Merge branch 'add-nonstandard'

src/CMakeLists.txt

@@ -3,6 +3,7 @@ add_library(migraph
     auto_contiguous.cpp
     dead_code_elimination.cpp
     eliminate_contiguous.cpp
+    fwd_conv_batchnorm_rewrite.cpp
     env.cpp
     generate.cpp
     program.cpp

src/fwd_conv_batchnorm_rewrite.cpp

+#include <migraph/fwd_conv_batchnorm_rewrite.hpp>
+#include <migraph/program.hpp>
+#include <migraph/instruction.hpp>
+#include <migraph/operators.hpp>
+#include <migraph/iterator_for.hpp>
+#include <migraph/dfor.hpp>
+
+namespace migraph {
+void fwd_conv_batchnorm_rewrite::apply(program& p) const
+{
+    for(auto ins : iterator_for(p))
+    {
+        if(ins->op.name() != "batch_norm_inference")
+            continue;
+        if(not std::all_of(ins->arguments.begin() + 1, ins->arguments.end(), [](auto arg) {
+               return arg->op.name() == "@literal";
+           }))
+            continue;
+
+        auto conv_ins = ins->arguments[0];
+        if(conv_ins->op.name() != "convolution")
+            continue;
+        if(conv_ins->arguments[1]->op.name() != "@literal")
+            continue;
+        // Get scale, bias, mean, variance from instruction_ref
+        const auto& gamma    = ins->arguments[1]->get_literal();
+        const auto& bias     = ins->arguments[2]->get_literal();
+        const auto& mean     = ins->arguments[3]->get_literal();
+        const auto& variance = ins->arguments[4]->get_literal();
+        // Get epsilon
+        auto bn_op   = any_cast<batch_norm_inference>(ins->op);
+        auto epsilon = bn_op.epsilon;
+        // Get convolution weights
+        const auto& weights = conv_ins->arguments[1]->get_literal();
+        // Get convolution op
+        auto conv_op      = conv_ins->op;
+        auto weights_lens = weights.get_shape().lens();
+        auto conv_lens    = conv_ins->get_shape().lens();
+        argument new_weights{weights.get_shape()};
+        argument new_bias{bias.get_shape()};
+        visit_all(weights, gamma, bias, mean, variance, new_weights, new_bias)(
+            [&](auto weights2,
+                auto gamma2,
+                auto bias2,
+                auto mean2,
+                auto variance2,
+                auto new_weights2,
+                auto new_bias2) {
+                dfor(weights_lens[0], weights_lens[1], weights_lens[2], weights_lens[3])(
+                    [&](std::size_t k, std::size_t c, std::size_t h, std::size_t w) {
+                        new_weights2(k, c, h, w) =
+                            gamma2(k) / std::sqrt(variance2(k) + epsilon) * weights2(k, c, h, w);
+                    });
+                dfor(new_bias.get_shape().elements())([&](std::size_t c) {
+                    new_bias2(c) = bias2(c) - (mean2(c) / std::sqrt(variance2(c) + epsilon));
+                });
+            });
+        // Replace convolution instruction with updated weights
+        auto l_weights = p.add_literal({weights.get_shape(), new_weights.data()});
+        auto l_bias    = p.add_literal({new_bias.get_shape(), new_bias.data()});
+        auto c = p.replace_instruction(conv_ins, conv_op, {conv_ins->arguments[0], l_weights});
+        auto b = p.insert_instruction(ins, broadcast{1}, c, l_bias);
+        p.replace_instruction(ins, add{}, {c, b});
+    }
+}
+} // namespace migraph

src/include/migraph/functional.hpp

@@ -5,6 +5,14 @@
 
 namespace migraph {
 
+struct swallow
+{
+    template <class... Ts>
+    constexpr swallow(Ts&&...)
+    {
+    }
+};
+
 namespace detail {
 
 template <class R, class F>
@@ -19,8 +27,48 @@ struct fix_f
     }
 };
 
+template <std::size_t...>
+struct seq
+{
+    using type = seq;
+};
+
+template <class, class>
+struct merge_seq;
+
+template <std::size_t... Xs, std::size_t... Ys>
+struct merge_seq<seq<Xs...>, seq<Ys...>> : seq<Xs..., (sizeof...(Xs) + Ys)...>
+{
+};
+
+template <std::size_t N>
+struct gens : merge_seq<typename gens<N / 2>::type, typename gens<N - N / 2>::type>
+{
+};
+
+template <>
+struct gens<0> : seq<>
+{
+};
+template <>
+struct gens<1> : seq<0>
+{
+};
+
+template <class F, std::size_t... Ns>
+constexpr void repeat_c_impl(F f, seq<Ns...>)
+{
+    swallow{(f(std::integral_constant<std::size_t, Ns>{}), 0)...};
+}
+
 } // namespace detail
 
+template <std::size_t N, class F>
+constexpr void repeat_c(F f)
+{
+    detail::repeat_c_impl(f, detail::gens<N>{});
+}
+
 /// Implements a fix-point combinator
 template <class R, class F>
 detail::fix_f<R, F> fix(F f)
@@ -35,7 +83,7 @@ auto fix(F f)
 }
 
 template <class... Ts>
-auto make_sequence(Ts... xs)
+auto pack(Ts... xs)
 {
     return [=](auto f) { return f(xs...); };
 }

src/include/migraph/fwd_conv_batchnorm_rewrite.hpp

+#ifndef MIGRAPH_GUARD_RTGLIB_FWD_CONV_BATCHNORM_REWRITE_HPP
+#define MIGRAPH_GUARD_RTGLIB_FWD_CONV_BATCHNORM_REWRITE_HPP
+
+#include <string>
+#include <migraph/instruction_ref.hpp>
+
+namespace migraph {
+
+struct program;
+
+struct fwd_conv_batchnorm_rewrite
+{
+    std::string name() const { return "fwd_conv_batchnorm_rewrite"; }
+    void apply(program& p) const;
+};
+
+} // namespace migraph
+
+#endif

src/include/migraph/generate.hpp

@@ -12,7 +12,11 @@ constexpr T normalize(unsigned long z)
 {
     if(z == 0)
         return 0;
-    return (2.0 / z) - 1.0;
+    const auto max     = 32768;
+    const double range = max / 2; // NOLINT
+    double result      = (z % max) / range;
+    result -= 1;
+    return result;
 }
 
 template <class T, MIGRAPH_REQUIRES(std::is_signed<T>{} and not std::is_floating_point<T>{})>
@@ -54,11 +58,29 @@ struct xorshf96_generator
     }
 };
 
+template <class T>
+struct xorshift_generator
+{
+    unsigned long x;
+
+    xorshift_generator(unsigned long seed = 0) : x(521288629ULL ^ seed) {}
+
+    constexpr T operator()() noexcept
+    {
+        x ^= x >> 12U;
+        x ^= x << 25U;
+        x ^= x >> 27U;
+        return normalize<T>(x * 0x2545F4914F6CDD1D);
+    }
+};
+
 template <class T>
 std::vector<T> generate_tensor_data(const migraph::shape& s, unsigned long seed = 0)
 {
     std::vector<T> result(s.elements());
     std::generate(result.begin(), result.end(), xorshf96_generator<T>{seed});
+    // std::generate(result.begin(), result.end(), [&]{ return seed % 7; });
+    // std::generate(result.begin(), result.end(), []{ return 1; });
     return result;
 }
 

src/include/migraph/instruction.hpp

@@ -115,6 +115,11 @@ struct instruction
     }
 
     shape get_shape() const { return result; }
+    const literal& get_literal() const
+    {
+        assert(op.name() == "@literal");
+        return lit;
+    }
 
     friend bool operator==(instruction_ref ref, const instruction& i) { return i == ref; }
 

