Allow overriding automatically selected bias memories (#106)

* AI87: reshape kernels for calcx4 mode; RTL sims
* Properly follow the bias group map when specified; --ignore-bias-groups

Author: Robert Muchsel

README.md

@@ -1,6 +1,6 @@
 # MAX78000 Model Training and Synthesis
 
-_February 6, 2021_
+_February 10, 2021_
 
 The Maxim Integrated AI project is comprised of four repositories:
 
@@ -650,6 +650,10 @@ The following example shows the weight memory layout for two layers. The first l
 
 ![Layers and Weight Memory](docs/KernelMemoryLayers.png)
 
+#### Bias Memories
+
+Bias values are stored in separate bias memories. There are four bias memory instances available, and a layer can access any bias memory instance where at least one processor is enabled. By default, bias memories are automatically allocated using a modified Fit-First Descending (FFD) algorithm. Before considering the required resource sizes in descending order, and placing values in the bias memory with most available resources, the algorithm places those bias values that require a single specified bias memory. The bias memory allocation can optionally be controlled using the [`bias_group`](#`bias_group` (Optional)) configuration option.
+
 
 ### Weight Storage Example
 
@@ -1192,6 +1196,7 @@ The following table describes the most important command line arguments for `ai8
 | `--debug-computation`    | Debug computation (SLOW)                                     |                                 |
 | `--stop-after`           | Stop after layer                                             | `--stop-after 2`                |
 | `--one-shot`             | Use layer-by-layer one-shot mechanism                        |                                 |
+| `--ignore-bias-groups`   | Do not force `bias_group` to only available x16 groups       |                                 |
 | *Streaming tweaks*       |                                                              |                                 |
 | `--overlap-data`         | Allow output to overwrite input                              |                                 |
 | `--override-start`       | Override auto-computed streaming start value (x8 hex)        |                                 |
@@ -1549,6 +1554,17 @@ Set `write_gap` to `1` to produce output for a subsequent two-input element-wise
 Example:
 	`write_gap: 1`
 
+##### `bias_group` (Optional)
+
+For layers that use a bias, this key can specify one or more bias memories that should be used. By default, the software uses a “Fit First Descending (FFD)” allocation algorithm that considers largest bias lengths first, and then the layer number, and places each bias in the available group with the most available space, descending to the smallest bias length.
+
+“Available groups” is layer specific and is a list of the groups that have enabled processors for the respective layer. `bias_group` must reference one or more of the available groups. This check can be overridden using the command line option `--ignore-bias-groups` that allows any group or list of groups for any layer.
+
+`bias_group` can be a list of integers or a single integer.
+
+Example:
+	`bias_group: 0`
+
 #### Example
 
 The following shows an example for a single “Fire” operation, the MAX78000/MAX78002 hardware layer numbers and its YAML description.

README.pdf

@@ -196,6 +196,8 @@ def get_parser():
                        help="set ext_rdy bit (default: false)")
     group.add_argument('--weight-start', type=int, metavar='N', default=0,
                        help="specify start offset for weights (debug, default: 0)")
+    group.add_argument('--ignore-bias-groups', action='store_true', default=False,
+                       help="do not force `bias_group` to use an active group (default: false)")
 
     # RTL sim
     group = parser.add_argument_group('RTL simulation')

izer/commandline.py

@@ -593,6 +593,7 @@ def main():
             input_pix_clk=args.input_pix_clk,
             fifo_go=args.fifo_go,
             pretend_zero_sram=args.pretend_zero_sram,
+            ignore_bias_groups=args.ignore_bias_groups,
         )
         if not args.embedded_code and args.autogen.lower() != 'none':
             rtlsim.append_regression(

izer/izer.py

@@ -179,7 +179,7 @@ def bias_sort(e):
     bias_map = sorted(bias_map, key=bias_sort)
 
     for _, (ll, gmap, blen) in enumerate(bias_map):
-        group = group_map[ll][argmin(group_bias_max[t] for t in gmap)]
+        group = gmap[argmin(group_bias_max[t] for t in gmap)]
         if group_bias_max[group] + blen > tc.dev.BIAS_SIZE:
             eprint(f'Layer {ll}: bias memory capacity exceeded - available groups: '
                    f'{gmap}, used so far: {group_bias_max}, needed: {blen}, '

