Skip to content

GitLab

Commit 3a4d36cf authored by charlie's avatar charlie
Browse files

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/AMDMIGraphX into dyn_model_test

parents 6bec381f e19f78ae
Hide whitespace changes
Inline Side-by-side
Showing with 223 additions and 1062 deletions
src/targets/gpu/device/sign.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/sign.hpp>
#include <migraphx/gpu/device/nary.hpp>
#include <migraphx/gpu/device/types.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void sign(hipStream_t stream, const argument& result, const argument& arg)
{
nary(stream, result, arg)([](auto x) __device__ { return (x > 0 ? 1 : ((x < 0) ? -1 : 0)); });
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/sin.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/sin.hpp>
#include <migraphx/gpu/device/nary.hpp>
#include <migraphx/gpu/device/types.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void sin(hipStream_t stream, const argument& result, const argument& arg)
{
nary(stream, result, arg)([](auto x) __device__ { return ::sin(to_hip_type(x)); });
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/sinh.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/sinh.hpp>
#include <migraphx/gpu/device/nary.hpp>
#include <migraphx/gpu/device/types.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void sinh(hipStream_t stream, const argument& result, const argument& arg)
{
nary(stream, result, arg)([](auto x) __device__ { return ::sinh(to_hip_type(x)); });
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/softmax.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/shape.hpp>
#include <migraphx/argument.hpp>
#include <migraphx/dfor.hpp>
#include <migraphx/gpu/device/softmax.hpp>
#include <migraphx/gpu/device/reduce.hpp>
#include <migraphx/gpu/device/tensor.hpp>
#include <migraphx/gpu/device/launch.hpp>
#include <migraphx/gpu/device/types.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void softmax(hipStream_t stream, const argument& result, const argument& arg, int64_t axis)
{
auto batch_lens = result.get_shape().lens();
index_int batch_item_num = batch_lens[axis];
batch_lens[axis] = 1;
migraphx::shape batch_shape{result.get_shape().type(), batch_lens};
hip_visit_all(result, arg, batch_shape)([&](auto output, auto input, auto batch) {
const index_int max_block_size = 128;
const index_int block_size = compute_block_size(batch_item_num, max_block_size);
using type = device_type<std::remove_cv_t<typename decltype(input)::value_type>>;
type init = lowest();
if(axis == batch_lens.size() - 1)
{
gs_launch(stream, batch_shape.elements() * block_size, block_size)(
[=](auto i, auto idx) __device__ {
auto start_loc = i / block_size * batch_item_num;
auto batch_max = block_reduce<max_block_size>(
idx, max{}, init, batch_item_num, [&](auto j) __device__ {
return input[start_loc + j];
});
auto batch_sum = block_reduce<max_block_size>(
idx, sum{}, 0, batch_item_num, [&](auto j) __device__ {
auto val = input[start_loc + j] - batch_max;
return ::exp(to_hip_type(val));
});
idx.local_stride(batch_item_num, [&](auto j) __device__ {
auto val = input[start_loc + j] - batch_max;
output[start_loc + j] = ::exp(to_hip_type(val)) / batch_sum;
});
});
}
else
{
gs_launch(stream, batch_shape.elements() * block_size, block_size)(
[=](auto i, auto idx) __device__ {
auto data_idx = batch.multi(i / block_size);
auto batch_max = block_reduce<max_block_size>(
idx, max{}, init, batch_item_num, [&](auto j) __device__ {
data_idx[axis] = j;
return input[data_idx];
});
auto batch_sum = block_reduce<max_block_size>(
idx, sum{}, 0, batch_item_num, [&](auto j) __device__ {
data_idx[axis] = j;
auto val = input[data_idx] - batch_max;
return ::exp(to_hip_type(val));
});
idx.local_stride(batch_item_num, [&](auto j) __device__ {
data_idx[axis] = j;
auto val = input[data_idx] - batch_max;
output[data_idx] = ::exp(to_hip_type(val)) / batch_sum;
});
});
}
});
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/sqdiff.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/sqdiff.hpp>
#include <migraphx/gpu/device/nary.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void sqdiff(hipStream_t stream, const argument& result, const argument& arg1, const argument& arg2)
{
nary(stream, result, arg1, arg2)([](auto x, auto y) __device__ { return (x - y) * (x - y); });
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/sqrt.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/sqrt.hpp>
#include <migraphx/gpu/device/nary.hpp>
#include <migraphx/gpu/device/types.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void sqrt(hipStream_t stream, const argument& result, const argument& arg)
{
nary(stream, result, arg)([](auto x) __device__ { return ::sqrt(to_hip_type(x)); });
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/sub.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/sub.hpp>
#include <migraphx/gpu/device/nary.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void sub(hipStream_t stream, const argument& result, const argument& arg1, const argument& arg2)
{
nary(stream, result, arg1, arg2)([](auto x, auto y) __device__ { return x - y; });
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/tan.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/tan.hpp>
#include <migraphx/gpu/device/nary.hpp>
#include <migraphx/gpu/device/types.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void tan(hipStream_t stream, const argument& result, const argument& arg)
{
nary(stream, result, arg)([](auto x) __device__ { return ::tan(to_hip_type(x)); });
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/tanh.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/tanh.hpp>
#include <migraphx/gpu/device/nary.hpp>
#include <migraphx/gpu/device/types.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void tanh(hipStream_t stream, const argument& result, const argument& arg)
{
nary(stream, result, arg)([](auto x) __device__ { return ::tanh(to_hip_type(x)); });
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/unary_not.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/unary_not.hpp>
#include <migraphx/gpu/device/nary.hpp>
#include <migraphx/type_traits.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
void unary_not(hipStream_t stream, const argument& result, const argument& arg)
{
nary(stream, result, arg)([](auto x) __device__ { return not x; });
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/device/where.cpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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.
