vulkan-coverage.md

Vulkan op coverage

Generated for chunk 6 of the Vulkan backend (docs/vulkan.md). What runs on Device::vulkan(i) today, slot by slot, and what each sibling's code calls.

Op table: 266 of 270 slots implemented, 4 null, all by design (chunk 9 filled the 17 training slots that remained: the four flash-attention backwards, group_norm_backward, resblock_backward, embedding_lookup_backward, scatter_rows_add, the two conv_transpose2d backwards, upsample_bilinear_2x_backward, interp2d_backward, pad2d_backward, adaptive_avg_pool2d_backward, max_pool2d_backward, lstm_forward_train and lstm_backward; chunk 8 the attention / BatchNorm / BCE ones). A null slot throws "not implemented on vulkan" when dispatched; none of the four is ever dispatched to Vulkan: two are host-only ops that always run on the CPU table, and filtered_lrelu's public entry points fall back to a composite of two ops Vulkan implements. Every op the siblings call, inference and training, runs on Vulkan (brotensor_test_vulkan checks the null set).

Outside the op table: conv1d*, causal_conv1d, conv1d_int8w_fp16 (wrappers over conv2d / conv_transpose1d / the INT8 conv), lora_forward / lora_backward (GEMM compositions) and the five fused_* ops (src/ops/fused.cpp; the stacked-weight SwiGLU GEMV takes the Vulkan GEMV epilogue, the rest compose dispatched ops) all run on Vulkan. jit::* traces on Vulkan compile to one SPIR-V kernel each (src/jit/trace_compiler_vulkan.cpp, docs/vulkan.md "Trace JIT"); one with no fusion is replayed op by op (src/jit/trace_eager.cpp). Device::CUDA aliases to Vulkan when no CUDA backend is registered (detail::resolve_device_alias; to Metal instead when Metal is registered and the default device is not Vulkan).

Null slots

Slot	Why
mse_scalar	host-only: no Tensor operand, ops.cpp always runs it on the CPU table
softmax_xent_segment	host-only: raw host pointers, ops.cpp always runs it on the CPU table
filtered_lrelu_forward	by design: composite fallback (bias_act + upfirdn2d), runs on Vulkan
filtered_lrelu_backward	by design: the public op falls back to its bias_act + upfirdn2d composite, which runs on Vulkan

Per sibling

Op-table ops each sibling's src/ and include/ call (an unqualified-name scan, so it over-approximates: a sibling function sharing an op's name counts). "Training" means the op is a gradient / optimiser / loss op or every call site is in a training file (LayaGrad, LoRA trainers, BC-ResNet / phoneme training, the Kokoro decoder backward, brogameagent's learners). Ops new in chunk 6 are marked *.

brolm

Inference: 54 ops, all implemented.

tanh_forward, sigmoid_forward, add_inplace, scale_inplace, clamp, mul_inplace, masked_mean_pool_forward, embedding_lookup_forward, copy_d2d, copy_d2d_strided, cast, layernorm_forward_inference_batched, linear_forward_batched, relu_forward_batched, conv2d_forward, silu_forward, gelu_forward, gelu_exact_forward, quick_gelu_forward, gather_rows, scatter_rows, linear_forward_batched_fp16, linear_forward_batched_ex, layernorm_forward_inference_batched_fp16, flash_attention_forward, flash_attention_gqa_forward, flash_attention_packed_qkv_forward, flash_attention_qkvo_forward, nchw_to_sequence, sequence_to_nchw, matmul, rope_forward, rms_norm_forward, swiglu_forward, kv_cache_append, flash_attention_decode, flash_attention_decode_masked, segment_softmax_stats, linear_forward_batched_int8w_fp16, dequant_q4k_to_fp16, linear_forward_batched_q4k_fp16, dequant_q8_0_to_fp16, linear_forward_batched_q8_0_fp16, dequant_q6k_to_fp16, linear_forward_batched_q6k_fp16, rope_apply, rope_qkv_packed_inplace, rope_apply_mrope, self_attention_bias_forward, self_attention_bias_int8w_fp16, causal_conv1d_update, l2_norm_forward, gated_delta_rule_chunked, gated_delta_rule_step

Training: 12 ops, all implemented (scatter_rows_add, flash_attention_packed_qkv_backward in chunk 9).

relu_backward, masked_mean_pool_backward, layernorm_forward_batched_with_caches, layernorm_backward_batched_with_caches, adam_step, xavier_init*, linear_backward_batched, silu_backward, gelu_exact_backward, scatter_rows_add, geglu_exact_backward, flash_attention_packed_qkv_backward

brodiffusion

Inference: 63 ops, all implemented.

