PTO ISA Manual And Reference¶

This directory is the canonical PTO ISA tree. It combines the architecture manual, the instruction set guides, the instruction set contracts, and the exact instruction-reference groupings in one place.

Textual Assembly Inside PTO ISA¶

Synchronization¶

• TSYNC - Synchronize PTO execution (wait on events or insert a per-op pipeline barrier).
• SYNCALL - Cross-core synchronization barrier (hardware FFTS or software GM polling).

Manual / Resource Binding¶

• TASSIGN - Bind a Tile object to an implementation-defined on-chip address (manual placement).
• SETFMATRIX - Set FMATRIX register(s) for IMG2COL-like ops.
• SET_IMG2COL_RPT - Set IMG2COL repeat metadata from an IMG2COL configuration tile.
• SET_IMG2COL_PADDING - Set IMG2COL padding metadata from an IMG2COL configuration tile.
• SET_QUANT_SCALAR - Set the scalar quantization parameter for subsequent TPUSh operations.
• SET_QUANT_VECTOR - Set the vector quantization parameter from a Scaling tile for subsequent TPUSH operations.

Elementwise (Tile-Tile)¶

• TADD - Elementwise add of two tiles.
• TABS - Elementwise absolute value of a tile.
• TAND - Elementwise bitwise AND of two tiles.
• TOR - Elementwise bitwise OR of two tiles.
• TSUB - Elementwise subtract of two tiles.
• TMUL - Elementwise multiply of two tiles.
• TMIN - Elementwise minimum of two tiles.
• TMAX - Elementwise maximum of two tiles.
• TCMP - Compare two tiles and write a packed predicate mask.
• TDIV - Elementwise division of two tiles.
• TSHL - Elementwise shift-left of two tiles.
• TSHR - Elementwise shift-right of two tiles.
• TXOR - Elementwise bitwise XOR of two tiles.
• TLOG - Elementwise natural logarithm of a tile.
• TRECIP - Elementwise reciprocal of a tile.
• TPRELU - Elementwise PReLU (parametric ReLU) with a per-element slope tile.
• TADDC - Elementwise ternary add: src0 + src1 + src2.
• TSUBC - Elementwise ternary op: src0 - src1 + src2.
• TCVT - Elementwise type conversion with a specified rounding mode.
• TSEL - Select between two tiles using a mask tile (per-element selection).
• TRSQRT - Elementwise reciprocal square root.
• TSQRT - Elementwise square root.
• TEXP - Elementwise exponential.
• TNOT - Elementwise bitwise NOT of a tile.
• TRELU - Elementwise ReLU of a tile.
• TNEG - Elementwise negation of a tile.
• TREM - Elementwise remainder of two tiles.
• TFMOD - Elementwise fmod of two tiles.
• TPOW - Elementwise power of two tiles.
• TMULADDDST - Elementwise ternary op: src0 * src1 + dst.
• TSUBRELU - Elementwise subtract then ReLU of two tiles.
• TFUSEDMULADD - Elementwise ternary op: src0 * dst + src1.
• TFUSEDMULADDRELU - Elementwise ternary op: ReLU(src0 * dst + src1).

Tile-Scalar / Tile-Immediate¶

• TEXPANDS - Broadcast a scalar into a destination tile.
• TCMPS - Compare a tile against a scalar and write per-element comparison results.
• TSELS - Select between source tile and scalar using a mask tile (per-element selection for source tile).
• TMINS - Elementwise minimum of a tile and a scalar.
• TADDS - Elementwise add a scalar to a tile.
• TSUBS - Elementwise subtract a scalar from a tile.
• TAXPY - In-place scaled accumulation (AXPY): dst = scalar * src0 + dst.
• TDIVS - Elementwise division with a scalar (tile/scalar or scalar/tile).
• TMULS - Elementwise multiply a tile by a scalar.
• TFMODS - Elementwise remainder with a scalar: fmod(src, scalar).
• TREMS - Elementwise remainder with a scalar: remainder(src, scalar).
• TMAXS - Elementwise max of a tile and a scalar: max(src, scalar).
• TANDS - Elementwise bitwise AND of a tile and a scalar.
• TORS - Elementwise bitwise OR of a tile and a scalar.
• TSHLS - Elementwise shift-left a tile by a scalar.
• TSHRS - Elementwise shift-right a tile by a scalar.
• TXORS - Elementwise bitwise XOR of a tile and a scalar.
• TLRELU - Leaky ReLU with a scalar slope.
• TADDSC - Elementwise fused add with scalar and a second tile: src0 + scalar + src1.
• TSUBSC - Elementwise fused op: src0 - scalar + src1.
• TPOWS - Elementwise power of a tile by a scalar.

