MLBEDSW-4838 Added basic TOSA support.

Added basic TOSA support, enabling Vela to
read and compile  a .tosa file corresponding to
CONV2D + Rescale + Clamp, and writing it to an
optimized .tflite file.

The optimized .tflite file, will in this case, hold
a commandstream where the Rescale and Clamp has been
fused into the CONV2D.

The optimized tflite file is not output from Vela.

  -Added support to read .tosa file into Vela
    internal structure.
      - Added tosa_reader.py, tosa_mapper.py and
        helper files stored under tosa/
      - Support for this limited to ~10 ops

    -Added reader_util.py for functions common
    for TOSA and TFLite

    -Added tosa_graph_optimiser.py
      -Added support to fuse Rescale into convolution
      -Modified handling for padding

    -Added support to fuse Clamp to previous op

    -Added graph_optimiser_util.py
      -Moved functions common for TOSA/TFLite graph
       optimization to this file.

    -Renamed graph_optimiser.py to tflite_graph_optmiser.py

    -Added separate tosa_supported_operators.py

    -Added supported_operator_util.py
       -For functions in common for TOSA/TFLite

Signed-off-by: Patrik Gustavsson <patrik.gustavsson@arm.com>
Change-Id: Ic3c540504ec8c5eb4771397fdc6882050ecf33ab

1 files changed, 10 insertions, 6 deletions

diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py
index 4ba3dee3..7e33e93b 100644
--- a/ethosu/vela/weight_compressor.py
+++ b/ethosu/vela/weight_compressor.py
@@ -203,7 +203,7 @@ def core_deinterleave(hwio, core, ncores):
     return ohwi[core : ohwi.shape[0] : ncores]
 
 
-def _prepare_scale_and_bias(arch, tens, rescale_for_faf):
+def _prepare_scale_and_bias(arch, tens, rescale_for_faf, explicit_scaling):
     assert tens.purpose in [TensorPurpose.FeatureMap, TensorPurpose.FSBias]
     assert tens.format == TensorFormat.NHWC
     # the connected operator should expect a bias input unless it is a FullyConnected
@@ -260,11 +260,15 @@ def _prepare_scale_and_bias(arch, tens, rescale_for_faf):
         else:
             raise UnsupportedFeatureError(f"Compression of {ifm_dtype} is not implemented; Tensor: '{tens.name}'")
 
-    # quantise all of the weight scales into (scale_factor, shift)
-    if ifm_dtype == DataType.int16:
-        quantised_scales = [reduced_quantise_scale(scale) for scale in scales]
+    if explicit_scaling:
+        assert len(explicit_scaling.shift) == len(explicit_scaling.multiplier)
+        quantised_scales = [(int(m), int(s)) for s, m in zip(explicit_scaling.shift, explicit_scaling.multiplier)]
     else:
-        quantised_scales = [quantise_scale(scale) for scale in scales]
+        # quantise all of the weight scales into (scale_factor, shift)
+        if ifm_dtype == DataType.int16:
+            quantised_scales = [reduced_quantise_scale(scale) for scale in scales]
+        else:
+            quantised_scales = [quantise_scale(scale) for scale in scales]
 
     # If only 1 quantised scale is used, repeat that value for the length of the biases
     if len(quantised_scales) == 1:
@@ -355,7 +359,7 @@ def encode_weight_and_scale_tensor(
 
     # Bias & scale
     if do_scales:
-        quantised_scales, biases = _prepare_scale_and_bias(arch, scale_tens, rescale_for_faf)
+        quantised_scales, biases = _prepare_scale_and_bias(arch, scale_tens, rescale_for_faf, op.explicit_scaling)
         scale_tens.element_size_bytes = 10
 
     # Slice the weight stream up depth-ways into bricks and compress