from tinygrad import Tensor, Context, GlobalCounters, dtypes from tinygrad.uop.ops import UOp, KernelInfo, sint, AxisType from tinygrad.dtype import AddrSpace from tinygrad.helpers import DEBUG, getenv N = getenv("N", 4096) M = getenv("M", N) K = getenv("K", N) NUM_RUNS = getenv("CNT", 5) # --------------------------- # launch/config constants # --------------------------- WARP_SIZE = 32 BLOCK_M, BLOCK_N, BLOCK_K = 128, 128, 8 TM, TN = 4, 4 LANES_PER_WAVE_M, LANES_PER_WAVE_N = 4, 8 assert N % BLOCK_N == 0 and M % BLOCK_M == 0 and K % BLOCK_K == 0 is_kernel5 = getenv("K5", 0) THREADS_PER_BLOCK = 128 if is_kernel5 else 256 WAVES_PER_BLOCK_N = 1 if is_kernel5 else 2 WAVES_PER_BLOCK_M = THREADS_PER_BLOCK // WARP_SIZE // WAVES_PER_BLOCK_N REG_TILES_PER_WAVE_N = BLOCK_N // (WAVES_PER_BLOCK_N * LANES_PER_WAVE_N * TN) REG_TILES_PER_WAVE_M = BLOCK_M // (WAVES_PER_BLOCK_M * LANES_PER_WAVE_M * TM) assert WAVES_PER_BLOCK_M*REG_TILES_PER_WAVE_M*LANES_PER_WAVE_M*TM == BLOCK_M, "M reshape is wrong" assert WAVES_PER_BLOCK_N*REG_TILES_PER_WAVE_N*LANES_PER_WAVE_N*TN == BLOCK_N, "N reshape is wrong" def rngs_for_shape(shape:tuple[sint, ...], rng:int, axis_type=AxisType.WEAK): return [UOp.range(s, rng+i, axis_type) for i,s in enumerate(shape)] def copy(dest:UOp, src:UOp, rng:int, upcast=False): assert dest.shape == src.shape rngs = rngs_for_shape(src.shape, rng, AxisType.UPCAST if upcast else AxisType.WEAK) return dest[*rngs].store(src[*rngs]).end(*rngs) def hand_spec_kernel3(c:UOp, a:UOp, b:UOp) -> UOp: # --------------------------- # block indices # --------------------------- block_id_n = UOp.special(N // BLOCK_N, "gidx0") block_id_m = UOp.special(M // BLOCK_M, "gidx1") # index the output with the globals c = c.reshape(M // BLOCK_M, BLOCK_M, N // BLOCK_N, BLOCK_N)[block_id_m, :, block_id_n, :] # open the main reduction range k_tile_range = UOp.range(K // BLOCK_K, 0, AxisType.REDUCE) a = a.reshape(M // BLOCK_M, BLOCK_M, K // BLOCK_K, BLOCK_K)[block_id_m, :, k_tile_range, :] b = b.reshape(K // BLOCK_K, BLOCK_K, N // BLOCK_N, BLOCK_N)[k_tile_range, :, block_id_n, :] # globals are no longer used, they are already in the indexes del block_id_m, block_id_n # --------------------------- # GLOBAL -> LOCAL (A_local, B_local) # --------------------------- tid = UOp.special(THREADS_PER_BLOCK, "lidx0") # A: read BM x BK tiles (permute on store into locals) BM_A_local_stride = (BLOCK_M + 4) if is_kernel5 else BLOCK_M A_local = UOp.placeholder((BLOCK_K, BM_A_local_stride), dtypes.float, slot=0, addrspace=AddrSpace.LOCAL).shrink_to((BLOCK_K, BLOCK_M)) A_local_store = copy(A_local.permute((1,0)).reshape(-1, THREADS_PER_BLOCK)[:, tid], a.