github/tinygrad
1
1
Fork 0
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-04-29 03:00:14 -04:00
Code Issues Packages Projects Releases Wiki Activity

fix prints, add third conv

Browse Source
This commit is contained in:
George Hotz
2023-01-28 14:10:27 -08:00
parent 92001a06e1
commit 381f3e92da
2 changed files with 12 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
test/external_test_gpu_ast.py
View File

@@ -55,6 +55,16 @@ class TestAST(unittest.TestCase):
ast = LazyOp(MovementOps.RESHAPE, (op8,), (64, 1024, 4))
compile_and_test_ast(ast)
def test_third_op_conv(self):
buf0 = GPUBuffer(shape=ShapeTracker(shape=(1, 64, 128, 4, 4, 1, 1, 8, 4), views=[View((1, 64, 128, 4, 4, 1, 1, 8, 4), (0, 4096, 32, 0, 0, 0, 0, 4, 1), 0)]), hostbuf=GPUBuffer(shape=(64, 1024, 4), force_create=True))
buf1 = GPUBuffer(shape=ShapeTracker(shape=(1, 64, 128, 4, 4, 1, 1, 8, 4), views=[View((1, 64, 128, 4, 4, 1, 1, 8, 4), (0, 0, 0, 128, 4, 0, 0, 16, 1), 0)]), hostbuf=GPUBuffer(shape=(4, 32, 4), force_create=True))
op0 = LazyOp(BinaryOps.MUL, (buf0,buf1,), None)
op1 = LazyOp(ReduceOps.SUM, (op0,), (1, 64, 128, 4, 4, 1, 1, 1, 1))
buf2 = GPUBuffer(shape=ShapeTracker(shape=(1, 64, 128, 4, 4, 1, 1, 1, 1), views=[View((1, 64, 128, 4, 4, 1, 1, 1, 1), (0, 0, 0, 4, 1, 1, 1, 1, 1), 0)]), hostbuf=GPUBuffer(shape=(16,), force_create=True))
op2 = LazyOp(BinaryOps.ADD, (op1,buf2,), None)
ast = LazyOp(MovementOps.RESHAPE, (op2,), (64, 512, 4))
compile_and_test_ast(ast)
# VALIDHACKS=1 IMAGE=2 DEBUG=4 PYTHONPATH="." GPU=1 OPT=2 python3 test/external_test_gpu_ast.py TestAST.test_reduce_op
# 164 time 27.75 ms running re_S128_4 with [128] None count 4 runtime 1016.06 us 2.07 GFLOPS () -> (128, 1)
# 169 time 22.51 ms running matmul with [4, 16, 128] [4, 16, 16] count 5 runtime 110.08 us 19.06 GFLOPS ('-DMATMUL',) -> (128, 1)

4
tinygrad/llops/ops_gpu.py
View File

@@ -193,7 +193,7 @@ class CLASTKernel(ASTKernel):
if self.first_reduce == 2 and isinstance(self.bufs[0]._buf, CLImage):
base_shape = self.bufs[0]._base_shape
if all([(base_shape[0]*base_shape[1])%st.shape[0] == 0 and st.shape[0]//base_shape[0] != 0 for st in self.sts]):
if DEBUG >= 3: print("split opencl", base_shape, self.shapes[0])
if DEBUG >= 3: print("split opencl", base_shape, self.sts[0].shape)
self.reshape_and_permute(lambda x: [base_shape[0], x[0]//base_shape[0]]+list(x[1:]), None)
self.simplify_ones()
@@ -305,7 +305,7 @@ class CLASTKernel(ASTKernel):
def print(self):
super().print()
for i in range(len(self.bufs)):
print(self.buftokens[i], self.bufs[i] in self.earlybufs, self.shapes[i], self.strides[i])
print(self.buftokens[i], self.bufs[i] in self.earlybufs, self.sts[i])
print(self.fxn.prg)
class GPUBuffer(ExplicitExecAST):