diff --git a/HalideIR b/HalideIR
index 87b089a0ba20f2e8257038ee9211d6816088ce95..e20e5e9abb3aa43147a90a4ffb3e190f62862970 160000
--- a/HalideIR
+++ b/HalideIR
@@ -1 +1 @@
-Subproject commit 87b089a0ba20f2e8257038ee9211d6816088ce95
+Subproject commit e20e5e9abb3aa43147a90a4ffb3e190f62862970
diff --git a/Jenkinsfile b/Jenkinsfile
index 4fc2285f507c9bd86ca209265d71eaf472dfe467..d74a207168292b1c1c9e8b036226bb5abe57fcb2 100644
--- a/Jenkinsfile
+++ b/Jenkinsfile
@@ -107,6 +107,7 @@ stage('Build') {
sh """
echo USE_ROCM=1 >> config.mk
echo ROCM_PATH=/opt/rocm >> config.mk
+ echo USE_VULKAN=1 >> config.mk
"""
make('gpu', '-j2')
}
diff --git a/src/codegen/llvm/intrin_rule_llvm.cc b/src/codegen/llvm/intrin_rule_llvm.cc
index 0338f007cac66ba7668bff718c88d6c5c58e067a..2e6bf061eb5e5884d4f92d87b97ee2e7c39a9af4 100644
--- a/src/codegen/llvm/intrin_rule_llvm.cc
+++ b/src/codegen/llvm/intrin_rule_llvm.cc
@@ -25,6 +25,27 @@ TVM_REGISTER_GLOBAL("tvm.intrin.rule.llvm.log")
TVM_REGISTER_GLOBAL("tvm.intrin.rule.llvm.sqrt")
.set_body(DispatchLLVMPureIntrin<::llvm::Intrinsic::sqrt, 1>);
+TVM_REGISTER_GLOBAL("tvm.intrin.rule.llvm.tanh")
+.set_body([](const TVMArgs& targs, TVMRetValue* rv) {
+ Expr e = targs[0];
+ const ir::Call* call = e.as<ir::Call>();
+ CHECK(call != nullptr);
+ const Expr& x = call->args[0];
+ Expr one = make_const(x.type(), 1);
+ Expr two = make_const(x.type(), 2);
+ Expr neg_two = make_const(x.type(), -2);
+
+ Expr exp_neg2x = ir::Call::make(
+ x.type(), "exp", {neg_two * x}, ir::Call::PureIntrinsic);
+ Expr exp_pos2x = ir::Call::make(
+ x.type(), "exp", {two * x}, ir::Call::PureIntrinsic);
+
+ Expr tanh_pos = (one - exp_neg2x) / (one + exp_neg2x);
+ Expr tanh_neg = (exp_pos2x - one) / (exp_pos2x + one);
+ *rv = ir::Select::make(
+ x >= make_zero(x.type()), tanh_pos, tanh_neg);
+});
+
TVM_REGISTER_GLOBAL("tvm.intrin.rule.llvm.pow")
.set_body(DispatchLLVMPureIntrin<::llvm::Intrinsic::pow, 1>);
diff --git a/src/codegen/spirv/ir_builder.cc b/src/codegen/spirv/ir_builder.cc
index 26f851031e5a5a68f1e86eb7a1c33509dd4c98da..500c88ae832de8e981d92ed8200288ec31c6d7d4 100644
--- a/src/codegen/spirv/ir_builder.cc
+++ b/src/codegen/spirv/ir_builder.cc
@@ -311,8 +311,8 @@ Value IRBuilder::GetConst_(const SType& dtype, const uint64_t* pvalue) {
}
CHECK_LE(dtype.type.bits(), 64);
Value ret = NewValue(dtype, kConstant);
- if (1 == dtype.type.bits() && dtype.is_uint()) {
- // Boolean types.