Axis Reduce / Expand¶

• TROWSUM - Reduce each row by summing across columns.
• TROWPROD - Reduce each row by multiplying across columns.
• TCOLSUM - Reduce each column by summing across rows.
• TCOLPROD - Reduce each column by multiplying across rows.
• TCOLMAX - Reduce each column by taking the maximum across rows.
• TROWMAX - Reduce each row by taking the maximum across columns.
• TROWMIN - Reduce each row by taking the minimum across columns.
• TROWARGMAX - Get the column index of the maximum element for each row.
• TROWARGMIN - Get the column index of the minimum element for each row.
• TCOLARGMAX - Get the row index /(value and row index) of the maximum element for each column.
• TCOLARGMIN - Get the row index /(value and row index) of the minimum element for each column.
• TROWEXPAND - Broadcast the first element of each source row across the destination row.
• TROWEXPANDDIV - Row-wise broadcast divide: divide each row of src0 by a per-row scalar vector src1.
• TROWEXPANDMUL - Row-wise broadcast multiply: multiply each row of src0 by a per-row scalar vector src1.
• TROWEXPANDSUB - Row-wise broadcast subtract: subtract a per-row scalar vector src1 from each row of src0.
• TROWEXPANDADD - Row-wise broadcast add: add a per-row scalar vector.
• TROWEXPANDMAX - Row-wise broadcast max with a per-row scalar vector.
• TROWEXPANDMIN - Row-wise broadcast min with a per-row scalar vector.
• TROWEXPANDEXPDIF - Row-wise exp-diff: compute exp(src0 - src1) with per-row scalars.
• TCOLMIN - Reduce each column by taking the minimum across rows.
• TCOLEXPAND - Broadcast the first element of each source column across the destination column.
• TCOLEXPANDDIV - Column-wise broadcast divide: divide each column by a per-column scalar vector.
• TCOLEXPANDMUL - Column-wise broadcast multiply: multiply each column by a per-column scalar vector.
• TCOLEXPANDADD - Column-wise broadcast add with per-column scalar vector.
• TCOLEXPANDMAX - Column-wise broadcast max with per-column scalar vector.
• TCOLEXPANDMIN - Column-wise broadcast min with per-column scalar vector.
• TCOLEXPANDSUB - Column-wise broadcast subtract: subtract a per-column scalar vector from each column.
• TCOLEXPANDEXPDIF - Column-wise exp-diff: compute exp(src0 - src1) with per-column scalars.

Memory (GM <-> Tile)¶

• TLOAD - Load data from a GlobalTensor (GM) into a Tile.
• TPREFETCH - Prefetch data from global memory into a tile-local cache/buffer (hint).
• TPREFETCH_ASYNC - Asynchronously prefetch a GlobalTensor region from GM into L2 cache via SDMA CMO.
• TSTORE - Store data from a Tile into a GlobalTensor (GM), optionally using atomic write or quantization parameters.
• TSTORE_FP - Store an accumulator tile into global memory using a scaling (fp) tile for vector quantization parameters.
• MGATHER - Gather-load elements from global memory into a tile using per-element indices.

• MSCATTER - Scatter-store elements from a tile into global memory using per-element indices.

