[RISCV64]: Implement CPU plugin just-in-time emitter for LogicalXor #30706
Conversation
Signed-off-by: 11happy <[email protected]>
| @@ -261,5 +261,23 @@ class jit_subtract_emitter : public jit_emitter { | |||
| void emit_isa(const std::vector<size_t>& in_vec_idxs, const std::vector<size_t>& out_vec_idxs) const; | |||
| }; | |||
|
|
|||
| class jit_xor_emitter : public jit_emitter { | |||
There was a problem hiding this comment.
Please align the emitter name with op name LogicalXor - it should be jit_logical_xor_emitter.
Ans also please align your changes with others in alphabetical order 😊
| switch (exec_prc_) { | ||
| case ov::element::i32: | ||
| h->vxor_vv(dst, src0, src1); | ||
| break; | ||
| case ov::element::f32: | ||
| h->vxor_vv(dst, src0, src1); | ||
| break; | ||
| default: | ||
| OV_CPU_JIT_EMITTER_THROW("Unsupported precision"); | ||
| } |
There was a problem hiding this comment.
The OpenVINO specification says: "LogicalXor op performs element-wise logical XOR operation".
But the RVV instruction performs "bitwise logical XOR operation". It means that your JIT Emitter works for "BitwiseXor" op - this is another operation with another logic.
So could you please update your implementation to support "elementwise" logical XOR operation?
The example of jit_logical_xor_emitter on ARM64.
And also for now please support only f32 case 😊
There was a problem hiding this comment.
Hii you please tell me the difference between bitwise xor & logical xor with an example it will help me with the implementation.
Thank you
There was a problem hiding this comment.
@11happy Hello! Let me describe simple example to distinguish bitwise xor and logical xor:
Bitwise XOR - we make XOR for each bits of values. The result of this operation new value where each bit is XOR between bits of 2 input values.
Let's imagine that we have 2 FP32 values: 1.2 and 1. Then:
[1.2] : 00111111 10011001 10011001 10011010
[1] : 00111111 10000000 00000000 00000000
bitwise XOR : 00000000 00011001 10011001 10011010 <- The FP32 decimal result is 2.350989e-39. This value should be in vector register
The result of bitwise xor will be number - in binary representation it is equal to xor of bits of two input values.
As for logical XOR - this is per element XOR operation. The result should be true (1) or false (0).
There was a problem hiding this comment.
so would it be fine to implement in this fashion, is they are not equal then its 1 else 0 , also thank you for this descriptive example
There was a problem hiding this comment.
I think that probably there was mistake in my example for logical xor 🤔 And this is not just not equal or equal.
I believe that xor - is logical operation which operates with boolean values (please see specification). It means that all values which are not zero - are automatically 1 or true. Then if only operand is zero and another is true (any not-zero value), then the result is true. Otherwise - false. Like in boolean algebra :)
You can take a look at the example with umbers in pytorch documentation.
| @@ -292,7 +292,8 @@ std::string EltwiseLayerCPUTest::getPrimitiveType(const utils::EltwiseTypes& elt | |||
| if ((eltwise_type == utils::EltwiseTypes::ADD) || | |||
| (eltwise_type == utils::EltwiseTypes::SUBTRACT) || | |||
| (eltwise_type == utils::EltwiseTypes::MULTIPLY) || | |||
| (eltwise_type == utils::EltwiseTypes::DIVIDE)) { | |||
| (eltwise_type == utils::EltwiseTypes::DIVIDE) || | |||
| (eltwise_type == utils::EltwiseTypes::BITWISE_XOR)) { | |||
There was a problem hiding this comment.
As I mentioned in the previous comment, the target operation is LogicalXor, not BitwiseXor.
logical ops are tested there. You don't need to update any checks for primitive type (getPrimitiveType) there as for eltwise/activation ops.
So please just check that the tests with --gtest_filter="*smoke*Logical*LogicalXor*" are successfully passed.
Overview:
Testing:
CC: