Add neureka support

.gitlab-ci.yml

@@ -20,25 +20,41 @@ stages:
   - lint
   - test
 
-format_python:
+python_format:
   stage: lint
   tags:
     - python-lint
   script:
     - black --check .
 
-static_check_python:
+python_sort_imports:
+  stage: lint
+  tags:
+    - python-lint
+  script:
+    - isort --check test
+
+python_static_check:
   stage: lint
   tags:
     - python-lint
   script:
     - pyright .
 
-run_test0:
+run_ne16_test:
   stage: test
   tags:
     - gap9-sdk
   artifacts:
     untracked: true
   script:
-    - cd test && pytest test.py --test-dir tests --recursive
+    - cd test && pytest test.py --test-dir tests --recursive -A ne16
+
+run_neureka_test:
+  stage: test
+  tags:
+    - siracusa-sdk
+  artifacts:
+    untracked: true
+  script:
+    - cd test && pytest test.py --test-dir tests --recursive -A neureka

README.md

@@ -39,46 +39,8 @@ _Note: The accelerator can provide additional helper functions if needed._
 
 ## Accelerators
 
-### NE16
-
-Github repo [link](https://github.com/pulp-platform/ne16).
-
-#### Implemented features
-
-- [x] Convolution w/ kernel shape 1x1
-- [x] Convolution w/ kernel shape 3x3
-- [x] Depthwise convolution w/ kernel shape 3x3
-- [x] Stride 1x1
-- [x] Stride 2x2
-- [ ] Normalization and quantization
-    - [x] With
-    - [ ] Without
-    - [x] Relu (w/ and w/o)
-    - [x] Bias (w/ and w/o)
-    - [ ] Per-channel shift
-    - [x] Per-layer shift
-    - [ ] Rounding
-- [ ] Input type
-    - [x] uint8
-    - [ ] uint16
-- [ ] Output type
-    - [x] int8
-    - [x] uint8 (only w/ Relu)
-    - [ ] int32
-    - [ ] uint32 (only w/ Relu)
-- [ ] Scale type
-    - [x] uint8
-    - [ ] uint16
-    - [ ] uint32
-- [x] Bias type
-    - [x] int32
-- [ ] Weight type
-    - [x] int8
-    - [ ] int2-7
-
-### Neureka
-
-**Untested and considered broken.**
+- [NE16](ne16/README.md)
+- [Neureka](neureka/README.md)
 
 ## Testing
 
@@ -101,7 +63,7 @@ The library will follow the [Semantic Versioning](https://semver.org/).
 
 * Luka Macan <[[email protected]](mailto:[email protected])>
 * Francesco Conti <[[email protected]](mailto:[email protected])>
-* Arpan Prasad <[[email protected]](mailto:[email protected])>
+* Arpan Suravi Prasad <[[email protected]](mailto:[email protected])>
 
 ## License
 

inc/pulp_nnx_ne16.h

@@ -43,7 +43,8 @@ void ne16_nnx_dispatch_wait(ne16_dev_t *dev);
 /** ne16_nnx_dispatch
  *
  * Dispatch a task to the accelerator.
- * Fails with return code 1 if the task cannot be dispatched. Otherwise returns 0.
+ * Fails with return code 1 if the task cannot be dispatched. Otherwise returns
+ * 0.
  */
 int ne16_nnx_dispatch(ne16_dev_t *dev, ne16_task_t *task);
 
@@ -59,7 +60,6 @@ int ne16_nnx_resolve_check(ne16_dev_t *dev, ne16_task_t *task);
  */
 void ne16_nnx_resolve_wait(ne16_dev_t *dev, ne16_task_t *task);
 
-
 /* Additional helper functions */
 
 /** ne16_nnx_dispatch_stride2x2
@@ -70,8 +70,8 @@ void ne16_nnx_resolve_wait(ne16_dev_t *dev, ne16_task_t *task);
  * Works only if the k_out is divisible by 2.
  */
 void ne16_nnx_dispatch_stride2x2(
-    ne16_dev_t *dev, ne16_task_t *task, const uint32_t w_in, const uint32_t k_in,
-    const uint32_t w_in_stride, const uint32_t k_in_stride,
+    ne16_dev_t *dev, ne16_task_t *task, const uint32_t w_in,
+    const uint32_t k_in, const uint32_t w_in_stride, const uint32_t k_in_stride,
     const uint32_t h_out, const uint32_t w_out, const uint32_t k_out,
     const uint32_t w_out_stride, const uint32_t k_out_stride,
     const uint8_t h_ker, const uint8_t w_ker);