relu_forward, tanh_forward, sigmoid_forward, add_inplace, axpby_inplace, add_scalar_inplace, scale_inplace, clamp, mul_inplace, mha_forward, concat_rows, concat_nchw_channels, copy_d2d, copy_d2d_strided, cast, layernorm_forward_inference_batched, linear_forward_batched, conv2d_forward, group_norm_forward, silu_forward, gelu_forward, gelu_exact_forward, upsample_nearest_2x, downsample_avg_2x, pad2d_forward, linear_forward_batched_fp16, geglu_exact_forward, cross_attention_forward_with_attn, layernorm_forward_inference_batched_fp16, flash_attention_forward, flash_attention_gqa_forward, flash_attention_windowed_forward, flash_attention_varlen_forward, flash_attention_qkvo_forward, flash_attention_project_kv, flash_attention_q_with_kv_cached_forward, nchw_to_sequence, sequence_to_nchw, spatial_merge_2x2_forward, pixel_shuffle_upsample_2x_forward, patch_unpack_forward, resblock_forward, resblock_forward_int8w_fp16, matmul, matmul_abt, rms_norm_forward, ddim_step, timestep_embedding, conv2d_int8w_fp16_forward, linear_forward_batched_int8w_fp16, flash_attention_project_kv_int8w_fp16, flash_attention_q_with_kv_cached_int8w_fp16, flash_attention_qkvo_int8w_fp16, modulate, broadcast_mul, rope_apply, rope_apply_perhead, l2_norm_forward, randn, rsqrt_forward, pixel_norm_forward, add_channel_bias_inplace, softmax_rows_forward

Training: 1 ops, all implemented.

adam_step*

brosoundml

Inference: 76 ops, all implemented.

linear_forward, relu_forward, tanh_forward, sigmoid_forward, add_inplace, axpby_inplace, add_scalar_inplace, scale_inplace, clamp, mul_inplace, embedding_lookup_forward*, concat_rows, concat_batched_rows, concat_nchw_channels, copy_d2d, copy_d2d_strided, cast, layernorm_forward_inference_batched, linear_forward_batched, relu_forward_batched, add_inplace_batched, conv2d_forward, group_norm_forward, silu_forward, gelu_forward, gelu_exact_forward, interp2d_align_corners_forward, pad2d_forward, slice2d_forward, top_k_rows, gather_rows, scatter_rows, linear_forward_batched_fp16, linear_forward_batched_ex, flash_attention_forward, flash_attention_gqa_forward, flash_attention_windowed_forward, flash_attention_varlen_forward, nchw_to_sequence, sequence_to_nchw, matmul, rms_norm_forward, sum_cols, argmax_rows, modulate, broadcast_mul, rope_apply, self_attention_bias_forward, rel_pos_bias_xl_forward, complex_mul, complex_abs, complex_angle, complex_from_polar, rfft, irfft, stft, istft, conv_transpose1d_forward, pad1d_forward, snake_forward, elu_forward, leaky_relu_forward, vq_encode_forward, resample1d_forward, log_forward, exp_forward, sample_logits_into, masked_diffusion_scores, masked_diffusion_commit, batch_norm_inference, randn, rand_uniform, sin_forward, add_channel_bias_inplace, add_row_bias_inplace, softmax_rows_forward

Training: 19 ops, all implemented (group_norm_backward in chunk 9; batch_norm_forward / _backward and bce_with_logits_fused_batched in chunk 8).

linear_backward, tanh_backward, adam_step, xavier_init, linear_backward_batched, relu_backward_batched, softmax_xent_fused_batched, bce_with_logits_fused_batched, conv2d_backward_input, conv2d_backward_weight*, conv2d_backward_bias, group_norm_backward, istft_backward, conv_transpose1d_backward_input, pad1d_backward, snake_backward, leaky_relu_backward, batch_norm_forward, batch_norm_backward

brovisionml

StyleGAN3 GAN inversion (stylegan3.cpp, fitting W+ to an image) is an inference-time feature that runs a backward: modulated_conv2d_backward (frozen weights), the filtered_lrelu / bias_act / upfirdn2d backwards and adam_step, all of which run on Vulkan now.

Inference: 54 ops, all implemented.

relu_forward, tanh_forward, sigmoid_forward, add_inplace, add_scalar_inplace, scale_inplace, clamp, mul_inplace, mha_forward, concat_rows, concat_batched_rows, concat_nchw_channels, copy_d2d, copy_d2d_strided, cast, layernorm_forward_inference_batched, linear_forward_batched, relu_forward_batched, conv2d_forward, deform_conv2d_forward, group_norm_forward, silu_forward, gelu_exact_forward, interp2d_forward, interp2d_align_corners_forward, pad2d_forward, slice2d_forward, l2_normalize_nchw_forward, convex_upsample_forward, adaptive_avg_pool2d_forward, max_pool2d_forward, gather_rows, conv_transpose2d_forward, linear_forward_batched_fp16, linear_forward_batched_fp16_act, flash_attention_varlen_forward, nchw_to_sequence, sequence_to_nchw, matmul, swiglu_forward, argmax_rows, rows_count_above, threshold_u8, broadcast_mul, rope_apply, self_attention_bias_forward, self_attention_decomposed_rel_pos_forward, self_attention_decomposed_rel_pos_windowed_forward, leaky_relu_forward, batch_norm_inference, pixel_norm_forward, modulated_conv2d_forward, bias_act_forward*, filtered_lrelu_forward (composite, runs)

Training: 4 ops, all implemented.

adam_step, matmul_backward, modulated_conv2d_backward, filtered_lrelu_backward (composite, runs)

brogameagent

The nets construct their weights with xavier_init (bit-identical to the CPU stream on Vulkan) and infer through the ops below; the learners' backwards run too (chunk 8).

