Skip to content

GitLab

Commit 5a14c0bf authored by umangyadav's avatar umangyadav
Browse files

Merge branch 'develop' into workspace_size

parents cb01e280 5fa42993
Hide whitespace changes
Inline Side-by-side
Showing with 528 additions and 629 deletions
src/targets/gpu/jit/pointwise.cpp
View file @ 5a14c0bf
......@@ -75,20 +75,16 @@ struct pointwise_compiler : compiler<pointwise_compiler>
options.virtual_inputs = reduce_dims(inputs);
options.params = "-Wno-float-equal";
auto axis = find_fast_axis(options.virtual_inputs);
auto vec = vectorize::elements(axis, options.virtual_inputs);
auto preloads = preload::broadcasts(axis, options.virtual_inputs);
auto vec = vectorize::elements(ctx, axis, options.virtual_inputs);
options.kernel_name = v.get("kernel", "kernel");
options.set_launch_params(
v,
compute_global_for(ctx,
options.output.elements() / vec.size,
oversubscribe_if(not preloads.is_preloading())));
v, compute_global_for(ctx, options.output.elements() / vec.size, 256));
auto src = interpolate_string(pointwise_kernel,
{{"kernel", options.kernel_name},
{"params", enum_params(inputs.size(), "void * private_p")},
{"args", enum_params(inputs.size(), "private_p")},
{"lambda", v.at("lambda").to<std::string>()},
{"transformers", make_transformer_args(preloads, vec)},
{"transformers", make_transformer_args(vec)},
{"preamble", v.get("preamble", std::string{})}});
return compile_hip_code_object(src, options);
}
......
src/targets/gpu/jit/reduce.cpp
View file @ 5a14c0bf
......@@ -121,7 +121,7 @@ struct reduce_compiler : compiler<reduce_compiler>
// Vectorize if the axis is a reduction axis
if(options.virtual_inputs.back().lens()[faxis] == 1)
{
vec = vectorize::elements(faxis, options.virtual_inputs);
vec = vectorize::elements(ctx, faxis, options.virtual_inputs);
}
auto relements = get_reduce_elements(options.virtual_inputs) / vec.size;
auto nelements = options.virtual_inputs.back().elements();
......
src/targets/gpu/jit/softmax.cpp
View file @ 5a14c0bf
......@@ -32,6 +32,8 @@ namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_USE_FAST_SOFTMAX)
using namespace migraphx::gpu::gen; // NOLINT
static const char* const softmax_kernel = R"__migraphx__(
......@@ -69,7 +71,7 @@ struct softmax_compiler : compiler<softmax_compiler>
// Vectorize if the axis is a reduction axis
if(faxis == axis)
{
vec = vectorize::elements(faxis, inputs);
vec = vectorize::elements(ctx, faxis, inputs);
}
auto relements = inputs[0].lens()[axis] / vec.size;
auto nelements = (inputs.back().elements() / inputs[0].lens()[axis]);
......@@ -81,6 +83,9 @@ struct softmax_compiler : compiler<softmax_compiler>
options.inputs = inputs;
options.kernel_name = "softmax_kernel";
if(enabled(MIGRAPHX_USE_FAST_SOFTMAX{}))
options.params = "-DMIGRAPHX_USE_FAST_SOFTMAX";
auto src = interpolate_string(
softmax_kernel,
{{"transformers", make_transformer_args(vec)}, {"axis", to_string(axis)}});
......
src/targets/gpu/kernels/include/migraphx/kernels/array.hpp
View file @ 5a14c0bf
......@@ -33,49 +33,95 @@
namespace migraphx {
// NOLINTNEXTLINE
#define MIGRAPHX_DEVICE_ARRAY_OP(op, binary_op) \
template <class U> \
constexpr array& operator op(const array<U, N>& x) \
{ \
for(index_int i = 0; i < N; i++) \
d[i] op x[i]; \
return *this; \
} \
template <class U, MIGRAPHX_REQUIRES(is_convertible<U, T>{})> \
constexpr array& operator op(const U& x) \
{ \
for(index_int i = 0; i < N; i++) \
d[i] op x; \
return *this; \
} \
template <class U> \
friend constexpr auto operator binary_op(const array& x, const array<U, N>& y) \
{ \
array<decltype(T {} binary_op U{}), N> z{}; \
for(index_int i = 0; i < N; i++) \
z[i] = x[i] binary_op y[i]; \
return z; \
} \
template <class U, MIGRAPHX_REQUIRES(is_convertible<U, T>{})> \
friend constexpr auto operator binary_op(const array& x, const U& y) \
{ \
array<decltype(T {} binary_op U{}), N> z{}; \
for(index_int i = 0; i < N; i++) \
z[i] = x[i] binary_op y; \
return z; \
} \
template <class U, MIGRAPHX_REQUIRES(is_convertible<U, T>{})> \
friend constexpr auto operator binary_op(const U& x, const array& y) \
{ \
array<decltype(T {} binary_op U{}), N> z{}; \
for(index_int i = 0; i < N; i++) \
z[i] = x binary_op y[i]; \
return z; \
#define MIGRAPHX_DEVICE_ARRAY_OP(op, binary_op) \
template <class U> \
constexpr array& operator op(const array<U, N>& x) \
{ \
array_detail::array_for_each(*this, x)([](auto& sy, auto sx) { sy op sx; }); \
return *this; \
} \
template <class U, MIGRAPHX_REQUIRES(is_convertible<U, T>{})> \
constexpr array& operator op(const U& x) \
{ \
array_detail::array_for_each (*this)([&](auto& sy) { sy op x; }); \
return *this; \
} \
template <class U> \
friend constexpr auto operator binary_op(const array& x, const array<U, N>& y) \
{ \
array<decltype(T {} binary_op U{}), N> z{}; \
array_detail::array_for_each(z, x, y)( \
[&](auto& sz, auto sx, auto sy) { sz = sx binary_op sy; }); \
return z; \
} \
template <class U, MIGRAPHX_REQUIRES(is_convertible<U, T>{})> \
friend constexpr auto operator binary_op(const array& x, const U& y) \
{ \
array<decltype(T {} binary_op U{}), N> z{}; \
array_detail::array_for_each(z, x)([&](auto& sz, auto sx) { sz = sx binary_op y; }); \
return z; \
} \
template <class U, MIGRAPHX_REQUIRES(is_convertible<U, T>{})> \
friend constexpr auto operator binary_op(const U& x, const array& y) \
{ \
array<decltype(T {} binary_op U{}), N> z{}; \
array_detail::array_for_each(z, y)([&](auto& sz, auto sy) { sz = x binary_op sy; }); \
return z; \
}
namespace array_detail {
template <class T>
constexpr auto is_vectorizable()
{
return not is_same<T, bool>{} and (is_fundamental<T>{} or is_same<T, half>{});
}
template <class T>
__device__ auto& array2vec(T& x)
{
using value_type = typename T::value_type;
constexpr auto size = decltype(x.size()){};
using type = vec<value_type, size>;
if constexpr(is_const<T>{})
return reinterpret_cast<const type&>(x);
else
return reinterpret_cast<type&>(x);
}
template <class T, class... Ts>
constexpr auto array_for_each(T& x, Ts&... xs)
{
MIGRAPHX_ASSERT(((x.size() == xs.size()) and ...));
return [&](auto f) {
constexpr auto size = decltype(x.size()){};
if constexpr((is_vectorizable<typename T::value_type>() or
(is_vectorizable<typename Ts::value_type>() or ...)) and
size <= 8 and size > 1 and (size % 2 == 0))
{
if(__builtin_is_constant_evaluated())
{
for(index_int i = 0; i < size; i++)
f(x[i], xs[i]...);
}
else
{
using vec_type = std::remove_reference_t<decltype(array2vec(x))>;
f(array2vec(x), __builtin_convertvector(array2vec(xs), vec_type)...);
}
}
else
{
for(index_int i = 0; i < size; i++)