src/include/migraph/tracer.hpp

@@ -2,17 +2,10 @@
 #define MIGRAPH_GUARD_RTGLIB_TRACER_HPP
 
 #include <ostream>
+#include <migraph/functional.hpp>
 
 namespace migraph {
 
-struct swallow
-{
-    template <class... Ts>
-    swallow(Ts&&...)
-    {
-    }
-};
-
 struct tracer
 {
     tracer() {}

src/targets/gpu/CMakeLists.txt

@@ -11,6 +11,7 @@ if(NOT TARGET MIOpen)
 endif()
 
 add_library(migraph_device 
+    device/add.cpp
     device/add_relu.cpp
     device/contiguous.cpp
 )

src/targets/gpu/device/add.cpp

+#include <migraph/gpu/device/add.hpp>
+#include <migraph/gpu/device/nary.hpp>
+
+namespace migraph {
+namespace gpu {
+namespace device {
+
+void add(const argument& result, const argument& arg1, const argument& arg2)
+{
+    nary(result, arg1, arg2)([](auto x, auto y) { return x + y; });
+}
+
+} // namespace device
+} // namespace gpu
+} // namespace migraph

src/targets/gpu/device/add_relu.cpp

@@ -5,10 +5,9 @@ namespace migraph {
 namespace gpu {
 namespace device {
 
-void add_relu(argument result, argument arg1, argument arg2)
+void add_relu(const argument& result, const argument& arg1, const argument& arg2)
 {
-    nary_standard(std::move(result), std::move(arg1), std::move(arg2))(
-        [](auto x, auto y) { return max(0, x + y); });
+    nary(result, arg1, arg2)([](auto x, auto y) { return std::max<decltype(x + y)>(0, x + y); });
 }
 
 } // namespace device

src/targets/gpu/device/include/migraph/gpu/device/launch.hpp

@@ -33,10 +33,10 @@ inline auto launch(std::size_t global, std::size_t local)
     };
 }
 
-inline auto gs_launch(std::size_t n, std::size_t local = 512)
+inline auto gs_launch(std::size_t n, std::size_t local = 1024)
 {
     std::size_t groups  = 1 + n / local;
-    std::size_t nglobal = std::min<std::size_t>(512, groups) * local;
+    std::size_t nglobal = std::min<std::size_t>(256, groups) * local;
 
     return [=](auto f) {
         launch(nglobal, local)([=](auto idx) {
@@ -48,6 +48,14 @@ inline auto gs_launch(std::size_t n, std::size_t local = 512)
     };
 }
 
+// Workaround hcc's broken tile_static macro
+#ifdef tile_static
+#undef tile_static
+#define MIGRAPH_DEVICE_SHARED __attribute__((tile_static))
+#else
+#define MIGRAPH_DEVICE_SHARED __shared__
+#endif
+
 } // namespace device
 } // namespace gpu
 } // namespace migraph

src/targets/gpu/device/include/migraph/gpu/device/nary.hpp

@@ -10,16 +10,25 @@ namespace migraph {
 namespace gpu {
 namespace device {
 
-template <class... Arguments>
-auto nary(argument result, Arguments... args)
+template <class T>
+using vec4 = T __attribute__((ext_vector_type(4)));
+
+template <class T>
+__device__ __host__ vec4<T>* as_vec4(T* x)
 {
-    return [=](auto f) {
-        if(all_of({args...}, [](const shape& s) { return s.standard(); }))
-            nary_standard(result, args...)(f);
-        else
-            nary_nonstandard(result, args...)(f);
+    return reinterpret_cast<vec4<T>*>(x);
+}
 
