src/cpu/l2_normalize.cpp
| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // ─── CPU L2 normalization over the channel axis (NCHW) ────────────────────── | ||
| 2 | // | ||
| 3 | // FP32 scalar host implementation of l2_normalize_nchw_forward. For every | ||
| 4 | // spatial position (n, h, w), normalizes the length-C channel vector to unit | ||
| 5 | // L2 norm with an epsilon floor on the divisor: | ||
| 6 | // Y[n,c,h,w] = X[n,c,h,w] / max(sqrt(sum_c X[n,c,h,w]^2), eps) | ||
| 7 | // The per-pixel vector normalize used by surface-normal / direction-field | ||
| 8 | // models (DSINE pred-normal normalize), feature-map L2 norm, and cosine-sim | ||
| 9 | // preprocessing. Distinct from l2_norm_forward (gated-deltanet per-head, last | ||
| 10 | // dim of a (L, H*D) layout) — this one is NCHW with the unit axis = channels. | ||
| 11 | // | ||
| 12 | // Reduction accumulates in double for stability; output is FP32. | ||
| 13 | // | ||
| 14 | // ── ACCUMULATION ── Y OVERWRITTEN. X and Y may alias. Inference-only: no | ||
| 15 | // backward (the bro pipeline never trains through this). | ||
| 16 | |||
| 17 | #include <brotensor/tensor.h> | ||
| 18 | |||
| 19 | #include <cmath> | ||
| 20 | #include <stdexcept> | ||
| 21 | #include <string> | ||
| 22 | |||
| 23 | namespace brotensor::detail::cpu { | ||
| 24 | |||
| 25 | namespace { | ||
| 26 | |||
| 27 | ✗ | [[noreturn]] inline void fail(const char* op, const std::string& reason) { | |
| 28 | ✗ | throw std::runtime_error(std::string("brotensor: ") + op + ": " + reason); | |
| 29 | ✗ | } | |
| 30 | |||
| 31 | 5 | inline void check_fp32(const ::brotensor::Tensor& t, | |
| 32 | const char* op, const char* name) { | ||
| 33 |
1/2✓ Branch 0 taken 5 times.
✗ Branch 1 not taken.
|
5 | if (t.dtype != Dtype::FP32) |
| 34 | ✗ | fail(op, std::string(name) + " must be FP32 (CPU backend is FP32-only)"); | |
| 35 | 5 | } | |
| 36 | |||
| 37 | } // namespace | ||
| 38 | |||
| 39 | 5 | void l2_normalize_nchw_forward(const ::brotensor::Tensor& X, | |
| 40 | int N, int C, int H, int W, | ||
| 41 | float eps, | ||
| 42 | ::brotensor::Tensor& Y) { | ||
| 43 | 5 | const char* op = "l2_normalize_nchw_forward"; | |
| 44 | 5 | check_fp32(X, op, "X"); | |
| 45 |
1/2✓ Branch 0 taken 5 times.
✗ Branch 1 not taken.
|
5 | if (N < 0 || C < 1 || H < 1 || W < 1) |
| 46 | ✗ | fail(op, "C/H/W must be >=1 and N >=0"); | |
| 47 |
1/2✓ Branch 0 taken 5 times.
✗ Branch 1 not taken.
|
5 | if (X.rows != N || X.cols != C * H * W) |
| 48 | ✗ | fail(op, "X shape must be (N, C*H*W)"); | |
| 49 |
1/6✗ Branch 0 not taken.
✓ Branch 1 taken 5 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
|
5 | if (Y.rows != N || Y.cols != C * H * W || Y.dtype != Dtype::FP32) |
| 50 | 5 | Y.resize(N, C * H * W, Dtype::FP32); | |
| 51 |
1/2✓ Branch 0 taken 5 times.
✗ Branch 1 not taken.
|
5 | if (N == 0) return; |
| 52 | |||
| 53 | 5 | const float* Xp = X.host_f32(); | |
| 54 | 5 | float* Yp = Y.host_f32_mut(); | |
| 55 | 5 | const int HW = H * W; | |
| 56 | 5 | const double epsd = static_cast<double>(eps); | |
| 57 | |||
| 58 |
2/2✓ Branch 0 taken 5 times.
✓ Branch 1 taken 6 times.
|
11 | for (int n = 0; n < N; ++n) { |
| 59 | 6 | const float* x_img = Xp + static_cast<long>(n) * C * HW; | |
| 60 | 6 | float* y_img = Yp + static_cast<long>(n) * C * HW; | |
| 61 |
2/2✓ Branch 0 taken 155 times.
✓ Branch 1 taken 6 times.
|
161 | for (int p = 0; p < HW; ++p) { |
| 62 | 155 | double ss = 0.0; | |
| 63 |
2/2✓ Branch 0 taken 4828 times.
✓ Branch 1 taken 155 times.
|
4983 | for (int c = 0; c < C; ++c) { |
| 64 | 4828 | const double v = x_img[static_cast<long>(c) * HW + p]; | |
| 65 | 4828 | ss += v * v; | |
| 66 | 4828 | } | |
| 67 | 155 | const double inv = 1.0 / std::max(std::sqrt(ss), epsd); | |
| 68 |
2/2✓ Branch 0 taken 4828 times.
✓ Branch 1 taken 155 times.
|
4983 | for (int c = 0; c < C; ++c) |
| 69 | 4828 | y_img[static_cast<long>(c) * HW + p] = | |
| 70 | 4828 | static_cast<float>(x_img[static_cast<long>(c) * HW + p] * inv); | |
| 71 | 155 | } | |
| 72 | 6 | } | |
| 73 | 5 | } | |
| 74 | |||
| 75 | } // namespace brotensor::detail::cpu | ||
| 76 |