ADR 0083: Tile conversion, layout, and partial operations

Context

ADR 0062 recorded a single repository-wide mnemonic audit. This record preserves the accepted decisions for this family as one decision-scoped owner. The former identifiers remain only in legacy_ids and the generated ADR index; current normative meaning is owned by the affected ASL/NDF clauses.

Decisions

Decision 133: `TQUANT` is the encoded single-destination affine quantizer

TQUANT is selected by TEPL carrier Mode=3, Function=10 and executes on the SFU engine. The architectural selector has one Local source Tile, one newly allocated Local destination Tile, and two per-PE scalar inputs. It does not directly expose parameter Tiles or auxiliary result Tiles. Quantization interfaces that require row-varying parameters, shared exponents, maxima, scaling Tiles, or other auxiliary results are compound software lowerings and do not add operands or results to this selector.

The source DataType is exactly FP32. The destination DataType is exactly S8 or U8. BSTART supplies the source DataType and a mandatory B.DATR supplies the destination DataType. Source and destination have the same nonzero valid shape and use row-major layout. When present, B.IOR.RegSrc0 supplies the raw FP32 quantization multiplier and B.IOR.RegSrc1 supplies the zero point in the selected destination integer encoding. RegSrc2 and RegDst are unused and must select zero. Omitting B.IOR supplies the operation defaults multiplier 1.0 and zero point zero. An explicitly encoded all-zero B.IOR instead reads the zero register for both inputs and is illegal because its multiplier is 0.0. Every nondefault multiplier must be finite, positive, and nonzero. The zero point must be in S8 range for an S8 destination and in U8 range for a U8 destination.

For each valid source element x, the unbounded quantized value is x * multiplier + zero_point. B.DATR.RMode selects the rounding rule and defaults to round-to-nearest-even. With Sat=1, the rounded value is clamped to the destination range; with Sat=0, conversion uses the selected integer width's modulo result. Floating exceptional inputs and numeric status follow the numeric conversion profile before the selected saturation rule. Canonicalize and PadValueOrByteId are inapplicable and must be zero. Elements outside the destination valid region remain Null padding.

B.IOS is illegal. Source and destination B.IOT bindings use the same PE_MASK; any subset is legal and mask zero is a strict no-op before GPR or Tile reads, allocation, or faults. The source persists. Complete descriptor, scalar, type, shape, layout, and capacity preflight precedes the source snapshot. Numeric status, destination payload, definedness, padding, and the destination descriptor publish atomically; rejection has no architectural effect.

Decision 134: `TDEQUANT` is the encoded single-destination affine dequantizer

TDEQUANT is selected by TEPL carrier Mode=3, Function=11 and executes on the SFU engine. The architectural selector has one Local source Tile, one newly allocated Local destination Tile, and two per-PE scalar inputs. It does not directly consume row-varying parameter Tiles. Interfaces that use scale or zero-point Tiles are compound software lowerings and do not change this selector's operand arity.

The source DataType is exactly S8 or U8; the destination DataType is exactly FP32. BSTART supplies the source DataType and a mandatory B.DATR supplies destination FP32. Source and destination have the same nonzero valid shape and use row-major layout. When present, B.IOR.RegSrc0 supplies the raw FP32 dequantization multiplier and B.IOR.RegSrc1 supplies the zero point in the source integer encoding. RegSrc2 and RegDst are unused and must select zero. Omitting B.IOR supplies the operation defaults multiplier 1.0 and zero point zero. An explicitly encoded all-zero B.IOR instead reads the zero register and is illegal because its multiplier is 0.0. Every nondefault multiplier must be finite, positive, and nonzero. The zero point must be in the selected source type's range.

For each valid source element q, the destination is the FP32 result of (q - zero_point) * multiplier. B.DATR.RMode selects floating rounding and defaults to round-to-nearest-even. Sat, Canonicalize, and PadValueOrByteId are inapplicable and must be zero. Numeric status follows the floating conversion and multiplication profile. Elements outside the destination valid region remain Null padding.

Decision 135: `TSORT` stably sorts independent row groups and returns indices

TSORT is selected by TEPL carrier Mode=3, Function=12 and executes on the SFU engine. It reads one Local source Tile and atomically creates two distinct Local destinations: a value destination with the source DataType and a U32 index destination. All three Tiles have the same nonzero valid shape and use row-major layout.

Each source row is partitioned from column zero into independent consecutive groups of sort_width elements. The final group contains only its remaining valid elements and never reads padding. Each group is stably sorted by the selected DataType. The value destination receives the reordered values; the index destination receives each value's original zero-based column offset within that group. Equal values preserve source order. descending=0 sorts ascending and descending=1 sorts descending. Numeric values precede NaNs in both directions; NaNs preserve source order. Signaling-NaN observation sets the selected numeric invalid status, and signed zeros compare equal so their source order is preserved.

