def find_gpu_target():
# Start with a target suitable for the machine you're running this on.
target = hl.get_host_target()
features_to_try = []
if target.os == hl.TargetOS.Windows:
# Try D3D12 first; if that fails, try OpenCL.
if struct.calcsize("P") == 8:
# D3D12Compute support is only available on 64-bit systems at present.
features_to_try.append(hl.TargetFeature.D3D12Compute)
features_to_try.append(hl.TargetFeature.OpenCL)
elif target.os == hl.TargetOS.OSX:
# OS X doesn't update its OpenCL drivers, so they tend to be broken.
# CUDA would also be a fine choice on machines with NVidia GPUs.
features_to_try.append(hl.TargetFeature.Metal)
else:
features_to_try.append(hl.TargetFeature.OpenCL)
# Uncomment the following lines to also try CUDA:
# features_to_try.append(hl.TargetFeature.CUDA);
for f in features_to_try:
new_target = target.with_feature(f)
if (hl.host_supports_target_device(new_target)):
return new_target
print(
"Requested GPU(s) are not supported. (Do you have the proper hardware and/or driver installed?)"
)
return target
def stereoBM(left_image, right_image, width, height, SADWindowSize,
minDisparity, numDisparities):
x, y, c = Var("x"), Var("y"), Var("c")
left, right = Func("left"), Func("right")
left[x, y, c] = left_image[x, y, c]
right[x, y, c] = right_image[x, y, c]
W, H, C = left_image.width(), left_image.height(), left_image.channels()
# Sobel filtering
filteredLeft = prefilterXSobel(left, W, H)
filteredRight = prefilterXSobel(right, W, H)
# Stereo block-matching
win2 = SADWindowSize / 2
maxDisparity = numDisparities - 1
xmin = maxDisparity + win2
xmax = width - win2 - 1
ymin = win2
ymax = height - win2 - 1
x_tile_size, y_tile_size = 32, 32
disp = findStereoCorrespondence(filteredLeft, filteredRight, SADWindowSize,
minDisparity, numDisparities, xmin, xmax,
ymin, ymax, x_tile_size, y_tile_size)
# Write to pretty html
args = h.ArgumentsVector()
disp.compile_to_lowered_stmt("disp.html", args, h.HTML)
# Compile / Profile
profile(disp, W, H)
# Start with a target suitable for the machine you're running
# this on.
target = h.get_host_target()
disp_image = disp.realize(W, H)
return disp_image
def schedule_for_cpu(self):
# Compute the look-up-table ahead of time.
self.lut.compute_root()
# Compute color channels innermost. Promise that there will
# be three of them and unroll across them.
self.curved.reorder(c, x, y) \
.bound(c, 0, 3) \
.unroll(c)
# Look-up-tables don't vectorize well, so just parallelize
# curved in slices of 16 scanlines.
self.curved.split(y, yo, yi, 16) \
.parallel(yo)
# Compute sharpen as needed per scanline of curved, reusing
# previous values computed within the same strip of 16
# scanlines.
self.sharpen.compute_at(self.curved, yi)
# Vectorize the sharpen. It's 16-bit so we'll vectorize it 8-wide.
self.sharpen.vectorize(x, 8)
# Compute the padded input at the same granularity as the
# sharpen. We'll leave the cast to 16-bit inlined into
# sharpen.
self.padded.store_at(self.curved, yo) \
.compute_at(self.curved, yi)
# Also vectorize the padding. It's 8-bit, so we'll vectorize
# 16-wide.
self.padded.vectorize(x, 16)
