Merge branch 'develop' into doc-standard

src/targets/gpu/kernels/include/migraphx/kernels/gathernd.hpp

@@ -53,35 +53,35 @@ __device__ void gathernd(const T& data_t, const U& indices_t, const V& output_t,
     auto indices_shape_lens = indices_shape.lens;
     auto data_shape_lens    = data_shape.lens;
     auto num_slice_dims     = indices_shape_lens.back();
-    std::size_t num_slices =
+    size_t num_slices =
         accumulate(indices_shape_lens.begin(), indices_shape_lens.end() - 1, 1, op::product{});
-    std::size_t slice_size = accumulate(data_shape_lens.begin() + num_slice_dims + batch_dims,
+    size_t slice_size = accumulate(data_shape_lens.begin() + num_slice_dims + batch_dims,
                                    data_shape_lens.end(),
                                    1,
                                    op::product{});
-    const std::size_t num_batches =
+    const size_t num_batches =
         accumulate(data_shape_lens.begin(), data_shape_lens.begin() + batch_dims, 1, op::product{});
-    const std::size_t data_batch_stride =
+    const size_t data_batch_stride =
         accumulate(data_shape_lens.begin() + batch_dims, data_shape_lens.end(), 1, op::product{});
     const auto num_slices_per_batch = num_slices / num_batches;
 
     ind.global_stride(output_shape.elements(), [&](auto i) {
         const auto* indices_ptr     = indices_t.data();
-        const std::size_t j         = i / slice_size;
-        const std::size_t batch_idx = j / num_slices_per_batch;
+        const size_t j              = i / slice_size;
+        const size_t batch_idx      = j / num_slices_per_batch;
 
         auto* slice_indices               = indices_ptr + (j * num_slice_dims);
-        std::size_t relative_slice_offset = 0;
-        for(std::size_t idx = 0; idx < num_slice_dims; ++idx)
+        size_t relative_slice_offset      = 0;
+        for(size_t idx = 0; idx < num_slice_dims; ++idx)
         {
             int64_t index                   = slice_indices[idx];
-            const std::size_t input_dim_idx = batch_dims + idx;
+            const size_t input_dim_idx      = batch_dims + idx;
             const auto input_dim            = data_shape_lens[input_dim_idx];
             MIGRAPHX_ASSERT(index >= -static_cast<int64_t>(input_dim) and
                             index < static_cast<int64_t>(input_dim));
             if(index < 0)
                 index += input_dim;
-            std::size_t size_from_slice_dims =
+            size_t size_from_slice_dims =
                 accumulate(data_shape_lens.begin() + batch_dims + idx + 1,
                            data_shape_lens.begin() + batch_dims + num_slice_dims,
                            slice_size,

src/targets/gpu/kernels/include/migraphx/kernels/layernorm.hpp

@@ -52,14 +52,17 @@ __device__ void generic_binary_layernorm(
     block::template run<reduce_output>([&](auto, auto r) {
         auto input       = r.inner([&](auto x1, auto x2) { return op(x1, x2); })(input1, input2);
         using value_type = typename Input1::type;
+        using vec_value_type       = vec_type<value_type>;
         constexpr auto relements   = r.template elements<Input1>();
-        constexpr auto relements_r = vec_type<value_type>{1.0 / relements};
+        constexpr auto relements_r = vec_value_type{1.0 / relements};
         auto relements_rsqrt       = sqrt(relements_r);
 
-        auto means = r.reduce(op::sum{}, make_array<vec_type<value_type>>(0, 0), [&](auto x) {
+        auto means = r.reduce(op::sum{},
+                              make_array<vec_value_type>(vec_value_type{0}, vec_value_type{0}),
+                              [&](auto x) {
                                   auto x_out = x * relements_r;
-            // dividing x by sqrt(relements) before squaring allows computing higher values
-            // before overflow in low precision
+                                  // dividing x by sqrt(relements) before squaring allows computing
+                                  // higher values before overflow in low precision
                                   auto x2_sqrt = x * relements_rsqrt;
                                   return make_array(x_out, x2_sqrt * x2_sqrt);
                               })(input);
@@ -67,7 +70,7 @@ __device__ void generic_binary_layernorm(
         auto mean_x        = means[0];
         auto mean_x2       = means[1];
         auto variance      = mean_x2 - (mean_x * mean_x);
-        value_type eps_val = eps; // implicit conversion for eps
+        value_type eps_val = implicit_conversion(eps);
 
