path: root/include/minimax_cnn1986.c

#include "minimax_cnn1986.h"

// ===================================================================
// Globals
// ===================================================================

const float infty = 3.0;
char ct1986_ptn[9];

// ===================================================================
// Implementation of a small convolutional neural network
// ===================================================================

static float flattened[CONV_NUM+2];
static float dense1[DENSE1_NUM];
static float dense2[DENSE2_NUM];

#ifndef DETERMINISTIC
union u_f {
	uint32_t u;
	float f;
};

static uint32_t state = 1;
static union u_f fudge;

#define DOXORSHIFT {		\
		state ^= state << 13; \
		state ^= state >> 17; \
		state ^= state << 5; \
		fudge.u = 0x3f800000 | state >> 10; \
		fudge.f = (fudge.f - 1.5) * 0.01; \
	}
#endif

#define RELU(x) ((x) = ((x)<0)?0:(x))

float
ct1986_evaluate_win(void) {
	/* ------------------ *
	 *  Convolution layer *
	 * ------------------ */
	// for each kernel
	for (uint8_t kern = 0; kern < KERN_NUM; kern++) {
		// the stride is 1, march across the board
		for (uint8_t bx = 0; bx < KERN_OSIZE; bx++) {
			for (uint8_t by = 0; by < KERN_OSIZE; by++) {
				flattened[kern+KERN_NUM*(bx+by*KERN_OSIZE)] =
					conv2d_biases[kern];
				// Compute the convolution for this position
				for (uint8_t ky = 0; ky < KERN_SIZE; ky++) {
					for (uint8_t kx = 0; kx < KERN_SIZE; kx++) {
						for (uint8_t c = 0; c < KERN_CHAN; c++) {
							// Where we are on the board
							const uint8_t loc = kx+bx+(ky+by)*5;
							// Look up what's on the board at this location, and
							// multiply it. For c=0 we have to do some extra work
							float lookup = 0;
							if (COUNT_AT(loc)>c) {
								if (c==0) {
									if (STONE_AT(loc) == STONE_STANDING) {
										lookup = (colours[loc] & 1) ? +0.25 : -0.25;
									} else if (STONE_AT(loc) == STONE_CAPSTONE) {
										lookup = (colours[loc] & 1) ? +1.00 : -1.00;
									} else {
										lookup = (colours[loc] & 1) ? +0.50 : -0.50;
									}
								} else {
									lookup = (colours[loc] & (1<<c)) ? +0.50 : -0.50;
								}
							}
							flattened[kern+KERN_NUM*(bx+by*KERN_OSIZE)]
								+= lookup*conv2d_weights[kern][ky][kx][c];
						}
					}
				}
				RELU(flattened[kern+KERN_NUM*(bx+by*KERN_OSIZE)]);
			}
		}
	}
	// Add input of flat counts
	flattened[CONV_NUM]   = (float)(white_count & 127)/21.0;
	flattened[CONV_NUM+1] = (float)(black_count & 127)/21.0;
	/* ------------------ *
	 *  First dense layer *
	 * ------------------ */
	for (uint8_t d1 = 0; d1 < DENSE1_NUM; d1++) {
		dense1[d1] = dense1_biases[d1];
		for (uint8_t fl = 0; fl < CONV_NUM+2; fl++) {
			dense1[d1] += flattened[fl]*dense1_weights[d1][fl];
		}
		RELU(dense1[d1]);
	}
	/* ------------------- *
	 *  Second dense layer *
	 * ------------------- */
	for (uint8_t d2 = 0; d2 < DENSE2_NUM; d2++) {
		dense2[d2] = dense2_biases[d2];
		for (uint8_t d1 = 0; d1 < DENSE1_NUM; d1++) {
			dense2[d2] += dense1[d1]*dense2_weights[d2][d1];
		}
		RELU(dense2[d2]);
	}
	/* ------------- *
	 *  Output layer *
	 * ------------- */
	float output = output_bias;
	for (uint8_t d2 = 0; d2 < DENSE2_NUM; d2++) {
		output += dense2[d2]*output_weights[d2];
	}
	// Truncated Pade approximant of logistic function
	output = (12.0+output+50.0*output/(output*output+10.0))/24.0;
#ifndef DETERMINISTIC
	DOXORSHIFT;
	output += fudge.f;
#endif
	if (output > 1.0) {
		if (ply & 1) return 1.0;
		else return -1.0;
	}
	else if (output < 0.0) {
		if (ply & 1) return -1.0;
		else return 1.0;
	}

	return (ply & 1) ? 2.0*output-1.0 : 1.0-2*output;
}

// ===================================================================
// α-β minimax using the above evaluator
// ===================================================================

static void
previous_ply(void) {
	if (ply>0) ply--;
	if (ply == 1) {
		current_colour = C_WHITE;
	} else {
		if (current_colour == C_BLACK) current_colour = C_WHITE;
		else current_colour = C_BLACK;
	}
}

static float val;
static enum WIN_TYPE w;

