thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

Automated work group parameters and tiling matmul

Browse source
This commit is contained in:
Corentin 2021-06-24 18:36:10 +09:00
parent 8c9ad5f2a4
commit 425380f1a1
3 changed files with 183 additions and 21 deletions
Download patch file Download diff file Expand all files Collapse all files

75
examples/python_naive_matmul/1_naive_matmul.py
View file

@ -5,35 +5,64 @@ import numpy as np
class MatMulOp:
def __init__(self, manager: kp.Manager, local_size_x: int = 1, local_size_y: int = 1):
def __init__(self, manager: kp.Manager, local_size_x: int = -1, local_size_y: int = -1):
self.mgr = manager
assert(local_size_x > 0)
assert(local_size_y > 0)
props = self.mgr.get_device_properties()
max_workgroup_invocation = props['max_work_group_invocations']
max_workgroup_size = props['max_work_group_size']
if local_size_x < 1:
if local_size_y > 0:
local_size_x = 1
while (2 * local_size_x * local_size_y <= max_workgroup_invocation
and 2 * local_size_x <= max_workgroup_size[0]):
local_size_x *= 2
else:
local_size_x = 1
local_size_y = 1
while 2 * local_size_x * local_size_y <= max_workgroup_invocation:
if 2 * local_size_x <= max_workgroup_size[0]:
local_size_x *= 2
if 2 * local_size_y <= max_workgroup_size[1]:
local_size_y *= 2
elif 2 * local_size_x > max_workgroup_size[0]: # stop if neither x nor y can be double
break
elif local_size_y < 0:
local_size_y = 1
while (2 * local_size_x * local_size_y <= max_workgroup_invocation
and 2 * local_size_x <= max_workgroup_size[0]):
local_size_y *= 2
assert local_size_x > 0
assert local_size_y > 0
assert local_size_x * local_size_y <= max_workgroup_invocation
assert local_size_x <= max_workgroup_size[0]
assert local_size_y <= max_workgroup_size[1]
self.local_size_x = local_size_x
self.local_size_y = local_size_y
self.shader = kp.Shader.compile_source(f'''
#version 450
#version 450
layout (local_size_x = {local_size_x}, local_size_y = {local_size_y}) in;
layout (local_size_x = {local_size_x}, 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 (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;
layout (constant_id = 0) const float tensor_size_f = 0;
void main()
{{
uint globalRow = gl_GlobalInvocationID.x;
uint globalCol = gl_GlobalInvocationID.y;
uint tensor_size = uint(tensor_size_f);
float acc = 0.0;
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;
}}''')
void main()
{{
uint globalRow = gl_GlobalInvocationID.x;
uint globalCol = gl_GlobalInvocationID.y;
uint tensor_size = uint(tensor_size_f);
float acc = 0.0;
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.tensor_shape: tuple[int, int] = (0, 0)
self.params: list[kp.Tensor] = []
self.algo = None
@ -62,7 +91,7 @@ class MatMulOp:
def main():
mgr = kp.Manager()
matmul_op = MatMulOp(mgr, 64, 64)
matmul_op = MatMulOp(mgr)
tensor_size = 512
tensor_shape = [tensor_size, tensor_size]
@ -92,5 +121,11 @@ def main():
f'{experiment_count * op_count / (1e9 * experiment_time):0.2f}GFLOPS')
def test():
main()
if __name__ == '__main__':
main()
else:
test()