Inference: 22 ops, all implemented.

linear_forward, relu_forward, tanh_forward, sigmoid_forward, add_inplace, add_scalar_inplace, scale_inplace, clamp, build_slot_mask, softmax_forward, layernorm_forward, attention_forward, mha_forward, masked_mean_pool_forward*, concat_rows, split_rows, copy_d2d, cast, linear_forward_batched, relu_forward_batched, tanh_forward_batched, add_inplace_batched

Training: 20 ops, all implemented (attention_backward / mha_backward in chunk 8).

linear_backward, relu_backward, tanh_backward, sigmoid_backward, softmax_backward, layernorm_backward, attention_backward, mha_backward, masked_mean_pool_backward, mse_scalar (host-only, runs), softmax_xent, softmax_xent_segment (host-only, runs), sgd_step, adam_step, xavier_init, linear_backward_batched, relu_backward_batched, tanh_backward_batched, mse_vec_per_sample*, softmax_xent_fused_batched

What the siblings still need (chunk 7)

Historical: written while the HIP backend still existed (removed 2026-10-03, docs/vulkan.md); the BROTENSOR_PREFER_HIP / HIP references below describe that time.

Status after chunk 7b. Everything in the list below is done: brolm, brosoundml and brogameagent gate on graph_capture_available and accept "vulkan" / "vk" (as do bro's bro.gpu and bro.motion); brolm's T5 runs an FP32-activation forward on Vulkan. The list missed some sites, fixed too: brolm laya_batch.cpp / laya_scheduler.cpp (CUDA / HIP-only graphs and replicas), brosoundml omnivoice_lm.cpp (a DeviceType test) and the JS OmniVoice precision default, brogameagent's JS device names, bro's native_motion.cpp. Found while doing it, still open:

• The null training slots now meet users. Vulkan is the default device with HIP and Vulkan both built, so the training code that runs on the default device reaches the null slots: brogameagent's attention / MHA / transformer layer backwards (attention_backward, mha_backward; 5 of its 33 tests) and brosoundml's BC-ResNet / phoneme-model training (batch_norm_forward in training mode and bce_with_logits_fused_batched; 2 of 48). They pass on HIP (BROTENSOR_PREFER_HIP=1). Done in chunk 8 (docs/vulkan-training.md): both siblings pass every test on Vulkan.
• STFT precision. The DFT-as-GEMM STFT (ops_spectral.cpp) carries ~5e-5 absolute error at n_fft 512 (FP32 partial sums of a DFT row that cancel), against the CPU FFT's ~1e-6, and the GEMM's K rotation makes the rounding depend on a frame's row position. brosoundml test_mel fails on it: log-mel differs from the CPU by 0.025 (HIP: 1e-6), and streaming differs from offline by 0.018 in the quietest bins. STT / KWS / wake results are unaffected in every test, but a front end compared bit for bit against the CPU will see it. Fix: an FFT, or FP64 (or compensated) accumulation in the DFT GEMM. Done in chunk 8: transforms up to L = 2048 run an FP64 direct DFT (dft64.comp); test_mel passes.
• Long graph submissions. A captured TripoSplat step (two 3.6 s flow forwards at 512 px, 8 steps) was one command buffer past the amdgpu ~10 s job limit: device lost. Graph captures are now cut into BROTENSOR_VK_BATCH-command segments, each its own submission (docs/vulkan.md, "Graphs").

The original list:

The op table is not the whole story: these sibling-side paths name a backend explicitly and either skip Vulkan or would misbehave on it. None of them is an op gap in brotensor.

