vela: Rename --keep-scale-placement CLI

 - Changed to --cache-bias-scale-tensor

Signed-off-by: Tim Hall <tim.hall@arm.com>
Change-Id: I285fe253f03ba98eff36dbe996ad3a57e2ee3d99

4 files changed, 15 insertions, 9 deletions

diff --git a/OPTIONS.md b/OPTIONS.md
index 156338d..4c5b3e8 100644
--- a/OPTIONS.md
+++ b/OPTIONS.md
@@ -98,13 +98,15 @@ mode.  More details can be found in the Configuration File section below.
 vela network.tflite --config my_vela_cfg1.ini --config my_vela_cfg2.ini --system-config My_Sys_Cfg --memory-mode My_Mem_Mode
 ```
 
-### Keep scale placement
+### Cache Bias Scale Tensor
 
-Prevents scheduler from placing scale tensors for IFM streamed passes in SRAM
-and keeps these in flash.  
+Controls whether the scheduler caches the bias & scale tensors in SRAM or if it
+leaves them in Flash.  This only affects IFM streamed passes.  
+**Type: Boolean**  
+**Default: True**  
 
 ```bash
-vela network.tflite --keep-scale-placement
+vela network.tflite --cache-bias-scale-tensor False
 ```
 
 ### Cascading
diff --git a/ethosu/vela/compiler_driver.py b/ethosu/vela/compiler_driver.py
index 7b1ea21..d17f1e5 100644
--- a/ethosu/vela/compiler_driver.py
+++ b/ethosu/vela/compiler_driver.py
@@ -174,7 +174,7 @@ def compiler_driver(nng, arch, options, scheduler_options):
     # block config, and calc and pack the scales and biases
     weight_compressor.update_pass_weight_and_scale_tensors(nng, arch)
 
-    if not scheduler_options.keep_scale_placement:
+    if scheduler_options.cache_bias_scale_tensor:
         scheduler.move_scales_to_fast_storage(nng, arch)
 
     # LiveRanges for constant tensors for all Npu subgraphs
diff --git a/ethosu/vela/scheduler.py b/ethosu/vela/scheduler.py
index 7347b5a..977eb58 100644
--- a/ethosu/vela/scheduler.py
+++ b/ethosu/vela/scheduler.py
@@ -65,7 +65,7 @@ class SchedulerOptions:
         use_ifm_streaming=True,
         pareto_metric=ParetoMetric.BwCycMem,
         use_nhcwb16_between_cascaded_passes=True,
-        keep_scale_placement=False,
+        cache_bias_scale_tensor=True,
     ):
         self.use_cascading = use_cascading
         self.verbose_schedule = verbose_schedule
@@ -73,7 +73,7 @@ class SchedulerOptions:
         self.use_ifm_streaming = use_ifm_streaming
         self.pareto_metric = pareto_metric
         self.use_nhcwb16_between_cascaded_passes = use_nhcwb16_between_cascaded_passes
-        self.keep_scale_placement = keep_scale_placement
+        self.cache_bias_scale_tensor = cache_bias_scale_tensor
 
     def __str__(self):
         return type(self).__name__ + ": " + str(self.__dict__)
diff --git a/ethosu/vela/vela.py b/ethosu/vela/vela.py
index b93774d..08ab483 100644
--- a/ethosu/vela/vela.py
+++ b/ethosu/vela/vela.py
@@ -245,7 +245,11 @@ def main(args=None):
         "--show-cpu-operations", action="store_true", help="Show the operations that fall back to the CPU"
     )
     parser.add_argument(
-        "--keep-scale-placement", action="store_true", help="Keep scale tensors memory placement during scheduling"
+        "--cache-bias-scale-tensor",
+        type=ast.literal_eval,
+        default=True,
+        choices=[True, False],
+        help="Controls the caching of the bias & scale tensors in SRAM (default: %(default)s)",
     )
     parser.add_argument(
         "--cascading",
@@ -416,7 +420,7 @@ def main(args=None):
         use_ifm_streaming=args.ifm_streaming,
         pareto_metric=args.pareto_metric,
         use_nhcwb16_between_cascaded_passes=args.nhcwb16_between_cascaded_passes,
-        keep_scale_placement=args.keep_scale_placement,
+        cache_bias_scale_tensor=args.cache_bias_scale_tensor,
     )
 
     model_reader_options = model_reader.ModelReaderOptions()