#define WIN_EVALUATE_OR_RECURSE(store,reset) {		\
		w = 0xFF; \
		if (ply >= 2*5 - 2) w = check_win(); \
		if (w < 0xFF) { \
			/* Somebody won, assign weights accordingly. Note in particular
				 that draws are only worth ∞/2 ;) \
			*/ \
			if (ply & 1) { \
				if (w == WIN_ROAD_WHITE || w == WIN_FLAT_WHITE) \
					val = -infty; \
				else if (w == WIN_DRAW) val = infty/2.0; \
				else val = infty; \
			} else { \
				if (w == WIN_ROAD_BLACK || w == WIN_FLAT_BLACK) \
					val = -infty; \
				else if (w == WIN_DRAW) val = infty/2.0; \
				else val = infty; \
			} \
		} else if (cur_depth == max_depth) { \
			/* We're at the bottom, evaluate */ \
			val = ct1986_evaluate_win(); \
		} else { \
			/* We're not at the bottom, recurse first */ \
			next_ply(); \
			val = -ct1986_negamax(cur_depth + 1, max_depth, -beta, -alpha); \
			previous_ply(); \
		} \
		{ reset }; \
		/* Prune */ \
		if (val >= beta) return val; \
		/* Update the optimal value, which alpha carries */ \
		if (val > optimal) { \
			optimal = val; \
			if (val > alpha)  alpha = val; \
			if (cur_depth == 0) { store }; \
		} \
	}

// UP DOWN LEFT RIGHT
static const int8_t deltas[4] = { +5, -5, -1, +1};

float
ct1986_negamax(const uint8_t cur_depth, const uint8_t max_depth,
							 float alpha, float beta) {
	const uint8_t black = (ply & 1),
		material = (black) ? black_count : white_count,
		flat = material & 127,
		cap = (ply > 2 && (material & 128)),
		standing = (ply > 2 && (material & 127));

	float optimal = -infty;

	// Step across the board
	for (uint8_t row = 0; row < 5; row++) {
		for (uint8_t col = 0; col < 5; col++) {
			// Try all valid actions for this square. Is it empty?
			const uint8_t loc = THE_COORDS(col, row);
			const uint8_t count = (COUNT_AT(loc) > 5) ? 5 : COUNT_AT(loc);
			// Only try moves after CPS
			if (count && ((colours[loc] & 1) == current_colour) && ply>2) {
				// There are stones, can we move them in a given direction?

				// Pre-compute end-stops
				uint8_t end_stops[4][2]; // (end, not_crush)
				// UP DOWN LEFT RIGHT
				end_stops[0][0] = (4-row > count) ? count : 4-row;
				end_stops[1][0] = (row > count) ? count : row;
				end_stops[2][0] = (col > count) ? count : col;
				end_stops[3][0] = (4-col > count) ? count : 4-col;
				const uint8_t cap_top = STONE_AT(loc) == STONE_CAPSTONE;
				for (uint8_t d = 0; d < 4; d++){
					end_stops[d][1] = 1;
					const uint8_t stop = end_stops[d][0];
					end_stops[d][0] = 0;
					for (uint8_t k = 1; k <= stop; k++) {
						const uint8_t stone = STONE_AT(loc+k*deltas[d]);
						if (stone == STONE_STANDING) {
							if (cap_top) {
								end_stops[d][1] = 0;
								end_stops[d][0]++;
							}
							break;
						} else if (stone == STONE_CAPSTONE) {
							break;
						}
						end_stops[d][0]++;
					}
				}

				uint16_t colours_backup[5];
				uint8_t celldat_backup[5], drops[5]; // we only use 4, the
				// fifth is to skip a
				// bounds check at (*)
				// Back up the row of the board
				for (uint8_t y = 0; y < 5; y++) {
					colours_backup[y] = colours[THE_COORDS(col, y)];
					celldat_backup[y] = celldat[THE_COORDS(col, y)];
				}

