Skip to content

GitLab

Commit c4b1102e authored by charlie's avatar charlie
Browse files

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

parents 5fc48e77 31065c7d
Hide whitespace changes
Inline Side-by-side
Showing with 443 additions and 294 deletions
src/driver/verify.cpp
View file @ c4b1102e
......@@ -145,7 +145,7 @@ void verify_reduced(program p,
auto* mm = p.get_main_module();
auto last = std::prev(mm->end(), n + 1);
mm->remove_instructions(last, mm->end());
std::cout << "Verify: " << std::endl;
std::cout << "Verify: " << n << std::endl;
std::cout << p << std::endl;
verify_program(std::to_string(n), p, t, options, quantize, inputs, tolerance);
}
......@@ -159,6 +159,7 @@ void verify_reduced_program(const program& p,
{
const auto* mm = p.get_main_module();
auto n = std::distance(mm->begin(), mm->end());
std::cout << "Verify steps: " << n << std::endl;
for(std::size_t i = 0; i < n; i++)
{
verify_reduced(p, i, t, options, quantize, inputs, tolerance);
......
src/fuse_pointwise.cpp
View file @ c4b1102e
......@@ -39,7 +39,7 @@ static literal get_scalar(instruction_ref ins)
if(ins->name() == "contiguous")
return get_scalar(ins->inputs().front());
const auto& s = ins->get_shape();
if(not(s.elements() == 1 or s.scalar()))
if(s.elements() != 1 && not(s.scalar()))
return {};
if(not ins->can_eval())
return {};
......
src/include/migraphx/common.hpp
View file @ c4b1102e
......@@ -36,6 +36,9 @@ struct operation;
std::vector<std::size_t> compute_broadcasted_lens(std::vector<std::size_t> s0,
std::vector<std::size_t> s1);
std::vector<shape::dynamic_dimension> compute_broadcasted_dyn_dims(shape s0, shape s1);
shape common_shape(const std::vector<shape>& shapes);
instruction_ref insert_common_op(module& m,
......
src/include/migraphx/context.hpp
View file @ c4b1102e
......@@ -66,6 +66,15 @@ any_ptr get_queue_context(T&)
{
return {};
}
template <class T>
void wait_for_context(T&, any_ptr)
{
}
template <class T>
void finish_on_context(T&, any_ptr)
{
}
#ifdef TYPE_ERASED_DECLARATION
......@@ -78,6 +87,10 @@ struct context
void from_value(const value& v);
// (optional)
any_ptr get_queue();
// (optional)
void wait_for(any_ptr queue);
// (optional)
void finish_on(any_ptr queue);
//
void finish() const;
};
......@@ -165,6 +178,18 @@ struct context
return (*this).private_detail_te_get_handle().get_queue();
}
void wait_for(any_ptr queue)
{
assert((*this).private_detail_te_handle_mem_var);
(*this).private_detail_te_get_handle().wait_for(queue);
}
void finish_on(any_ptr queue)
{
assert((*this).private_detail_te_handle_mem_var);
(*this).private_detail_te_get_handle().finish_on(queue);
}
void finish() const
{
assert((*this).private_detail_te_handle_mem_var);
......@@ -187,6 +212,8 @@ struct context
virtual value to_value() const = 0;
virtual void from_value(const value& v) = 0;
virtual any_ptr get_queue() = 0;
virtual void wait_for(any_ptr queue) = 0;
virtual void finish_on(any_ptr queue) = 0;
virtual void finish() const = 0;
};
......@@ -231,6 +258,33 @@ struct context
return get_queue_context(private_detail_te_self);
}
template <class T>
static auto private_detail_te_default_wait_for(char, T&& private_detail_te_self, any_ptr queue)
-> decltype(private_detail_te_self.wait_for(queue))
{
private_detail_te_self.wait_for(queue);
}
template <class T>
static void private_detail_te_default_wait_for(float, T&& private_detail_te_self, any_ptr queue)
{
wait_for_context(private_detail_te_self, queue);
}
template <class T>
static auto private_detail_te_default_finish_on(char, T&& private_detail_te_self, any_ptr queue)
-> decltype(private_detail_te_self.finish_on(queue))
{
private_detail_te_self.finish_on(queue);
}
template <class T>
static void
private_detail_te_default_finish_on(float, T&& private_detail_te_self, any_ptr queue)
{
finish_on_context(private_detail_te_self, queue);
}
template <typename PrivateDetailTypeErasedT>
struct private_detail_te_handle_type : private_detail_te_handle_base_type
{
......@@ -248,7 +302,7 @@ struct context
PrivateDetailTypeErasedT value,
typename std::enable_if<not std::is_reference<PrivateDetailTypeErasedU>::value,
int>::type* = nullptr) noexcept
: private_detail_te_value(std::move(value))
: private_detail_te_value(value)
{
}
......@@ -277,6 +331,18 @@ struct context
return private_detail_te_default_get_queue(char(0), private_detail_te_value);
}
void wait_for(any_ptr queue) override
{
private_detail_te_default_wait_for(char(0), private_detail_te_value, queue);
}
void finish_on(any_ptr queue) override
{
private_detail_te_default_finish_on(char(0), private_detail_te_value, queue);
}
void finish() const override { private_detail_te_value.finish(); }
PrivateDetailTypeErasedT private_detail_te_value;
......
