Ran mnist_cnn.py from keras/examples after adding plaidml as the backend. This issue affects many others, but this is the simplest example.

Will run fine for a while, then loss will hit nan and acc will plummet until it hits 0, where it stays.

Andys-iMac-2:examples andy$ python mnist_cnn.py x_train shape: (60000, 28, 28, 1) 60000 train samples 10000 test samples INFO:plaidml:Opening device "amd_radeon_pro_580_compute_engine.0 Train on 60000 samples, validate on 10000 samples Epoch 1/12 59776/60000 [============================>.] - ETA: 0s - loss: 0.3177 - acc: 0.9025INFO:plaidml:Analyzing Ops: 85 of 285 operations complete 60000/60000 [==============================] - 27s - loss: 0.3172 - acc: 0.9026 - val_loss: 0.2699 - val_acc: 0.9217 Epoch 2/12 60000/60000 [==============================] - 18s - loss: 0.1104 - acc: 0.9666 - val_loss: 0.2247 - val_acc: 0.9308 Epoch 3/12 60000/60000 [==============================] - 19s - loss: nan - acc: 0.5408 - val_loss: nan - val_acc: 0.0000e+00 Epoch 4/12 60000/60000 [==============================] - 19s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00 Epoch 5/12 60000/60000 [==============================] - 18s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00 Epoch 6/12 60000/60000 [==============================] - 18s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00 Epoch 7/12 60000/60000 [==============================] - 18s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00 Epoch 8/12 60000/60000 [==============================] - 18s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00 Epoch 9/12 60000/60000 [==============================] - 18s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00 Epoch 10/12 60000/60000 [==============================] - 18s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00 Epoch 11/12 60000/60000 [==============================] - 18s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00 Epoch 12/12 60000/60000 [==============================] - 18s - loss: nan - acc: 0.0000e+00 - val_loss: nan - val_acc: 0.0000e+00 Test loss: nan Test accuracy: 0.0

finally I implemented it in one op for B,H,W,C

class ReflectionPadding2D(PMLTile.Operation):
    def __init__(self, input, h_pad, w_pad):
        if K.image_data_format() == 'channels_last':
            if input.shape.ndims == 4:
                H, W = input.shape.dims[1:3]
                if (type(H) == int and h_pad >= H) or \
                   (type(W) == int and w_pad >= W):
                    raise ValueError("Paddings must be less than dimensions.")
                c = """ function (I[B, H, W, C] ) -> (O) {{
                        WE = W + {w_pad}*2;
                        HE = H + {h_pad}*2;
                    """.format(h_pad=h_pad, w_pad=w_pad)
                if w_pad > 0:
                    c += """
                        LEFT_PAD [b, h, w , c : B, H, WE, C ] = =(I[b, h, {w_pad}-w,            c]), w < {w_pad} ;
                        HCENTER  [b, h, w , c : B, H, WE, C ] = =(I[b, h, w-{w_pad},            c]), w < W+{w_pad}-1 ;
                        RIGHT_PAD[b, h, w , c : B, H, WE, C ] = =(I[b, h, 2*W - (w-{w_pad}) -2, c]);
                        LCR = LEFT_PAD+HCENTER+RIGHT_PAD;
                    """.format(h_pad=h_pad, w_pad=w_pad)
                else:
                    c += "LCR = I;"
                if h_pad > 0:
                    c += """
                        TOP_PAD   [b, h, w , c : B, HE, WE, C ] = =(LCR[b, {h_pad}-h,            w, c]), h < {h_pad};
                        VCENTER   [b, h, w , c : B, HE, WE, C ] = =(LCR[b, h-{h_pad},            w, c]), h < H+{h_pad}-1 ;
                        BOTTOM_PAD[b, h, w , c : B, HE, WE, C ] = =(LCR[b, 2*H - (h-{h_pad}) -2, w, c]);
                        TVB = TOP_PAD+VCENTER+BOTTOM_PAD;
                    """.format(h_pad=h_pad, w_pad=w_pad)
                else:
                    c += "TVB = LCR;"
                c += "O = TVB; }"
                inp_dims = input.shape.dims
                out_dims = (inp_dims[0], inp_dims[1]+h_pad*2, inp_dims[2]+w_pad*2, inp_dims[3])
            else:
                raise NotImplemented
        else:
            raise NotImplemented
        super(ReflectionPadding2D, self).__init__(c, [('I', input) ],
                [('O', PMLTile.Shape(input.shape.dtype, out_dims ) )])

also I implemented it via slice and concat but I suppose it will consume more VRAM for this? or am I wrong??

