[feat] aclnnHardswishBackwardV2 has the same boundary values with the community-op-plugin-AtomGit

已合并

[feat] aclnnHardswishBackwardV2 has the same boundary values with the community #4978

culechan创建于 5月19日从culechan/op-plugin： feat519 合入到 Ascend/op-plugin： master

[feat] aclnnHardswishBackwardV2 has the same boundary values with the community #4978

已合并

culechan创建于 5月19日

从culechan/op-plugin feat519合入到Ascend/op-plugin master

共 3 个文件变更+54-9

Mop_plugin/ops/opapi/HardSwishBackwardKernelNpuOpApi.cpp+12-7

	@@ -34,17 +34,22 @@ at::Tensor hardswish_backward(const at::Tensor & grad_output, const at::Tensor &
34	at::Tensor hardswish_backward(const at::Tensor & grad_output, const at::Tensor & self)	34	at::Tensor hardswish_backward(const at::Tensor & grad_output, const at::Tensor & self)
35	{	35	{
36	DO_COMPATIBILITY(aclnnHardswishBackward, acl_op::hardswish_backward(grad_output, self));	36	DO_COMPATIBILITY(aclnnHardswishBackward, acl_op::hardswish_backward(grad_output, self));
		37	+ static const bool checkv2_is_available = check_aclnn_kernel_available("aclnnHardswishBackwardV2");
37	auto output_size_0 = self.sizes();	38	auto output_size_0 = self.sizes();
38	auto output_dtype_0 = self.scalar_type();	39	auto output_dtype_0 = self.scalar_type();
39	at::Tensor out = npu_preparation::apply_tensor_without_format(output_size_0,	40	at::Tensor out = npu_preparation::apply_tensor_without_format(output_size_0,
40	grad_output.options().dtype(output_dtype_0));	41	grad_output.options().dtype(output_dtype_0));
41	- EXEC_NPU_CMD(aclnnHardswishBackward, grad_output, self, out);	42	+ if (checkv2_is_available) {
42	- at::Tensor values_le3 = at::empty({}, self.options());	43	+ EXEC_NPU_CMD(aclnnHardswishBackwardV2, grad_output, self, out);
43	- at::Tensor values_ge3 = at::empty({}, self.options());	44	+ } else {
44	- op_api::fill_(values_le3, 0.0f);	45	+ EXEC_NPU_CMD(aclnnHardswishBackward, grad_output, self, out);
45	- op_api::fill_(values_ge3, 1.0f);	46	+ at::Tensor values_le3 = at::empty({}, self.options());
46	- out.index_put_({self.eq(-3.0f)}, values_le3);	47	+ at::Tensor values_ge3 = at::empty({}, self.options());
47	- out.index_put_({self.eq(3.0f)}, values_ge3);	48	+ op_api::fill_(values_le3, 0.0f);
		49	+ op_api::fill_(values_ge3, 1.0f);
		50	+ out.index_put_({self.eq(-3.0f)}, values_le3);
		51	+ out.index_put_({self.eq(3.0f)}, values_ge3);
		52	+ }
48	return out;	53	return out;
49	}	54	}
50	#endif	55	#endif

Mtest/test_base_ops/test_hard_swish.py+18-0

             self.assertRtolEqual(cpu_output, npu_output)
++    def _boundary_values(self):
++        """Values around hardswish boundaries -3 and 3."""
++        return [-3.5, -3.1, -3.0, -2.999, -2.5, -1.0, 0.0, 1.0, 2.5, 2.999, 3.0, 3.1, 3.5]
++
++    def test_hardswish_boundary_fp32(self):
++        boundary = torch.tensor([self._boundary_values()], dtype=torch.float32)
++        cpu_output = self.cpu_op_exec(boundary)
++        npu_input = boundary.to("npu")
++        npu_output = self.npu_op_exec(npu_input)
++        self.assertRtolEqual(cpu_output, npu_output)
++
++    def test_hardswish_boundary_fp16(self):
++        cpu_input = torch.tensor([self._boundary_values()], dtype=torch.float32)
++        cpu_output = self.cpu_op_exec(cpu_input)
++        npu_input = torch.tensor([self._boundary_values()], dtype=torch.float16).to("npu")
++        npu_output = self.npu_op_exec(npu_input)
++        cpu_output = cpu_output.astype(npu_output.dtype)
++        self.assertRtolEqual(cpu_output, npu_output)
 if __name__ == "__main__":
     run_tests()