# JIT-compile the pipeline for the CPU.
target = hl.get_host_target()
self.curved.compile_jit(target)
return
def stereoBM(left_image, right_image, width, height,
SADWindowSize, minDisparity, numDisparities):
x, y, c = Var("x"), Var("y"), Var("c")
left, right = Func("left"), Func("right")
left[x, y, c] = left_image[x, y, c]
right[x, y, c] = right_image[x, y, c]
W, H, C = left_image.width(), left_image.height(), left_image.channels()
# Sobel filtering
filteredLeft = prefilterXSobel(left, W, H)
filteredRight = prefilterXSobel(right, W, H)
# Stereo block-matching
win2 = SADWindowSize / 2
maxDisparity = numDisparities - 1
xmin = maxDisparity + win2
xmax = width - win2 - 1
ymin = win2
ymax = height - win2 - 1
x_tile_size, y_tile_size = 32, 32
disp = findStereoCorrespondence(filteredLeft, filteredRight, SADWindowSize, minDisparity, numDisparities,
xmin, xmax, ymin, ymax, x_tile_size, y_tile_size)
# Write to pretty html
args = h.ArgumentsVector()
disp.compile_to_lowered_stmt("disp.html", args, h.HTML)
# Compile / Profile
profile(disp, W, H)
# Start with a target suitable for the machine you're running
# this on.
target = h.get_host_target()
disp_image = disp.realize(W, H)
return disp_image
def schedule_for_gpu(self):
# We make the decision about whether to use the GPU for each
# hl.Func independently. If you have one hl.Func computed on the
# CPU, and the next computed on the GPU, Halide will do the
# copy-to-gpu under the hood. For this pipeline, there's no
# reason to use the CPU for any of the stages. Halide will
# copy the input image to the GPU the first time we run the
# pipeline, and leave it there to reuse on subsequent runs.
# As before, we'll compute the LUT once at the start of the
# pipeline.
self.lut.compute_root()
# Let's compute the look-up-table using the GPU in 16-wide
# one-dimensional thread blocks. First we split the index
# into blocks of size 16:
block, thread = hl.Var("block"), hl.Var("thread")
self.lut.split(i, block, thread, 16)
# Then we tell cuda that our Vars 'block' and 'thread'
# correspond to CUDA's notions of blocks and threads, or
# OpenCL's notions of thread groups and threads.
self.lut.gpu_blocks(block) \
.gpu_threads(thread)
# This is a very common scheduling pattern on the GPU, so
# there's a shorthand for it:
# lut.gpu_tile(i, ii, 16)
# hl.Func::gpu_tile method is similar to hl.Func::tile, except that
# it also specifies that the tile coordinates correspond to
# GPU blocks, and the coordinates within each tile correspond
# to GPU threads.
# Compute color channels innermost. Promise that there will
# be three of them and unroll across them.
self.curved.reorder(c, x, y) \
.bound(c, 0, 3) \
.unroll(c)
# Compute curved in 2D 8x8 tiles using the GPU.
self.curved.gpu_tile(x, y, xi, yi, 8, 8)
# This is equivalent to:
# curved.tile(x, y, xo, yo, xi, yi, 8, 8)
# .gpu_blocks(xo, yo)
# .gpu_threads(xi, yi)
# We'll leave sharpen as inlined into curved.
# Compute the padded input as needed per GPU block, storing the
# intermediate result in shared memory. hl.Var::gpu_blocks, and
# hl.Var::gpu_threads exist to help you schedule producers within
# GPU threads and blocks.
self.padded.compute_at(self.curved, x)
# Use the GPU threads for the x and y coordinates of the
# padded input.
self.padded.gpu_threads(x, y)
# JIT-compile the pipeline for the GPU. CUDA or OpenCL are
# not enabled by default. We have to construct a hl.Target
# object, enable one of them, and then pass that target
# object to compile_jit. Otherwise your CPU will very slowly
# pretend it's a GPU, and use one thread per output pixel.
# Start with a target suitable for the machine you're running
# this on.
target = hl.get_host_target()
# Then enable OpenCL or CUDA.
#use_opencl = False
use_opencl = True
if use_opencl:
# We'll enable OpenCL here, because it tends to give better
# performance than CUDA, even with NVidia's drivers, because
# NVidia's open source LLVM backend doesn't seem to do all
# the same optimizations their proprietary compiler does.
target.set_feature(hl.TargetFeature.OpenCL)
print("(Using OpenCL)")
else:
# Uncomment the next line and comment out the line above to
# try CUDA instead.
target.set_feature(hl.TargetFeature.CUDA)
print("(Using CUDA)")
