TXOR

Instruction Diagram

Introduction

Performs element-wise bitwise XOR on two tiles.

Mathematical Semantics

For each element (i, j) in the valid region:

$$\mathrm{dst}_{i,j} = \mathrm{src0}_{i,j} \oplus \mathrm{src1}_{i,j}$$

Assembly Syntax

Synchronous form:

%dst = txor %src0, %src1 : !pto.tile<...>

AS Level 1 (SSA)

%dst = pto.txor %src0, %src1 : (!pto.tile<...>, !pto.tile<...>) -> !pto.tile<...>

AS Level 2 (DPS)

pto.txor ins(%src0, %src1 : !pto.tile_buf<...>, !pto.tile_buf<...>) outs(%dst : !pto.tile_buf<...>)

C++ Built-in APIs

Declared in include/pto/common/pto_instr.hpp:

The public include header is <pto/pto-inst.hpp>, and the internal declaration is located in pto/common/pto_instr.hpp.

template <typename TileDataDst, typename TileDataSrc0, typename TileDataSrc1, typename TileDataTmp,
          typename... WaitEvents>
PTO_INST RecordEvent TXOR(TileDataDst &dst, TileDataSrc0 &src0, TileDataSrc1 &src1, TileDataTmp &tmp, WaitEvents &... events);

Constraints

• The operation iterates over dst.GetValidRow()/dst.GetValidCol().
• Implementation check (Ascend 950PR/Ascend 950DT):
  • The element types of dst, src0, and src1 must be consistent.
  • The supported element types are uint8_t, int8_t, uint16_t, int16_t, uint32_t, and int32_t.
  • dst, src0, and src1 must be row-major.
  • src0.GetValidRow()/GetValidCol() and src1.GetValidRow()/GetValidCol() must be consistent with dst.
• Implementation check (Atlas A2/A3 training products/Atlas A2/A3 inference products):
  • The element types of dst, src0, src1, and tmp must be consistent.
  • The supported element types are uint8_t, int8_t, uint16_t, int16_t, uint32_t, and int32_t.
  • dst, src0, src1, and tmp must be row-major.
  • The valid shapes of src0, src1, and tmp must be consistent with that of dst.
  • In manual mode, the memory regions of dst, src0, src1, and tmp must not overlap.

Temporary Space

Atlas A2/A3 Training Products/Atlas A2/A3 Inference Products

tmp is used as intermediate scratch storage. The implementation of Atlas A2/A3 training products/Atlas A2/A3 inference products computes XOR by decomposition: XOR(a,b) = AND(NOT(AND(a,b)), OR(a,b)), which requires tmp to hold the intermediate result OR(a,b).

• tmp must have the same element type as dst/src0/src1.
• tmp must be row-major.
• The valid shape of tmp must be consistent with dst (tmp.GetValidRow() == dst.GetValidRow() and tmp.GetValidCol() == dst.GetValidCol()).
• In manual mode, the memory region of tmp must not overlap with dst, src0, or src1.

Ascend 950PR/Ascend 950DT

tmp is accepted by the API but not used by the Ascend 950PR/Ascend 950DT implementation. The Ascend 950PR/Ascend 950DT backend directly uses the vxor vector instruction and does not require temporary tile storage. tmp is retained in the C++ built-in API signature only for API compatibility with Atlas A2/A3 training products/Atlas A2/A3 inference products.

Examples

#include <pto/pto-inst.hpp>

using namespace pto;

void example() {
  using TileDst = Tile<TileType::Vec, uint32_t, 16, 16>;
  using TileSrc0 = Tile<TileType::Vec, uint32_t, 16, 16>;
  using TileSrc1 = Tile<TileType::Vec, uint32_t, 16, 16>;
  using TileTmp = Tile<TileType::Vec, uint32_t, 16, 16>;
  TileDst dst;
  TileSrc0 src0;
  TileSrc1 src1;
  TileTmp tmp;
  TXOR(dst, src0, src1, tmp);
}

ASM Examples

Automatic Mode

# Automatic mode: the compiler/runtime is responsible for resource placement and scheduling.
%dst = pto.txor %src0, %src1 : (!pto.tile<...>, !pto.tile<...>) -> !pto.tile<...>

Manual Mode

# Manual mode: explicitly bind resources first, then issue the instruction.
# Optional (when the instruction contains tile operands):
# pto.tassign %arg0, @tile(0x1000)
# pto.tassign %arg1, @tile(0x2000)
%dst = pto.txor %src0, %src1 : (!pto.tile<...>, !pto.tile<...>) -> !pto.tile<...>

PTO Assembly Form

%dst = txor %src0, %src1 : !pto.tile<...>
# AS Level 2 (DPS)
pto.txor ins(%src0, %src1 : !pto.tile_buf<...>, !pto.tile_buf<...>) outs(%dst : !pto.tile_buf<...>)