[REFACTOR] topi -> tvm/topi (apache#6186)

This PR migrates the topi library as a sub namespace of tvm.

CMakeLists.txt

@@ -259,7 +259,7 @@ file(GLOB DATATYPE_SRCS src/target/datatype/*.cc)
 list(APPEND COMPILER_SRCS ${DATATYPE_SRCS})
 
 file(GLOB TOPI_SRCS
-    topi/src/*.cc
+    src/topi/*.cc
 )
 
 file(GLOB RUNTIME_SRCS
@@ -356,7 +356,7 @@ else()
   set(CMAKE_CUDA_STANDARD 14)
 endif()
 
-add_library(tvm SHARED ${COMPILER_SRCS} ${RUNTIME_SRCS})
+add_library(tvm SHARED ${COMPILER_SRCS} ${RUNTIME_SRCS} ${TOPI_SRCS})
 add_library(tvm_topi SHARED ${TOPI_SRCS})
 add_library(tvm_runtime SHARED ${RUNTIME_SRCS})
 add_library(tvm_runtime_static STATIC ${RUNTIME_SRCS})
@@ -419,15 +419,6 @@ if (HIDE_PRIVATE_SYMBOLS AND NOT ${CMAKE_SYSTEM_NAME} MATCHES "Darwin")
   target_link_libraries(tvm_runtime ${HIDE_SYMBOLS_LINKER_FLAGS})
 endif()
 
-# Related headers
-target_include_directories(
-  tvm
-  PUBLIC "topi/include")
-target_include_directories(
-  tvm_topi
-  PUBLIC "topi/include")
-
-
 # Tests
 set(TEST_EXECS "")
 file(GLOB TEST_SRCS tests/cpp/*.cc)
@@ -472,11 +463,6 @@ if (INSTALL_DEV)
     FILES_MATCHING
     PATTERN "*.h"
   )
-  install(
-    DIRECTORY "topi/include/." DESTINATION "include"
-    FILES_MATCHING
-    PATTERN "*.h"
-  )
   install(
     DIRECTORY "3rdparty/dlpack/include/." DESTINATION "include"
     FILES_MATCHING

Makefile

@@ -77,14 +77,12 @@ $(OUTPUTDIR)/libtvm_web_runtime.js: $(OUTPUTDIR)/libtvm_web_runtime.bc
 # Lint scripts
 cpplint:
 	python3 3rdparty/dmlc-core/scripts/lint.py vta cpp vta/include vta/src
-	python3 3rdparty/dmlc-core/scripts/lint.py topi cpp topi/include;
 	python3 3rdparty/dmlc-core/scripts/lint.py tvm cpp \
 	 include src \
 	 examples/extension/src examples/graph_executor/src
 
 pylint:
 	python3 -m pylint python/tvm --rcfile=$(ROOTDIR)/tests/lint/pylintrc
-	python3 -m pylint topi/python/topi --rcfile=$(ROOTDIR)/tests/lint/pylintrc
 	python3 -m pylint vta/python/vta --rcfile=$(ROOTDIR)/tests/lint/pylintrc
 
 jnilint:

apps/android_camera/app/src/main/jni/Android.mk

@@ -39,8 +39,6 @@ LOCAL_LDFLAGS := -L$(SYSROOT)/usr/lib/ -llog
 LOCAL_C_INCLUDES := $(ROOT_PATH)/include \
                     $(ROOT_PATH)/3rdparty/dlpack/include \
                     $(ROOT_PATH)/3rdparty/dmlc-core/include \
-                    $(ROOT_PATH)/3rdparty/HalideIR/src \
-                    $(ROOT_PATH)/topi/include
 
