[TOPI] add nn schedulers for HLS backends (#1663)

* [TOPI] add nn schedulers for HLS backends

* fix pylint

* fix topi transform test

nnvm/python/nnvm/testing/config.py

@@ -10,5 +10,5 @@ def ctx_list():
     device_list = (device_list.split(",") if device_list
                    else ["llvm", "cuda"])
     device_list = set(device_list)
-    res = [("llvm", tvm.cpu(0)), ("cuda", tvm.gpu(0))]
-    return [x for x in res if x[1].exist and x[0] in device_list]
+    res = [(device, tvm.context(device, 0)) for device in device_list]
+    return [x for x in res if x[1].exist]

src/codegen/codegen_aocl.cc

@@ -33,6 +33,8 @@ runtime::Module BuildAOCL(Array<LoweredFunc> funcs, std::string target_str,
 
   // Compile the .cl file.
   std::string cmd = "aoc aocl.cl";
+  // AOCL supports fp64.
+  cmd += " -Dcl_khr_fp64";
   Target target = Target::create(target_str);
   if (target->device_name != "") {
     cmd += " -board=" + target->device_name;

topi/python/topi/hls/__init__.py

@@ -3,3 +3,4 @@
 from __future__ import absolute_import as _abs
 
 from .injective import schedule_injective, schedule_elemwise, schedule_broadcast
+from .nn import *

topi/tests/python/common.py

@@ -9,4 +9,4 @@ def get_all_backend():
         A list of all supported targets
     """
     return ['llvm', 'cuda', 'opencl', 'metal', 'rocm', 'vulkan', 'nvptx',
-            'llvm -device=arm_cpu']
+            'llvm -device=arm_cpu', 'aocl_sw_emu']

topi/tests/python/test_topi_pooling.py

@@ -5,6 +5,8 @@ import topi
 import math
 from topi.util import get_const_tuple
 
+from common import get_all_backend
+
 def verify_pool(n, ic, ih, kh, sh, padding, pool_type, ceil_mode, count_include_pad=True):
     iw = ih
     kw = kh
@@ -64,7 +66,7 @@ def verify_pool(n, ic, ih, kh, sh, padding, pool_type, ceil_mode, count_include_
         f(a, b)
         np.testing.assert_allclose(b.asnumpy(), b_np, rtol=1e-5)
 
-    for device in ['cuda', 'opencl', 'metal', 'rocm', 'vulkan', 'nvptx']:
+    for device in get_all_backend():
         check_device(device)
 
 def test_pool():
@@ -109,7 +111,7 @@ def verify_global_pool(n, c, h, w, pool_type):
         f(a, b)
         np.testing.assert_allclose(b.asnumpy(), b_np, rtol=1e-5)
 
-    for device in ['cuda', 'opencl', 'metal', 'rocm', 'vulkan', 'nvptx']:
+    for device in get_all_backend():
         check_device(device)
 
 def test_global_pool():

topi/tests/python/test_topi_reduce.py

@@ -4,6 +4,8 @@ import numpy as np
 import tvm
 import topi
 
+from common import get_all_backend
+
 def _my_npy_argmax(arr, axis, keepdims):
     if not keepdims:
         return arr.argmax(axis=axis)
@@ -90,7 +92,7 @@ def verify_reduce_map_ele(in_shape, axis, keepdims, type="sum", dtype="float32")
                 np.testing.assert_allclose(out_tvm_val, in_npy_map.min(axis=axis), 1E-3, 1E-3)
         else:
             np.testing.assert_allclose(out_tvm.asnumpy(), out_npy, 1E-3, 1E-3)
-    for device in ["cuda", "opencl", "metal", "llvm", "rocm", "vulkan", "nvptx"]:
+    for device in get_all_backend():
         check_device(device)
 
 

topi/tests/python/test_topi_relu.py

@@ -5,6 +5,8 @@ import tvm
 import topi
 from topi.util import get_const_tuple
 
+from common import get_all_backend
+
 def verify_relu(m, n):
     A = tvm.placeholder((m, n), name='A')
     B = topi.nn.relu(A)
@@ -27,7 +29,7 @@ def verify_relu(m, n):
         foo(a, b)
         np.testing.assert_allclose(b.asnumpy(), b_np, rtol=1e-5)
 
-    for device in ['cuda', 'opencl', 'metal', 'rocm', 'vulkan', 'nvptx', 'sdaccel']:
+    for device in get_all_backend():
         check_device(device)
 
 

topi/tests/python/test_topi_softmax.py

@@ -7,6 +7,8 @@ import topi.testing
 import logging
 from topi.util import get_const_tuple
 
