Merge branch 'master' of https://github.com/ROCmSoftwarePlatform/MIGraph into imagescaler

dev-requirements.txt

@@ -2,8 +2,5 @@ pfultz2/rocm-recipes
 pcre
 danmar/cppcheck@f965e5873 -DHAVE_RULES=1
 ROCm-Developer-Tools/HIP@3a41f286203968421c557338d6fb39c36f3c717c
-# Needed for clang-ocl
-RadeonOpenCompute/rocm-cmake@6240bb3 --build
-RadeonOpenCompute/clang-ocl@799713643b5591a3b877c586ef2c7fbc012af819
 # python/cpython@v3.6.6 -X autotools -H sha256:92aa914572c695c0aeb01b0a214813f414da4b51a371234df514a74761f2bb36
 -f requirements.txt

requirements.txt

 google/protobuf -DCMAKE_POSITION_INDEPENDENT_CODE=On
 RadeonOpenCompute/rocm-cmake@6240bb3 --build
-ROCmSoftwarePlatform/MIOpen,http://gitlab1.amd.com/pfultz/miopen/-/archive/f9baaabdf7c26b70648966306b2e1adb3ba3cab5/miopen-f9baaabdf7c26b70648966306b2e1adb3ba3cab5.tar.bz2
-ROCmSoftwarePlatform/rocBLAS@v14.0.1
+ROCmSoftwarePlatform/MIOpen,http://gitlab1.amd.com/pfultz/miopen/-/archive/04a9a6a6cb3a3e9457fdc207669992c08b7df5d4/miopen-04a9a6a6cb3a3e9457fdc207669992c08b7df5d4.tar.bz2
+ROCmSoftwarePlatform/rocBLAS@v14.3.0
 blaze,https://bitbucket.org/blaze-lib/blaze/get/f0755dea0e03.tar.gz -X header -DHEADER_DIR=blaze

src/include/migraph/ranges.hpp

@@ -45,6 +45,12 @@ bool contains(const C& c, const T& x)
     return generic_find(c, x) != c.end();
 }
 
+template <class T>
+bool contains(const std::initializer_list<T>& c, const T& x)
+{
+    return generic_find(c, x) != c.end();
+}
+
 template <class T, class U>
 bool contains(const std::initializer_list<T>& c, const U& x)
 {

src/program.cpp

@@ -427,7 +427,7 @@ void program::perf_report(std::ostream& os, std::size_t n, parameter_map params)
     }
 
     double total_time             = common_average(total_vec);
-    double rate                   = std::ceil(1000.0 / total_time);
+    double rate                   = 1000.0 / total_time;
     double overhead_time          = common_average(overhead_vec);
     double overhead_percent       = overhead_time * 100.0 / total_time;
     double total_instruction_time = 0.0;

src/targets/gpu/fuse_ops.cpp

@@ -134,15 +134,12 @@ MIGRAPH_PRED_MATCHER(fusable_conv, instruction_ref ins)
         return false;
     auto wei = ins->inputs().at(1)->get_shape();
     assert(wei.lens().size() == 4);
-    auto channels = wei.lens()[1] * wei.lens()[0];
-    if(wei.lens()[0] > 64 and channels > 32768)
-        return false;
     auto conv = any_cast<miopen_convolution>(ins->get_operator());
-    if(conv.algo == miopenConvolutionFwdAlgoWinograd)
+    if(wei.lens()[1] > 512 and conv.algo != miopenConvolutionFwdAlgoWinograd)
         return false;
     auto op = conv.op;
-    return op.padding == make_array<size_t>(0, 0) and op.stride == make_array<size_t>(1, 1) and
-           op.dilation == make_array<size_t>(1, 1);
+    return contains({{0, 0}, {1, 1}, {2, 2}}, op.padding) and
+           contains({{0, 0}, {1, 1}}, op.stride) and op.dilation == make_array<size_t>(1, 1);
 }
 
 struct hip_triadd