chapters/type_conversion.adoc

//
// This confidential and proprietary software may be used only as
// authorised by a licensing agreement from ARM Limited
// (C) COPYRIGHT 2020-2022 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.

=== Type Conversion

==== CAST

Casts a tensor from one data type to another.

*Arguments:*

|===
|Argument|Type|Name|Shape|Description

|Input|in_t*|input|shape|Input tensor
|Output|out_t*|output|shape|Output tensor
|===

*Operation Function:*

[source,c++]
----
for_each(index in shape) {
    in_t in = tensor_read<in_t>(input, shape, index);
    out_t out;
    if (out_t == bool_t) {
        out = (in != 0) ? true : false;
    } else if (in_t == bool_t) {
        out = (in) ? 1 : 0;
    } else if (out_t == fp16_t || out_t == bf16_t || out_t == fp32_t) {
        out = round_to_nearest_float(in);
    } else if (in_t == fp16_t || in_t == bf16_t || in_t == fp32_t) {
        out = apply_clip<out_t>(round_to_nearest_int(in), minimum<out_t>, maximum<out_t>);
    } else if (sizeof(out_t) >= sizeof(in_t)) {
        out = sign_extend(in);
    } else {
        out = truncate(in);
    }
    tensor_write<out_t>(output, shape, index, out)
}
----

*Supported Data Types:*

|===
|Profile|Mode|in_t|out_t

|Any|bool to signed 8|bool_t|int8_t
|Any|bool to signed 16|bool_t|int16_t
|Any|bool to signed 32|bool_t|int32_t
|Any|signed 8 to bool|int8_t|bool_t
|Any|signed 8 to signed 16|int8_t|int16_t
|Any|signed 8 to signed 32|int8_t|int32_t
|MI, MT|signed 8 to fp16|int8_t|fp16_t
|MI, MT|signed 8 to bf16|int8_t|bf16_t
|MI, MT|signed 8 to fp32|int8_t|fp32_t
|Any|signed 16 to bool|int16_t|bool_t
|Any|signed 16 to signed 8|int16_t|int8_t
|Any|signed 16 to signed 32|int16_t|int32_t
|MI, MT|signed 16 to fp16|int16_t|fp16_t
|MI, MT|signed 16 to bf16|int16_t|bf16_t
|MI, MT|signed 16 to fp32|int16_t|fp32_t
|Any|signed 32 to bool|int32_t|bool_t
|Any|signed 32 to signed 8|int32_t|int8_t
|Any|signed 32 to signed 16|int32_t|int16_t
|MI, MT|signed 32 to fp16|int32_t|fp16_t
|MI, MT|signed 32 to bf16|int32_t|bf16_t
|MI, MT|signed 32 to fp32|int32_t|fp32_t
|MI, MT|fp16 to signed 8|fp16_t|int8_t
|MI, MT|fp16 to signed 16|fp16_t|int16_t
|MI, MT|fp16 to signed 32|fp16_t|int32_t
|MI, MT|bf16 to signed 8|bf16_t|int8_t
|MI, MT|bf16 to signed 16|bf16_t|int16_t
|MI, MT|bf16 to signed 32|bf16_t|int32_t
|MI, MT|fp32 to signed 8|fp32_t|int8_t
|MI, MT|fp32 to signed 16|fp32_t|int16_t
|MI, MT|fp32 to signed 32|fp32_t|int32_t
|===

==== RESCALE

Rescale quantized values into a new domain. This function scales by factor: multiplier * 2^-shift^.

*Arguments:*

|===
|Argument|Type|Name|Shape|Description

|Input|in_t*|input|shape|Input tensor from 1 to 4 dims
|Output|out_t*|output|shape|Output tensor with the same shape as input
|Attribute|in_t|input_zp|-|Input tensor zero point. Must be zero for non-int8 types.
|Attribute|out_t|output_zp|-|Output tensor zero point. Must be zero for non-int8 types.
|Input (MT profile) Attribute (BI/MI profiles)|mul_t*|multiplier|[NC]|Scaling multiplier array
|Input (MT profile) Attribute (BI/MI profiles)|uint6_t*|shift|[NC]|Scaling shift array
|Attribute|bool_t|scale32|-|if (scale32) mul_t=int32_t else mul_t=int16_t
|Attribute|bool_t|double_round|-|Select double round mode
|Attribute|bool_t|per_channel|-|if (per_channel) NC=shape[dims-1] else NC=1
|===

*Operation Function:*

[source,c++]
----
for_each(index in shape) {
    // uint16 values can have zero_point 0 or 32768
    // int8/uint8 can have zero point within their valid range
    // No other types can have zero point != 0
    ERROR_IF(in_t != int8_t &&
             in_t != uint8_t &&
             in_t != uint16_t && input_zp != 0);
    ERROR_IF(out_t != int8_t &&
             out_t != uint8_t &&
             out_t != uint16_t && output_zp != 0);
    ERROR_IF(in_t == uint16_t && (input_zp != 0 || input_zp != 32768));
    ERROR_IF(out_t == uint16_t && (output_zp != 0 || output_zp != 32768));
    ERROR_IF(scale32 && in_t == int48_t);
    ERROR_IF(!scale32 && double_round);
    int48_t value = tensor_read<in_t>(input, shape, index);
    value = value - input_zp;
    int c = (per_channel) ? index[dims-1] : 0;
    int32_t result = (scale32) ?
        apply_scale_32(value, multiplier[c], shift[c], double_round) :
        apply_scale_16(value, multiplier[c], shift[c]);
    result = (out_t)apply_clip<int32_t>(result + output_zp, minimum<out_t>, maximum<out_t>);
    tensor_write<out_t>(output, shape, index, result);
}
----

*Supported Data Types:*

|===
|Profile|Mode|in_t|out_t

|Any|signed 8 to signed 8|int8_t|int8_t
|Any|signed 8 to signed 16|int8_t|int16_t
|Any|signed 8 to signed 32|int8_t|int32_t
|Any|signed 8 to unsigned 8|int8_t|uint8_t
|Any|signed 16 to signed 8|int16_t|int8_t
|Any|signed 16 to signed 16|int16_t|int16_t
|Any|signed 16 to signed 32|int16_t|int32_t
|Any|signed 16 to unsigned 8|int16_t|uint8_t
|Any|signed 16 to unsigned 16|int16_t|uint16_t
|Any|signed 32 to signed 8|int32_t|int8_t
|Any|signed 32 to signed 16|int32_t|int16_t
|Any|signed 32 to signed 32|int32_t|int32_t
|Any|signed 48 to signed 8|int48_t|int8_t
|Any|signed 48 to signed 16|int48_t|int16_t
|Any|signed 48 to signed 32|int48_t|int32_t
|Any|unsigned 8 to signed 8|uint8_t|int8_t
|Any|unsigned 8 to signed 16|uint8_t|int16_t
|Any|unsigned 16 to signed 16|uint16_t|int16_t
|===