-    };
+template <class T>
+__device__ __host__ T* as_pointer(vec4<T>* x)
+{
+    return reinterpret_cast<T*>(x);
+}
+
+template <class... Ts>
+auto pack_vec4(Ts... xs)
+{
+    return [=](auto f, std::size_t n) { return f(as_vec4(xs)[n]...); };
 }
 
 template <class F, class... Arguments>
@@ -28,14 +37,12 @@ auto nary_nonstandard_impl(F f, argument result, Arguments... args)
     const auto& output_shape = result.get_shape();
     visit_all(result, args...)([&](auto output, auto... inputs) {
         visit_tensor_size(output_shape.lens().size(), [&](auto ndim) {
-            auto data = make_sequence(
-                std::make_pair(hip_tensor_descriptor<ndim>{inputs.get_shape().lens(),
-                                                           inputs.get_shape().strides()},
-                               inputs.data())...);
-            hip_tensor_descriptor<ndim> out_desc(output_shape.lens(), output_shape.strides());
+            auto data = pack(
+                std::make_pair(hip_tensor_descriptor<ndim>{inputs.get_shape()}, inputs.data())...);
+            hip_tensor_descriptor<ndim> out_desc(output_shape);
             auto* outp = output.data();
             gs_launch(output_shape.elements())([=](auto i) {
-                data([&](auto... ps) {
+                data([&](auto&&... ps) {
                     auto outidx = out_desc.multi(i);
                     outp[i]     = f(ps.second[ps.first.linear(outidx)]...);
                 });
@@ -44,24 +51,199 @@ auto nary_nonstandard_impl(F f, argument result, Arguments... args)
     });
 }
 