+ if (dtype.type == UInt(1)) {
+ // bool types.
if (*pvalue) {
ib_.Begin(spv::OpConstantTrue).AddSeq(ret);
} else {
diff --git a/tests/ci_build/Dockerfile.gpu b/tests/ci_build/Dockerfile.gpu
index b71b4cb118ec0099866489033933cd11c9e0d512..e49e498b8d406a26031a47e46f317cd8f4a6b334 100644
--- a/tests/ci_build/Dockerfile.gpu
+++ b/tests/ci_build/Dockerfile.gpu
@@ -43,6 +43,11 @@ RUN bash /install/ubuntu_install_opengl.sh
# Enable doxygen for c++ doc build
RUN apt-get install -y doxygen graphviz
+# Install vulkan
+COPY install/ubuntu_install_vulkan.sh /install/ubuntu_install_vulkan.sh
+RUN bash /install/ubuntu_install_vulkan.sh
+
+
# Environment variables
ENV PATH=/node_modules/.bin:${PATH}
ENV PATH=/usr/local/nvidia/bin:${PATH}
@@ -53,3 +58,10 @@ ENV C_INCLUDE_PATH=/usr/local/cuda/include:${C_INCLUDE_PATH}
ENV LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/nvidia/lib64:${LIBRARY_PATH}
ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/nvidia/lib64:${LD_LIBRARY_PATH}
ENV LD_LIBRARY_PATH=/opt/rocm/lib:${LD_LIBRARY_PATH}
+
+ENV PATH=/node_modules/.bin:${PATH}
+
+ENV VULKAN_SDK=/usr/local/VulkanSDK/1.0.65.0/x86_64
+ENV PATH=${PATH}:${VULKAN_SDK}/bin
+ENV LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:${VULKAN_SDK}/lib
+ENV VK_LAYER_PATH=${VULKAN_SDK}/etc/explicit_layer.d
diff --git a/tests/ci_build/install/ubuntu_install_vulkan.sh b/tests/ci_build/install/ubuntu_install_vulkan.sh
new file mode 100644
index 0000000000000000000000000000000000000000..a4155da496518af14e305ad5ddaf2f081c4ec46d
--- /dev/null
+++ b/tests/ci_build/install/ubuntu_install_vulkan.sh
@@ -0,0 +1,9 @@
+#/bin/bash
+
+wget https://sdk.lunarg.com/sdk/download/1.0.65.0/linux/vulkansdk-linux-x86_64-1.0.65.0.run
+
+bash vulkansdk-linux-x86_64-1.0.65.0.run
+mv VulkanSDK /usr/local/VulkanSDK
+cd /usr/local/VulkanSDK/1.0.65.0
+./build_tools.sh
+./build_samples.sh
diff --git a/topi/tests/python/test_topi_math.py b/topi/tests/python/test_topi_math.py
index 937639785b0b5b80fe6717f35be8ebbde0bcdc88..b1b989179f08c6314fe6b63815c0729db9177cfa 100644
--- a/topi/tests/python/test_topi_math.py
+++ b/topi/tests/python/test_topi_math.py
@@ -15,16 +15,38 @@ def test_ewise():
l = tvm.var('l')
A = tvm.placeholder((m, l), name='A')
- def test_apply(func, name):
+ shape = (20, 3)
+
+ def test_apply(func, name, f_numpy):
B = func(A)
assert tuple(B.shape) == tuple(A.shape)
assert B.op.body[0].name == name
+ a_np = np.random.uniform(size=shape).astype(A.dtype)
+ a_np = np.abs(a_np)
+ b_np = f_numpy(a_np)
+
+ def check_device(device):
+ ctx = tvm.context(device, 0)
+ if not ctx.exist:
+ print("Skip because %s is not enabled" % device)
+ return
+ print("Running on target: %s" % device)
+ with tvm.target.create(device):
+ s = topi.generic.schedule_injective(B)
+ a = tvm.nd.array(a_np, ctx)
+ b = tvm.nd.array(np.zeros_like(b_np), ctx)
+ foo = tvm.build(s, [A, B], device, name=name)
+ foo(a, b)
+ np.testing.assert_allclose(b.asnumpy(), b_np, rtol=1e-5)
+
+ for device in ['cuda', 'opencl', 'metal', 'rocm', 'vulkan', 'llvm']:
+ check_device(device)
- test_apply(topi.exp, "exp")
- test_apply(topi.tanh, "tanh")
- test_apply(topi.sigmoid, "sigmoid")
- test_apply(topi.log, "log")
- test_apply(topi.sqrt, "sqrt")
+ test_apply(topi.exp, "exp", np.exp)
+ test_apply(topi.tanh, "tanh", np.tanh)
+ test_apply(topi.sigmoid, "sigmoid", lambda x:1/(1+np.exp(-x)))
+ test_apply(topi.log, "log", np.log)
+ test_apply(topi.sqrt, "sqrt", np.sqrt)
if __name__ == "__main__":
test_util()