izer/kbias.py

@@ -132,6 +132,16 @@ def load(  # pylint: disable=too-many-branches,too-many-statements
 
             in_exp = 1
             in_chan = in_expand_thresh[ll]
+        elif calcx4[ll]:
+            kernel_reshaped = kernel[ll].reshape(
+                output_chan[ll],
+                in_expand[ll],
+                -1,
+            ).swapaxes(0, 1).reshape(
+                kernel[ll].shape,
+            )
+            in_exp = in_expand[ll]
+            in_chan = input_chan[ll]
         else:
             kernel_reshaped = kernel[ll]
             in_exp = in_expand[ll]

izer/kernels.py

@@ -159,6 +159,7 @@ def create_net(  # pylint: disable=too-many-arguments,too-many-locals,too-many-b
         input_pix_clk=9,
         fifo_go=False,
         pretend_zero_sram=False,
+        ignore_bias_groups=False,
 ):
     """
     Chain multiple CNN layers, create and save input and output
@@ -550,10 +551,11 @@ def create_net(  # pylint: disable=too-many-arguments,too-many-locals,too-many-b
         group_map[ll] = this_map
 
         if bias_group_map[ll] is not None:
-            for _, e in bias_group_map[ll]:
+            for _, e in enumerate(bias_group_map[ll]):
                 if e not in group_map[ll]:
-                    eprint(f'Layer {ll}: `bias_group` references an unused group. Used groups for '
-                           f'this layer are: {group_map[ll]}.')
+                    eprint(f'Layer {ll}: `bias_group` references the unused group {e}. '
+                           f'Used groups for this layer are: {group_map[ll]}.',
+                           error=not ignore_bias_groups)
 
         # Ensure input and output map are the same for passthrough layers
         if operator[ll] == op.NONE:

izer/max7800x.py

@@ -10,8 +10,16 @@
 ./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-mppool-256 --config-file tests/test-mppool-256.yaml --device "$DEVICE" "$@"
 
 ./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-flatten-bias --config-file tests/test-flatten-bias.yaml --device "$DEVICE" "$@"
+./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-tfrock-bias --config-file tests/test-tfrock-bias.yaml --ignore-bias-groups --device "$DEVICE" "$@"
+
+./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-riscv-csv-qfastfifostream-x4-likecifar --config-file tests/test-ffsx4-likecifar10-hwc.yaml --fast-fifo-quad --riscv --device "$DEVICE" "$@"
 
 ./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-qfastfifostream-readahead-multipass --config-file tests/test-ffsreadahead-multipass.yaml --fast-fifo-quad --riscv --device "$DEVICE" "$@"
 ./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-fastfifostream-readahead-multipass --config-file tests/test-ffsreadahead-multipass.yaml --fast-fifo --riscv --device "$DEVICE" "$@"
 ./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-qfastfifostream-x4-readahead-multipass --config-file tests/test-ffsx4readahead-multipass.yaml --fast-fifo-quad --riscv --device "$DEVICE" "$@"
 ./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-fastfifostream-x4-readahead-multipass --config-file tests/test-ffsx4readahead-multipass.yaml --fast-fifo --riscv --device "$DEVICE" "$@"
+
+./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-qfastfifostream-readahead-multipass-bias --config-file tests/test-ffsreadahead-multipass-bias.yaml --fast-fifo-quad --riscv --device "$DEVICE" "$@"
+./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-fastfifostream-readahead-multipass-bias --config-file tests/test-ffsreadahead-multipass-bias.yaml --fast-fifo --riscv --device "$DEVICE" "$@"
+./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-qfastfifostream-x4-readahead-multipass-bias --config-file tests/test-ffsx4readahead-multipass-bias.yaml --fast-fifo-quad --riscv --device "$DEVICE" "$@"
+./ai8xize.py --rtl"$PRELOAD" --verbose --autogen $TARGET --log --test-dir $TARGET --prefix $PREFIX-fastfifostream-x4-readahead-multipass-bias --config-file tests/test-ffsx4readahead-multipass-bias.yaml --fast-fifo --riscv --device "$DEVICE" "$@"