Skip to content

GitLab

Commit 870a396b authored by Khalique Ahmed's avatar Khalique Ahmed
Browse files

manual merge

parents 228b665c d309e02f
Hide whitespace changes
Inline Side-by-side
Showing with 424 additions and 1015 deletions
src/targets/gpu/device/where.cpp deleted 100644 → 0
View file @ 228b665c
/*
* 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/elu.cpp deleted 100644 → 0
View file @ 228b665c
/*
* 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/elu.hpp>
#include <migraphx/gpu/context.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
shape miopen_elu::compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, *this}.has(2).not_broadcasted();
return inputs.at(1);
}
argument miopen_elu::compute(context& ctx,
const shape& output_shape,
const std::vector<argument>& args) const
{
float alpha = 1;
float beta = 0;
auto x_desc = make_tensor(args[0].get_shape());
auto y_desc = make_tensor(output_shape);
miopenActivationForward(ctx.get_stream().get_miopen(),
ad.get(),
&alpha,
x_desc.get(),
args[0].implicit(),
&beta,
y_desc.get(),
args[1].implicit());
return args[1];
}
void miopen_elu::finalize(context&, const shape&, const std::vector<shape>&)
{
ad = make_elu(op.alpha);
}
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/fuse_mlir.cpp
View file @ 870a396b
......@@ -49,7 +49,7 @@ struct mlir_conv
std::string name() const { return "gpu::mlir_conv"; }
shape compute_shape(std::vector<shape> inputs, const std::vector<module_ref>& mods) const
{
check_shapes{inputs, *this}.standard();
check_shapes{inputs, *this}.packed_or_broadcasted();
if(mods.size() != 1)
MIGRAPHX_THROW("should have one submodule.");
if(inputs.size() < 2)
......@@ -61,13 +61,28 @@ 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;
// Avoid MLIR assertion: Index < Length && "Invalid index!"
if(ins->get_shape().lens().size() != 4)
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")));
}
......@@ -84,9 +99,10 @@ struct find_conv_pointwise
i.name());
}))
return;
// Only fuse with fp32 for now
// Only fuse with fp32/fp16
if(std::any_of(ins->inputs().begin(), ins->inputs().end(), [&](auto i) {
return i->get_shape().type() != shape::type_t::float_type;
return not contains({shape::type_t::float_type, shape::type_t::half_type},
i->get_shape().type());
}))
return;
std::sort(names.begin(), names.end());
......
src/targets/gpu/fuse_ops.cpp
View file @ 870a396b
......@@ -26,24 +26,9 @@
#include <migraphx/gpu/fuse_ops.hpp>
#include <migraphx/matcher.hpp>
#include <migraphx/gpu/miopen.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>
......@@ -204,10 +189,12 @@ MIGRAPHX_PRED_MATCHER(fusable_conv, instruction_ref ins)
return false;
auto wei = ins->inputs().at(1)->get_shape();
assert(wei.lens().size() == 4);
auto conv = any_cast<miopen_convolution>(ins->get_operator());
if(conv.op.group > 1)
auto miopen_conv_op = ins->get_operator().to_value();
auto algo = miopen_conv_op.at("algo").to<miopenConvFwdAlgorithm_t>();
auto conv_op = from_value<op::convolution>(miopen_conv_op["op"]);
if(conv_op.group > 1)
return false;
if(wei.lens()[1] > 512 and conv.algo != miopenConvolutionFwdAlgoWinograd)
if(wei.lens()[1] > 512 and algo != miopenConvolutionFwdAlgoWinograd)
return false;
// Do not fuse non-symmetric input
......@@ -215,109 +202,14 @@ MIGRAPHX_PRED_MATCHER(fusable_conv, instruction_ref ins)
if(input_lens[2] != input_lens[3] or wei.lens()[2] != wei.lens()[3])
return false;
auto op = conv.op;
// Dont fuse winograd for non-3x3s since there is no fused windograd for those configs
if(conv.algo == miopenConvolutionFwdAlgoWinograd and wei.lens()[2] != 3 and
wei.lens()[3] != 3 and contains({{1, 1}}, op.stride))
if(algo == miopenConvolutionFwdAlgoWinograd and wei.lens()[2] != 3 and wei.lens()[3] != 3 and