class ReflectionPadding2D():
    def __init__(self, h_pad, w_pad):
        self.h_pad, self.w_pad = h_pad, w_pad
    def __call__(self, inp):
        h_pad, w_pad = self.h_pad, self.w_pad
        if K.image_data_format() == 'channels_last':
            if inp.shape.ndims == 4:
                w = K.concatenate ([ inp[:,:,w_pad:0:-1,:],
                                     inp,
                                     inp[:,:,-2:-w_pad-2:-1,:] ], axis=2 )
                h = K.concatenate ([ w[:,h_pad:0:-1,:,:],
                                     w,
                                     w[:,-2:-h_pad-2:-1,:,:] ], axis=1 )
                return h
            else:
                raise NotImplemented
        else:
            raise NotImplemented

Traceback (most recent call last): File "C:\Program Files\WindowsApps\PythonSoftwareFoundation.Python.3.7_3.7.1264.0_x64__qbz5n2kfra8p0\lib\runpy.py", line 193, in run_module_as_main "main", mod_spec) File "C:\Program Files\WindowsApps\PythonSoftwareFoundation.Python.3.7_3.7.1264.0_x64__qbz5n2kfra8p0\lib\runpy.py", line 85, in run_code exec(code, run_globals) File "C:\Users\andre\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.7_qbz5n2kfra8p0\LocalCache\local-packages\Python37\Scripts\plaidml-setup.exe_main.py", line 5, in File "C:\Users\andre\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.7_qbz5n2kfra8p0\LocalCache\local-packages\Python37\site-packages\plaidml_init.py", line 50, in import plaidml.settings File "C:\Users\andre\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.7_qbz5n2kfra8p0\LocalCache\local-packages\Python37\site-packages\plaidml\settings.py", line 33, in _setup_config('PLAIDML_EXPERIMENTAL_CONFIG', 'experimental.json') File "C:\Users\andre\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.7_qbz5n2kfra8p0\LocalCache\local-packages\Python37\site-packages\plaidml\settings.py", line 30, in _setup_config 'Could not find PlaidML configuration file: "{}".'.format(filename)) plaidml.exceptions.PlaidMLError: Could not find PlaidML configuration file: "experimental.json".

Hi I am trying plaid ml on AMD Vega 10 : gfx900

I get the following error:

prj47-rack-06@PRJ47-RACK-06:~/biswa/plaidbench$ python plaidbench.py mobilenet Using PlaidML backend. INFO:plaidml:Initializing device gfx900.0: "gfx900", vendor "Advanced Micro Devi ces, Inc." INFO:plaidml:Initializing device gfx900.1: "gfx900", vendor "Advanced Micro Devi ces, Inc." INFO:plaidml:Initializing device gfx900.2: "gfx900", vendor "Advanced Micro Devi ces, Inc." INFO:plaidml:Initializing device gfx900.3: "gfx900", vendor "Advanced Micro Devi ces, Inc." INFO:plaidml:Opening device "gfx900.3": "Advanced Micro Devices, Inc. gfx900"

Model loaded. Compiling and running initial batch, batch_size=1 Warmup Memory access fault by GPU node-7 on address 0x4408bd6000. Reason: Page not pres ent or supervisor privilege. Aborted (core dumped)

Any idea how to resolve this?

Thanks, Biswa

Hello again, I'm experiencing a new issue with the 0.6.0 rc1 version of the plaidml. Using 0.5 led to this issue: https://github.com/plaidml/plaidml/issues/73. Any luck of solving it?

