fix: Manual float16 conversions to extend platform support

[iwoplaza]

React Native (Hermes) unfortunately doesn't yet support Float16Array natively, this PR uses the existing toHalfBits and fromHalfBits instead.

I also fixed faulty behavior of toHalfBits in some edge cases and added tests validation it behaves like Float16Array would.

[github-actions]

pkg.pr.new

packages
Ready to be installed by your favorite package manager ⬇️

https://pkg.pr.new/software-mansion/TypeGPU/eslint-plugin-typegpu@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/tgpu-gen@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/tinyest-for-wgsl@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/typegpu@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/@typegpu/cli@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/@typegpu/color@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/@typegpu/gl@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/@typegpu/noise@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/@typegpu/radiance-cascades@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/@typegpu/react@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/@typegpu/sdf@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/@typegpu/three@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

https://pkg.pr.new/software-mansion/TypeGPU/unplugin-typegpu@e7eb8d49956a2c64b843ac76ba7dc37d5d51baed

benchmark
view benchmark

commit
view commit

[github-actions]

Bundle size comparison (import * as ... in PR vs import * as ... in target):

🟢 Decreased	➖ Unchanged	🔴 Increased (max 1.00%)	❔ Unknown
0	37	285	0

import * as ... in PR vs import * as ... in target (did bundle size increase?):

Click to reveal the results table (152 entries).