*/
#include <migraphx/gpu/device/where.hpp>
#include <migraphx/gpu/device/tensor.hpp>
#include <migraphx/gpu/device/types.hpp>
#include <migraphx/gpu/device/launch.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace device {
template <class Shape>
constexpr auto get_rank(const Shape&)
{
return decltype(typename Shape::hip_index{}.size()){};
}
void where(hipStream_t stream,
const argument& result,
const argument& arg0,
const argument& arg1,
const argument& arg2)
{
hip_visit_all(result, arg1, arg2)([&](auto output, auto x, auto y) {
hip_visit_all(arg0)([&](auto cond) {
if constexpr(get_rank(cond.get_shape()) == get_rank(output.get_shape()))
{
gs_launch(stream, arg1.get_shape().elements())([=](auto idx) __device__ {
auto i = output.get_shape().multi(idx);
output[i] = cond[i] ? x[i] : y[i];
});
}
});
});
}
} // namespace device
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/driver/compile_op.cpp
View file @ 3a4d36cf
......@@ -38,8 +38,11 @@ struct compile_op : action<compile_op>
context ctx;
auto inputs = p.parse_shapes(v.at("inputs"));
auto op = gpu::compile_op(v.at("name").to<std::string>(), ctx, inputs, v);
double t = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
std::cout << op << ": " << t << "ms" << std::endl;
auto [host_time, device_time] = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
std::cout << op << ": " << host_time << "ms";
if(device_time > 0)
std::cout << ", " << device_time << "ms";
std::cout << std::endl;
}
};
......
src/targets/gpu/driver/include/migraphx/gpu/driver/perf.hpp
View file @ 3a4d36cf
......@@ -33,7 +33,8 @@ inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
namespace driver {
double time_op(context& ctx, operation op, const std::vector<shape>& inputs, int n = 100);
std::pair<double, double>
time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n = 100);
} // namespace driver
} // namespace gpu
......
src/targets/gpu/driver/perf.cpp
View file @ 3a4d36cf
......@@ -42,22 +42,31 @@ std::vector<argument> generate_arguments(const std::vector<shape>& shapes, unsig
}
using milliseconds = std::chrono::duration<double, std::milli>;
double time_op(context& ctx, operation op, const std::vector<shape>& inputs, int n)
std::pair<double, double>
time_op(context& ictx, operation op, const std::vector<shape>& inputs, int n)
{
// TODO: Use std::ref
migraphx::context gctx = ctx;
auto output = op.compute_shape(inputs);
op.finalize(gctx, output, inputs);
migraphx::context ctx = ictx;
auto& gctx = any_cast<migraphx::gpu::context>(ctx);
auto output = op.compute_shape(inputs);
op.finalize(ctx, output, inputs);
auto args = generate_arguments(inputs);
auto run = [&] {
op.compute(gctx, output, args);
gctx.finish();
op.compute(ctx, output, args);
ctx.finish();
};
gctx.enable_perf_measurement();
run();
auto r = range(n);
double t = std::accumulate(
r.begin(), r.end(), double{0.0}, [&](auto x, auto) { return x + time<milliseconds>(run); });
return t / n;
double host_time = 0.0;
double device_time = 0.0;
for(auto i : range(n))
{
(void)i;
host_time += time<milliseconds>(run);
device_time += gctx.get_elapsed_ms();
}
return std::make_pair(host_time / n, device_time / n);
}
} // namespace driver
......
