|
|
e865c9 |
#ifndef LAYER_CONV_INC_CPP
|
|
|
e865c9 |
#define LAYER_CONV_INC_CPP
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
#include "layer.simple.inc.cpp"
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
struct Kernel {
|
|
|
e865c9 |
int sx, sy;
|
|
|
e865c9 |
int dx, dy;
|
|
|
e865c9 |
int ox, oy;
|
|
|
e865c9 |
|
|
|
e865c9 |
inline Kernel():
|
|
|
e865c9 |
sx(), sy(), dx(), dy(), ox(), oy() { }
|
|
|
e865c9 |
inline Kernel(int sx, int sy, int dx, int dy, int ox, int oy):
|
|
|
e865c9 |
sx(sx), sy(sy), dx(dx), dy(dy), ox(ox), oy(oy) { }
|
|
|
e865c9 |
inline Kernel(int s, int d, int o):
|
|
|
e865c9 |
sx(s), sy(s), dx(d), dy(d), ox(o), oy(o) { }
|
|
|
e865c9 |
inline operator bool() const
|
|
|
e865c9 |
{ return sx > 0 && sy > 0 && dx > 0 && dy > 0; }
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void print(const char *prefix = nullptr) const {
|
|
|
e865c9 |
if (prefix && *prefix) printf("%s: ", prefix);
|
|
|
e865c9 |
printf("x(sdo): %d %d %d, y(sdo): %d %d %d\n", sx, dx, ox, sy, dy, oy);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
void printYX(const char *prefix = nullptr) const {
|
|
|
e865c9 |
if (prefix && *prefix) printf("%s: ", prefix);
|
|
|
e865c9 |
printf("y(sdo): %d %d %d, x(sdo): %d %d %d\n", sy, dy, oy, sx, dx, ox);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
};
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
template<typename iter=""></typename>
|
|
|
e865c9 |
void iterateTestConvolution(Layout cl, Layout pl, Kernel k, Neuron *c_neurons, Neuron *p_neurons, Weight *weights) {
|
|
|
e865c9 |
if (!cl) return;
|
|
|
e865c9 |
assert(pl);
|
|
|
e865c9 |
assert(k);
|
|
|
e865c9 |
assert(c_neurons);
|
|
|
e865c9 |
assert(p_neurons);
|
|
|
e865c9 |
assert(weights);
|
|
|
e865c9 |
assert(pl.x0 + k.ox >= 0 && pl.x0 + (cl.getW()-1)*k.dx + k.ox + k.sx <= pl.sx);
|
|
|
e865c9 |
assert(pl.y0 + k.oy >= 0 && pl.y0 + (cl.getH()-1)*k.dy + k.oy + k.sy <= pl.sy);
|
|
|
e865c9 |
|
|
|
e865c9 |
for(int cy = cl.y0; cy < cl.y1; ++cy)
|
|
|
e865c9 |
for(int cx = cl.x0; cx < cl.x1; ++cx)
|
|
|
e865c9 |
for(int cz = cl.z0; cz < cl.z1; ++cz) {
|
|
|
e865c9 |
int ci = (cy*cl.sx + cx)*cl.sz + cz;
|
|
|
e865c9 |
Neuron &cn = c_neurons[ci];
|
|
|
e865c9 |
typename Iter::AccumType a = {};
|
|
|
e865c9 |
Iter::init(cn, a);
|
|
|
e865c9 |
|
|
|
e865c9 |
for(int ky = 0; ky < k.sy; ++ky)
|
|
|
e865c9 |
for(int kx = 0; kx < k.sx; ++kx)
|
|
|
e865c9 |
for(int pz = pl.z0; pz < pl.z1; ++pz) {
|
|
|
e865c9 |
int wi = ((cy - cl.y0)*cl.getW() + cx - cl.x0)*cl.getD() + cz - cl.z0;
|
|
|
e865c9 |
wi = ((wi*k.sy + ky)*k.sx + kx)*pl.getD() + pz - pl.z0;
|
|
|
e865c9 |
Weight &w = weights[wi];
|
|
|
e865c9 |
|
|
|
e865c9 |
int px = pl.x0 + (cx - cl.x0)*k.dx + k.ox + kx;
|
|
|
e865c9 |
int py = pl.y0 + (cy - cl.y0)*k.dy + k.oy + ky;
|
|
|
e865c9 |
int pi = (py*pl.sx + px)*pl.sz + pz;
|
|
|
e865c9 |
Neuron &pn = p_neurons[pi];
|
|
|
e865c9 |
|
|
|
e865c9 |
Iter::iter(pn, w, a);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
