diff --git a/examples/python_naive_matmul/benchmark.py b/examples/python_naive_matmul/benchmark.py new file mode 100644 index 000000000..b10369d7c --- /dev/null +++ b/examples/python_naive_matmul/benchmark.py @@ -0,0 +1,49 @@ +import time + +import kp +import numpy as np +from imp1_naive import MatMulOp as MatMulOp1 +from imp2_tiled import MatMulOp as MatMulOp2 +from imp3_better_tiling import MatMulOp as MatMulOp3 + + +def main(): + experiment_count = 1000 + tensor_size = 512 + tensor_shape = [tensor_size, tensor_size] + mat_1 = np.triu(np.ones(tensor_shape)) + mat_2 = np.triu(np.ones(tensor_shape)) + mat_result = mat_1 @ mat_2 + tensor_shape = [tensor_size, tensor_size] + + print(f'{tensor_shape} input tensors:\n' + f'{mat_1}\n' + f'{mat_2}\n') + print(f'Output :\n{mat_result}') + + mgr = kp.Manager() + tensor_in_1 = mgr.tensor(mat_1) + tensor_in_2 = mgr.tensor(mat_2) + tensor_out = mgr.tensor(np.zeros(tensor_shape)) + for MatMulOp in [MatMulOp1, MatMulOp2, MatMulOp3]: + matmul_op = MatMulOp(mgr) + matmul_op(tensor_shape, tensor_in_1, tensor_in_2, tensor_out) + + 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) + + if (tensor_out.data().reshape(tensor_shape) == mat_result).all(): + print(f'From {MatMulOp.__module__} : {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') + else: + print(f'Test failed => output tensor is wrong :\n{tensor_out.data().reshape(tensor_shape)}') + + +if __name__ == '__main__': + main() diff --git a/examples/python_naive_matmul/1_naive_matmul.py b/examples/python_naive_matmul/imp1_naive.py similarity index 92% rename from examples/python_naive_matmul/1_naive_matmul.py rename to examples/python_naive_matmul/imp1_naive.py index d5fe70cba..faefec563 100644 --- a/examples/python_naive_matmul/1_naive_matmul.py +++ b/examples/python_naive_matmul/imp1_naive.py @@ -41,7 +41,7 @@ class MatMulOp: self.local_size_x = local_size_x self.local_size_y = local_size_y - self.shader = kp.Shader.compile_source(f''' + self.shader = f''' #version 450 layout (local_size_x = {local_size_x}, local_size_y = {local_size_y}) in; @@ -62,7 +62,8 @@ void main() for(uint k = 0u; k < tensor_size; k++) acc += in_tensor_1[(k * tensor_size) + globalRow] * in_tensor_2[(globalCol * tensor_size) + k]; out_tensor[(globalCol * tensor_size) + globalRow] = acc; -}}''') +}}''' + self.compiled_shader = kp.Shader.compile_source(self.shader) self.tensor_shape: tuple[int, int] = (0, 0) self.params: list[kp.Tensor] = [] self.algo = None @@ -74,17 +75,18 @@ void main() if self.algo is None or self.tensor_shape != tensor_shape or self.params != params: self.tensor_shape = tensor_shape self.params = params + workgroup = (tensor_shape[0] // self.local_size_x, tensor_shape[1] // self.local_size_y, 1) self.algo = self.mgr.algorithm( params, # params - self.shader, # spirv - (tensor_shape[0] // self.local_size_x, tensor_shape[1] // self.local_size_y, 1), # workgroup + self.compiled_shader, # spirv + workgroup, # 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)) + .record(kp.OpTensorSyncLocal([tensor_out])) .eval()) @@ -121,11 +123,5 @@ def main(): f'{experiment_count * op_count / (1e9 * experiment_time):0.2f}GFLOPS') -def test(): - main() - - if __name__ == '__main__': main() -else: - test() diff --git a/examples/python_naive_matmul/2_tiled_matmul.py b/examples/python_naive_matmul/imp2_tiled.py similarity index 79% rename from examples/python_naive_matmul/2_tiled_matmul.py rename to examples/python_naive_matmul/imp2_tiled.py index 839b07989..ed6a1ddf8 100644 --- a/examples/python_naive_matmul/2_tiled_matmul.py +++ b/examples/python_naive_matmul/imp2_tiled.py @@ -24,7 +24,7 @@ class MatMulOp: assert tile_size <= max_workgroup_size[1] self.tile_size = tile_size - self.shader = kp.Shader.compile_source(f''' + self.shader = f''' #version 450 layout (local_size_x = {tile_size}, local_size_y = {tile_size}) in; @@ -40,20 +40,21 @@ 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 row = gl_LocalInvocationID.x; // 0 .. tile_size + uint col = gl_LocalInvocationID.y; // 0 .. tile_size + // gl_WorkGroupID : 0 .. tensor_size / tile_size + uint globalRow = ({tile_size} * gl_WorkGroupID.x) + row; + uint globalCol = ({tile_size} * gl_WorkGroupID.y) + col; uint tensor_size = uint(tensor_size_f); float acc = 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][row] = in_tensor_1[tiledCol * tensor_size + globalRow]; - sub_tensor_2[col][row] = in_tensor_2[globalCol * tensor_size + tiledRow]; + uint tiledRow = ({tile_size} * t) + row; + uint tiledCol = ({tile_size} * t) + col; + sub_tensor_1[col][row] = in_tensor_1[(tiledCol * tensor_size) + globalRow]; + sub_tensor_2[col][row] = in_tensor_2[(globalCol * tensor_size) + tiledRow]; memoryBarrierShared(); barrier(); @@ -63,8 +64,10 @@ void main() barrier(); }} - out_tensor[(globalCol * tensor_size) + globalRow] = acc; -}}''') + uint globalIndex = (tensor_size * globalCol) + globalRow; + out_tensor[globalIndex] = acc; +}}''' + self.compiled_shader = kp.Shader.compile_source(self.shader) self.tensor_shape: tuple[int, int] = (0, 0) self.params: list[kp.Tensor] = [] self.algo = None @@ -76,17 +79,18 @@ void main() if self.algo is None or self.tensor_shape != tensor_shape or self.params != params: self.tensor_shape = tensor_shape self.params = params + workgroup = (tensor_shape[0] // self.tile_size, tensor_shape[1] // self.tile_size, 1) self.algo = self.mgr.algorithm( params, # params - self.shader, # spirv - (tensor_shape[0] // self.tile_size, tensor_shape[1] // self.tile_size, 1), # workgroup + self.compiled_shader, # spirv + workgroup, # 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)) + .record(kp.OpTensorSyncLocal([tensor_out])) .eval()) diff --git a/examples/python_naive_matmul/imp2_tiled_debug.py b/examples/python_naive_matmul/imp2_tiled_debug.py new file mode 100644 index 000000000..b10cbfd7d --- /dev/null +++ b/examples/python_naive_matmul/imp2_tiled_debug.py @@ -0,0 +1,156 @@ +import time + +import kp +import numpy as np + + +class MatMulOp: + def __init__(self, manager: kp.Manager, tile_size: int = -1): + 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 + 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 + + assert tile_size > 0 + assert tile_size * tile_size <= max_workgroup_invocation + assert tile_size <= max_workgroup_size[0] + assert tile_size <= max_workgroup_size[1] + self.tile_size = tile_size + + print(f'{tile_size=}') + self.shader = f''' +#version 450 + +layout (local_size_x = {tile_size}, local_size_y = {tile_size}) 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 (set = 0, binding = 3) writeonly buffer buf_test1_tensor {{ float test1_tensor[]; }}; +layout (set = 0, binding = 4) writeonly buffer buf_test2_tensor {{ float test2_tensor[]; }}; +layout (set = 0, binding = 5) writeonly buffer buf_test3_tensor {{ float test3_tensor[]; }}; +layout (set = 0, binding = 6) writeonly buffer buf_test4_tensor {{ float test4_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_LocalInvocationID.x; // 0 .. tile_size + uint col = gl_LocalInvocationID.y; // 0 .. tile_size + // gl_WorkGroupID : 0 .. tensor_size / tile_size + uint globalRow = ({tile_size} * gl_WorkGroupID.x) + row; + uint globalCol = ({tile_size} * gl_WorkGroupID.y) + col; + + uint tensor_size = uint(tensor_size_f); + float acc = 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][row] = in_tensor_1[(tiledCol * tensor_size) + globalRow]; + sub_tensor_2[col][row] = in_tensor_2[(globalCol * tensor_size) + tiledRow]; + + memoryBarrierShared(); + barrier(); + + for(uint k = 0u; k < {tile_size}; k++) + acc += sub_tensor_1[k][row] * sub_tensor_2[col][k]; + + barrier(); + }} + uint globalIndex = (tensor_size * globalCol) + globalRow; + out_tensor[globalIndex] = acc; + test1_tensor[globalIndex] = row; + test2_tensor[globalIndex] = col; + test3_tensor[globalIndex] = gl_WorkGroupID.x; + test4_tensor[globalIndex] = gl_WorkGroupID.y; +}}''' + print(self.shader) + self.compiled_shader = kp.Shader.compile_source(self.shader) + 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, tensor_test_1: kp.Tensor, tensor_test_2: kp.Tensor, + tensor_test_3: kp.Tensor, tensor_test_4: kp.Tensor): + # params = [tensor_in_1, tensor_in_2, tensor_out] + params = [tensor_in_1, tensor_in_2, tensor_out, tensor_test_1, tensor_test_2, tensor_test_3, tensor_test_4] + + if self.algo is None or self.tensor_shape != tensor_shape or self.params != params: + self.tensor_shape = tensor_shape + self.params = params + workgroup = (tensor_shape[0] // self.tile_size, tensor_shape[1] // self.tile_size, 1) + # workgroup = (2, 2, 1) + print(f'{float(tensor_shape[0])=} {self.tile_size=} {workgroup=}') + self.algo = self.mgr.algorithm( + params, # params + self.compiled_shader, # spirv + workgroup, # 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[2:])) + # .record(kp.OpTensorSyncLocal([tensor_out])) + .eval()) + + +def main(): + mgr = kp.Manager() + + matmul_op = MatMulOp(mgr, 4) + + tensor_size = 8 + 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)) + tensor_test_1 = mgr.tensor(np.zeros(tensor_shape)) + tensor_test_2 = mgr.tensor(np.zeros(tensor_shape)) + tensor_test_3 = mgr.tensor(np.zeros(tensor_shape)) + tensor_test_4 = 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) + matmul_op(tensor_shape, tensor_in_1, tensor_in_2, + tensor_out, tensor_test_1, tensor_test_2, tensor_test_3, tensor_test_4) + + # experiment_count = 10 + # 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'test_1 :\n{tensor_test_1.data().reshape(tensor_shape)}') + print(f'test_2 :\n{tensor_test_2.data().reshape(tensor_shape)}') + print(f'test_3 :\n{tensor_test_3.data().reshape(tensor_shape)}') + print(f'test_4 :\n{tensor_test_4.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() diff --git a/examples/python_naive_matmul/3_better_tiling.py b/examples/python_naive_matmul/imp3_better_tiling.py similarity index 78% rename from examples/python_naive_matmul/3_better_tiling.py rename to examples/python_naive_matmul/imp3_better_tiling.py index 5dcc52349..6b3ada314 100644 --- a/examples/python_naive_matmul/3_better_tiling.py +++ b/examples/python_naive_matmul/imp3_better_tiling.py @@ -30,11 +30,12 @@ class MatMulOp: 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''' + self.local_size_x = tile_size + self.local_size_y = tile_size // thread_work_ratio + self.shader = f''' #version 450 -layout (local_size_x = {tile_size}, local_size_y = {local_size_y}) in; +layout (local_size_x = {tile_size}, local_size_y = {self.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[]; }}; @@ -47,8 +48,8 @@ shared float sub_tensor_2[{tile_size}][{tile_size}]; void main() {{ - uint row = gl_GlobalInvocationID.x; - uint col = gl_GlobalInvocationID.y; + uint row = gl_LocalInvocationID.x; + uint col = gl_LocalInvocationID.y; uint globalRow = {tile_size} * gl_WorkGroupID.x + row; uint globalCol = {tile_size} * gl_WorkGroupID.y + row; @@ -62,23 +63,24 @@ void main() {{ 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]; + sub_tensor_1[col + t * {self.local_size_y}][row] = in_tensor_1[ + (tiledCol + t * {self.local_size_y}) * tensor_size + globalRow]; + sub_tensor_2[col + t * {self.local_size_y}][row] = in_tensor_2[ + (globalCol + t * {self.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]; + acc[l] += sub_tensor_1[k][row] * sub_tensor_2[col + l * {self.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]; -}}''') + out_tensor[(globalCol + l * {self.local_size_y}) * tensor_size + globalRow] = acc[l]; +}}''' + self.compiled_shader = kp.Shader.compile_source(self.shader) self.tensor_shape: tuple[int, int] = (0, 0) self.params: list[kp.Tensor] = [] self.algo = None @@ -90,17 +92,20 @@ void main() if self.algo is None or self.tensor_shape != tensor_shape or self.params != params: self.tensor_shape = tensor_shape self.params = params + print( + tensor_shape, self.local_size_x, self.local_size_y, + (tensor_shape[0] // self.local_size_x, tensor_shape[1] // self.local_size_y, 1)) self.algo = self.mgr.algorithm( params, # params - self.shader, # spirv - (tensor_shape[0] // self.tile_size, tensor_shape[1] // self.tile_size, 