src/targets/gpu/driver/run_op.cpp
View file @ 3a4d36cf
......@@ -43,8 +43,8 @@ struct run_op : action<run_op>
auto op = make_op(name);
if(v.contains("fields"))
op.from_value(v.at("fields"));
double t = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
std::cout << op << ": " << t << "ms" << std::endl;
auto [host_time, device_time] = time_op(ctx, op, inputs, p.get(v, "iterations", 100));
std::cout << op << ": " << host_time << "ms" << std::endl;
}
};
......
src/targets/gpu/fuse_mlir.cpp
View file @ 3a4d36cf
......@@ -61,13 +61,25 @@ struct mlir_conv
MIGRAPHX_REGISTER_OP(mlir_conv);
namespace {
MIGRAPHX_PRED_MATCHER(is_mlir_conv, instruction_ref ins)
{
if(ins->name() != "convolution")
return false;
value v = ins->get_operator().to_value();
auto group = v.at("group").to<int>();
if(group != 1)
return false;
return true;
}
struct find_conv_pointwise
{
// Find a convolution followed by a pointwise operation.
auto matcher() const
{
auto convolution =
match::skip(match::name("contiguous"))(match::name("convolution").bind("convolution"));
match::skip(match::name("contiguous"))(is_mlir_conv().bind("convolution"));
return match::name("pointwise")(match::any_of[match::inputs()](convolution.bind("x")));
}
......
src/targets/gpu/fuse_ops.cpp
View file @ 3a4d36cf
......@@ -26,30 +26,15 @@
#include <migraphx/gpu/fuse_ops.hpp>
#include <migraphx/matcher.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/gpu/clip.hpp>
#include <migraphx/gpu/convolution.hpp>
#include <migraphx/gpu/device_name.hpp>
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/add.hpp>
#include <migraphx/gpu/mul.hpp>
#include <migraphx/gpu/gemm.hpp>
#include <migraphx/gpu/device/layernorm.hpp>
#include <migraphx/gpu/device/gelu.hpp>
#include <migraphx/gpu/device/mul_add.hpp>
#include <migraphx/gpu/device/add_clip.hpp>
#include <migraphx/gpu/device/add_relu.hpp>
#include <migraphx/gpu/device/add_sigmoid.hpp>
#include <migraphx/gpu/device/add_tanh.hpp>
#include <migraphx/gpu/device/mul_add_relu.hpp>
#include <migraphx/gpu/device/add.hpp>
#include <migraphx/match/layernorm.hpp>
#include <migraphx/match/gelu_erf.hpp>
#include <migraphx/match/gelu_tanh.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/register_op.hpp>
#include <migraphx/array.hpp>
#include <migraphx/permutation.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/op/clip.hpp>
#include <cmath>
#include <set>
......@@ -225,95 +210,6 @@ MIGRAPHX_PRED_MATCHER(fusable_conv, instruction_ref ins)
contains({{0, 0}, {1, 1}}, op.stride) and contains({{1, 1}}, op.dilation);
}
struct hip_triadd : ternary_device<hip_triadd, &device::add>
{
};
MIGRAPHX_REGISTER_OP(hip_triadd)
struct hip_triadd_clip : quinary_device<hip_triadd_clip, &device::add_clip>
{
};
MIGRAPHX_REGISTER_OP(hip_triadd_clip)
struct hip_add_clip : quaternary_device<hip_add_clip, &device::add_clip>
{
};
MIGRAPHX_REGISTER_OP(hip_add_clip)
struct hip_triadd_relu : ternary_device<hip_triadd_relu, &device::add_relu>
{
};
MIGRAPHX_REGISTER_OP(hip_triadd_relu)
struct hip_triadd_sigmoid : ternary_device<hip_triadd_sigmoid, &device::add_sigmoid>
{
};
MIGRAPHX_REGISTER_OP(hip_triadd_sigmoid)
struct hip_triadd_tanh : ternary_device<hip_triadd_tanh, &device::add_tanh>
{
};
MIGRAPHX_REGISTER_OP(hip_triadd_tanh)
struct hip_add_relu : binary_device<hip_add_relu, &device::add_relu>
{
};
MIGRAPHX_REGISTER_OP(hip_add_relu)
struct hip_add_sigmoid : binary_device<hip_add_relu, &device::add_sigmoid>
{
};
MIGRAPHX_REGISTER_OP(hip_add_sigmoid)
struct hip_add_tanh : binary_device<hip_add_tanh, &device::add_tanh>
{
};
MIGRAPHX_REGISTER_OP(hip_add_tanh)
struct hip_layernorm : unary_device<hip_layernorm, &device::layernorm>
{
// Empty finalize to skip dimension reduction
void finalize(context&, const shape&, const std::vector<shape>&) {}
};
MIGRAPHX_REGISTER_OP(hip_layernorm)
struct hip_triadd_layernorm : ternary_device<hip_triadd_layernorm, &device::triadd_layernorm>
{
// Empty finalize to skip dimension reduction