src/targets/gpu/include/migraphx/gpu/leaky_relu.hpp → src/include/migraphx/dyn_output.hpp
View file @ c4b1102e
......@@ -21,44 +21,55 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_RTGLIB_LEAKY_RELU_HPP
#define MIGRAPHX_GUARD_RTGLIB_LEAKY_RELU_HPP
#ifndef MIGRAPHX_GUARD_MIGRAPHLIB_DYN_OUTPUT_HPP
#define MIGRAPHX_GUARD_MIGRAPHLIB_DYN_OUTPUT_HPP
#include <migraphx/op/leaky_relu.hpp>
#include <migraphx/shape.hpp>
#include <migraphx/reflect.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/argument.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct context;
struct dyn_output
{
// original shape from the instruction
shape ins_shape;
// shape computed at eval time using input arguments
shape computed_shape;
};
struct miopen_leaky_relu
/**
* Handle dynamic and static shape at evaluation time.
* If converted to shape type, returns original ins_shape.
* If converted to dyn_output type, will compute an output shape using the input arguments.
*/
template <class F>
struct compute_output_shape
{
op::leaky_relu op;
shared<activation_descriptor> ad;
F ins_inputs;
template <class Self, class F>
static auto reflect(Self& self, F f)
operator dyn_output() const
{
return migraphx::reflect(self.op, f);
return ins_inputs([](const auto& x, shape ins_shape, const std::vector<argument>& inputs) {
if(ins_shape.dynamic())
return dyn_output{ins_shape, compute_shape(x, to_shapes(inputs))};
return dyn_output{ins_shape, ins_shape};
});
}
std::string name() const { return "gpu::leaky_relu"; }
shape compute_shape(const std::vector<shape>& inputs) const;
argument
compute(context& ctx, const shape& output_shape, const std::vector<argument>& args) const;
void finalize(context&, const shape&, const std::vector<shape>&);
std::ptrdiff_t output_alias(const std::vector<shape>& shapes) const
operator shape() const
{
return shapes.size() - 1;
return ins_inputs(
[](const auto&, shape ins_shape, const std::vector<argument>&) { return ins_shape; });
}
};
} // namespace gpu
template <class F>
compute_output_shape<F> make_compute_output_shape(F f)
{
return {f};
}
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/include/migraphx/rewrite_batchnorm.hpp → src/include/migraphx/execution_environment.hpp
View file @ c4b1102e
......@@ -21,28 +21,21 @@
* 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_MIGRAPHLIB_EXECUTION_ENV_HPP
#define MIGRAPHX_GUARD_MIGRAPHLIB_EXECUTION_ENV_HPP
#include <string>
#include <migraphx/instruction_ref.hpp>
#include <migraphx/config.hpp>
#include <migraphx/any_ptr.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
struct module;
/**
* Rewrite batchnorm to a multiply and add.