         r.inner([&](auto& y, auto x, auto... xs) {
             auto m = x - mean_x;

src/targets/gpu/kernels/include/migraphx/kernels/math.hpp

@@ -29,11 +29,15 @@
 #include <migraphx/kernels/functional.hpp>
 #include <migraphx/kernels/type_traits.hpp>
 #include <migraphx/kernels/hip.hpp>
+#include <migraphx/kernels/float8.hpp>
 
 namespace migraphx {
 
 namespace math {
 constexpr float as_float(migraphx::half x) { return x; }
+
+constexpr float as_float(migraphx::fp8::fp8e4m3fnuz x) { return x; }
+
 template <class T>
 constexpr T as_float(T x)
 {
@@ -57,14 +61,14 @@ constexpr T as_float(T x)
 // NOLINTNEXTLINE
 #define MIGRAPHX_DEVICE_MATH_FOR(type, name, fname)                    \
     template <class... Ts, MIGRAPHX_REQUIRES(not is_any_vec<Ts...>())> \
-    auto __device__ name(type x, Ts... xs)->type                       \
+    auto __device__ name(type x, Ts... xs) -> type                     \
     {                                                                  \
         return fname(x, xs...);                                        \
     }
 
 // NOLINTNEXTLINE
 #define MIGRAPHX_DEVICE_MATH_BINARY_FOR(type, name, fname) \
-    inline auto __device__ name(type x, type y)->type { return fname(x, y); }
+    inline auto __device__ name(type x, type y) -> type { return fname(x, y); }
 
 // NOLINTNEXTLINE
 #define MIGRAPHX_DEVICE_MATH_HALF(name, fname)                         \
@@ -72,6 +76,12 @@ constexpr T as_float(T x)
     auto __device__ name(migraphx::half x, Ts... xs)                   \
         MIGRAPHX_RETURNS(fname(math::as_float(x), math::as_float(xs)...))
 
+// NOLINTNEXTLINE
+#define MIGRAPHX_DEVICE_MATH_FP8(name, fname)                                      \
+    template <class... Ts, MIGRAPHX_REQUIRES(not is_any_vec<Ts...>())>             \
+    auto __device__ name(migraphx::fp8::fp8e4m3fnuz x, Ts... xs) MIGRAPHX_RETURNS( \
+        migraphx::fp8::fp8e4m3fnuz(fname(math::as_float(x), math::as_float(xs)...)))
+
 // Template with two overloads for math functions, one for half2 type and one for more generic
 // <half, N> vectorization where N is 4 or another even number.
 
@@ -162,6 +172,33 @@ MIGRAPHX_DEVICE_MATH_HALF(tan, ::tan)
 MIGRAPHX_DEVICE_MATH_HALF(tanh, ::tanh)
 MIGRAPHX_DEVICE_MATH_HALF(fmod, ::fmod)
 