I've got PlaidML running on my AMD Bonaire on Arch Linux with Python 2.7 in a Conda environment. Every other Python package I have runs with 3.6 and my goal is to keep it that way. ;-)

There doesn't seem to even be a pip package for 3.6, so the pip install -U plaidml-keras fails with Python 3.6. If you can post build-from-GitHub-source instructions, I can make a local package and install it.

P.S.: Let me know if you want Arch setup instructions for AMD GPUs. Most of it is on the Arch User Repository wiki but I've got some scripts that do the work.

P.P.S.: Benchmark results

Using PlaidML backend.
INFO:plaidml:Initializing device bonaire.0: "Bonaire", vendor "Advanced Micro Devices, Inc."
INFO:plaidml:Opening device "bonaire.0": "Advanced Micro Devices, Inc. Bonaire"
Downloading data from https://github.com/fchollet/deep-learning-models/releases/download/v0.6/mobilenet_1_0_224_tf.h5
16793600/17225924 [============================>.] - ETA: 0s 
Model loaded.
Compiling and running initial batch, batch_size=1
Warmup
Doing the main timing
Example finished, elapsed: 6.821215868 (compile), 15.0223557949 (execution)

iMac 2017 with a Radeon Pro 580 and a Core i5-7600K. Compiled and installed PlaidML from source. Installed via the pip wheel.

Ran plaidml-setup:

PlaidML Setup (0.0.0.dev0)

Thanks for using PlaidML!

Some Notes:

• Bugs and other issues: https://github.com/plaidml/plaidml
• Questions: https://stackoverflow.com/questions/tagged/plaidml
• Say hello: https://groups.google.com/forum/#!forum/plaidml-dev
• PlaidML is licensed under the GNU AGPLv3

Default Config Devices: No devices.

Experimental Config Devices: intel(r)_core(tm)i5-7600k_cpu@_3.80ghz.0 : Intel Intel(R) Core(TM) i5-7600K CPU @ 3.80GHz

Using experimental devices can cause poor performance, crashes, and other nastiness. Enable experimental device support? (y,n)[n]:y

PlaidML sends anonymous usage statistics to help guide improvements. We'd love your help making it better.

Enable telemetry reporting? (y,n)[y]:y

Almost done. Multiplying some matrices... Tile code: function (B[X,Z], C[Z,Y]) -> (A) { A[x,y : X,Y] = +(B[x,z] * C[z,y]); } ERROR:plaidml:OpenCL: [CL_INVALID_WORK_GROUP_SIZE] : OpenCL Error : clEnqueueNDRangeKernel failed: total work group size (32) is greater than the device can support (1) (cb=12) Whew. That worked.

Save settings to /Users/andy/.plaidml? (y,n)[y]:y Success!

Should a gpu be detected at this point? Is there somewhere I can lower total work group size manually?

New to submitting git issues. Sorry if I'm missing anything.

Hi, Running plaidml-setup gives me the following:

No OpenCL devices found. Check driver installation. Read the helpful, easy driver installation instructions from our README: http://github.com/plaidml/plaidml

This is the output from clinfo: Number of platforms: 1 Platform Profile: FULL_PROFILE Platform Version: OpenCL 2.1 AMD-APP (2766.5) Platform Name: AMD Accelerated Parallel Processing Platform Vendor: Advanced Micro Devices, Inc. Platform Extensions: cl_khr_icd cl_khr_d3d10_sharing cl_khr_d3d11_sharing cl_khr_dx9_media_sharing cl_amd_event_callback cl_amd_offline_devices