• Programming model

• Machine model
• Syntax and operands
• Type system
• Location intent and legality
• Memory model

Data Movement / Layout¶

• TEXTRACT - Extract a sub-tile from a source tile.
• TEXTRACT_FP - Extract with fp/scaling tile (vector-quantization parameters).
• TIMG2COL - Image-to-column transform for convolution-like workloads.
• TINSERT - Insert a sub-tile into a destination tile at an (indexRow, indexCol) offset.
• TINSERT_FP - Insert with fp/scaling tile (vector-quantization parameters).
• TFILLPAD - Copy+pad a tile outside the valid region with a compile-time pad value.
• TFILLPAD_INPLACE - In-place fill/pad variant.
• TFILLPAD_EXPAND - Fill/pad while allowing dst to be larger than src.
• TMOV - Move/copy between tiles, optionally applying implementation-defined conversion modes.
• TMOV_FP - Move/convert from an accumulator tile into a destination tile, using a scaling (fp) tile for vector quantization parameters.
• TRESHAPE - Reinterpret a tile as another tile type/shape while preserving the underlying bytes.
• TTRANS - Transpose with an implementation-defined temporary tile.
• TSUBVIEW - Reinterpret a tile as a subtile of another tile.
• TGET_SCALE_ADDR - Bind the on-chip address of output tile to a scaled factor of that of input tile.
• TCONCAT - Concatenate two tiles horizontally along the column dimension.
• TInterleave - Interleave two source tiles into an alternating even/odd element stream, split into two destination halves.
• TDeInterleave - De-interleave source tiles back into even-position and odd-position element streams (inverse of TInterleave).
• TPAIRREDUCESUM - Pair-reduction sum: add every 2 adjacent elements and write results to the lower half of dst.

Complex¶

• TPRINT - Debug/print elements from a tile (implementation-defined).
• TMRGSORT - Merge sort for multiple sorted lists (implementation-defined element format and layout).
• TSORT32 - Sort each 32-element block of src together with the corresponding indices from idx, and write the sorted value-index pairs into dst.
• TGATHER - Gather/select elements using either an index tile or a compile-time mask pattern.
• TCI - Generate a contiguous integer sequence into a destination tile.
• TTRI - Generate a triangular (lower/upper) mask tile.
• TRANDOM - Generates random numbers in the destination tile using a counter-based cipher algorithm.
• TPARTADD - Partial elementwise add with implementation-defined handling of mismatched valid regions.
• TPARTMUL - Partial elementwise multiply with implementation-defined handling of mismatched valid regions.
• TPARTMAX - Partial elementwise max with implementation-defined handling of mismatched valid regions.
• TPARTMIN - Partial elementwise min with implementation-defined handling of mismatched valid regions.
• TPARTARGMAX - Partial elementwise max selection returning corresponding index (argmax), with implementation-defined handling of mismatched valid regions.
• TPARTARGMIN - Partial elementwise min selection returning corresponding index (argmin), with implementation-defined handling of mismatched valid regions.
• TGATHERB - Gather elements using byte offsets.
• TSCATTER - Scatter rows of a source tile into a destination tile using per-element row indices.
• TQUANT - Quantize a tile (e.g. FP32 to FP8) producing exponent/scaling/max outputs.
• TDEQUANT - Affine dequantization of a quantized tile (S8/S16 -> FP32): dst = (src - offset) * scale.
• THISTOGRAM - Per-byte histogram (256 bins) over a selected byte of each source element, with optional cascaded upper-byte filtering; the radix-sort bucket-count primitive.

Cross-core Communication¶

• TALLOC - Allocate a TPipe FIFO slot as a GlobalTensor view.
• TPUSH - Push a producer tile into a TPipe FIFO for Cube-Vector communication.
• TPOP - Pop a consumer tile from a TPipe FIFO for Cube-Vector communication.
• TFREE - Release FIFO space for a TPipe entry; no-op for TileData TPOP flow.

Communication¶

Cross-Core Communication¶

• TALLOC - Allocate a TPipe FIFO slot as a GlobalTensor view.
• TPUSH - Push a producer tile into a TPipe FIFO for Cube-Vector communication.
• TPOP - Pop a consumer tile from a TPipe FIFO for Cube-Vector communication.
• TFREE - Release FIFO space for a TPipe entry; no-op for TileData TPOP flow.

The grouped instruction set trees under tile/, vector/, scalar/, comm/, and system/ are the canonical PTO ISA paths.

• docs/isa/tile/ops/
• docs/isa/vector/ops/
• docs/isa/scalar/ops/
• docs/isa/comm/
• docs/isa/system/ops/