 LOCAL_MODULE = tvm4j_runtime_packed
 

apps/android_deploy/app/src/main/jni/Android.mk

@@ -38,8 +38,7 @@ LOCAL_LDFLAGS := -L$(SYSROOT)/usr/lib/ -llog
 
 LOCAL_C_INCLUDES := $(ROOT_PATH)/include \
                     $(ROOT_PATH)/3rdparty/dlpack/include \
-                    $(ROOT_PATH)/3rdparty/dmlc-core/include \
-                    $(ROOT_PATH)/topi/include
+                    $(ROOT_PATH)/3rdparty/dmlc-core/include
 
 LOCAL_MODULE = tvm4j_runtime_packed
 

apps/android_rpc/app/src/main/jni/Android.mk

@@ -38,8 +38,7 @@ LOCAL_LDFLAGS := -L$(SYSROOT)/usr/lib/ -llog
 
 LOCAL_C_INCLUDES := $(ROOT_PATH)/include \
                     $(ROOT_PATH)/3rdparty/dlpack/include \
-                    $(ROOT_PATH)/3rdparty/dmlc-core/include \
-                    $(ROOT_PATH)/topi/include
+                    $(ROOT_PATH)/3rdparty/dmlc-core/include
 
 LOCAL_MODULE = tvm4j_runtime_packed
 

apps/sgx/build.rs

@@ -35,9 +35,7 @@ fn main() {
             concat!(
                 mf_dir!("/../../python"),
                 ":",
-                mf_dir!("/../../nnvm/python"),
-                ":",
-                mf_dir!("/../../topi/python")
+                mf_dir!("/../../nnvm/python")
             ),
         )
         .output()

topi/README.md → apps/topi_recipe/README.md

@@ -15,7 +15,7 @@
 <!--- specific language governing permissions and limitations -->
 <!--- under the License. -->
 
-# TOPI: TVM Operator Inventory
+# TOPI Recipe: TVM Operator Optimization Recipes
 
 TOPI is the operator collection library for TVM intended at sharing the effort of crafting
 and optimizing tvm generated kernels. The goal:
@@ -24,11 +24,6 @@ and optimizing tvm generated kernels. The goal:
 - Give common primitives for fused op creation.
 - Provide commonly used schedules under each architectures
 
-## Organization
-- [include](include) C++ library, header only
-- [python](python) python library
-- [recipe](recipe) Recipe collections containing useful operator examples.
-
 ## Guidelines
 - Use numpy-style naming convention for known ops
 - Seperate operator declaration from schedule when possible.
@@ -39,10 +34,6 @@ and optimizing tvm generated kernels. The goal:
 - Data layout aware, if not specified in argument or in function, assume NCHW by default.
 
 
-## Testcase
-- Add testcases to testout the schedule and dataflow in the TOPI workflow
-- Only do correctness testing without attaching compiler flags and only run it once.
-
 ## Performance Tuning Workflow
 Since TVM is work in progress, some optimization might not be perfect.
 One quick way I find useful is to do codegen plus manual modification.

topi/recipe/broadcast/test_broadcast_map.py → apps/topi_recipe/broadcast/test_broadcast_map.py

@@ -20,7 +20,7 @@
 from tvm.contrib import nvcc
 import numpy as np
 
-import topi
+from tvm import topi
 
 
 TASK = "reduce_map"

topi/recipe/conv/depthwise_conv2d_test.py → apps/topi_recipe/conv/depthwise_conv2d_test.py

@@ -21,9 +21,9 @@
 from scipy import signal
 from tvm.contrib import nvcc
 
-import topi
-from topi.util import get_const_tuple
-from topi.cuda.depthwise_conv2d import schedule_depthwise_conv2d_nchw, schedule_depthwise_conv2d_nhwc
+from tvm import topi
+from tvm.topi.util import get_const_tuple
+from tvm.topi.cuda.depthwise_conv2d import schedule_depthwise_conv2d_nchw, schedule_depthwise_conv2d_nhwc
 