+template <class F>
+void binary_broadcast_vec_impl(F f,
+                               const argument& result,
+                               const argument& arg1,
+                               const argument& arg2)
+{
+    const auto& output_shape = result.get_shape();
+    const auto& b_shape      = arg2.get_shape();
+    auto bdim =
+        std::distance(b_shape.strides().begin(),
+                      std::find_if(b_shape.strides().begin(), b_shape.strides().end(), [](auto x) {
+                          return x != 0;
+                      }));
+    auto bdim_len         = output_shape.lens()[bdim];
+    auto bdim_stride      = output_shape.strides()[bdim];
+    auto bdim_next_stride = bdim_stride * bdim_len;
+
+    visit_all(result, arg1, arg2)([&](auto output, auto input1, auto input2) {
+        using type = std::remove_cv_t<typename decltype(output)::value_type>;
+        auto* xp   = as_vec4(input1.data());
+        auto* yp   = as_vec4(input2.data());
+        auto* outp = as_vec4(output.data());
+
+        const std::size_t vec_size     = 4;
+        const std::size_t nlocal       = 1024;
+        const std::size_t nglobal      = 256 * nlocal;
+        const std::size_t n            = output.size() / vec_size;
+        const std::size_t bdim_vec_len = bdim_len / vec_size;
+
+        launch(nglobal, nlocal)([=](auto idx) __device__ {
+            MIGRAPH_DEVICE_SHARED vec4<type> buffer[2048 / vec_size];
+            // Load bias into LDS
+            for(size_t i = idx.local; i < bdim_vec_len; i += nlocal)
+            {
+                buffer[i] = yp[i];
+            }
+            __syncthreads();
+            auto* bp = as_pointer(buffer);
+            // Process the data
+            for(size_t i = idx.global; i < n; i += nglobal)
+            {
+                auto bidx      = ((i * vec_size) % bdim_next_stride) / bdim_stride;
+                auto b         = bp[bidx];
+                vec4<type> x   = xp[i];
+                vec4<type> out = outp[i];
+                for(std::size_t j = 0; j < vec_size; j++)
+                {
+                    out[j] = f(x[j], b);
+                }
+                outp[i] = out;
+            }
+        });
+    });
+}
+
+template <class F>
+void binary_broadcast_impl(F f, const argument& result, const argument& arg1, const argument& arg2)
+{
+    const auto& output_shape = result.get_shape();
+    const auto& b_shape      = arg2.get_shape();
+    auto bdim =
+        std::distance(b_shape.strides().begin(),
+                      std::find_if(b_shape.strides().begin(), b_shape.strides().end(), [](auto x) {
+                          return x != 0;
+                      }));
+    auto bdim_len         = output_shape.lens()[bdim];
+    auto bdim_stride      = output_shape.strides()[bdim];
+    auto bdim_next_stride = bdim_stride * bdim_len;
+
+    visit_all(result, arg1, arg2)([&](auto output, auto input1, auto input2) {
+        using type = std::remove_cv_t<typename decltype(output)::value_type>;
+        auto* xp   = input1.data();
+        auto* yp   = input2.data();
+        auto* outp = output.data();
+
+        const std::size_t nlocal  = 1024;
+        const std::size_t nglobal = 256 * nlocal;
+        const std::size_t n       = output.size();
+
+        launch(nglobal, nlocal)([=](auto idx) __device__ {
+            MIGRAPH_DEVICE_SHARED type buffer[2048];
+            // Load bias into LDS
+            for(size_t i = idx.local; i < bdim_len; i += nlocal)
+            {
+                buffer[i] = yp[i];
+            }
+            __syncthreads();
+            // Process the data
+            for(size_t i = idx.global; i < n; i += nglobal)
+            {
+                auto bidx = (i % bdim_next_stride) / bdim_stride;
+                auto b    = buffer[bidx];
+                type x    = xp[i];
+                outp[i]   = f(x, b);
+            }
+        });
+    });
+}
+
+template <class F, class... Arguments>
+void nary_standard_vec_impl(F f, argument result, Arguments... args)
+{
+    // assert(x.get_shape().elements() == y.get_shape().elements());
+    const auto& output_shape = result.get_shape();
+    visit_all(result, args...)([&](auto output, auto... inputs) {
+        using type                 = std::remove_cv_t<typename decltype(output)::value_type>;
+        const std::size_t vec_size = 4;
+        auto data                  = pack_vec4(inputs.data()...);
+        auto* outp                 = as_vec4(output.data());
+        gs_launch(output_shape.elements() / vec_size)([=](auto i) {
+            vec4<type> out = outp[i];
+            data(
+                [&](auto... xs) {
+                    for(std::size_t j = 0; j < vec_size; j++)
+                    {
+                        out[j] = f(xs[j]...);
+                    }
+                },
+                i);
+            outp[i] = out;
+        });
+    });
+}
+
+template <class F, class... Arguments>
+void nary_standard_impl(F f, argument result, Arguments... args)
+{
+    // assert(x.get_shape().elements() == y.get_shape().elements());
+    const auto& output_shape = result.get_shape();
+    visit_all(result, args...)([&](auto output, auto... inputs) {
+        auto data  = pack(inputs.data()...);
+        auto* outp = output.data();
+        gs_launch(output_shape.elements())(
+            [=](auto i) { data([&](auto... xps) { outp[i] = f(xps[i]...); }); });
+    });
+}
+
+template <class F, class... Arguments>
+void nary_impl(F f, argument result, Arguments... args)
+{
+    bool standard = all_of({args.get_shape()...}, [](const shape& s) { return s.standard(); });
+    bool packed   = all_of({args.get_shape()...}, [](const shape& s) { return s.packed(); });
+    bool same_shapes =
+        all_of({args.get_shape()...}, [&](const shape& s) { return s == result.get_shape(); });
+    if(standard or (packed and same_shapes))
+        nary_standard_impl(f, result, args...);
+    else
+        nary_nonstandard_impl(f, result, args...);
+}
+
 template <class... Arguments>
 auto nary_nonstandard(argument result, Arguments... args)
 {
-    return [=](auto f) { return nary_nonstandard_impl(f, result, args...); };
+    return [=](auto f) { nary_nonstandard_impl(f, result, args...); };
 }
 