f(x[i], xs[i]...);
}
};
}
} // namespace array_detail
template <class T, index_int N>
struct array
{
using value_type = T;
T d[N];
constexpr T& operator[](index_int i)
{
......@@ -108,18 +154,13 @@ struct array
constexpr T dot(const array& x) const
{
T result = 0;
for(index_int i = 0; i < N; i++)
result += x[i] * d[i];
return result;
auto r = x * (*this);
return r.reduce([](auto a, auto b) { return a + b; }, 0);
}
constexpr T product() const
{
T result = 1;
for(index_int i = 0; i < N; i++)
result *= d[i];
return result;
return reduce([](auto x, auto y) { return x * y; }, 1);
}
constexpr T single(index_int width = 100) const
......@@ -134,6 +175,24 @@ struct array
return result;
}
template <class F>
constexpr auto apply(F f) const
{
array<decltype(f(d[0])), N> result;
for(index_int i = 0; i < N; i++)
result[i] = f(d[i]);
return result;
}
template <class F>
constexpr auto reduce(F f, T init) const
{
T result = init;
for(index_int i = 0; i < N; i++)
result = f(result, d[i]);
return result;
}
MIGRAPHX_DEVICE_ARRAY_OP(+=, +)
MIGRAPHX_DEVICE_ARRAY_OP(-=, -)
MIGRAPHX_DEVICE_ARRAY_OP(*=, *)
......@@ -201,6 +260,11 @@ struct array
}
};
template <class T, class... Ts>
constexpr array<T, sizeof...(Ts) + 1> make_array(T x, Ts... xs)
{
return {x, static_cast<T>(xs)...};
}
template <class T, T... Xs>
struct integral_const_array : array<T, sizeof...(Xs)>
{
......
src/targets/gpu/include/migraphx/gpu/max.hpp → src/targets/gpu/kernels/include/migraphx/kernels/concat.hpp
View file @ 5a14c0bf
......@@ -21,22 +21,46 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#ifndef MIGRAPHX_GUARD_RTGLIB_MAX_HPP
#define MIGRAPHX_GUARD_RTGLIB_MAX_HPP
#include <migraphx/gpu/oper.hpp>
#include <migraphx/gpu/device/max.hpp>
#include <migraphx/kernels/index.hpp>
#include <migraphx/kernels/functional.hpp>
#include <migraphx/kernels/tensor_view.hpp>
#ifndef MIGRAPHX_GUARD_KERNELS_CONCAT_HPP
#define MIGRAPHX_GUARD_KERNELS_CONCAT_HPP
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
struct hip_max : binary_device<hip_max, device::max>
template <index_int Axis, class Output, class Input, class Start>
constexpr auto concat_slice(Output out, Input, Start)
{
};
constexpr auto lens = get_shape_c<Input>{}.lens;
constexpr auto strides = get_shape_c<Output>{}.strides;
constexpr auto offset = return_c([] {
constexpr auto output_shape = get_shape_c<Output>{};
return Start{} * output_shape.strides[Axis];
});
constexpr auto s = make_shape(lens, strides);
return make_tensor_view(&out[offset], s);
}
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
template <index_int Axis, class Input>
constexpr auto concat_ends(Input)
{
constexpr auto lens = get_shape_c<Input>{}.lens;
return _c<lens[Axis]>;
}
#endif
template <index_int Axis, class Output, class... Inputs>
__device__ void concat(Output output, Inputs... inputs)
{
auto idx = make_index();
fold([&](auto start, auto input) {
auto y = concat_slice<Axis>(output, input, start);
idx.global_stride(input.get_shape().elements(), [&](auto i) { y[i] = input[i]; });
return start + concat_ends<Axis>(input);
})(_c<0>, inputs...);
}
} // namespace migraphx
#endif // MIGRAPHX_GUARD_KERNELS_CONCAT_HPP
src/targets/gpu/kernels/include/migraphx/kernels/index.hpp
View file @ 5a14c0bf
......@@ -28,9 +28,60 @@
#include <migraphx/kernels/types.hpp>
#include <migraphx/kernels/integral_constant.hpp>
#include <migraphx/kernels/type_traits.hpp>
#include <migraphx/kernels/debug.hpp>
namespace migraphx {
#if defined(MIGRAPHX_NGLOBAL) && defined(MIGRAPHX_NLOCAL)
#define MIGRAPHX_NGROUP ((MIGRAPHX_NGLOBAL + MIGRAPHX_NLOCAL - 1) / MIGRAPHX_NLOCAL)
#endif
inline __device__ __attribute__((const)) index_int compute_global_size()
{
#ifdef MIGRAPHX_NGLOBAL
return MIGRAPHX_NGLOBAL;
#else
// This actualy works even when global is not divisible by local size.
// This doesnt actually do a multiplicatiosn. Instead it calls a device
// function to get the global size, which is why it works.
return blockDim.x * gridDim.x; // NOLINT
#endif
}
// We cant just use blockDim.x to get the local size since its broken on hip
// when global is not divisible by local size. In this case, we calulate the
// size for the last group.
inline __device__ __attribute__((const)) index_int compute_local_size()
{
#ifdef MIGRAPHX_NLOCAL
const auto nlocal = MIGRAPHX_NLOCAL;
#else
const auto nlocal = blockDim.x; // NOLINT
#endif
#ifdef MIGRAPHX_NGROUP
const auto ngroup = MIGRAPHX_NGROUP;
#else
const auto ngroup = gridDim.x; // NOLINT
#endif
const auto group_id = blockIdx.x; // NOLINT
const auto nglobal = compute_global_size();
if(group_id == ngroup - 1)
{
return 1 + (nglobal - 1) % nlocal;
}
else
{
return nlocal; // NOLINT
}
}
#ifdef MIGRAPHX_NGROUP
// If global is divisible by local then local can be a const
#if(MIGRAPHX_NGLOBAL % MIGRAPHX_NLOCAL == 0) || (MIGRAPHX_NGROUP == 1)
#define MIGRAPHX_HAS_CONST_LOCAL 1
#endif
#endif
struct index
{
index_int global = 0;
......@@ -38,20 +89,44 @@ struct index
index_int group = 0;
#ifdef MIGRAPHX_NGLOBAL
constexpr index_constant<MIGRAPHX_NGLOBAL> nglobal() const { return {}; }
constexpr index_constant<MIGRAPHX_NGLOBAL> nglobal() const
{
static_assert(MIGRAPHX_NGLOBAL > 0, "Global size must be greater than 0");
return {};
}
#else
__device__ index_int nglobal() const
{
return blockDim.x * gridDim.x; // NOLINT
MIGRAPHX_ASSERT(compute_global_size() > 0);
return compute_global_size(); // NOLINT
}
#endif
#ifdef MIGRAPHX_NLOCAL
constexpr index_constant<MIGRAPHX_NLOCAL> nlocal() const { return {}; }
#ifdef MIGRAPHX_HAS_CONST_LOCAL
constexpr index_constant<MIGRAPHX_NLOCAL> nlocal() const
{
static_assert(MIGRAPHX_NLOCAL > 0, "Local size must be greater than 0");
return {};
}
#else
__device__ index_int nlocal() const
{
return blockDim.x; // NOLINT
#ifdef MIGRAPHX_NGROUP
static_assert((MIGRAPHX_NGLOBAL % MIGRAPHX_NLOCAL != 0) and (MIGRAPHX_NGROUP > 1),
"Local size should be const");
#endif
MIGRAPHX_ASSERT(compute_local_size() > 0);
return compute_local_size(); // NOLINT
}
#endif
#ifdef MIGRAPHX_NLOCAL
constexpr index_constant<MIGRAPHX_NLOCAL> max_nlocal() const { return {}; }
#else
__device__ index_int max_nlocal() const
{
MIGRAPHX_ASSERT(blockDim.x > 0);
return blockDim.x;
}
#endif
template <class N, class Stride>
......@@ -63,6 +138,7 @@ struct index
template <class F, class N, class Stride>
static constexpr void for_stride(index_int start, N n, Stride stride, F f)
{
MIGRAPHX_ASSERT(start < stride);
if constexpr(not is_integral<N>{} and not is_integral<Stride>{} and
max_stride_iterations(n, stride) == 1)
{
......