 TASK = "depthwise_conv2d"
 USE_MANUAL_CODE = False
@@ -118,7 +118,7 @@ def check_device(device):
         print("average time cost of 1000 runs (depthwise_conv2d + scale_shift) = %g us" % (tcost_2*1e6))
         print("average time cost of 1000 runs (depthwise_conv2d + scale_shift + relu) = %g us" % (tcost_3*1e6))
         # correctness
-        depthwise_conv2d_scipy = topi.testing.depthwise_conv2d_python_nchw(input_np, filter_np, stride=[stride_h, stride_w], padding=padding)
+        depthwise_conv2d_scipy = tvm.topi.testing.depthwise_conv2d_python_nchw(input_np, filter_np, stride=[stride_h, stride_w], padding=padding)
         scale_shift_scipy = np.zeros(shape=get_const_tuple(ScaleShift.shape))
         for c in range(in_channel * channel_multiplier):
             scale_shift_scipy[:,c,:,:] = depthwise_conv2d_scipy[:,c,:,:] * scale_np[c] + shift_np[c]
@@ -207,7 +207,7 @@ def check_device(device):
         print("average time cost of 1000 runs (depthwise_conv2d + scale_shift) = %g us" % (tcost_2*1e6))
         print("average time cost of 1000 runs (depthwise_conv2d + scale_shift + relu) = %g us" % (tcost_3*1e6))
         # correctness
-        depthwise_conv2d_scipy = topi.testing.depthwise_conv2d_python_nhwc(input_np, filter_np, stride=[stride_h, stride_w], padding=padding)
+        depthwise_conv2d_scipy = tvm.topi.testing.depthwise_conv2d_python_nhwc(input_np, filter_np, stride=[stride_h, stride_w], padding=padding)
         scale_shift_scipy = np.zeros(shape=get_const_tuple(ScaleShift.shape))
         for c in range(in_channel * channel_multiplier):
             scale_shift_scipy[:,:,:,c] = depthwise_conv2d_scipy[:,:,:,c] * scale_np[c] + shift_np[c]

topi/recipe/conv/test_conv2d_hwcn_map.py → apps/topi_recipe/conv/test_conv2d_hwcn_map.py

@@ -21,8 +21,8 @@
 import tvm
 from tvm import te
 from tvm.contrib import nvcc
-import topi
-from topi.util import get_const_tuple
+from tvm import topi
+from tvm.topi.util import get_const_tuple
 
 TASK = "conv2d_hwcn_map"
 USE_MANUAL_CODE = False
@@ -65,7 +65,7 @@ def test_conv2d_hwcn_map():
 
     a_np = np.random.uniform(size=get_const_tuple(A.shape)).astype(A.dtype)
     w_np = np.random.uniform(size=get_const_tuple(W.shape)).astype(W.dtype)
-    b_np = topi.testing.conv2d_hwcn_python(a_np, w_np, stride, padding)
+    b_np = tvm.topi.testing.conv2d_hwcn_python(a_np, w_np, stride, padding)
     c_np = np.maximum(b_np, 0)
 
     def check_device(device):

topi/recipe/conv/test_conv_int8_arm.py → apps/topi_recipe/conv/test_conv_int8_arm.py

@@ -21,7 +21,7 @@
 import numpy as np
 import tvm
 from tvm import te
-import topi
+from tvm import topi
 
 logging.basicConfig(stream=sys.stdout, level=logging.INFO)
 LOGGER = logging.getLogger('test_conv_int8_intel')

topi/recipe/conv/test_conv_int8_intel.py → apps/topi_recipe/conv/test_conv_int8_intel.py

@@ -21,7 +21,7 @@
 import numpy as np
 import tvm
 from tvm import te
-import topi
+from tvm import topi
 
 logging.basicConfig(stream=sys.stdout, level=logging.INFO)
 LOGGER = logging.getLogger('test_conv_int8_intel')

topi/recipe/gemm/gemm_int8.py → apps/topi_recipe/gemm/gemm_int8.py

@@ -21,7 +21,7 @@
 import tvm
 from tvm import te
 from tvm import autotvm
-from topi.cuda.tensor_intrin import dp4a
+from tvm.topi.cuda.tensor_intrin import dp4a
 