contains({{1, 1}}, conv_op.stride))
return false;
return contains({{0, 0, 0, 0}, {1, 1, 1, 1}, {2, 2, 2, 2}}, op.padding) and
contains({{0, 0}, {1, 1}}, op.stride) and contains({{1, 1}}, op.dilation);
return contains({{0, 0, 0, 0}, {1, 1, 1, 1}, {2, 2, 2, 2}}, conv_op.padding) and
contains({{0, 0}, {1, 1}}, conv_op.stride) and contains({{1, 1}}, conv_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_sigmoid, &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>
{
shape compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, *this}.has(4).standard();
return inputs[0];
}
// 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
......@@ -341,256 +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
......@@ -820,7 +462,7 @@ void apply_conv_bias(context& ctx, module& m, const match::matcher_result& r)
auto ins = r.result;
auto input_ins = conv_ins->inputs().at(0);
auto weights_ins = conv_ins->inputs().at(1);
auto conv_op = any_cast<miopen_convolution>(conv_ins->get_operator()).op;
auto conv_op = from_value<op::convolution>((conv_ins->get_operator()).to_value()["op"]);
auto alloc_ins = ins->inputs().back();
auto old_ws_ins = conv_ins->inputs().at(2);
......@@ -886,7 +528,7 @@ struct find_conv_pointwise
auto ins = r.result;
auto input_ins = conv_ins->inputs().at(0);
auto weights_ins = conv_ins->inputs().at(1);
auto conv_op = any_cast<miopen_convolution>(conv_ins->get_operator()).op;
auto conv_op = from_value<op::convolution>(conv_ins->get_operator().to_value()["op"]);
auto alloc_ins = ins->inputs().back();
module_ref pm = ins->module_inputs().front();
......@@ -907,46 +549,6 @@ struct find_conv_pointwise
}
};
struct find_gemm_add
{
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")));
}
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 gemm = any_cast<rocblas_gemm<op::dot>>(gemm_ins->get_operator());
// Already fused gemm
if(not float_equal(gemm.beta, 0))
return;
auto inputs = gemm_ins->inputs();
inputs.pop_back();
auto copy_ins = c_ins;
// Insert copy
if(ins == m.end() or c_ins->outputs().size() > 1 or c_ins->inputs().empty())
{
copy_ins = m.insert_instruction(ins, hip_copy{}, c_ins, ins->inputs().back());
}
inputs.push_back(copy_ins);
inputs.push_back(copy_ins);
gemm.beta = 1;
m.replace_instruction(ins, gemm, inputs);
}
};
struct find_gemm_pointwise
{
auto matcher() const
......@@ -1170,11 +772,9 @@ struct find_layernorm_pointwise
{
auto ins = r.result;
auto layernorm = r.instructions["layernorm"];
auto* pm = ins->module_inputs().front();
if(not layernorm->module_inputs().empty())
return;
auto* pm = ins->module_inputs().front();
auto inputs = layernorm->inputs();
inputs.pop_back();
inputs.insert(inputs.end(), ins->inputs().begin() + 1, ins->inputs().end());
......@@ -1183,29 +783,46 @@ struct find_layernorm_pointwise
}
};
struct find_concat_pointwise
{
auto matcher() const
{
return precompile_name("pointwise")(
match::arg(0)(precompile_name("concat").bind("concat")));
}
void apply(module& m, const match::matcher_result& r) const
{
auto ins = r.result;
auto concat = r.instructions["concat"];
if(not concat->module_inputs().empty())
return;
// TODO: Handle type conversions
if(ins->get_shape().type() != concat->get_shape().type())
return;
auto* pm = ins->module_inputs().front();
auto inputs = concat->inputs();
inputs.pop_back();
inputs.insert(inputs.end(), ins->inputs().begin() + 1, ins->inputs().end());
auto op = concat->get_operator();
op.from_value({{"additional_args", ins->inputs().size() - 1}, {"ignore_modules", true}});
m.replace_instruction(ins, op, 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_concat_pointwise{},
find_gemm_pointwise{},
find_contiguous_tranpose_gemm{},
find_commutative_broadcast{});
......