Test	tsdown
d_bool.ts	13.62 kB ($${\color{red}+1.0\%}$$)
d_f16.ts	13.62 kB ($${\color{red}+1.0\%}$$)
d_f32.ts	13.62 kB ($${\color{red}+1.0\%}$$)
d_i32.ts	13.62 kB ($${\color{red}+1.0\%}$$)
d_u32.ts	13.62 kB ($${\color{red}+1.0\%}$$)
d_u16.ts	13.65 kB ($${\color{red}+1.0\%}$$)
d_textureDepth2d.ts	14.07 kB ($${\color{red}+1.0\%}$$)
d_textureDepthCube.ts	14.07 kB ($${\color{red}+1.0\%}$$)
d_texture1d.ts	14.08 kB ($${\color{red}+1.0\%}$$)
d_texture2d.ts	14.08 kB ($${\color{red}+1.0\%}$$)
d_texture3d.ts	14.08 kB ($${\color{red}+1.0\%}$$)
d_textureCube.ts	14.08 kB ($${\color{red}+1.0\%}$$)
d_textureDepth2dArray.ts	14.08 kB ($${\color{red}+1.0\%}$$)
d_textureDepthCubeArray.ts	14.09 kB ($${\color{red}+1.0\%}$$)
d_textureDepthMultisampled2d.ts	14.09 kB ($${\color{red}+1.0\%}$$)
d_texture2dArray.ts	14.09 kB ($${\color{red}+1.0\%}$$)
d_textureCubeArray.ts	14.10 kB ($${\color{red}+1.0\%}$$)
d_textureMultisampled2d.ts	14.10 kB ($${\color{red}+1.0\%}$$)
std_discard.ts	14.92 kB ($${\color{red}+0.9\%}$$)
std_isBeingTranspiled.ts	15.01 kB ($${\color{red}+0.9\%}$$)
std_getTargetShaderLanguage.ts	15.08 kB ($${\color{red}+0.9\%}$$)
std_extensionEnabled.ts	15.13 kB ($${\color{red}+0.9\%}$$)
std_copy.ts	15.16 kB ($${\color{red}+0.9\%}$$)
std_arrayLength.ts	15.16 kB ($${\color{red}+0.9\%}$$)
std_range.ts	15.40 kB ($${\color{red}+0.9\%}$$)
d_disarrayOf.ts	15.57 kB ($${\color{red}+0.9\%}$$)
std_dpdx.ts	15.88 kB ($${\color{red}+0.9\%}$$)
std_dpdxCoarse.ts	15.88 kB ($${\color{red}+0.9\%}$$)
std_dpdxFine.ts	15.88 kB ($${\color{red}+0.9\%}$$)
std_dpdy.ts	15.88 kB ($${\color{red}+0.9\%}$$)
std_dpdyCoarse.ts	15.88 kB ($${\color{red}+0.9\%}$$)
std_dpdyFine.ts	15.88 kB ($${\color{red}+0.9\%}$$)
std_fwidth.ts	15.88 kB ($${\color{red}+0.9\%}$$)
std_fwidthCoarse.ts	15.88 kB ($${\color{red}+0.9\%}$$)
std_fwidthFine.ts	15.88 kB ($${\color{red}+0.9\%}$$)
std_bitcastF32toU32.ts	47.75 kB ($${\color{red}+0.8\%}$$)
std_bitcastU32toF32.ts	47.75 kB ($${\color{red}+0.8\%}$$)
std_bitcastU32toI32.ts	47.75 kB ($${\color{red}+0.8\%}$$)
std_bitcast.ts	47.76 kB ($${\color{red}+0.8\%}$$)
std_atomicLoad.ts	16.68 kB ($${\color{red}+0.8\%}$$)
std_atomicStore.ts	16.68 kB ($${\color{red}+0.8\%}$$)
std_textureBarrier.ts	16.68 kB ($${\color{red}+0.8\%}$$)
std_atomicAdd.ts	16.69 kB ($${\color{red}+0.8\%}$$)
std_atomicAnd.ts	16.69 kB ($${\color{red}+0.8\%}$$)
std_atomicMax.ts	16.69 kB ($${\color{red}+0.8\%}$$)
std_atomicMin.ts	16.69 kB ($${\color{red}+0.8\%}$$)
std_atomicOr.ts	16.69 kB ($${\color{red}+0.8\%}$$)
std_atomicSub.ts	16.69 kB ($${\color{red}+0.8\%}$$)
std_atomicXor.ts	16.69 kB ($${\color{red}+0.8\%}$$)
std_storageBarrier.ts	16.69 kB ($${\color{red}+0.8\%}$$)
std_workgroupBarrier.ts	16.69 kB ($${\color{red}+0.8\%}$$)
d_vec2b.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec2f.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec2h.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec2i.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec2u.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec3b.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec3f.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec3h.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec3i.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec3u.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec4b.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec4f.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec4h.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec4i.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_vec4u.ts	20.07 kB ($${\color{red}+0.7\%}$$)
d_formatToWGSLType.ts	21.56 kB ($${\color{red}+0.6\%}$$)
d_uint8.ts	21.56 kB ($${\color{red}+0.6\%}$$)
d_float16.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_float16x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_float16x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_float32.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_float32x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_float32x3.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_float32x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint16.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint16x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint16x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint32.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint32x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint32x3.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint32x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint8.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint8x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_sint8x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_snorm16.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_snorm16x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_snorm16x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_snorm8.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_snorm8x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_snorm8x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_uint16.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_uint16x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_uint16x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_uint32.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_uint32x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_uint32x3.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_uint32x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_uint8x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_uint8x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_unorm10_10_10_2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_unorm16.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_unorm16x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_unorm16x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_unorm8.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_unorm8x2.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_unorm8x4.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_unorm8x4_bgra.ts	21.57 kB ($${\color{red}+0.6\%}$$)
d_packedFormats.ts	21.59 kB ($${\color{red}+0.6\%}$$)
d_isPackedData.ts	21.63 kB ($${\color{red}+0.6\%}$$)
d_alignmentOf.ts	22.52 kB ($${\color{red}+0.6\%}$$)
std_subgroupAdd.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupAll.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupAnd.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupAny.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupBallot.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupBroadcast.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupBroadcastFirst.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupElect.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupExclusiveAdd.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupExclusiveMul.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupInclusiveAdd.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupInclusiveMul.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupMax.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupMin.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupMul.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupOr.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupShuffle.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupShuffleDown.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupShuffleUp.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupShuffleXor.ts	25.02 kB ($${\color{red}+0.5\%}$$)
std_subgroupXor.ts	25.02 kB ($${\color{red}+0.5\%}$$)
d_isBuiltin.ts	25.25 kB ($${\color{red}+0.5\%}$$)
d_sizeOf.ts	25.29 kB ($${\color{red}+0.5\%}$$)
d_isContiguous.ts	25.30 kB ($${\color{red}+0.5\%}$$)
d_getLongestContiguousPrefix.ts	25.31 kB ($${\color{red}+0.5\%}$$)
std_textureDimensions.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureGather.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureLoad.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureSample.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureSampleBaseClampToEdge.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureSampleBias.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureSampleCompare.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureSampleCompareLevel.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureSampleGrad.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureSampleLevel.ts	26.63 kB ($${\color{red}+0.5\%}$$)
std_textureStore.ts	26.63 kB ($${\color{red}+0.5\%}$$)
d_arrayOf.ts	26.73 kB ($${\color{red}+0.5\%}$$)
d_size.ts	26.98 kB ($${\color{red}+0.5\%}$$)
d_align.ts	26.98 kB ($${\color{red}+0.5\%}$$)
d_location.ts	26.99 kB ($${\color{red}+0.5\%}$$)
d_interpolate.ts	26.99 kB ($${\color{red}+0.5\%}$$)

import { ... } in PR vs import * as ... in PR (is the library tree-Shakeable?):