+// use float to compute fp8 overload
+MIGRAPHX_DEVICE_MATH_FP8(abs, ::abs)
+MIGRAPHX_DEVICE_MATH_FP8(acos, ::acos)
+MIGRAPHX_DEVICE_MATH_FP8(acosh, ::acosh)
+MIGRAPHX_DEVICE_MATH_FP8(asin, ::asin)
+MIGRAPHX_DEVICE_MATH_FP8(asinh, ::asinh)
+MIGRAPHX_DEVICE_MATH_FP8(atan, ::atan)
+MIGRAPHX_DEVICE_MATH_FP8(atanh, ::atanh)
+MIGRAPHX_DEVICE_MATH_FP8(ceil, ::ceil)
+MIGRAPHX_DEVICE_MATH_FP8(cos, ::cos)
+MIGRAPHX_DEVICE_MATH_FP8(cosh, ::cosh)
+MIGRAPHX_DEVICE_MATH_FP8(erf, ::erf)
+MIGRAPHX_DEVICE_MATH_FP8(exp, ::exp)
+MIGRAPHX_DEVICE_MATH_FP8(floor, ::floor)
+MIGRAPHX_DEVICE_MATH_FP8(isnan, ::isnan)
+MIGRAPHX_DEVICE_MATH_FP8(log, ::log)
+MIGRAPHX_DEVICE_MATH_FP8(pow, ::pow)
+MIGRAPHX_DEVICE_MATH_FP8(remainder, ::remainder)
+MIGRAPHX_DEVICE_MATH_FP8(round, ::round)
+MIGRAPHX_DEVICE_MATH_FP8(rsqrt, ::rsqrt)
+MIGRAPHX_DEVICE_MATH_FP8(sin, ::sin)
+MIGRAPHX_DEVICE_MATH_FP8(sinh, ::sinh)
+MIGRAPHX_DEVICE_MATH_FP8(sqrt, ::sqrt)
+MIGRAPHX_DEVICE_MATH_FP8(tan, ::tan)
+MIGRAPHX_DEVICE_MATH_FP8(tanh, ::tanh)
+MIGRAPHX_DEVICE_MATH_FP8(fmod, ::fmod)
+
 // Map math functions to hip half2 functions
 // The half2 type is defined in include/hip/amd_detail/hip_fp16_gcc.h and is 2 16-bit floats
 // packed into a 32-bit number.  See include/hip/amd_detail/hip_fp16_math_fwd.h for the HIP names
@@ -253,7 +290,7 @@ MIGRAPHX_DEVICE_MATH_VEC(where)
 template <class T, class U>
 constexpr auto convert(U v)
 {
-    return vec_transform(v)([](auto x) -> T { return x; });
+    return vec_transform(v)([](auto x) -> T { return static_cast<T>(x); });
 }
 
 } // namespace migraphx

src/targets/gpu/kernels/include/migraphx/kernels/pad.hpp

@@ -28,6 +28,7 @@
 #include <migraphx/kernels/index.hpp>
 #include <migraphx/kernels/algorithm.hpp>
 #include <migraphx/kernels/ranges.hpp>
+#include <migraphx/kernels/vec.hpp>
 
 namespace migraphx {
 
@@ -53,9 +54,9 @@ __device__ void pad(const index& idx,
         if(any_of(range_multi.begin(), range_multi.end(), [&](auto j) {
                return multi[j] < offsets[j] or input_idx[j] >= input_bounds[j];
            }))
-            output[multi] = pad_val;
+            output[multi] = implicit_conversion(pad_val);
         else
-            output[multi] = input[input_idx];
+            output[multi] = implicit_conversion(input[input_idx]);
     });
 }
 

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

@@ -106,7 +106,7 @@ __device__ auto block_reduce(index idx, Op op, T init, Index n, F f)
 #endif
     using type = decltype(index::invoke_loop(f, 0, _c<0>));
     __shared__ type buffer[idx.max_nlocal() / lanes_per_thread];
-    type x = init;
+    type x = type(init);
     idx.local_stride(n, [&](auto i, auto d) { x = op(x, index::invoke_loop(f, i, d)); });
     dpp_reduce(x, op);
 
@@ -117,7 +117,7 @@ __device__ auto block_reduce(index idx, Op op, T init, Index n, F f)
     }
     __syncthreads();
 
-    type y = init;
+    type y = type(init);
     for(index_int i = 0; i < idx.nlocal() / lanes_per_thread; i++)
     {
         y = op(y, buffer[i]);
@@ -244,9 +244,8 @@ struct reducer_base
         {
             auto&& derived = static_cast<const Derived&>(*this);
             auto t         = derived.slice(x);
-            return make_storage_access<typename decltype(t)::type>([=](auto i, auto...) -> auto& {
-                return t[i];
-            });
+            return make_storage_access<typename decltype(t)::type>(
+                [=](auto i, auto...) -> auto& { return t[i]; });
         }
     }
 
@@ -393,7 +392,7 @@ struct block
         {
             using max_iterations = decltype(idx.max_local_stride_iterations(n));
             inner_storage<R, max_iterations{}, N> storage;
-            idx.local_stride(n, [&](auto j, auto d) { storage(j, d) = f(xs(j, d)...); });
+            idx.local_stride(n, [&](auto j, auto d) { storage(j, d) = R{f(xs(j, d)...)}; });
             return storage;
         }
     };
@@ -482,7 +481,7 @@ struct lane
         __device__ auto reduce_impl(Op op, T init, Read read, N n, U&& x, Us&&... xs) const
         {
             using type = remove_reference_t<decltype(x(0, _c<0>))>;
-            type r     = init;
+            type r     = type(init);
             for(index_int j = 0; j < n; j++)
             {
                 r = op(r, read(x(j, _c<0>), xs(j, _c<0>)...));

src/targets/gpu/kernels/include/migraphx/kernels/roialign.hpp

@@ -62,7 +62,7 @@ struct avg_pool
     template <class T>
     MIGRAPHX_DEVICE_CONSTEXPR T final(T x, index_int y)
     {
-        return (y == 0) ? 0.0 : (x / y);
+        return (y == 0) ? T{0.0} : T{x / y};
     }
 };
 
