GCC Code Coverage Report

Directory: ./
Coverage: low: ≥ 0% medium: ≥ 75.0% high: ≥ 90.0%
Coverage Exec / Excl / Total
Lines: 90.1% 82 / 0 / 91
Functions: 100.0% 6 / 0 / 6
Branches: 53.7% 72 / 0 / 134
src/cpu/bias_act.cpp
Line Branch Exec Source
1 // ─── CPU bias_act (StyleGAN3) ───────────────────────────────────────────────
2 //
3 // Fused per-channel bias + activation + gain + clamp. FP32 reference mirroring
4 // NVlabs `_bias_act_ref`. CPU is FP32-only.
5 //
6 // X: (N, C*HW) — HW is the flattened spatial size; channel c owns the
7 // contiguous block [c*HW, (c+1)*HW) within each row.
8 // b: (C,1) or null. act: 0 = linear, 1 = lrelu. clamp < 0 ⇒ no clamp.
9 //
10 // Forward: t = X + b[c]; y = act(t); y *= gain; if clamp>=0: clip(±clamp).
11 // Backward: recompute t; dt = dY*gain*act'(t)*(clamp active ? 0 : 1);
12 // dX = dt (overwrite); dB[c] += Σ dt (accumulate — caller zeros).
13 //
14 // The clamp gradient mask uses the PRE-clamp value y_pre = gain*act(t): where
15 // |y_pre| > clamp the output saturated, so the gradient is zero there.
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 constexpr int ACT_LINEAR = 0;
28 constexpr int ACT_LRELU = 1;
29
30 135 inline void check_fp32(const ::brotensor::Tensor& t,
31 const char* op, const char* name) {
32
1/2
✓ Branch 0 taken 135 times.
✗ Branch 1 not taken.
 135 if (t.dtype != Dtype::FP32) {
33 ✗ throw std::runtime_error(std::string("brotensor: ") + op + ": " +
34 ✗ name + " must be FP32 (CPU backend is "
35 "FP32-only)");
36 }
37 135 }
38
39 60 inline void check_act(int act, const char* op) {
40
3/4
✓ Branch 0 taken 32 times.
✓ Branch 1 taken 28 times.
✓ Branch 2 taken 32 times.
✗ Branch 3 not taken.
 60 if (act != ACT_LINEAR && act != ACT_LRELU) {
41 ✗ throw std::runtime_error(std::string("brotensor: ") + op +
42 ": act must be 0 (linear) or 1 (lrelu)");
43 }
44 60 }
45
46 6958 inline float apply_act(float t, int act, float alpha) {
47
4/4
✓ Branch 0 taken 5403 times.
✓ Branch 1 taken 1555 times.
✓ Branch 2 taken 2621 times.
✓ Branch 3 taken 2782 times.
 6958 if (act == ACT_LRELU) return t > 0.0f ? t : alpha * t;
48 1555 return t; // linear
49 6958 }
50
51 4814 inline float act_grad(float t, int act, float alpha) {
52
4/4
✓ Branch 0 taken 3499 times.
✓ Branch 1 taken 1315 times.
✓ Branch 2 taken 1691 times.
✓ Branch 3 taken 1808 times.
 4814 if (act == ACT_LRELU) return t > 0.0f ? 1.0f : alpha;
53 1315 return 1.0f; // linear
54 4814 }
55
56 } // namespace
57
58 30 void bias_act_forward(const ::brotensor::Tensor& X, const ::brotensor::Tensor* b,
59 int N, int C, int HW, int act, float alpha,
60 float gain, float clamp, ::brotensor::Tensor& Y) {
61 30 check_fp32(X, "bias_act_forward", "X");
62
2/2
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
 30 if (b) check_fp32(*b, "bias_act_forward", "b");
63 30 check_act(act, "bias_act_forward");
64 30 const int cols = C * HW;
65
1/2
✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
 30 if (X.rows != N || X.cols != cols) {
66 ✗ throw std::runtime_error("bias_act_forward: X shape mismatch");
67 }
68
3/4
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
✓ Branch 2 taken 15 times.
✗ Branch 3 not taken.
 30 if (b && b->size() != C) {
69 ✗ throw std::runtime_error("bias_act_forward: b must have C elements");
70 }
71
1/6
✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
 30 if (Y.rows != N || Y.cols != cols || Y.dtype != Dtype::FP32) {
72 30 Y.resize(N, cols, Dtype::FP32);
73 30 }
74
2/4
✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 30 times.
 30 if (N == 0 || cols == 0) return;
75
76 30 const float* Xp = X.host_f32();
77
2/2
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
 30 const float* Bp = b ? b->host_f32() : nullptr;
78 30 float* Yp = Y.host_f32_mut();
79
80
2/2
✓ Branch 0 taken 30 times.
✓ Branch 1 taken 72 times.
 102 for (int n = 0; n < N; ++n) {
81
2/2
✓ Branch 0 taken 254 times.
✓ Branch 1 taken 72 times.
 326 for (int c = 0; c < C; ++c) {
82
2/2
✓ Branch 0 taken 135 times.
✓ Branch 1 taken 119 times.