src/targets/gpu/kernels/include/migraphx/kernels/layernorm.hpp
View file @ 5a14c0bf
......@@ -29,6 +29,12 @@
namespace migraphx {
template <class T, index_int N, class Op>
constexpr auto vec_reduce(const array<T, N>& a, Op op)
{
return a.apply([&](auto x) { return vec_reduce(x, op); });
}
template <index_int Axis,
class F,
class BinOp,
......@@ -37,46 +43,46 @@ template <index_int Axis,
class Input2,
class... Inputs>
__device__ void generic_binary_layernorm(
F compute, BinOp op, Output output, Input1 input1, Input2 input2, Inputs... inputs)
F compute, BinOp op, float eps, Output output, Input1 input1, Input2 input2, Inputs... inputs)
{
using reduce_output = reduce::with_axis<Input1, Axis>;
reduce::block::run<reduce_output>([&](auto, auto r) {
using value_type = typename Input1::type;
constexpr auto relements = r.template elements<Input1>();
auto mean = [&](auto f) {
return r.reduce(op::sum{}, 0, [&](auto x1, auto x2) {
return f(x1, x2) / value_type{relements};
auto means =
r.reduce(op::sum{}, make_array<vec_type<value_type>>(0, 0), [&](auto x1, auto x2) {
auto x = op(x1, x2);
return make_array(x, x * x) * vec_type<value_type>{1.0 / relements};
})(input1, input2);
};
// mean(x)
auto mean_x = mean(op);
// mean(m ^ 2)
auto mean_m2 = mean([&](auto x1, auto x2) {
auto m = op(x1, x2) - mean_x;
return m * m;
});
auto mean_x = means[0];
auto mean_x2 = means[1];
auto variance = mean_x2 - (mean_x * mean_x);
value_type eps_val = eps; // implicit conversion for eps
r.inner([&](auto& y, auto x1, auto x2, auto... xs) {
auto m = op(x1, x2) - mean_x;
// m * rsqrt(mean(m ^ 2) + 1e-12)
y = compute(m * rsqrt(mean_m2 + value_type{1e-12}), xs...);
auto x = op(x1, x2);
auto m = x - mean_x;
// m * rsqrt(mean(m ^ 2) + epsilon)
y = compute(m * rsqrt(variance + eps_val), xs...);
})(output, input1, input2, inputs...);
});
}
template <index_int Axis, class F, class Output, class Input, class... Inputs>
__device__ void layernorm(F compute, Output output, Input input, Inputs... inputs)
__device__ void layernorm(F compute, float eps, Output output, Input input, Inputs... inputs)
{
generic_binary_layernorm<Axis>(
compute, [](auto x, auto) { return x; }, output, input, input, inputs...);
compute, [](auto x, auto) { return x; }, eps, output, input, input, inputs...);
}
template <index_int Axis, class F, class Output, class Input1, class Input2, class... Inputs>
__device__ void
add_layernorm(F compute, Output output, Input1 input1, Input2 input2, Inputs... inputs)
add_layernorm(F compute, float eps, Output output, Input1 input1, Input2 input2, Inputs... inputs)
{
generic_binary_layernorm<Axis>(
compute, [](auto x1, auto x2) { return x1 + x2; }, output, input1, input2, inputs...);
compute, [](auto x1, auto x2) { return x1 + x2; }, eps, output, input1, input2, inputs...);
}
} // namespace migraphx
......
src/targets/gpu/kernels/include/migraphx/kernels/math.hpp
View file @ 5a14c0bf
......@@ -104,6 +104,7 @@ MIGRAPHX_DEVICE_MATH(floor, ::floor)
MIGRAPHX_DEVICE_MATH(isnan, ::isnan)
MIGRAPHX_DEVICE_MATH(log, ::log)
MIGRAPHX_DEVICE_MATH(pow, ::pow)
MIGRAPHX_DEVICE_MATH(remainder, ::remainder)
MIGRAPHX_DEVICE_MATH(round, ::round)
MIGRAPHX_DEVICE_MATH(rsqrt, ::rsqrt)
MIGRAPHX_DEVICE_MATH(sin, ::sin)
......@@ -111,6 +112,7 @@ MIGRAPHX_DEVICE_MATH(sinh, ::sinh)
MIGRAPHX_DEVICE_MATH(sqrt, ::sqrt)
MIGRAPHX_DEVICE_MATH(tan, ::tan)
MIGRAPHX_DEVICE_MATH(tanh, ::tanh)
MIGRAPHX_DEVICE_MATH(fmod, ::fmod)
// Float overloads
MIGRAPHX_DEVICE_MATH_FOR(float, acos, ::acosf)
......@@ -126,6 +128,7 @@ MIGRAPHX_DEVICE_MATH_FOR(float, sin, ::sinf)
MIGRAPHX_DEVICE_MATH_FOR(float, sinh, ::sinhf)
MIGRAPHX_DEVICE_MATH_FOR(float, tan, ::tanf)
MIGRAPHX_DEVICE_MATH_FOR(float, tanh, ::tanhf)
MIGRAPHX_DEVICE_MATH_FOR(float, fmod, ::fmodf)
// Builtin half functions
MIGRAPHX_DEVICE_MATH_FOR(migraphx::half, abs, ::__habs)
......@@ -148,11 +151,13 @@ MIGRAPHX_DEVICE_MATH_HALF(erf, ::erf)
MIGRAPHX_DEVICE_MATH_HALF(floor, ::floor)
MIGRAPHX_DEVICE_MATH_HALF(isnan, ::isnan)
MIGRAPHX_DEVICE_MATH_HALF(pow, ::pow)
MIGRAPHX_DEVICE_MATH_HALF(remainder, ::remainder)
MIGRAPHX_DEVICE_MATH_HALF(round, ::round)
MIGRAPHX_DEVICE_MATH_HALF(sin, ::sin)
MIGRAPHX_DEVICE_MATH_HALF(sinh, ::sinh)
MIGRAPHX_DEVICE_MATH_HALF(tan, ::tan)
MIGRAPHX_DEVICE_MATH_HALF(tanh, ::tanh)
MIGRAPHX_DEVICE_MATH_HALF(fmod, ::fmod)
// Map math functions to hip half2 functions
// The half2 type is defined in include/hip/amd_detail/hip_fp16_gcc.h and is 2 16-bit floats
......@@ -226,11 +231,13 @@ MIGRAPHX_DEVICE_MATH_VEC(cosh)
MIGRAPHX_DEVICE_MATH_VEC(erf)
MIGRAPHX_DEVICE_MATH_VEC(exp)
MIGRAPHX_DEVICE_MATH_VEC(floor)
MIGRAPHX_DEVICE_MATH_VEC(fmod)
MIGRAPHX_DEVICE_MATH_VEC(isnan)
MIGRAPHX_DEVICE_MATH_VEC(log)
MIGRAPHX_DEVICE_MATH_VEC(max)
MIGRAPHX_DEVICE_MATH_VEC(min)
MIGRAPHX_DEVICE_MATH_VEC(pow)
MIGRAPHX_DEVICE_MATH_VEC(remainder)
MIGRAPHX_DEVICE_MATH_VEC(round)
MIGRAPHX_DEVICE_MATH_VEC(rsqrt)
MIGRAPHX_DEVICE_MATH_VEC(sin)
......