• Device tests by equality. X.device == Device::CUDA || X.device == Device::HIP (brosoundml modules.cpp, whisper_modules.cpp, qwen_tts_talker.cpp, qwen_tts_code_predictor.cpp, supertonic.cpp; brodiffusion pipeline.cpp, triposplat/sampler.cpp, flow_model.cpp; brolm detail/dense_decoder.cpp, t5.cpp) gates GPU-only fast paths and graph capture. A tensor created through the Device::CUDA alias is tagged vulkan, so these tests are false on Vulkan and the code takes its CPU-style / ungraphed path (correct, slower). They should test !device.is_cpu() or add Device::VULKAN. Capture sites need no rewrite: CudaGraphCapture is device-neutral since chunk 7a (it records on the default device, Vulkan included; graph_capture_available(Device) is the gate, docs/vulkan.md "Device-neutral capture"). Done for brodiffusion in 7a.
• Device parsing. The JS hosts map "cuda" / "hip" / "rocm" strings to devices (brolm api/host_lm_internal.h, brosoundml api/host_soundml_internal.h, brodiffusion native_diffusion_triposplat.cpp, brovisionml native_vision_ops.cpp) and pick a default GPU from CUDA / HIP only; "vulkan" / "vk" needs adding (brotensor's own BROTENSOR_DEFAULT_DEVICE parser already accepts both).
• T5 on Vulkan. brolm's T5 re-casts to BF16 only on CUDA / HIP and keeps FP16 + the residual clamp elsewhere. BF16 would not help on Vulkan (the cooperative-matrix GEMM converts BF16 operands to FP16 as it stages them, so a BF16 activation beyond 65504 becomes inf there), but the FP16 path is not usable either: chunk 7a ran PixArt-Sigma with T5-XXL on Vulkan and got noise, while the same DiT and VAE on Vulkan with the T5 encoding computed on HIP produced the HIP image. T5 on Vulkan needs FP32 activations through the matrix ops (the SIMT GEMM) or a range-safe BF16 GEMM. Resolved: chunk 7b took the FP32 route; the range-safe BF16 GEMM (docs/vulkan-bf16.md) then let T5 take its BF16 path on Vulkan too, 2x faster than HIP.
• Sibling GPU kernels: must be fixed before Vulkan tensors reach them. brodiffusion fused_resblock.cu, fused_transformer.cu (fused linear + GeGLU, row-bias / vector adds) and triposplat/flow_rope.cu, brovisionml dsine_ops.cu (ray ReLU, angular propagation) and dpt_preprocess_gpu.cu are CUDA / HIP kernels, and their entry points route any non-CPU tensor to them (fused.cpp's on_gpu(t) is t.device != CPU; dsine checks == CPU). With a Vulkan tensor that is a HIP kernel handed Vulkan buffer addresses in a HIP + Vulkan build, and the host CPU implementation dereferencing them in a Vulkan-only build: a fault either way. Each needs a Vulkan branch: the brodiffusion fusions can compose brotensor ops Vulkan has (resblock_forward, a linear then geglu_exact_forward — not the GEMM's GeGLU epilogue, which wants interleaved rows where brodiffusion stacks the halves — add_row_bias_inplace, add_inplace); dsine's two ops, DPT's preprocessing and TripoSplat's RoPE tables (whose host fallback syncs, which a capture forbids) need small GLSL kernels. Done in 7a (custom kernels: docs/vulkan.md "Custom kernels").
• BF16 activations past FP16's range. Not an op gap, a dtype-policy one: Sana's linear attention sums N products into its (hd, hd) summaries in BF16 on CUDA / HIP; on Vulkan the next GEMM stages them as FP16 and they become inf (all-NaN output). 7a runs that core in FP32 on Vulkan. Flux / Krea2 / Qwen-Image (BF16 on every GPU, flux_compute_dtype) pass their fixture tests but were not run with real weights on Vulkan.
• BF16 checkpoints. safetensors::upload_compute(_checked) / upload_as already do the policy's cast at load on Vulkan (raw BF16 bits up, device cast to FP16; tested). safetensors::upload and gguf::upload_raw keep the file dtype by contract, so a BF16 weight loaded through them stays BF16 and runs through the GEMM's slower convert-at-load path; upload_fp16 rejects BF16 on every backend. Siblings that hand-convert on the host are unaffected.

Every slot

Grouped as in include/brotensor/detail/op_table.h; the file is the Vulkan op file that registers the slot.

Dense layers + elementwise activations

Slot	Vulkan	File
linear_forward	implemented	ops_linear.cpp
linear_backward	implemented	ops_linear.cpp
relu_forward	implemented	ops_elementwise.cpp
relu_backward	implemented	ops_elementwise.cpp
tanh_forward	implemented	ops_elementwise.cpp
tanh_backward	implemented	ops_elementwise.cpp
sigmoid_forward	implemented	ops_elementwise.cpp
sigmoid_backward	implemented	ops_elementwise.cpp
add_inplace	implemented	ops_elementwise.cpp
axpby_inplace	implemented	ops_elementwise.cpp
add_scalar_inplace	implemented	ops_elementwise.cpp
scale_inplace	implemented	ops_elementwise.cpp
clamp	implemented	ops_elementwise.cpp
mul_inplace	implemented	ops_elementwise.cpp
div_inplace	implemented	ops_elementwise.cpp
build_slot_mask	implemented (chunk 6)	ops_misc.cpp

Reductions / norm / softmax / attention (training)

Slot	Vulkan	File
softmax_forward	implemented	ops_norm.cpp
softmax_backward	implemented	ops_norm.cpp
layernorm_forward	implemented	ops_norm.cpp
layernorm_backward	implemented	ops_norm.cpp
attention_forward	implemented (chunk 6)	ops_attention_proj.cpp
attention_backward	implemented	ops_attention_bwd.cpp
mha_forward	implemented	ops_attention_proj.cpp
mha_backward	implemented	ops_attention_bwd.cpp

