aboutsummaryrefslogtreecommitdiff
path: root/nn.c
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-10-20 11:44:30 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-10-20 11:44:30 +0200
commit9b9005101e6170a226acb265428aac429be6edbf (patch)
treef89c76f4a5c2a00a72abe9b92468ac9f15083b94 /nn.c
downloadfacedet-9b9005101e6170a226acb265428aac429be6edbf.tar.gz
facedet-9b9005101e6170a226acb265428aac429be6edbf.zip
Init
Diffstat (limited to 'nn.c')
-rw-r--r--nn.c206
1 files changed, 206 insertions, 0 deletions
diff --git a/nn.c b/nn.c
new file mode 100644
index 0000000..aa4638e
--- /dev/null
+++ b/nn.c
@@ -0,0 +1,206 @@
+#include "gemm.h"
+#include "nn.h"
+
+#include <math.h>
+#include <string.h>
+
+void yunet_nn_conv2d(const float *x,
+ int cin,
+ int ih,
+ int iw,
+ const float *w,
+ const float *bias,
+ int cout,
+ int k,
+ int stride,
+ int pad,
+ int group,
+ int relu,
+ float *y,
+ int oh,
+ int ow) {
+ int cin_g = cin / group;
+ int cout_g = cout / group;
+ for (int oc = 0; oc < cout; oc++) {
+ int g = oc / cout_g;
+ const float *woc = w + (size_t)oc * cin_g * k * k;
+ float *yc = y + (size_t)oc * oh * ow;
+ for (int oy = 0; oy < oh; oy++) {
+ for (int ox = 0; ox < ow; ox++) {
+ float acc = bias[oc];
+ int iy0 = oy * stride - pad;
+ int ix0 = ox * stride - pad;
+ for (int ic = 0; ic < cin_g; ic++) {
+ const float *xc = x + (size_t)(g * cin_g + ic) * ih * iw;
+ const float *wk = woc + (size_t)ic * k * k;
+ for (int ky = 0; ky < k; ky++) {
+ int iy = iy0 + ky;
+ if (iy < 0 || iy >= ih)
+ continue;
+ const float *row = xc + (size_t)iy * iw;
+ const float *wr = wk + (size_t)ky * k;
+ for (int kx = 0; kx < k; kx++) {
+ int ix = ix0 + kx;
+ if (ix < 0 || ix >= iw)
+ continue;
+ acc += row[ix] * wr[kx];
+ }
+ }
+ }
+ yc[(size_t)oy * ow + ox] = relu && acc < 0.0f ? 0.0f : acc;
+ }
+ }
+ }
+}
+
+void yunet_nn_im2col3(const float *x, int cin, int ih, int iw, int stride, float *col, int oh, int ow) {
+ size_t n = (size_t)oh * ow;
+ for (int ic = 0; ic < cin; ic++) {
+ const float *xc = x + (size_t)ic * ih * iw;
+ for (int ky = 0; ky < 3; ky++) {
+ for (int kx = 0; kx < 3; kx++) {
+ float *dst = col + ((size_t)(ic * 9 + ky * 3 + kx)) * n;
+ for (int oy = 0; oy < oh; oy++) {
+ int iy = oy * stride - 1 + ky;
+ float *d = dst + (size_t)oy * ow;
+ if (iy < 0 || iy >= ih) {
+ memset(d, 0, sizeof(float) * ow);
+ continue;
+ }
+ const float *row = xc + (size_t)iy * iw;
+ if (stride == 1) {
+ int ox = 0;
+ if (kx == 0) {
+ d[0] = 0.0f;
+ ox = 1;
+ }
+ int last = ow;
+ if (kx == 2)
+ last = ow - 1;
+ if (last > ox)
+ memcpy(d + ox, row + ox - 1 + kx, sizeof(float) * (size_t)(last - ox));
+ if (kx == 2)
+ d[ow - 1] = 0.0f;
+ } else {
+ for (int ox = 0; ox < ow; ox++) {
+ int ix = ox * stride - 1 + kx;
+ d[ox] = (ix < 0 || ix >= iw) ? 0.0f : row[ix];
+ }
+ }
+ }
+ }
+ }
+ }
+}
+
+#if defined(__aarch64__) && !defined(YUNET_SCALAR)
+# include <arm_neon.h>
+void yunet_nn_dwconv3(
+ const float *x, int c, int ih, int iw, const float *w, const float *bias, int relu, float *y) {
+ for (int ch = 0; ch < c; ch++) {
+ const float *xc = x + (size_t)ch * ih * iw;
+ const float *wc = w + (size_t)ch * 9;
