hps_accel: Add standalone pool test

Adds standalone test for data pooling.

Signed-off-by: Alan Green <[email protected]>

proj/hps_accel/src/conv2d_call.cc

@@ -36,11 +36,12 @@ void test_conv2d(const Conv2DData* data) {
 
   const tflite::RuntimeShape& output_shape =
       *(reinterpret_cast<const tflite::RuntimeShape*>(data->output_shape));
+
   tflite::reference_integer_ops::ConvPerChannel(
       *(reinterpret_cast<const tflite::ConvParams*>(data->params)),
       reinterpret_cast<const int32_t*>(data->output_multiplier),
       reinterpret_cast<const int32_t*>(data->output_shift), input_shape,
-      reinterpret_cast<const int8_t*>(arena_input),
+      arena_input,
       *(reinterpret_cast<const tflite::RuntimeShape*>(data->filter_shape)),
       reinterpret_cast<const int8_t*>(data->filter_data),
       *(reinterpret_cast<const tflite::RuntimeShape*>(data->bias_shape)),

proj/hps_accel/src/pool_03.h

@@ -0,0 +1,26 @@
+/*
+ * Copyright 2021 The CFU-Playground Authors
+ *
+ * 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.
+ */
+
+#ifndef _POOL_03_H
+#define _POOL_03_H
+
+#include <cstdint>
+
+#include "pool_call.h"
+
+extern PoolData pool_03_data;
+
+#endif  // _POOL_03_H

proj/hps_accel/src/pool_call.cc

@@ -0,0 +1,69 @@
+/*
+ * Copyright 2022 The CFU-Playground Authors
+ *
+ * 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.
+ */
+
+#include "pool_call.h"
+
+#include <cstdio>
+
+#include "playground_util/dump.h"
+#include "tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h"
+#include "tflite.h"
+
+void test_pool(const PoolData* data) {
+  printf("Testing Pool %s\n", data->name);
+  // Copy input arena
+  int8_t* arena_input = reinterpret_cast<int8_t*>(tflite_tensor_arena);
+  auto input_shape =
+      *(reinterpret_cast<const tflite::RuntimeShape*>(data->input_shape));
+  for (int i = 0; i < input_shape.FlatSize(); i++) {
+    arena_input[i] = data->input_data[i];
+  }
+  // Set up output arena
+  int8_t* arena_output =
+      reinterpret_cast<int8_t*>(tflite_tensor_arena) + 128 * 1024;
+  const tflite::RuntimeShape& output_shape =
+      *(reinterpret_cast<const tflite::RuntimeShape*>(data->output_shape));
+
+  tflite::reference_integer_ops::MaxPool(
+      *(reinterpret_cast<const tflite::PoolParams*>(data->params)), input_shape,
+      arena_input, output_shape, arena_output);
+
+  // Check for differences with output
+  int diff_count = 0;
+  int first_diff = 0;
+  int num_words = output_shape.FlatSize() / 4;
+  const int32_t* arena_words = reinterpret_cast<const int32_t*>(arena_output);
+  const int32_t* expected_words =
+      reinterpret_cast<const int32_t*>(data->output_data);
+  for (int i = 0; i < num_words; i++) {
+    if (arena_words[i] != expected_words[i]) {
+      diff_count++;
+      if (diff_count == 1) {
+        first_diff = i;
+      }
+    }
+  }
+
+  if (diff_count == 0) {
+    printf("OK - output identical to golden output\n");
+  } else {
+    printf("FAIL - %d differences, first at word %d\n", diff_count, first_diff);
+    printf("actual:\n");
+    dump_hex(arena_words + first_diff, 16);
+    printf("expected:\n");
+    dump_hex(expected_words + first_diff, 16);
+  }
+}

proj/hps_accel/src/pool_call.h

@@ -0,0 +1,36 @@
+/*
+ * Copyright 2022 The CFU-Playground Authors
+ *
+ * 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.
+ */
+
+#ifndef _POOL_CALL_H
+#define _POOL_CALL_H
+
+#include <cstdint>
+
+// Functionality for calling MaxPool from a test harnesses
+
+struct PoolData {
+  const char* name;
+  const uint8_t* params;
+  const uint8_t* input_shape;
+  const uint8_t* input_data;
+  const uint8_t* output_shape;
+  const uint8_t* output_data;
+};
+
+// Tests MaxPool with the data in the given structure
+void test_pool(const PoolData* data);
+
+#endif  // _POOL_CALL_H

proj/hps_accel/src/proj_menu.cc

@@ -28,10 +28,11 @@
 #include "conv2d_23.h"
 #include "conv2d_23p.h"
 #include "conv2d_call.h"
-#include "fixedpoint/fixedpoint.h"
 #include "hps_cfu.h"
 #include "menu.h"
 #include "playground_util/random.h"
+#include "pool_03.h"
+#include "pool_call.h"
 
 namespace {
 
@@ -78,6 +79,7 @@ void do_test_layer_04(void) { test_conv2d(&conv2d_layer_04_data); }
 void do_test_layer_05(void) { test_conv2d(&conv2d_layer_05_data); }
 void do_test_layer_06(void) { test_conv2d(&conv2d_layer_06_data); }
 void do_test_layer_23p(void) { test_conv2d(&conv2d_layer_23p_data); }
+void do_test_pool_03(void) { test_pool(&pool_03_data); }
 
 struct Menu MENU = {
     "Project Menu",
@@ -97,7 +99,8 @@ struct Menu MENU = {
         MENU_ITEM('4', "test layer 04", do_test_layer_04),
         MENU_ITEM('5', "test layer 05", do_test_layer_05),
         MENU_ITEM('6', "test layer 06", do_test_layer_06),
-        MENU_ITEM('P', "test layer 23p", do_test_layer_23p),
+        MENU_ITEM('7', "test layer 23p", do_test_layer_23p),
+        MENU_ITEM('a', "test pool 03", do_test_pool_03),
         MENU_END,
     },
 };

proj/hps_accel/src/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h

@@ -226,7 +226,7 @@ inline void ConvPerChannel(
                                 output_shape, output_data);
   }
 
