Merge pull request #330 from ROCmSoftwarePlatform/pooling

Replace pooling with reduce_mean where possible

src/CMakeLists.txt

@@ -14,6 +14,7 @@ add_library(migraphx
     eliminate_pad.cpp
     fwd_conv_batchnorm_rewrite.cpp
     rewrite_rnn.cpp
+    rewrite_pooling.cpp
     env.cpp
     generate.cpp
     instruction.cpp

src/include/migraphx/op/reshape.hpp

@@ -59,7 +59,9 @@ struct reshape
 
         shape s{inputs.front().type(), rdims};
         if(s.elements() != inputs.front().elements())
-            MIGRAPHX_THROW("Wrong number of elements for reshape");
+            MIGRAPHX_THROW("Wrong number of elements for reshape: reshape has " +
+                           std::to_string(s.elements()) + " elements whereas the input has " +
+                           std::to_string(inputs.front().elements()));
         return s;
     }
     argument compute(shape output_shape, std::vector<argument> args) const

src/include/migraphx/rewrite_pooling.hpp

+#ifndef MIGRAPHX_GUARD_RTGLIB_REWRITE_POOLING_HPP
+#define MIGRAPHX_GUARD_RTGLIB_REWRITE_POOLING_HPP
+
+#include <string>
+#include <migraphx/config.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct program;
+
+/**
+ * Rewrite pooling to reduce_mean
+ */
+struct rewrite_pooling
+{
+    std::string name() const { return "rewrite_pooling"; }
+    void apply(program& prog) const;
+};
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+
+#endif

src/include/migraphx/tensor_view.hpp

@@ -132,7 +132,11 @@ struct tensor_view
             return m_data + this->size();
     }
 
-    std::vector<T> to_vector() const { return std::vector<T>(this->begin(), this->end()); }
+    template <class U = T>
+    std::vector<U> to_vector() const
+    {
+        return std::vector<U>(this->begin(), this->end());
+    }
 
     friend std::ostream& operator<<(std::ostream& os, const tensor_view<T>& x)
     {

src/rewrite_pooling.cpp

+#include <migraphx/rewrite_pooling.hpp>
+#include <migraphx/instruction.hpp>
+#include <migraphx/iterator_for.hpp>
+#include <migraphx/op/pooling.hpp>
+#include <migraphx/op/reshape.hpp>
+#include <migraphx/op/reduce_mean.hpp>
+#include <migraphx/program.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+void rewrite_pooling::apply(program& prog) const
+{
+    for(auto ins : iterator_for(prog))
+    {
+        if(ins->name() != "pooling")
+            continue;
+        if(ins->get_shape().lens().size() != 4)
+            continue;
+        if(ins->inputs().empty())
+            continue;
+        auto&& s  = ins->inputs().front()->get_shape();
+        auto&& op = any_cast<op::pooling>(ins->get_operator());
+        if(op.mode != "average")
+            continue;
+        if(op.padding[0] != 0 and op.padding[1] != 0)
+            continue;
+        if(op.stride[0] != 1 and op.stride[1] != 1)
+            continue;
+        if(s.lens()[2] != op.lengths[0] and s.lens()[3] != op.lengths[1])
+            continue;
+        std::int64_t n = s.lens()[0];
+        std::int64_t c = s.lens()[1];
+        auto reshape =
+            prog.insert_instruction(ins, op::reshape{{n * c, -1}}, ins->inputs().front());
+        auto pooling = prog.insert_instruction(ins, op::reduce_mean{{1}}, reshape);
+        prog.replace_instruction(ins, op::reshape{{n, c, 1, 1}}, pooling);
+    }
+}
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx

src/targets/gpu/device/include/migraphx/gpu/device/array.hpp

@@ -16,6 +16,12 @@ struct hip_array
     MIGRAPHX_DEVICE_CONSTEXPR T& operator[](std::size_t i) { return d[i]; }
     MIGRAPHX_DEVICE_CONSTEXPR const T& operator[](std::size_t i) const { return d[i]; }
 