+from common import get_all_backend
+
 def verify_softmax(m, n, dtype="float32"):
     A = tvm.placeholder((m, n), dtype=dtype, name='A')
     B = topi.nn.softmax(A)
@@ -63,7 +65,7 @@ def verify_log_softmax(m, n, dtype="float32"):
         foo(a, b)
         np.testing.assert_allclose(b.asnumpy(), b_np, rtol=1e-5)
 
-    for device in ["cuda", "opencl", "metal", "rocm", "vulkan", "nvptx"]:
+    for device in get_all_backend():
         check_device(device)
 
 

topi/tests/python/test_topi_transform.py

@@ -3,6 +3,8 @@ import numpy as np
 import tvm
 import topi
 
+from common import get_all_backend
+
 def verify_expand_dims(in_shape, out_shape, axis, num_newaxis):
     A = tvm.placeholder(shape=in_shape, name="A")
     B = topi.expand_dims(A, axis, num_newaxis)
@@ -22,7 +24,7 @@ def verify_expand_dims(in_shape, out_shape, axis, num_newaxis):
         foo(data_nd, out_nd)
         np.testing.assert_allclose(out_nd.asnumpy(), out_npy)
 
-    for device in ["llvm", "nvptx", "cuda", "opencl", "metal", "rocm", "vulkan", "sdaccel"]:
+    for device in get_all_backend():
         check_device(device)
 
 
@@ -45,7 +47,7 @@ def verify_tranpose(in_shape, axes):
         foo(data_nd, out_nd)
         np.testing.assert_allclose(out_nd.asnumpy(), out_npy)
 
-    for device in ["llvm", "nvptx", "cuda", "opencl", "metal", "rocm", "vulkan", "sdaccel"]:
+    for device in get_all_backend():
         check_device(device)
 
 
@@ -68,7 +70,7 @@ def verify_reshape(src_shape, dst_shape):
         foo(data_nd, out_nd)
         np.testing.assert_allclose(out_nd.asnumpy(), out_npy)
 
-    for device in ["llvm", "nvptx", "cuda", "opencl", "metal", "rocm", "vulkan", "sdaccel"]:
+    for device in get_all_backend():
         check_device(device)
 
 
@@ -96,7 +98,7 @@ def verify_squeeze(src_shape, axis):
         foo(data_nd, out_nd)
         np.testing.assert_allclose(out_nd.asnumpy(), out_npy)
 
-    for device in ["llvm", "nvptx", "cuda", "opencl", "metal", "rocm", "vulkan", "sdaccel"]:
+    for device in get_all_backend():
         check_device(device)
 
 def verify_concatenate(shapes, axis):
@@ -121,7 +123,7 @@ def verify_concatenate(shapes, axis):
         foo(*(data_nds + [out_nd]))
         np.testing.assert_allclose(out_nd.asnumpy(), out_npy)
 
-    for device in ["llvm", "nvptx", "cuda", "opencl", "metal", "rocm", "vulkan", "sdaccel"]:
+    for device in get_all_backend():
         check_device(device)
 
 
@@ -146,7 +148,7 @@ def verify_split(src_shape, indices_or_sections, axis):
         for out_nd, out_npy in zip(out_nds, out_npys):
             np.testing.assert_allclose(out_nd.asnumpy(), out_npy)
 
-    for device in ["llvm", "nvptx", "cuda", "opencl", "metal", "rocm", "vulkan", "sdaccel"]:
+    for device in get_all_backend():
         check_device(device)
 
 
@@ -204,7 +206,7 @@ def verify_flip(in_shape, axis):
         foo(data_nd, out_nd)
         np.testing.assert_allclose(out_nd.asnumpy(), out_npy)
 
-    for device in ["llvm", "cuda", "opencl", "sdaccel"]:
+    for device in ["llvm", "cuda", "opencl", "sdaccel", "aocl_sw_emu"]:
         check_device(device)
 
 def verify_take(src_shape, indices_src, axis=None):
@@ -243,7 +245,7 @@ def verify_take(src_shape, indices_src, axis=None):
         foo(data_nd, indices_nd, out_nd)
         np.testing.assert_allclose(out_nd.asnumpy(), out_npys)
 
-    for device in ["llvm", "opencl", "sdaccel"]:
+    for device in ["llvm", "opencl", "sdaccel", "aocl_sw_emu"]:
         check_device(device)
 
 def verify_strided_slice(in_shape, begin, end, stride=None):
@@ -270,7 +272,7 @@ def verify_strided_slice(in_shape, begin, end, stride=None):
         foo(data_nd, out_nd)
         np.testing.assert_allclose(out_nd.asnumpy(), out_npy)
 
-    for device in ["llvm", "opencl", "sdaccel"]:
+    for device in ["llvm", "opencl", "sdaccel", "aocl_sw_emu"]:
         check_device(device)
 
 def test_strided_slice():