src/targets/gpu/kernels/include/migraphx/kernels/reduce.hpp
View file @ 5a14c0bf
......@@ -94,16 +94,17 @@ MIGRAPHX_DPP_REDUCE(op::max, v_max)
MIGRAPHX_DPP_REDUCE(op::min, v_min)
MIGRAPHX_DPP_REDUCE(op::product, v_mul)
template <class Op, class T, class F>
__device__ auto block_reduce(index idx, Op op, T init, index_int n, F f)
template <class Op, class T, class Index, class F>
__device__ auto block_reduce(index idx, Op op, T init, Index n, F f)
{
MIGRAPHX_ASSERT(idx.max_nlocal() == idx.nlocal());
#if __AMDGCN_WAVEFRONT_SIZE == 32
constexpr index_int lanes_per_thread = 16;
#else
constexpr index_int lanes_per_thread = 64;
#endif
using type = decltype(f(0));
__shared__ type buffer[idx.nlocal() / lanes_per_thread];
__shared__ type buffer[idx.max_nlocal() / lanes_per_thread];
type x = init;
idx.local_stride(n, [&](auto i) { x = op(x, f(i)); });
dpp_reduce(x, op);
......@@ -123,12 +124,12 @@ __device__ auto block_reduce(index idx, Op op, T init, index_int n, F f)
return y;
}
#else
template <class Op, class T, class F>
__device__ auto block_reduce(index idx, Op op, T init, index_int n, F f)
template <class Op, class T, class Index, class F>
__device__ auto block_reduce(index idx, Op op, T init, Index n, F f)
{
MIGRAPHX_ASSERT(idx.max_nlocal() == idx.nlocal());
using type = decltype(f(0));
__shared__ type buffer[idx.nlocal()];
__shared__ type buffer[idx.max_nlocal()];
type x = init;
idx.local_stride(n, [&](auto i) { x = op(x, f(i)); });
buffer[idx.local] = x;
......@@ -196,17 +197,14 @@ struct block
struct reducer
{
index idx;
Slicer slicer;
Slicer slice;
template <class Op, class T, class Read>
__device__ auto reduce(Op op, T init, Read read) const
{
return sliced(slicer, [=](auto x, auto... xs) {
return vec_reduce(block_reduce(idx,
op,
init,
x.get_shape().elements(),
[&](auto j) { return read(x[j], xs[j]...); }),
op);
return sliced(slice, [=](auto x, auto... xs) {
return block_reduce(idx, op, init, x.get_shape().elements(), [&](auto j) {
return vec_reduce(read(x[j], xs[j]...), op);
});
});
}
......@@ -220,7 +218,7 @@ struct block
template <class F>
__device__ auto inner(F f) const
{
return sliced(slicer, [=](auto x, auto... xs) {
return sliced(slice, [=](auto x, auto... xs) {
idx.local_stride(x.get_shape().elements(), [&](auto j) { f(x[j], xs[j]...); });
});
}
......@@ -228,7 +226,7 @@ struct block
template <class Input>
constexpr auto elements() const
{
using reduce_type = decltype(slicer(Input{}));
using reduce_type = decltype(slice(Input{}));
using value_type = typename Input::type;
constexpr auto relements = get_shape_c<reduce_type>{}.elements();
if constexpr(vec_size<value_type>() > 1)
......@@ -262,11 +260,11 @@ struct lane
struct reducer
{
index idx;
Slicer slicer;
Slicer slice;
template <class Op, class T, class Read>
__device__ auto reduce(Op op, T init, Read read) const
{
return sliced(slicer, [=](auto x, auto... xs) {
return sliced(slice, [=](auto x, auto... xs) {
using type = typename decltype(x)::type;
type r = init;
for(index_int j = 0; j < x.get_shape().elements(); j++)
......@@ -286,7 +284,7 @@ struct lane
template <class F>
__device__ auto inner(F f) const
{
return sliced(slicer, [=](auto x, auto... xs) {
return sliced(slice, [=](auto x, auto... xs) {
for(index_int j = 0; j < x.get_shape().elements(); j++)
{
f(x[j], xs[j]...);
......@@ -297,7 +295,7 @@ struct lane
template <class Input>
constexpr auto elements() const
{
using reduce_type = decltype(slicer(Input{}));
using reduce_type = decltype(slice(Input{}));
return get_shape_c<reduce_type>{}.elements();
}
};
......
src/targets/gpu/kernels/include/migraphx/kernels/softmax.hpp
View file @ 5a14c0bf
......@@ -33,11 +33,15 @@ template <index_int Axis, class Input, class Output>
__device__ void softmax(Input input, Output output)
{
reduce::block::run<reduce::with_axis<Input, Axis>>([&](auto, auto r) {
auto batch_max = r.reduce(op::max{}, lowest{}, op::id{})(input);
auto batch_sum =
r.reduce(op::sum{}, 0, [&](auto x) { return migraphx::exp(x - batch_max); })(input);
r.inner([&](auto& y, auto x) { y = migraphx::exp(x - batch_max) / batch_sum; })(output,
input);
#ifdef MIGRAPHX_USE_FAST_SOFTMAX
const auto c = vec_at(r.slice(input)[0], 0);
#else
const auto c = r.reduce(op::max{}, lowest{}, op::id{})(input);
#endif
auto batch_sum = r.reduce(op::sum{}, 0, [&](auto x) {
return migraphx::convert<float>(migraphx::exp(x - c));
})(input);
r.inner([&](auto& y, auto x) { y = migraphx::exp(x - c) / batch_sum; })(output, input);
});
}
......
