Occupancy Calculator

Problem

Compute blocks per SM and occupancy from register, shared memory and thread limits. reg_blocks = regs_per_sm // (threads_per_block * registers_per_thread); smem_blocks = smem_per_sm // smem_per_block (or blocks_limit when smem_per_block is 0); thread_blocks = threads_per_sm // threads_per_block; blocks = min of all and blocks_limit; occupancy = blocks * threads_per_block / threads_per_sm.

The signature is occupancy_calculator(threads_per_block, registers_per_thread, regs_per_sm, threads_per_sm, smem_per_block, smem_per_sm, blocks_limit=32). Return [blocks, occupancy].

Hint

Higher occupancy hides memory latency but is not always faster than a register-rich kernel.

Test cases

case 1

in: [256,32,65536,2048,0,100000,32]

expected: [8,1]

case 2

in: [128,64,65536,2048,16384,65536,32]

expected: [4,0.25]

case 3

in: [1024,32,65536,2048,0,100000,32]

expected: [2,1]

case 4

in: [64,128,65536,2048,0,100000,32]

expected: [8,0.25]

Starter codePython

def occupancy_calculator(threads_per_block, registers_per_thread, regs_per_sm, threads_per_sm, smem_per_block, smem_per_sm, blocks_limit=32):
    # Your code here
    pass

4 test cases · run your code against them in the editor