Repository navigation
Tags
Tags
release: v0.19.0 — enable swizzle_assignment (gogpu#90) * docs: add star history chart (starhistory.io) * fix(lint): remove dead nil checks on zeroLiteralForScalar (SA4023) zeroLiteralForScalar always returns a non-nil LiteralValue (all scalar kinds covered + default case). The nil checks were unreachable. * fix(lint): remove remaining dead nil check in MatrixType path (SA4023) * feat(wgsl): implement swizzle_assignment extension (gogpu#89, ADR-003) enable swizzle_assignment; allows multi-component swizzle on assignment LHS: result.rgb = c.rgb; result.rgb += c.rgb * c.a; Lowerer decomposes to load → per-component extract → construct → store (Dawn pattern). All 5 backends work without changes. Spec-compliant per W3C gpuweb/gpuweb#5268 (merged 2026-08-25). Ahead of Rust naga. 22 test cases + 2 integration tests. Without extension: clear error. * test(wgsl): all 11 Dawn swizzle assignment patterns + false positive fix Fix parseSwizzleMap treating struct member names (e.g. 'arr') as RGBA swizzle patterns by rejecting duplicate indices. Restores pointers.wgsl SPIR-V validation (172/172). 48 swizzle tests covering all Dawn patterns: single/multi/chained/indexed (constant + dynamic), compound assignment, vec2/vec3/vec4, i32/u32/f32. * docs: prepare release v0.19.0 * fix(lint): remove unused nolint directive in darwin test
release: v0.18.0 (gogpu#86) * feat(wgsl): capability-based scalar width validation (gogpu#85) Add Capabilities type (ir.CapFloat64, ir.CapShaderInt64, ir.CapShaderFloat16) matching Rust naga valid::Capabilities bitflags. WGSL lowerer now rejects f64/i64/u64/f16 unless the caller enables the corresponding capability. LowerWithSource remains permissive (CapAll) for backward compatibility. LowerWithCapabilities provides strict mode for runtime use. CompileOptions.Capabilities defaults to 0 (strict). 32 test cases covering reject/accept per type, struct fields, function params, constructors, vec/mat aliases, atomics, CapAll, combined caps. * refactor: move capability validation to unified ir.Validator (ADR-002) Capability checks now live in ir.ValidateWithCapabilities — the single gate between lowerer and backends. Matches Rust naga architecture: lowerer is permissive, validator enforces capabilities, backends trust validated IR. Removed checkScalarWidth from WGSL lowerer, removed LowerWithCapabilities and LowerWithWarningsAndCapabilities. Validator checks Scalar, Vector, Matrix, and Atomic types for width-capability compliance. +84/-450 lines — less code, better architecture, zero duplication. * docs: prepare release v0.18.0
release: v0.17.16 (gogpu#84) * fix(msl): parenthesize binary expressions inside swizzle (gogpu#83) writeSwizzle now wraps binary/select sub-expressions in parentheses, matching Rust naga is_scoped=false semantics. Without this, C++ member access (.) bound tighter than arithmetic operators, emitting mat * vec.xyz instead of (mat * vec).xyz — a type error rejected by Metal compiler. Discovered via g3d standard.wgsl on Apple M5. * test(msl): add i64 rejection case for 64-bit atomic policy Covers signed branch explicitly — atomicAdd on atomic<i64> must be rejected the same as u64. Width == 8 handles both uniformly, but the test makes the contract visible for future readers. * docs: prepare release v0.17.16