Test	tsdown
tgpu_init.ts	260.15 kB ($${\color{green}-3.2\%}$$)
tgpu_initFromDevice.ts	259.62 kB ($${\color{green}-3.4\%}$$)
tgpu_resolve.ts	165.55 kB ($${\color{green}-38.4\%}$$)
tgpu_resolveWithContext.ts	165.49 kB ($${\color{green}-38.4\%}$$)
tgpu_bindGroupLayout.ts	69.40 kB ($${\color{green}-74.2\%}$$)
tgpu_mutableAccessor.ts	66.41 kB ($${\color{green}-75.3\%}$$)
tgpu_accessor.ts	66.39 kB ($${\color{green}-75.3\%}$$)
tgpu_privateVar.ts	65.74 kB ($${\color{green}-75.5\%}$$)
tgpu_workgroupVar.ts	65.74 kB ($${\color{green}-75.5\%}$$)
tgpu_const.ts	64.98 kB ($${\color{green}-75.8\%}$$)
tgpu_fn.ts	38.59 kB ($${\color{green}-85.6\%}$$)
tgpu_fragmentFn.ts	38.59 kB ($${\color{green}-85.6\%}$$)
tgpu_vertexFn.ts	38.40 kB ($${\color{green}-85.7\%}$$)
tgpu_computeFn.ts	38.11 kB ($${\color{green}-85.8\%}$$)
tgpu_vertexLayout.ts	27.22 kB ($${\color{green}-89.9\%}$$)
tgpu_comptime.ts	14.91 kB ($${\color{green}-94.5\%}$$)
tgpu_unroll.ts	1.66 kB ($${\color{green}-99.4\%}$$)
tgpu_slot.ts	1.54 kB ($${\color{green}-99.4\%}$$)
tgpu_lazy.ts	1.19 kB ($${\color{green}-99.6\%}$$)

If you wish to run a comparison for other, slower bundlers, run the 'Tree-shake test' from the GitHub Actions menu.

[Copilot]

Pull request overview

This PR replaces Float16Array-dependent float16 read/write paths with a manual IEEE-754 binary16 conversion layer to support runtimes like React Native (Hermes) that don’t implement Float16Array.

Changes:

• Introduces float16Conversion.ts with helpers to encode/decode float16 values via uint16 payloads.
• Switches std/packing and dataIO float16 serialization to use the new helpers instead of typed-binary float16 APIs.
• Updates std/bitcast CPU implementation to avoid constructing Float16Array, using a Uint16Array view plus encode/decode.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
packages/typegpu/src/std/packing.ts	Uses manual float16 read/write wrappers for pack/unpack builtins.
packages/typegpu/src/std/bitcast.ts	Removes Float16Array buffer view; adds f16 encode/decode path via Uint16Array.
packages/typegpu/src/data/float16Conversion.ts	Adds new float16 encoding/decoding utilities used across the codebase.
packages/typegpu/src/data/dataIO.ts	Routes all f16 scalar/vector serialization through the new conversion helpers.

Comments suppressed due to low confidence (1)

packages/typegpu/src/data/float16Conversion.ts:33

• decodeUint16AsFloat16 handles subnormals incorrectly: for exponent === 0, IEEE-754 binary16 uses an exponent of -14 (no implicit leading 1). Returning mantissa / 1024 makes subnormal values ~16384× too large (e.g. 0x0001).

  if (exponent === 0) {
    return sign === 0 ? mantissa / 1024 : -mantissa / 1024;
  }