src/targets/gpu/lowering.cpp
View file @ 5a14c0bf
......@@ -26,6 +26,8 @@
#include <migraphx/manage_ptr.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/instruction_ref.hpp>
#include <migraphx/stringutils.hpp>
#include <migraphx/op/convolution.hpp>
#include <migraphx/op/deconvolution.hpp>
......@@ -35,15 +37,12 @@
#include <migraphx/op/quant_convolution.hpp>
#include <migraphx/op/quant_dot.hpp>
#include <migraphx/gpu/batch_norm_inference.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/convolution.hpp>
#include <migraphx/gpu/deconvolution.hpp>
#include <migraphx/gpu/device_name.hpp>
#include <migraphx/gpu/gemm.hpp>
#include <migraphx/gpu/int8_conv_pack.hpp>
#include <migraphx/gpu/miopen.hpp>
#include <migraphx/gpu/quant_convolution.hpp>
#include <migraphx/gpu/rocblas.hpp>
#include <migraphx/gpu/compiler.hpp>
#include <migraphx/iterator_for.hpp>
......@@ -81,78 +80,21 @@ struct miopen_apply
(void)i;
}
const std::unordered_set<std::string>& get_rocblas_fp32_archs()
{
static std::unordered_set<std::string> supported_archs{"gfx908", "gfx90a"};
return supported_archs;
}
void init()
{
assert(mod != nullptr);
assert(pass != nullptr);
#if ROCBLAS_VERSION_MAJOR >= 2 && ROCBLAS_VERSION_MINOR >= 38
auto& ctx = get_context();
const auto device_name = trim(split_string(get_device_name(), ':').front());
if(contains(get_rocblas_fp32_archs(), device_name))
compute_fp32 = true;
rocblas_gemm_flags flag;
rocblas_query_int8_layout_flag(ctx.get_stream().get_rocblas(), &flag);
int8_x4_format = (flag == rocblas_gemm_flags_pack_int8x4);
#endif
auto& ctx = get_context();
int8_x4_format = get_int8_x4_format(ctx);
compute_fp32 = get_compute_fp32_flag();
offload_copy = (mod->name() == "main") ? pass->offload_copy : false;
add_generic_op("acos");
add_generic_op("acosh");
add_generic_op("add");
add_generic_op("asin");
add_generic_op("asinh");
add_generic_op("atan");
add_generic_op("atanh");
add_generic_op("ceil");
add_generic_op("contiguous");
add_generic_op("cos");
add_generic_op("cosh");
add_generic_op("div");
add_generic_op("equal");
add_generic_op("erf");
add_generic_op("exp");
add_generic_op("floor");
add_generic_op("greater");
add_generic_op("less");
add_generic_op("log");
add_generic_op("logical_and");
add_generic_op("logical_or");
add_generic_op("logical_xor");
add_generic_op("max");
add_generic_op("min");
add_generic_op("mul");
add_generic_op("not");
add_generic_op("pow");
add_generic_op("prelu");
add_generic_op("recip");
add_generic_op("relu");
add_generic_op("round");
add_generic_op("rsqrt");
add_generic_op("sigmoid");
add_generic_op("sign");
add_generic_op("sin");
add_generic_op("sinh");
add_generic_op("sqdiff");
add_generic_op("sqrt");
add_generic_op("sub");
add_generic_op("tan");
add_generic_op("tanh");
add_generic_op("where");
add_extend_op("abs");
add_extend_op("argmax");
add_extend_op("argmin");
add_extend_op("clip");
add_extend_op("concat");
add_extend_op("convert");
add_extend_op("elu");
add_extend_op("gather");
add_extend_op("leaky_relu");
......@@ -170,16 +112,15 @@ struct miopen_apply
add_extend_op("scatter_none");
add_extend_op("topk");
add_batch_norm_inference_op();
add_convolution_op();
add_deconvolution_op();
add_convolution_op<op::convolution>("convolution");
add_convolution_op<op::deconvolution>("deconvolution");
add_convolution_op<op::quant_convolution>("quant_convolution");
add_gemm_op<op::dot>("dot");
add_gemm_op<op::quant_dot>("quant_dot");
add_if_op();
add_loop_op();
add_neg_op();
add_nms_op();
add_quant_convolution_op();
}
void copy_params() const
......@@ -228,7 +169,8 @@ struct miopen_apply
init();
for(auto it = mod->begin(); it != mod->end(); it++)
{
auto s = it->get_shape();
auto s = it->get_shape();
auto attrs = it->get_operator().attributes();
if(apply_map.count(it->name()) > 0)
{
check_shape(s, apply_map.at(it->name())(it));
......@@ -237,11 +179,37 @@ struct miopen_apply
{
check_shape(s, insert_precompile_op(it));
}
else if(attrs.contains("target"))
{
check_shape(s, insert_custom_op(it, attrs));
}
}
copy_params();
}
instruction_ref insert_custom_op(instruction_ref ins, const value& attrs) const
{
const auto& custom_op = ins->get_operator();
if(attrs.at("target") == "cpu")
{
auto s = ins->get_shape();
std::vector<instruction_ref> cpu_inputs;
auto inputs = ins->inputs();
auto output = inputs.back();
std::transform(
inputs.begin(), inputs.end(), std::back_inserter(cpu_inputs), [&](auto in) {
return mod->insert_instruction(ins, make_op("hip::copy_from_gpu"), in);
});
cpu_inputs.front() =
mod->insert_instruction(ins, make_op("hip::sync_stream"), cpu_inputs);
auto cpu_out = mod->insert_instruction(ins, custom_op, cpu_inputs);
auto gpu_out =
mod->insert_instruction(ins, make_op("hip::copy_to_gpu"), cpu_out, output);
return mod->replace_instruction(ins, gpu_out);
}
return ins;
}
instruction_ref insert_precompile_op(instruction_ref ins) const
{
auto output = insert_allocation(ins, ins->get_shape());
......@@ -260,38 +228,6 @@ struct miopen_apply
return mod->insert_instruction(ins, make_op("allocate", {{"shape", to_value(s)}}));
}
void add_convolution_op()
{
apply_map.emplace("convolution", [=](instruction_ref ins) {
auto&& op = any_cast<op::convolution>(ins->get_operator());
auto conv = miopen_convolution{op, make_conv(op)};
auto ws = conv.find(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
auto workspace = insert_allocation(ins, ws);
auto output = insert_allocation(ins, ins->get_shape());
return mod->replace_instruction(
ins, conv, ins->inputs().at(0), ins->inputs().at(1), workspace, output);
});
}
void add_deconvolution_op()
{
apply_map.emplace("deconvolution", [=](instruction_ref ins) {
auto&& op = any_cast<op::deconvolution>(ins->get_operator());
auto conv = miopen_deconvolution{op, make_deconv(op)};
auto ws = conv.find(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
auto workspace = insert_allocation(ins, ws);
auto output = insert_allocation(ins, ins->get_shape());
return mod->replace_instruction(
ins, conv, ins->inputs().at(0), ins->inputs().at(1), workspace, output);
});
}
template <typename Op>
void add_gemm_op(const std::string& name)
{
......@@ -305,31 +241,33 @@ struct miopen_apply
});
}
void add_quant_convolution_op()
template <typename Op>
void add_convolution_op(const std::string& name)
{
apply_map.emplace("quant_convolution", [=](instruction_ref ins) {
auto&& op = any_cast<op::quant_convolution>(ins->get_operator());
shape ws;
miopen_quant_convolution conv;
auto compile_quant_conv_with_format = [&](bool format) {
conv = miopen_quant_convolution{op, format, make_conv(op)};
ws = conv.find(get_context(), ins->get_shape(), to_shapes(ins->inputs()));
apply_map.emplace(name, [=](instruction_ref ins) {
operation conv =
miopen_convolution<Op>{any_cast<Op>(ins->get_operator()), int8_x4_format};
migraphx::context ctx = get_context();
size_t ws_bytes = 0;
auto compile_conv_with_format = [&](bool format) {
conv = miopen_convolution<Op>{any_cast<Op>(ins->get_operator()), format};
auto ws = conv.compile(ctx, ins->get_shape(), to_shapes(ins->inputs()));
ws_bytes = ws.get("workspace", 0);
};
try
{
compile_quant_conv_with_format(int8_x4_format);
{ // for the regular convolution and deconvolution, this try would always succeed
compile_conv_with_format(int8_x4_format);
}
catch(migraphx::exception&)
{
// In case no solver supports the default format, retry using the other format.