fix(dxil): PHI node ordering + coverage waves 3-4 (gogpu#75) * fix(dxil): PHI nodes grouped at top of basic block — path_count.wgsl now VALID mem2reg Phase B phi expressions had high handles, emitted after low-handle regular expressions in merged StmtEmit runs. splitPhiRanges separates phi ranges and emits them first. Fixes 'PHI nodes not grouped at top of basic block' IDxcValidator error for Vello path_count.wgsl compute shader. * docs: update for v0.17.13 — PHI ordering fix, coverage waves 3-4
refactor: ARCH-001 internal packages + 12/18 coverage ≥80% + 13 panic… …s→errors (gogpu#72) * refactor(glsl): move implementation to internal/codegen — DXIL pattern (ARCH-001) All codegen moved to glsl/internal/codegen/ via git mv. Thin public wrapper at glsl/glsl.go with type aliases (zero overhead). Public API unchanged: Compile, Options, DefaultOptions, TranslationInfo, BindingMapKey. Writer and Features types now internal. 0 lint issues, all tests pass. * refactor(msl): move implementation to internal/codegen — DXIL pattern (ARCH-001) All codegen moved to msl/internal/codegen/ via git mv. Thin public wrapper at msl/msl.go with type aliases (zero overhead). Public API unchanged: Compile, CompileWithPipeline, Options, DefaultOptions, BoundsCheckPolicies, VertexFormat constants. 0 lint issues, all tests pass. * refactor(hlsl): move implementation to internal/codegen — DXIL pattern (ARCH-001) All codegen moved to hlsl/internal/codegen/ via git mv. Shared namer logic (NeedsTrailingUnderscore, keyword maps) moved to internal/backend/ to avoid circular imports — both hlsl/internal/codegen and dxil/internal/emit import from shared location. Public API unchanged via type aliases. 0 lint issues, all tests pass including DXIL. * refactor(spirv): move implementation to internal/codegen — DXIL pattern (ARCH-001) All codegen moved to spirv/internal/codegen/ via git mv. Thin public wrapper at spirv/spirv.go with type aliases. Public API unchanged: Backend, Options, DefaultOptions, ModuleBuilder, capabilities, opcodes. SPIR-V val 172/172 (100%). 0 lint issues, all tests pass. * refactor: extract shared textutil.IndentWriter — DRY across glsl/hlsl/msl (ARCH-001 Phase 0) internal/textutil/textutil.go with IndentWriter (WriteLine, WriteIndent, PushIndent, PopIndent). Embedded in all 3 text backend Writers. Eliminates 3× duplication of indent/output methods. 5 unit tests. 0 lint issues. * refactor: replace type aliases with real types in public API — DXIL pattern (ARCH-001) glsl: 8/8 aliases → real types + toCodegenOptions/fromCodegenTranslationInfo msl: 22/22 aliases → real types + conversion functions hlsl: 15/16 aliases → real types (Writer stays alias — implementation type) spirv: 4/30 aliases → real types (implementation types stay — tightly coupled) go doc now shows full struct definitions. 0 lint issues, all tests pass. * refactor(wgsl): move to internal/parser + internal/lower — DXIL pattern (ARCH-001 Phase 2a) parser/ contains lexer, token, parser, ast, deporder, errors. lower/ contains the 16K LOC lowerer (imports parser for AST types). Public API: REAL types (Module, Lexer, Tokens, Parser, ParseError, Span). No type aliases. go doc shows full struct definitions. 0 lint issues, all tests pass. * refactor(ir): move TypeRegistry to internal/registry — ARCH-001 Phase 2b Registry moved to module-level internal/registry/ (not ir/internal/) because wgsl/internal/lower/ needs access and Go internal visibility restricts ir/internal/ to ir/ subtree only. compact.go and validate.go stay in ir/ — circular import prevents moving. 