void finalize(context&, const shape&, const std::vector<shape>&) {}
};
MIGRAPHX_REGISTER_OP(hip_triadd_layernorm)
struct hip_gelu : unary_device<hip_gelu, &device::gelu>
{
};
MIGRAPHX_REGISTER_OP(hip_gelu)
struct hip_add_gelu : binary_device<hip_add_gelu, &device::add_gelu>
{
};
MIGRAPHX_REGISTER_OP(hip_add_gelu)
struct hip_gelu_new : unary_device<hip_gelu_new, &device::gelu_new>
{
};
MIGRAPHX_REGISTER_OP(hip_gelu_new)
struct hip_add_gelu_new : binary_device<hip_add_gelu_new, &device::add_gelu_new>
{
};
MIGRAPHX_REGISTER_OP(hip_add_gelu_new)
struct hip_mul_add : ternary_device<hip_mul_add, &device::mul_add>
{
};
MIGRAPHX_REGISTER_OP(hip_mul_add)
struct hip_mul_add_relu : ternary_device<hip_mul_add_relu, &device::mul_add_relu>
{
};
MIGRAPHX_REGISTER_OP(hip_mul_add_relu)
void move_broadcasted_back(std::vector<instruction_ref>& args)
{
// Ensure the last arguments is the broadcasted one
......@@ -337,256 +233,6 @@ void move_standard_front(std::vector<instruction_ref>& args)
auto gpu_name(const std::string& s) { return match::name("gpu::" + s); }
namespace {
struct find_layernorm
{
auto matcher() const { return match::layernorm(&gpu_name); }
void apply(module& m, const match::matcher_result& r) const
{
auto ins = r.result;
auto x_ins = r.instructions["x"];
auto args = ins->inputs();
// We dont fuse for non-standard layouts
if(not x_ins->get_shape().standard())
return;
auto relements = x_ins->get_shape().lens().back();
if(relements > 1024 or (relements % 4 != 0 and relements > 256))
return;
m.replace_instruction(ins, hip_layernorm{}, x_ins, args.back());
}
};
struct find_triadd_layernorm
{
auto matcher() const
{
return match::name("gpu::layernorm")(match::arg(0)(match::name("gpu::triadd")(
match::used_once(), match::all_of[match::inputs()](match::standard_shape()))));
}
void apply(module& m, const match::matcher_result& r) const
{
auto ins = r.result;
auto triadd = ins->inputs().front();
m.replace_instruction(ins, hip_triadd_layernorm{}, triadd->inputs());
}
};
struct find_gelu
{
auto matcher() const { return match::gelu_erf(&gpu_name); }
void apply(module& m, const match::matcher_result& r) const
{
auto ins = r.result;
auto x_ins = r.instructions["x"];
auto args = ins->inputs();
m.replace_instruction(ins, hip_gelu{}, x_ins, args.back());
}
};
struct find_add_gelu
{
auto matcher() const
{
return match::name("gpu::gelu")(match::arg(0)(match::name("gpu::add").bind("add")));
}
void apply(module& m, const match::matcher_result& r) const
{
auto add_ins = r.instructions["add"];
auto ins = r.result;
auto args = add_ins->inputs();
move_standard_front(args);
move_broadcasted_back(args);
args.back() = ins->inputs().back();
m.replace_instruction(ins, hip_add_gelu{}, args);
}
};
struct find_gelu_new
{
bool fast_math = true;
auto matcher() const { return match::gelu_tanh(&gpu_name); }
void apply(module& m, const match::matcher_result& r) const
{
auto ins = r.result;
auto x_ins = r.instructions["x"];
auto args = ins->inputs();
if(fast_math)
m.replace_instruction(ins, hip_gelu{}, x_ins, args.back());
else
m.replace_instruction(ins, hip_gelu_new{}, x_ins, args.back());
}
};
struct find_add_gelu_new
{
auto matcher() const
{
return match::name("gpu::gelu_new")(match::arg(0)(match::name("gpu::add").bind("add")));
}
void apply(module& m, const match::matcher_result& r) const
{
auto add_ins = r.instructions["add"];
auto ins = r.result;
auto args = add_ins->inputs();
move_standard_front(args);
move_broadcasted_back(args);
args.back() = ins->inputs().back();
m.replace_instruction(ins, hip_add_gelu_new{}, args);
}
};
struct find_add_clip
{
auto matcher() const
{
return match::name(std::unordered_set<std::string>{"gpu::clip", "gpu::clipped_relu"})(
match::arg(0)(match::any_of(match::name("gpu::add"),
match::name("gpu::triadd"),
match::any_of[match::inputs()](match::standard_shape()))
.bind("add")));
}
void apply(module& m, const match::matcher_result& r) const
{
