[mlir][tosa] Align operator descriptions with the spec #128046
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Jerry-Ge
requested review from
GeorgeARM and
eric-k256
and removed request for
GeorgeARM
|
@llvm/pr-subscribers-mlir Author: Jerry-Ge (Jerry-Ge) ChangesThis commit aligns the operator descriptions to the v1.0 spec (https://www.mlplatform.org/tosa/tosa_spec.html) Co-authored by:
Full diff: https://github.com/llvm/llvm-project/pull/128046.diff 1 Files Affected:
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
index 7cdf79f4dc59d..695a8b4700090 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
@@ -37,7 +37,8 @@ def Tosa_ArgMaxOp : Tosa_InferShapedTypeOp<"argmax"> {
let description = [{
This returns the index with the largest value across the given axis of the
- input tensor.
+ input tensor. If multiple locations have equal values, returns the first
+ match along the search axis.
}];
let arguments = (ins
@@ -98,7 +99,8 @@ def Tosa_Conv2DOp : Tosa_ConvOp<"conv2d"> {
let description = [{
Performs a 2D convolution over the given tensor input, using the weight
- tensor.
+ tensor. Implementations may choose to skip calculation of multiplies in
+ the padding area.
}];
let arguments = (ins
@@ -129,7 +131,8 @@ def Tosa_Conv3DOp : Tosa_ConvOp<"conv3d"> {
let summary = "3D Convolution operator";
let description = [{
- Performs a 3D convolution over the given input tensor.
+ Performs a 3D convolution over the given input tensor. Implementations
+ may choose to skip calculation of multiplies in the padding area.
}];
let arguments = (ins
@@ -161,7 +164,8 @@ def Tosa_DepthwiseConv2DOp : Tosa_ConvOp<"depthwise_conv2d"> {
let description = [{
Performs 2D convolutions separately over each channel of the given tensor
- input, using the weight tensor.
+ input, using the weight tensor. Implementations may choose to skip
+ calculation of multiplies in the padding area.
}];
let arguments = (ins
@@ -323,7 +327,8 @@ def Tosa_TransposeConv2DOp : Tosa_ConvOp<"transpose_conv2d"> {
let description = [{
Performs a 2D transposed convolution over the given tensor input, using the
- weights tensor.
+ weights tensor. Implementations may choose to skip calculation of multiplies
+ by zero at fractional input positions.
}];
let arguments = (ins
@@ -388,10 +393,14 @@ def Tosa_SigmoidOp : Tosa_ElementwiseUnaryOp<"sigmoid"> {
let summary = "Computes elementwise sigmoid of input.";
let description = [{
- Sigmoid function: output = 1 / (1 + exp(-input))
- For quantized integer data types, the TABLE operator should be used instead
- with the following definition. The sigmoid table has 513 entries each of
- 16-bit precision and covering the input range -16.0 to +16.0
+ Applies the sigmoid logistic function to each element of the input tensor:
+ $ sigmoid(x) = \frac{1}{1 + e^{-x}} $.
+
+ For quantized integer data types, the TABLE operator should be used instead.
+ Each implementation may choose an appropriate TABLE given the scale and zero
+ point of the input data. Eight or sixteen bit precision tables may be used
+ based on the input tensor to the sigmoid function. The sigmoid table has 513
+ entries each of 16-bit precision and covering the input range -16.0 to +16.0
in steps of 1/16.
}];
@@ -411,10 +420,14 @@ def Tosa_TanhOp : Tosa_ElementwiseUnaryOp<"tanh"> {
let summary = "Computes elementwise hyperbolic tangent of input";
let description = [{
- Parameterized hyperbolic tangent.
- For quantized integer data types, the TABLE operator should be used instead
- with the following definition. The tanh_table has 513 entries each of
- 16-bit precision and covering the input range -8.0 to +8.0 in steps of 1/32.
+ Parameterized hyperbolic tangent: $ tanh(x) = \frac{1 - e^{-2x}}{1 + e^{-2x}} $.
+
+ For quantized integer data types, the TABLE operator should be used instead.
+ Each implementation may choose an appropriate TABLE given the scale and zero
+ point of the input data. Eight or sixteen bit precision tables may be used
+ based on the input tensor to the tanh function. The tanh_table has 513
+ entries each of 16-bit precision and covering the input range -8.0 to +8.0
+ in steps of 1/32.