0 lint issues. * test(glsl): add 207 tests — coverage 57%→74.3% (COV-002) 5 test files, 5139 LOC. Covers: compile integration (100 tests), statements (22), writer utils (24), features (33), types (28). Table-driven, named subtests, real WGSL→GLSL output verification. * test(msl): add 34 tests — coverage 38%→41.8% (COV-001 wave 1) 1341 LOC. Covers: Version.Less, BoundsCheckPolicies, blockEndsWithReturn, builtinInputAttribute (19 builtins), resolveInterpolation, compile integration (workgroup init, dual-source blend, storage buffers, etc.). * test(glsl): wave 2 — coverage 74.3%→80.4%, 67 targeted tests (COV-002) writeHelperFunctions 12%→100%, resolveTypeInner 18%→tested, scanBlockForStages 16%→91%, tryConstEvalUnary 23%→100%, writeUniformBlock 28%→improved, plus writeCallResult, writeAtomicResult, formatConstResult, computeBlockNames, expandBoolVectorOp all from 0%. * test(msl): wave 2 — coverage 41.8%→56.6%, 748 tests across 4 files (COV-001) Integration tests via compileWGSL: math functions, expressions, control flow, textures, atomics, pipeline constants, vertex pulling, bounds checking. 3919 LOC new tests. Remaining gap: ray queries, subgroups, external textures — specialized WGSL extensions. * test(backend): coverage 35.1%→97.1% — namer, interface ordering, sig packing hlsl_namer_test: EndsWithDigit, IsASCIIAlphanumeric, NeedsTrailingUnderscore (44 cases). interface_order_test: MemberInterfaceLess, SortedArgIndices, SortFlatBindings. sig_pack_binding_test: componentDimensions, SigElementInfoForBinding (13 cases), PackStructMembers, PackSignatureElements edge cases. * test(dxil/module): coverage 52%→100% — types, constants, serialization, float16 60+ tests: type system dedup, instruction constructors, aggregate/undef constants, global variables, float32ToF16Bits (12 cases), full serialization coverage (all instruction kinds, branches, phi, atomics, GEP, value symtab, function decls). Enterprise quality. * test(dxil/viewid): coverage 49.9%→83.0% — dataflow, taint, walker analysis_test: scalarSet, componentTaint, giveUp fallback, expandTaint. expressions_test: computeTaint for 15+ expression kinds, merge ops. walker_test: walkStatement (if/switch/loop/block), handleStore, rootLocalVar, resolveFlatRange. All tests verify actual scalar dataflow. * test(ir): coverage 70.1%→81.1% — overrides, resolve, inline, validation process_overrides: literalToFloat, evalBinary/Unary, cloneModule, filterEmitsInBlock, overrideRemapExprHandles (18 kinds). resolve: Load/Access/Splay/Swizzle type resolution, modf/frexp struct finding, atomicResult, workGroupUniformLoad. inline: blockContainsReturn, shouldAliasArgType. Core: StorageFormat, TypeSize, Validate errors. * test(dxil/passes): DCE 88.6%, mem2reg 92.9%, SROA 83.0% — all above 80% DCE: 27 tests — dead loop, side-effect roots, visitExprHandles (14 kinds), stmtHasSideEffects (22 types), shared consumer tracking. mem2reg: 17 tests — Phase B phi, loop disqualification, collectStores, classifyStatements, promoteBlocks recursion, isBoolLocal. SROA: 19 tests — nested stores, resolveStructLocal, allMembersDecomposable, classify with control flow recursion, Compose-store. * test(hlsl): coverage 32.1%→62.3%, 1308 tests across 5 files Integration: entry points, math, textures, storage buffers, derivatives. Unit: binaryOpStr, scalarKindToHLSL, continueCtx, mathFunction dispatch. Types: hlslTypeSize, isDynamicallySized, struct padding, matCx2. Advanced: 8 storage texture formats, interpolation, force loop bounding. Remaining gap: ray tracing, subgroups, external textures — specialized IR. * test(wgsl): parser 62%→83.8%, lower 31%→54.5% Parser: overrides, bitcast, increment/decrement, 11 assign ops, all type specs (atomic, sampler, 16 textures), DependencyOrder (12 cases), error recovery, Token.String, SourceError formatting, AST