Platform Name: AMD Accelerated Parallel Processing Number of devices: 1 Device Type: CL_DEVICE_TYPE_GPU Vendor ID: 1002h Board name: Radeon RX 580 Series Device Topology: PCI[ B#1, D#0, F#0 ] Max compute units: 36 Max work items dimensions: 3 Max work items[0]: 1024 Max work items[1]: 1024 Max work items[2]: 1024 Max work group size: 256 Preferred vector width char: 4 Preferred vector width short: 2 Preferred vector width int: 1 Preferred vector width long: 1 Preferred vector width float: 1 Preferred vector width double: 1 Native vector width char: 4 Native vector width short: 2 Native vector width int: 1 Native vector width long: 1 Native vector width float: 1 Native vector width double: 1 Max clock frequency: 1340Mhz Address bits: 64 Max memory allocation: 4244635648 Image support: Yes Max number of images read arguments: 128 Max number of images write arguments: 64 Max image 2D width: 16384 Max image 2D height: 16384 Max image 3D width: 2048 Max image 3D height: 2048 Max image 3D depth: 2048 Max samplers within kernel: 16 Max size of kernel argument: 1024 Alignment (bits) of base address: 2048 Minimum alignment (bytes) for any datatype: 128 Single precision floating point capability Denorms: No Quiet NaNs: Yes Round to nearest even: Yes Round to zero: Yes Round to +ve and infinity: Yes IEEE754-2008 fused multiply-add: Yes Cache type: Read/Write Cache line size: 64 Cache size: 16384 Global memory size: 8589934592 Constant buffer size: 4244635648 Max number of constant args: 8 Local memory type: Scratchpad Local memory size: 32768 Max pipe arguments: 16 Max pipe active reservations: 16 Max pipe packet size: 4244635648 Max global variable size: 3820172032 Max global variable preferred total size: 8589934592 Max read/write image args: 64 Max on device events: 1024 Queue on device max size: 8388608 Max on device queues: 1 Queue on device preferred size: 262144 SVM capabilities: Coarse grain buffer: Yes Fine grain buffer: Yes Fine grain system: No Atomics: No Preferred platform atomic alignment: 0 Preferred global atomic alignment: 0 Preferred local atomic alignment: 0 Kernel Preferred work group size multiple: 64 Error correction support: 0 Unified memory for Host and Device: 0 Profiling timer resolution: 1 Device endianess: Little Available: Yes Compiler available: Yes Execution capabilities: Execute OpenCL kernels: Yes Execute native function: No Queue on Host properties: Out-of-Order: No Profiling : Yes Queue on Device properties: Out-of-Order: Yes Profiling : Yes Platform ID: 00007FFEC2C66FD0 Name: Ellesmere Vendor: Advanced Micro Devices, Inc. Device OpenCL C version: OpenCL C 2.0 Driver version: 2766.5 Profile: FULL_PROFILE Version: OpenCL 2.0 AMD-APP (2766.5) Extensions: cl_khr_fp64 cl_amd_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_gl_sharing cl_khr_gl_depth_images cl_amd_device_attribute_query cl_amd_vec3 cl_amd_printf cl_amd_media_ops cl_amd_media_ops2 cl_amd_popcnt cl_khr_d3d10_sharing cl_khr_d3d11_sharing cl_khr_dx9_media_sharing cl_khr_image2d_from_buffer cl_khr_spir cl_khr_subgroups cl_khr_gl_event cl_khr_depth_images cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_amd_liquid_flash cl_amd_planar_yuv

Shouldn't it be working? I just switched to a new computer, so I used to use NVIDIA with CUDA. Any help is appreciated!

Note: I do have the most recent AMD driver installed.

I just installed plaidml and i tried to run this example:

#!/usr/bin/env python

import plaidml.keras
plaidml.keras.install_backend() 

import numpy as np
import matplotlib.pyplot as plt
from keras.models import Sequential
from keras.layers.core import Dense, Activation, Dropout
from keras.datasets import mnist
from keras.utils import np_utils

# fix a random seed for reproducibility
np.random.seed(9)

# user inputs
nb_epoch = 25
num_classes = 10
batch_size = 128
train_size = 60000
test_size = 10000
v_length = 784

# split the mnist data into train and test
(trainData, trainLabels), (testData, testLabels) = mnist.load_data()


# reshape the dataset
trainData = trainData.reshape(train_size, v_length)
testData = testData.reshape(test_size, v_length)
trainData = trainData.astype("float32")
testData = testData.astype("float32")
trainData /= 255
testData /= 255


# convert class vectors to binary class matrices --> one-hot encoding
mTrainLabels = np_utils.to_categorical(trainLabels, num_classes)
mTestLabels = np_utils.to_categorical(testLabels, num_classes)