1), # workgroup + self.compiled_shader, # spirv + (tensor_shape[0] // self.local_size_x, tensor_shape[1] // self.local_size_y, 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)) + .record(kp.OpTensorSyncLocal([tensor_out])) .eval()) diff --git a/examples/python_naive_matmul/matmul_plot.py b/examples/python_naive_matmul/matmul_plot.py new file mode 100644 index 000000000..81763a439 --- /dev/null +++ b/examples/python_naive_matmul/matmul_plot.py @@ -0,0 +1,97 @@ +from argparse import ArgumentParser + +import cv2 +import numpy as np + + +def plot_tensor(window_name: str, tensor: np.ndarray, coord_highlight: tuple[int, int] = None): + font_size = 48 + image = np.zeros((tensor.shape[1] * font_size, tensor.shape[0] * font_size, 3), dtype=np.uint8) + + for y in range(tensor.shape[1]): + for x in range(tensor.shape[0]): + if coord_highlight and x == coord_highlight[1] and y == coord_highlight[0]: + cv2.putText( + image, str(int(tensor[y, x])), (x * font_size, int((y + 0.8) * font_size)), + cv2.FONT_HERSHEY_TRIPLEX, 1., (127, 127, 255)) + else: + cv2.putText( + image, str(int(tensor[y, x])), (x * font_size, int((y + 0.8) * font_size)), + cv2.FONT_HERSHEY_TRIPLEX, 1., (255, 255, 255)) + + cv2.imshow(window_name, image) + + +def main(): + parser = ArgumentParser() + parser.add_argument('tensor_size', type=int, help='Size of the square tensors') + parser.add_argument('tile_size', type=int) + parser.add_argument('local_size', type=int, nargs=2) + parser.add_argument('workgroup', type=int, nargs=2) + arguments = parser.parse_args() + + tensor_size: int = arguments.tensor_size + tile_size: int = arguments.tile_size + local_size: tuple[int, int, int] = tuple(arguments.local_size) + workgroup: tuple[int, int, int] = tuple(arguments.workgroup) + + tensor_shape = (tensor_size, tensor_size) + tensor_1 = np.triu(np.ones(tensor_shape)) + tensor_2 = np.triu(np.ones(tensor_shape)) + tensor_out = np.zeros(tensor_shape) + tensor_test_1 = np.zeros(tensor_shape) + tensor_test_2 = np.zeros(tensor_shape) + tensor_test_3 = np.zeros(tensor_shape) + tensor_test_4 = np.zeros(tensor_shape) + tensor_test_5 = np.zeros(tensor_shape) + + plot_tensor('tensor_1', tensor_1) + plot_tensor('tensor_2', tensor_2) + plot_tensor('tensor_out', tensor_out) + plot_tensor('tensor_test_1', tensor_test_1) + plot_tensor('tensor_test_2', tensor_test_2) + plot_tensor('tensor_test_3', tensor_test_3) + plot_tensor('tensor_test_4', tensor_test_4) + plot_tensor('tensor_test_5', tensor_test_5) + cv2.waitKey(-1) + + print(f'{workgroup=} {local_size=}') + for workgroup_x in range(workgroup[0]): + for workgroup_y in range(workgroup[1]): + for invocation_x in range(workgroup_x * local_size[0], (workgroup_x + 1) * local_size[0]): + for invocation_y in range(workgroup_y * local_size[1], (workgroup_y + 1) * local_size[1]): + row = invocation_x + col = invocation_y + globalRow = (tile_size * workgroup_x) + row + globalCol = (tile_size * workgroup_y) + col + try: + tensor_out[row, col] = row + tensor_test_1[row, col] = col + tensor_test_2[row, col] = workgroup_x + tensor_test_3[row, col] = workgroup_y + tensor_test_4[row, col] = globalRow + tensor_test_5[row, col] = globalCol + plot_tensor('tensor_out', tensor_out, (row, col)) + plot_tensor('tensor_test_1', tensor_test_1, (row, col)) + plot_tensor('tensor_test_2', tensor_test_2, (row, col)) + plot_tensor('tensor_test_3', tensor_test_3, (row, col)) + plot_tensor('tensor_test_4', tensor_test_4, (row, col)) + plot_tensor('tensor_test_5', tensor_test_5, (row, col)) + cv2.waitKey(-1) + except IndexError as error: + print(f'{workgroup_x=} {workgroup_y=} {row=} {col=}') + raise error + + plot_tensor('tensor_1', tensor_1) + plot_tensor('tensor_2', tensor_2) + plot_tensor('tensor_out', tensor_out) + plot_tensor('tensor_test_1', tensor_test_1) + plot_tensor('tensor_test_2', tensor_test_2) + plot_tensor('tensor_test_3', tensor_test_3) + plot_tensor('tensor_test_4', tensor_test_4) + plot_tensor('tensor_test_5', tensor_test_5) + cv2.waitKey(-1) + + +if __name__ == '__main__': + main()