compile_quant_conv_with_format(not int8_x4_format);
compile_conv_with_format(not int8_x4_format);
}
auto args = ins->inputs();
auto workspace = insert_allocation(ins, ws);
auto output = insert_allocation(ins, ins->get_shape());
auto workspace = insert_allocation(ins, shape{shape::int8_type, {ws_bytes}});
return mod->replace_instruction(ins, conv, args[0], args[1], workspace, output);
});
}
......@@ -364,43 +302,6 @@ struct miopen_apply
});
}
void add_batch_norm_inference_op()
{
apply_map.emplace("batch_norm_inference", [=](instruction_ref ins) {
auto&& op = any_cast<op::batch_norm_inference>(ins->get_operator());
auto output = insert_allocation(ins, ins->get_shape());
shape old_shape = ins->inputs().at(1)->get_shape();
auto input = ins->inputs()[0];
auto input_lens = input->get_shape().lens();
std::vector<int64_t> rsp_lens(input_lens.size(), 1);
// for per_activation case, also need to reshape input
if(op.bn_mode == op::batch_norm_inference::per_activation)
{
std::copy(input_lens.begin() + 1, input_lens.end(), rsp_lens.begin() + 1);
}
else
{
rsp_lens[1] = static_cast<int64_t>(old_shape.elements());
}
auto reshape_op = op::reshape{rsp_lens};
std::vector<instruction_ref> reshapes;
std::transform(ins->inputs().begin() + 1,
ins->inputs().end(),
std::back_inserter(reshapes),
[&](auto i) { return mod->insert_instruction(ins, reshape_op, i); });
return mod->replace_instruction(ins,
miopen_batch_norm_inference{op},
input,
reshapes[0],
reshapes[1],
reshapes[2],
reshapes[3],
output);
});
}
// use 0 - input to represent neg
void add_neg_op()
{
......
src/targets/gpu/mlir.cpp
View file @ 5a14c0bf
......@@ -21,6 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include "migraphx/make_op.hpp"
#include <migraphx/gpu/mlir.hpp>
#ifdef MIGRAPHX_MLIR
......@@ -43,8 +44,9 @@
#include <migraphx/gpu/code_object_op.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/device_name.hpp>
#include <migraphx/iterator_for.hpp>
#include <migraphx/gpu/perfdb.hpp>
#include <migraphx/iterator_for.hpp>
#include <migraphx/permutation.hpp>
#include <deque>
#include <variant>
......@@ -194,7 +196,6 @@ struct mlir_program
MlirType make_tensor(const shape& s) const
{
assert(s.standard());
std::vector<int64_t> lens(s.lens().begin(), s.lens().end());
return mlirRankedTensorTypeGet(
lens.size(), lens.data(), make_type(s.type()), mlirAttributeGetNull());
......@@ -502,11 +503,12 @@ struct mlir_program
{
pp =
problem_params{ins->get_operator(), to_shapes(ins->inputs()), ins->get_shape()};
std::string tuned = get_tune_params();
// check if HW supports xdlops
bool xdlops = contains(get_xdlops_archs(), target_name);
std::string tuned = get_tune_params(xdlops);
if(not tuned.empty())
ops.add_attributes({{"perf_config", tuned}});
// check if HW supports xdlops
if(contains(get_xdlops_archs(), target_name))
if(xdlops)
ops.add_attributes({{"xdlopsV2", true}});
}
......@@ -571,7 +573,7 @@ struct mlir_program
MIGRAPHX_THROW("Failed to compile mlir program");
}
std::string get_tune_params() { return get_mlir_perf_for_conv(pp); }
std::string get_tune_params(bool xdlops) { return get_mlir_perf_for_conv(pp, xdlops); }
mlir_context ctx;
MlirLocation location;
......@@ -589,8 +591,54 @@ std::string dump_mlir(const module& m)
return mlir_print(&mlirOperationPrint, mod_op);
}
code_object_op compile_mlir(const context&, const module& m)
void adjust_param_shapes(module& m, const std::vector<instruction_ref>& inputs)
{
auto names = m.get_parameter_names();
std::sort(names.begin(), names.end());
for(auto i : range(names.size()))
{
const auto& name = names[i];
const auto& input = inputs[i]->get_shape();
auto param = m.get_parameter(name);
if(input.standard())
continue;
auto lens = input.lens();
auto strides = input.strides();
std::vector<operation> ops;
if(input.transposed())
{
auto perm = find_permutation(input);
auto iperm = invert_permutation(perm);
lens = reorder_dims(lens, iperm);
strides = reorder_dims(strides, iperm);
ops.push_back(make_op("transpose", {{"permutation", perm}}));
}
if(input.broadcasted())
{
std::transform(lens.begin(),
lens.end(),
strides.begin(),
lens.begin(),
[](auto len, auto stride) -> std::size_t {
if(stride == 0)
return 1;
return len;
});
ops.push_back(make_op("multibroadcast", {{"out_lens", input.lens()}}));
}
auto new_param =
std::accumulate(ops.begin(),
ops.end(),
m.add_parameter(name + ".0", shape{input.type(), lens}),
[&](auto x, auto op) { return m.insert_instruction(param, op, x); });
m.replace_instruction(param, new_param);
m.remove_instruction(param);
}
}
code_object_op compile_mlir(const context&, module m, const std::vector<instruction_ref>& inputs)
{
adjust_param_shapes(m, inputs);
const bool trace = enabled(MIGRAPHX_TRACE_MLIR{});
if(trace)
std::cout << m << std::endl;
......@@ -662,13 +710,19 @@ instruction_ref insert_mlir(module& m,
std::string dump_mlir(const module&) { return {}; }
code_object_op compile_mlir(const context&, const module&) { return {}; }
template <class T>
void use(T&)
{
}
// Disabling clang-tidy warning on non-real useage.
// NOLINTBEGIN(performance-unnecessary-value-param)
code_object_op compile_mlir(const context&, module, const std::vector<instruction_ref>&)
{
return {};
}
// NOLINTEND(performance-unnecessary-value-param)
instruction_ref
// cppcheck-suppress funcArgNamesDifferent
insert_mlir(module& m, instruction_ref, code_object_op co, const std::vector<instruction_ref>&)
......
src/targets/gpu/perfdb.cpp
View file @ 5a14c0bf
......@@ -108,16 +108,17 @@ auto query_miopen_db(const std::string& query)
} // namespace
std::string get_mlir_perf_for_conv(const problem_params& pp)
std::string get_mlir_perf_for_conv(const problem_params& pp, bool xdlops)
{
std::string query = "select P.* \
std::string solver = xdlops ? "ConvMlirIgemmFwdXdlops" : "ConvMlirIgemmFwd";
std::string query = "select P.* \
from perf_db P, config C \
where P.config = C.id AND \
P.solver = 'ConvMlirIgemmFwdXdlops' AND \
P.solver = '${solver}' AND \
${config}";
auto results =
query_miopen_db(interpolate_string(query, {{"config", generate_miopen_config(pp)}}));
auto results = query_miopen_db(
interpolate_string(query, {{"config", generate_miopen_config(pp)}, {"solver", solver}}));
if(results.empty())
return "";
return results.front().at("params");
......
