RFR: 8366444: Add support for add/mul reduction operations for Float16
Bhavana Kilambi
bkilambi at openjdk.org
Thu Dec 11 12:24:00 UTC 2025
On Fri, 26 Sep 2025 12:00:31 GMT, Bhavana Kilambi <bkilambi at openjdk.org> wrote:
> This patch adds mid-end support for vectorized add/mul reduction operations for half floats. It also includes backend aarch64 support for these operations. Only vectorization support through autovectorization is added as VectorAPI currently does not support Float16 vector species.
>
> Both add and mul reduction vectorized through autovectorization mandate the implementation to be strictly ordered. The following is how each of these reductions is implemented for different aarch64 targets -
>
> **For AddReduction :**
> On Neon only targets (UseSVE = 0): Generates scalarized additions using the scalar `fadd` instruction for both 8B and 16B vector lengths. This is because Neon does not provide a direct instruction for computing strictly ordered floating point add reduction.
>
> On SVE targets (UseSVE > 0): Generates the `fadda` instruction which computes add reduction for floating point in strict order.
>
> **For MulReduction :**
> Both Neon and SVE do not provide a direct instruction for computing strictly ordered floating point multiply reduction. For vector lengths of 8B and 16B, a scalarized sequence of scalar `fmul` instructions is generated and multiply reduction for vector lengths > 16B is not supported.
>
> Below is the performance of the two newly added microbenchmarks in `Float16OperationsBenchmark.java` tested on three different aarch64 machines and with varying `MaxVectorSize` -
>
> Note: On all machines, the score (ops/ms) is compared with the master branch without this patch which generates a sequence of loads (`ldrsh`) to load the FP16 value into an FPR and a scalar `fadd/fmul` to add/multiply the loaded value to the running sum/product. The ratios given below are the ratios between the throughput with this patch and the throughput without this patch.
> Ratio > 1 indicates the performance with this patch is better than the master branch.
>
> **N1 (UseSVE = 0, max vector length = 16B):**
>
> Benchmark vectorDim Mode Cnt 8B 16B
> ReductionAddFP16 256 thrpt 9 1.41 1.40
> ReductionAddFP16 512 thrpt 9 1.41 1.41
> ReductionAddFP16 1024 thrpt 9 1.43 1.40
> ReductionAddFP16 2048 thrpt 9 1.43 1.40
> ReductionMulFP16 256 thrpt 9 1.22 1.22
> ReductionMulFP16 512 thrpt 9 1.21 1.23
> ReductionMulFP16 1024 thrpt 9 1.21 1.22
> ReductionMulFP16 2048 thrpt 9 1.20 1.22
>
>
> On N1, the scalarized sequence of `fadd/fmul` are generated for both `MaxVectorSize` of 8B and 16B for add reduction ...
Apologies, I missed to address the assertion failure you pointed out in my previous comment. It seems to exist because gdb showed that the combined stress flags somehow set the vector length to 4B which is not allowed. The assertion failure itself can be fixed by adding `length < 8` to this condition in aarch64_vector.ad file -
` if (length < 8 || length_in_bytes > 16 || !is_feat_fp16_supported()) {
return false;
}
`
which would avoid vectorization for 4B vector length.
But after this change, the IR rules for reduction fail because now the vector reduction nodes are not generated but the IR rule is expecting them. I'll look into this but I actually noticed that this test fail even on master branch with the following IR failures -
One or more @IR rules failed:
Failed IR Rules (11) of Methods (11)
------------------------------------
1) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorAddConstInputFloat16()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#ADD_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(AddVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
2) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorAddFloat16()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#ADD_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(AddVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
3) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorDivFloat16()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#DIV_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(DivVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
4) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorFmaFloat16()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#FMA_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(FmaVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
5) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorFmaFloat16MixedConstants()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#FMA_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(FmaVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
6) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorFmaFloat16ScalarMixedConstants()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#FMA_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(FmaVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
7) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorMaxFloat16()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#MAX_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(MaxVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
8) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorMinFloat16()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#MIN_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(MinVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
9) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorMulFloat16()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#MUL_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(MulVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
10) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorSqrtFloat16()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#SQRT_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(SqrtVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
11) Method "public void compiler.vectorization.TestFloat16VectorOperations.vectorSubFloat16()" - [Failed IR rules: 1]:
* @IR rule 2: "@compiler.lib.ir_framework.IR(phase={DEFAULT}, applyIfPlatformAnd={}, applyIfCPUFeatureOr={}, counts={"_#V#SUB_VHF#_", " >0 "}, failOn={}, applyIfPlatform={}, applyIfPlatformOr={}, applyIfOr={}, applyIfCPUFeatureAnd={"fphp", "true", "asimdhp", "true"}, applyIf={}, applyIfCPUFeature={}, applyIfAnd={}, applyIfNot={})"
> Phase "PrintIdeal":
- counts: Graph contains wrong number of nodes:
* Constraint 1: "(\d+(\s){2}(SubVHF.*)+(\s){2}===.*vector[A-Za-z]<S,8>)"
- Failed comparison: [found] 0 > 0 [given]
- No nodes matched!
Mostly looks like the expected shape (the default is the `VECTOR_SIZE_MAX`) is not found in the IR graph (as the stress flags might have resulted in a change in vector length) and these failures seem to exist on both aarch64 and x86_64.
-------------
PR Comment: https://git.openjdk.org/jdk/pull/27526#issuecomment-3641660105
More information about the core-libs-dev
mailing list