Python 2 Player Tetris Library implemented in C++

Following tetris guidelines, implemented in C++ for speed, actions are taken simultaneously, players turn-based with matching keys instead piece, can be used to train AI/whatever. with instant softdrop after specified time and short softdrop like DAS (not implemented in python lib yet)

Two versions to choose from, tetris_c requires SDL2 but have render available, tetris_c_nore has no render therefore no SDL2 install necessary(or if you're installing from wheel it might not be an issue)

Usage Brief view

import tetris

x=tetris.Container() # game object
y=x.copy() # deep copy if needed
x.reset() # reset state
x.seed_reset(n) # reset with rng seed
state=x.get_state() # shape (500,) numpy array, 1D so easier to use mp shared memory, more info below
shapes=x.get_shapes() # (7,4,4,4) 7 pieces 4 rotations and 4x4 shape 

x.step(ac1,ac2)

In Depth Info

Game Step

action is scalar int

0:hold
1:hard drop
2:cw
3:ccw
4:left tap
5:right tap
6:left das
7:right das
8:soft drop
9:180

States

p1_state=state[0:232]
p2_state=state[232:464]

sum attack state[464], negative is to p1 individual attacks state[465:]

Reward

state[0] is reward

127 is lose, 126 is win, otherwise

if harddrop (action==1)
    if lines_cleared
        rew=10 * (attack+ b2b *2)+2*lines
    rew+=b2b_not_broke
else rew=wallkick

Python for extracting

lines_cleared = (reward%10)//2
atk = reward//10 # b2b rewards 2 atk but is 1 in-game
wallkick=reward%5 if action!=1 else 0
b2b_not_broken=reward%5 if action==1 else 0

game doesn't auto reset on game over, check those signals and call reset on your own

States

i piece goes to -2 so the value is +2 by default, if you want to draw current piece subtract that back

x = state[1]

Board is 30 high, bottom 21 is returned, y spawns at 9 normally but 8 if occupied, otherwise game over.
Value returned to python starts at 8 so pieces spawn at 1 normally, could be negative if you navigate the piece off screen

y = state[2]

Distance to bottom

softdropdist = state[3]

Default is 10 actions will force harddrop

action_count = state[4]

Garbage cleared from last action, use for reward etc

garbage = state[5]

hold_used = state[6]

rotation = state[7]

active = state[8]

Draw active piece on current location(you have to check x negative yourself):

board[x-2,y]=shape[active][rotation]

Held can be -1 meaning not used

held_piece = state[9]

next_pieces = state[10:15]

Cheat data(hidden queue)

hidden_queue=state[15:22]

Less cheating because you can manually count the bag:

hidden_queue=sort(state[15:22])

Board state[22:232].reshape(21,10)

Feedback

Any ideas for improvements are welcome.

State is mostly boolean/binary so it's possible I can optimise the code further to increase speed as use cases probably will be AI training that needs millions of iterations.

if you want to read/change the code, tetrisboard is main logic including clear rewards, and env.cpp is the python wrapper+rendering+returning state