def loadKernel(self, device):
"""Load the kernel and initialize the device."""
self.context = cl.Context([device], None, None)
# These definitions are required for the kernel to function.
self.defines += (' -DOUTPUT_SIZE=' + str(self.OUTPUT_SIZE))
self.defines += (' -DOUTPUT_MASK=' + str(self.OUTPUT_SIZE - 1))
# If the user wants to mine with vectors, enable the appropriate code
# in the kernel source.
if self.VECTORS:
self.defines += ' -DVECTORS'
# Some AMD devices support a special "bitalign" instruction that makes
# bitwise rotation (required for SHA-256) much faster.
if (device.extensions.find('cl_amd_media_ops') != -1):
self.defines += ' -DBITALIGN'
#enable the expierimental BFI_INT instruction optimization
if self.BFI_INT:
self.defines += ' -DBFI_INT'
else:
#since BFI_INT requires cl_amd_media_ops, disable it
if self.BFI_INT:
self.BFI_INT = False
# Locate and read the OpenCL source code in the kernel's directory.
kernelFileDir, pyfile = os.path.split(__file__)
kernelFilePath = os.path.join(kernelFileDir, 'kernel.cl')
kernelFile = open(kernelFilePath, 'r')
kernel = kernelFile.read()
kernelFile.close()
# For fast startup, we cache the compiled OpenCL code. The name of the
# cache is determined as the hash of a few important,
# compilation-specific pieces of information.
m = md5()
m.update(device.platform.name)
m.update(device.platform.version)
m.update(device.name)
m.update(self.defines)
m.update(kernel)
cacheName = '%s.elf' % m.hexdigest()
fileName = os.path.join(kernelFileDir, cacheName)
# Finally, the actual work of loading the kernel...
try:
binary = open(fileName, 'rb')
except IOError:
binary = None
try:
if binary is None:
self.kernel = cl.Program(
self.context, kernel).build(self.defines)
#apply BFI_INT if enabled
if self.BFI_INT:
#patch the binary output from the compiler
patcher = BFIPatcher(self.interface)
binaryData = patcher.patch(self.kernel.binaries[0])
self.interface.debug("Applied BFI_INT patch")
#reload the kernel with the patched binary
self.kernel = cl.Program(
self.context, [device],
[binaryData]).build(self.defines)
#write the kernel binaries to file
binaryW = open(fileName, 'wb')
binaryW.write(self.kernel.binaries[0])
binaryW.close()
else:
binaryData = binary.read()
self.kernel = cl.Program(
self.context, [device], [binaryData]).build(self.defines)
except cl.LogicError:
self.interface.fatal("Failed to compile OpenCL kernel!")
return
except PatchError:
self.interface.fatal('Failed to apply BFI_INT patch to kernel! '
'Is BFI_INT supported on this hardware?')
return
finally:
if binary: binary.close()
cl.unload_compiler()
# If the user didn't specify their own worksize, use the maxium
# supported by the device.
maxSize = self.kernel.search.get_work_group_info(
cl.kernel_work_group_info.WORK_GROUP_SIZE, self.device)
if self.WORKSIZE is None:
self.WORKSIZE = maxSize
else:
if self.WORKSIZE > maxSize:
self.interface.log('Warning: Worksize exceeds the maximum of '
+ str(maxSize) + ', using default.')
if self.WORKSIZE < 1:
self.interface.log('Warning: Invalid worksize, using default.')
self.WORKSIZE = min(self.WORKSIZE, maxSize)
self.WORKSIZE = max(self.WORKSIZE, 1)
#if the worksize is not a power of 2, round down to the nearest one
if (self.WORKSIZE & (self.WORKSIZE - 1)) != 0:
self.WORKSIZE = 1 << int(math.floor(math.log(X)/math.log(2)))
self.interface.setWorkFactor(self.WORKSIZE)
def loadKernel(self, device):
#Load the kernel and initialize the device.