}];
let arguments = (ins
@@ -433,10 +446,10 @@ def Tosa_ErfOp : Tosa_ElementwiseUnaryOp<"erf"> {
let summary = "Computes gauss error function of input";
let description = [{
- Gauss error function: $ erf(x) = \frac{2}{\sqrt(\pi)} \int_{0}^{x} e^{-t^2} \ dt $
+ Gauss error function: $ erf(x) = \frac{2}{\sqrt{\pi}} \int_{0}^{x} e^{-t^2} dt $
For quantized integer data types, the TABLE operator should be used instead
- with the following definition. The erf_table has 513 entries each of
- 16-bit/8-bit precision and covering the input range -4.0 to +4.0 in steps of 1/64.
+ with the following definition. The ERF table has 513 entries each of
+ 16-bit precision and covering the input range -4.0 to +4.0 in steps of 1/64.
}];
let arguments = (ins
@@ -466,7 +479,7 @@ def Tosa_AddOp : Tosa_ElementwiseOp<"add", [
let description = [{
Elementwise addition of input1 and input2. Axis of size 1 will be broadcast,
- as necessary.
+ as necessary. Rank of input tensors must match.
Example:
@@ -500,7 +513,8 @@ def Tosa_ArithmeticRightShiftOp : Tosa_ElementwiseOp<"arithmetic_right_shift",
let description = [{
Elementwise arithmetic right shift of input1 by the amount specified in
- input2. Axis of size 1 will be broadcast, as necessary.
+ input2. Axis of size 1 will be broadcast, as necessary. Rank of input tensors
+ must match.
}];
let arguments = (ins
@@ -524,7 +538,7 @@ def Tosa_BitwiseAndOp : Tosa_ElementwiseOp<"bitwise_and", [
let description = [{
Elementwise bitwise AND of input1 and input2. Axis of size 1
- will be broadcast as necessary.
+ will be broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -547,7 +561,7 @@ def Tosa_BitwiseOrOp : Tosa_ElementwiseOp<"bitwise_or", [
let description = [{
Elementwise bitwise OR of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -570,7 +584,7 @@ def Tosa_BitwiseXorOp : Tosa_ElementwiseOp<"bitwise_xor", [
let description = [{
Elementwise bitwise XOR of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -590,8 +604,10 @@ def Tosa_IntDivOp : Tosa_ElementwiseOp<"int_div", [SameOperandsAndResultElementT
let summary = "Integer divide operator";
let description = [{
- Elementwise integer divide operator of input1 by input2. Axis of size 1
- will be broadcast, as necessary.
+ Elementwise integer divide operator of input1 by input2. The result of the divide
+ is truncated towards zero. Expected use is for operations on non-scaled integers.
+ Floating point divide should use RECIPROCAL and MUL. Quantized integer divide
+ should use TABLE (for 1/x) and MUL.
}];
let arguments = (ins
@@ -616,7 +632,7 @@ def Tosa_LogicalAndOp : Tosa_ElementwiseOp<"logical_and", [
let description = [{
Elementwise logical AND of input1 and input2. Axis of size 1 will be
- broadcast, as necessary.
+ broadcast, as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -638,7 +654,7 @@ def Tosa_LogicalLeftShiftOp : Tosa_ElementwiseOp<"logical_left_shift",
let description = [{
Elementwise left shift of input1 and input2. Axis of size 1 will be
- broadcast, as necessary.
+ broadcast, as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -660,7 +676,8 @@ def Tosa_LogicalRightShiftOp : Tosa_ElementwiseOp<"logical_right_shift",
let description = [{
Elementwise logical right shift of input1 by the amount specified in input2.
- Axis of size 1 will be broadcast, as necessary.
+ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must
+ match.
}];
let arguments = (ins
@@ -683,7 +700,7 @@ def Tosa_LogicalOrOp : Tosa_ElementwiseOp<"logical_or", [
let description = [{
Elementwise logical OR of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -705,8 +722,8 @@ def Tosa_LogicalXorOp : Tosa_ElementwiseOp<"logical_xor", [
let summary = "Returns the truth value of x XOR y element-wise.";
let description = [{
- Elementwise logical XOR of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ Elementwise logical XOR of input1 and input2. Axis of size 1 will be
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -729,7 +746,7 @@ def Tosa_MaximumOp : Tosa_ElementwiseOp<"maximum", [
let description = [{
Elementwise max of input1 and input2. Axis of size 1 will be broadcast, as
- necessary.
+ necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -753,7 +770,7 @@ def Tosa_MinimumOp : Tosa_ElementwiseOp<"minimum", [
let description = [{
Elementwise minimum of input1 and input2. Axis of size 1
- will be broadcast, as necessary.
+ will be broadcast, as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -784,8 +801,8 @@ def Tosa_MulOp : Tosa_Op<"mul", [
let description = [{
Elementwise multiplication (Hadamard product) of input1 and input2.
- Axis of size 1 will be broadcast, as necessary.
- i8/i16 input type can be promoted to i32 result type.
+ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must
+ match.
}];
let arguments = (ins
@@ -814,7 +831,8 @@ def Tosa_PowOp : Tosa_ElementwiseOp<"pow", [SameOperandsAndResultElementType]> {
let description = [{
Elementwise input1 raised to the power of input2.
- Axis of size 1 will be broadcast, as necessary.
+ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must
+ match.
}];
let arguments = (ins
@@ -835,7 +853,7 @@ def Tosa_SubOp : Tosa_ElementwiseOp<"sub", [SameOperandsAndResultElementType]> {
let description = [{
Elementwise subtraction of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -857,20 +875,20 @@ def Tosa_TableOp : Tosa_InferShapedTypeOp<"table"> {
let summary = "Table lookup op";
let description = [{
- Interpolated table lookup operation. Input values are scaled to create a
- fixed-point 9.7 value. The high 9 bits are used to index into the table.
- The fractional bits are used to interpolate based on the looked up value and
- the index+1 value in the table. The TABLE operator then returns a 16.7
- interpolated value. Note that there must be 513 values to handle the full
- range of inputs.
+ Table lookup operation. For int8_t TABLE operation, perform a 256 entry
+ table lookup returning an int8_t value. For int16_t tables, the int16_t
+ input is treated as a fixed-point 9.7 value. The most significant 9 bits
+ are used to index into the table. The fractional 7 bits are used to
+ interpolate based on table[index] and table[index+1]. For int16_t inputs,
+ the TABLE operator returns a 16.7 interpolated value in an int32_t. This
+ value can then be input to the RESCALE operator to scale to the required
+ output data type. Note that int16_t table has 513 values to handle
+ table[index+1] when index=511.
- The TABLE operator is expected to be used as follows:
- * A RESCALE node is expected before the TABLE operator to scale the input
- to a full int16_t range for the table lookup
- * If an int16_t result is required then follow the TABLE operator with a
- RESCALE with a right shift of 7
- * If an int8_t result is required then follow the TABLE operator with a
- RESCALE with a right shift of 15
+ An int16_t to int16_t table lookup can be constructed in TOSA as follows:
+ * Use the TABLE operator to produce a fixed point 16.7 interpolated result
+ * Use RESCALE (in_t=int32_t, out_t=int16_t, scale=1<<14, shift=21) to
+ scale the output to int16_t range (or alternate scale as required)
}];
let arguments = (ins
@@ -902,7 +920,7 @@ def Tosa_AbsOp : Tosa_ElementwiseUnaryOp<"abs"> {
let summary = "Elementwise abs op";
let description = [{
- Elementwise absolute value operation
+ Elementwise absolute value operation.
Example:
@@ -1026,7 +1044,7 @@ def Tosa_FloorOp : Tosa_ElementwiseUnaryOp<"floor"> {
let summary = "Elementwise floor op";
let description = [{
- Elementwise floor operation
+ Elementwise floor operation.
}];
let arguments = (ins
@@ -1085,7 +1103,7 @@ def Tosa_NegateOp : Tosa_ElementwiseUnaryOp<"negate"> {
let summary = "Elementwise negate op";
let description = [{
- Elementwise negation operation
+ Elementwise negation operation.
}];
let arguments = (ins
@@ -1543,9 +1561,9 @@ def Tosa_PadOp : Tosa_InferShapedTypeOp<"pad"> {
let summary = "Pads a tensor with value specified.";
let description = [{
- The `tosa.pad` operation pads a tensor along borders of each dimension with
- `pad_const` (defaults to zero), given a padding configuration `padding`
- specifying low and high values along the dimensions.