Mtest/test_base_ops/test_hard_swish_backward.py+24-2

	@@ -8,7 +8,7 @@ from torch_npu.testing.common_utils import create_common_tensor
8		8
9	class TestHardSwishBackWard(TestCase):	9	class TestHardSwishBackWard(TestCase):
10	def cpu_op_exec(self, input1):	10	def cpu_op_exec(self, input1):
11	- input1.requires_grad = True	11	+ input1 = input1.clone().detach().requires_grad_(True)
12	cpu_output = torch.nn.functional.hardswish(input1, inplace=False)	12	cpu_output = torch.nn.functional.hardswish(input1, inplace=False)
13	cpu_output.backward(torch.ones_like(cpu_output))	13	cpu_output.backward(torch.ones_like(cpu_output))
14	output_grad = input1.grad	14	output_grad = input1.grad
	@@ -18,7 +18,7 @@ class TestHardSwishBackWard(TestCase):
18	return cpu_output, output_grad	18	return cpu_output, output_grad
19		19
20	def npu_op_exec(self, input1):	20	def npu_op_exec(self, input1):
21	- input1.requires_grad = True	21	+ input1 = input1.clone().detach().requires_grad_(True)
22	output = torch.nn.functional.hardswish(input1, inplace=False)	22	output = torch.nn.functional.hardswish(input1, inplace=False)
23	output.backward(torch.ones_like(output))	23	output.backward(torch.ones_like(output))
24	output = output.to("cpu")	24	output = output.to("cpu")
	@@ -72,6 +72,28 @@ class TestHardSwishBackWard(TestCase):
72	self.assertRtolEqual(cpu_output, npu_output)	72	self.assertRtolEqual(cpu_output, npu_output)
73	self.assertRtolEqual(cpu_output_grad, npu_output_grad)	73	self.assertRtolEqual(cpu_output_grad, npu_output_grad)
74		74
		75	+ def _boundary_values(self):
		76	+ """Values around hardswish boundaries -3 and 3."""
		77	+ return [-3.5, -3.1, -3.0, -2.999, -2.5, -1.0, 0.0, 1.0, 2.5, 2.999, 3.0, 3.1, 3.5]
		78	+
		79	+ def test_hardswish_boundary_fp32(self):
		80	+ boundary = torch.tensor([self._boundary_values()], dtype=torch.float32)
		81	+ cpu_output, cpu_grad = self.cpu_op_exec(boundary)
		82	+ npu_input = boundary.to("npu")
		83	+ npu_output, npu_grad = self.npu_op_exec(npu_input)
		84	+ self.assertRtolEqual(cpu_output, npu_output)
		85	+ self.assertRtolEqual(cpu_grad, npu_grad)
		86	+
		87	+ def test_hardswish_boundary_fp16(self):
		88	+ cpu_input = torch.tensor([self._boundary_values()], dtype=torch.float32)
		89	+ cpu_output, cpu_grad = self.cpu_op_exec(cpu_input)
		90	+ npu_input = torch.tensor([self._boundary_values()], dtype=torch.float16).to("npu")
		91	+ npu_output, npu_grad = self.npu_op_exec(npu_input)
		92	+ cpu_output = cpu_output.astype(npu_output.dtype)
		93	+ cpu_grad = cpu_grad.astype(npu_grad.dtype)
		94	+ self.assertRtolEqual(cpu_output, npu_output)
		95	+ self.assertRtolEqual(cpu_grad, npu_grad)
		96	+
75		97
76	if __name__ == "__main__":	98	if __name__ == "__main__":
77	run_tests()	99	run_tests()