Pooling / losses / embeddings / concat

Slot	Vulkan	File
masked_mean_pool_forward	implemented (chunk 6)	ops_misc.cpp
masked_mean_pool_backward	implemented (chunk 6)	ops_misc.cpp
mse_vec_forward	implemented (chunk 6)	ops_misc.cpp
mse_vec_backward	implemented (chunk 6)	ops_misc.cpp
mse_scalar	null	-
softmax_xent	implemented	ops_xent.cpp
softmax_xent_segment	null	-
softmax_xent_fused	implemented	ops_xent.cpp
embedding_lookup_forward	implemented (chunk 6)	ops_misc.cpp
embedding_lookup_backward	implemented (chunk 9)	ops_scatter.cpp
concat_rows	implemented	ops_copy.cpp
split_rows	implemented	ops_copy.cpp
concat_batched_rows	implemented	ops_spatial.cpp
concat_nchw_channels	implemented	ops_spatial.cpp
concat_nchw_channels_backward	implemented	ops_spatial.cpp
copy_d2d	implemented	ops_copy.cpp
copy_d2d_strided	implemented	ops_copy.cpp
cast	implemented	ops_copy.cpp

Inference batched + optim

Slot	Vulkan	File
layernorm_forward_inference_batched	implemented	ops_norm.cpp
layernorm_forward_batched_with_caches	implemented	ops_norm.cpp
layernorm_backward_batched_with_caches	implemented	ops_norm.cpp
sgd_step	implemented (chunk 6)	ops_misc.cpp
adam_step	implemented (chunk 6)	ops_misc.cpp
xavier_init	implemented (chunk 6)	ops_misc.cpp

Batched (inference-only) variants

Slot	Vulkan	File
linear_forward_batched	implemented	ops_linear.cpp
relu_forward_batched	implemented	ops_elementwise.cpp
tanh_forward_batched	implemented	ops_elementwise.cpp
add_inplace_batched	implemented	ops_elementwise.cpp

Batched (training) backward variants

Slot	Vulkan	File
linear_backward_batched	implemented	ops_linear.cpp
relu_backward_batched	implemented	ops_elementwise.cpp
tanh_backward_batched	implemented	ops_elementwise.cpp

Batched per-sample loss kernels

Slot	Vulkan	File
mse_vec_per_sample	implemented (chunk 6)	ops_misc.cpp
softmax_xent_fused_batched	implemented	ops_xent.cpp
bce_with_logits_fused_batched	implemented	ops_xent.cpp

Conv2d (forward + backwards)

Slot	Vulkan	File
conv2d_forward	implemented	ops_conv.cpp
conv2d_backward_input	implemented	ops_conv.cpp
conv2d_backward_weight	implemented (chunk 6)	ops_vision.cpp
conv2d_backward_bias	implemented	ops_conv.cpp

Modulated deformable conv2d (torchvision deform_conv2d v2, fwd)

Slot	Vulkan	File
deform_conv2d_forward	implemented (chunk 6)	ops_vision.cpp

Conv3d (forward + W8A16 variant for Qwen3-VL patch embed)

Slot	Vulkan	File
conv3d_forward	implemented	ops_conv.cpp
conv3d_int8w_fp16_forward	implemented	ops_quant.cpp

GroupNorm

Slot	Vulkan	File
group_norm_forward	implemented	ops_gnorm.cpp
group_norm_backward	implemented (chunk 9)	ops_gnorm.cpp

Activations: silu, gelu (tanh-approx + exact), quick_gelu

Slot	Vulkan	File
silu_forward	implemented	ops_elementwise.cpp
silu_backward	implemented	ops_elementwise.cpp
gelu_forward	implemented	ops_elementwise.cpp
gelu_backward	implemented	ops_elementwise.cpp
gelu_exact_forward	implemented	ops_elementwise.cpp
gelu_exact_backward	implemented	ops_elementwise.cpp
quick_gelu_forward	implemented	ops_elementwise.cpp
quick_gelu_backward	implemented	ops_elementwise.cpp

Resample (NN / bilinear / avgpool)

Slot	Vulkan	File
upsample_nearest_2x	implemented	ops_spatial.cpp
upsample_bilinear_2x	implemented	ops_spatial.cpp
downsample_avg_2x	implemented	ops_spatial.cpp
upsample_nearest_2x_backward	implemented	ops_spatial.cpp
upsample_bilinear_2x_backward	implemented (chunk 9)	ops_spatial_bwd.cpp
downsample_avg_2x_backward	implemented	ops_spatial.cpp

Arbitrary-scale 2D resample (nearest / bilinear / bicubic)

Slot	Vulkan	File
interp2d_forward	implemented	ops_spatial.cpp
interp2d_backward	implemented (chunk 9)	ops_spatial_bwd.cpp
interp2d_align_corners_forward	implemented	ops_spatial.cpp

2D padding (zero / reflect / replicate) — NCHW

Slot	Vulkan	File
pad2d_forward	implemented	ops_spatial.cpp
pad2d_backward	implemented (chunk 9)	ops_spatial_bwd.cpp

2D spatial slice / crop on NCHW

Slot	Vulkan	File
slice2d_forward	implemented	ops_spatial.cpp
slice2d_backward	implemented	ops_spatial.cpp