*/
struct rewrite_batchnorm
struct execution_environment
{
std::string name() const { return "rewrite_batchnorm"; }
void apply(module& m) const;
any_ptr queue = any_ptr{};
bool async = false;
};
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
#endif /* MIGRAPHX_GUARD_MIGRAPHLIB_EXECUTION_ENV_HPP */
src/include/migraphx/op/batch_norm_inference.hpp deleted 100644 → 0
View file @ 5fc48e77
/*
* 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_OPERATORS_BATCH_NORM_HPP
#define MIGRAPHX_GUARD_OPERATORS_BATCH_NORM_HPP
#include <migraphx/check_shapes.hpp>
#include <migraphx/config.hpp>
#include <cmath>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
struct batch_norm_inference
{
float epsilon = 1.0e-6f;
float momentum = 0.9f;
std::string name() const { return "batch_norm_inference"; }
enum bn_infer_mode_t
{
per_activation,
spatial,
};
bn_infer_mode_t bn_mode = spatial;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return pack(
f(self.epsilon, "epsilon"), f(self.momentum, "momentum"), f(self.bn_mode, "bn_mode"));
}
shape compute_shape(std::vector<shape> inputs) const
{
check_shapes{inputs, *this}.has(5);
check_shapes{inputs.data(), inputs.data() + 1, *this}.same_ndims();
check_shapes{inputs.data() + 1, inputs.data() + inputs.size(), *this}.same_shape();
return inputs.front();
}
};
} // namespace op
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
#endif
src/include/migraphx/op/binary.hpp
View file @ c4b1102e
......@@ -28,6 +28,7 @@
#include <migraphx/check_shapes.hpp>
#include <migraphx/argument.hpp>
#include <migraphx/value.hpp>
#include <migraphx/dyn_output.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......@@ -60,10 +61,19 @@ struct binary : op_name<Derived>
value attributes() const { return base_attributes(); }
shape compute_shape(std::vector<shape> inputs) const
{
check_shapes{inputs, static_cast<const Derived&>(*this)}.has(2).same_type().same_dims();
check_shapes{inputs, static_cast<const Derived&>(*this), true}
.has(2)
.same_type()
.same_dims();
auto s0 = inputs.at(0);
auto s1 = inputs.at(1);
if(s0 == s1 and s0.packed())
if(s0.dynamic() or s1.dynamic())
{
if(s0 == s1)
return s0;
MIGRAPHX_THROW("BINARY: " + point_function() + ": fixed-dyn shape for inputs");
}
else if(s0 == s1 and s0.packed())
{
return s0;
}
......@@ -81,9 +91,9 @@ struct binary : op_name<Derived>
}
}
argument compute(const shape& output_shape, std::vector<argument> args) const
argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
{
argument result{output_shape};
argument result{dyn_out.computed_shape};
visit_all(result, args[0], args[1])([&](auto output, auto input1, auto input2) {
std::transform(input1.begin(),
input1.end(),
......
src/include/migraphx/op/broadcast.hpp
View file @ c4b1102e
......@@ -27,23 +27,30 @@
#include <migraphx/check_shapes.hpp>
#include <migraphx/argument.hpp>
#include <migraphx/config.hpp>
#include <migraphx/dyn_output.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
/// The broadcast operator performs the numpy-style broadcasting of an axis of a given tensor. This
/// is achieved primarily by setting the stride of the broadcasted axis to zero. Linear indicies are
/// computed from multi-indicies by computing the inner product on the multi-index with the strides.
/// For example, if we have a tensor A(2,3) it has lengths of (2,3) and strides of (3,1). If we want
/// to compute the linear offset that corresponds to the element on the 2nd row (i = 1) and 3rd
/// column (j = 2), we compute the following inner product (1,2) dot (3, 1) = 1*3 + 2*1 = 5. It is
/// obvious from there that we can negate the effects of a given axis by setting the stride of that
/// axis to zero.
/**
* 1 input version:
* Broadcasts a tensor from the original shape to the broadcast_lens by setting the stride of
* broadcasted dimensions to zero. `axis` attribute for a 1D input shape is the output dimension
* that stays the same. ex: broadcasting shape [1024] -> [4, 1024, 3] has axis = 1 For higher rank
* input shapes, axis is an offset parameter for the broadcasting. Such that this operator would
* work in the opposite direction of NumPy broadcasting. ex: broadcasting shape [2, 2] -> [2, 2, 3]
* with axis = 0
*
* 2 input version:
* Broadcast the first input 1D shape into the second input shape based on the axis parameter.
* Handles broadcasting a 1D static shape into a higher rank dynamic shape.