Pos() (33 nodes). Lower: overrides, globals, control flow, 30+ texture ops, atomics (10), barriers, constant folding (40+ math builtins), type constructors, abstract concretization, entry points, error cases, interpolation. * test(spirv): coverage 70.4%→75.4%, 107 tests StorageFormatToImageFormat (36 formats), float32ToF16Bits edge cases, ModuleBuilder methods (12), atomicOpcode (12 ops), 50+ WGSL compilation tests with SPIR-V binary assertions (OpSelect, OpKill, derivatives, pack/unpack polyfills, interpolation, barriers, textures, builtins). Remaining gap: subgroups, ray queries — WGSL extensions not supported. * fix(glsl,msl): panic→error, unsupported stage validation GLSL: exitLoop/exitSwitch return error instead of panic. writeEntryPoint rejects StageMesh/StageTask with clear error. 6 unsupported feature tests with hand-crafted IR (ray query, mesh/task shader, binding array). MSL: writeEntryPoint rejects unsupported stages. 2 unsupported tests. * fix(hlsl,spirv): 11 panics→errors — enterprise error handling HLSL: exitLoop/exitSwitch return error instead of panic (2 panics). SPIR-V: emitScalarType, addressSpaceToStorageClass, emitScalarConstant, resolveScalarType, findTypeHandleByID, resolveTypeResolution, emitInlineType + cascading caller updates (~200 sites). 8 error-path tests. 0 panics remaining in hlsl, 4 remaining in spirv ray_query (deep signature cascade, separate task). * docs: update for v0.17.12 — ARCH-001 refactoring, 12/18 packages ≥80% coverage, 13 panics→errors * fix: dead code cleanup + nilerr fixes — 11 unused functions removed, 2 swallowed errors fixed Removed 11 unused functions from wgsl/internal/lower/lower.go. Fixed 5 unused params (prefix _). Fixed 3 unused returns. Fixed 2 nilerr in spirv/backend.go (silently swallowing errors). 0 unused/unparam/nilerr lint issues. * style: fix gofmt in module_test.go, backend tests * fix: update .golangci.yml paths for internal/ package structure Exclusion paths updated after git mv: wgsl/parser.go → wgsl/internal/parser/, wgsl/lower.go → wgsl/internal/lower/, msl/ → msl/internal/codegen/, hlsl/ → hlsl/internal/codegen/, etc. Added internal/registry and internal/backend/hlsl_keywords exclusions. 0 lint issues.
fix(dxil): gg 61/61 — array-of-vector flattening, fine.wgsl fixed (BU… …G-DXIL-041) (gogpu#71) * fix(wgsl): validate math builtin argument count — SIGSEGV → error (gogpu#170) Add mathMinArgs() that maps MathFunction to required argument count. dot/cross/distance/reflect/step/pow/atan2 require 2, mix/smoothstep/ clamp/fma require 3. Catches dot(v) with 1 arg at lowering time instead of crashing in SPIR-V backend with nil pointer dereference. Fixes gogpu/wgpu#170 * fix(dxil): array-of-vector flattening — GEP scaling + multi-scalar load (BUG-DXIL-041) fine.wgsl "Invalid record" caused by 3 bugs in flattened array<vec4,N>: (1) GEP index not scaled by vecWidth, (2) single-scalar load instead of vec4 load, (3) garbage component IDs from missing pendingComponents. gg production: 58/59 → 61/61 (0 failures). Fixes BUG-DXIL-041. * docs: update public docs — gg 61/61 (100%), fine.wgsl fixed, 105 diff=0 * fix: resolve all golangci-lint issues — goconst constants, nolint cleanup, tmp/ exclusion Extract constants: errExpectedParameterName/MemberName/VariableName/Type (parser.go), semPosition/semTarget (semantic.go), hlslFloat (resources.go), samplerHeapName/comparisonSamplerHeapName (sampler_heap.go), outcomeSkip (corpus.go). Remove 4 unused //nolint:gosec. Exclude wgsl/lexer.go and wgsl/token.go keyword maps from goconst. Exclude tmp/ from lint.