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Resolution Time Benchmark

---
config:
  themeVariables:
    xyChart:
      plotColorPalette: "#E63946, #3B82F6, #059669"
---
xychart
  title "Random Branching (🔴 PR | 🔵 main | 🟢 release)"
  x-axis "max depth" [1, 2, 3, 4, 5, 6, 7, 8]
  y-axis "time (ms)"
  line [0.78, 1.53, 3.27, 5.67, 6.38, 10.78, 18.27, 23.47]
  line [0.81, 1.56, 3.50, 5.73, 5.94, 11.18, 18.87, 21.47]
  line [0.85, 1.72, 3.58, 5.66, 6.21, 10.12, 18.82, 20.02]

Loading

---
config:
  themeVariables:
    xyChart:
      plotColorPalette: "#E63946, #3B82F6, #059669"
---
xychart
  title "Linear Recursion (🔴 PR | 🔵 main | 🟢 release)"
  x-axis "max depth" [1, 2, 3, 4, 5, 6, 7, 8]
  y-axis "time (ms)"
  line [0.31, 0.47, 0.59, 0.70, 0.96, 0.99, 1.17, 1.29]
  line [0.34, 0.51, 0.63, 0.78, 1.05, 1.02, 1.24, 1.31]
  line [0.33, 0.50, 0.61, 0.73, 0.96, 1.00, 1.24, 1.40]

Loading

---
config:
  themeVariables:
    xyChart:
      plotColorPalette: "#E63946, #3B82F6, #059669"
---
xychart
  title "Full Tree (🔴 PR | 🔵 main | 🟢 release)"
  x-axis "max depth" [1, 2, 3, 4, 5, 6, 7, 8]
  y-axis "time (ms)"
  line [0.82, 1.78, 3.37, 5.72, 10.35, 23.14, 47.39, 96.34]
  line [0.82, 1.82, 3.43, 5.58, 10.67, 22.60, 48.38, 98.09]
  line [0.77, 1.85, 3.99, 6.40, 11.35, 22.65, 47.92, 98.67]

Loading

[pullfrog]

Caution

The new float16 conversion functions in float16Conversion.ts have algorithmic bugs that corrupt subnormal values — decoded 16384× too large, encoded as garbage bit patterns — and lose negative zero. packages/typegpu/src/data/numeric.ts already exports correct, well-tested toHalfBits/fromHalfBits that do not depend on Float16Array; reusing those would fix all three issues at once.

Reviewed changes

• New float16Conversion.ts — writeFloat16/readFloat16 wrappers for binary I/O plus standalone encodeFloat16AsUint16/decodeUint16AsFloat16 functions implementing IEEE 754 binary16 conversion.
• dataIO.ts — All output.writeFloat16/input.readFloat16 calls replaced with the new writeFloat16/readFloat16 helpers.
• bitcast.ts — Float16Array buffer view replaced with Uint16Array; new writeFloat16ToBuffer/readFloat16FromBuffer functions handle the f16 bitcast path with manual encode/decode.
• packing.ts — writeFloat16/readFloat16 helpers used in pack2x16float/unpack2x16float CPU paths.

---

  ^{｜ Fix all ➔ ｜ Fix 👍s ➔ ｜ View workflow run ｜ Using DeepSeek Pro (free via Pullfrog for OSS) (Kimi K2 not used — the program covers this model; add its provider key to run your pick) ｜ 𝕏}

[pullfrog]

✅ No new issues found.

Reviewed changes

• Removed buggy conversion functions — encodeFloat16AsUint16/decodeUint16AsFloat16 deleted from float16Conversion.ts; writeFloat16/readFloat16 now delegate to toHalfBits/fromHalfBits from numeric.ts.
• Improved toHalfBits rounding — Both the normal and subnormal paths in numeric.ts now implement proper round-ties-to-even (using round-bit + sticky-bit) instead of the previous add-0.5-and-truncate approach. The normal path also gained an exponent-overflow guard when rounding causes a mantissa carry.
• Removed Float16Array from bitcast — bufViews replaced f16: Float16Array with u16: Uint16Array; new writeFloat16ToBuffer/readFloat16FromBuffer functions use toHalfBits/fromHalfBits for encode/decode. Float16Array purged from writeToBuffer/readFromBuffer type signatures.
• Added comprehensive tests — halfBits.test.ts validates toHalfBits/fromHalfBits against native Float16Array across zeros, NaNs, infinities, normals, subnormals, overflow, underflow, and round-ties-to-even (14 tests, all passing).

  ^{｜ View workflow run ｜ Using DeepSeek Pro (free via Pullfrog for OSS) (Kimi K2 not used — the program covers this model; add its provider key to run your pick) ｜ 𝕏}