* broadcast_lens is not used
*/
struct broadcast
{
uint64_t axis = 0;
std::vector<std::size_t> broadcast_lens;
uint64_t axis = 0;
std::vector<std::size_t> broadcast_lens = {};
template <class Self, class F>
static auto reflect(Self& self, F f)
......@@ -54,37 +61,88 @@ struct broadcast
std::string name() const { return "broadcast"; }
shape compute_shape(std::vector<shape> inputs) const
{
auto input = inputs.at(0);
auto t = input.type();
std::vector<size_t> bcast_strides(broadcast_lens.size(), 0);
// the broacast op is deprecated now, so not handling the negative
// value of axis anymore
if(axis >= broadcast_lens.size())
check_shapes{inputs, *this, true}.has(1, 2);
auto s0 = inputs.at(0);
auto t = s0.type();
if(inputs.size() == 1)
{
MIGRAPHX_THROW("BROADCAST : axis is out of range");
}
// the ONNX broadcast op is deprecated now, so not handling the negative
// value of axis anymore
if(axis >= broadcast_lens.size())
{
MIGRAPHX_THROW("BROADCAST : axis " + migraphx::to_string(axis) +
" is out of range");
}
if(broadcast_lens.size() - axis < s0.lens().size())
{
MIGRAPHX_THROW("BROADCAST: (broadcast ndims - axis) is less than s0 ndims");
}
if(not std::equal(s0.lens().begin(), s0.lens().end(), broadcast_lens.begin() + axis))
{
MIGRAPHX_THROW("BROADCAST: when broadcasting, succeeding sizes must match");
}
if(broadcast_lens.size() - axis < input.lens().size())
{
MIGRAPHX_THROW("BROADCAST: (broadcast ndims - axis) is less than input ndims");
std::vector<size_t> bcast_strides(broadcast_lens.size(), 0);
std::copy(s0.strides().begin(), s0.strides().end(), bcast_strides.begin() + axis);
shape output{t, broadcast_lens, std::move(bcast_strides)};
if(output.elements() < s0.elements())
{
// don't think this can occur?
MIGRAPHX_THROW("BROADCAST: output size must be greater than or equal to s0 size");
}
return output;
}
if(not std::equal(input.lens().begin(), input.lens().end(), broadcast_lens.begin() + axis))
else
{
MIGRAPHX_THROW("BROADCAST: when broadcasting, succeeding sizes must match");
}
std::copy(input.strides().begin(), input.strides().end(), bcast_strides.begin() + axis);
// two inputs
auto s1 = inputs.at(1);
if(s0.dynamic())
{
MIGRAPHX_THROW("BROADCAST_2in: s0 is a dynamic shape, does not handle broadcasting "
"a dynamic shape");
}
if(s0.ndim() != 1)
{
MIGRAPHX_THROW("BROADCAST_2in: s0 has ndim " + migraphx::to_string(s0.ndim()) +
", only handle ndim = 1");
}
if(axis >= s1.ndim())
{
MIGRAPHX_THROW("BROADCAST_2in: axis " + migraphx::to_string(axis) +
" is out of range");
}
if(s1.dynamic())
{
s0 = s0.to_dynamic();
if(s0.dyn_dims()[0] != s1.dyn_dims()[axis])
{
MIGRAPHX_THROW("BROADCAST_2in: s0 length doesn't match with dynamic s1 axis "
"dimension length (" +
migraphx::to_string(s0.dyn_dims()[0]) +
" != " + migraphx::to_string(s1.dyn_dims()[axis]) + ")");
}
return s1;
}
shape output{t, broadcast_lens, std::move(bcast_strides)};
if(output.elements() < input.elements())
MIGRAPHX_THROW("BROADCAST: output size must be greater than or equal to input size");
return output;
if(s0.lens()[0] != s1.lens()[axis])
{
MIGRAPHX_THROW("BROADCAST_2in: s0 length doesn't match with static s1 axis "
"dimension length (" +
migraphx::to_string(s0.dyn_dims()[0]) +
" != " + migraphx::to_string(s1.dyn_dims()[axis]) + ")");
}
std::vector<size_t> bcast_strides(s1.ndim(), 0);
std::copy(s0.strides().begin(), s0.strides().end(), bcast_strides.begin() + axis);
shape output{t, s1.lens(), std::move(bcast_strides)};
return output;
}
}
argument compute(shape output_shape, std::vector<argument> args) const
argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
{
return args[0].reshape(output_shape);
return args[0].reshape(dyn_out.computed_shape);
}
std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 0; }
};
......
src/include/migraphx/op/common.hpp
View file @ c4b1102e
......@@ -33,7 +33,7 @@ namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
// Padding mode is default_ for all constant padding.
// Padding mode is default_ for fixed shape padding.