# create the model
model = Sequential()
model.add(Dense(512, input_shape=(784,)))
model.add(Activation("relu"))
model.add(Dropout(0.2))
model.add(Dense(256))
model.add(Activation("relu"))
model.add(Dropout(0.2))
model.add(Dense(num_classes))
model.add(Activation("softmax"))

# summarize the model
model.summary()

# compile the model
model.compile(loss="categorical_crossentropy",
			  optimizer="adam",
			  metrics=["accuracy"])

# fit the model
history = model.fit(trainData, 
				 	mTrainLabels,
					validation_data=(testData, mTestLabels),
					batch_size=batch_size,
					nb_epoch=nb_epoch,
					verbose=2)

# print the history keys


# evaluate the model
scores = model.evaluate(testData, mTestLabels, verbose=0)

# history plot for accuracy
plt.plot(history.history["acc"])
plt.plot(history.history["val_acc"])
plt.title("Model Accuracy")
plt.xlabel("Epoch")
plt.ylabel("Accuracy")
plt.legend(["train", "test"], loc="upper left")
plt.show()

# history plot for accuracy
plt.plot(history.history["loss"])
plt.plot(history.history["val_loss"])
plt.title("Model Loss")
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.legend(["train", "test"], loc="upper left")
plt.show()

and I got this error

could not broadcast input array from shape (3,2048) into shape (6144)

Then I tried running Hello VGG example from plaidml github page and I got the same error.

I am using plaidml 0.3.4 on ubuntu in virtualenv and I am trying to run this code on rx 480.

Tnx for help.

Setup:

sudo apt-get install clinfo
clinfo [sees 1080ti]
sudo pip install -U plaidml-keras
plaidml-setup
[insert before keras import:]
import plaidml.keras
plaidml.keras.install_backend()

But, intermediate problem:

 ImportError: No module named plaidml.keras
$ which python
/home/phobrain/anaconda2/bin//python

Fix:

sys.path.append('/usr/local/lib/python2.7/dist-packages/')
import plaidml.keras
plaidml.keras.install_backend()

'Real' issue being reported:

File "siaconv.py", line 919, in doit epochs=epochs) File "/home/phobrain/anaconda2/lib/python2.7/site-packages/keras/legacy/interfaces.py", line 87, in wrapper return func(*args, **kwargs) File "/home/phobrain/anaconda2/lib/python2.7/site-packages/keras/engine/training.py", line 1926, in fit_generator self._make_train_function() File "/home/phobrain/anaconda2/lib/python2.7/site-packages/keras/engine/training.py", line 967, in _make_train_function **self._function_kwargs) File "/usr/local/lib/python2.7/dist-packages/plaidml/keras/backend.py", line 1718, in function return _Function(inputs, outputs, updates, name) File "/usr/local/lib/python2.7/dist-packages/plaidml/keras/backend.py", line 931, in init c.add_update(_plaidml_val(var), _plaidml_val(newval)) File "/usr/local/lib/python2.7/dist-packages/plaidml/init.py", line 1289, in add_update _lib().plaidml_add_composer_update(self, dest, src) File "/usr/local/lib/python2.7/dist-packages/plaidml/init.py", line 674, in _check_err self.raise_last_status() File "/usr/local/lib/python2.7/dist-packages/plaidml/library.py", line 136, in raise_last_status raise self.last_status() plaidml.exceptions.Unknown: Duplicate updates

model.fit_generator(
        myGen('data', tr_pairs, tr_y, batch_size, True),
        (len(tr_pairs)-1) / batch_size,
        validation_data=myGen('valid', te_pairs, te_y, batch_size, True),
        validation_steps=1,
        max_queue_size=2,
        workers=1,
        epochs=epochs)

Net:

KERNEL_INIT = 'glorot_normal'

    seq.add(Dense(dense_size, input_shape=input_shape,
                activation='relu', kernel_initializer=KERNEL_INIT))
    seq.add(BatchNormalization())
    seq.add(Dense((dense_size*2)/3,
            activation='relu',
            kernel_initializer=KERNEL_INIT))
    seq.add(Dropout(0.1, seed=SEED))
    seq.add(Dense(dense_size/4,
            activation='relu',
            kernel_initializer=KERNEL_INIT))
    seq.add(Dense((dense_size*2)/3,
            activation='relu',
            kernel_initializer=KERNEL_INIT))
    seq.add(Dense(dense_size,
            activation='relu',
            kernel_initializer=KERNEL_INIT))
    seq.add(Dense(512,
                activation='relu',
                kernel_initializer=KERNEL_INIT))
    seq.add(Dense(256,
                activation='relu',
                kernel_initializer=KERNEL_INIT))
    seq.add(Dense(128,
                activation='relu',
                kernel_initializer=KERNEL_INIT))
    seq.add(Dense(256,
                activation='relu',
                kernel_initializer=KERNEL_INIT))
    seq.add(Dense(128,
                activation='relu',
                kernel_initializer=KERNEL_INIT))

Hi,

I've build plaidml 0.3.5 to use on Odroid XU4 with Mali-T628 GPU with debian stretch. I manage to install the wheel, when I run plaidml-setup, I get:

"No supported devices found. Run 'clinfo' and file an issue containing the full output."

However, with plaidml 0.3.0rc1 latest available with pip install plaidml, my devices can be configured and I have 2 mali-t628 reported. "experimental.json" appears quite similar in both cases.

Any clue with what I may have done wrong building plaidml ? (used basel 0.18.1 with --config linux_arm_32v7) or what change might explain 0.3.5 not recognizing my devices where 0.3.0rc1 did ?

Thanks

Here's my clinfo report:

Number of platforms 1 Platform Name ARM Platform Platform Vendor ARM Platform Version OpenCL 1.2 v1.r12p0-04rel0.03af15950392f3702b248717f4938b82 Platform Profile FULL_PROFILE Platform Extensions cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_byte_addressable_store cl_khr_3d_image_writes cl_khr_fp64 cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_fp16 cl_khr_gl_sharing cl_khr_icd cl_khr_egl_event cl_khr_egl_image cl_arm_core_id cl_arm_printf cl_arm_thread_limit_hint cl_arm_non_uniform_work_group_size cl_arm_import_memory Platform Extensions function suffix ARM

Platform Name ARM Platform Number of devices 2 Device Name Mali-T628 Device Vendor ARM Device Vendor ID 0x6200010 Device Version OpenCL 1.2 v1.r12p0-04rel0.03af15950392f3702b248717f4938b82 Driver Version 1.2 Device OpenCL C Version OpenCL C 1.2 v1.r12p0-04rel0.03af15950392f3702b248717f4938b82 Device Type GPU Device Profile FULL_PROFILE Max compute units 4 Max clock frequency 600MHz Device Partition (core) Max number of sub-devices 0 Supported partition types None Max work item dimensions 3 Max work item sizes 256x256x256 Max work group size 256 Preferred work group size multiple 4 Preferred / native vector sizes
char 16 / 16
short 8 / 8
int 4 / 4
long 2 / 2
half 8 / 8 (cl_khr_fp16) float 4 / 4
double 2 / 2 (cl_khr_fp64) Half-precision Floating-point support (cl_khr_fp16) Denormals Yes Infinity and NANs Yes Round to nearest Yes Round to zero Yes Round to infinity Yes IEEE754-2008 fused multiply-add Yes Support is emulated in software No Correctly-rounded divide and sqrt operations No Single-precision Floating-point support (core) Denormals Yes Infinity and NANs Yes Round to nearest Yes Round to zero Yes Round to infinity Yes IEEE754-2008 fused multiply-add Yes Support is emulated in software No Correctly-rounded divide and sqrt operations No Double-precision Floating-point support (cl_khr_fp64) Denormals Yes Infinity and NANs Yes Round to nearest Yes Round to zero Yes Round to infinity Yes IEEE754-2008 fused multiply-add Yes Support is emulated in software No Correctly-rounded divide and sqrt operations No Address bits 64, Little-Endian Global memory size 2090405888 (1.947GiB) Error Correction support No Max memory allocation 522601472 (498.4MiB) Unified memory for Host and Device Yes Minimum alignment for any data type 128 bytes Alignment of base address 1024 bits (128 bytes) Global Memory cache type Read/Write Global Memory cache size <printDeviceInfo:89: get CL_DEVICE_GLOBAL_MEM_CACHE_SIZE : error -30> Global Memory cache line 64 bytes Image support Yes Max number of samplers per kernel 16 Max size for 1D images from buffer 65536 pixels Max 1D or 2D image array size 2048 images Max 2D image size 65536x65536 pixels Max 3D image size 65536x65536x65536 pixels Max number of read image args 128 Max number of write image args 8 Local memory type Global Local memory size 32768 (32KiB) Max constant buffer size 65536 (64KiB) Max number of constant args 8 Max size of kernel argument 1024 Queue properties
Out-of-order execution Yes Profiling Yes Prefer user sync for interop No Profiling timer resolution 1000ns Execution capabilities
Run OpenCL kernels Yes Run native kernels No printf() buffer size 1048576 (1024KiB) Built-in kernels
Device Available Yes Compiler Available Yes Linker Available Yes Device Extensions cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_byte_addressable_store cl_khr_3d_image_writes cl_khr_fp64 cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_fp16 cl_khr_gl_sharing cl_khr_icd cl_khr_egl_event cl_khr_egl_image cl_arm_core_id cl_arm_printf cl_arm_thread_limit_hint cl_arm_non_uniform_work_group_size cl_arm_import_memory