2D neighborhood unfold (spatial-preserving im2col) on NCHW

Slot	Vulkan	File
unfold2d_forward	implemented	ops_spatial.cpp

L2 normalize over channel axis (NCHW), per-pixel unit direction

Slot	Vulkan	File
l2_normalize_nchw_forward	implemented	ops_gnorm.cpp

Convex (mask-based) upsample (RAFT-style) on NCHW

Slot	Vulkan	File
convex_upsample_forward	implemented	ops_spatial.cpp

Per-row top-k (descending values + int32 indices)

Slot	Vulkan	File
top_k_rows	implemented	ops_topk.cpp

Adaptive avg pool 2D (NCHW), arbitrary output spatial size

Slot	Vulkan	File
adaptive_avg_pool2d_forward	implemented	ops_spatial.cpp
adaptive_avg_pool2d_backward	implemented (chunk 9)	ops_spatial_bwd.cpp

Max pool 2D (NCHW): forward returns Y + int32 flat-spatial Idx

Slot	Vulkan	File
max_pool2d_forward	implemented	ops_spatial.cpp
max_pool2d_backward	implemented (chunk 9)	ops_spatial_bwd.cpp

Row gather / scatter-add (general 2D — superset of embedding_lookup)

Slot	Vulkan	File
gather_rows	implemented	ops_spatial.cpp
scatter_rows_add	implemented (chunk 9)	ops_scatter.cpp
scatter_rows	implemented	ops_spatial.cpp

2D transposed convolution (NCHW) — forward + three backwards

Slot	Vulkan	File
conv_transpose2d_forward	implemented	ops_conv.cpp
conv_transpose2d_backward_input	implemented (chunk 9)	ops_conv.cpp
conv_transpose2d_backward_weight	implemented (chunk 9)	ops_conv.cpp
conv_transpose2d_backward_bias	implemented	ops_conv.cpp

SAM-style window partition / reverse (NCHW <-> windowed batch)

Slot	Vulkan	File
window_partition_forward	implemented	ops_spatial.cpp
window_reverse_forward	implemented	ops_spatial.cpp

FP16 linear (inference-only) + GEGLU family

Slot	Vulkan	File
linear_forward_batched_fp16	implemented	ops_linear.cpp
linear_forward_batched_ex	implemented	ops_linear.cpp
linear_forward_batched_fp16_act	implemented	ops_linear.cpp
geglu_forward	implemented	ops_glu.cpp
geglu_backward	implemented	ops_glu.cpp
geglu_exact_forward	implemented	ops_glu.cpp
geglu_exact_backward	implemented	ops_glu.cpp

Causal mask helper

Slot	Vulkan	File
build_causal_mask_row	implemented (chunk 6)	ops_misc.cpp

Cross-attention family (FP16 inference + FP32 train)

Slot	Vulkan	File
cross_attention_forward	implemented	ops_attention_proj.cpp
cross_attention_forward_with_attn	implemented	ops_attention_proj.cpp
self_attention_forward_train	implemented	ops_attention_proj.cpp
self_attention_backward	implemented	ops_attention_bwd.cpp
attention_token_moments	implemented (chunk 6)	ops_misc.cpp
cross_attention_forward_train	implemented (chunk 6)	ops_attention_proj.cpp
cross_attention_backward	implemented	ops_attention_bwd.cpp

FP16 LayerNorm inference + FP16 self-attention

Slot	Vulkan	File
layernorm_forward_inference_batched_fp16	implemented	ops_norm.cpp
self_attention_forward	implemented	ops_attention_proj.cpp

Flash attention family

Slot	Vulkan	File
flash_attention_forward	implemented	ops_attention.cpp
flash_attention_gqa_forward	implemented	ops_attention.cpp
flash_attention_windowed_forward	implemented	ops_attention.cpp
flash_attention_varlen_forward	implemented	ops_attention.cpp
flash_attention_packed_qkv_forward	implemented	ops_attention.cpp
flash_attention_packed_qkv_backward	implemented (chunk 9)	ops_fa_bwd.cpp
flash_attention_varlen_backward	implemented (chunk 9)	ops_fa_bwd.cpp
flash_attention_qkvo_forward	implemented	ops_attention.cpp
flash_attention_qkvo_backward	implemented (chunk 9)	ops_fa_bwd.cpp
flash_attention_backward	implemented (chunk 9)	ops_fa_bwd.cpp
flash_attention_project_kv	implemented	ops_attention.cpp
flash_attention_q_with_kv_cached_forward	implemented	ops_attention.cpp

NCHW <-> sequence transposes

Slot	Vulkan	File
nchw_to_sequence	implemented	ops_copy.cpp
sequence_to_nchw	implemented	ops_copy.cpp

2x2 pixel-unshuffle: (N,C,H,W) -> (N,4C,H/2,W/2)

Slot	Vulkan	File
spatial_merge_2x2_forward	implemented	ops_spatial.cpp

DC-AE up-shortcut: repeat_interleave + 2x pixel-shuffle

Slot	Vulkan	File
pixel_shuffle_upsample_2x_forward	implemented	ops_spatial.cpp

DiT unpatchify: token rows -> image (depth-to-space + channel keep)