// same_lower and same_upper used for dynamic padding.
enum padding_mode_t
{
......
src/include/migraphx/op/convolution.hpp
View file @ c4b1102e
......@@ -58,8 +58,8 @@ struct convolution
void check_attribute_size() const
{
if(not((padding.size() == stride.size() or (padding.size() / 2) == stride.size()) and
stride.size() == dilation.size()))
if((padding.size() != stride.size() and (padding.size() / 2) != stride.size()) or
stride.size() != dilation.size())
{
MIGRAPHX_THROW("CONVOLUTION: inconsistent attribute sizes");
}
......@@ -74,7 +74,8 @@ struct convolution
// num of dims of input and attribute should match
const auto input_size = inputs[0].max_lens().size();
const auto padding_size = padding.size();
if(not(input_size == padding_size / 2 + 2 or input_size == padding_size + 2))
if(input_size != padding_size / 2 + 2 && input_size != padding_size + 2)
{
MIGRAPHX_THROW("CONVOLUTION: input and attribute size mismatch!");
}
......
src/include/migraphx/op/deconvolution.hpp
View file @ c4b1102e
......@@ -61,8 +61,8 @@ struct deconvolution
void check_attribute_size() const
{
if(not((padding.size() == stride.size() or (padding.size() / 2) == stride.size()) and
stride.size() == dilation.size()))
if((padding.size() != stride.size() and (padding.size() / 2) != stride.size()) or
stride.size() != dilation.size())
{
MIGRAPHX_THROW("deconvolution: inconsistent attribute sizes");
}
......
src/include/migraphx/op/elu.hpp
View file @ c4b1102e
......@@ -38,7 +38,7 @@ struct elu : unary<elu>
std::string point_op() const
{
return "${function:where}(${0} > 0, ${0}, ${alpha} * (migraphx::exp(${0}) - 1))";
return "${function:where}(${0} > 0, ${0}, ${alpha} * (${function:exp}(${0}) - 1))";
}
template <class Self, class F>
......
src/include/migraphx/op/multibroadcast.hpp
View file @ c4b1102e
......@@ -26,64 +26,105 @@
#include <migraphx/check_shapes.hpp>
#include <migraphx/argument.hpp>
#include <migraphx/dyn_output.hpp>
#include <migraphx/common.hpp>
#include <migraphx/config.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
/**
* Broadcast multiple dimensions between two tensors.
* Two versions of this operator: one input and two inputs.
* One input version uses output_lens attribute and broadcasts to it.
* Two inputs version broadcasts both inputs to the common shape at evaluation time.
*/
struct multibroadcast
{
std::vector<std::size_t> output_lens;
std::vector<std::size_t> output_lens = {};
// optional attribute
std::vector<shape::dynamic_dimension> output_dyn_dims = {};
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return pack(f(self.output_lens, "out_lens"));
return pack(f(self.output_lens, "out_lens"), f(self.output_dyn_dims, "out_dyn_dims"));
}
std::string name() const { return "multibroadcast"; }
shape compute_shape(std::vector<shape> inputs) const
{
check_shapes{inputs, *this}.has(1);
auto t = inputs.at(0).type();
auto input = inputs.at(0);
check_shapes{inputs, *this, true}.has(1, 2);
if(input.lens().empty())
{
MIGRAPHX_THROW("MULTIBROADCAST: inputs dimensions should be > 0");
}
auto t = inputs.at(0).type();
auto s0 = inputs.at(0);
if(input.lens().size() > output_lens.size())
if(s0.max_lens().empty())
{
MIGRAPHX_THROW("MULTIBROADCAST: inputs dimensions should <= output size");
MIGRAPHX_THROW("MULTIBROADCAST: input dimensions should be > 0");
}
auto offset = output_lens.size() - input.lens().size();
for(std::ptrdiff_t i = input.lens().size() - 1; i >= 0; i--)
auto make_bcast_strides = [&](std::vector<std::size_t> bcast_lens, std::size_t offset) {
std::vector<size_t> bcast_strides(bcast_lens.size(), 0);
for(std::ptrdiff_t i = s0.lens().size() - 1; i >= 0; i--)
{
if(bcast_lens[i + offset] == s0.lens()[i])
{
bcast_strides[i + offset] = s0.strides()[i];
}
}
return bcast_strides;
};
if(inputs.size() == 1)
{
if(output_lens[i + offset] != input.lens()[i] and input.lens()[i] != 1)
if(s0.lens().size() > output_lens.size())
{
MIGRAPHX_THROW("MULTIBROADCAST: input shape {" + to_string_range(input.lens()) +
"} cannot be broadcasted to {" + to_string_range(output_lens) +
"}!");
MIGRAPHX_THROW("MULTIBROADCAST: input dimensions should <= output size");
}
}
std::vector<size_t> bcast_strides(output_lens.size(), 0);
for(std::ptrdiff_t i = input.lens().size() - 1; i >= 0; i--)
auto offset = output_lens.size() - s0.lens().size();
for(std::ptrdiff_t i = s0.lens().size() - 1; i >= 0; i--)
{
if(output_lens[i + offset] != s0.lens()[i] and s0.lens()[i] != 1)
{
MIGRAPHX_THROW("MULTIBROADCAST: input shape {" + to_string_range(s0.lens()) +
"} cannot be broadcasted to {" + to_string_range(output_lens) +
"}!");
}
}
auto bcast_strides = make_bcast_strides(output_lens, offset);
return {t, output_lens, std::move(bcast_strides)};
}
else
{
if(output_lens[i + offset] == input.lens()[i])
// two inputs
auto s1 = inputs.at(1);
if(s0.dynamic() or s1.dynamic())
{
bcast_strides[i + offset] = input.strides()[i];
if(not output_dyn_dims.empty())
{
return {t, output_dyn_dims};
}
return {t, compute_broadcasted_dyn_dims(s0, s1)};
}
else
{
auto bcast_lens = compute_broadcasted_lens(s0.lens(), s1.lens());
auto offset = bcast_lens.size() - s0.lens().size();
auto bcast_strides = make_bcast_strides(bcast_lens, offset);
return {t, std::move(bcast_lens), std::move(bcast_strides)};
}
}
return {t, output_lens, bcast_strides};
}
argument compute(shape output_shape, std::vector<argument> args) const
argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
{
return args[0].reshape(output_shape);
return args[0].reshape(dyn_out.computed_shape);
}
std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 0; }
};
......