auto add_ins = r.instructions["add"];
auto ins = r.result;
auto ins_args = ins->inputs();
auto add_args = add_ins->inputs();
move_standard_front(add_args);
move_broadcasted_back(add_args);
// Use the allocation from the clip operator
add_args.pop_back();
add_args.insert(add_args.end(), std::next(ins_args.begin()), ins_args.end());
if(add_ins->name() == "gpu::add")
m.replace_instruction(ins, hip_add_clip{}, add_args);
else if(add_ins->name() == "gpu::triadd")
m.replace_instruction(ins, hip_triadd_clip{}, add_args);
}
};
struct find_add_unary
{
std::string op_name;
operation binary_add_op;
operation ternary_add_op;
auto matcher() const
{
return match::name(op_name)(match::arg(0)(
match::used_once(),
match::any_of(match::name("gpu::add"),
match::name("gpu::triadd"),
match::any_of(match::name("@literal"),
match::any_of[match::inputs()](match::standard_shape())))
.bind("add")));
}
void apply(module& m, const match::matcher_result& r) const
{
auto add_ins = r.instructions["add"];
auto ins = r.result;
auto args = add_ins->inputs();
move_standard_front(args);
move_broadcasted_back(args);
// Use the allocation from the relu operator
args.back() = ins->inputs().back();
if(add_ins->name() == "gpu::add")
m.replace_instruction(ins, binary_add_op, args);
else if(add_ins->name() == "gpu::triadd")
m.replace_instruction(ins, ternary_add_op, args);
}
};
struct find_triadd
{
auto matcher() const
{
return match::name("gpu::add")(match::either_arg(0, 1)(
match::name("gpu::add")(match::used_once()).bind("add"),
match::any(match::any_of(match::name("@literal"),
match::any_of[match::inputs()](match::standard_shape())))
.bind("input")));
}
void apply(module& m, const match::matcher_result& r) const
{
auto add_ins = r.instructions["add"];
auto input_ins = r.instructions["input"];
auto ins = r.result;
auto args = add_ins->inputs();
auto is_broadcasted = [](auto arg) { return arg->get_shape().broadcasted(); };
if(std::count_if(args.begin(), args.end(), is_broadcasted) > 2)
return;
args.insert(args.begin(), input_ins);
move_standard_front(args);
move_broadcasted_back(args);
args.back() = ins->inputs().back();
m.replace_instruction(ins, hip_triadd{}, args);
}
};
struct find_mul_add
{
auto matcher() const
{
return match::name("gpu::add")(match::either_arg(0, 1)(
match::name("gpu::mul")(match::used_once()).bind("mul"), match::any().bind("b")));
}
void apply(module& m, const match::matcher_result& r) const
{
auto mul_ins = r.instructions["mul"];
auto b_ins = r.instructions["b"];
auto ins = r.result;
auto args = mul_ins->inputs();
assert(mul_ins != b_ins);
move_standard_front(args);
move_broadcasted_back(args);
args.insert(std::prev(args.end()), b_ins);
args.back() = ins->inputs().back();
m.replace_instruction(ins, hip_mul_add{}, args);
}
};
struct find_mul_add_relu
{
auto matcher() const
{
return match::name("gpu::relu")(
match::arg(0)(match::name("gpu::mul_add")(match::used_once()).bind("mul_add")));
}
void apply(module& m, const match::matcher_result& r) const
{
auto mul_add_ins = r.instructions["mul_add"];
auto ins = r.result;
auto args = mul_add_ins->inputs();
// Use the allocation from the relu operator
args.back() = ins->inputs().back();
m.replace_instruction(ins, hip_mul_add_relu{}, args);
}
};
struct miopen_fusion
{
struct fuse_op_data
......@@ -827,13 +473,14 @@ void apply_conv_bias(context& ctx, module& m, const match::matcher_result& r)
m.replace_instruction(ins, cb, input_ins, weights_ins, old_ws_ins, bias_ins, alloc_ins);
}
inline auto precompile_name(std::string s) // NOLINT
template <class... Strings>
inline auto precompile_name(Strings... names) // NOLINT
{
return match::make_basic_pred_matcher([=](instruction_ref ins) {
if(ins->name() != "gpu::precompile_op")
return false;
auto op = from_value<operation>(ins->get_operator().to_value().at("op"));
return (op.name() == s);
return (contains({names...}, op.name()));
});
}
......@@ -902,14 +549,70 @@ struct find_conv_pointwise