src/targets/gpu/prefuse_ops.cpp
View file @ 5a14c0bf
......@@ -35,6 +35,12 @@ namespace {
template <class Derived, std::size_t N>
struct layernorm_base
{
float epsilon = 1e-12f;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return pack(f(self.epsilon, "epsilon"));
}
shape compute_shape(std::vector<shape> inputs, std::vector<module_ref> mods) const
{
std::size_t nargs = 1;
......@@ -62,6 +68,7 @@ struct layernorm_base
struct layernorm : layernorm_base<layernorm, 0>
{
std::string name() const { return "gpu::prelayernorm"; }
};
MIGRAPHX_REGISTER_OP(layernorm);
......@@ -80,8 +87,9 @@ struct find_layernorm
{
auto ins = r.result;
auto x_ins = r.instructions["x"];
auto eps = r.instructions["eps"]->eval().at<float>();
m.replace_instruction(ins, layernorm{}, x_ins);
m.replace_instruction(ins, layernorm{eps}, x_ins);
}
};
......@@ -96,8 +104,9 @@ struct find_add_layernorm
{
auto ins = r.result;
auto add_ins = r.instructions["add"];
auto eps = r.instructions["eps"]->eval().at<float>();
m.replace_instruction(ins, add_layernorm{}, add_ins->inputs());
m.replace_instruction(ins, add_layernorm{eps}, add_ins->inputs());
}
};
} // namespace
......
src/targets/gpu/quant_convolution.cpp deleted 100644 → 0
View file @ cb01e280
/*
* 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/quant_convolution.hpp>
#include <migraphx/gpu/device/convert.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/generate.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
shape miopen_quant_convolution::compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, *this}.has(4).standard();
return op.normalize_compute_shape({inputs.at(0), inputs.at(1)});
}
argument miopen_quant_convolution::compute(context& ctx,
const shape& output_shape,
const std::vector<argument>& args) const
{
auto x_desc = make_tensor(args[0].get_shape(), int8_x4_format);
auto w_desc = make_tensor(args[1].get_shape(), int8_x4_format);
auto y_desc = make_tensor(output_shape);
if(solution_id == 0)
MIGRAPHX_THROW("MIOpen Convolution: invalid solution ID");
auto status = miopenConvolutionForwardImmediate(ctx.get_stream().get_miopen(),
w_desc.get(),
args[1].implicit(),
x_desc.get(),
args[0].implicit(),
cd.get(),
y_desc.get(),
args[3].implicit(),
args[2].implicit(),
args[2].get_shape().bytes(),
solution_id);
if(status != miopenStatusSuccess)
MIGRAPHX_THROW("QUANT_CONVOLUTION: run convolution forward failed");
return args[3];
}
shape miopen_quant_convolution::find(context& ctx,
const shape& output_shape,
std::vector<shape> inputs)
{
shape workspace_shape{};
auto x_desc = make_tensor(inputs[0], int8_x4_format);
auto w_desc = make_tensor(inputs[1], int8_x4_format);
auto y_desc = make_tensor(output_shape);
std::size_t workspace_size = 0;
miopenConvolutionForwardGetWorkSpaceSize(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
&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;
auto 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 Quant Convolution: find convolution failed");
size_t solution_count = 1;
miopenConvSolution_t qconv_solution;
status = miopenConvolutionForwardGetSolution(ctx.get_stream().get_miopen(),
w_desc.get(),
x_desc.get(),
cd.get(),
y_desc.get(),
1,
&solution_count,
&qconv_solution);
if(status != miopenStatusSuccess or solution_count != 1)
MIGRAPHX_THROW("MIOpen Quant Convolution: get solution failed");
solution_id = qconv_solution.solution_id;
algo = qconv_solution.algorithm;
return shape{shape::int8_type, {qconv_solution.workspace_size}};
}
void miopen_quant_convolution::finalize(context& ctx,
const shape& output_shape,
std::vector<shape> inputs)
{
if(cd == nullptr)
cd = make_conv(op);
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 Quant Convolution: workspace has changed during finalization.");
}
auto x_desc = make_tensor(inputs[0], int8_x4_format);
auto w_desc = make_tensor(inputs[1], int8_x4_format);
auto y_desc = make_tensor(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 Quant Convolution: compile solution failed");
}
shape miopen_quant_convolution::pack_int8_shape(const shape& s) const
{
if(s.type() != shape::int8_type)
{
MIGRAPHX_THROW("PACK_INT8_SHAPE: only process int8_type");
}
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
src/targets/gpu/rocblas.cpp
View file @ 5a14c0bf
......@@ -21,7 +21,13 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
*/
#include <unordered_set>
#include <migraphx/ranges.hpp>
#include <migraphx/stringutils.hpp>
#include <migraphx/gpu/device_name.hpp>
#include <migraphx/gpu/rocblas.hpp>
#include <migraphx/gpu/context.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......@@ -41,6 +47,33 @@ rocblas_handle_ptr create_rocblas_handle_ptr(hipStream_t s)
return rb;
}
const std::unordered_set<std::string>& get_rocblas_fp32_archs()
{
static std::unordered_set<std::string> supported_archs{"gfx908", "gfx90a"};
return supported_archs;
}
bool get_compute_fp32_flag()
{
bool compute_fp32 = false;
#if ROCBLAS_VERSION_MAJOR >= 2 && ROCBLAS_VERSION_MINOR >= 38
const auto device_name = trim(split_string(get_device_name(), ':').front());
if(contains(get_rocblas_fp32_archs(), device_name))
compute_fp32 = true;
#endif
return compute_fp32;
}
bool get_int8_x4_format(context& ctx)
{
bool int8_x4_format = true;
#if ROCBLAS_VERSION_MAJOR >= 2 && ROCBLAS_VERSION_MINOR >= 38
rocblas_gemm_flags flag;
rocblas_query_int8_layout_flag(ctx.get_stream().get_rocblas(), &flag);
int8_x4_format = (flag == rocblas_gemm_flags_pack_int8x4);
#endif
return int8_x4_format;
}
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/softmax.cpp deleted 100644 → 0
View file @ cb01e280
/*
* 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/softmax.hpp>
#include <migraphx/gpu/device/softmax.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/tune_axis.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
shape hip_softmax::compute_shape(const std::vector<shape>& inputs) const
{
check_shapes{inputs, *this}.has(2).standard();
return op.normalize_compute_shape({inputs.at(0)});
}
argument hip_softmax::compute(context& ctx, const shape&, const std::vector<argument>& args) const
{
auto n_dim = args.front().get_shape().lens().size();
auto tuned_axis = tune_axis(n_dim, op.axis, op.name());
device::softmax(ctx.get_stream().get(), args.back(), args.front(), tuned_axis);
return args.back();
}
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/target.cpp
View file @ 5a14c0bf
......@@ -41,7 +41,6 @@
#include <migraphx/propagate_constant.hpp>
#include <migraphx/register_target.hpp>
#include <migraphx/replace_allocate.hpp>
#include <migraphx/rewrite_batchnorm.hpp>
#include <migraphx/rewrite_gelu.hpp>
#include <migraphx/rewrite_pooling.hpp>
#include <migraphx/rewrite_quantization.hpp>
......@@ -110,8 +109,6 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
dead_code_elimination{},
insert_pad{},
dead_code_elimination{},
rewrite_batchnorm{},
dead_code_elimination{},
rewrite_rnn{},
dead_code_elimination{},
inline_module{},
......