@@ -76,7 +76,7 @@ MIGRAPHX_DEVICE_CONSTEXPR typename Iterator::value_type bilinear_interpolate(
     {
         if(xy[ii] < -1.0f or xy[ii] > dims[ii])
         {
-            return 0;
+            return implicit_conversion(0);
         }
 
         xy[ii]   = migraphx::max(xy[ii], 0.0f);
@@ -96,11 +96,12 @@ MIGRAPHX_DEVICE_CONSTEXPR typename Iterator::value_type bilinear_interpolate(
     float lx = xy[1] - low[1];
     float hy = 1.0f - ly;
     float hx = 1.0f - lx;
-    array<typename Iterator::value_type, 4> ws = {hy * hx, hy * lx, ly * hx, ly * lx};
+    // do calculations in floating point and convert final result to required type
+    array<float, 4> ws = {hy * hx, hy * lx, ly * hx, ly * lx};
 
     auto v01 = pooling(data[locs[0]] * ws[0], data[locs[1]] * ws[1]);
     auto v23 = pooling(data[locs[2]] * ws[2], data[locs[3]] * ws[3]);
-    return pooling(v01, v23);
+    return implicit_conversion(pooling(v01, v23));
 }
 
 template <class Iterator, class Op>
@@ -113,7 +114,8 @@ MIGRAPHX_DEVICE_CONSTEXPR auto calc_pooling(const Iterator& data,
                                             float roi_offset,
                                             Op op)
 {
-    typename Iterator::value_type output_val = op.init();
+    using in_dtype      = typename Iterator::value_type;
+    in_dtype output_val = in_dtype{op.init()};
     const int64_t count = bin_grid_size[0] * bin_grid_size[1];
     dfor(bin_grid_size[0], bin_grid_size[1])([&](auto iy, auto ix) {
         array<index_int, 2> id = {iy, ix};
@@ -148,7 +150,6 @@ __device__ void roialign(const T& x_t, const U& rois_t, const V& ind_t, W& y_t,
     const auto x    = x_t.begin();
     const auto rois = rois_t.begin();
     const auto ind  = ind_t.begin();
-
     // input shape
     auto x_lens      = x_t.get_shape().lens;
     auto channel_num = x_lens[1];
@@ -176,10 +177,12 @@ __device__ void roialign(const T& x_t, const U& rois_t, const V& ind_t, W& y_t,
         const auto offset_rois = rois + (n * roi_column_num);
         const int batch_ind    = ind[n];
 
-        array<float, 2> roi_starts = {offset_rois[1] * s.spatial_scale,
-                                      offset_rois[0] * s.spatial_scale};
-        array<float, 2> roi_ends   = {offset_rois[3] * s.spatial_scale,
-                                    offset_rois[2] * s.spatial_scale};
+        array<float, 2> roi_starts = {
+            static_cast<float>(offset_rois[1]) * static_cast<float>(s.spatial_scale),
+            static_cast<float>(offset_rois[0]) * static_cast<float>(s.spatial_scale)};
+        array<float, 2> roi_ends = {
+            static_cast<float>(offset_rois[3]) * static_cast<float>(s.spatial_scale),
+            static_cast<float>(offset_rois[2]) * static_cast<float>(s.spatial_scale)};
 
         array<float, 2> roi_size{};
         array<float, 2> bin_size{};

