// SPDX-FileCopyrightText: © 2026 Tenstorrent USA, Inc.
// SPDX-License-Identifier: Apache-2.0
//
// SuperPoint NMS, step 3 + keypoint candidate compaction (data movement only, ttnn.generic_op).
// M = 9x9 window max in strip layout [H, SW, 32] (this core's L1 shard), P = padded strip scores
// [H, SW + 2*PAD, 32] (this core's L1 shard). Writes the natural dense NMS map row by row:
//   out[y, x] = P[y, xw + PAD, l] if it equals M[y, xw, l] else 0,  x = l*SW + xw
// (bit-identical to where(s == maxpool9x9(s), s, 0) for s >= 0). out: [H, W] ROW_MAJOR interleaved.
//
// Additionally every RISC appends the keypoint candidates of its rows, in raster order, to its own
// slot of the candidate tensor C [NSLOT, CAP + 1] uint32 (one DRAM page per slot):
//   C[slot, 0] = number of candidates (may exceed CAP -> overflow, the host falls back),
//   C[slot, 1 + k] = (bf16 score bits << 16) | ((y - y_first) * W + x)
// with the host's extract_keypoints rule: BORDER <= y < H - BORDER, BORDER <= x < W - BORDER and
// float(score) > threshold  <=>  int16(score bits) > int16(THR) (THR = fp32 bits of the threshold
// >> 16; exact for bf16 scores, see postprocess._bf16_threshold_bits).
// THR and BORDER are per-request values (taxonomy class C): read from the 64-byte parameter tensor
// PRM = [THR, BORDER, 0...] (uint32, DRAM) that the host rewrites only when they change.
// KP_REC (records for the DMA-only compaction kp_compact3.cpp): every RISC also writes its candidates
// as final 4-word keypoint-header entries into its slot of the records tensor R (L1, this core's shard,
// slot = PROC: 16 B count word + CAP x 16 B): ((y << 16) | x, score bits, (y0c*wc << 16) | y1c*wc,
// (x0c << 16) | x1c); the cell indices of the bilinear taps are  q = floor((2v - 7)(c - 1) / (16c - 9)),
// v0c = clamp(q), v1c = clamp(q + 1)  (the host asserts this equals postprocess.SampleTables).
// Extra RT arg 9: records address.  Defines: KP_WC, KP_HC (cells per row / column).
// (the unfold itself is nms_unfold_fn.inc, prepended to this source by models/tt/nms_kernels.py)

void kernel_main() {
    constexpr uint32_t SW = get_compile_time_arg_val(1);
    unfold_rows<0, get_compile_time_arg_val(0), SW * 16>(reinterpret_cast<const uint32_t*>(get_arg_val<uint32_t>(0)), 0);
}