Device Name Mali-T628 Device Vendor ARM Device Vendor ID 0x6200010 Device Version OpenCL 1.2 v1.r12p0-04rel0.03af15950392f3702b248717f4938b82 Driver Version 1.2 Device OpenCL C Version OpenCL C 1.2 v1.r12p0-04rel0.03af15950392f3702b248717f4938b82 Device Type GPU Device Profile FULL_PROFILE Max compute units 2 Max clock frequency 600MHz Device Partition (core) Max number of sub-devices 0 Supported partition types None Max work item dimensions 3 Max work item sizes 256x256x256 Max work group size 256 Preferred work group size multiple 4 Preferred / native vector sizes
char 16 / 16
short 8 / 8
int 4 / 4
long 2 / 2
half 8 / 8 (cl_khr_fp16) float 4 / 4
double 2 / 2 (cl_khr_fp64) Half-precision Floating-point support (cl_khr_fp16) Denormals Yes Infinity and NANs Yes Round to nearest Yes Round to zero Yes Round to infinity Yes IEEE754-2008 fused multiply-add Yes Support is emulated in software No Correctly-rounded divide and sqrt operations No Single-precision Floating-point support (core) Denormals Yes Infinity and NANs Yes Round to nearest Yes Round to zero Yes Round to infinity Yes IEEE754-2008 fused multiply-add Yes Support is emulated in software No Correctly-rounded divide and sqrt operations No Double-precision Floating-point support (cl_khr_fp64) Denormals Yes Infinity and NANs Yes Round to nearest Yes Round to zero Yes Round to infinity Yes IEEE754-2008 fused multiply-add Yes Support is emulated in software No Correctly-rounded divide and sqrt operations No Address bits 64, Little-Endian Global memory size 2090405888 (1.947GiB) Error Correction support No Max memory allocation 522601472 (498.4MiB) Unified memory for Host and Device Yes Minimum alignment for any data type 128 bytes Alignment of base address 1024 bits (128 bytes) Global Memory cache type Read/Write Global Memory cache size <printDeviceInfo:89: get CL_DEVICE_GLOBAL_MEM_CACHE_SIZE : error -30> Global Memory cache line 64 bytes Image support Yes Max number of samplers per kernel 16 Max size for 1D images from buffer 65536 pixels Max 1D or 2D image array size 2048 images Max 2D image size 65536x65536 pixels Max 3D image size 65536x65536x65536 pixels Max number of read image args 128 Max number of write image args 8 Local memory type Global Local memory size 32768 (32KiB) Max constant buffer size 65536 (64KiB) Max number of constant args 8 Max size of kernel argument 1024 Queue properties
Out-of-order execution Yes Profiling Yes Prefer user sync for interop No Profiling timer resolution 1000ns Execution capabilities
Run OpenCL kernels Yes Run native kernels No printf() buffer size 1048576 (1024KiB) Built-in kernels
Device Available Yes Compiler Available Yes Linker Available Yes Device Extensions cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_byte_addressable_store cl_khr_3d_image_writes cl_khr_fp64 cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_fp16 cl_khr_gl_sharing cl_khr_icd cl_khr_egl_event cl_khr_egl_image cl_arm_core_id cl_arm_printf cl_arm_thread_limit_hint cl_arm_non_uniform_work_group_size cl_arm_import_memory