}
};
struct find_gemm_add
struct find_gemm_pointwise
{
auto matcher() const
{
return match::name("gpu::add")(
match::all_of[match::inputs()](match::standard_shape()),
match::either_arg(0, 1)(match::used_once().bind("c"),
match::name("gpu::gemm")(match::nargs(3)).bind("gemm")));
return precompile_name("pointwise")(
match::nargs(3),
match::either_arg(0, 1)(
match::any_of(match::standard_shape(), match::is_constant()).bind("c"),
match::name("gpu::gemm")(match::nargs(3), match::used_once()).bind("gemm")));
}
// TODO: Move to matcher.hpp
static auto match_param(const std::string& name)
{
return match::make_basic_pred_matcher([=](auto ins) {
if(ins->name() != "@param")
return false;
auto p = any_cast<builtin::param>(ins->get_operator());
return p.parameter == name;
});
}
template <class M>
static auto match_mul_const(M m, const std::string& var)
{
return match::name("mul")(match::either_arg(0, 1)(match::name("@literal").bind(var), m))
.bind(var + "_mul");
}
static auto match_add(const std::string& input, const std::string& output)
{
auto param = match::name("@param");
auto add = match::name("add")(match::args(param, param));
auto inner_mul = match::any_of(match_mul_const(match_param(input), "alpha"),
match_mul_const(match_param(output), "beta"));
auto mul_add = match::name("add")(match::either_arg(0, 1)(inner_mul, param));
auto add_mul = match_mul_const(add, "gamma");
return match::name("@return")(match::args(match::any_of(add, mul_add, add_mul)));
}
static float get_float(instruction_ref ins) { return ins->get_literal().at<float>(); }
template <class Gemm>
static bool update_gemm(Gemm& gemm, module_ref pm, unsigned input)
{
auto names = pm->get_parameter_names();
if(names.size() != 2)
return false;
std::sort(names.begin(), names.end());
unsigned output = input == 0 ? 1 : 0;
auto mr = match::match_instruction(
*pm, std::prev(pm->end()), match_add(names[input], names[output]));
if(mr.result == pm->end())
return false;
if(contains(mr.instructions, "alpha_mul"))
gemm.alpha *= get_float(mr.instructions["alpha"]);
else if(contains(mr.instructions, "beta_mul"))
gemm.beta *= get_float(mr.instructions["beta"]);
else if(contains(mr.instructions, "gamma_mul"))
{
gemm.alpha *= get_float(mr.instructions["gamma"]);
gemm.beta *= get_float(mr.instructions["gamma"]);
}
return true;
}
void apply(module& m, const match::matcher_result& r) const
......@@ -923,69 +626,85 @@ struct find_gemm_add
// Already fused gemm
if(not float_equal(gemm.beta, 0))
return;
gemm.beta = 1;
auto inputs = gemm_ins->inputs();
inputs.pop_back();
auto copy_ins = c_ins;
if(not update_gemm(
gemm, ins->module_inputs().front(), ins->inputs().front() == gemm_ins ? 0 : 1))
return;
// Insert copy
if(ins == m.end() or c_ins->outputs().size() > 1 or c_ins->inputs().empty())
// const-fold input if not standard shape since rocblas can't handle it
if(not c_ins->get_shape().standard())
{
copy_ins = m.insert_instruction(ins, hip_copy{}, c_ins, ins->inputs().back());
auto c = make_op("contiguous");
auto l = c.compute(c.compute_shape({c_ins->get_shape()}), {c_ins->eval()});
c_ins = m.add_literal(l.get_shape(), l.data());
}
inputs.push_back(copy_ins);
inputs.push_back(copy_ins);
gemm.beta = 1;
auto inputs = gemm_ins->inputs();
inputs.pop_back();
inputs.push_back(c_ins);
inputs.push_back(ins->inputs().back());
m.replace_instruction(ins, gemm, inputs);
}
};
auto pointwise_name(const std::string& s)
{
return precompile_name("pointwise")(match::make_basic_pred_matcher([=](auto ins) {
module_ref pm = ins->module_inputs().front();
auto n = std::count_if(pm->begin(), pm->end(), [&](auto& i) { return i.name() == s; });
if(n != 1)
return false;
return std::all_of(pm->begin(), pm->end(), [&](auto& i) {
return starts_with(i.name(), "@") or i.name() == s;
});
}));
}
struct find_gemm_pointwise
struct find_contiguous_tranpose_gemm
{
auto matcher() const
{
return pointwise_name("add")(
match::nargs(3),
match::all_of[match::inputs()](match::standard_shape()),