The exact supported value DataTypes are FP32 and FP16. B.DATR is inapplicable and every field must remain zero. B.DIM LB0 supplies sort_width in 1..64; omission and an encoded zero both select the operation default 32, while nonzero values outside 1..64 are illegal. When present, B.IOR.RegSrc0 supplies descending and must contain exactly zero or one. Omitting B.IOR selects ascending order. RegSrc1, RegSrc2, and RegDst are unused and must select zero.

B.IOS is illegal. All source and destination B.IOT bindings use the same PE_MASK; any subset is legal and mask zero is a strict no-op before reads, allocation, comparison status, or faults. The source persists. Complete descriptor, control, type, shape, layout, capacity, distinct-destination, and definedness preflight precedes the source snapshot. Both destination payloads, definedness, Null padding, numeric status, and descriptors publish as one atomic transaction; rejection has no architectural effect.

Decision 136: `TMRGSORT` stably merges two single-row sorted value streams

TMRGSORT is selected by TEPL carrier Mode=3, Function=13 and executes on the SFU engine. The architectural selector reads exactly two Local source Tiles and creates one newly allocated Local destination. Interfaces that merge three or four lists, merge packed records, return per-list consumption counts, use temporary Tiles, stop on exhaustion, or merge four blocks from one Tile are compound software lowerings and do not add operands or results to this selector.

Each source is a nonempty single-row, row-major, already-sorted value stream. The destination is single-row and row-major, with ValidCol equal to the sum of both source ValidCol values. Source and destination share one DataType. The exact supported DataTypes are FP32 and FP16. descending=0 requires ascending sources and produces an ascending merge; descending=1 requires descending sources and produces a descending merge. If either source stream is not sorted in the selected order, the complete block is illegal before any architectural effect.

Equal values select the left source first. Numeric values precede NaNs in both directions; NaNs preserve their source order and left-source precedence. Signaling-NaN observation sets numeric invalid status, and signed zeros compare equal. Both sources persist.

B.DATR is inapplicable and every field must remain zero. When present, B.IOR.RegSrc0 supplies descending and must contain exactly zero or one. Omitting B.IOR selects ascending order. RegSrc1, RegSrc2, and RegDst are unused and must select zero. B.IOS is illegal. All source and destination B.IOT bindings use the same PE_MASK; any subset is legal and mask zero is a strict no-op before source reads, sortedness checks, allocation, numeric status, or faults.

Complete descriptor, control, type, shape, layout, capacity, definedness, and input-order preflight precedes source snapshots. Destination payload, definedness, Null padding, numeric status, and descriptor publish atomically; rejection has no architectural effect.

Decision 137: `TTRANS` is a logical two-dimensional transpose

TTRANS is selected by TEPL carrier Mode=3, Function=14 and executes on the SFU engine. It reads one Local source Tile and creates one newly allocated Local destination. For every source valid coordinate [r,c], the destination coordinate [c,r] receives the same logical element bit-for-bit. The destination therefore has ValidRow=source.ValidCol and ValidCol=source.ValidRow. The complete valid source rectangle must be defined. The source persists.

Required nonzero B.DIM LB0 supplies destination ValidCol; omitted LB1 supplies destination ValidRow=1; omitted LB2 supplies destination physical Col=ValidCol; physical rows derive from destination capacity, physical columns, and DataType. The resulting destination dimensions must equal the transposed source dimensions before effects. Source and destination use the same DataType. Every assigned Tile DataType except HiF4X2 is supported; globally reserved encodings and HiF4X2 reject before effects.

The source is addressed through its architectural layout. An assigned B.DATR.Layout transformation governs destination physical placement without changing the logical transpose. Every reserved Layout code rejects before effects. DataType, PadValueOrByteId, CMode, RMode, Sat, and Canonicalize are inapplicable and must remain at their operation-default values. Physical destination elements outside the transposed valid rectangle remain Null padding.

TTRANS has no architectural temporary-Tile operand and does not define a separate convolution-layout conversion. A backend scratch Tile or a compound format-conversion interface is a lowering detail and does not change the selector's one-source, one-destination schema. B.IOR and B.IOS are illegal. Source and destination B.IOT bindings use the same PE_MASK; any subset is legal and mask zero is a strict no-op before reads, allocation, or faults.

Complete schema, descriptor, type, dimension, layout, capacity, allocation, mask, and source-definedness preflight precedes the source snapshot. Destination payload, definedness, Null padding, and descriptor publish atomically; rejection has no architectural effect.