 template <class... Arguments>
 auto nary_standard(argument result, Arguments... args)
+{
+    return [=](auto f) { nary_standard_impl(f, result, args...); };
+}
+
+template <class... Arguments>
+auto nary(argument result, Arguments... args)
+{
+    return [=](auto f) { nary_impl(f, result, args...); };
+}
+
+inline auto nary(const argument& result, const argument& arg1, const argument& arg2)
 {
     return [=](auto f) {
-        // assert(x.get_shape().elements() == y.get_shape().elements());
-        const auto& output_shape = result.get_shape();
-        visit_all(result, args...)([&](auto output, auto... inputs) {
-            auto data  = make_sequence(inputs.data()...);
-            auto* outp = output.data();
-            gs_launch(output_shape.elements())(
-                [=](auto i) { data([&](auto... xps) { outp[i] = f(xps[i]...); }); });
-        });
+        // TODO: Check result and arg1 shape is the same
+        if(arg1.get_shape().standard() and arg2.get_shape().broadcasted())
+        {
+            auto not_zero       = [](auto x) { return x != 0; };
+            const auto& strides = arg2.get_shape().strides();
+            auto b_it           = std::find_if(strides.begin(), strides.end(), not_zero);
+            auto b_idx          = std::distance(strides.begin(), b_it);
+            auto b_len          = result.get_shape().lens()[b_idx];
+            auto b_stride       = result.get_shape().strides()[b_idx];
+            assert(arg2.get_shape().lens()[b_idx] == b_len);
+            if(b_len <= 2048 and std::none_of(std::next(b_it), strides.end(), not_zero))
+            {
+                const bool divisible_by_4 = (b_len % 4 == 0) and (b_stride % 4 == 0) and
+                                            (arg1.get_shape().elements() % 4 == 0);
+                if(divisible_by_4)
+                    binary_broadcast_vec_impl(f, result, arg1, arg2);
+                else
+                    binary_broadcast_impl(f, result, arg1, arg2);
+                return;
+            }
+        }
+        nary_impl(f, result, arg1, arg2);
     };
 }
 

src/targets/gpu/device/include/migraph/gpu/device/tensor.hpp

@@ -2,6 +2,7 @@
 #define MIGRAPH_GUARD_RTGLIB_DEAVICE_TENSOR_HPP
 
 #include <hip/hip_runtime.h>
+#include <migraph/functional.hpp>
 
 namespace migraph {
 namespace gpu {
@@ -53,14 +54,13 @@ template <size_t NDim>
 struct hip_tensor_descriptor
 {
     __device__ __host__ hip_tensor_descriptor() = default;
-    template <typename T, typename V>
-    __device__ __host__ hip_tensor_descriptor(const T& lens_ext, const V& strides_ext)
+
+    hip_tensor_descriptor(const shape& s)
     {
-        for(size_t i = 0; i < NDim; i++)
-            lens[i] = lens_ext[i];
-        for(size_t i = 0; i < NDim; i++)
-            strides[i] = strides_ext[i];
+        std::copy(s.lens().begin(), s.lens().end(), lens);
+        std::copy(s.strides().begin(), s.strides().end(), strides);
     }
+
     __device__ __host__ hip_index<NDim> multi(size_t idx) const
     {
         hip_index<NDim> result{};

src/targets/gpu/eliminate_allocation.cpp

@@ -20,7 +20,7 @@ void eliminate_allocation::apply(program& p) const
             continue;
         allocs.emplace_back(ins, n);
         std::size_t size = ins->get_shape().bytes();
-        n += size + (size % 4);
+        n += size + (size % 32);
     }
     auto mem = p.add_parameter("memory", shape{shape::int8_type, {n}});
     for(auto&& pp : allocs)

src/targets/gpu/fuse_ops.cpp

@@ -12,7 +12,7 @@ struct hip_add_relu
     std::string name() const { return "hip::add_relu"; }
     shape compute_shape(const std::vector<shape>& inputs) const
     {
-        check_shapes{inputs}.has(3).standard();
+        check_shapes{inputs, *this}.has(3);
         return inputs.front();
     }
     argument compute(context&, const shape&, const std::vector<argument>& args) const