-#ifdef SHOW_OUTPUT_HASHES
+#ifdef SHOW_CONV_HASHES
   static int hash_layer = 0;
   int32_t hash = murmurhash3_32(reinterpret_cast<uint8_t*>(output_data),
                                 output_shape.FlatSize());

proj/hps_accel/src/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h

@@ -18,6 +18,7 @@ limitations under the License.
 #include <cstdio>
 #include <limits>
 
+#include "playground_util/murmurhash.h"
 #include "tensorflow/lite/kernels/internal/common.h"
 #include "tensorflow/lite/kernels/internal/reference/integer_ops/pooling_accel.h"
 
@@ -84,40 +85,11 @@ inline bool AveragePool(const PoolParams& params,
   return true;
 }
 
-inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
-                    const int8_t* input_data, const RuntimeShape& output_shape,
-                    int8_t* output_data) {
-#if SHOW_MAX_POOL_PARAMS
-  // padding_width, padding_height,
-  // stride_width, stride_height, filter_height, filter_width,
-  // quantized_activation_min, quantized_activation_max,
-  // input_shape[0], input_shape[1], input_shape[2], input_shape[3],
-  // output_shape[0], output_shape[1], output_shape[2], output_shape[3]
-  printf("\n");
-  const auto& padding = params.padding_values;
-  printf("%d, %d, ", padding.width, padding.height);
-  printf("%d, %d, %d, %d, ", params.stride_height, params.stride_width,
-         params.filter_height, params.filter_width);
-  printf("%ld, %ld, ", params.quantized_activation_min,
-         params.quantized_activation_max);
-  printf("%ld, %ld, %ld, %ld, ", input_shape.Dims(0), input_shape.Dims(1),
-         input_shape.Dims(2), input_shape.Dims(3));
-  printf("%ld, %ld, %ld, %ld, ", output_shape.Dims(0), output_shape.Dims(1),
-         output_shape.Dims(2), output_shape.Dims(3));
-
-  printf("\n");
-#endif
-
-#ifdef ACCEL_MAX_POOL
-#if GATEWARE_GEN != 2
-#error MAX_POOL op requires gateware gen 2
-#endif
-  if (CanAccelerateMaxPool(params, input_shape, output_shape)) {
-    return AccelerateMaxPool(params, input_shape, input_data, output_shape,
-                             output_data);
-  }
-#endif
-
+inline void UnacceleratedMaxPool(const PoolParams& params,
+                                 const RuntimeShape& input_shape,
+                                 const int8_t* input_data,
+                                 const RuntimeShape& output_shape,
+                                 int8_t* output_data) {
   TFLITE_DCHECK_LE(params.quantized_activation_min,
                    params.quantized_activation_max);
   TFLITE_DCHECK_GE(params.quantized_activation_min,
@@ -172,6 +144,58 @@ inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
   }
 }
 
+inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
+                    const int8_t* input_data, const RuntimeShape& output_shape,
+                    int8_t* output_data) {
+#if SHOW_MAX_POOL_PARAMS
+  // padding_width, padding_height,
+  // stride_width, stride_height, filter_height, filter_width,
+  // quantized_activation_min, quantized_activation_max,
+  // input_shape[0], input_shape[1], input_shape[2], input_shape[3],
+  // output_shape[0], output_shape[1], output_shape[2], output_shape[3]
+  printf("\n");
+  const auto& padding = params.padding_values;
+  printf("%d, %d, ", padding.width, padding.height);
+  printf("%d, %d, %d, %d, ", params.stride_height, params.stride_width,
+         params.filter_height, params.filter_width);
+  printf("%ld, %ld, ", params.quantized_activation_min,
+         params.quantized_activation_max);
+  printf("%ld, %ld, %ld, %ld, ", input_shape.Dims(0), input_shape.Dims(1),
+         input_shape.Dims(2), input_shape.Dims(3));
+  printf("%ld, %ld, %ld, %ld, ", output_shape.Dims(0), output_shape.Dims(1),
+         output_shape.Dims(2), output_shape.Dims(3));
+
+  printf("\n");
+#endif
+
+  bool accelerated = false;
+
+#ifdef ACCEL_MAX_POOL
+#if GATEWARE_GEN != 2
+#error MAX_POOL op requires gateware gen 2
+#endif
+  accelerated = CanAccelerateMaxPool(params, input_shape, output_shape);
+  if (accelerated) {
+    AccelerateMaxPool(params, input_shape, input_data, output_shape,
+                      output_data);
+  }
+#endif
+  if (!accelerated) {
+    UnacceleratedMaxPool(params, input_shape, input_data, output_shape,
+                         output_data);
+  }
+
+#ifdef SHOW_POOL_HASHES
+  static int hash_layer = 0;
+  int32_t input_hash = murmurhash3_32(reinterpret_cast<const uint8_t*>(input_data),
+                                      input_shape.FlatSize());
+  int32_t output_hash = murmurhash3_32(reinterpret_cast<const uint8_t*>(output_data),
+                                       output_shape.FlatSize());
+  printf("%3d, %08lx, %08lx\n", hash_layer, input_hash, output_hash);
+  hash_layer++;
+#endif
+}
+
 inline bool AveragePool(const PoolParams& params,
                         const RuntimeShape& input_shape,
                         const int16_t* input_data,