src/targets/gpu/gemm_impl.cpp
View file @ 870a396b
......@@ -21,7 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <rocblas.h>
#include <rocblas/rocblas.h>
#include <migraphx/gpu/gemm_impl.hpp>
#include <migraphx/reduce_dims.hpp>
#include <migraphx/permutation.hpp>
......
src/targets/gpu/hip.cpp
View file @ 870a396b
......@@ -183,8 +183,8 @@ argument register_on_gpu(const argument& arg)
{
auto arg_shared = arg.share();
auto p = register_on_gpu(arg_shared.data(), arg_shared.get_shape().bytes());
return {arg_shared.get_shape(),
[p, a = std::move(arg_shared)]() mutable { return get_device_ptr(p.get()); }};
auto s = arg_shared.get_shape();
return {s, [p, a = std::move(arg_shared)]() mutable { return get_device_ptr(p.get()); }};
}
argument to_gpu(const argument& arg, bool host)
......@@ -196,12 +196,21 @@ argument to_gpu(const argument& arg, bool host)
argument from_gpu(const argument& arg)
{
argument result;
arg.visit([&](auto x) {
using type = typename decltype(x)::value_type;
auto v = read_from_gpu<type>(arg.data(), x.get_shape().bytes() / sizeof(type));
// cppcheck-suppress returnDanglingLifetime
result = {x.get_shape(), [v]() mutable { return v.data(); }};
});
arg.visit(
[&](auto x) {
using type = typename decltype(x)::value_type;
auto v = read_from_gpu<type>(arg.data(), x.get_shape().bytes() / sizeof(type));
// cppcheck-suppress returnDanglingLifetime
result = {x.get_shape(), [v]() mutable { return v.data(); }};
},
[&](const auto& xs) {
std::vector<argument> args;
std::transform(xs.begin(), xs.end(), std::back_inserter(args), [&](auto x) {
return from_gpu(x);
});
result = argument{args};
});
return result;
}
......
src/targets/gpu/include/migraphx/gpu/acos.hpp deleted 100644 → 0
View file @ 228b665c
/*
* 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_ACOS_HPP
#define MIGRAPHX_GUARD_RTGLIB_ACOS_HPP
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/device/acos.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct hip_acos : unary_device<hip_acos, device::acos>
{
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/acosh.hpp deleted 100644 → 0
View file @ 228b665c
/*
* 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_ACOSH_HPP
#define MIGRAPHX_GUARD_RTGLIB_ACOSH_HPP
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/device/acosh.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct hip_acosh : unary_device<hip_acosh, device::acosh>
{
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/asin.hpp deleted 100644 → 0
View file @ 228b665c
/*
* 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 @ 228b665c
/*
* 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
src/targets/gpu/include/migraphx/gpu/atan.hpp deleted 100644 → 0
View file @ 228b665c
/*
* 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_ATAN_HPP
#define MIGRAPHX_GUARD_RTGLIB_ATAN_HPP
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/device/atan.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct hip_atan : unary_device<hip_atan, device::atan>
{
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/atanh.hpp deleted 100644 → 0
View file @ 228b665c
/*
* 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_ATANH_HPP
#define MIGRAPHX_GUARD_RTGLIB_ATANH_HPP
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/device/atanh.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct hip_atanh : unary_device<hip_atanh, device::atanh>
{
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/ceil.hpp deleted 100644 → 0
View file @ 228b665c
/*
* 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_CEIL_HPP
#define MIGRAPHX_GUARD_RTGLIB_CEIL_HPP
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/device/ceil.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct hip_ceil : unary_device<hip_ceil, device::ceil>
{
