Fix instance_norm、conv2d_xpu、inplace optimizer bugs. #52627
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
你的PR提交成功，感谢你对开源项目的贡献! |
| @@ -186,6 +186,5 @@ PD_REGISTER_KERNEL(assign_value, | |||
| int, | |||
| float, | |||
| double, | |||
| int64_t, | |||
| phi::dtype::float16) {} | |||
There was a problem hiding this comment.
assign kernel 为什么不支持 fp16?
There was a problem hiding this comment.
本身assign_value 这个 kernel 就没有为 fp16 设计，kernel 类型实现和属性绑定的，例如 fp32的 kernel 需要有一个 fp32_values 的属性，int64类型的 kernel 需要有一个 int64_values 的属性，但是 fp16类型没有fp16_values的属性，不适合新增fp16类型注册。
zhupengyang
reviewed
Apr 10, 2023
| saved_mean->data<float>(), | ||
| saved_var->data<float>(), | ||
| true); | ||
| DenseTensor scale_data; |
There was a problem hiding this comment.
float* scale_data_ptr = nullptr;
DenseTensor one_scale; // or dummy_scale
if (!scale_ptr) {
one_scale.Resize({c});
// xdnn interface expects float data
dev_ctx.template Alloc<float>(&one_scale);
phi::funcs::set_constant(dev_ctx, &one_scale, static_cast<float>(1));
scale_data_ptr = one_scale.data<float>();
} else {
scale_data_ptr = scale_ptr->data<float>();
}
| h, | ||
| w, | ||
| epsilon, | ||
| scale_ptr == nullptr ? scale_data.data<float>() |
| epsilon, | ||
| scale_ptr == nullptr ? scale_data.data<float>() | ||
| : scale_ptr->data<float>(), | ||
| bias_ptr == nullptr ? bias_data.data<float>() : bias_ptr->data<float>(), |
csy0225
force-pushed
the
instance_norm_fix
branch
2 times, most recently
from
3823f84 to
d496241
Compare
| const float* bias_data_fp32 = nullptr; | ||
| const auto* bias_ptr = bias.get_ptr(); | ||
| if (bias_ptr == nullptr) { | ||
| DenseTensor bias_data; |
There was a problem hiding this comment.
能定义在{}内部吗？bias_data_fp32 会不会变成野指针？
| scale_data_fp32 = scale_data.data<float>(); | ||
| } else if (scale_ptr->dtype() == | ||
| phi::CppTypeToDataType<phi::dtype::float16>::Type()) { | ||
| float* scale_data_temp = |
There was a problem hiding this comment.
scale_data_temp 这个内存什么时候被释放？RAII_GUARD 析构的时候？
| dev_ctx.template Alloc<float>(&scale_data); | ||
| phi::funcs::set_constant(dev_ctx, &scale_data, static_cast<float>(1)); | ||
| scale_data_fp32 = scale_data.data<float>(); | ||
| } else if (scale_ptr->dtype() == |
There was a problem hiding this comment.
条件是不是应该是：
scale_ptr->dtype() != phi::CppTypeToDataType<phi::dtype::float>::Type()
csy0225
force-pushed
the
instance_norm_fix
branch
from
d496241 to
e3962e2
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR types
Bug fixes
PR changes
OPs
Describe
修复 instance_norm 算子在 scale == null、bias == null 时出现 segment fault 的错误。
修复 inplace 优化算子的前一个算子是 feed 算子时， model input variable shape 被修改，导致第二次推理失败的问题。
修复 conv2d_xpu 算子 has_bias 错误的问题。
取消了 assign_value 算子在 xpu 后端 fp16类型的注册。