match::either_arg(0, 1)(match::used_once().bind("c"),
match::name("gpu::gemm")(match::nargs(3)).bind("gemm")));
return match::name("gpu::contiguous")(match::arg(0)(
match::name("transpose")(
match::arg(0)(match::name("gpu::gemm")(match::used_once()).bind("gemm")))
.bind("transpose")));
}
template <class Vector>
static bool is_swapped(const Vector& perm, std::size_t i, std::size_t j)
{
if(i >= perm.size() or j >= perm.size())
return false;
auto perm2 = perm;
std::iota(perm2.begin(), perm2.end(), 0);
std::swap(perm2[i], perm2[j]);
return perm2 == perm;
}
void apply(module& m, const match::matcher_result& r) const
{
auto ins = r.result;
auto gemm_ins = r.instructions["gemm"];
auto c_ins = r.instructions["c"];
auto ins = r.result;
auto gemm = r.instructions["gemm"];
auto alloc = gemm->inputs().back();
auto transpose = r.instructions["transpose"];
auto perm = transpose->get_operator().to_value()["permutation"].to_vector<int64_t>();
auto iperm = invert_permutation(perm);
auto gemm = any_cast<rocblas_gemm<op::dot>>(gemm_ins->get_operator());
if(perm.size() < 3)
return;
// Already fused gemm
if(not float_equal(gemm.beta, 0))
if(not is_swapped(perm, perm.size() - 3, perm.size() - 2))
return;
auto inputs = gemm_ins->inputs();
inputs.pop_back();
auto lens = gemm->get_shape().lens();
if(lens.size() > 3 and
not std::all_of(lens.begin(), lens.end() - 3, [](auto i) { return i == 1; }))
return;
inputs.push_back(c_ins);
inputs.push_back(ins->inputs().back());
auto gemmv = gemm->get_operator().to_value();
gemmv["trans_batch"] = 1;
gemm.beta = 1;
m.replace_instruction(ins, gemm, inputs);
auto s = shape{alloc->get_shape().type(), reorder_dims(alloc->get_shape().lens(), iperm)};
auto new_alloc = m.insert_instruction(gemm, make_op("allocate", {{"shape", to_value(s)}}));
auto alloc_transpose =
m.insert_instruction(gemm, make_op("transpose", {{"permutation", perm}}), new_alloc);
auto inputs = gemm->inputs();
inputs.back() = alloc_transpose;
auto new_gemm = m.insert_instruction(gemm, make_op("gpu::gemm", gemmv), inputs);
auto gemm_transpoe = m.insert_instruction(gemm, transpose->get_operator(), new_gemm);
m.replace_instruction(ins, gemm_transpoe);
}
};
......@@ -1041,29 +760,41 @@ struct find_contiguous_pointwise
}
};
struct find_layernorm_pointwise
{
auto matcher() const
{
return precompile_name("pointwise")(match::arg(0)(
precompile_name("gpu::prelayernorm", "gpu::preadd_layernorm").bind("layernorm")));
}
void apply(module& m, const match::matcher_result& r) const
{
auto ins = r.result;
auto layernorm = r.instructions["layernorm"];
auto* pm = ins->module_inputs().front();
if(not layernorm->module_inputs().empty())
return;
auto inputs = layernorm->inputs();
inputs.pop_back();
inputs.insert(inputs.end(), ins->inputs().begin() + 1, ins->inputs().end());
m.replace_instruction(ins, layernorm->get_operator(), inputs, {pm});
}
};
void fuse_ops::apply(module& m) const
{
match::find_matches(m, find_contiguous_pointwise{}, find_gelu{}, find_gelu_new{fast_math});
match::find_matches(m, find_contiguous_pointwise{});
run_passes(m, {dead_code_elimination{}});
match::find_matches(m, find_triadd{});
match::find_matches(m,
find_layernorm{},
find_conv_pointwise{ctx},
find_conv_bias_relu{ctx},
find_conv_bias{ctx},
find_add_gelu{},
find_add_gelu_new{},
find_mul_add{},
find_mul_add_relu{},
find_add_unary{"gpu::relu", hip_add_relu{}, hip_triadd_relu{}},
find_add_unary{"gpu::sigmoid", hip_add_sigmoid{}, hip_triadd_sigmoid{}},
find_add_unary{"gpu::tanh", hip_add_tanh{}, hip_triadd_tanh{}},
find_add_clip{});
match::find_matches(m, find_conv_pointwise{ctx}, find_conv_bias_relu{ctx}, find_conv_bias{ctx});
run_passes(m, {dead_code_elimination{}});
match::find_matches(m,
find_triadd_layernorm{},
find_gemm_add{},
find_layernorm_pointwise{},
find_gemm_pointwise{},
find_contiguous_tranpose_gemm{},
find_commutative_broadcast{});
match::find_matches(m, find_contiguous{});
}
......