# If you want to see all of the OpenCL or CUDA API calls done
# by the pipeline, you can also enable the Debug
# flag. This is helpful for figuring out which stages are
# slow, or when CPU -> GPU copies happen. It hurts
# performance though, so we'll leave it commented out.
# target.set_feature(hl.TargetFeature.Debug)
self.curved.compile_jit(target)
def test_target():
# Target("") should be exactly like get_host_target().
t1 = hl.get_host_target()
t2 = hl.Target("")
assert t1 == t2, "Default ctor failure"
assert t1.supported()
# to_string roundtripping
t1 = hl.Target()
ts = t1.to_string()
assert ts == "arch_unknown-0-os_unknown"
# Note, this should *not* validate, since validate_target_string
# now returns false if any of arch-bits-os are undefined
assert not hl.Target.validate_target_string(ts)
# Don't attempt to roundtrip this: trying to create
# a Target with unknown portions will now assert-fail.
#
# t2 = hl.Target(ts)
# assert t2 == t1
# repr() and str()
assert str(t1) == "arch_unknown-0-os_unknown"
assert repr(t1) == "<halide.Target arch_unknown-0-os_unknown>"
assert t1.os == hl.TargetOS.OSUnknown
assert t1.arch == hl.TargetArch.ArchUnknown
assert t1.bits == 0
# Full specification round-trip:
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32,
[hl.TargetFeature.SSE41])
ts = t1.to_string()
assert ts == "x86-32-linux-sse41"
assert hl.Target.validate_target_string(ts)
# Full specification (without features) round-trip:
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32)
ts = t1.to_string()
assert ts == "x86-32-linux"
assert hl.Target.validate_target_string(ts)
# Full specification round-trip, crazy features
t1 = hl.Target(hl.TargetOS.Android, hl.TargetArch.ARM, 32, [
hl.TargetFeature.JIT, hl.TargetFeature.SSE41, hl.TargetFeature.AVX,
hl.TargetFeature.AVX2, hl.TargetFeature.CUDA, hl.TargetFeature.OpenCL,
hl.TargetFeature.OpenGL, hl.TargetFeature.OpenGLCompute,
hl.TargetFeature.Debug
])
ts = t1.to_string()
assert ts == "arm-32-android-avx-avx2-cuda-debug-jit-opencl-opengl-openglcompute-sse41"
assert hl.Target.validate_target_string(ts)
# Expected failures:
ts = "host-unknowntoken"
assert not hl.Target.validate_target_string(ts)
ts = "x86-23"
assert not hl.Target.validate_target_string(ts)
# bits == 0 is allowed only if arch_unknown and os_unknown are specified,
# and no features are set
ts = "x86-0"
assert not hl.Target.validate_target_string(ts)
ts = "0-arch_unknown-os_unknown-sse41"
assert not hl.Target.validate_target_string(ts)
# "host" is only supported as the first token
ts = "opencl-host"
assert not hl.Target.validate_target_string(ts)
# set_feature
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32,
[hl.TargetFeature.SSE41])
assert t1.has_feature(hl.TargetFeature.SSE41)
assert not t1.has_feature(hl.TargetFeature.AVX)
t1.set_feature(hl.TargetFeature.AVX)
t1.set_feature(hl.TargetFeature.SSE41, False)
assert t1.has_feature(hl.TargetFeature.AVX)
assert not t1.has_feature(hl.TargetFeature.SSE41)
# set_features
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32,
[hl.TargetFeature.SSE41])
assert t1.has_feature(hl.TargetFeature.SSE41)
assert not t1.has_feature(hl.TargetFeature.AVX)
t1.set_features([hl.TargetFeature.SSE41], False)
t1.set_features([hl.TargetFeature.AVX, hl.TargetFeature.AVX2], True)
assert t1.has_feature(hl.TargetFeature.AVX)
assert t1.has_feature(hl.TargetFeature.AVX2)
assert not t1.has_feature(hl.TargetFeature.SSE41)
# with_feature
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32,
[hl.TargetFeature.SSE41])
t2 = t1.with_feature(hl.TargetFeature.NoAsserts).with_feature(
hl.TargetFeature.NoBoundsQuery)
ts = t2.to_string()
assert ts == "x86-32-linux-no_asserts-no_bounds_query-sse41"
# without_feature
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32,
[hl.TargetFeature.SSE41, hl.TargetFeature.NoAsserts])
# Note that NoBoundsQuery wasn't set here, so 'without' is a no-op
t2 = t1.without_feature(hl.TargetFeature.NoAsserts).without_feature(