src/targets/gpu/kernels/include/migraphx/kernels/scatter_reduction_modes.hpp

+/*
+ * The MIT License (MIT)
+ *
+ * Copyright (c) 2015-2023 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+ * THE SOFTWARE.
+ */
+#ifndef MIGRAPHX_GUARD_KERNELS_SCATTER_REDUCTION_MODES_HPP
+#define MIGRAPHX_GUARD_KERNELS_SCATTER_REDUCTION_MODES_HPP
+
+#include <migraphx/kernels/types.hpp>
+
+namespace migraphx {
+
+struct assign_none
+{
+    template <class T, class U>
+    MIGRAPHX_DEVICE_CONSTEXPR void operator()(T& x, U y) const
+    {
+        x = y;
+    }
+};
+
+struct assign_add
+{
+    template <class T, class U>
+    MIGRAPHX_DEVICE_CONSTEXPR void operator()(T& x, U y) const
+    {
+        atomicAdd(&x, y);
+    }
+};
+
+struct assign_mul
+{
+    template <class T, class U>
+    MIGRAPHX_DEVICE_CONSTEXPR void operator()(T& x, U y) const
+    {
+        T old = x;
+        T assumed;
+        do
+        {
+            assumed = old;
+            old     = atomicCAS(&x, assumed, assumed * y);
+        } while(assumed != old);
+    }
+};
+
+struct assign_max
+{
+    template <typename T, typename U>
+    MIGRAPHX_DEVICE_CONSTEXPR void operator()(T& x, U y) const
+    {
+        atomicMax(&x, y);
+    }
+};
+
+struct assign_min
+{
+    template <typename T, typename U>
+    MIGRAPHX_DEVICE_CONSTEXPR void operator()(T& x, U y) const
+    {
+        atomicMin(&x, y);
+    }
+};
+
+} // namespace migraphx
+#endif

src/targets/gpu/kernels/include/migraphx/kernels/scatternd.hpp

@@ -26,36 +26,10 @@
 
 #include <migraphx/kernels/index.hpp>
 #include <migraphx/kernels/algorithm.hpp>
+#include <migraphx/kernels/scatter_reduction_modes.hpp>
 
 namespace migraphx {
 
-struct assign_none
-{
-    template <class T, class U>
-    MIGRAPHX_DEVICE_CONSTEXPR void operator()(T& x, U y) const
-    {
-        x = y;
-    }
-};
-
-struct assign_add
-{
-    template <class T, class U>
-    MIGRAPHX_DEVICE_CONSTEXPR void operator()(T& x, U y) const
-    {
-        x += y;
-    }
-};
-
-struct assign_mul
-{
-    template <class T, class U>
-    MIGRAPHX_DEVICE_CONSTEXPR void operator()(T& x, U y) const
-    {
-        x *= y;
-    }
-};
-
 template <class T, class U, class V, class F>
 __device__ void scatternd(const T& indices_t, const U& updates_t, const V& output_t, F f)
 {

src/targets/gpu/kernels/include/migraphx/kernels/softmax.hpp

@@ -43,7 +43,7 @@ __device__ void softmax(Input input1, Output output)
         auto exp_in = r.inner([&](auto x) { return migraphx::exp(x - c); })(input);
         auto batch_sum =
             r.reduce(op::sum{}, 0, [](auto x) { return migraphx::convert<float>(x); })(exp_in);
-        r.inner([&](auto& y, auto x) { y = x / batch_sum; })(output, exp_in);
+        r.inner([&](auto& y, auto x) { y = implicit_conversion(x / batch_sum); })(output, exp_in);
     });
 }
 

src/targets/gpu/kernels/include/migraphx/kernels/tensor_view.hpp

@@ -27,6 +27,7 @@
 #include <migraphx/kernels/shape.hpp>
 #include <migraphx/kernels/debug.hpp>
 #include <migraphx/kernels/iota_iterator.hpp>
+#include <migraphx/kernels/float8.hpp>
 
 namespace migraphx {
 

src/targets/gpu/kernels/include/migraphx/kernels/type_traits.hpp

@@ -251,7 +251,7 @@ constexpr T numeric_max()
 }
 
 template <class T>
-constexpr T numeric_lowest()
+constexpr auto numeric_lowest() -> decltype(numeric_max<T>())
 {
     if constexpr(is_integral<T>{})
     {

src/targets/gpu/kernels/include/migraphx/kernels/vec.hpp

@@ -207,7 +207,7 @@ struct implicit_conversion_op
     template <class U>
     constexpr operator U() const
     {
-        return x;
+        return static_cast<U>(x);
     }
 };
 

src/targets/gpu/mlir.cpp

@@ -73,6 +73,7 @@ namespace gpu {
 
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_MLIR);
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNE_EXHAUSTIVE);
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNE_LIMIT);
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNING_DB);
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNING_CFG);
 
@@ -796,7 +797,9 @@ struct mlir_program
         if(enabled(MIGRAPHX_MLIR_TUNE_EXHAUSTIVE{}))
             tuning_mode = RocmlirTuningParamSetKindExhaustive;
         mlir_tuning_space params{mlirRockTuningSpaceCreate(mmodule.get(), tuning_mode)};
-        for(auto i : range(mlirRockTuningGetNumParams(params.get())))
+        const auto limit =
+            value_of(MIGRAPHX_MLIR_TUNE_LIMIT{}, std::numeric_limits<std::size_t>::max());
+        for(auto i : range(std::min<std::size_t>(limit, mlirRockTuningGetNumParams(params.get()))))
         {
             mlir_tuning_param param{mlirRockTuningParamCreate()};
             if(not mlirRockTuningParamGet(params.get(), i, param.get()))
@@ -1032,6 +1035,15 @@ tuning_config get_tuning_config_mlir(const context& migraphx_ctx,
     mlir_program mp;
     mp.set_gpu_properties(migraphx_ctx);
     mp.parse(m);
+
+    const bool trace = enabled(MIGRAPHX_TRACE_MLIR{});
+    static std::mutex mutex;
+    if(trace)
+    {
+        const std::lock_guard<std::mutex> lock(mutex);
+        auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
+        std::cout << mlir_print(&mlirOperationPrint, mod_op) << std::endl;
+    }
     return mp.get_tuning_config(exhaustive);
 }
 

src/targets/gpu/prefuse_ops.cpp

@@ -31,6 +31,7 @@
 #ifdef MIGRAPHX_USE_COMPOSABLEKERNEL
 #include <migraphx/gpu/ck.hpp>
 #endif
+#include <migraphx/gpu/fuse_mlir.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {
@@ -124,34 +125,55 @@ struct find_add_layernorm
     }
 };
 
-#ifdef MIGRAPHX_USE_COMPOSABLEKERNEL
-
 struct pre_gemm_softmax_gemm : gemm_softmax_gemm
 {
     std::string name() const { return "gpu::pre_gemm_softmax_gemm"; }
 };
 MIGRAPHX_REGISTER_OP(pre_gemm_softmax_gemm);
 
-MIGRAPHX_PRED_MATCHER(is_ck_gemm, instruction_ref ins)
+auto is_ck_gemm()
 {
+    return match::make_basic_pred_matcher([=](instruction_ref ins) {
+#ifdef MIGRAPHX_USE_COMPOSABLEKERNEL
+        if(not enabled(MIGRAPHX_ENABLE_CK{}))
+            return false;
         if(ins->name() != "dot")
             return false;
         if(not pre_gemm_softmax_gemm::is_ck_supported_type(ins->get_shape().type()))
             return false;
         return true;
+#else
+        (void)ins;
+        return false;
+#endif
+    });
+}
+
+auto is_mlir_gemm()
+{
+    return match::make_basic_pred_matcher([=](instruction_ref ins) {
+        if(not mlir_attention_enabled())
+            return false;
+        if(ins->name() != "dot")
+            return false;
+        return std::all_of(ins->inputs().begin(), ins->inputs().end(), [&](auto i) {
+            return pre_gemm_softmax_gemm::is_mlir_supported_type(i->get_shape().type());
+        });
+    });
 }
 
 struct find_gemm_softmax_gemm
 {
     auto matcher() const
     {
-        auto gemm1 =
-            match::skip(match::name("contiguous"))(match::name("dot")(is_ck_gemm().bind("gemm1")));
+        auto gemm1 = match::skip(match::name("contiguous"))(
+            match::name("dot")(match::any_of(is_ck_gemm(), is_mlir_gemm()).bind("gemm1")));
         auto mul = match::name("mul")(
             match::nargs(2), match::either_arg(0, 1)(match::is_constant().bind("scale"), gemm1));
         auto softmax = match::name("softmax")(match::arg(0)(mul)).bind("softmax");
 
-        return match::name("dot")(is_ck_gemm().bind("gemm2"))(match::arg(0)(softmax));
+        return match::name("dot")(match::any_of(is_ck_gemm(), is_mlir_gemm()).bind("gemm2"))(
+            match::arg(0)(softmax));
     }
 
     void apply(module_pass_manager& mpm, const match::matcher_result& r) const
@@ -179,8 +201,6 @@ struct find_gemm_softmax_gemm
     }
 };
 
-#endif
-
 } // namespace
 
 void prefuse_ops::apply(module_pass_manager& mpm) const
@@ -188,10 +208,7 @@ void prefuse_ops::apply(module_pass_manager& mpm) const
     match::find_matches(mpm.get_module(), find_layernorm{});
     mpm.run_pass(dead_code_elimination{});
     match::find_matches(mpm.get_module(), find_add_layernorm{});
-#ifdef MIHRAPHX_USE_COMPOSABLEKERNEL
-    if(enabled(MIGRAPHX_ENABLE_CK{}))
     match::find_matches(mpm, find_gemm_softmax_gemm{});
-#endif
 }
 
 } // namespace gpu

src/targets/gpu/target.cpp

@@ -98,6 +98,7 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
     ctx.set_exhaustive_tune_flag(options.exhaustive_tune);
     std::set<shape::type_t> unsupported_types(shape::types().begin(), shape::types().end());
     unsupported_types.erase(shape::type_t::float_type);
+    unsupported_types.erase(shape::type_t::fp8e4m3fnuz_type);
     unsupported_types.erase(shape::type_t::half_type);
     unsupported_types.erase(shape::type_t::bool_type);
     unsupported_types.erase(shape::type_t::int8_type);

test/gpu/fuse_mlir.cpp

@@ -144,10 +144,12 @@ TEST_CASE(int_quant_dot_tanh_fails)
         auto tanh = add_pointwise(p1, "main:pointwise0", {dot}, single_pointwise("tanh"));
         mm->add_return({tanh});
     }
-    migraphx::program p2(p1);
-    // This pass should do nothing as int32_t tanh isn't supported.
+    // This pass should not fuse as int32_t tanh isn't supported.
     run_pass(p1);
-    EXPECT(p1 == p2);
+    auto* mm = p1.get_main_module();
+    bool has_pointwise =
+        std::any_of(mm->begin(), mm->end(), [&](const auto& i) { return i.name() == "pointwise"; });
+    EXPECT(has_pointwise);
 }
 
 int main(int argc, const char* argv[])

test/gpu/jit.cpp

@@ -139,7 +139,8 @@ const std::string math_template = R"__migraphx__(
 #include <migraphx/kernels/pointwise.hpp>
 #include <migraphx/kernels/math.hpp>
 #include <migraphx/kernels/types.hpp>
-using namespace migraphx;
+
+namespace migraphx {
 extern "C" {
 __global__ void kernel(${type}* p) 
 {
@@ -148,6 +149,7 @@ __global__ void kernel(${type}* p)
 
 }
 }
+}
 
 int main() {}
 
@@ -348,19 +350,20 @@ TEST_CASE(compile_math)
     auto vec_sizes = {2, 4, 6};
     for(auto&& t : migraphx::shape::types())
     {
-        if(contains({migraphx::shape::bool_type,
-                     migraphx::shape::fp8e4m3fnuz_type,
-                     migraphx::shape::tuple_type},
-                    t))
+        if(contains({migraphx::shape::bool_type, migraphx::shape::tuple_type}, t))
             continue;
         auto name = migraphx::shape::cpp_type(t);
         if(t == migraphx::shape::half_type)
             name.insert(0, "migraphx::");
         data_types.push_back(name);
+        // fp8 doesn't have vectorization support yet, therefore skip it for now.
+        if(t != migraphx::shape::fp8e4m3fnuz_type)
+        {
             migraphx::transform(vec_sizes, std::back_inserter(data_types), [&](auto i) {
                 return "migraphx::vec<" + name + ", " + std::to_string(i) + ">";
             });
         }
+    }
     migraphx::shape input{migraphx::shape::float_type, {5, 2}};
     migraphx::gpu::hip_compile_options options;
     options.global = 1024;
@@ -429,7 +432,6 @@ TEST_CASE(assert_type_min_max)
                 min = std::to_string(as.min());
                 max = std::to_string(as.max());
             }
-
             auto src = migraphx::interpolate_string(assert_template,
                                                     {{"type", name}, {"max", max}, {"min", min}});
             migraphx::shape input{migraphx::shape::float_type, {5, 2}};