Iter::done(cn, a);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
template<typename iter=""></typename>
|
|
|
e865c9 |
void iterateConvolution(Layout cl, Layout pl, Layout wl, Kernel k, Neuron *c_neurons, Neuron *p_neurons, Weight *weights) {
|
|
|
e865c9 |
if (!cl) return;
|
|
|
e865c9 |
assert(pl);
|
|
|
e865c9 |
assert(wl);
|
|
|
e865c9 |
assert(k);
|
|
|
e865c9 |
assert(c_neurons);
|
|
|
e865c9 |
assert(p_neurons);
|
|
|
e865c9 |
assert(weights);
|
|
|
e865c9 |
assert(cl.isSubLayoutOf(wl));
|
|
|
e865c9 |
assert(pl.x0 + k.ox >= 0 && pl.x0 + (wl.getW()-1)*k.dx + k.ox + k.sx <= pl.sx);
|
|
|
e865c9 |
assert(pl.y0 + k.oy >= 0 && pl.y0 + (wl.getH()-1)*k.dy + k.oy + k.sy <= pl.sy);
|
|
|
e865c9 |
|
|
|
e865c9 |
int c_h = cl.getH();
|
|
|
e865c9 |
int c_w = cl.getW();
|
|
|
e865c9 |
int c_d = cl.getD();
|
|
|
e865c9 |
int c_swz = c_w*cl.sz;
|
|
|
e865c9 |
int c_shxz = c_h*cl.sx*cl.sz;
|
|
|
e865c9 |
int c_dx = cl.sz - c_d;
|
|
|
e865c9 |
int c_dy = (cl.sx - c_w)*cl.sz;
|
|
|
e865c9 |
|
|
|
e865c9 |
int p_d = pl.getD();
|
|
|
e865c9 |
int p_dx = k.dx*pl.sz;
|
|
|
e865c9 |
int p_dy = k.dy*pl.sx*pl.sz - c_w*p_dx;
|
|
|
e865c9 |
|
|
|
e865c9 |
int k_sxd = k.sx*p_d;
|
|
|
e865c9 |
int k_syxd = k.sy*k_sxd;
|
|
|
e865c9 |
int p_ddy = (pl.sx - k.sx)*pl.sz;
|
|
|
e865c9 |
int p_ddx = pl.sz - p_d;
|
|
|
e865c9 |
|
|
|
e865c9 |
int w_w = wl.getW();
|
|
|
e865c9 |
int w_d = wl.getD();
|
|
|
e865c9 |
int w_dx = (w_d - c_d)*k_syxd;
|
|
|
e865c9 |
int w_dy = (w_w - c_w)*w_d*k_syxd;
|
|
|
e865c9 |
|
|
|
e865c9 |
int cx0 = cl.x0 - wl.x0;
|
|
|
e865c9 |
int cy0 = cl.y0 - wl.y0;
|
|
|
e865c9 |
int cz0 = cl.z0 - wl.z0;
|
|
|
e865c9 |
|
|
|
e865c9 |
Neuron *icn = c_neurons + (cl.y0*cl.sx + cl.x0)*cl.sz + cl.z0;
|
|
|
e865c9 |
Neuron *ipn = p_neurons + ((pl.y0 + cy0*k.dy + k.oy)*pl.sx + pl.x0 + cx0*k.dx + k.ox)*pl.sz + pl.z0;
|
|
|
e865c9 |
Weight *iw = weights + ((cy0*w_w + cx0)*w_d + cz0)*k_syxd;
|
|
|
e865c9 |
|
|
|
e865c9 |
for(Neuron *e = icn + c_shxz; icn < e; icn += c_dy, ipn += p_dy, iw += w_dy)
|
|
|
e865c9 |
for(Neuron *e = icn + c_swz; icn < e; icn += c_dx, ipn += p_dx, iw += w_dx)
|
|
|
e865c9 |
for(Neuron *e = icn + c_d; icn < e; ++icn) {
|
|
|
e865c9 |
typename Iter::AccumType a;
|
|
|
e865c9 |
Iter::init(*icn, a);
|
|
|
e865c9 |
|
|
|
e865c9 |
Neuron *iipn = ipn;
|
|
|
e865c9 |
for(Weight *e = iw + k_syxd; iw < e; iipn += p_ddy)
|
|
|
e865c9 |
for(Weight *e = iw + k_sxd; iw < e; iipn += p_ddx)
|
|
|
e865c9 |
for(Weight *e = iw + p_d; iw < e; ++iw, ++iipn)
|
|
|
e865c9 |
Iter::iter(*iipn, *iw, a);
|
|
|
e865c9 |
|
|
|
e865c9 |
Iter::done(*icn, a);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
template<typename iter=""></typename>
|
|
|
e865c9 |
void iterateConvolutionPoint(Layout cl, Layout pl, Layout wl, Kernel k, int kx, int ky, Neuron *c_neurons, Neuron *p_neurons, Weight *weights) {
|
|
|
e865c9 |
if (!cl) return;
|
|
|
e865c9 |
assert(pl);
|
|
|
e865c9 |
assert(wl);
|
|
|
e865c9 |
assert(k);
|
|
|
e865c9 |
assert(c_neurons);