inc/pulp_nnx_neureka.h

@@ -0,0 +1,61 @@
+/*
+ * Luka Macan <[email protected]>
+ *
+ * Copyright 2023 ETH Zurich and University of Bologna
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#include "neureka.h"
+#include "neureka_siracusa_bsp.h"
+#include "neureka_task.h"
+#include <stdint.h>
+
+/* PULP-NNX interface */
+
+void neureka_nnx_init(neureka_dev_t *dev, neureka_siracusa_conf_t *conf);
+void neureka_nnx_term(neureka_dev_t *dev);
+
+/** neureka_nnx_dispatch_check
+ *
+ * Check whether you can dispatch to the accelerator.
+ */
+int neureka_nnx_dispatch_check(neureka_dev_t *dev);
+
+/** neureka_nnx_dispatch_wait
+ *
+ * Block until you can dispatch to the accelerator.
+ */
+void neureka_nnx_dispatch_wait(neureka_dev_t *dev);
+
+/** neureka_nnx_dispatch
+ *
+ * Dispatch a task to the accelerator.
+ * Fails with return code 1 if the task cannot be dispatched. Otherwise returns
+ * 0.
+ */
+int neureka_nnx_dispatch(neureka_dev_t *dev, neureka_task_t *task);
+
+/** neureka_nnx_resolve_check
+ *
+ * Check whether the task has been resolved.
+ */
+int neureka_nnx_resolve_check(neureka_dev_t *dev, neureka_task_t *task);
+
+/** neureka_nnx_resolve_wait
+ *
+ * Block until you can resolve the task.
+ */
+void neureka_nnx_resolve_wait(neureka_dev_t *dev, neureka_task_t *task);

ne16/README.md

@@ -0,0 +1,37 @@
+# NE16
+
+## Docs
+
+- Github repo [link](https://github.com/pulp-platform/ne16).
+
+## Implemented features
+
+- [x] Convolution w/ kernel shape 1x1
+- [x] Convolution w/ kernel shape 3x3
+- [x] Depthwise convolution w/ kernel shape 3x3
+- [x] Stride 2x2
+- [ ] Normalization and quantization
+    - [x] With
+    - [ ] Without
+    - [x] Relu (w/ and w/o)
+    - [x] Bias (w/ and w/o)
+    - [ ] Per-channel shift
+    - [x] Per-layer shift
+    - [ ] Rounding
+- [ ] Input type
+    - [x] uint8
+    - [ ] uint16
+- [ ] Output type
+    - [x] int8
+    - [x] uint8 (only w/ Relu)
+    - [ ] int32
+    - [ ] uint32 (only w/ Relu)
+- [ ] Scale type
+    - [x] uint8
+    - [ ] uint16
+    - [ ] uint32
+- [x] Bias type
+    - [x] int32
+- [ ] Weight type
+    - [x] int8
+    - [ ] int2-7

ne16/gvsoc/ne16_gvsoc.h

@@ -42,13 +42,17 @@ typedef enum ne16_gvsoc_log_level_e {
 static void ne16_gvsoc_log_activate(ne16_dev_t *dev,
                                     ne16_gvsoc_log_level_e log_level,
                                     ne16_gvsoc_log_format_e format) {
+#if __PLATFORM__ == ARCHI_PLATFORM_GVSOC
   hwpe_task_reg_write(&dev->hwpe_dev, NE16_REG_GVSOC_LOG_LEVEL, log_level);
   hwpe_task_reg_write(&dev->hwpe_dev, NE16_REG_GVSOC_LOG_FORMAT, format);
+#endif
 }
 
 static void ne16_gvsoc_log_deactivate(ne16_dev_t *dev) {
+#if __PLATFORM__ == ARCHI_PLATFORM_GVSOC
   hwpe_task_reg_write(&dev->hwpe_dev, NE16_REG_GVSOC_LOG_LEVEL,
                       NE16_GVSOC_LOG_LEVEL_CONFIG);
+#endif
 }
 
 #endif // __NE16_GVSOC_H__

ne16/hal/ne16_task.c

@@ -114,11 +114,7 @@ void ne16_task_set_strides(ne16_task_t *task, const uint32_t k_in,
   const uint32_t num_k_in = divnceil(k_in, NE16_INPUT_CHANNEL_THROUGHPUT);
 
   const ne16_stride_t input_stride = {
-      .d0 = k_in_stride,
-      .d1 = k_in_stride * w_in_stride,
-      .d2 = task->depthwise ? 0
-                            : k_in_stride * NE16_FILTER_BUFFER_SIZE *
-                                  NE16_FILTER_BUFFER_SIZE};
+      .d0 = k_in_stride, .d1 = k_in_stride * w_in_stride, .d2 = 0};
   task->data.cfg.input_stride = input_stride;
 
