SSPU.dll Specification

1. Overview

This DLL provides SSPU communication plus Sloppy operations. Each operation sends the input data to the SSPU device, performs the calculation with the specified parameters, and receives output data.

2. Runtime environment

3. Import method

3.1 Prerequisites

• Install .NET 8.0 SDK or later.
• Run the application with administrator privileges.
• Place SSPU.dll in an accessible path.

3.2 Add a reference

// Add a reference to SSPU.dll
// Option A: Visual Studio: right-click project -> Add Reference -> Browse
// Option B: add this to .csproj:
<Reference Include="SSPU">
  <HintPath>path\to\SSPU.dll</HintPath>
</Reference>

3.3 Integration steps

using SSPU;

// 1. Create an instance
var sspu = new SSPUDevice();

// 2. Connect the device
int result = sspu.Connect();
if (result != 0) {
    Console.WriteLine($"Connection failed: {result}");
    return;
}

// 3. Run an operation (SloppySigmoid example)
byte[] inBuffer = PrepareInputData(...);
byte[] outBuffer = new byte[expectedOutputSize];
result = sspu.SloppySigmoid(inBuffer, outBuffer, param_n, param_b);

// 4. Prefer asynchronous execution
await Task.Run(() => sspu.SloppySigmoid(inBuffer, outBuffer, param_n, param_b));

4. API specification

4.1 Method list

int Connect()

int SloppySigmoid(byte[] inBuffer, byte[] outBuffer, byte param_n, byte param_b)

int SloppyTanh(byte[] inBuffer, byte[] outBuffer, byte param_n, byte param_b)

int SloppySiLU(byte[] inBuffer, byte[] outBuffer, byte param_n, byte param_b)

int SloppyGeLU(byte[] inBuffer, byte[] outBuffer, byte param_n, byte param_b)

int SloppySoftmax(byte[] inBuffer, byte[] outBuffer, byte param_n, byte param_b, byte param_d, byte param_mode = 0)

int IntAddSat(byte[] aBuffer, byte[] bBuffer, byte[] outBuffer, byte param_n)

int IntSubSat(byte[] aBuffer, byte[] bBuffer, byte[] outBuffer, byte param_n)

int IntMulSat(byte[] aBuffer, byte[] bBuffer, byte[] outBuffer, byte param_n, byte param_k)

4.2 Error information

5. Input/output data specification

5.1 Terminology

5.2 Input/output types

5.3 Input data

Input-data generation flow

1. Mask the value to n bits (for example, use only the lower 22 bits).
2. Sign-extend it: when the highest bit is 1, treat it as negative and fill higher bits with 1.
3. Pack the bits LSB first.
4. Store each item in (n + 7) / 8 bytes; unused bits in the final byte are not used.

Bit layout example (n = 22 -> 3 bytes/item)

Input example: {45, -73} (n = 22)

// Byte sequence for {45, -73} (n = 22, 3 bytes/data):
// 45  -> 0x00002D -> Byte0: 0x2D, Byte1: 0x00, Byte2: 0x00
// -73 -> 0x3FFFB7 -> Byte0: 0xB7, Byte1: 0xFF, Byte2: 0x3F
//
// Final inBuffer: 2D 00 00 B7 FF 3F

SloppySoftmax input/output notes

• Input data consists of consecutive Softmax groups; every group contains 2^d consecutive elements.
• Output is returned in the same group order.
• The number of input elements must be a multiple of 2^d (complete groups).
• This is an approximate operation; output can deviate by approximately +/-1 least-significant bit.

5.4 Output data

Output-data reconstruction flow

1. Restore the output-width bits LSB first.
2. Interpret the sign according to the output type.
3. Signed (Tanh / SiLU / GeLU): use the highest bit as the sign bit and sign-extend it.
4. Unsigned (Sigmoid / Softmax): do not sign-extend (fixed-point fraction at or above 0).

Buffer allocation

5.5 Capacity limit

One calculation call can process at most one device page: an effective area of 15872 elements. Calls exceeding the limit must be split by the caller.

SloppySoftmax maximum group count examples

Important: When the limit is exceeded, error code -12 is returned. SloppySoftmax also returns -12 when its input element count is not a multiple of 2^d.