|
|
|
e865c9 |
assert(p_neurons);
|
|
|
e865c9 |
assert(weights);
|
|
|
e865c9 |
assert(cl.isSubLayoutOf(wl));
|
|
|
e865c9 |
assert(kx >= 0 && kx < k.sx);
|
|
|
e865c9 |
assert(ky >= 0 && ky < k.sy);
|
|
|
e865c9 |
assert(pl.x0 + k.ox >= 0 && pl.x0 + (wl.getW()-1)*k.dx + k.ox + k.sx <= pl.sx);
|
|
|
e865c9 |
assert(pl.y0 + k.oy >= 0 && pl.y0 + (wl.getH()-1)*k.dy + k.oy + k.sy <= pl.sy);
|
|
|
e865c9 |
|
|
|
e865c9 |
int c_h = cl.getH();
|
|
|
e865c9 |
int c_w = cl.getW();
|
|
|
e865c9 |
int c_d = cl.getD();
|
|
|
e865c9 |
int c_swz = c_w*cl.sz;
|
|
|
e865c9 |
int c_shxz = c_h*cl.sx*cl.sz;
|
|
|
e865c9 |
int c_dx = cl.sz - c_d;
|
|
|
e865c9 |
int c_dy = (cl.sx - c_w)*cl.sz;
|
|
|
e865c9 |
|
|
|
e865c9 |
int p_d = pl.getD();
|
|
|
e865c9 |
int p_dx = k.dx*pl.sz;
|
|
|
e865c9 |
int p_dy = k.dy*pl.sx*pl.sz - c_w*p_dx;
|
|
|
e865c9 |
|
|
|
e865c9 |
int k_sxd = k.sx*p_d;
|
|
|
e865c9 |
int k_syxd = k.sy*k_sxd;
|
|
|
e865c9 |
|
|
|
e865c9 |
int w_w = wl.getW();
|
|
|
e865c9 |
int w_d = wl.getD();
|
|
|
e865c9 |
int w_dz = k_syxd - p_d;
|
|
|
e865c9 |
int w_dx = (w_d - c_d)*k_syxd;
|
|
|
e865c9 |
int w_dy = (w_w - c_w)*w_d*k_syxd;
|
|
|
e865c9 |
|
|
|
e865c9 |
int cx0 = cl.x0 - wl.x0;
|
|
|
e865c9 |
int cy0 = cl.y0 - wl.y0;
|
|
|
e865c9 |
int cz0 = cl.z0 - wl.z0;
|
|
|
e865c9 |
|
|
|
e865c9 |
Neuron *icn = c_neurons + (cl.y0*cl.sx + cl.x0)*cl.sz + cl.z0;
|
|
|
e865c9 |
Neuron *ipn = p_neurons + ((pl.y0 + cy0*k.dy + k.oy + ky)*pl.sx + pl.x0 + cx0*k.dx + k.ox + kx)*pl.sz + pl.z0;
|
|
|
e865c9 |
Weight *iw = weights + ((cy0*w_w + cx0)*w_d + cz0)*k_syxd + ky*k_sxd + kx*p_d;
|
|
|
e865c9 |
|
|
|
e865c9 |
for(Neuron *e = icn + c_shxz; icn < e; icn += c_dy, ipn += p_dy, iw += w_dy)
|
|
|
e865c9 |
for(Neuron *e = icn + c_swz; icn < e; icn += c_dx, ipn += p_dx, iw += w_dx)
|
|
|
e865c9 |
for(Neuron *e = icn + c_d; icn < e; ++icn, ipn -= p_d, iw += w_dz)
|
|
|
e865c9 |
for(Neuron *e = ipn + p_d; ipn < e; ++ipn, ++iw)
|
|
|
e865c9 |
Iter::iter2(*icn, *ipn, *iw);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
template<func func=""></func>
|
|
|
e865c9 |
class LayerConv: public Layer {
|
|
|
e865c9 |
public:
|
|
|
e865c9 |
Kernel kernel;
|
|
|
e865c9 |
|
|
|
e865c9 |
LayerConv(Layer &prev, const Layout &layout, const Kernel &kernel, Weight *weights = nullptr):
|
|
|
e865c9 |
Layer(&prev, layout, layout.getActiveCount()*kernel.sx*kernel.sy*prev.back().layout.getD(), weights),
|
|
|
e865c9 |
kernel(kernel)
|
|
|
e865c9 |
{
|
|
|
e865c9 |
assert(kernel);
|
|
|
e865c9 |
if (ownWeights) fillWeights(-1, 1);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void pass(Barrier &barrier) override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
static inline void init(Neuron&, AccumType &a) { a.v = 0; }
|
|
|
e865c9 |
static inline void iter(Neuron &n, Weight &w, AccumType &a) { a.v += n.v * w.w; }
|
|
|
e865c9 |
static inline void done(Neuron &n, AccumType &a) { func(n, a.v); }
|
|
|
e865c9 |
};
|
|
|
e865c9 |