Slot	Vulkan	File
patch_unpack_forward	implemented	ops_spatial.cpp

Diffusion ResBlock (forward + W8A16 + backward)

Slot	Vulkan	File
resblock_forward	implemented	ops_diffusion.cpp
resblock_forward_int8w_fp16	implemented	ops_diffusion.cpp
resblock_backward	implemented (chunk 9)	ops_diffusion.cpp

Matmul + RoPE + RMSNorm + SwiGLU + KV-cache + Llama family

Slot	Vulkan	File
matmul	implemented	ops_linear.cpp
matmul_abt	implemented	ops_linear.cpp
matmul_backward	implemented	ops_linear.cpp
lstm_forward_train	implemented (chunk 9)	ops_lstm.cpp
lstm_backward	implemented (chunk 9)	ops_lstm.cpp
rope_forward	implemented	ops_rope.cpp
rope_backward	implemented	ops_rope.cpp
rms_norm_forward	implemented	ops_norm.cpp
rms_norm_backward	implemented	ops_norm.cpp
swiglu_forward	implemented	ops_glu.cpp
swiglu_backward	implemented	ops_glu.cpp
kv_cache_append	implemented	ops_attention.cpp
flash_attention_decode	implemented	ops_attention.cpp
flash_attention_decode_masked	implemented	ops_attention.cpp

Public reductions

Slot	Vulkan	File
sum_rows	implemented	ops_reduce.cpp
sum_cols	implemented	ops_reduce.cpp
argmax_rows	implemented	ops_reduce.cpp
rows_count_above	implemented (chunk 6)	ops_misc.cpp
segment_softmax_stats	implemented	ops_topk.cpp
threshold_u8	implemented (chunk 6)	ops_misc.cpp

Diffusion sampler steps + timestep embedding

Slot	Vulkan	File
ddim_step	implemented	ops_diffusion.cpp
euler_step	implemented	ops_diffusion.cpp
dpmpp_2m_step	implemented	ops_diffusion.cpp
timestep_embedding	implemented	ops_diffusion.cpp

INT8 weight-only quantisation (W8A16)

Slot	Vulkan	File
matmul_int8w_fp16	implemented	ops_quant.cpp
conv2d_int8w_fp16_forward	implemented	ops_quant.cpp
linear_forward_batched_int8w_fp16	implemented	ops_quant.cpp
dequant_q4k_to_fp16	implemented	ops_quant.cpp
linear_forward_q4k_fp16	implemented	ops_quant.cpp
linear_forward_batched_q4k_fp16	implemented	ops_quant.cpp
dequant_q8_0_to_fp16	implemented	ops_quant.cpp
linear_forward_q8_0_fp16	implemented	ops_quant.cpp
linear_forward_batched_q8_0_fp16	implemented	ops_quant.cpp
dequant_q6k_to_fp16	implemented	ops_quant.cpp
linear_forward_q6k_fp16	implemented	ops_quant.cpp
linear_forward_batched_q6k_fp16	implemented	ops_quant.cpp
flash_attention_project_kv_int8w_fp16	implemented	ops_quant_attention.cpp
flash_attention_q_with_kv_cached_int8w_fp16	implemented	ops_quant_attention.cpp
flash_attention_qkvo_int8w_fp16	implemented	ops_quant_attention.cpp

DiT / diffusion extras: AdaLN modulation, axial RoPE, T5-bias attention

Slot	Vulkan	File
modulate	implemented	ops_glu.cpp
broadcast_mul	implemented	ops_glu.cpp
rope_apply	implemented	ops_rope.cpp
rope_apply_perhead	implemented	ops_rope.cpp
rope_apply_backward	implemented	ops_rope.cpp
rope_qkv_packed_inplace	implemented	ops_rope.cpp
rope_apply_mrope	implemented	ops_rope.cpp
self_attention_bias_forward	implemented	ops_attention_proj.cpp
rel_pos_bias_xl_forward	implemented	ops_attention_proj.cpp
self_attention_bias_int8w_fp16	implemented	ops_quant_attention.cpp
self_attention_decomposed_rel_pos_forward	implemented (chunk 6)	ops_vision.cpp
self_attention_decomposed_rel_pos_windowed_forward	implemented (chunk 6)	ops_vision.cpp

Spectral / FFT core (brosoundml)

Slot	Vulkan	File
complex_mul	implemented	ops_spectral.cpp
complex_mul_backward	implemented	ops_spectral.cpp
complex_abs	implemented	ops_spectral.cpp
complex_abs_backward	implemented	ops_spectral.cpp
complex_angle	implemented	ops_spectral.cpp
complex_from_polar	implemented	ops_spectral.cpp
fft	implemented	ops_spectral.cpp
ifft	implemented	ops_spectral.cpp
rfft	implemented	ops_spectral.cpp
irfft	implemented	ops_spectral.cpp
rfft_backward	implemented	ops_spectral.cpp
irfft_backward	implemented	ops_spectral.cpp

STFT / iSTFT (brosoundml)

Slot	Vulkan	File
stft	implemented	ops_spectral.cpp
stft_backward	implemented	ops_spectral.cpp
istft	implemented	ops_spectral.cpp
istft_backward	implemented	ops_spectral.cpp