feat(dxil): 105 golden diff=0, 55.1% parity — ViewID, workgroup decom… …p, scheduling infra (gogpu#70) * feat(dxil): alignment encoding fix, bitcast constant fold, TBAA normalizer — +47 lines parity Alignment: log2(bytes)+1 for store/load/alloca/globalvar. Workgroup vars get 8-byte alignment for i64/f64. Bitcast constant fold: float constants in raw buffer stores folded to i32 literals via Float32bits. Binary constant folding infrastructure. TBAA metadata stripped. 104 diff=0, 54.7% parity (was 54.5%). * feat(dxil): constant fold int-to-float casts, mul-to-shl in CBV/UAV — +33 lines parity tryFoldNumericCast for sitofp/uitofp const folding. addMulOrShlInstr for CBV/UAV byte-offset stride. 104 diff=0, 54.8% parity (was 54.7%). Remaining 78 shaders need architectural passes. * feat(dxil): instruction scheduling infrastructure — eval-right-first, resource read detection emitBinaryOperands evaluates right operand first when it leads to resource read. exprLeadsToResourceRead walks Load/AccessIndex/GlobalVariable chains. leafEmitPriority classifies reassociate leaves. No metric change yet — blocked by StmtEmit range boundaries. Fix: merge consecutive emit ranges within same basic block (future work). * feat(dxil): merge consecutive StmtEmit ranges for cross-range reassociation emitBlock detects adjacent StmtEmit statements and processes them via emitMergedStmtEmit with gap-aware membership set. Enables reassociate pass to see chains spanning emit range boundaries. +4 lines parity. Full instruction scheduling needs scheduler pass for multi-use intermediates. * feat(dxil): ViewID ExprAlias + ExprPhi taint — precise post-mem2reg dataflow, 105 diff=0 ViewID dataflow now handles ExprAlias (forward taint) and ExprPhi (union incoming edges) instead of falling back to conservative all-inputs→all-outputs. Fixes per-scalar tracking after mem2reg SSA promotion. 105 diff=0 (was 104), 55.1% parity (was 54.8%). Newly diff=0: interface/vertex_two_structs. * feat(dxil): workgroup struct decomposition — per-member globals with MSVC names Struct-typed workgroup globals decomposed into N separate addrspace(3) globals with MSVC-mangled member suffix names matching DXC. Linkage override for external decomposed members. +13 lines parity. 105 diff=0, 55.1% parity. * docs: update public docs — 105 golden diff=0, 55.1% parity
feat(dxil): 104 golden diff=0, 54.5% parity — loadInput DCE, Quantize… …F16, sigId fixes (gogpu#69) * feat(dxil): zero-store promotion, cast elimination, strength reduction — +2 golden diff=0 Zero-store local promotion: unassigned struct members resolve to zero. Same-type integer cast elimination: bitcast i32↔i32 is no-op. Sub→add canonicalization: sub X,C → add X,-C for positive constants. Mul→shl strength reduction: mul X,2^N → shl X,N. Normalizer: strip nsw/nuw/exact integer flags, fast on fcmp. 96 diff=0 (was 94), 48.6% parity (was 48.1%), +125 lines. * feat(dxil): per-member loadInput DCE via backwards reachability — +2 golden diff=0 usedStructMembers() with computeLiveExprs() backwards analysis eliminates dx.op.loadInput calls for struct input members never consumed by shader body. Handles post-inlining patterns where intermediate expressions fall outside StmtEmit ranges. 98 diff=0 (was 96), 49.6% parity (was 48.6%), +288 lines. Newly diff=0: skybox/fs_main, msl-vpt-formats-x1/render_vertex. * feat(dxil): input sigId fix, ViewID StartCol, cross-arg ordering — 100 golden diff=0 Three bugs: (1) sigId used register row instead of element index — broke packed inputs sharing same row. (2) ViewID inScalarToComp ignored StartCol offset. (3) Cross-argument struct inputs emitted in declaration order, DXC emits reverse signature order. 100 diff=0 (was 98), 49.6% parity. Newly diff=0: unconsumed_vertex_outputs_frag/fs_main, mesh-shader/fs_main. * feat(dxil): QuantizeF16 legacy ops, Int64 flag from bitcode, MSVC groupshared names — 103 golden diff=0 QuantizeF16 via dx.op.legacyF32ToF16/F16ToF32 (not fptrunc/fpext) — eliminates NativeLowPrecision flag, fixes 13-shader header cascade. Int64 flag scanned from emitted module, not IR types. AtomicInt64OnHeapResource flag for 64-bit atomics on non-workgroup resources. MSVC name decoration for workgroup variables. 103 diff=0 (was 100), 54.3% parity (was 49.6%). * feat(dxil): createHandle legacy, SM version normalization — 104 golden diff=0 Always use createHandle (opcode 57), removed unused createHandleFromBinding path + 9 dead functions. SM version normalized in golden comparison. AtomicInt64OnHeapResource flag test. 104 diff=0 (was 103), 54.5% parity. * docs: update public docs — 104 golden diff=0, 54.5% parity * style: fix gofmt in opcodes.go and reassociate_test.go
PreviousNext