self.context = cl.Context([device], None, None)
# get the maximum worksize of the device
maxWorkSize = self.device.get_info(cl.device_info.MAX_WORK_GROUP_SIZE)
# If the user didn't specify their own worksize, use the maximum supported worksize of the device
if self.WORKSIZE is None:
self.interface.error('WORKSIZE not supplied, using HW max. of ' + str(maxWorkSize))
self.WORKSIZE = maxWorkSize
else:
# If the worksize is larger than the maximum supported worksize of the device
if (self.WORKSIZE > maxWorkSize):
self.interface.error('WORKSIZE out of range, using HW max. of ' + str(maxWorkSize))
self.WORKSIZE = maxWorkSize
# If the worksize is not a power of 2
if (self.WORKSIZE & (self.WORKSIZE - 1)) != 0:
self.interface.error('WORKSIZE invalid, using HW max. of ' + str(maxWorkSize))
self.WORKSIZE = maxWorkSize
# These definitions are required for the kernel to function.
self.defines += (' -DOUTPUT_SIZE=' + str(self.OUTPUT_SIZE))
self.defines += (' -DOUTPUT_MASK=' + str(self.OUTPUT_SIZE - 1))
self.defines += (' -DWORKSIZE=' + str(self.WORKSIZE))
# If the user wants to mine with vectors, enable the appropriate code
# in the kernel source.
if self.VECTORS:
self.defines += ' -DVECTORS'
self.rateDivisor = 2
elif self.VECTORS4:
self.defines += ' -DVECTORS4'
self.rateDivisor = 4
else:
self.rateDivisor = 1
# Some AMD devices support a special "bitalign" instruction that makes
# bitwise rotation (required for SHA-256) much faster.
if (device.extensions.find('cl_amd_media_ops') != -1):
self.defines += ' -DBITALIGN'
#enable the expierimental BFI_INT instruction optimization
if self.BFI_INT:
self.defines += ' -DBFI_INT'
# Locate and read the OpenCL source code in the kernel's directory.
kernelFileDir, pyfile = os.path.split(__file__)
kernelFilePath = os.path.join(kernelFileDir, 'kernel.cl')
kernelFile = open(kernelFilePath, 'r')
kernel = kernelFile.read()
kernelFile.close()
# For fast startup, we cache the compiled OpenCL code. The name of the
# cache is determined as the hash of a few important,
# compilation-specific pieces of information.
m = md5()
m.update(device.platform.name)
m.update(device.platform.version)
m.update(device.name)
m.update(self.defines)
m.update(kernel)
cacheName = '%s.elf' % m.hexdigest()
fileName = os.path.join(kernelFileDir, cacheName)
# Finally, the actual work of loading the kernel...
try:
binary = open(fileName, 'rb')
except IOError:
binary = None
try:
if binary is None:
self.kernel = cl.Program(
self.context, kernel).build(self.defines)
#apply BFI_INT if enabled
if self.BFI_INT:
#patch the binary output from the compiler
patcher = BFIPatcher(self.interface)
binaryData = patcher.patch(self.kernel.binaries[0])
self.interface.debug("Applied BFI_INT patch")
#reload the kernel with the patched binary
self.kernel = cl.Program(
self.context, [device],
[binaryData]).build(self.defines)
#write the kernel binaries to file
binaryW = open(fileName, 'wb')
binaryW.write(self.kernel.binaries[0])
binaryW.close()
else:
binaryData = binary.read()
self.kernel = cl.Program(
self.context, [device], [binaryData]).build(self.defines)
except cl.LogicError:
self.interface.fatal("Failed to compile OpenCL kernel!")
return
except PatchError:
self.interface.fatal('Failed to apply BFI_INT patch to kernel! '
'Is BFI_INT supported on this hardware?')
return
finally:
if binary: binary.close()
#unload the compiler to reduce memory usage
cl.unload_compiler()
# Finally, the actual work of loading the kernel...
try:
binary = open(fileName, 'rb')
except IOError:
binary = None
try:
if binary is None:
self.kernel = cl.Program(
self.context, kernel).build(self.defines)
#apply BFI_INT if enabled
if self.BFI_INT:
#patch the binary output from the compiler
patcher = BFIPatcher(self.interface)
binaryData = patcher.patch(self.kernel.binaries[0])
self.interface.debug("Applied BFI_INT patch")
#reload the kernel with the patched binary
self.kernel = cl.Program(
self.context, [device],
[binaryData]).build(self.defines)
#write the kernel binaries to file
binaryW = open(fileName, 'wb')
binaryW.write(self.kernel.binaries[0])
binaryW.close()
else:
binaryData = binary.read()