hl.TargetFeature.NoBoundsQuery)
ts = t2.to_string()
assert ts == "x86-32-linux-sse41"
# natural_vector_size
# SSE4.1 is 16 bytes wide
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32,
[hl.TargetFeature.SSE41])
assert t1.natural_vector_size(hl.UInt(8)) == 16
assert t1.natural_vector_size(hl.Int(16)) == 8
assert t1.natural_vector_size(hl.UInt(32)) == 4
assert t1.natural_vector_size(hl.Float(32)) == 4
# has_gpu_feature
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32,
[hl.TargetFeature.OpenCL])
t2 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [])
assert t1.has_gpu_feature()
assert not t2.has_gpu_feature()
# has_large_buffers & maximum_buffer_size
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 64,
[hl.TargetFeature.LargeBuffers])
t2 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 64, [])
assert t1.has_large_buffers()
assert t1.maximum_buffer_size() == 9223372036854775807
assert not t2.has_large_buffers()
assert t2.maximum_buffer_size() == 2147483647
# supports_device_api
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 64,
[hl.TargetFeature.CUDA])
t2 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 64)
assert t1.supports_device_api(hl.DeviceAPI.CUDA)
assert not t2.supports_device_api(hl.DeviceAPI.CUDA)
# supports_type (deprecated version)
t1 = hl.Target(hl.TargetOS.OSX, hl.TargetArch.X86, 64,
[hl.TargetFeature.Metal])
t2 = hl.Target(hl.TargetOS.OSX, hl.TargetArch.X86, 64)
assert not t1.supports_type(hl.Float(64))
assert t2.supports_type(hl.Float(64))
# supports_type (preferred version)
t1 = hl.Target(hl.TargetOS.OSX, hl.TargetArch.X86, 64,
[hl.TargetFeature.Metal])
t2 = hl.Target(hl.TargetOS.OSX, hl.TargetArch.X86, 64)
assert not t1.supports_type(hl.Float(64), hl.DeviceAPI.Metal)
assert not t2.supports_type(hl.Float(64), hl.DeviceAPI.Metal)
# target_feature_for_device_api
assert hl.target_feature_for_device_api(
hl.DeviceAPI.OpenCL) == hl.TargetFeature.OpenCL
# with_feature with non-convertible lists
try:
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32,
["this is a string"])
except TypeError as e:
assert "incompatible constructor arguments" in str(e)
else:
assert False, 'Did not see expected exception!'
def test_target():
# Target("") should be exactly like get_host_target().
t1 = hl.get_host_target()
t2 = hl.Target("")
assert t1 == t2, "Default ctor failure"
assert t1.supported();
# to_string roundtripping
t1 = hl.Target()
ts = t1.to_string()
assert ts == "arch_unknown-0-os_unknown"
assert hl.validate_target_string(ts)
t2 = hl.Target(ts)
# equality
assert t2 == t1
# repr() and str()
assert str(t1) == "arch_unknown-0-os_unknown"
assert repr(t1) == "<halide.Target arch_unknown-0-os_unknown>"
assert t1.os == hl.TargetOS.OSUnknown
assert t1.arch == hl.TargetArch.ArchUnknown
assert t1.bits == 0
# Full specification round-trip:
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [ hl.TargetFeature.SSE41 ])
ts = t1.to_string()
assert ts == "x86-32-linux-sse41"
assert hl.validate_target_string(ts)
# Full specification (without features) round-trip:
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32)
ts = t1.to_string()
assert ts == "x86-32-linux"
assert hl.validate_target_string(ts)
# Full specification round-trip, crazy features
t1 = hl.Target(hl.TargetOS.Android, hl.TargetArch.ARM, 32,
[hl.TargetFeature.JIT, hl.TargetFeature.SSE41, hl.TargetFeature.AVX, hl.TargetFeature.AVX2,
hl.TargetFeature.CUDA, hl.TargetFeature.OpenCL, hl.TargetFeature.OpenGL, hl.TargetFeature.OpenGLCompute,
hl.TargetFeature.Debug])
ts = t1.to_string()
assert ts == "arm-32-android-avx-avx2-cuda-debug-jit-opencl-opengl-openglcompute-sse41"
assert hl.validate_target_string(ts)
# Expected failures:
ts = "host-unknowntoken"
assert not hl.validate_target_string(ts)
ts = "x86-23"
assert not hl.validate_target_string(ts)