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/clip.hpp deleted 100644 → 0
View file @ 228b665c
/*
* 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_CLIP_HPP
#define MIGRAPHX_GUARD_RTGLIB_CLIP_HPP
#include <migraphx/argument.hpp>
#include <migraphx/reflect.hpp>
#include <migraphx/op/clip.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
struct hip_clip
{
op::clip op;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return migraphx::reflect(self.op, f);
}
std::string name() const { return "gpu::clip"; }
shape compute_shape(std::vector<shape> inputs) const;
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
{
return shapes.size() - 1;
}
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/compile_gen.hpp
View file @ 870a396b
......@@ -36,6 +36,9 @@ inline namespace MIGRAPHX_INLINE_NS {
struct shape;
namespace gpu {
struct context;
namespace gen {
struct vectorize
......@@ -43,6 +46,10 @@ struct vectorize
std::size_t size = 1;
std::size_t axis = 0;
static vectorize elements(std::size_t axis, const std::vector<shape>& inputs);
static vectorize elements(context& ctx, std::size_t axis, const std::vector<shape>& inputs);
static vectorize elements(std::size_t axis,
const std::vector<shape>& inputs,
const std::vector<std::size_t>& sizes);
std::string str() const;
};
struct preload
......
src/include/migraphx/rewrite_batchnorm.hpp → src/targets/gpu/include/migraphx/gpu/compile_miopen.hpp
View file @ 870a396b
......@@ -21,28 +21,31 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_RTGLIB_FWD_CONV_BATCHNORM_REWRITE_HPP
#define MIGRAPHX_GUARD_RTGLIB_FWD_CONV_BATCHNORM_REWRITE_HPP
#ifndef MIGRAPHX_GUARD_GPU_COMPILE_MIOPEN_HPP
#define MIGRAPHX_GUARD_GPU_COMPILE_MIOPEN_HPP
#include <string>
#include <migraphx/instruction_ref.hpp>
#include <migraphx/config.hpp>
#include <migraphx/instruction_ref.hpp>
#include <string>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
struct module;
struct context;
struct operation;
/**
* Rewrite batchnorm to a multiply and add.
*/
struct rewrite_batchnorm
namespace gpu {
struct compile_miopen
{
std::string name() const { return "rewrite_batchnorm"; }
context* ctx = nullptr;
std::string name() const { return "gpu::compile_miopen"; }
void apply(module& m) const;
std::size_t compile(operation& op, instruction_ref ins, bool format) const;
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
#endif // MIGRAPHX_GUARD_GPU_COMPILE_MIOPEN_HPP
src/targets/gpu/include/migraphx/gpu/concat.hpp deleted 100644 → 0
View file @ 228b665c
/*
* 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_CONCAT_HPP
#define MIGRAPHX_GUARD_RTGLIB_CONCAT_HPP
#include <migraphx/argument.hpp>
#include <migraphx/reflect.hpp>
#include <migraphx/op/concat.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
struct hip_concat
{
op::concat op;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return migraphx::reflect(self.op, f);
}
std::string name() const { return "gpu::concat"; }
shape compute_shape(std::vector<shape> inputs) const;
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
{
return shapes.size() - 1;
}
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/targets/gpu/include/migraphx/gpu/concat_gpu_opt.hpp
View file @ 870a396b
......@@ -24,8 +24,9 @@
#ifndef MIGRAPHX_GUARD_RTGLIB_CONCAT_GPU_OPT_HPP
#define MIGRAPHX_GUARD_RTGLIB_CONCAT_GPU_OPT_HPP
#include <migraphx/gpu/concat.hpp>
#include <migraphx/op/concat.hpp>
#include <migraphx/operation.hpp>
#include <migraphx/serialize.hpp>
namespace migraphx {
namespace gpu {
......@@ -36,7 +37,8 @@ struct concat_gpu_optimization
std::string allocate() const { return "hip::allocate"; }
migraphx::op::concat get_concat(const migraphx::operation& op) const
{
return migraphx::any_cast<migraphx::gpu::hip_concat>(op).op;
auto v = op.to_value();
return from_value<migraphx::op::concat>(v.at("op"));
}
};
......