iterateConvolution(mtLayouts[barrier.tid], prev->layout, layout, kernel, neurons, prev->neurons, weights);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void backpassWeights(Barrier &barrier) override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
static inline void init(Neuron &n, AccumType &a) { a.v = n.d; }
|
|
|
e865c9 |
static inline void iter(Neuron &n, Weight &w, AccumType &a) { w.w += n.v * a.v; }
|
|
|
e865c9 |
};
|
|
|
e865c9 |
iterateConvolution(mtLayouts[barrier.tid], prev->layout, layout, kernel, neurons, prev->neurons, weights);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void backpassDeltas(Barrier &barrier) override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
static inline void iter2(Neuron &cn, Neuron &pn, Weight &w) { pn.a.v += cn.d * w.w; }
|
|
|
e865c9 |
static inline void iter3(Neuron &n) { n.d *= n.a.v; n.a.v = 0; }
|
|
|
e865c9 |
};
|
|
|
e865c9 |
int ksx = kernel.sx, ksy = kernel.sy;
|
|
|
e865c9 |
for(int kx = 0; kx < ksx; ++kx)
|
|
|
e865c9 |
for(int ky = 0; ky < ksy; ++ky) {
|
|
|
e865c9 |
iterateConvolutionPoint(mtLayouts[barrier.tid], prev->layout, layout, kernel, kx, ky, neurons, prev->neurons, weights);
|
|
|
e865c9 |
barrier.wait();
|
|
|
e865c9 |
}
|
|
|
036a8f |
iterateNeurons(mtPrevLayouts[barrier.tid], prev->neurons);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void testPass() override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
static inline void init(Neuron&, AccumType &a) { a.v = 0; }
|
|
|
e865c9 |
static inline void iter(Neuron &n, Weight &w, AccumType &a) { a.v += n.v * w.w; }
|
|
|
e865c9 |
static inline void done(Neuron &n, AccumType &a) { func(n, a.v); }
|
|
|
e865c9 |
};
|
|
|
e865c9 |
iterateTestConvolution(layout, prev->layout, kernel, neurons, prev->neurons, weights);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void testBackpass() override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
static inline void init(Neuron &n, AccumType &a) { a.v = n.d; }
|
|
|
e865c9 |
static inline void iter(Neuron &n, Weight &w, AccumType &a) { n.a.v += a.v * w.w; w.w += a.v * n.v; }
|
|
|
e865c9 |
static inline void iter3(Neuron &n) { n.d *= n.a.v; n.a.v = 0; }
|
|
|
e865c9 |
};
|
|
|
e865c9 |
clearAccum();
|
|
|
e865c9 |
iterateTestConvolution(layout, prev->layout, kernel, neurons, prev->neurons, weights);
|
|
|
e865c9 |
iterateNeurons(prev->layout, prev->neurons);
|
|
|
e865c9 |
clearAccum();
|
|
|
e865c9 |
}
|
|
|
e865c9 |
};
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
template<func func=""></func>
|
|
|
e865c9 |
class LayerDeconv: public Layer {
|
|
|
e865c9 |
public:
|
|
|
e865c9 |
Kernel kernel;
|
|
|
e865c9 |
|
|
|
e865c9 |
LayerDeconv(Layer &prev, const Layout &layout, const Kernel &kernel, Weight *weights = nullptr):
|
|
|
e865c9 |
Layer(&prev, layout, prev.back().layout.getActiveCount()*kernel.sx*kernel.sy*layout.getD(), weights),
|
|
|
e865c9 |
kernel(kernel)
|
|
|
e865c9 |
{
|
|
|
e865c9 |
assert(kernel);
|
|
|
e865c9 |