src/include/migraphx/op/pooling.hpp
View file @ c4b1102e
......@@ -64,8 +64,8 @@ struct pooling
void check_attribute_size() const
{
if(not((padding.size() == stride.size() or (padding.size() / 2) == stride.size()) and
stride.size() == lengths.size()))
if((padding.size() != stride.size() and (padding.size() / 2) != stride.size()) or
stride.size() != lengths.size())
{
MIGRAPHX_THROW("POOLING: inconsistent attribute sizes");
}
......@@ -83,7 +83,7 @@ struct pooling
size_t kdims = input_lens.size() - 2;
auto input_size = inputs[0].lens().size();
auto padding_size = padding.size();
if(not(input_size == padding_size / 2 + 2 or input_size == padding_size + 2))
if(input_size != padding_size / 2 + 2 and input_size != padding_size + 2)
{
MIGRAPHX_THROW("POOLING: input and attribute size mismatch!");
}
......
src/include/migraphx/op/quant_convolution.hpp
View file @ c4b1102e
......@@ -63,8 +63,8 @@ struct quant_convolution
void check_attribute_size() const
{
if(not((padding.size() == stride.size() or (padding.size() / 2) == stride.size()) and
stride.size() == dilation.size()))
if((padding.size() != stride.size() and (padding.size() / 2) != stride.size()) or
stride.size() != dilation.size())
{
MIGRAPHX_THROW("QUANT_CONVOLUTION: inconsistent attribute sizes");
}
......
src/include/migraphx/op/squeeze.hpp
View file @ c4b1102e
......@@ -29,6 +29,7 @@
#include <migraphx/config.hpp>
#include <migraphx/value.hpp>
#include <migraphx/op/normalize_attribute.hpp>
#include <migraphx/dyn_output.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......@@ -54,52 +55,90 @@ struct squeeze
std::string name() const { return "squeeze"; }
shape normalize_compute_shape(std::vector<shape> inputs) const
{
check_shapes{inputs, *this}.has(1);
check_shapes{inputs, *this, true}.has(1);
auto input_shape = inputs[0];
auto type = input_shape.type();
auto old_lens = input_shape.lens();
auto old_strides = input_shape.strides();
if(std::any_of(axes.begin(), axes.end(), [&](auto axis) { return old_lens[axis] != 1; }))
if(input_shape.dynamic())
{
MIGRAPHX_THROW("squeeze axis dimension should be equal to 1");
}
std::vector<std::size_t> new_lens;
std::vector<std::size_t> new_strides;
if(axes.empty())
{
for(auto i : range(old_lens.size()))
std::vector<shape::dynamic_dimension> one_dyn_dims{{1, 1, 0}, {1, 1, 1}};
if(std::any_of(axes.begin(), axes.end(), [&](auto axis) {
return not contains(one_dyn_dims, input_shape.dyn_dims()[axis]);
}))
{
MIGRAPHX_THROW(
"SQUEEZE: dynamic axis dimension should be equal to {1, 1, 0} or {1, 1, 1}");
}
std::vector<shape::dynamic_dimension> dyn_dims = {};
if(axes.empty())
{
if(old_lens[i] != 1)
for(auto i : range(input_shape.ndim()))
{
new_lens.push_back(old_lens[i]);
new_strides.push_back(old_strides[i]);
auto dd = input_shape.dyn_dims()[i];
if(not contains(one_dyn_dims, dd))
{
dyn_dims.push_back(dd);
}
}
}
}
else
{
for(auto i : range(old_lens.size()))
else
{
if(std::find(axes.begin(), axes.end(), i) == axes.end())
for(auto i : range(input_shape.ndim()))
{
new_lens.push_back(old_lens[i]);
new_strides.push_back(old_strides[i]);
if(std::find(axes.begin(), axes.end(), i) == axes.end())
{
dyn_dims.push_back(input_shape.dyn_dims()[i]);
}
}
}
}
if(new_lens.empty())
{
return shape{type};
return {input_shape.type(), dyn_dims};
}
else
{
return shape{type, new_lens, new_strides};
auto type = input_shape.type();
auto old_lens = input_shape.lens();
auto old_strides = input_shape.strides();
if(std::any_of(