+ Pads a tensor along the borders of each dimension with a supplied value.
+ Returns a new tensor with the padding included. The pad_const value includes
+ the zero point if the tensor uses a zero point.
Example:
@@ -1705,7 +1723,9 @@ def Tosa_TransposeOp : Tosa_InferShapedTypeOp<"transpose",
let summary = "Transpose operator";
let description = [{
- Permutes the dimensions based on perm.
+ Permutes the dimensions of the input tensor input1 based on the perms
+ argument. Each value in the perms list must be a valid dimension of the
+ input tensor and may not be repeated.
}];
let arguments = (ins
@@ -1738,8 +1758,10 @@ def Tosa_GatherOp : Tosa_InferShapedTypeOp<"gather"> {
let summary = "Gather operation,";
let description = [{
- Generate a tensor for which each element in the output is a slice of the
- values tensor based on the value of indices.
+ Generate a tensor for which each element in the output is a subtensor of the
+ values tensor based on the indices. N is the number of batches, W the number
+ of indices in each batch, K the range of each index and C the number data
+ channels for each index.
}];
let arguments = (ins
@@ -1759,8 +1781,14 @@ def Tosa_ScatterOp : Tosa_InferShapedTypeOp<"scatter"> {
let summary = "Scatter operation,";
let description = [{
- The values_out tensor is set to the values_in tensor with data modified as follows:
- data from the input tensor is inserted at the positions specified by the indices tensor.
+ The values_out tensor is set to the values_in tensor with data modified as
+ follows: data from the input tensor is inserted at the positions specified
+ by the indices tensor. N is the number of batches, W the number of indices
+ in each batch, K the range of each index and C the number data channels for
+ each index. It is not permitted to repeat the same output index within a
+ single SCATTER operation and so each output index occurs at most once. It
+ follows that K >= W. In use cases that require multiple updates to the same
+ output position, these must be decomposed into multiple SCATTER operations.
}];
let arguments = (ins
@@ -1786,12 +1814,30 @@ def Tosa_ResizeOp : Tosa_InferShapedTypeOp<"resize"> {
let summary = "Resize operation, supports various resize/upsample modes";
let description = [{
- Resizes a tensor. Resize is only allowed in the H and W dimensions. In
- expected use, The height dimension is scaled by factor (scale_y_n/scale_y_d).
- And the width dimension is scaled by factor (scale_x_n/scale_x_d). Thus the
- output dimensions can be derived from the input dimensions by inverting the
- scale. And the [order_y, border_x] values adjust the output size to allow
- fractional sampling beyond integer input position (IH-1,IW-1).
+ Resizes a tensor. Resize is only allowed in the H and W dimensions.
+
+ The height dimension is scaled by factor (scale_y_n/scale_y_d). The width
+ dimension is scaled by factor (scale_x_n/scale_x_d).
+
+ The NEAREST_NEIGHBOR mode returns the value of the input tensor closest to
+ the calculated sample position for both floating-point and integer data
+ formats.
+
+ Floating-point BILINEAR mode returns a bilinearly interpolated output value
+ based on the four closest input sample positions.
+
+ For integer BILINEAR interpolation mode, the output value must be scaled by
+ 1/(scale_y_n * scale_x_n) in a following operation to complete the
+ interpolation (for example with a RESCALE operator).
+
+ The output dimensions can be derived from the input dimensions by inverting
+ the scale as described in the pseudocode. The [border_y, border_x] values
+ adjust the output size to allow fractional sampling beyond integer input
+ position (IH - 1,IW - 1).
+
+ The limit MAX_SCALE is applied to each scale ratio after reduction of the
+ ratio. Individual scale numerator and denominaor values are allowed to be
+ larger than MAX_SCALE.
}];
let arguments = (ins
|
|
@llvm/pr-subscribers-mlir-tosa Author: Jerry-Ge (Jerry-Ge) ChangesThis commit aligns the operator descriptions to the v1.0 spec (https://www.mlplatform.org/tosa/tosa_spec.html) Co-authored by:
Full diff: https://github.com/llvm/llvm-project/pull/128046.diff 1 Files Affected:
diff --git a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
index 7cdf79f4dc59d..695a8b4700090 100644
--- a/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
+++ b/mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td
@@ -37,7 +37,8 @@ def Tosa_ArgMaxOp : Tosa_InferShapedTypeOp<"argmax"> {
let description = [{
This returns the index with the largest value across the given axis of the
- input tensor.
