Crate jxl_encoder_simd 

SIMD-accelerated primitives for jxl_encoder.

This crate wraps platform-specific SIMD intrinsics behind safe public functions. The main encoder crate (jxl_encoder) maintains #![forbid(unsafe_code)] and calls into these safe wrappers.

Uses archmage for token-based SIMD dispatch and magetypes for cross-platform vector types.

Direct variant access

Each kernel is available in three forms:

• A dispatching function (e.g. dct_8x8) that picks the best at runtime
• Concrete _avx2(token, ...) / _neon(token, ...) / _scalar(...) variants

For hot loops, callers should summon a token once, then call the concrete variant directly from an #[arcane] function so LLVM can inline across the target-feature boundary.

Structs

EntropyCoeffResult

Results from vectorized entropy coefficient processing.

X64V3Token

Proof that AVX2 + FMA + BMI1/2 + F16C + LZCNT are available (x86-64-v3 level).

Traits

SimdToken

Marker trait for SIMD capability tokens.

Functions

compute_block_l2_errors

Compute per-block masked weighted L2 error between original and reconstructed XYB planes.

compute_block_l2_errors_avx2

compute_block_l2_errors_scalar

compute_mask1x1

Compute per-pixel masking field from XYB Y channel.

compute_mask1x1_avx2

compute_mask1x1_scalar

dct_8x8

Compute scaled 8x8 forward DCT with SIMD acceleration.

dct_8x8_avx2

dct_8x8_scalar

dct_8x16

Compute scaled 8x16 forward DCT with SIMD acceleration.

dct_8x16_avx2

dct_8x16_scalar

dct_16x8

Compute scaled 16x8 forward DCT with SIMD acceleration.

dct_16x8_avx2

dct_16x8_scalar

dct_16x16

Compute 16x16 forward DCT with SIMD acceleration.

dct_16x16_avx2

AVX2 16x16 forward DCT: process 8 rows at a time via batched 16-point DCT.

dct_16x16_scalar

dequant_block_dct8

Dequantize a DCT8 block and apply CfL (chroma-from-luma) in one pass.

dequant_dct8_avx2

dequant_dct8_scalar

entropy_coeffs_avx2

entropy_coeffs_scalar

entropy_estimate_coeffs

Vectorized entropy coefficient processing.

epf_step1

Apply EPF Step 1 to 3-channel XYB planes.

epf_step2

Apply EPF Step 2 to 3-channel XYB planes.

epf_step1_avx2

epf_step1_scalar

epf_step2_avx2

epf_step2_scalar

forward_xyb_avx2

forward_xyb_scalar

gab_smooth_avx2

AVX2+FMA gab smooth: processes 8 pixels per iteration in interior rows. Border rows/columns use scalar fallback.

gab_smooth_channel

Apply 3x3 weighted gaborish smooth to a single channel in-place.

gab_smooth_scalar

gaborish_5x5_avx2

AVX2+FMA gaborish 5x5: processes 8 pixels per iteration in interior region. Border pixels (within 2 of edge) use scalar fallback.

gaborish_5x5_channel

Apply the 5x5 gaborish inverse kernel to a single channel.

gaborish_5x5_scalar

idct_8x8

Compute scaled 8x8 inverse DCT with SIMD acceleration.

idct_8x8_avx2

idct_8x8_scalar

idct_8x16

Compute 8x16 inverse DCT with SIMD acceleration.

idct_8x16_avx2

idct_8x16_scalar

idct_16x8

Compute 16x8 inverse DCT with SIMD acceleration.

idct_16x8_avx2

idct_16x8_scalar

idct_16x16

Compute 16x16 inverse DCT with SIMD acceleration.

idct_16x16_avx2

AVX2 16x16 IDCT: process 8 rows at a time via batched 16-point IDCT.

idct_16x16_scalar

inverse_xyb_avx2

inverse_xyb_planar_avx2

inverse_xyb_planar_scalar

inverse_xyb_scalar

linear_rgb_to_xyb_batch

Convert separate R, G, B channel buffers to separate X, Y, B channel buffers.

pixel_domain_loss

Compute pixel-domain loss for one channel of a block.

pixel_domain_loss_avx2

pixel_domain_loss_scalar

quantize_block_dct8

Quantize a DCT8 block (64 coefficients) with dead-zone thresholding.

quantize_dct8_avx2

quantize_dct8_scalar

transpose_8x8

Transpose an 8x8 f32 matrix.

transpose_8x8_avx2

AVX2 8x8 transpose using unpack/shuffle/permute instructions.

xyb_to_linear_rgb_batch

Convert separate X, Y, B channel buffers to interleaved linear RGB.

xyb_to_linear_rgb_planar

Convert separate X, Y, B channel buffers to planar linear RGB.