if (ownWeights) fillWeights(-1, 1);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void pass(Barrier &barrier) override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
static inline void iter2(Neuron &cn, Neuron &pn, Weight &w) { pn.a.v += cn.v * w.w; }
|
|
|
e865c9 |
static inline void iter3(Neuron &n) { func(n, n.a.v); n.a.v = 0; }
|
|
|
e865c9 |
};
|
|
|
e865c9 |
int k_sx = kernel.sx, k_sy = kernel.sy;
|
|
|
e865c9 |
for(int kx = 0; kx < k_sx; ++kx)
|
|
|
e865c9 |
for(int ky = 0; ky < k_sy; ++ky) {
|
|
|
036a8f |
iterateConvolutionPoint(mtPrevLayouts[barrier.tid], layout, prev->layout, kernel, kx, ky, prev->neurons, neurons, weights);
|
|
|
e865c9 |
barrier.wait();
|
|
|
e865c9 |
}
|
|
|
e865c9 |
iterateNeurons(mtLayouts[barrier.tid], neurons);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void backpassWeights(Barrier &barrier) override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
static inline void init(Neuron &n, AccumType &a) { a.v = n.v; }
|
|
|
e865c9 |
static inline void iter(Neuron &n, Weight &w, AccumType &a) { w.w += n.d * a.v; }
|
|
|
e865c9 |
};
|
|
|
8e5348 |
iterateConvolution(mtPrevLayouts[barrier.tid], layout, prev->layout, kernel, prev->neurons, neurons, weights);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void backpassDeltas(Barrier &barrier) override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
static inline void init(Neuron&, AccumType &a) { a.v = 0; }
|
|
|
e865c9 |
static inline void iter(Neuron &n, Weight &w, AccumType &a) { a.v += n.d * w.w; }
|
|
|
e865c9 |
static inline void done(Neuron &n, AccumType &a) { n.d *= a.v; }
|
|
|
e865c9 |
};
|
|
|
8e5348 |
iterateConvolution(mtPrevLayouts[barrier.tid], layout, prev->layout, kernel, prev->neurons, neurons, weights);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void testPass() override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
static inline void init(Neuron &n, AccumType &a) { a.v = n.v; }
|
|
|
e865c9 |
static inline void iter(Neuron &n, Weight &w, AccumType &a) { n.a.v += a.v * w.w; }
|
|
|
e865c9 |
static inline void iter3(Neuron &n) { func(n, n.a.v); n.a.v = 0; }
|
|
|
e865c9 |
};
|
|
|
e865c9 |
clearAccum();
|
|
|
e865c9 |
iterateTestConvolution(prev->layout, layout, kernel, prev->neurons, neurons, weights);
|
|
|
e865c9 |
iterateNeurons(layout, neurons);
|
|
|
e865c9 |
clearAccum();
|
|
|
e865c9 |
}
|
|
|
e865c9 |
|
|
|
e865c9 |
|
|
|
e865c9 |
void testBackpass() override {
|
|
|
e865c9 |
struct I: public Iter {
|
|
|
e865c9 |
struct AccumType: public Accum { NeuronReal vv; };
|
|
|
e865c9 |
static inline void init(Neuron &n, AccumType &a) { a.v = 0; a.vv = n.v; }
|
|
|
e865c9 |
static inline void iter(Neuron &n, Weight &w, AccumType &a) { a.v += n.d * w.w; w.w += n.d * a.vv; }
|
|
|
e865c9 |
static inline void done(Neuron &n, AccumType &a) { n.d *= a.v; }
|
|
|
e865c9 |
};
|
|
|
e865c9 |
iterateTestConvolution(prev->layout, layout, kernel, prev->neurons, neurons, weights);
|
|
|
e865c9 |
}
|
|
|
e865c9 |
};
|
|
|
e865c9 |
|
|
|
e865c9 |
#endif
|