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=100) # B: read BK x BN tiles B_local = UOp.placeholder((BLOCK_K, BLOCK_N), dtypes.float, slot=1, addrspace=AddrSpace.LOCAL) B_local_store = copy(B_local.reshape(-1, THREADS_PER_BLOCK)[:, tid], b.reshape(-1, THREADS_PER_BLOCK)[:, tid], rng=200) # NOTE: no explicit barrier needed, the AFTER on the LOCAL buffers implies it in late codegen A_local, B_local = A_local.after(A_local_store, B_local_store), B_local.after(A_local_store, B_local_store) # open inner k range k = UOp.range(BLOCK_K, 3, AxisType.REDUCE) # --------------------------- # LOCAL -> REG (per-wave tiles) # --------------------------- warp, lane = tid // WARP_SIZE, tid % WARP_SIZE waveIdx, waveIdy = warp % WAVES_PER_BLOCK_N, warp // WAVES_PER_BLOCK_N laneIdx, laneIdy = lane % LANES_PER_WAVE_N, lane // LANES_PER_WAVE_N assert waveIdy.vmax+1 == WAVES_PER_BLOCK_M and laneIdy.vmax+1 == LANES_PER_WAVE_M A_col = UOp.placeholder((REG_TILES_PER_WAVE_M, TM), dtypes.float, slot=0, addrspace=AddrSpace.REG) A_local_slice = A_local[k, :].reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM)[waveIdy, :, laneIdy, :] A_col = A_col.after(copy(A_col, A_local_slice, 300, upcast=True)) B_row = UOp.placeholder((REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=1, addrspace=AddrSpace.REG) B_local_slice = B_local[k, :].reshape(WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN)[waveIdx, :, laneIdx, :] B_row = B_row.after(copy(B_row, B_local_slice, 400, upcast=True)) # --------------------------- # FMA: c_regs += A_col * B_row # --------------------------- c_regs = UOp.placeholder((REG_TILES_PER_WAVE_M, TM, REG_TILES_PER_WAVE_N, TN), dtypes.float, slot=2, addrspace=AddrSpace.REG) i = UOp.range(c_regs.size, 16) c_regs = c_regs.after(c_regs.flatten()[i].store(0.0).end(i)) # TODO: why don't these work as upcast? # why if the ranges merge is it slow?!? (if you change the order on end, they will merge. big slowdown on METAL) iter_m, t_m, iter_n, t_n = rngs = rngs_for_shape(c_regs.shape, 500) sink = c_regs[*rngs].store(c_regs.after(k)[*rngs] + A_col[iter_m, t_m] * B_row[iter_n, t_n]).end(iter_m, iter_n, t_m, t_n) # Close k, sync, and close K tiles sink = sink.end(k).end(k_tile_range) # --------------------------- # REG -> GLOBAL (epilogue) # --------------------------- c = c.reshape(WAVES_PER_BLOCK_M, REG_TILES_PER_WAVE_M, LANES_PER_WAVE_M, TM, WAVES_PER_BLOCK_N, REG_TILES_PER_WAVE_N, LANES_PER_WAVE_N, TN) c = c[waveIdy, :, laneIdy, :, waveIdx, :, laneIdx, :] sink = copy(c, c_regs.after(sink), rng=600) return sink.sink(arg=KernelInfo(opts_to_apply=())).simplify() def eval_custom_matmul(fxn, dt=dtypes.float): a = Tensor.randn(M, K, dtype=dt) b = Tensor.randn(K, N, dtype=dt) c = Tensor.empty(M, N, dtype=dtypes.float) with Context(DEBUG=0): Tensor.realize(a, b) ets = [] with Context(DEBUG=max(2, DEBUG.value)): for _ in range(NUM_RUNS): GlobalCounters.reset() tst = Tensor.custom_kernel(c, a, b, fxn=fxn)[0].realize() ets.append(GlobalCounters.time_sum_s) print(f"REAL TFLOPS {M * N * K * 2 / min(ets) * 1e-12:.2f}") if getenv("VERIFY", 1): GlobalCounters.reset() with Context(DEBUG=2): tc = (a.float() @ b.float()).realize() with Context(DEBUG=0): err = (tc - tst).square().mean().item() print(f"mean squared error {err}") if err > (1e-2 if dt == dtypes.half else 1e-6): raise RuntimeError("matmul is wrong!") if __name__ == "__main__": eval_custom_matmul(hand_spec_kernel3)