src/targets/gpu/include/migraphx/gpu/context.hpp
View file @ 870a396b
......@@ -197,7 +197,9 @@ struct hip_device
struct context
{
context(std::size_t device_id = 0, std::size_t n = value_of(MIGRAPHX_NSTREAMS{}, 1))
: current_device(std::make_shared<hip_device>(device_id, n))
: current_device(std::make_shared<hip_device>(device_id, n)),
begin_event(create_event()),
finish_event(create_event())
{
}
......@@ -274,6 +276,24 @@ struct context
this->current_device = std::make_shared<hip_device>(0, n_streams);
}
void wait_for(any_ptr queue)
{
auto status = hipEventRecord(begin_event.get(), queue.get<hipStream_t>());
if(status != hipSuccess)
MIGRAPHX_THROW("failed to record " + hip_error(status));
get_stream().wait(begin_event.get());
}
void finish_on(any_ptr queue)
{
get_stream().record(finish_event.get());
auto status = hipStreamWaitEvent(queue.get<hipStream_t>(), finish_event.get(), 0);
if(status != hipSuccess)
MIGRAPHX_THROW("Failed to wait on event " + hip_error(status));
}
any_ptr get_queue() { return get_stream().get(); }
void enable_perf_measurement(bool b = true)
......@@ -316,9 +336,13 @@ struct context
// TODO: Make this a vector to support multiple devices
std::shared_ptr<hip_device> current_device;
std::vector<shared<hip_event_ptr>> events;
bool measure_perf = false;
bool measure_perf = false;
// for event perf timing
shared<hip_event_ptr> start_event = nullptr;
shared<hip_event_ptr> stop_event = nullptr;
// for stream syncronization
shared<hip_event_ptr> begin_event = nullptr;
shared<hip_event_ptr> finish_event = nullptr;
};
inline void migraphx_to_value(value& v, const context& ctx) { v = ctx.to_value(); }
......
src/targets/gpu/include/migraphx/gpu/convolution.hpp
View file @ 870a396b
......@@ -25,45 +25,318 @@
#define MIGRAPHX_GUARD_RTGLIB_CONVOLUTION_HPP
#include <migraphx/shape.hpp>
#include <migraphx/op/convolution.hpp>
#include <migraphx/generate.hpp>
#include <migraphx/operation.hpp>
#include <migraphx/register_op.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/op/identity.hpp>
#include <migraphx/op/convolution.hpp>
#include <migraphx/op/quant_convolution.hpp>
#include <migraphx/op/deconvolution.hpp>
#include <unordered_map>
#include <migraphx/reflect.hpp>
#include <migraphx/gpu/context.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
inline shape reshape_if_1d(const shape& input)
{
shape new_shape{input};
auto dims = new_shape.lens();
if(dims.size() == 3)
{
std::vector<size_t> new_dims = dims;
new_dims.insert(new_dims.begin() + 2, 1);
new_shape = shape{input.type(), new_dims};
}
return new_shape;
}
template <class Op>
struct miopen_convolution
{
op::convolution op;
Op op;
bool int8_x4_format = false;
shared<convolution_descriptor> cd = nullptr;
miopenConvFwdAlgorithm_t algo{};
#ifdef MIGRAPHX_HAS_FIND_2_API
value::binary solution_object{};
shared<miopen_solution> solution_ptr = nullptr;
#endif
uint64_t solution_id = 0;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return pack(f(self.op.padding, "padding"),
f(self.op.stride, "stride"),
f(self.op.dilation, "dilation"),
f(self.op.group, "group"),
f(self.op.padding_mode, "padding_mode"),
return pack(f(self.op, "op"),
#ifdef MIGRAPHX_HAS_FIND_2_API
f(self.solution_object, "solution_object"),
#endif
f(self.algo, "algo"),
f(self.int8_x4_format, "int8_x4_format"),
f(self.solution_id, "solution_id"));
}
std::string name() const { return "gpu::convolution"; }
shape compute_shape(const std::vector<shape>& inputs) const;
std::string name() const { return "gpu::" + op.name(); }
inline shape compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, op}.has(4);