axes.begin(), axes.end(), [&](auto axis) { return old_lens[axis] != 1; }))
{
MIGRAPHX_THROW("SQUEEZE: static axis dimension should be equal to 1");
}
std::vector<std::size_t> new_lens;
std::vector<std::size_t> new_strides;
if(axes.empty())
{
for(auto i : range(old_lens.size()))
{
if(old_lens[i] != 1)
{
new_lens.push_back(old_lens[i]);
new_strides.push_back(old_strides[i]);
}
}
}
else
{
for(auto i : range(old_lens.size()))
{
if(std::find(axes.begin(), axes.end(), i) == axes.end())
{
new_lens.push_back(old_lens[i]);
new_strides.push_back(old_strides[i]);
}
}
}
if(new_lens.empty())
{
return shape{type};
}
else
{
return shape{type, new_lens, new_strides};
}
}
}
argument compute(shape output_shape, std::vector<argument> args) const
argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
{
return args[0].reshape(output_shape);
return args[0].reshape(dyn_out.computed_shape);
}
std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 0; }
};
......
src/include/migraphx/op/unary.hpp
View file @ c4b1102e
......@@ -30,7 +30,7 @@
#include <migraphx/argument.hpp>
#include <migraphx/stringutils.hpp>
#include <migraphx/value.hpp>
#include <migraphx/operation.hpp>
#include <migraphx/dyn_output.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......
src/include/migraphx/op/unsqueeze.hpp
View file @ c4b1102e
......@@ -29,11 +29,20 @@
#include <migraphx/config.hpp>
#include <migraphx/value.hpp>
#include <migraphx/op/normalize_attribute.hpp>
#include <migraphx/dyn_output.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace op {
/**
* Adds dimensions to a tensor based on the axes attribute.
* `axes` are based on the number of output shape dimensions and should not contain duplicates.
* `steps` are for modifying dimensions added to the middle of the original shape.
* Each step must be a factor of the original dimension.
* ex: unsqueeze(shape = [3, 4, 10], axes = [2, 4, 5], steps = [2]) -> shape = [3, 4, 2, 5, 1, 1]
* Dynamic shape version does not handle `steps`.
*/
struct unsqueeze
{
std::vector<int64_t> axes;
......@@ -56,63 +65,89 @@ struct unsqueeze
std::string name() const { return "unsqueeze"; }
shape normalize_compute_shape(std::vector<shape> inputs) const
{
check_shapes{inputs, *this}.has(1);
check_shapes{inputs, *this, true}.has(1);
auto input_shape = inputs[0];
auto type = input_shape.type();
auto old_lens = input_shape.lens();
auto old_strides = input_shape.strides();
if(input_shape.scalar())
if(input_shape.dynamic())
{
if(old_lens.size() == 1 and old_lens.front() == 1)
return shape{type, old_lens};
else
MIGRAPHX_THROW("UNSQUEEZE: Input must be a scalar");
if(not steps.empty())
{
MIGRAPHX_THROW("UNSQUEEZE_dyn: nonempty steps attribute");
}
std::vector<shape::dynamic_dimension> dyn_dims = {};
auto new_ndim = input_shape.ndim() + axes.size();
std::size_t k = 0;
for(auto i : range(new_ndim))
{
if(std::find(axes.begin(), axes.end(), i) != axes.end())
{
dyn_dims.push_back({1, 1, 0});
}
else
{
dyn_dims.push_back(input_shape.dyn_dims().at(k++));
}
}
return {input_shape.type(), dyn_dims};
}
else
{
auto type = input_shape.type();
auto old_lens = input_shape.lens();
auto old_strides = input_shape.strides();
if(input_shape.scalar())
{
if(old_lens.size() == 1 and old_lens.front() == 1)
return shape{type, old_lens};
else
MIGRAPHX_THROW("UNSQUEEZE: Input must be a scalar");
}
if(steps.size() > axes.size())
MIGRAPHX_THROW("UNSQUEEZE: Steps provided with no axis");
if(steps.size() > axes.size())
MIGRAPHX_THROW("UNSQUEEZE: Steps provided with no axis");
