Compilation Options¶
The BiShengIR compiler controls the compilation process, feature switches, optimization strategies, and target hardware adaptation through a series of command-line compilation options. The following sections describe the specific configuration rules and parameter meanings of each option by category.
BiShengIR Feature Control Options¶
Applicable products:
Ascend 950PR/Ascend 950DT
Atlas A2 training products/Atlas A2 inference products
Atlas A3 training products/Atlas A3 inference products
Option |
Description |
Type |
Default Value |
|---|---|---|---|
–enable-triton-kernel-compile |
Enables Triton kernel compilation. |
bool |
false |
–enable-torch-compile |
Enables Torch-MLIR compilation (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products). |
bool |
false (this option is provided only when the |
–enable-hivm-compile |
Enables BiShengHIR HIVM compilation. |
bool |
true |
–enable-hfusion-compile |
Enables BiShengHIR HFusion compilation. |
bool |
false |
–enable-symbol-analysis |
Enables symbol analysis (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products). |
bool |
false |
–enable-multi-kernel |
When disabled, the computation graph must be fused into a single kernel; when enabled, supports hoisting to generate multiple kernels (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products). |
bool |
false |
–enable-manage-host-resources |
Enables resource management for host functions (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products). |
bool |
false |
–ensure-no-implicit-broadcast |
Determines whether to ensure that no implicit broadcast semantics exist. If a broadcast from a dynamic dimension to a dynamic dimension exists, a runtime error is triggered (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products). |
bool |
false (this option is provided only when the |
–enable-hivm-graph-sync-solver |
Uses the HIVM computation graph synchronization solver instead of the synchronization injection mechanism. |
bool |
true |
–disable-auto-cv-work-space-manage |
Must be used together with the |
bool |
false |
–disable-hivm-auto-inject-sync |
Disables automatic intra-core synchronization injection. |
bool |
false |
–disable-hivm-tensor-compile |
Disables BiShengHIR HIVM tensor compilation (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products). |
bool |
false |
BiShengIR General Optimization Options¶
Applicable products:
Ascend 950PR/Ascend 950DT
Atlas A2 training products/Atlas A2 inference products
Atlas A3 training products/Atlas A3 inference products
Option |
Description |
Type |
Default Value |
|---|---|---|---|
–enable-auto-multi-buffer |
Enables automatic multi-buffering optimization. For details, see Multi-buffering |
bool |
true |
–limit-auto-multi-buffer-only-for-local-buffer |
After enabling automatic multi-buffering, restricts this optimization to on-chip local buffers (UB, L1, L0C) only, that is, skips multi-buffering of the GM Workspace. This takes effect only for functions whose core type is MIX |
bool |
false |
–enable-tuning-mode |
Enables tuning mode. When memory planning fails, compilation is not retried multiple times |
bool |
false |
–block-dim=<uint> |
Specifies the number of blocks to use |
unsigned |
1 |
BiShengIR HFusion Optimization Options¶
Applicable products:
Ascend 950PR/Ascend 950DT
Atlas A2 training products/Atlas A2 inference products
Atlas A3 training products/Atlas A3 inference products
Option |
Description |
Type |
Default Value |
|---|---|---|---|
–enable-deterministic-computing |
When enabled, the computation results are deterministic; when disabled, additional optimizations are enabled to improve performance (for example, binding reduction operations to multiple cores for execution), but the computation results are not deterministic (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products) |
bool |
true |
–enable-ops-reorder |
Enables operator reordering optimization in the optimization pipeline |
bool |
true |
–hfusion-max-horizontal-fusion-size=<int> |
Maximum number of attempts for horizontal fusion (unlimited by default) (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products) |
int32_t |
-1 |
–hfusion-max-buffer-count-tuning=<long> |
Maximum buffer count tuning threshold in HFusion automatic scheduling (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products) |
int64_t |
0 |
–cube-tiling-tuning=<long> |
Cube tiling size tuning parameter in HFusion automatic scheduling (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products) |
list int64_t |
“” |
–enable-hfusion-count-buffer-dma-opt |
When enabled, the buffers used by DMA operations are not reused by vector operations (only supports Atlas A2 training products/Atlas A2 inference products and Atlas A3 training products/Atlas A3 inference products) |
bool |
false |
BiShengIR HIVM Optimization Options¶
Applicable products:
Ascend 950PR/Ascend 950DT
Atlas A2 training products/Atlas A2 inference products
Atlas A3 training products/Atlas A3 inference products
Option |
Description |
Type |
Default Value |
|---|---|---|---|
–limit-auto-multi-buffer-of-local-buffer=<value> |
After enabling automatic multi-buffering, limits the multi-buffering scope of the local buffer. |
enum |
no-l0c |
–limit-auto-multi-buffer-buffer=<value> |
After enabling automatic multi-buffering, limits the side on which multi-buffering takes effect in MIX operators. |
enum |
|
–set-workspace-multibuffer=<uint> |
Specifies the number of multi-buffers for the GM Workspace |
unsigned |
2 for Ascend 950PR/Ascend 950DT; 4 for Atlas A3 training products/Atlas A3 inference products and Atlas A2 training products/Atlas A2 inference products |
BiShengIR Target Platform Options¶
The compilation option format is --target=Ascend<Name>, which specifies the target hardware platform for MLIR compilation. Here, <Name> is a placeholder that must be replaced with the actual value obtained through the corresponding query command based on the actual AI processor model.
The AI processor models and their corresponding query methods are as follows:
Method 1: Query using the npu-smi info -t board -i <id> -c <chip_id> command
Applicable products:
Ascend 950PR/Ascend 950DT
Atlas A2 training products/Atlas A2 inference products
Atlas A3 training products/Atlas A3 inference products
Run this command on the server where the AI processor is installed to obtain the Chip Name and NPU Name information. The actual configuration value is <Chip Name>_<NPU Name>. For example, if the Chip Name is Ascendxxx and the NPU Name is yyy, the configuration value is Ascendxxx_yyy.
Command parameter description:
id: Device ID. TheNPU IDobtained through thenpu-smi info -lcommand is the device ID.chip_id: Chip ID. TheChip IDobtained through thenpu-smi info -mcommand is the chip ID.
Method 2: Query through the npu-smi info command
Applicable products:
Atlas A2 training products/Atlas A2 inference products
Run this command on the server where the AI processor is installed to obtain the corresponding value of <Name>. The complete configuration value is Ascend<Name>. Example: If the value of <Name> is xxx, the configuration value is Ascendxxx.