src/targets/gpu/include/migraph/gpu/device/add.hpp

+
+#ifndef MIGRAPH_GUARD_RTGLIB_DEVICE_ADD_HPP
+#define MIGRAPH_GUARD_RTGLIB_DEVICE_ADD_HPP
+
+#include <migraph/argument.hpp>
+
+namespace migraph {
+namespace gpu {
+namespace device {
+
+void add(const argument& result, const argument& arg1, const argument& arg2);
+
+} // namespace device
+} // namespace gpu
+} // namespace migraph
+
+#endif

src/targets/gpu/include/migraph/gpu/device/add_relu.hpp

@@ -8,7 +8,7 @@ namespace migraph {
 namespace gpu {
 namespace device {
 
-void add_relu(argument result, argument arg1, argument arg2);
+void add_relu(const argument& result, const argument& arg1, const argument& arg2);
 
 } // namespace device
 } // namespace gpu

src/targets/gpu/lowering.cpp

@@ -9,6 +9,7 @@
 #include <migraph/gpu/hip.hpp>
 #include <migraph/dfor.hpp>
 #include <migraph/gpu/device/contiguous.hpp>
+#include <migraph/gpu/device/add.hpp>
 #include <migraph/iterator_for.hpp>
 #include <migraph/gpu/rocblas.hpp>
 #include <migraph/gpu/context.hpp>
@@ -168,6 +169,23 @@ struct miopen_pooling
     }
 };
 
+struct hip_add
+{
+    std::string name() const { return "gpu::add"; }
+    shape compute_shape(const std::vector<shape>& inputs) const
+    {
+        // check_shapes{inputs, *this}.has(3).standard();
+        check_shapes{inputs, *this}.has(3);
+        return inputs.at(0);
+    }
+
+    argument compute(context&, const shape&, const std::vector<argument>& args) const
+    {
+        device::add(args[2], args[0], args[1]);
+        return args[2];
+    }
+};
+
 struct miopen_add
 {
     std::string name() const { return "gpu::add"; }
@@ -390,7 +408,7 @@ struct miopen_apply
     {
         auto output = insert_allocation(ins, ins->result);
         return prog->replace_instruction(
-            ins, miopen_add{}, ins->arguments.at(0), ins->arguments.at(1), output);
+            ins, hip_add{}, ins->arguments.at(0), ins->arguments.at(1), output);
     }
 
     instruction_ref apply_gemm(instruction_ref ins)

src/targets/gpu/target.cpp

@@ -10,6 +10,7 @@
 #include <migraph/dead_code_elimination.hpp>
 #include <migraph/simplify_reshapes.hpp>
 #include <migraph/eliminate_contiguous.hpp>
+#include <migraph/fwd_conv_batchnorm_rewrite.hpp>
 
 namespace migraph {
 namespace gpu {
@@ -20,6 +21,8 @@ std::vector<pass> target::get_passes(migraph::context& gctx) const
     // clang-format off
     return
     {
+        dead_code_elimination{},
+        fwd_conv_batchnorm_rewrite{},
         dead_code_elimination{},
         auto_contiguous{},
         simplify_reshapes{},

test/gpu/miopen.cpp

@@ -77,6 +77,12 @@ struct auto_print
 };
 std::array<std::function<void()>, 2> auto_print::handlers = {};
 
+template <class T>
+auto get_hash(const T& x)
+{
+    return std::hash<T>{}(x);
+}
+
 void compile_check(migraph::program& p, const migraph::target& t)
 {
     auto name = t.name();
@@ -100,7 +106,7 @@ migraph::argument run_cpu()
     migraph::program::parameter_map m;
     for(auto&& x : p.get_parameter_shapes())
     {
-        m[x.first] = migraph::generate_argument(x.second);
+        m[x.first] = migraph::generate_argument(x.second, get_hash(x.first));
     }
     return p.eval(m);
 }
@@ -112,11 +118,10 @@ migraph::argument run_gpu()
     auto p = v.create_program();
     auto_print pp{p, 1};
     compile_check(p, migraph::gpu::target{});
-
     migraph::program::parameter_map m;
     for(auto&& x : p.get_parameter_shapes())
     {
-        m[x.first] = migraph::gpu::to_gpu(migraph::generate_argument(x.second));
+        m[x.first] = migraph::gpu::to_gpu(migraph::generate_argument(x.second, get_hash(x.first)));
     }
 
     return migraph::gpu::from_gpu(p.eval(m));
@@ -131,8 +136,10 @@ void verify_args(const std::string& name,
         {
             // TODO: Check for nans
             std::cout << "FAILED: " << name << std::endl;
-            // std::cout << cpu << std::endl;
-            // std::cout << gpu << std::endl;
+            if(cpu.size() < 32)
+                std::cout << "cpu:" << cpu << std::endl;
+            if(gpu.size() < 32)
+                std::cout << "gpu:" << gpu << std::endl;
             if(migraph::range_zero(cpu))
                 std::cout << "Cpu data is all zeros" << std::endl;
             if(migraph::range_zero(gpu))
@@ -154,6 +161,7 @@ void verify_args(const std::string& name,
             if(gpu_nan_idx >= 0)
                 std::cout << "Non finite number found in gpu at " << gpu_nan_idx << ": "
                           << gpu[gpu_nan_idx] << std::endl;
+            std::cout << std::endl;
         }
     });
 }
@@ -210,6 +218,62 @@ struct test_add_broadcast
     }
 };
 
+struct test_add_broadcast2
+{
+    migraph::program create_program() const
+    {
+        migraph::program p;
+        migraph::shape s{migraph::shape::float_type, {3}};
+        auto x  = p.add_parameter("x", {migraph::shape::float_type, {2, 3, 4}});
+        auto y  = p.add_parameter("y", {migraph::shape::float_type, {3}});
+        auto by = p.add_instruction(migraph::broadcast{1}, x, y);
+        p.add_instruction(migraph::add{}, x, by);
+        return p;
+    }
+};
+
+struct test_add_broadcast3
+{
+    migraph::program create_program() const
+    {
+        migraph::program p;
+        migraph::shape s{migraph::shape::float_type, {3}};
+        auto x  = p.add_parameter("x", {migraph::shape::float_type, {2, 4, 5}});
+        auto y  = p.add_parameter("y", {migraph::shape::float_type, {4}});
+        auto by = p.add_instruction(migraph::broadcast{1}, x, y);
+        p.add_instruction(migraph::add{}, x, by);
+        return p;
+    }
+};
+
+struct test_add_broadcast4
+{
+    migraph::program create_program() const
+    {
+        migraph::program p;
+        migraph::shape s{migraph::shape::float_type, {3}};
+        auto x  = p.add_parameter("x", {migraph::shape::float_type, {2, 3, 5}});
+        auto y  = p.add_parameter("y", {migraph::shape::float_type, {3}});
+        auto by = p.add_instruction(migraph::broadcast{1}, x, y);
+        p.add_instruction(migraph::add{}, x, by);
+        return p;
+    }
+};
+
+struct test_add_broadcast5
+{
+    migraph::program create_program() const
+    {
+        migraph::program p;
+        migraph::shape s{migraph::shape::float_type, {3}};
+        auto x  = p.add_parameter("x", {migraph::shape::float_type, {2, 4, 8}});
+        auto y  = p.add_parameter("y", {migraph::shape::float_type, {4}});
+        auto by = p.add_instruction(migraph::broadcast{1}, x, y);
+        p.add_instruction(migraph::add{}, x, by);
+        return p;
+    }
+};
+
 struct test_conv_relu
 {
     migraph::program create_program() const
@@ -418,6 +482,10 @@ int main()
 {
     verify_program<test_add>();
     verify_program<test_add_broadcast>();
+    verify_program<test_add_broadcast2>();
+    verify_program<test_add_broadcast3>();
+    verify_program<test_add_broadcast4>();
+    verify_program<test_add_broadcast5>();
     verify_program<test_conv_relu>();
     verify_program<test_add_relu>();
     verify_program<test_conv_pooling>();