 254 const float bias_v = Bp ? Bp[c] : 0.0f;
83 254 const size_t base = (static_cast<size_t>(n) * C + c) * HW;
84
2/2
✓ Branch 0 taken 4814 times.
✓ Branch 1 taken 254 times.
 5068 for (int k = 0; k < HW; ++k) {
85 4814 float y = apply_act(Xp[base + k] + bias_v, act, alpha) * gain;
86
2/2
✓ Branch 0 taken 2670 times.
✓ Branch 1 taken 2144 times.
 4814 if (clamp >= 0.0f) {
87
2/2
✓ Branch 0 taken 219 times.
✓ Branch 1 taken 1925 times.
 2144 if (y < -clamp) y = -clamp;
88
2/2
✓ Branch 0 taken 1379 times.
✓ Branch 1 taken 546 times.
 1925 else if (y > clamp) y = clamp;
89 2144 }
90 4814 Yp[base + k] = y;
91 4814 }
92 254 }
93 72 }
94 30 }
95
96 30 void bias_act_backward(const ::brotensor::Tensor& dY, const ::brotensor::Tensor& X,
97 const ::brotensor::Tensor* b,
98 int N, int C, int HW, int act, float alpha,
99 float gain, float clamp,
100 ::brotensor::Tensor& dX, ::brotensor::Tensor* dB) {
101 30 check_fp32(dY, "bias_act_backward", "dY");
102 30 check_fp32(X, "bias_act_backward", "X");
103
2/2
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
 30 if (b) check_fp32(*b, "bias_act_backward", "b");
104
2/2
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
 30 if (dB) check_fp32(*dB, "bias_act_backward", "dB");
105 30 check_act(act, "bias_act_backward");
106 30 const int cols = C * HW;
107
1/2
✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
 30 if (X.rows != N || X.cols != cols) {
108 ✗ throw std::runtime_error("bias_act_backward: X shape mismatch");
109 }
110
1/2
✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
 30 if (dY.rows != N || dY.cols != cols) {
111 ✗ throw std::runtime_error("bias_act_backward: dY shape mismatch");
112 }
113
3/4
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
✓ Branch 2 taken 15 times.
✗ Branch 3 not taken.
 30 if (b && b->size() != C) {
114 ✗ throw std::runtime_error("bias_act_backward: b must have C elements");
115 }
116
3/4
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
✓ Branch 2 taken 15 times.
✗ Branch 3 not taken.
 30 if (dB && dB->size() != C) {
117 ✗ throw std::runtime_error("bias_act_backward: dB must have C elements");
118 }
119
1/6
✗ Branch 0 not taken.
✓ Branch 1 taken 30 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
 30 if (dX.rows != N || dX.cols != cols || dX.dtype != Dtype::FP32) {
120 30 dX.resize(N, cols, Dtype::FP32);
121 30 }
122
2/4
✓ Branch 0 taken 30 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 30 times.
 30 if (N == 0 || cols == 0) return;
123
124 30 const float* dYp = dY.host_f32();
125 30 const float* Xp = X.host_f32();
126
2/2
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
 30 const float* Bp = b ? b->host_f32() : nullptr;
127 30 float* dXp = dX.host_f32_mut();
128
2/2
✓ Branch 0 taken 15 times.
✓ Branch 1 taken 15 times.
 30 float* dBp = dB ? dB->host_f32_mut() : nullptr;
129
130
2/2
✓ Branch 0 taken 30 times.
✓ Branch 1 taken 72 times.
 102 for (int n = 0; n < N; ++n) {
131
2/2
✓ Branch 0 taken 254 times.
✓ Branch 1 taken 72 times.
 326 for (int c = 0; c < C; ++c) {
132
2/2
✓ Branch 0 taken 135 times.
✓ Branch 1 taken 119 times.
 254 const float bias_v = Bp ? Bp[c] : 0.0f;
133 254 const size_t base = (static_cast<size_t>(n) * C + c) * HW;
134 254 float db_acc = 0.0f;
135
2/2
✓ Branch 0 taken 4814 times.
✓ Branch 1 taken 254 times.
 5068 for (int k = 0; k < HW; ++k) {
136 4814 const float t = Xp[base + k] + bias_v;
137 4814 float dt = dYp[base + k] * gain * act_grad(t, act, alpha);
138
2/2
✓ Branch 0 taken 2670 times.
✓ Branch 1 taken 2144 times.
 4814 if (clamp >= 0.0f) {
139 2144 const float y_pre = gain * apply_act(t, act, alpha);
140
4/4
✓ Branch 0 taken 1925 times.
✓ Branch 1 taken 219 times.
✓ Branch 2 taken 546 times.
✓ Branch 3 taken 1379 times.
 2144 if (y_pre < -clamp || y_pre > clamp) dt = 0.0f;
141 2144 }
142 4814 dXp[base + k] = dt; // overwrite
143 4814 db_acc += dt;
144 4814 }
145
2/2
✓ Branch 0 taken 119 times.
✓ Branch 1 taken 135 times.
 254 if (dBp) dBp[c] += db_acc; // accumulate — caller zeros
146 254 }
147 72 }
148 30 }
149
150 } // namespace brotensor::detail::cpu
151