std::vector<shape> conv_inputs(inputs.begin(), inputs.begin() + 2);
check_shapes{conv_inputs, *this}.max_ndims(5).packed_layouts(
{{0, 1, 2}, {0, 1, 2, 3}, {0, 2, 3, 1}, {0, 1, 2, 3, 4}});
return migraphx::compute_shape<Op>(op, conv_inputs);
}
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
shape find(context& ctx, const shape& output_shape, std::vector<shape> inputs);
void finalize(context& ctx, const shape& output_shape, std::vector<shape> inputs);
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const
{
auto x_desc = make_tensor(reshape_if_1d(args[0].get_shape()), int8_x4_format);
auto w_desc = make_tensor(reshape_if_1d(args[1].get_shape()), int8_x4_format);
auto y_desc = make_tensor(reshape_if_1d(output_shape));
auto* miopen_stream_handle = ctx.get_stream().get_miopen();
auto workspace_size = args[2].get_shape().bytes();
#ifdef MIGRAPHX_HAS_FIND_2_API
{
const miopenTensorArgument_t tensor_args[3] = {
{miopenTensorConvolutionX, nullptr, args[0].implicit()},
{miopenTensorConvolutionW, nullptr, args[1].implicit()},
{miopenTensorConvolutionY, nullptr, args[3].implicit()},
};
if(solution_ptr.get() == nullptr)
MIGRAPHX_THROW("MIOpen " + op.name() + " : Load MIOpen Solution before running it");
auto status = miopenRunSolution(miopen_stream_handle,
solution_ptr.get(),
3,
tensor_args,
args[2].implicit(),
workspace_size);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() +
" : running convolution using find_2.0 failed");
return args[3];
}
#else
// else use immediate mode
if(solution_id == 0)
MIGRAPHX_THROW("MIOpen " + op.name() + " : invalid solution ID");
auto status = miopenConvolutionForwardImmediate(miopen_stream_handle,
w_desc.get(),
args[1].implicit(),
x_desc.get(),
args[0].implicit(),
cd.get(),
y_desc.get(),
args[3].implicit(),
args[2].implicit(),
workspace_size,
solution_id);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + ": running convolution failed");
return args[3];
#endif
}
void set_conv_descriptor()
{
cd = (op.name() == "deconvolution") ? make_deconv(op) : make_conv(op);
}
value compile(migraphx::context& ctx, const shape& output, const std::vector<shape>& input)
{
set_conv_descriptor();
auto ws = find(any_cast<migraphx::gpu::context>(ctx), output, input);
return {{"workspace", ws.bytes()}};
}
shape find(context& ctx, const shape& output_shape, const std::vector<shape>& inputs)
{
shape workspace_shape{};
auto x_desc = make_tensor(reshape_if_1d(inputs[0]), int8_x4_format);
auto w_desc = make_tensor(reshape_if_1d(inputs[1]), int8_x4_format);
auto y_desc = make_tensor(reshape_if_1d(output_shape));
std::size_t workspace_size = 0;
#ifdef MIGRAPHX_HAS_FIND_2_API
{
auto conv_problem = make_obj<miopen_problem>(
&miopenCreateConvProblem, cd.get(), miopenProblemDirectionForward);
set_tensor_descriptor(miopenTensorConvolutionX, x_desc, conv_problem);
set_tensor_descriptor(miopenTensorConvolutionW, w_desc, conv_problem);
set_tensor_descriptor(miopenTensorConvolutionY, y_desc, conv_problem);
auto* miopen_stream_handle = ctx.get_stream().get_miopen();
solution_ptr = find_solution(miopen_stream_handle, conv_problem.get());
auto status = miopenGetSolutionWorkspaceSize(solution_ptr.get(), &workspace_size);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen" + op.name() + " : failed to get solution's workspace size");
std::size_t solution_size;
status = miopenGetSolutionSize(solution_ptr.get(), &solution_size);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen" + op.name() + ": Failed to fetch solution size");
auto solution_binary = std::vector<char>{};
solution_binary.resize(solution_size);
status = miopenSaveSolution(solution_ptr.get(), solution_binary.data());
