summaryrefslogtreecommitdiffstats
path: root/third_party/aom/av1/encoder/ml.c
diff options
context:
space:
mode:
authorMoonchild <mcwerewolf@gmail.com>2018-10-24 05:58:24 +0200
committerGitHub <noreply@github.com>2018-10-24 05:58:24 +0200
commitd1a35c3fa6a59f622becc328bf00eff98732dc53 (patch)
tree6792772d3cb4e22e4bac907376ba17d3030bd008 /third_party/aom/av1/encoder/ml.c
parent81acc4099a515cc1b74ec2b0669aa85fe078aabc (diff)
parent192199b03fa2e56d2728b0de1dbe4bedfc1edc50 (diff)
downloadUXP-d1a35c3fa6a59f622becc328bf00eff98732dc53.tar
UXP-d1a35c3fa6a59f622becc328bf00eff98732dc53.tar.gz
UXP-d1a35c3fa6a59f622becc328bf00eff98732dc53.tar.lz
UXP-d1a35c3fa6a59f622becc328bf00eff98732dc53.tar.xz
UXP-d1a35c3fa6a59f622becc328bf00eff98732dc53.zip
Merge pull request #850 from trav90/add-av1-support
Add initial support for AV1 video.
Diffstat (limited to 'third_party/aom/av1/encoder/ml.c')
-rw-r--r--third_party/aom/av1/encoder/ml.c73
1 files changed, 73 insertions, 0 deletions
diff --git a/third_party/aom/av1/encoder/ml.c b/third_party/aom/av1/encoder/ml.c
new file mode 100644
index 000000000..d21def43a
--- /dev/null
+++ b/third_party/aom/av1/encoder/ml.c
@@ -0,0 +1,73 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <math.h>
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "av1/encoder/ml.h"
+
+void av1_nn_predict(const float *features, const NN_CONFIG *nn_config,
+ float *output) {
+ int num_input_nodes = nn_config->num_inputs;
+ int buf_index = 0;
+ float buf[2][NN_MAX_NODES_PER_LAYER];
+ const float *input_nodes = features;
+
+ // Propagate hidden layers.
+ const int num_layers = nn_config->num_hidden_layers;
+ assert(num_layers <= NN_MAX_HIDDEN_LAYERS);
+ for (int layer = 0; layer < num_layers; ++layer) {
+ const float *weights = nn_config->weights[layer];
+ const float *bias = nn_config->bias[layer];
+ float *output_nodes = buf[buf_index];
+ const int num_output_nodes = nn_config->num_hidden_nodes[layer];
+ assert(num_output_nodes < NN_MAX_NODES_PER_LAYER);
+ for (int node = 0; node < num_output_nodes; ++node) {
+ float val = 0.0f;
+ for (int i = 0; i < num_input_nodes; ++i)
+ val += weights[i] * input_nodes[i];
+ val += bias[node];
+ // ReLU as activation function.
+ val = val > 0.0f ? val : 0.0f; // Could use AOMMAX().
+ output_nodes[node] = val;
+ weights += num_input_nodes;
+ }
+ num_input_nodes = num_output_nodes;
+ input_nodes = output_nodes;
+ buf_index = 1 - buf_index;
+ }
+
+ // Final output layer.
+ const float *weights = nn_config->weights[num_layers];
+ for (int node = 0; node < nn_config->num_outputs; ++node) {
+ const float *bias = nn_config->bias[num_layers];
+ float val = 0.0f;
+ for (int i = 0; i < num_input_nodes; ++i)
+ val += weights[i] * input_nodes[i];
+ output[node] = val + bias[node];
+ weights += num_input_nodes;
+ }
+}
+
+void av1_nn_softmax(const float *input, float *output, int n) {
+ // Softmax function is invariant to adding the same constant
+ // to all input values, so we subtract the maximum input to avoid
+ // possible overflow.
+ float max_inp = input[0];
+ for (int i = 1; i < n; i++) max_inp = AOMMAX(max_inp, input[i]);
+ float sum_out = 0.0f;
+ for (int i = 0; i < n; i++) {
+ output[i] = (float)exp(input[i] - max_inp);
+ sum_out += output[i];
+ }
+ for (int i = 0; i < n; i++) output[i] /= sum_out;
+}