   // WARNING: Stride works only for even output channel sizes (divisible by 2)
@@ -133,20 +129,18 @@ void ne16_task_set_strides(ne16_task_t *task, const uint32_t k_in,
     task->data.cfg.weights_stride.d0 = task->weight_d0_stride * task->qw;
     task->data.cfg.weights_stride.d1 =
         task->weight_d0_stride * task->qw * num_k_in;
-    task->data.cfg.weights_stride.d2 = 0;
   } else if (!task->depthwise) {
     task->data.cfg.weights_stride.d0 =
         NE16_FILTER_SIZE * NE16_FILTER_SIZE * task->weight_d0_stride;
     task->data.cfg.weights_stride.d1 = NE16_FILTER_SIZE * NE16_FILTER_SIZE *
                                        task->weight_d0_stride * task->qw *
                                        num_k_in;
-    task->data.cfg.weights_stride.d2 = 0;
   } else {
     task->data.cfg.weights_stride.d0 =
         NE16_FILTER_SIZE * NE16_FILTER_SIZE * task->weight_d0_stride;
     task->data.cfg.weights_stride.d1 = 0;
-    task->data.cfg.weights_stride.d2 = 0;
   }
+  task->data.cfg.weights_stride.d2 = 0;
 }
 
 void ne16_task_set_counters(ne16_task_t *task, const uint32_t k_in,
@@ -195,8 +189,9 @@ void ne16_task_set_dims(ne16_task_t *task, const uint32_t w_in,
                         const uint32_t k_in, const uint32_t w_in_stride,
                         const uint32_t k_in_stride, const uint32_t h_out,
                         const uint32_t w_out, const uint32_t k_out,
-                        const uint32_t w_out_stride, const uint32_t k_out_stride,
-                        const uint8_t padding_top, const uint8_t padding_bottom,
+                        const uint32_t w_out_stride,
+                        const uint32_t k_out_stride, const uint8_t padding_top,
+                        const uint8_t padding_bottom,
                         const uint8_t padding_right,
                         const uint8_t padding_left) {
   ne16_task_set_strides(task, k_in, w_in_stride, k_in_stride, w_out_stride,
@@ -220,7 +215,8 @@ void ne16_task_set_dims_stride2x2(
   ne16_task_set_strides(task, k_in, w_in_stride, k_in_stride, w_out_stride,
                         k_out_stride);
   ne16_task_set_counters(task, k_in, h_out > 1 ? 3 : 1, w_out > 1 ? 3 : 1,
-                         k_out, h_in + padding_top >= 5 ? 0 : padding_bottom, 0);
+                         k_out, h_in + padding_top >= 5 ? 0 : padding_bottom,
+                         0);
 
   const uint8_t padding_bottom_new =
       (h_in + padding_top - h_ker) % stride == 0 ? 0 : padding_bottom;

ne16/hal/ne16_task.h

@@ -156,8 +156,9 @@ void ne16_task_set_dims(ne16_task_t *task, const uint32_t w_in,
                         const uint32_t k_in, const uint32_t w_in_stride,
                         const uint32_t k_in_stride, const uint32_t h_out,
                         const uint32_t w_out, const uint32_t k_out,
-                        const uint32_t w_out_stride, const uint32_t k_out_stride,
-                        const uint8_t padding_top, const uint8_t padding_bottom,
+                        const uint32_t w_out_stride,
+                        const uint32_t k_out_stride, const uint8_t padding_top,
+                        const uint8_t padding_bottom,
                         const uint8_t padding_right,
                         const uint8_t padding_left);
 void ne16_task_set_dims_stride2x2(

neureka/README.md

@@ -0,0 +1,35 @@
+# Neureka
+
+## Docs
+
+Github repo [link](https://github.com/siracusa-soc/ne).
+
+## Implemented features
+
+- [x] Convolution w/ kernel shape 1x1
+- [x] Convolution w/ kernel shape 3x3
+- [x] Depthwise convolution w/ kernel shape 3x3
+- [ ] Normalization and quantization
+    - [x] With
+    - [ ] Without
+    - [x] Relu (w/ and w/o)
+    - [x] Bias (w/ and w/o)
+    - [ ] Per-channel shift
+    - [x] Per-layer shift
+    - [ ] Rounding
+- [ ] Input type
+    - [x] uint8
+    - [x] int8
+- [ ] Output type
+    - [x] int8
+    - [x] uint8 (only w/ Relu)
+    - [ ] int32
+    - [ ] uint32 (only w/ Relu)
+- [ ] Scale type
+    - [x] uint8
+    - [ ] uint32
+- [x] Bias type
+    - [x] int32
+- [ ] Weight type
+    - [x] int8
+    - [ ] int2-7