NULL platform behavior clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...) ARM Platform clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...) Success [ARM] clCreateContext(NULL, ...) [default] Success [ARM] clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU) No devices found in platform clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU) Success (2) Platform Name ARM Platform Device Name Mali-T628 Device Name Mali-T628 clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR) No devices found in platform clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM) No devices found in platform clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL) Success (2) Platform Name ARM Platform Device Name Mali-T628 Device Name Mali-T628

ICD loader properties ICD loader Name OpenCL ICD Loader ICD loader Vendor OCL Icd free software ICD loader Version 2.2.11 ICD loader Profile OpenCL 2.1

In the last publicly available version (0.7.0) of PlaidML the Tile language was used to write operators. The latest documentation talks about a C++/Python EDSL being developed as well, however, documentation there is a bit lacking, and only C++ is presented, not any Python version. I was wondering whether that is meant to substitute the Tile language or will both of them be kept in the future? I used to have some test project built using the Tile language from Python, which used to work well for me, and I'd like to improve that, but was wondering if I should port it to the EDSL approach, or is it okay to stay with the Tile language approach.

Also, I was wondering if the v1 branch still supports the Tile language, and will it continue to do so when it is finally released? If I take the path of compiling it from source, and use that instead of the pip installable v0.7, will my project that uses it break or continue working?

Is there any idea of a time-line for the next public release? I have been following this promising project for a while, hoping that a new release will come sooner or later.

Hello.

I try to perform stencil on the following code

affine.store %1, %arg2[%arg3, %arg4, %arg5, %arg6] : memref<?x?x?x?xi64>
affine.yield %arg2 : memref<?x?x?x?xi64>

by using the matchPattern function:

matchPattern(yield, m_Op<AffineYieldOp>(m_Capture(&store, m_Op<AffineStoreOp>(m_Any(), m_Any())))

But it seems m_Capture function and m_Op function used in existing examples, such as StencilGEMM, can not be used to capture operation without a return val, like affine.store here. Can I just use existing structure to match this pattern and capture the affine.store op ?

This patch fix the pass "--x86-stencil-tpp-unary" so that all the reduce patterns in RN50 get stenciled with correct TPP parameters and unary flags.

This wip patch modifies scoped allocs to allocas using PromoteBuffersToStackPass as well as pxa localization pass. As of now, the first pass does not seem to be scoping allocs other than weights. On the other hand, pxa localization pass throws a segfault at runtime for threads>1. This patch also parallelizes layers along batch dimension barring those which don't have batch dimension as the outer loop's induction variable.

This patch adds support for parallelization directives to be specified in a file using the environment <PLAIDML_PARALLELIZATION_CONFIG_FILE>. This patch adds a rule parser which matches the shapes of convolution based on the equalities/inequalities in the config file and applies the rules that follow.
It is important to note that collapse directive also adds a parallelize directive by default and can only be applied to 2 loop levels corresponding to a perfect loop nest (validity of reordering of loops in order to support the requested order is not verified).

Hey all,

I've decided to try some mL projects and since I have a amd gpu (5700xt) I decided to use Plaidml. On the main website theres a test code for VGG-19 and I'm trying to run it right now but I run into the error attached in the screenshot. I tried to simply change collections to collections.abc but it looks like python 3.10 already does that? I'm pretty stuck, any help would be appreciated. Thanks!