+ input tensor. If multiple locations have equal values, returns the first
+ match along the search axis.
}];
let arguments = (ins
@@ -98,7 +99,8 @@ def Tosa_Conv2DOp : Tosa_ConvOp<"conv2d"> {
let description = [{
Performs a 2D convolution over the given tensor input, using the weight
- tensor.
+ tensor. Implementations may choose to skip calculation of multiplies in
+ the padding area.
}];
let arguments = (ins
@@ -129,7 +131,8 @@ def Tosa_Conv3DOp : Tosa_ConvOp<"conv3d"> {
let summary = "3D Convolution operator";
let description = [{
- Performs a 3D convolution over the given input tensor.
+ Performs a 3D convolution over the given input tensor. Implementations
+ may choose to skip calculation of multiplies in the padding area.
}];
let arguments = (ins
@@ -161,7 +164,8 @@ def Tosa_DepthwiseConv2DOp : Tosa_ConvOp<"depthwise_conv2d"> {
let description = [{
Performs 2D convolutions separately over each channel of the given tensor
- input, using the weight tensor.
+ input, using the weight tensor. Implementations may choose to skip
+ calculation of multiplies in the padding area.
}];
let arguments = (ins
@@ -323,7 +327,8 @@ def Tosa_TransposeConv2DOp : Tosa_ConvOp<"transpose_conv2d"> {
let description = [{
Performs a 2D transposed convolution over the given tensor input, using the
- weights tensor.
+ weights tensor. Implementations may choose to skip calculation of multiplies
+ by zero at fractional input positions.
}];
let arguments = (ins
@@ -388,10 +393,14 @@ def Tosa_SigmoidOp : Tosa_ElementwiseUnaryOp<"sigmoid"> {
let summary = "Computes elementwise sigmoid of input.";
let description = [{
- Sigmoid function: output = 1 / (1 + exp(-input))
- For quantized integer data types, the TABLE operator should be used instead
- with the following definition. The sigmoid table has 513 entries each of
- 16-bit precision and covering the input range -16.0 to +16.0
+ Applies the sigmoid logistic function to each element of the input tensor:
+ $ sigmoid(x) = \frac{1}{1 + e^{-x}} $.
+
+ For quantized integer data types, the TABLE operator should be used instead.
+ Each implementation may choose an appropriate TABLE given the scale and zero
+ point of the input data. Eight or sixteen bit precision tables may be used
+ based on the input tensor to the sigmoid function. The sigmoid table has 513
+ entries each of 16-bit precision and covering the input range -16.0 to +16.0
in steps of 1/16.
}];
@@ -411,10 +420,14 @@ def Tosa_TanhOp : Tosa_ElementwiseUnaryOp<"tanh"> {
let summary = "Computes elementwise hyperbolic tangent of input";
let description = [{
- Parameterized hyperbolic tangent.
- For quantized integer data types, the TABLE operator should be used instead
- with the following definition. The tanh_table has 513 entries each of
- 16-bit precision and covering the input range -8.0 to +8.0 in steps of 1/32.
+ Parameterized hyperbolic tangent: $ tanh(x) = \frac{1 - e^{-2x}}{1 + e^{-2x}} $.
+
+ For quantized integer data types, the TABLE operator should be used instead.
+ Each implementation may choose an appropriate TABLE given the scale and zero
+ point of the input data. Eight or sixteen bit precision tables may be used
+ based on the input tensor to the tanh function. The tanh_table has 513
+ entries each of 16-bit precision and covering the input range -8.0 to +8.0
+ in steps of 1/32.
}];
let arguments = (ins
@@ -433,10 +446,10 @@ def Tosa_ErfOp : Tosa_ElementwiseUnaryOp<"erf"> {
let summary = "Computes gauss error function of input";
let description = [{
- Gauss error function: $ erf(x) = \frac{2}{\sqrt(\pi)} \int_{0}^{x} e^{-t^2} \ dt $
+ Gauss error function: $ erf(x) = \frac{2}{\sqrt{\pi}} \int_{0}^{x} e^{-t^2} dt $
For quantized integer data types, the TABLE operator should be used instead
- with the following definition. The erf_table has 513 entries each of
- 16-bit/8-bit precision and covering the input range -4.0 to +4.0 in steps of 1/64.
