pub const SIEVE_WGSL: &str = "// GPU/NPU Sieve \u{2014} portable bitmask-output version.\n// Splits each 64-bit Quin field into lo/hi u32 pairs (no shader-int64 required,\n// works on Vulkan, Metal, and DX12 without the shader-int64 capability).\n// Output: dense 27-u32 bitmask \u{2014} one bit per Quin, 850 Quins max per block.\n// Core 2 decodes the bitmask on the CPU via trailing-zero bit-scan.\n\nstruct Quin {\n subject_lo: u32, subject_hi: u32,\n predicate_lo: u32, predicate_hi: u32,\n object_lo: u32, object_hi: u32,\n context_lo: u32, context_hi: u32,\n meta_lo: u32, meta_hi: u32,\n parity_lo: u32, parity_hi: u32,\n}\n\nstruct FilterMask {\n lo: u32,\n hi: u32,\n}\n\n// Binding 0: flat Quin array (variable length \u{2014} one full block = 850 Quins)\n@group(0) @binding(0) var<storage, read> quins: array<Quin>;\n// Binding 1: output bitmask \u{2014} ceil(850 / 32) = 27 u32 words, atomically written\n@group(0) @binding(1) var<storage, read_write> out_bitmask: array<atomic<u32>, 27>;\n// Binding 2: 64-bit filter mask split into lo/hi u32 (matches Rust FilterMask64)\n@group(0) @binding(2) var<uniform> sieve_filter: FilterMask;\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) global_id: vec3<u32>) {\n let index = global_id.x;\n if index >= arrayLength(&quins) { return; }\n\n let meta_lo = quins[index].meta_lo;\n let meta_hi = quins[index].meta_hi;\n\n // Both halves of the 64-bit metadata field must satisfy the mask\n if (meta_lo & sieve_filter.lo) == sieve_filter.lo && (meta_hi & sieve_filter.hi) == sieve_filter.hi {\n let bucket = index / 32u;\n let bit_shift = index % 32u;\n atomicOr(&out_bitmask[bucket], 1u << bit_shift);\n }\n}\n";Expand description
Legacy 27-word bitmask sieve compute shader retained for platform::npu_ffi compatibility.