 DO_TUNING = True
 PRETUNED_INDEX = 75333

topi/recipe/reduce/test_reduce_map.py → apps/topi_recipe/reduce/test_reduce_map.py

@@ -20,7 +20,7 @@
 from tvm.contrib import nvcc
 import numpy as np
 
-import topi
+from tvm import topi
 
 
 TASK = "reduce_map"

conda/tvm/build.sh

@@ -22,7 +22,3 @@ set -u
 cd python
 $PYTHON setup.py install --single-version-externally-managed --record=/tmp/record.txt
 cd ..
-
-cd topi/python
-$PYTHON setup.py install --single-version-externally-managed --record=/tmp/record.txt
-cd ../..

docker/Dockerfile.demo_android

@@ -70,5 +70,5 @@ RUN cd /usr && \
     make -j10
 
 # Environment variables
-ENV PYTHONPATH=/usr/tvm/python:/usr/tvm/topi/python:/usr/tvm/vta/python:${PYTHONPATH}
+ENV PYTHONPATH=/usr/tvm/python:/usr/tvm/vta/python:${PYTHONPATH}
 ENV ANDROID_HOME=/opt/android-sdk-linux/

docker/Dockerfile.demo_cpu

@@ -30,4 +30,4 @@ COPY install/install_tvm_cpu.sh /install/install_tvm_cpu.sh
 RUN bash /install/install_tvm_cpu.sh
 
 # Environment variables
-ENV PYTHONPATH=/usr/tvm/python:/usr/tvm/topi/python:/usr/tvm/vta/python:${PYTHONPATH}
+ENV PYTHONPATH=/usr/tvm/python:/usr/tvm/vta/python:${PYTHONPATH}

docker/Dockerfile.demo_gpu

@@ -28,7 +28,7 @@ COPY install/install_tvm_gpu.sh /install/install_tvm_gpu.sh
 RUN bash /install/install_tvm_gpu.sh
 
 # Environment variables
-ENV PYTHONPATH=/usr/tvm/python:/usr/tvm/topi/python:/usr/tvm/vta/python:${PYTHONPATH}
+ENV PYTHONPATH=/usr/tvm/python:/usr/tvm/vta/python:${PYTHONPATH}
 ENV PATH=/usr/local/nvidia/bin:${PATH}
 ENV PATH=/usr/local/cuda/bin:${PATH}
 ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/nvidia/lib64:${LD_LIBRARY_PATH}

docker/Dockerfile.demo_opencl

@@ -76,6 +76,5 @@ RUN mkdir -p ${TVM_BUILD_DIR} && \
 	make -j6
 
 RUN echo "Building Python package"
-ENV PYTHONPATH=${TVM_HOME}/python:${TVM_HOME}/topi/python:${PYTHONPATH}
+ENV PYTHONPATH=${TVM_HOME}/python:${PYTHONPATH}
 RUN cd ${TVM_HOME}/python && python3 setup.py install --user
-RUN cd ${TVM_HOME}/topi/python && python3 setup.py install --user

docker/bash.sh

@@ -70,7 +70,7 @@ else
 fi
 
 if [[ "${DOCKER_IMAGE_NAME}" == *"ci"* ]]; then
-    CI_PY_ENV="-e PYTHONPATH=/workspace/python:/workspace/topi/python"
+    CI_PY_ENV="-e PYTHONPATH=/workspace/python"
 else
     CI_PY_ENV=""
 fi

docs/Doxyfile

@@ -770,7 +770,7 @@ WARN_LOGFILE           =
 # spaces.
 # Note: If this tag is empty the current directory is searched.
 
-INPUT                  = include/tvm topi/include/topi
+INPUT                  = include/tvm
 
 # This tag can be used to specify the character encoding of the source files
 # that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses

docs/api/python/index.rst

@@ -44,5 +44,5 @@ Python API
    micro
    contrib
    graph_runtime
-   vta/index
    topi
+   vta/index