# bits == 0 is allowed only if arch_unknown and os_unknown are specified,
# and no features are set
ts = "x86-0"
assert not hl.validate_target_string(ts)
ts = "0-arch_unknown-os_unknown-sse41"
assert not hl.validate_target_string(ts)
# "host" is only supported as the first token
ts = "opencl-host"
assert not hl.validate_target_string(ts)
# set_feature
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [hl.TargetFeature.SSE41])
assert t1.has_feature(hl.TargetFeature.SSE41)
assert not t1.has_feature(hl.TargetFeature.AVX)
t1.set_feature(hl.TargetFeature.AVX)
t1.set_feature(hl.TargetFeature.SSE41, False)
assert t1.has_feature(hl.TargetFeature.AVX)
assert not t1.has_feature(hl.TargetFeature.SSE41)
# set_features
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [hl.TargetFeature.SSE41])
assert t1.has_feature(hl.TargetFeature.SSE41)
assert not t1.has_feature(hl.TargetFeature.AVX)
t1.set_features([hl.TargetFeature.SSE41], False)
t1.set_features([hl.TargetFeature.AVX, hl.TargetFeature.AVX2], True)
assert t1.has_feature(hl.TargetFeature.AVX)
assert t1.has_feature(hl.TargetFeature.AVX2)
assert not t1.has_feature(hl.TargetFeature.SSE41)
# with_feature
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [hl.TargetFeature.SSE41])
t2 = t1.with_feature(hl.TargetFeature.NoAsserts).with_feature(hl.TargetFeature.NoBoundsQuery)
ts = t2.to_string()
assert ts == "x86-32-linux-no_asserts-no_bounds_query-sse41"
# without_feature
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [hl.TargetFeature.SSE41, hl.TargetFeature.NoAsserts])
# Note that NoBoundsQuery wasn't set here, so 'without' is a no-op
t2 = t1.without_feature(hl.TargetFeature.NoAsserts).without_feature(hl.TargetFeature.NoBoundsQuery)
ts = t2.to_string()
assert ts == "x86-32-linux-sse41"
# natural_vector_size
# SSE4.1 is 16 bytes wide
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [hl.TargetFeature.SSE41])
assert t1.natural_vector_size(hl.UInt(8)) == 16
assert t1.natural_vector_size(hl.Int(16)) == 8
assert t1.natural_vector_size(hl.UInt(32)) == 4
assert t1.natural_vector_size(hl.Float(32)) == 4
# has_gpu_feature
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [hl.TargetFeature.OpenCL])
t2 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [])
assert t1.has_gpu_feature()
assert not t2.has_gpu_feature()
# has_large_buffers & maximum_buffer_size
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 64, [hl.TargetFeature.LargeBuffers])
t2 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 64, [])
assert t1.has_large_buffers()
assert t1.maximum_buffer_size() == 9223372036854775807
assert not t2.has_large_buffers()
assert t2.maximum_buffer_size() == 2147483647
# supports_device_api
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 64, [hl.TargetFeature.CUDA])
t2 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 64)
assert t1.supports_device_api(hl.DeviceAPI.CUDA)
assert not t2.supports_device_api(hl.DeviceAPI.CUDA)
# supports_type (deprecated version)
t1 = hl.Target(hl.TargetOS.OSX, hl.TargetArch.X86, 64, [hl.TargetFeature.Metal])
t2 = hl.Target(hl.TargetOS.OSX, hl.TargetArch.X86, 64)
assert not t1.supports_type(hl.Float(64))
assert t2.supports_type(hl.Float(64))
# supports_type (preferred version)
t1 = hl.Target(hl.TargetOS.OSX, hl.TargetArch.X86, 64, [hl.TargetFeature.Metal])
t2 = hl.Target(hl.TargetOS.OSX, hl.TargetArch.X86, 64)
assert not t1.supports_type(hl.Float(64), hl.DeviceAPI.Metal)
assert not t2.supports_type(hl.Float(64), hl.DeviceAPI.Metal)
# target_feature_for_device_api
assert hl.target_feature_for_device_api(hl.DeviceAPI.OpenCL) == hl.TargetFeature.OpenCL
# with_feature with non-convertible lists
try:
t1 = hl.Target(hl.TargetOS.Linux, hl.TargetArch.X86, 32, [ "this is a string" ])
except TypeError as e:
assert str(e) == "No registered converter was able to produce a C++ rvalue of type Halide::Target::Feature from this Python object of type str"
else:
assert False, 'Did not see expected exception!'
