From 3309a5362a13f840e84a2f67b9ba7141aae58cc4 Mon Sep 17 00:00:00 2001 From: Eric Kunze Date: Thu, 1 Oct 2020 18:50:46 -0700 Subject: TOSA spec 0.20.0, initial public version Change-Id: I7c528ebf7e78759ada4bb951b2471c367d3f4525 --- chapters/ewise_binary.adoc | 611 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 611 insertions(+) create mode 100644 chapters/ewise_binary.adoc (limited to 'chapters/ewise_binary.adoc') diff --git a/chapters/ewise_binary.adoc b/chapters/ewise_binary.adoc new file mode 100644 index 0000000..92c4926 --- /dev/null +++ b/chapters/ewise_binary.adoc @@ -0,0 +1,611 @@ +// +// This confidential and proprietary software may be used only as +// authorised by a licensing agreement from ARM Limited +// (C) COPYRIGHT 2020 ARM Limited +// ALL RIGHTS RESERVED +// The entire notice above must be reproduced on all authorised +// copies and copies may only be made to the extent permitted +// by a licensing agreement from ARM Limited. + +=== Elementwise Binary Operators + +==== ADD + +Elementwise addition of input1 and input2. Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = apply_add(value1, value2) + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 32|int32 +|MI, MT|float|float +|=== + +==== ARITHMETIC_RIGHT_SHIFT + +Elementwise arithmetic right shift of input1 by the amount specified in input2. Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + assert(value2 <= 31) + in_t acc = value1 >> value2 + acc = apply_clip(acc, minimum, maximum) + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 8|int8 +|Any|signed 16|int16 +|Any|signed 32|int32 +|=== + +==== BITWISE_AND + +Elementwise bitwise AND of input tensor 0 and input tensor 1. Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor of same type as the input tensors, with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = value1 & value2 + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 8|aint8 +|Any|signed 16|int16 +|Any|signed 32|int32 +|=== + +==== BITWISE_OR + +Elementwise bitwise OR of input1 and input2. Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = value1 | value2 + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 8|aint8 +|Any|signed 16|int16 +|Any|signed 32|int32 +|=== + +==== BITWISE_XOR + +Elementwise bitwise XOR of input1 and input2. Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = value1 ^ value2 + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 8|aint8 +|Any|signed 16|int16 +|Any|signed 32|int32 +|=== + +==== LOGICAL_AND + +Elementwise logical AND of input1 and input2. Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Quantization Parameters:* + +None + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = value1 && value2 + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|Bool|Bool +|=== + +==== LOGICAL_LEFT_SHIFT + +Elementwise left shift of input1 and input2. Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + assert(value2 <= 31) + in_t acc = value1 << value2 + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 8|int8 +|Any|signed 16|int16 +|Any|signed 32|int32 +|=== + +==== LOGICAL_RIGHT_SHIFT + +Elementwise logical right shift of input1 by the amount specified in input2. Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + assert(value2 <= 31) + in_t acc = (unsigned in_t)value1 >> value2 + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 8|int8 +|Any|signed 16|int16 +|Any|signed 32|int32 +|=== + +==== LOGICAL_OR + +Elementwise logical OR of input1 and input2. Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = value1 || value2 + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|Bool|Bool +|=== + +==== LOGICAL_XOR + +Elementwise logical XOR of input tensor 0 and input tensor 1. Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor from 1 to 4 dims +|Input|in_t*|input2|shape2|Input tensor with the same rank as Input 0 +|Output|in_t*|output|shape|Output tensor of same type as the input tensors, with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = value1 != value2 + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|Bool|Bool +|=== + +==== MAXIMUM + +Elementwise max of input1 and input2. Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = apply_max(value1, value2) + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 32|int32 +|MI, MT|float|float +|=== + +==== MINIMUM + +Elementwise minimum of input tensor 0 and input tensor 1. Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as input1 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = apply_min(value1, value2) + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 32|int32 +|MI, MT|float|float +|=== + +==== MUL + +Elementwise multiplication (Hadamard product) of input tensor 0 and input tensor 1. Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as Input 0 +|Output|out_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = value1 * value2 // takes low bits for int32_t + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* +|=== +|Profile|Mode|in_t|out_t + +|Any|signed 8|int8|int32 +|Any|signed 16|int16|int32 +|Any|signed 32|int32|int32 +|MI, MT|float|float|float +|=== + +==== POW + +Elementwise input tensor 0 value raised to the power of input 1 tensor. Axis of size 1 will be broadcast, as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor from 1 to 4 dims +|Input|in_t*|input2|shape2|Input tensor with the same rank as Input 0 +|Output|in_t*|output|shape|Output tensor of same type as the input tensors, with broadcast shape if necessary +|=== + +*Quantization Parameters:* + +Only supported with floating point values. + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|MI, MT|float|float +|=== + +==== SUB + +Elementwise subtraction of input tensor 0 and input tensor 1. Axis of size 1 will be broadcast as necessary. Rank of input tensors must match. + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|input1|shape1|Input tensor +|Input|in_t*|input2|shape2|Input tensor with the same rank as Input 0 +|Output|in_t*|output|shape|Output tensor with broadcast shape if necessary +|=== + +*Operation Function:* + +[source,c] +---- +for_each (index in shape) { + index1 = apply_broadcast(shape, shape1, index) + index2 = apply_broadcast(shape, shape2, index) + in_t value1 = tensor_read(input1, shape1, index1) + in_t value2 = tensor_read(input2, shape2, index2) + in_t acc = apply_sub(value1, value2); + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t + +|Any|signed 32|int32 +|MI, MT|float|float +|=== + +==== TABLE + +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. + +The TABLE operator is expected to be used as follows: + +* A RECALE 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 + +*Arguments:* + +|=== +|Argument|Type|Name|Shape|Description + +|Input|in_t*|Input|shape|Input tensor +|Input|table_t*|table|[513]|Lookup table tensor +|Output|out_t*|output|shape|Output tensor +|=== + +*Quantization Parameters:* + +None + +*Operation Function:* + +[source,c] +---- +assert(dimensions(shape)<=4) +for_each (index in shape) { + in_t value = tensor_read(input, shape, index) + out_t acc = apply_lookup(table, value) + tensor_write(output, shape, index, acc) +} +---- + +*Supported Data Types:* + +|=== +|Profile|Mode|in_t|table_t|out_t + +|Any|signed 16|int16|int16|int32 +|=== + -- cgit v1.2.1