+ float *yc = y + (size_t)ch * ih * iw;
+ float b = bias[ch];
+ float32x4_t wk[9];
+ for (int i = 0; i < 9; i++)
+ wk[i] = vdupq_n_f32(wc[i]);
+ for (int oy = 0; oy < ih; oy++) {
+ const float *pm = oy > 0 ? xc + (size_t)(oy - 1) * iw : NULL;
+ const float *pc = xc + (size_t)oy * iw;
+ const float *pp = oy < ih - 1 ? xc + (size_t)(oy + 1) * iw : NULL;
+ float *out = yc + (size_t)oy * iw;
+ int x1 = iw - 1;
+ const int edges[2] = {0, x1};
+ for (int e = 0; e < 2; e++) {
+ int ex = edges[e];
+ float acc = b;
+ for (int ky = 0; ky < 3; ky++) {
+ const float *row = ky == 0 ? pm : ky == 1 ? pc : pp;
+ if (!row)
+ continue;
+ for (int kx = 0; kx < 3; kx++) {
+ int ix = ex - 1 + kx;
+ if (ix < 0 || ix >= iw)
+ continue;
+ acc += wc[ky * 3 + kx] * row[ix];
+ }
+ }
+ out[ex] = relu && acc < 0.0f ? 0.0f : acc;
+ }
+ int ox = 1;
+ for (; ox + 4 <= x1; ox += 4) {
+ float32x4_t acc = vdupq_n_f32(b);
+ if (pm) {
+ acc = vfmaq_f32(acc, wk[0], vld1q_f32(pm + ox - 1));
+ acc = vfmaq_f32(acc, wk[1], vld1q_f32(pm + ox));
+ acc = vfmaq_f32(acc, wk[2], vld1q_f32(pm + ox + 1));
+ }
+ acc = vfmaq_f32(acc, wk[3], vld1q_f32(pc + ox - 1));
+ acc = vfmaq_f32(acc, wk[4], vld1q_f32(pc + ox));
+ acc = vfmaq_f32(acc, wk[5], vld1q_f32(pc + ox + 1));
+ if (pp) {
+ acc = vfmaq_f32(acc, wk[6], vld1q_f32(pp + ox - 1));
+ acc = vfmaq_f32(acc, wk[7], vld1q_f32(pp + ox));
+ acc = vfmaq_f32(acc, wk[8], vld1q_f32(pp + ox + 1));
+ }
+ if (relu)
+ acc = vmaxq_f32(acc, vdupq_n_f32(0.0f));
+ vst1q_f32(out + ox, acc);
+ }
+ for (; ox < x1; ox++) {
+ float acc = b;
+ for (int ky = 0; ky < 3; ky++) {
+ const float *row = ky == 0 ? pm : ky == 1 ? pc : pp;
+ if (!row)
+ continue;
+ for (int kx = 0; kx < 3; kx++)
+ acc += wc[ky * 3 + kx] * row[ox - 1 + kx];
+ }
+ out[ox] = relu && acc < 0.0f ? 0.0f : acc;
+ }
+ }
+ }
+}
+#else
+void yunet_nn_dwconv3(
+ const float *x, int c, int ih, int iw, const float *w, const float *bias, int relu, float *y) {
+ yunet_nn_conv2d(x, c, ih, iw, w, bias, c, 3, 1, 1, c, relu, y, ih, iw);
+}
+#endif
+
+void yunet_nn_maxpool2(const float *x, int c, int ih, int iw, float *y) {
+ int oh = ih / 2, ow = iw / 2;
+ for (int ch = 0; ch < c; ch++) {
+ const float *xc = x + (size_t)ch * ih * iw;
+ float *yc = y + (size_t)ch * oh * ow;
+ for (int oy = 0; oy < oh; oy++)
+ for (int ox = 0; ox < ow; ox++) {
+ const float *p = xc + (size_t)(oy * 2) * iw + ox * 2;
+ float a = p[0] > p[1] ? p[0] : p[1];
+ float b = p[iw] > p[iw + 1] ? p[iw] : p[iw + 1];
+ yc[(size_t)oy * ow + ox] = a > b ? a : b;
+ }
+ }
+}
+
+void yunet_nn_resize2x(const float *x, int c, int ih, int iw, float *y) {
+ int oh = ih * 2, ow = iw * 2;
+ for (int ch = 0; ch < c; ch++) {
+ const float *xc = x + (size_t)ch * ih * iw;
+ float *yc = y + (size_t)ch * oh * ow;
+ for (int oy = 0; oy < oh; oy++) {
+ const float *row = xc + (size_t)(oy / 2) * iw;
+ float *out = yc + (size_t)oy * ow;
+ for (int ox = 0; ox < ow; ox++)
+ out[ox] = row[ox / 2];
+ }
+ }
+}
+
+void yunet_nn_add(const float *a, const float *b, float *y, size_t n) {
+ for (size_t i = 0; i < n; i++)
+ y[i] = a[i] + b[i];
+}
+
+float yunet_nn_sigmoid(float v) { return 1.0f / (1.0f + expf(-v)); }