+    MIGRAPHX_DEVICE_CONSTEXPR T& front() { return d[0]; }
+    MIGRAPHX_DEVICE_CONSTEXPR const T& front() const { return d[0]; }
+
+    MIGRAPHX_DEVICE_CONSTEXPR T& back() { return d[N - 1]; }
+    MIGRAPHX_DEVICE_CONSTEXPR const T& back() const { return d[N - 1]; }
+
     MIGRAPHX_DEVICE_CONSTEXPR T* data() { return d; }
     MIGRAPHX_DEVICE_CONSTEXPR const T* data() const { return d; }
 

src/targets/gpu/device/include/migraphx/gpu/device/reduce.hpp

@@ -209,28 +209,15 @@ constexpr std::size_t compute_block_size(std::size_t n, std::size_t max_block_si
 }
 
 template <class Op, class T, class Input, class Output>
-void reduce(hipStream_t stream,
-            const argument& result,
-            const argument& arg,
-            Op op,
-            T init,
-            Input read_input,
-            Output read_output)
+void reduce_multi_impl(hipStream_t stream,
+                       const argument& result,
+                       const argument& arg,
+                       Op op,
+                       T init,
+                       Input read_input,
+                       Output read_output,
+                       const shape& reduce_slice)
 {
-    auto&& output_shape = result.get_shape();
-    auto&& input_shape  = arg.get_shape();
-    std::vector<std::size_t> reduce_lens;
-    std::transform(output_shape.lens().begin(),
-                   output_shape.lens().end(),
-                   input_shape.lens().begin(),
-                   std::back_inserter(reduce_lens),
-                   [](auto x, auto y) -> std::size_t {
-                       if(x == y)
-                           return 1;
-                       else
-                           return y;
-                   });
-    shape reduce_slice{output_shape.type(), reduce_lens};
     hip_visit_all(result, arg, reduce_slice)([&](auto output, auto input, auto reduce_shape) {
         auto nelements = result.get_shape().elements();
         auto relements = reduce_slice.elements();
@@ -250,6 +237,83 @@ void reduce(hipStream_t stream,
     });
 }
 
+template <class Op, class T, class Input, class Output>
+void reduce_standard_impl(hipStream_t stream,
+                          const argument& result,
+                          const argument& arg,
+                          Op op,
+                          T init,
+                          Input read_input,
+                          Output read_output,
+                          std::size_t relements,
+                          std::size_t stride)
+{
+    hip_visit_all(result, arg)([&](auto output, auto input) {
+        auto nelements = result.get_shape().elements();
+
+        const std::size_t max_block_size = 256;
+        const std::size_t block_size     = compute_block_size(relements, max_block_size);
+        gs_launch(stream, nelements * block_size, block_size)([=](auto i, auto idx) __device__ {
+            const auto out_idx  = i / block_size;
+            const auto base_idx = out_idx * stride;
+            auto r = block_reduce<max_block_size>(idx, op, init, relements, [&](auto j) __device__ {
+                return read_input(input.data()[base_idx + j]);
+            });
+            if(idx.local == 0)
+                output.data()[out_idx] = read_output(r);
+        });
+    });
+}
+
+template <class Op, class T, class Input, class Output>
+void reduce(hipStream_t stream,
+            const argument& result,
+            const argument& arg,
+            Op op,
+            T init,
+            Input read_input,
+            Output read_output)
+{
+    auto&& output_shape = result.get_shape();
+    auto&& input_shape  = arg.get_shape();
+    if(input_shape.standard() and output_shape.standard() and
+       output_shape.lens().back() != input_shape.lens().back() and
+       std::equal(output_shape.lens().begin(),
+                  std::prev(output_shape.lens().end()),
+                  input_shape.lens().begin()))
+    {
+        std::size_t stride = std::accumulate(input_shape.strides().begin(),
+                                             input_shape.strides().end(),
+                                             1,
+                                             std::multiplies<size_t>());
+        reduce_standard_impl(stream,
+                             result,
+                             arg,
+                             op,
+                             init,
+                             read_input,
+                             read_output,
+                             input_shape.lens().back(),
+                             stride);
+    }
+    else
+    {
+        std::vector<std::size_t> reduce_lens;
+        std::transform(output_shape.lens().begin(),
+                       output_shape.lens().end(),
+                       input_shape.lens().begin(),
+                       std::back_inserter(reduce_lens),
+                       [](auto x, auto y) -> std::size_t {
+                           if(x == y)
+                               return 1;
+                           else
+                               return y;
+                       });
+        shape reduce_slice{output_shape.type(), reduce_lens};
+        reduce_multi_impl(stream, result, arg, op, init, read_input, read_output, reduce_slice);
+    }
+}
+
 } // namespace device
 } // namespace gpu
 } // namespace MIGRAPHX_INLINE_NS

