fix(windows): vendor audiopus_sys + patch libopus for clang-cl SIMD

cargo-xwin drives the Windows MSVC cross-compile via clang-cl, under
which CMake sets MSVC=1 — causing libopus 1.3.1's `if(NOT MSVC)` guards
to skip the per-file `-msse4.1` / `-mssse3` COMPILE_FLAGS that its x86
SIMD source files need. Clang-cl (unlike real cl.exe) still honors
Clang's target-feature system, so those files then fail to compile
with "always_inline function '_mm_cvtepi16_epi32' requires target
feature 'sse4.1'" errors across silk/NSQ_sse4_1.c, NSQ_del_dec_sse4_1.c,
and VQ_WMat_EC_sse4_1.c.

Earlier attempts to fix this downstream (cargo-xwin toolchain file,
override.cmake CMAKE_C_COMPILE_OBJECT <FLAGS> replace, CFLAGS env vars)
all failed because cargo-xwin rewrites override.cmake from scratch on
every `cargo xwin build` invocation and cmake-rs's -DCMAKE_C_FLAGS=
assembly happens before toolchain FORCE sets propagate.

Fixing it upstream at the source: vendor audiopus_sys 0.2.2 into
vendor/audiopus_sys, patch its bundled opus/CMakeLists.txt to introduce
an MSVC_CL var (true only when CMAKE_C_COMPILER_ID == "MSVC", i.e. real
cl.exe), and flip the eight `if(NOT MSVC)` SIMD guards to
`if(NOT MSVC_CL)`. Clang-cl then gets the GCC-style per-file flags and
the SSE4.1 sources build cleanly. Also flip the `if(MSVC)` global /arch
block at line 445 to `if(MSVC_CL)` so only cl.exe applies /arch:AVX and
clang-cl relies purely on per-file flags (no global/per-file mixing).

Wire via [patch.crates-io] in the workspace root Cargo.toml; the patch
is resolved relative to the workspace root as `vendor/audiopus_sys`.

Upstream context: xiph/opus#256, xiph/opus PR #257 (both stale).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

vendor/audiopus_sys/opus/training/rnn_dump.py

@@ -0,0 +1,66 @@
+#!/usr/bin/python
+
+from __future__ import print_function
+
+from keras.models import Sequential
+from keras.models import Model
+from keras.layers import Input
+from keras.layers import Dense
+from keras.layers import LSTM
+from keras.layers import GRU
+from keras.models import load_model
+from keras import backend as K
+import sys
+
+import numpy as np
+
+def printVector(f, vector, name):
+    v = np.reshape(vector, (-1));
+    #print('static const float ', name, '[', len(v), '] = \n', file=f)
+    f.write('static const opus_int8 {}[{}] = {{\n   '.format(name, len(v)))
+    for i in range(0, len(v)):
+        f.write('{}'.format(max(-128,min(127,int(round(128*v[i]))))))
+        if (i!=len(v)-1):
+            f.write(',')
+        else:
+            break;
+        if (i%8==7):
+            f.write("\n   ")
+        else:
+            f.write(" ")
+    #print(v, file=f)
+    f.write('\n};\n\n')
+    return;
+
+def binary_crossentrop2(y_true, y_pred):
+        return K.mean(2*K.abs(y_true-0.5) * K.binary_crossentropy(y_pred, y_true), axis=-1)
+
+
+#model = load_model(sys.argv[1], custom_objects={'binary_crossentrop2': binary_crossentrop2})
+main_input = Input(shape=(None, 25), name='main_input')
+x = Dense(32, activation='tanh')(main_input)
+x = GRU(24, activation='tanh', recurrent_activation='sigmoid', return_sequences=True)(x)
+x = Dense(2, activation='sigmoid')(x)
+model = Model(inputs=main_input, outputs=x)
+model.load_weights(sys.argv[1])
+
+weights = model.get_weights()
+
+f = open(sys.argv[2], 'w')
+
+f.write('/*This file is automatically generated from a Keras model*/\n\n')
+f.write('#ifdef HAVE_CONFIG_H\n#include "config.h"\n#endif\n\n#include "mlp.h"\n\n')
+
+printVector(f, weights[0], 'layer0_weights')
+printVector(f, weights[1], 'layer0_bias')
+printVector(f, weights[2], 'layer1_weights')
+printVector(f, weights[3], 'layer1_recur_weights')
+printVector(f, weights[4], 'layer1_bias')
+printVector(f, weights[5], 'layer2_weights')
+printVector(f, weights[6], 'layer2_bias')
+
+f.write('const DenseLayer layer0 = {\n   layer0_bias,\n   layer0_weights,\n   25, 32, 0\n};\n\n')
+f.write('const GRULayer layer1 = {\n   layer1_bias,\n   layer1_weights,\n   layer1_recur_weights,\n   32, 24\n};\n\n')
+f.write('const DenseLayer layer2 = {\n   layer2_bias,\n   layer2_weights,\n   24, 2, 1\n};\n\n')
+
+f.close()