+ with the following definition. The ERF table has 513 entries each of
+ 16-bit precision and covering the input range -4.0 to +4.0 in steps of 1/64.
}];
let arguments = (ins
@@ -466,7 +479,7 @@ def Tosa_AddOp : Tosa_ElementwiseOp<"add", [
let description = [{
Elementwise addition of input1 and input2. Axis of size 1 will be broadcast,
- as necessary.
+ as necessary. Rank of input tensors must match.
Example:
@@ -500,7 +513,8 @@ def Tosa_ArithmeticRightShiftOp : Tosa_ElementwiseOp<"arithmetic_right_shift",
let description = [{
Elementwise arithmetic right shift of input1 by the amount specified in
- input2. Axis of size 1 will be broadcast, as necessary.
+ input2. Axis of size 1 will be broadcast, as necessary. Rank of input tensors
+ must match.
}];
let arguments = (ins
@@ -524,7 +538,7 @@ def Tosa_BitwiseAndOp : Tosa_ElementwiseOp<"bitwise_and", [
let description = [{
Elementwise bitwise AND of input1 and input2. Axis of size 1
- will be broadcast as necessary.
+ will be broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -547,7 +561,7 @@ def Tosa_BitwiseOrOp : Tosa_ElementwiseOp<"bitwise_or", [
let description = [{
Elementwise bitwise OR of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -570,7 +584,7 @@ def Tosa_BitwiseXorOp : Tosa_ElementwiseOp<"bitwise_xor", [
let description = [{
Elementwise bitwise XOR of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -590,8 +604,10 @@ def Tosa_IntDivOp : Tosa_ElementwiseOp<"int_div", [SameOperandsAndResultElementT
let summary = "Integer divide operator";
let description = [{
- Elementwise integer divide operator of input1 by input2. Axis of size 1
- will be broadcast, as necessary.
+ Elementwise integer divide operator of input1 by input2. The result of the divide
+ is truncated towards zero. Expected use is for operations on non-scaled integers.
+ Floating point divide should use RECIPROCAL and MUL. Quantized integer divide
+ should use TABLE (for 1/x) and MUL.
}];
let arguments = (ins
@@ -616,7 +632,7 @@ def Tosa_LogicalAndOp : Tosa_ElementwiseOp<"logical_and", [
let description = [{
Elementwise logical AND of input1 and input2. Axis of size 1 will be
- broadcast, as necessary.
+ broadcast, as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -638,7 +654,7 @@ def Tosa_LogicalLeftShiftOp : Tosa_ElementwiseOp<"logical_left_shift",
let description = [{
Elementwise left shift of input1 and input2. Axis of size 1 will be
- broadcast, as necessary.
+ broadcast, as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -660,7 +676,8 @@ def Tosa_LogicalRightShiftOp : Tosa_ElementwiseOp<"logical_right_shift",
let description = [{
Elementwise logical right shift of input1 by the amount specified in input2.
- Axis of size 1 will be broadcast, as necessary.
+ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must
+ match.
}];
let arguments = (ins
@@ -683,7 +700,7 @@ def Tosa_LogicalOrOp : Tosa_ElementwiseOp<"logical_or", [
let description = [{
Elementwise logical OR of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -705,8 +722,8 @@ def Tosa_LogicalXorOp : Tosa_ElementwiseOp<"logical_xor", [
let summary = "Returns the truth value of x XOR y element-wise.";
let description = [{
- Elementwise logical XOR of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ Elementwise logical XOR of input1 and input2. Axis of size 1 will be
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -729,7 +746,7 @@ def Tosa_MaximumOp : Tosa_ElementwiseOp<"maximum", [
let description = [{
Elementwise max of input1 and input2. Axis of size 1 will be broadcast, as
- necessary.
+ necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -753,7 +770,7 @@ def Tosa_MinimumOp : Tosa_ElementwiseOp<"minimum", [
let description = [{
Elementwise minimum of input1 and input2. Axis of size 1
- will be broadcast, as necessary.
+ will be broadcast, as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -784,8 +801,8 @@ def Tosa_MulOp : Tosa_Op<"mul", [
let description = [{
Elementwise multiplication (Hadamard product) of input1 and input2.