src/targets/gpu/target.cpp

@@ -16,6 +16,7 @@
 #include <migraphx/common_subexpression_elimination.hpp>
 #include <migraphx/fwd_conv_batchnorm_rewrite.hpp>
 #include <migraphx/rewrite_rnn.hpp>
+#include <migraphx/rewrite_pooling.hpp>
 #include <migraphx/eliminate_concat.hpp>
 #include <migraphx/eliminate_identity.hpp>
 #include <migraphx/gpu/concat_gpu_opt.hpp>
@@ -46,6 +47,7 @@ std::vector<pass> target::get_passes(migraphx::context& gctx) const
         fwd_conv_batchnorm_rewrite{},
         dead_code_elimination{},
         rewrite_rnn{},
+        rewrite_pooling{},
         dead_code_elimination{},
         //common_subexpression_elimination{},
         //dead_code_elimination{},

src/tf/tf.cpp

@@ -574,23 +574,18 @@ struct tf_parser
     parse_mean(const std::string&, attribute_map attributes, std::vector<instruction_ref> args)
     {
         bool keep_dims = attributes.at("keep_dims").b();
-        std::vector<int32_t> hw_axes{2, 3};
-        // check if conditions for GlobalAvgPool are met
-        auto lens = args[0]->get_shape().lens();
-        auto axes = parse_axes(args[1]->eval().get<int32_t>().to_vector(), lens.size());
+        auto lens      = args[0]->get_shape().lens();
+        auto axes = parse_axes(args[1]->eval().get<int32_t>().to_vector<int64_t>(), lens.size());
 
-        if(axes == hw_axes and lens.size() == 4)
+        if(keep_dims)
         {
-            op::pooling op{"average"};
-            op.lengths[0] = lens[2];
-            op.lengths[1] = lens[3];
-            auto l0       = prog.add_instruction(op, args.front());
-            if(keep_dims)
-                return l0;
-            return prog.add_instruction(
-                op::squeeze{std::vector<int64_t>(hw_axes.begin(), hw_axes.end())}, l0);
+            return prog.add_instruction(op::reduce_mean{axes}, args[0]);
+        }
+        else
+        {
+            auto ins = prog.add_instruction(op::reduce_mean{axes}, args[0]);
+            return prog.add_instruction(op::squeeze{axes}, ins);
         }
-        MIGRAPHX_THROW("MIGraphX does not support mean outside of GlobalAvgPool transformation");
     }
 
     instruction_ref parse_pack(const std::string&,

test/tf/tf_test.cpp

@@ -257,8 +257,7 @@ TEST_CASE(mean_test)
     auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
     p.add_literal(l);
     p.add_literal(l);
-    migraphx::op::pooling op;
-    op.lengths = {16, 16};
+    migraphx::op::reduce_mean op{{2, 3}};
     p.add_instruction(op, l0);
     auto l3 = p.add_instruction(op, l0);
     p.add_instruction(migraphx::op::squeeze{{2, 3}}, l3);
@@ -272,9 +271,8 @@ TEST_CASE(mean_test_nhwc)
     migraphx::program p;
     migraphx::literal l{migraphx::shape{migraphx::shape::int32_type, {2}}, {1, 2}};
     auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {1, 3, 16, 16}});
-    migraphx::op::pooling op;
-    op.lengths = {16, 16};
-    auto l3    = p.add_instruction(op, l0);
+    migraphx::op::reduce_mean op{{2, 3}};
+    auto l3 = p.add_instruction(op, l0);
     p.add_instruction(migraphx::op::squeeze{{2, 3}}, l3);
     auto prog = optimize_tf("mean_test_nhwc.pb", true);