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen" + op.name() + ": Saving solution failed");
solution_object = value::binary{solution_binary.data(), solution_size};
return shape{shape::int8_type, {workspace_size}};
}
#else
auto status = miopenConvolutionForwardGetWorkSpaceSize(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
&workspace_size);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen" + op.name() + " : Failed to get forward workspace size");
workspace_shape = shape{shape::int8_type, {workspace_size}};
auto x_shape = inputs[0];
auto w_shape = inputs[1];
if(int8_x4_format)
{
x_shape = pack_int8_shape(x_shape);
w_shape = pack_int8_shape(w_shape);
}
auto x = to_gpu(generate_argument(x_shape));
auto w = to_gpu(generate_argument(w_shape));
auto y = allocate_gpu(output_shape);
auto workspace = allocate_gpu(workspace_shape);
int algo_count = 1;
miopenConvAlgoPerf_t perf;
status = miopenFindConvolutionForwardAlgorithm(ctx.get_stream().get_miopen(),
x_desc.get(),
x.implicit(),
w_desc.get(),
w.implicit(),
cd.get(),
y_desc.get(),
y.implicit(),
1,
&algo_count,
&perf,
workspace.implicit(),
workspace_size,
false);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + " : find convolution failed");
algo = perf.fwd_algo;
size_t solution_count;
status = miopenConvolutionForwardGetSolutionCount(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
&solution_count);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + ": get solution count failed");
std::vector<miopenConvSolution_t> solutions(solution_count);
status = miopenConvolutionForwardGetSolution(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
solution_count,
&solution_count,
solutions.data());
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + ": get solution failed");
solution_id = solutions.front().solution_id;
return shape{shape::int8_type, {perf.memory}};
#endif
}
void finalize(context& ctx, const shape& output_shape, const std::vector<shape>& inputs)
{
#ifdef MIGRAPHX_HAS_FIND_2_API
{
(void)(ctx); // avoid warnings
(void)(output_shape);
(void)(inputs);
// load solution
if(solution_ptr == nullptr)
{
miopenSolution_t ptr;
auto status =
miopenLoadSolution(&ptr,
reinterpret_cast<const char*>(solution_object.data()),
solution_object.size());
solution_ptr = miopen_solution{ptr};
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen " + op.name() + ": loading convolution solution failed");
}
}
#else
// Use immediate mode API
{
set_conv_descriptor();
if(solution_id == 0)
{
// Check that workspace hasn't changed
auto size = inputs.at(2).bytes();
auto ws = find(ctx, output_shape, inputs);
if(ws.bytes() > size)
MIGRAPHX_THROW("MIOpen " + op.name() +
": workspace has changed during finalization.");
}
auto x_desc = make_tensor(reshape_if_1d(inputs[0]), int8_x4_format);
auto w_desc = make_tensor(reshape_if_1d(inputs[1]), int8_x4_format);
auto y_desc = make_tensor(reshape_if_1d(output_shape));
auto status = miopenConvolutionForwardCompileSolution(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
solution_id);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("MIOpen Convolution: compile solution failed");
}
#endif
}
inline std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
{
return shapes.size() - 1;
}
};
inline shape pack_int8_shape(const shape& s) const
{
if(s.type() != shape::int8_type)
{
return s;
}
auto lens = s.lens();
auto strides = s.strides();
lens[1] = (lens[1] + 3) / 4 * 4;
strides[0] = strides[1] * lens[1];
return {s.type(), lens, strides};
}
};
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
......
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