src/targets/gpu/gemm_impl.cpp
View file @ 3a4d36cf
......@@ -24,6 +24,7 @@
#include <rocblas.h>
#include <migraphx/gpu/gemm_impl.hpp>
#include <migraphx/reduce_dims.hpp>
#include <migraphx/permutation.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......@@ -67,6 +68,19 @@ void blas_shape(const shape& s)
MIGRAPHX_THROW("GPU_GEMM: Batch dimension is not collapsible");
}
shape transpose_batch(const shape& s, unsigned trans_batch)
{
if(trans_batch == 0)
return s;
if(s.lens().size() < 3)
return s;
auto batch = s.lens().size() - 3;
std::vector<int64_t> perm(s.lens().size());
std::iota(perm.begin(), perm.end(), 0);
std::swap(perm[batch], perm[batch + trans_batch]);
return shape::from_permutation(s.type(), s.lens(), perm);
}
template <class R, class... Ts, class... Us>
R rocblas_invoke(R (*f)(Ts...), Us... xs)
{
......@@ -97,6 +111,12 @@ void gemm_impl(context& ctx,
bool int8_x4_format,
bool compute_fp32)
{
const bool is_3inputs = (args.size() == 4);
if(not is_3inputs)
{
beta = 0;
}
bool transa = is_transposed(args[0].get_shape());
bool transb = is_transposed(args[1].get_shape());
auto n_dim = output_shape.lens().size();
......@@ -105,12 +125,8 @@ void gemm_impl(context& ctx,
rocblas_int lda = args[0].get_shape().strides()[transa ? dim_1 : dim_0];
rocblas_int ldb = args[1].get_shape().strides()[transb ? dim_1 : dim_0];
rocblas_int ldc = args[2].get_shape().strides()[dim_0];
rocblas_int ldd = is_3inputs ? args[3].get_shape().strides()[dim_0] : ldc;
bool is_3inputs = (args.size() == 4);
if(!is_3inputs)
{
beta = 0;
}
rocblas_datatype arg_type = get_type(args[0].get_shape().type());
auto output_type = arg_type;
if(output_type == rocblas_datatype_i8_r)
......@@ -160,8 +176,13 @@ void gemm_impl(context& ctx,
auto num_matrices = std::accumulate(
out_lens.rbegin() + 2, out_lens.rend(), std::size_t{1}, std::multiplies<std::size_t>());
if(num_matrices == 1)
if(num_matrices == 1 or (num_matrices > 1 and get_batch_stride(args[1]) == 0))
{
// If the batch dimension of B is broadcasted, then we can
// multiply m by the batch_size and use rocblas_gemm_ex
// instead of rocblas_gemm_strided_batched_ex.
m *= num_matrices;
// the rocblas_gemm API handles inputs and output matrices as
// column-major format. When doing a C = A * B, we actually do
// C^T = (B^T) * (A^T). That is the reason we input args[1] as
......@@ -186,7 +207,7 @@ void gemm_impl(context& ctx,
ldc,
is_3inputs ? to_pointer(args[3]) : to_pointer(args[2]),
output_type,
ldc,
ldd,
compute_type,
rocblas_gemm_algo_standard,
0,
......@@ -197,6 +218,7 @@ void gemm_impl(context& ctx,
auto a_stride = get_batch_stride(args[0]);
auto b_stride = get_batch_stride(args[1]);
auto c_stride = get_batch_stride(args[2]);
auto d_stride = is_3inputs ? get_batch_stride(args[3]) : c_stride;
rocblas_invoke(&rocblas_gemm_strided_batched_ex,
ctx.get_stream().get_rocblas(),
transb ? rocblas_operation_transpose : rocblas_operation_none,
......@@ -220,8 +242,8 @@ void gemm_impl(context& ctx,
c_stride,
is_3inputs ? to_pointer(args[3]) : to_pointer(args[2]),
output_type,
ldc,
c_stride,
ldd,
d_stride,
num_matrices,
compute_type,
rocblas_gemm_algo_standard,
......
src/targets/gpu/include/migraphx/gpu/asin.hpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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_RTGLIB_ASIN_HPP
#define MIGRAPHX_GUARD_RTGLIB_ASIN_HPP
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/device/asin.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct hip_asin : unary_device<hip_asin, device::asin>
{
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/asinh.hpp deleted 100644 → 0
View file @ 6bec381f
/*
* The MIT License (MIT)
*
* Copyright (c) 2015-2022 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_RTGLIB_ASINH_HPP
#define MIGRAPHX_GUARD_RTGLIB_ASINH_HPP
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/device/asinh.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct hip_asinh : unary_device<hip_asinh, device::asinh>
{
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment