\nThe biggest difference between this and MIN_COMBINED is that the minimum range\nis rounded first, before it\'s subtracted from the rounded value. With\nMIN_COMBINED, a small bias is introduced where repeated iterations of quantizing\nand dequantizing will introduce a larger and larger error.\n\n*SCALED mode Example*\n\n`SCALED` mode matches the quantization approach used in\n`QuantizeAndDequantize{V2|V3}`.\n\nIf the mode is `SCALED`, we do not use the full range of the output type,\nchoosing to elide the lowest possible value for symmetry (e.g., output range is\n-127 to 127, not -128 to 127 for signed 8 bit quantization), so that 0.0 maps to\n0.\n\nWe first find the range of values in our tensor. The\nrange we use is always centered on 0, so we find m such that\n```c++\n  m = max(abs(input_min), abs(input_max))\n```\n\nOur input tensor range is then `[-m, m]`.\n\nNext, we choose our fixed-point quantization buckets, `[min_fixed, max_fixed]`.\nIf T is signed, this is\n```\n  num_bits = sizeof(T) * 8\n  [min_fixed, max_fixed] =\n      [-(1 << (num_bits - 1) - 1), (1 << (num_bits - 1)) - 1]\n```\n\nOtherwise, if T is unsigned, the fixed-point range is\n```\n  [min_fixed, max_fixed] = [0, (1 << num_bits) - 1]\n```\n\nFrom this we compute our scaling factor, s:\n```c++\n  s = (max_fixed - min_fixed) / (2 * m)\n```\n\nNow we can quantize the elements of our tensor:\n```c++\nresult = round(input * s)\n```\n\nOne thing to watch out for is that the operator may choose to adjust the\nrequested minimum and maximum values slightly during the quantization process,\nso you should always use the output ports as the range for further calculations.\nFor example, if the requested minimum and maximum values are close to equal,\nthey will be separated by a small epsilon value to prevent ill-formed quantized\nbuffers from being created. Otherwise, you can end up with buffers where all the\nquantized values map to the same float value, which causes problems for\noperations that have to perform further calculations on them."
 }
 op {
   name: "QuantizedAdd"