Decision 139: `TSCATTER` uses each index as a destination-row selector

TSCATTER is selected by TEPL carrier Mode=3, Function=16 and executes on the SFU engine. It reads one Local value source and one Local index source, then creates one newly allocated Local destination. The two sources have the same nonzero valid shape. The destination has the same ValidCol and a nonzero ValidRow large enough for every selected row. For every source coordinate [r,c], the selected index value k=index[r,c] names a logical destination row and the operation writes destination[k,c]=source[r,c].

The index Tile DataType and value DataType pair is restricted by element width. FP32, S32, and U32 values use S32 or U32 indices. FP16, BF16, S16, and U16 values use S16 or U16 indices. S8 and U8 values also use S16 or U16 indices. A signed index must be nonnegative, and every index must be less than destination.ValidRow. Two source elements must not select the same destination coordinate; duplicate destinations make the complete block illegal before effects.

Required nonzero B.DIM LB0 supplies destination ValidCol; omitted LB1 supplies ValidRow=1; omitted LB2 supplies destination physical Col=ValidCol; physical rows derive from capacity. destination.ValidCol must equal both source ValidCol values, while both sources have equal ValidRow. The complete valid rectangles of both sources must be defined.

An assigned B.DATR.Layout transformation governs destination physical placement without changing row-index semantics. DataType, CMode, RMode, Sat, and Canonicalize are inapplicable. PadValueOrByteId is fixed to encoded zero for this operation: before applying the scatter writes, every physical destination element is initialized to the selected value DataType's positive or integer zero. Thus valid positions not selected by an index and physical padding are defined zero rather than preserved old contents.

B.IOR and B.IOS are illegal. All three B.IOT bindings use the same PE_MASK; any subset is legal and mask zero is a strict no-op before Tile reads, index and duplicate checks, allocation, or faults. Both sources persist. The destination is a renamed result and never reads a previous destination value.

Complete schema, descriptor, type-pair, shape, layout, capacity, index-range, duplicate-destination, definedness, mask, and allocation preflight precedes source snapshots. Zero initialization, scattered payload, definedness, and destination descriptor publish atomically; rejection has no architectural effect.

Decision 138: `TGATHER` uses each index as a source-row selector

TGATHER is selected by TEPL carrier Mode=3, Function=15 and executes on the SFU engine. It reads one Local value source and one Local index source, then creates one newly allocated Local destination. The index source and destination have the same nonzero valid shape. The value source has at least the destination ValidCol columns. For every destination coordinate [r,c], the selected index value k=index[r,c] names a logical source row and the result is destination[r,c]=source[k,c].

The index Tile DataType is exactly S32 or U32. A signed index must be nonnegative, and every index must be less than source.ValidRow. An invalid index makes the complete block illegal before source reads or destination effects; indices never wrap, clamp, or produce an implementation-defined value. The complete index valid rectangle and every selected value-source element must be defined.

The exact value DataTypes are FP32, FP16, S32, S16, U32, and U16. Source and destination use the same value DataType. Required nonzero B.DIM LB0 supplies destination ValidCol; omitted LB1 supplies ValidRow=1; omitted LB2 supplies destination physical Col=ValidCol; physical rows derive from capacity. The dimensions must match the index Tile and fit the source column extent before effects.

An assigned B.DATR.Layout transformation governs destination physical placement without changing row-index semantics. DataType, PadValueOrByteId, CMode, RMode, Sat, and Canonicalize are inapplicable and must remain at their operation-default values. Destination physical elements outside the valid rectangle remain Null padding.

The selector has no architectural mask-pattern or temporary-Tile operand. Mask-pattern gather and scratch-Tile interfaces are compound lowerings and do not change the encoded three-Tile schema. B.IOR and B.IOS are illegal. All three B.IOT bindings use the same PE_MASK; any subset is legal and mask zero is a strict no-op before Tile reads, index checks, allocation, or faults. Both sources persist.

Complete schema, descriptor, type, shape, layout, capacity, index-range, definedness, mask, and allocation preflight precedes source snapshots. Destination payload, definedness, Null padding, and descriptor publish atomically; rejection has no architectural effect.

Decision 140: `TPARTADD` combines two origin-anchored partial rectangles

TPARTADD is selected by TEPL carrier Mode=3, Function=17 and executes on the SFU engine. It reads two Local sources and creates one newly allocated Local destination. All three Tiles use one DataType and row-major layout. Each source valid rectangle is anchored at logical coordinate [0,0] and must fit within the destination valid rectangle. At least one source valid rectangle must equal the complete destination valid rectangle, so every destination coordinate is covered by at least one source.

For each destination valid coordinate, if both sources cover that coordinate, the result is their selected-DataType addition. If exactly one source covers the coordinate, that source element is copied bit-for-bit. No coordinate may be uncovered. Integer addition wraps modulo the selected element width; floating addition, rounding, exceptional values, and numeric status follow the same typed rules as TADD.

