[michalharakal]

Summary

Seventh step of the JVM ↔ wasm parity series (after #1363, #1364, #1368, #1369, #1426, #1427), and the last compute gap on the list: convolutions. On the JVM conv1d / conv2d / conv3d run as im2col + the Vector API (JvmVectorKernels); in a browser they were the common per-element loop — the only FP32 op a forward pass spends time in that was still scalar Kotlin on wasm after the fused-op step.

skainet_f32_conv (f32_conv.c, shared native tree, every native build) — one kernel for all three ranks over three spatial dims (a lower rank passes the identity geometry for the leading ones). Per batch and group it gathers the receptive fields patch-major into a caller-provided scratch (col[p, :] is patch p's window, zero in the padding; whole rows copied with memcpy when the width dilation is 1) and multiplies them against the weight rows with the existing skainet_fp32_matmul_wt — every output is a contiguous dot product of a weight row and a patch row, so the SIMD128 / NEON body that kernel already has serves the convolution, accumulated in ascending k. Gather and GEMM thread over patches through the shared row pool; a bias of [c_out] is added in the kernel. Groups (depthwise included), stride, padding and dilation are all geometry. The 22-int geometry block travels as one pointer (SKAINET_CONV_GEO_INTS).

WasmCpuOps.conv1d/conv2d/conv3d — dense FP32 input, weight and optional bias ([c_out] or [1, c_out, 1, …]) go to the sidecar: the weight is a parameter and stays resident after its first call (denseWeightHandle), a heap input is staged per call and a resident one read in place, the output is resident under the usual budget — so conv → relu → … never leaves linear memory. Anything the common validation would reject stays on the common path, which raises its message. Sidecar rebuilt (SidecarArtifacts.kt); exports grow by one.

Matrix, bench, docs — the Kernel packs table gains Float32 conv1d / conv2d / conv3d (im2col + GEMM): panama-vector on the JVM (the one row whose JVM cell is the ops layer itself, not a provider — noted in the test), wasm-simd128 on JS/WASM. engine-conv2d (a 3 × 3 same-padding layer, 64 → 64 channels over 56 × 56, with bias) joins jvm-cpu-publish and wasm-bench with identical shapes. eager-execution.adoc states what still runs as common Kotlin on wasm (memory movement: reshape / permute / concat / gather / unfold / pooling, and the integer and narrow-float element paths); the browser-lane page's status line and the CHANGELOG carry the step.

Verification

• :skainet-backends:skainet-backend-wasm-simd:wasmJsNodeTest (13/13) — new convolutions run as im2col and GEMM in the sidecar and match the scalar lane: conv2d with and without bias and with the 4-D bias form, a 2 × 3 kernel with stride, dilation and asymmetric padding, depthwise (groups = c_in), two groups with stride, 1 × 1; conv1d with stride, dilation and groups; conv3d with stride and padding; a conv → relu chain and a resident input, every output resident and the input never downloaded; the weight resident after its first call and not re-uploaded by the second. All against DefaultCpuOps on the same inputs.
• :skainet-backends:skainet-backend-native-cpu:jvmTest — KernelSupportMatrixTest (the CMake build compiles the new file into libskainet_kernels), NativeFusedKernelsTest; generateKernelMatrix; the host gcc -Wall -Wextra compile of f32_conv.c is clean.
• scripts/pr-gate.sh --quick, full gate.

Measured

engine-conv2d, batch 1, 64 → 64 channels, 56 × 56, 3 × 3, pad 1, bias (231 MFLOP per call); 12-core laptop.

Tier	GFLOP/s
wasm scalar (before)	0.04
wasm simd128-st	6.76
wasm simd128-mt (4 workers)	22.0
JVM panama (im2col + Vector API, DefaultCpuOpsJvm)	10.7
JVM ffm (same path — no native conv binding on the JVM)	11.6

The sidecar convolution lands at roughly half the sidecar's dense GEMM rate on the same tier (14.1 st / 43.6 mt for the 1024³ GEMM): the gather is memory-bound and the dot kernel is the one-row-at-a-time x · Wᵀ body, not the 4 × 8 register-blocked GEMM. The JVM path is single-threaded and keeps its Vector API dot per output; the four-worker sidecar now outruns it. Binding skainet_f32_conv on the JVM through FFM (an Fp32ConvKernel face beside the fused-op faces) would close that in the other direction; left out of this PR as agreed for the series — noted as a follow-up.