vendor/audiopus_sys/opus/training/rnn_train.py

@@ -0,0 +1,177 @@
+#!/usr/bin/python3
+
+from __future__ import print_function
+
+from keras.models import Sequential
+from keras.models import Model
+from keras.layers import Input
+from keras.layers import Dense
+from keras.layers import LSTM
+from keras.layers import GRU
+from keras.layers import CuDNNGRU
+from keras.layers import SimpleRNN
+from keras.layers import Dropout
+from keras import losses
+import h5py
+from keras.optimizers import Adam
+
+from keras.constraints import Constraint
+from keras import backend as K
+import numpy as np
+
+import tensorflow as tf
+from keras.backend.tensorflow_backend import set_session
+config = tf.ConfigProto()
+config.gpu_options.per_process_gpu_memory_fraction = 0.44
+set_session(tf.Session(config=config))
+
+def binary_crossentrop2(y_true, y_pred):
+    return K.mean(2*K.abs(y_true-0.5) * K.binary_crossentropy(y_true, y_pred), axis=-1)
+
+def binary_accuracy2(y_true, y_pred):
+    return K.mean(K.cast(K.equal(y_true, K.round(y_pred)), 'float32') + K.cast(K.equal(y_true, 0.5), 'float32'), axis=-1)
+
+def quant_model(model):
+    weights = model.get_weights()
+    for k in range(len(weights)):
+        weights[k] = np.maximum(-128, np.minimum(127, np.round(128*weights[k])*0.0078125))
+    model.set_weights(weights)
+
+class WeightClip(Constraint):
+    '''Clips the weights incident to each hidden unit to be inside a range
+    '''
+    def __init__(self, c=2):
+        self.c = c
+
+    def __call__(self, p):
+        return K.clip(p, -self.c, self.c)
+
+    def get_config(self):
+        return {'name': self.__class__.__name__,
+            'c': self.c}
+
+reg = 0.000001
+constraint = WeightClip(.998)
+
+print('Build model...')
+
+main_input = Input(shape=(None, 25), name='main_input')
+x = Dense(32, activation='tanh', kernel_constraint=constraint, bias_constraint=constraint)(main_input)
+#x = CuDNNGRU(24, return_sequences=True, kernel_constraint=constraint, recurrent_constraint=constraint, bias_constraint=constraint)(x)
+x = GRU(24, recurrent_activation='sigmoid', activation='tanh', return_sequences=True, kernel_constraint=constraint, recurrent_constraint=constraint, bias_constraint=constraint)(x)
+x = Dense(2, activation='sigmoid', kernel_constraint=constraint, bias_constraint=constraint)(x)
+model = Model(inputs=main_input, outputs=x)
+
+batch_size = 2048
+
+print('Loading data...')
+with h5py.File('features10b.h5', 'r') as hf:
+    all_data = hf['data'][:]
+print('done.')
+
+window_size = 1500
+
+nb_sequences = len(all_data)//window_size
+print(nb_sequences, ' sequences')
+x_train = all_data[:nb_sequences*window_size, :-2]
+x_train = np.reshape(x_train, (nb_sequences, window_size, 25))
+
+y_train = np.copy(all_data[:nb_sequences*window_size, -2:])
+y_train = np.reshape(y_train, (nb_sequences, window_size, 2))
+
+print("Marking ignores")
+for s in y_train:
+    for e in s:
+        if (e[1] >= 1):
+            break
+        e[0] = 0.5
+
+all_data = 0;
+x_train = x_train.astype('float32')
+y_train = y_train.astype('float32')
+
+print(len(x_train), 'train sequences. x shape =', x_train.shape, 'y shape = ', y_train.shape)
+
+model.load_weights('newweights10a1b_ep206.hdf5')
+
+#weights = model.get_weights()
+#for k in range(len(weights)):
+#    weights[k] = np.round(128*weights[k])*0.0078125
+#model.set_weights(weights)
+
+# try using different optimizers and different optimizer configs
+model.compile(loss=binary_crossentrop2,
+              optimizer=Adam(0.0001),
+              metrics=[binary_accuracy2])
+
+print('Train...')
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=10, validation_data=(x_train, y_train))
+model.save("newweights10a1c_ep10.hdf5")
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=50, initial_epoch=10)
+model.save("newweights10a1c_ep50.hdf5")
+
+model.compile(loss=binary_crossentrop2,
+              optimizer=Adam(0.0001),
+              metrics=[binary_accuracy2])
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=100, initial_epoch=50)
+model.save("newweights10a1c_ep100.hdf5")
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=150, initial_epoch=100)
+model.save("newweights10a1c_ep150.hdf5")
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=200, initial_epoch=150)
+model.save("newweights10a1c_ep200.hdf5")
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=201, initial_epoch=200)
+model.save("newweights10a1c_ep201.hdf5")
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=202, initial_epoch=201, validation_data=(x_train, y_train))
+model.save("newweights10a1c_ep202.hdf5")
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=203, initial_epoch=202, validation_data=(x_train, y_train))
+model.save("newweights10a1c_ep203.hdf5")
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=204, initial_epoch=203, validation_data=(x_train, y_train))
+model.save("newweights10a1c_ep204.hdf5")
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=205, initial_epoch=204, validation_data=(x_train, y_train))
+model.save("newweights10a1c_ep205.hdf5")
+
+quant_model(model)
+model.fit(x_train, y_train,
+          batch_size=batch_size,
+          epochs=206, initial_epoch=205, validation_data=(x_train, y_train))
+model.save("newweights10a1c_ep206.hdf5")
+

vendor/audiopus_sys/opus/training/txt2hdf5.py

@@ -0,0 +1,12 @@
+#!/usr/bin/python
+
+from __future__ import print_function
+
+import numpy as np
+import h5py
+import sys
+
+data = np.loadtxt(sys.argv[1], dtype='float32')
+h5f = h5py.File(sys.argv[2], 'w');
+h5f.create_dataset('data', data=data)
+h5f.close()