diff --git a/examples/python_naive_matmul/2_tiled_matmul.py b/examples/python_naive_matmul/2_tiled_matmul.py index be403a46a..839b07989 100644 --- a/examples/python_naive_matmul/2_tiled_matmul.py +++ b/examples/python_naive_matmul/2_tiled_matmul.py @@ -13,7 +13,7 @@ class MatMulOp: max_workgroup_size = props['max_work_group_size'] if tile_size < 0: tile_size = 1 - while (2 * tile_size * tile_size <= max_workgroup_invocation + while (4 * tile_size * tile_size <= max_workgroup_invocation and 2 * tile_size <= max_workgroup_size[0] and 2 * tile_size <= max_workgroup_size[1]): tile_size *= 2 diff --git a/examples/python_naive_matmul/3_better_tiling.py b/examples/python_naive_matmul/3_better_tiling.py new file mode 100644 index 000000000..5dcc52349 --- /dev/null +++ b/examples/python_naive_matmul/3_better_tiling.py @@ -0,0 +1,141 @@ +import time + +import kp +import numpy as np + + +class MatMulOp: + def __init__(self, manager: kp.Manager, tile_size: int = -1, thread_work_ratio: int = 8): + self.mgr = manager + + props = self.mgr.get_device_properties() + max_workgroup_invocation = props['max_work_group_invocations'] + max_workgroup_size = props['max_work_group_size'] + if tile_size < 0: + tile_size = 1 + local_size_y = tile_size // thread_work_ratio + while (4 * tile_size * local_size_y <= max_workgroup_invocation + and 2 * tile_size <= max_workgroup_size[0] + and 2 * local_size_y <= max_workgroup_size[1]): + tile_size *= 2 + local_size_y = tile_size // thread_work_ratio + else: + local_size_y = tile_size // thread_work_ratio + + assert tile_size > 0 + assert thread_work_ratio > 0 + assert tile_size * local_size_y <= max_workgroup_invocation + assert tile_size <= max_workgroup_size[0] + assert local_size_y <= max_workgroup_size[1] + self.tile_size = tile_size + self.thread_work_ratio = thread_work_ratio + + local_size_y = tile_size // thread_work_ratio + self.shader = kp.Shader.compile_source(f''' +#version 450 + +layout (local_size_x = {tile_size}, local_size_y = {local_size_y}) in; + +layout (set = 0, binding = 0) readonly buffer buf_in_tensor_1 {{ float in_tensor_1[]; }}; +layout (set = 0, binding = 1) readonly buffer buf_in_tensor_2 {{ float in_tensor_2[]; }}; +layout (set = 0, binding = 2) writeonly buffer buf_out_tensor {{ float out_tensor[]; }}; + +layout (constant_id = 0) const float tensor_size_f = 0; + +shared float sub_tensor_1[{tile_size}][{tile_size}]; +shared float sub_tensor_2[{tile_size}][{tile_size}]; + +void main() +{{ + uint row = gl_GlobalInvocationID.x; + uint col = gl_GlobalInvocationID.y; + uint globalRow = {tile_size} * gl_WorkGroupID.x + row; + uint globalCol = {tile_size} * gl_WorkGroupID.y + row; + + uint tensor_size = uint(tensor_size_f); + float acc[{thread_work_ratio}]; + for (uint l = 0u; l < {thread_work_ratio}; l++) + acc[l] = 0.0; + + uint numTiles = tensor_size / {tile_size}; + for(uint t = 0u; t < numTiles; t++) + {{ + uint tiledRow = {tile_size} * t + row; + uint tiledCol = {tile_size} * t + col; + sub_tensor_1[col + t * {local_size_y}][row] = in_tensor_1[ + (tiledCol + t * {local_size_y}) * tensor_size + globalRow]; + sub_tensor_2[col + t * {local_size_y}][row] = in_tensor_2[ + (globalCol + t * {local_size_y})* tensor_size + tiledRow]; + + memoryBarrierShared(); + barrier(); + + for(uint k = 0u; k < {tile_size}; k++) + for(uint l = 0u; l < {thread_work_ratio}; l++) + acc[l] += sub_tensor_1[k][row] * sub_tensor_2[col + l * {local_size_y}][k]; + + barrier(); + }} + for(uint l = 0u; l < {thread_work_ratio}; l++) + out_tensor[(globalCol + l * {local_size_y}) * tensor_size + globalRow] = acc[l]; +}}''') + self.tensor_shape: tuple[int, int] = (0, 0) + self.params: list[kp.Tensor] = [] + self.algo = None + + def __call__(self, tensor_shape: tuple[int, int], tensor_in_1: kp.Tensor, tensor_in_2: kp.Tensor, + tensor_out: kp.Tensor): + params = [tensor_in_1, tensor_in_2, tensor_out] + + if self.algo is None or self.tensor_shape != tensor_shape or self.params != params: + self.tensor_shape = tensor_shape + self.params = params + self.algo = self.mgr.algorithm( + params, # params + self.shader, # spirv + (tensor_shape[0] // self.tile_size, tensor_shape[1] // self.tile_size, 1), # workgroup + [float(tensor_shape[0])], # spec_consts + []) # push_consts + + (self.mgr.sequence() + .record(kp.OpTensorSyncDevice(self.params)) + .record(kp.OpAlgoDispatch(self.algo)) + .record(kp.OpTensorSyncLocal(self.params)) + .eval()) + + +def main(): + mgr = kp.Manager() + + matmul_op = MatMulOp(mgr) + + tensor_size = 512 + tensor_shape = [tensor_size, tensor_size] + tensor_in_1 = mgr.tensor(np.triu(np.ones(tensor_shape))) + tensor_in_2 = mgr.tensor(np.triu(np.ones(tensor_shape))) + tensor_out = mgr.tensor(np.zeros(tensor_shape)) + + print(f'{tensor_shape} input tensors:\n' + f'{tensor_in_1.data().reshape(tensor_shape)}\n' + f'{tensor_in_2.data().reshape(tensor_shape)}\n') + + matmul_op(tensor_shape, tensor_in_1, tensor_in_2, tensor_out) + + experiment_count = 1000 + start_time = time.time() + for _ in range(experiment_count): + matmul_op(tensor_shape, tensor_in_1, tensor_in_2, tensor_out) + end_time = time.time() + experiment_time = end_time - start_time + op_count = tensor_shape[0] * tensor_shape[1] * (tensor_shape[1] - 1) + + print(f'Output :\n{tensor_out.data().reshape(tensor_shape)}') + + print(f'{experiment_count} matmul time : ' + f'{experiment_time * 1000:0.2f}ms => ' + f'{experiment_count / experiment_time:0.2f}op/s or ' + f'{experiment_count * op_count / (1e9 * experiment_time):0.2f}GFLOPS') + + +if __name__ == '__main__': + main()