The exact supported DataTypes are FP32, FP16, BF16, S32, S16, S8, U32, U16, and U8. Required nonzero B.DIM LB0 supplies destination ValidCol; omitted LB1 supplies ValidRow=1; omitted LB2 supplies physical Col=ValidCol; physical rows derive from capacity. Source physical shapes may differ, but their row-major descriptors and complete valid rectangles must be legal and defined.

B.DATR, B.IOR, and B.IOS are illegal. Destination physical elements outside its valid rectangle remain Null padding. All three B.IOT bindings use the same PE_MASK; any subset is legal and mask zero is a strict no-op before reads, allocation, numeric status, or faults. Both sources persist and may alias each other or the destination because source payloads are snapshotted before any result write.

Complete schema, descriptor, type, shape, coverage, capacity, definedness, mask, and allocation preflight precedes source snapshots. Destination payload, definedness, Null padding, numeric status, and descriptor publish atomically; rejection has no architectural effect.

Decision 142: `TPARTMAX` combines two origin-anchored partial rectangles

TPARTMAX is selected by TEPL carrier Mode=3, Function=19 and executes on the SFU engine. It reads two Local sources and creates one newly allocated Local destination. All three Tiles use one DataType and row-major layout. Each source valid rectangle is anchored at logical coordinate [0,0] and must fit within the destination valid rectangle. At least one source valid rectangle must equal the complete destination valid rectangle, so every destination coordinate is covered by at least one source.

For each destination valid coordinate, if both sources cover that coordinate, the result is their selected-DataType maximum. If exactly one source covers the coordinate, that source element is copied bit-for-bit. No coordinate may be uncovered. Signed integers use signed ordering and unsigned integers use unsigned ordering. Floating one-NaN, two-NaN, signaling-NaN, canonical-NaN, and signed-zero behavior follows TMAX exactly; a mixed-sign zero maximum is positive zero.

Complete schema, descriptor, type, shape, coverage, capacity, source-encoding, definedness, mask, and allocation preflight precedes source snapshots. Destination payload, definedness, Null padding, numeric status, and descriptor publish atomically; rejection has no architectural effect.

Decision 143: `TPARTMIN` combines two origin-anchored partial rectangles

TPARTMIN is selected by TEPL carrier Mode=3, Function=20 and executes on the SFU engine. It reads two Local sources and creates one newly allocated Local destination. All three Tiles use one DataType and row-major layout. Each source valid rectangle is anchored at logical coordinate [0,0] and must fit within the destination valid rectangle. At least one source valid rectangle must equal the complete destination valid rectangle, so every destination coordinate is covered by at least one source.

For each destination valid coordinate, if both sources cover that coordinate, the result is their selected-DataType minimum. If exactly one source covers the coordinate, that source element is copied bit-for-bit. No coordinate may be uncovered. Signed integers use signed ordering and unsigned integers use unsigned ordering. Floating one-NaN, two-NaN, signaling-NaN, canonical-NaN, and signed-zero behavior follows TMIN exactly; a mixed-sign zero minimum is negative zero.

Decision 141: `TPARTMUL` combines two origin-anchored partial rectangles

TPARTMUL is selected by TEPL carrier Mode=3, Function=18 and executes on the SFU engine. It reads two Local sources and creates one newly allocated Local destination. All three Tiles use one DataType and row-major layout. Each source valid rectangle is anchored at logical coordinate [0,0] and must fit within the destination valid rectangle. At least one source valid rectangle must equal the complete destination valid rectangle, so every destination coordinate is covered by at least one source.

For each destination valid coordinate, if both sources cover that coordinate, the result is their selected-DataType multiplication. If exactly one source covers the coordinate, that source element is copied bit-for-bit. No coordinate may be uncovered. Integer multiplication wraps modulo the selected element width; floating multiplication, rounding, exceptional values, signed zeros, and numeric status follow the same typed rules as TMUL.

Context​

Decisions​

Decision 133: TQUANT is the encoded single-destination affine quantizer​

Decision 134: TDEQUANT is the encoded single-destination affine dequantizer​

Decision 135: TSORT stably sorts independent row groups and returns indices​

Decision 136: TMRGSORT stably merges two single-row sorted value streams​

Decision 137: TTRANS is a logical two-dimensional transpose​

Decision 139: TSCATTER uses each index as a destination-row selector​

Decision 138: TGATHER uses each index as a source-row selector​

Decision 140: TPARTADD combines two origin-anchored partial rectangles​

Decision 142: TPARTMAX combines two origin-anchored partial rectangles​

Decision 143: TPARTMIN combines two origin-anchored partial rectangles​

Decision 141: TPARTMUL combines two origin-anchored partial rectangles​