- Axis of size 1 will be broadcast, as necessary.
- i8/i16 input type can be promoted to i32 result type.
+ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must
+ match.
}];
let arguments = (ins
@@ -814,7 +831,8 @@ def Tosa_PowOp : Tosa_ElementwiseOp<"pow", [SameOperandsAndResultElementType]> {
let description = [{
Elementwise input1 raised to the power of input2.
- Axis of size 1 will be broadcast, as necessary.
+ Axis of size 1 will be broadcast, as necessary. Rank of input tensors must
+ match.
}];
let arguments = (ins
@@ -835,7 +853,7 @@ def Tosa_SubOp : Tosa_ElementwiseOp<"sub", [SameOperandsAndResultElementType]> {
let description = [{
Elementwise subtraction of input1 and input2. Axis of size 1 will be
- broadcast as necessary.
+ broadcast as necessary. Rank of input tensors must match.
}];
let arguments = (ins
@@ -857,20 +875,20 @@ def Tosa_TableOp : Tosa_InferShapedTypeOp<"table"> {
let summary = "Table lookup op";
let description = [{
- Interpolated table lookup operation. Input values are scaled to create a
- fixed-point 9.7 value. The high 9 bits are used to index into the table.
- The fractional bits are used to interpolate based on the looked up value and
- the index+1 value in the table. The TABLE operator then returns a 16.7
- interpolated value. Note that there must be 513 values to handle the full
- range of inputs.
+ Table lookup operation. For int8_t TABLE operation, perform a 256 entry
+ table lookup returning an int8_t value. For int16_t tables, the int16_t
+ input is treated as a fixed-point 9.7 value. The most significant 9 bits
+ are used to index into the table. The fractional 7 bits are used to
+ interpolate based on table[index] and table[index+1]. For int16_t inputs,
+ the TABLE operator returns a 16.7 interpolated value in an int32_t. This
+ value can then be input to the RESCALE operator to scale to the required
+ output data type. Note that int16_t table has 513 values to handle
+ table[index+1] when index=511.
- The TABLE operator is expected to be used as follows:
- * A RESCALE node is expected before the TABLE operator to scale the input
- to a full int16_t range for the table lookup
- * If an int16_t result is required then follow the TABLE operator with a
- RESCALE with a right shift of 7
- * If an int8_t result is required then follow the TABLE operator with a
- RESCALE with a right shift of 15
+ An int16_t to int16_t table lookup can be constructed in TOSA as follows:
+ * Use the TABLE operator to produce a fixed point 16.7 interpolated result
+ * Use RESCALE (in_t=int32_t, out_t=int16_t, scale=1<<14, shift=21) to
+ scale the output to int16_t range (or alternate scale as required)
}];
let arguments = (ins
@@ -902,7 +920,7 @@ def Tosa_AbsOp : Tosa_ElementwiseUnaryOp<"abs"> {
let summary = "Elementwise abs op";
let description = [{
- Elementwise absolute value operation
+ Elementwise absolute value operation.
Example:
@@ -1026,7 +1044,7 @@ def Tosa_FloorOp : Tosa_ElementwiseUnaryOp<"floor"> {
let summary = "Elementwise floor op";
let description = [{
- Elementwise floor operation
+ Elementwise floor operation.
}];
let arguments = (ins
@@ -1085,7 +1103,7 @@ def Tosa_NegateOp : Tosa_ElementwiseUnaryOp<"negate"> {
let summary = "Elementwise negate op";
let description = [{
- Elementwise negation operation
+ Elementwise negation operation.
}];
let arguments = (ins
@@ -1543,9 +1561,9 @@ def Tosa_PadOp : Tosa_InferShapedTypeOp<"pad"> {
let summary = "Pads a tensor with value specified.";
let description = [{
- The `tosa.pad` operation pads a tensor along borders of each dimension with
- `pad_const` (defaults to zero), given a padding configuration `padding`
- specifying low and high values along the dimensions.
+ Pads a tensor along the borders of each dimension with a supplied value.
+ Returns a new tensor with the padding included. The pad_const value includes
+ the zero point if the tensor uses a zero point.