std::size_t new_size = old_lens.size() + axes.size();
std::size_t new_size = old_lens.size() + axes.size();
std::vector<std::size_t> new_lens(new_size);
std::vector<std::size_t> new_strides(new_size);
std::size_t p = 0;
for(auto i : range(new_size))
{
auto axis_idx = std::find(axes.begin(), axes.end(), i) - axes.begin();
if(axis_idx < axes.size())
std::vector<std::size_t> new_lens(new_size);
std::vector<std::size_t> new_strides(new_size);
std::size_t p = 0;
for(auto i : range(new_size))
{
std::int64_t step = 1;
if(axis_idx < steps.size())
step = steps[axis_idx];
if(step == 0)
MIGRAPHX_THROW("UNSQUEEZE: step must be non-zero");
new_lens[i] = step;
if(p < old_strides.size())
auto axis_idx = std::find(axes.begin(), axes.end(), i) - axes.begin();
if(axis_idx < axes.size())
{
if((old_lens[p] % step) != 0)
MIGRAPHX_THROW("UNSQUEEZE: Axis dimenstion is not divisible by step");
old_lens[p] /= step;
new_strides[i] = old_strides[p] * old_lens[p];
std::int64_t step = 1;
if(axis_idx < steps.size())
step = steps[axis_idx];
if(step == 0)
MIGRAPHX_THROW("UNSQUEEZE: step must be non-zero");
new_lens[i] = step;
if(p < old_strides.size())
{
if((old_lens[p] % step) != 0)
MIGRAPHX_THROW("UNSQUEEZE: Axis dimenstion is not divisible by step");
old_lens[p] /= step;
new_strides[i] = old_strides[p] * old_lens[p];
}
else
{
if(step != 1)
MIGRAPHX_THROW("UNSQUEEZE: Step must be 1 for extra axes");
new_strides[i] = 1;
}
}
else
{
if(step != 1)
MIGRAPHX_THROW("UNSQUEEZE: Step must be 1 for extra axes");
new_strides[i] = 1;
new_lens[i] = old_lens[p];
new_strides[i] = old_strides[p++];
}
}
else
{
new_lens[i] = old_lens[p];
new_strides[i] = old_strides[p++];
}
return shape{type, new_lens, new_strides};
}
return shape{type, new_lens, new_strides};
}
argument compute(shape output_shape, std::vector<argument> args) const
argument compute(const dyn_output& dyn_out, std::vector<argument> args) const
{
return args[0].reshape(output_shape);
return args[0].reshape(dyn_out.computed_shape);
}
std::ptrdiff_t output_alias(const std::vector<shape>&) const { return 0; }
};
......
src/include/migraphx/operation.hpp
View file @ c4b1102e
......@@ -32,6 +32,7 @@
#include <utility>
#include <unordered_map>
#include <migraphx/reflect.hpp>
#include <migraphx/dyn_output.hpp>
#include <migraphx/functional.hpp>
#include <migraphx/streamutils.hpp>
#include <migraphx/normalize_attributes.hpp>
......@@ -95,46 +96,6 @@ bool has_finalize(const operation& x);
#else
struct dyn_output
{
// original shape from the instruction
shape ins_shape;
// shape computed at eval time using input arguments
shape computed_shape;
};
/**
* Handle dynamic and static shape at evaluation time.
* If converted to shape type, returns original ins_shape.
* If converted to dyn_output type, will compute an output shape using the input arguments.
*/
template <class F>
struct compute_output_shape
{
F ins_inputs;
operator dyn_output() const
{
return ins_inputs([](const auto& x, shape ins_shape, const std::vector<argument>& inputs) {
if(ins_shape.dynamic())
return dyn_output{ins_shape, compute_shape(x, to_shapes(inputs))};
return dyn_output{ins_shape, ins_shape};
});
}
operator shape() const
{
return ins_inputs(
[](const auto&, shape ins_shape, const std::vector<argument>&) { return ins_shape; });
}
};
template <class F>
compute_output_shape<F> make_compute_output_shape(F f)
{
return {f};
}
namespace detail {
namespace operation_operators {
......
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