def schedule_for_gpu(self):
# We make the decision about whether to use the GPU for each
# hl.Func independently. If you have one hl.Func computed on the
# CPU, and the next computed on the GPU, Halide will do the
# copy-to-gpu under the hood. For this pipeline, there's no
# reason to use the CPU for any of the stages. Halide will
# copy the input image to the GPU the first time we run the
# pipeline, and leave it there to reuse on subsequent runs.
# As before, we'll compute the LUT once at the start of the
# pipeline.
self.lut.compute_root()
# Let's compute the look-up-table using the GPU in 16-wide
# one-dimensional thread blocks. First we split the index
# into blocks of size 16:
block, thread = hl.Var("block"), hl.Var("thread")
self.lut.split(i, block, thread, 16)
# Then we tell cuda that our Vars 'block' and 'thread'
# correspond to CUDA's notions of blocks and threads, or
# OpenCL's notions of thread groups and threads.
self.lut.gpu_blocks(block) \
.gpu_threads(thread)
# This is a very common scheduling pattern on the GPU, so
# there's a shorthand for it:
# lut.gpu_tile(i, ii, 16)
# hl.Func::gpu_tile method is similar to hl.Func::tile, except that
# it also specifies that the tile coordinates correspond to
# GPU blocks, and the coordinates within each tile correspond
# to GPU threads.
# Compute color channels innermost. Promise that there will
# be three of them and unroll across them.
self.curved.reorder(c, x, y) \
.bound(c, 0, 3) \
.unroll(c)
# Compute curved in 2D 8x8 tiles using the GPU.
self.curved.gpu_tile(x, y, xi, yi, 8, 8)
# This is equivalent to:
# curved.tile(x, y, xo, yo, xi, yi, 8, 8)
# .gpu_blocks(xo, yo)
# .gpu_threads(xi, yi)
# We'll leave sharpen as inlined into curved.
# Compute the padded input as needed per GPU block, storing the
# intermediate result in shared memory. hl.Var::gpu_blocks, and
# hl.Var::gpu_threads exist to help you schedule producers within
# GPU threads and blocks.
self.padded.compute_at(self.curved, x)
# Use the GPU threads for the x and y coordinates of the
# padded input.
self.padded.gpu_threads(x, y)
# JIT-compile the pipeline for the GPU. CUDA or OpenCL are
# not enabled by default. We have to construct a hl.Target
# object, enable one of them, and then pass that target
# object to compile_jit. Otherwise your CPU will very slowly
# pretend it's a GPU, and use one thread per output pixel.
# Start with a target suitable for the machine you're running
# this on.
target = hl.get_host_target()
# Then enable OpenCL or CUDA.
#use_opencl = False
use_opencl = True
if use_opencl:
# We'll enable OpenCL here, because it tends to give better
# performance than CUDA, even with NVidia's drivers, because
# NVidia's open source LLVM backend doesn't seem to do all
# the same optimizations their proprietary compiler does.
target.set_feature(hl.TargetFeature.OpenCL)
print("(Using OpenCL)")
else:
# Uncomment the next line and comment out the line above to
# try CUDA instead.
target.set_feature(hl.TargetFeature.CUDA)
print("(Using CUDA)")
# If you want to see all of the OpenCL or CUDA API calls done
# by the pipeline, you can also enable the Debug
# flag. This is helpful for figuring out which stages are
# slow, or when CPU -> GPU copies happen. It hurts
# performance though, so we'll leave it commented out.
# target.set_feature(hl.TargetFeature.Debug)
self.curved.compile_jit(target)