Example:
@@ -1705,7 +1723,9 @@ def Tosa_TransposeOp : Tosa_InferShapedTypeOp<"transpose",
let summary = "Transpose operator";
let description = [{
- Permutes the dimensions based on perm.
+ Permutes the dimensions of the input tensor input1 based on the perms
+ argument. Each value in the perms list must be a valid dimension of the
+ input tensor and may not be repeated.
}];
let arguments = (ins
@@ -1738,8 +1758,10 @@ def Tosa_GatherOp : Tosa_InferShapedTypeOp<"gather"> {
let summary = "Gather operation,";
let description = [{
- Generate a tensor for which each element in the output is a slice of the
- values tensor based on the value of indices.
+ Generate a tensor for which each element in the output is a subtensor of the
+ values tensor based on the indices. N is the number of batches, W the number
+ of indices in each batch, K the range of each index and C the number data
+ channels for each index.
}];
let arguments = (ins
@@ -1759,8 +1781,14 @@ def Tosa_ScatterOp : Tosa_InferShapedTypeOp<"scatter"> {
let summary = "Scatter operation,";
let description = [{
- The values_out tensor is set to the values_in tensor with data modified as follows:
- data from the input tensor is inserted at the positions specified by the indices tensor.
+ The values_out tensor is set to the values_in tensor with data modified as
+ follows: data from the input tensor is inserted at the positions specified
+ by the indices tensor. N is the number of batches, W the number of indices
+ in each batch, K the range of each index and C the number data channels for
+ each index. It is not permitted to repeat the same output index within a
+ single SCATTER operation and so each output index occurs at most once. It
+ follows that K >= W. In use cases that require multiple updates to the same
+ output position, these must be decomposed into multiple SCATTER operations.
}];
let arguments = (ins
@@ -1786,12 +1814,30 @@ def Tosa_ResizeOp : Tosa_InferShapedTypeOp<"resize"> {
let summary = "Resize operation, supports various resize/upsample modes";
let description = [{
- Resizes a tensor. Resize is only allowed in the H and W dimensions. In
- expected use, The height dimension is scaled by factor (scale_y_n/scale_y_d).
- And the width dimension is scaled by factor (scale_x_n/scale_x_d). Thus the
- output dimensions can be derived from the input dimensions by inverting the
- scale. And the [order_y, border_x] values adjust the output size to allow
- fractional sampling beyond integer input position (IH-1,IW-1).
+ Resizes a tensor. Resize is only allowed in the H and W dimensions.
+
+ The height dimension is scaled by factor (scale_y_n/scale_y_d). The width
+ dimension is scaled by factor (scale_x_n/scale_x_d).
+
+ The NEAREST_NEIGHBOR mode returns the value of the input tensor closest to
+ the calculated sample position for both floating-point and integer data
+ formats.
+
+ Floating-point BILINEAR mode returns a bilinearly interpolated output value
+ based on the four closest input sample positions.
+
+ For integer BILINEAR interpolation mode, the output value must be scaled by
+ 1/(scale_y_n * scale_x_n) in a following operation to complete the
+ interpolation (for example with a RESCALE operator).
+
+ The output dimensions can be derived from the input dimensions by inverting
+ the scale as described in the pseudocode. The [border_y, border_x] values
+ adjust the output size to allow fractional sampling beyond integer input
+ position (IH - 1,IW - 1).
+
+ The limit MAX_SCALE is applied to each scale ratio after reduction of the
+ ratio. Individual scale numerator and denominaor values are allowed to be
+ larger than MAX_SCALE.
}];
let arguments = (ins
|
Jerry-Ge
force-pushed
the
description
branch
2 times, most recently
from
92db3e1 to
e5c8136
Compare
lhutton1
reviewed
Feb 21, 2025
This commit aligns the operator descriptions to the v1.0 spec. Change-Id: I4774faa861e04748d411ba99fe61fe5d28475bc4 Co-authored-by: Peng Sun <[email protected]> Signed-off-by: Luke Hutton <[email protected]>
Jerry-Ge
force-pushed
the
description
branch
from
e5c8136 to
0a3feec
Compare
GeorgeARM
approved these changes
Feb 21, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This commit aligns the operator descriptions to the v1.0 spec (https://www.mlplatform.org/tosa/tosa_spec.html)
Co-authored-by: Peng Sun <[email protected]>
Signed-off-by: Luke Hutton <[email protected]>