				// I'm not a huge fan of looping through enums, but it's
				// better than manually unrolling this. Sufficiently smart
				// compilers?
				for (enum MOVE_DIRECTION dir = M_UP; dir <= M_RIGHT; dir++) {
					// Back-up the column once we start looking horizontally
					if (dir == M_LEFT) {
						for (uint8_t x = 0; x < 5; x++) {
							colours_backup[x] = colours[THE_COORDS(x, row)];
							celldat_backup[x] = celldat[THE_COORDS(x, row)];
						}
					}
					/*
					 * We don't do anything terribly efficient here just try
					 * all the ordered partitions of num ∈ {1 … end_stop}, and
					 * skip the partition if it calls for multiple stones at
					 * the end with a crush.
					 */
					uint8_t gaps, t, idx, mask;
					for (uint8_t num = 1; num <= count; num++) {
						for (uint8_t steps = 1;
								 steps <= end_stops[dir][0] && steps <= num;
								 steps++) {
							gaps = 0b00000111 >> (4-steps);
							do  {
								// Translate to a drop sequence
								drops[0] = 1; mask = 1; idx = 0;
								for (uint8_t d = 0; d + 1 < num; d++) {
									if (gaps & mask) {
										idx++;
										drops[idx] = 1; // (*) we don't need to bounds check
									} else {
										drops[idx] += 1;
									}
									mask <<= 1;
								}
								// TODO: Work out what this should be before partition

								// Ensure legal move if we have to crush
								if (end_stops[dir][1] || drops[steps-1] <= 1) {
									// Try it, and manually check for win if it's valid
									uint8_t j = num;
									for (uint8_t k = 0; k < steps; k++) {
										// Dear future me, i'm sorry
										j -= drops[k];
										colours[loc+(k+1)*deltas[dir]]  = (colours[loc+(k+1)*deltas[dir]] << drops[k])
											| ((colours[loc] >> j) & (0xFFFF >> (0x10 - drops[k])));
										celldat[loc+(k+1)*deltas[dir]]  = (k == steps - 1) ? STONE_AT(loc) : STONE_FLAT
											| ((celldat[loc+(k+1)*deltas[dir]] + ((drops[k] << NUM_SHIFT))) & NUM_MASK);
									}
									// Then we drop them from the source
									colours[loc] >>= num;
									const uint8_t dec_count = celldat[loc] - (num << NUM_SHIFT);
									celldat[loc] = dec_count & NUM_MASK;
									// First check for wins, if we're at the bottom
									// evaluate, otherwise recurse
									WIN_EVALUATE_OR_RECURSE({
											// If we did update the optimal value, store
											// this move
											generate_move(loc, dir, steps, drops, ct1986_ptn);
										},{
											// Reset the board data after recursing or
											// before returning
											if (dir <= M_DOWN) {
												for (uint8_t y = 0; y < 5; y++) {
													colours[THE_COORDS(col, y)] = colours_backup[y];
													celldat[THE_COORDS(col, y)] = celldat_backup[y];
												}
											} else {
												for (uint8_t x = 0; x < 5; x++) {
													colours[THE_COORDS(x, row)] = colours_backup[x];
													celldat[THE_COORDS(x, row)] = celldat_backup[x];
												}
											}
										});
								}
								/*
								 * With thanks to
								 * https://graphics.stanford.edu/~seander/bithacks.html#NextBitPermutation
								 * we have the following magic to generate the next
								 * permutation of steps-many set bits
								 */
								t = (gaps | (gaps - 1));
								gaps = (t + 1) | (((~t & -~t) - 1) >> (__builtin_ctz(gaps) + 1));
							} while (gaps && (gaps + 1 <= (1<<(num-1))));
						}
					}
				}
			} else if (material && count == 0) {
				// Empty square, try placements

				if (flat) {
					// Generate the placement
					if (black) black_count--;
					else white_count--;
					colours[loc] = current_colour;
					celldat[loc] = NUM_INC | STONE_FLAT;
					WIN_EVALUATE_OR_RECURSE({
							// If we did update the optimal value, store
							generate_place(loc, STONE_FLAT, ct1986_ptn);
						},{
							// Reset the state
							celldat[loc] = 0;
							if (black) black_count++;
							else white_count++;
						});

					// Do the same for walls, can't happen without flats
					if (standing) {
						if (black) black_count--;
						else white_count--;
						colours[loc] = current_colour;
						celldat[loc] = NUM_INC | STONE_STANDING;
						WIN_EVALUATE_OR_RECURSE({
								generate_place(loc, STONE_STANDING, ct1986_ptn);
							},{
								celldat[loc] = 0;
								if (black) black_count++;
								else white_count++;
							});
					}
				}

				// and for caps
				if (cap) {
					if (black) black_count &= 127;
					else white_count &= 127;
					colours[loc] = current_colour;
					celldat[loc] = NUM_INC | STONE_CAPSTONE;
					WIN_EVALUATE_OR_RECURSE({
							generate_place(loc, STONE_CAPSTONE, ct1986_ptn);
						},{
							celldat[loc] = 0;
							if (black) black_count |= 128;
							else white_count |= 128;
						});
				}
			}
			ct1986_display_progress(cur_depth);
		}
	}
	return alpha;
}

inline float
ct1986_generate(const uint8_t max_depth) {
	return ct1986_negamax(0, max_depth, -infty, infty);
}