1D convolution family (brosoundml)

Slot	Vulkan	File
conv_transpose1d_forward	implemented	ops_audio.cpp
conv_transpose1d_backward_input	implemented	ops_audio.cpp
conv_transpose1d_backward_weight	implemented	ops_audio.cpp
conv_transpose1d_backward_bias	implemented	ops_audio.cpp
causal_conv1d_update	implemented	ops_audio.cpp
pad1d_forward	implemented	ops_audio.cpp
pad1d_backward	implemented	ops_audio.cpp

Vocoder / codec activations (brosoundml)

Slot	Vulkan	File
snake_forward	implemented	ops_audio.cpp
snake_backward	implemented	ops_audio.cpp
elu_forward	implemented	ops_elementwise.cpp
elu_backward	implemented	ops_elementwise.cpp
leaky_relu_forward	implemented	ops_elementwise.cpp
leaky_relu_backward	implemented	ops_elementwise.cpp

Codec quantization (brosoundml CHUNK 5, family D)

Slot	Vulkan	File
vq_encode_forward	implemented	ops_audio.cpp
vq_encode_backward	implemented	ops_audio.cpp
fsq_quantize_forward	implemented	ops_audio.cpp
fsq_quantize_backward	implemented	ops_audio.cpp

1D resampling (brosoundml CHUNK 6, family E)

Slot	Vulkan	File
resample1d_forward	implemented	ops_audio.cpp
resample1d_backward	implemented	ops_audio.cpp

log / exp / round elementwise (brosoundml CHUNK 6, family G)

Slot	Vulkan	File
log_forward	implemented	ops_elementwise.cpp
log_backward	implemented	ops_elementwise.cpp
exp_forward	implemented	ops_elementwise.cpp
exp_backward	implemented	ops_elementwise.cpp
round_forward	implemented	ops_elementwise.cpp
round_backward	implemented	ops_elementwise.cpp

Autoregressive logit sampling (brosoundml CHUNK 7, family F)

Slot	Vulkan	File
sample_logits	implemented	ops_sampling.cpp
sample_logits_into	implemented	ops_sampling.cpp

Masked-diffusion token selection (OmniVoice codebook grids)

Slot	Vulkan	File
masked_diffusion_scores	implemented	ops_sampling.cpp
masked_diffusion_commit	implemented	ops_sampling.cpp

L2 norm + Gated Delta Rule (linear-attention text path)

Slot	Vulkan	File
l2_norm_forward	implemented	ops_norm.cpp
l2_norm_backward	implemented	ops_norm.cpp
gated_delta_rule_chunked	implemented (chunk 6)	ops_delta.cpp
gated_delta_rule_step	implemented (chunk 6)	ops_delta.cpp

BatchNorm (vision backbones: ResNet, DETR-R50, classic Mask2Former)

Slot	Vulkan	File
batch_norm_forward	implemented	ops_gnorm.cpp
batch_norm_inference	implemented	ops_gnorm.cpp
batch_norm_backward	implemented	ops_gnorm.cpp

Image preprocessing helpers (vision-model inference)

Slot	Vulkan	File
image_normalize	implemented	ops_diffusion.cpp
image_u8_to_f32_nhwc_to_nchw	implemented	ops_diffusion.cpp

Counter-based noise generation (Philox 4x32-10)

Slot	Vulkan	File
randn	implemented	ops_diffusion.cpp
rand_uniform	implemented	ops_diffusion.cpp
rand_bernoulli	implemented	ops_diffusion.cpp
randn_truncated	implemented	ops_diffusion.cpp

StyleGAN3-R synthesis primitives — sin/cos/rsqrt + pixel_norm

Slot	Vulkan	File
sin_forward	implemented	ops_elementwise.cpp
sin_backward	implemented	ops_elementwise.cpp
cos_forward	implemented	ops_elementwise.cpp
cos_backward	implemented	ops_elementwise.cpp
rsqrt_forward	implemented	ops_elementwise.cpp
rsqrt_backward	implemented	ops_elementwise.cpp
pixel_norm_forward	implemented	ops_norm.cpp
pixel_norm_backward	implemented	ops_norm.cpp

StyleGAN3-R core ops — modulated_conv2d / upfirdn2d / bias_act

Slot	Vulkan	File
modulated_conv2d_forward	implemented (chunk 6)	ops_vision.cpp
modulated_conv2d_backward	implemented (chunk 6)	ops_vision.cpp
upfirdn2d_forward	implemented (chunk 6)	ops_misc.cpp
upfirdn2d_backward	implemented (chunk 6)	ops_misc.cpp
bias_act_forward	implemented (chunk 6)	ops_misc.cpp
bias_act_backward	implemented (chunk 6)	ops_misc.cpp
filtered_lrelu_forward	null	-
filtered_lrelu_backward	null	-
add_channel_bias_inplace	implemented	ops_elementwise.cpp
add_row_bias_inplace	implemented	ops_elementwise.cpp
softmax_rows_forward	implemented	ops_norm.cpp