src/targets/ref/lowering.cpp
View file @ 5a14c0bf
......@@ -26,7 +26,6 @@
#include <migraphx/instruction.hpp>
#include <migraphx/dfor.hpp>
#include <migraphx/op/identity.hpp>
#include <migraphx/op/batch_norm_inference.hpp>
#include <migraphx/op/convolution.hpp>
#include <migraphx/op/deconvolution.hpp>
#include <migraphx/op/quant_convolution.hpp>
......@@ -75,84 +74,6 @@ typename std::conditional_t<std::is_integral<T>{}, std::make_signed<T>, std::ena
return x;
}
//
// ref implemenataion of batch norm for inference
//
// inputs are:
// args[0] -> input data buffer
// args[1] -> mini batch mean
// args[2] -> mini batch variance
// args[3] -> gamma
// args[4] -> bias
//
// The equation to compute batch norm for inference is:
//
// output[i] = bias + gamma * (input[i] + mean) / sqrt(variance + epsilon)
//
// the input data format should be nchw
//
struct ref_batch_norm_inference
{
op::batch_norm_inference op;
template <class Self, class F>
static auto reflect(Self& self, F f)
{
return migraphx::reflect(self.op, f);
}
std::string name() const { return "ref::batch_norm_inference"; }
shape compute_shape(const std::vector<shape>& inputs) const { return op.compute_shape(inputs); }
argument compute(context&, const shape& output_shape, std::vector<argument> args) const
{
argument output{output_shape};
double epsilon = op.epsilon;
auto input = args[0];
auto arg_gamma = args[1];
auto arg_bias = args[2];
auto mini_batch_mean = args[3];
auto mini_batch_variance = args[4];
if(op.bn_mode == op::batch_norm_inference::spatial)
{
visit_all(output, input, mini_batch_mean, mini_batch_variance, arg_gamma, arg_bias)(
[&](auto result, auto buffer, auto mean, auto variance, auto gamma, auto bias) {
par_for(output_shape.elements(), [&](auto i) {
auto idx = output_shape.multi(i);
auto c = idx[1];
assert((variance[c] + epsilon) > 0);
result[i] =
gamma[c] * (buffer[i] - mean[c]) / std::sqrt(variance[c] + epsilon) +
bias[c];
});
});
}
if(op.bn_mode == op::batch_norm_inference::per_activation)
{
visit_all(output, input, mini_batch_mean, mini_batch_variance, arg_gamma, arg_bias)(
[&](auto result, auto buffer, auto mean, auto variance, auto gamma, auto bias) {
par_for(output_shape.elements(), [&](auto i) {
auto idx = output_shape.multi(i);
idx[0] = 0;
auto index = output_shape.index(idx);
assert((variance[index] + epsilon) > 0);
result[i] = gamma[index] * (buffer[i] - mean[index]) /
std::sqrt(variance[index] + epsilon) +
bias[index];
});
});
}
return output;
}
};
MIGRAPHX_REGISTER_OP(ref_batch_norm_inference)
struct ref_lrn
{
op::lrn op;
......@@ -237,15 +158,16 @@ struct ref_convolution : auto_register_op<ref_convolution<Op>>
argument compute(context&, shape output_shape, std::vector<argument> args) const
{
std::vector<std::size_t> padding;
if(op.use_dynamic_same_auto_pad)
if(op.padding_mode != op::padding_mode_t::default_)
{
auto input_lens = args[0].get_shape().lens();
std::vector<std::size_t> img_lens{input_lens.begin() + 2, input_lens.end()};
auto input_lens = args[0].get_shape().lens();
auto weights_lens = args[1].get_shape().lens();
std::vector<std::size_t> k_lens{weights_lens.begin() + 2, weights_lens.end()};
padding = calc_dyn_auto_pad(img_lens, k_lens, op.stride, op.dilation);
output_shape =
compute_padded_shape({args.at(0).get_shape(), args.at(1).get_shape()}, padding);
padding =
op.padding_mode == op::same_upper
? calc_dyn_auto_pad(input_lens, weights_lens, op.stride, op.dilation, true)
: calc_dyn_auto_pad(input_lens, weights_lens, op.stride, op.dilation, false);
output_shape = compute_padded_shape(
args[0].get_shape(), args[1].get_shape(), padding, op.stride, op.dilation);
}
else
{
......@@ -313,34 +235,6 @@ struct ref_convolution : auto_register_op<ref_convolution<Op>>
});
return result;
}
private:
/*!
* Used for dynamic auto padding since padding needs to be computed at evaulation time.
* \param inputs two fixed shape inputs [input_tensor, weights]
* \param padding from auto_pad calculation
*/
shape compute_padded_shape(const std::vector<shape>& inputs,
const std::vector<std::size_t>& padding) const
{
const shape& input = inputs.at(0);
const shape& weights = inputs.at(1);
const size_t num_spatial_dims = input.lens().size() - 2;
std::vector<size_t> output_lens{input.lens()[0], weights.lens()[0]};
// calculate the output shape of the convolution: ((W - K + 2P) / S) + 1
for(size_t i = 0; i < num_spatial_dims; i++)
{
auto padding_factor = padding[i] + padding[i + num_spatial_dims];
output_lens.push_back(std::size_t(std::max<std::ptrdiff_t>(
1,
(input.lens()[i + 2] - (1 + op.dilation[i] * (weights.lens()[i + 2] - 1)) +
padding_factor) /
op.stride[i] +
1)));
}
return inputs[0].with_lens(output_lens);
}
};
struct ref_im2col
......@@ -731,8 +625,6 @@ struct ref_apply
void init()
{
apply_map["batch_norm_inference"] =
extend_op<ref_batch_norm_inference, op::batch_norm_inference>();
apply_map["convolution"] = extend_op<ref_convolution<op::convolution>, op::convolution>();
apply_map["dot"] = extend_op<ref_gemm, op::dot>();
apply_map["quant_dot"] = extend_op<ref_quant_gemm, op::quant_dot>();
......
src/tf/parse_batchnorm.cpp
View file @ 5a14c0bf
......@@ -23,6 +23,7 @@
*/
#include <migraphx/tf/op_parser.hpp>
#include <migraphx/tf/tf_parser.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/ranges.hpp>
#include <migraphx/make_op.hpp>
......@@ -38,16 +39,37 @@ struct parse_batchnorm : op_parser<parse_batchnorm>
instruction_ref parse(const op_desc& /*opd*/,
const tf_parser& /*parser*/,
tf_parser::node_info info,
const std::vector<instruction_ref>& args) const
std::vector<instruction_ref> args) const
{
float epsilon = 1e-5f;
float momentum = 0.9f;
// different default epsilon than from ONNX
float epsilon = 1e-4f;
if(contains(info.attributes, "epsilon"))
{
epsilon = info.attributes.at("epsilon").f();
}
auto op = make_op("batch_norm_inference", {{"epsilon", epsilon}, {"momentum", momentum}});
return info.add_instruction(op, args);
auto x_lens = args[0]->get_shape().lens();
auto x_type = args[0]->get_shape().type();
// unsqueeze tensors of shape (C) to broadcast correctly
auto rt = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {0.5}});
auto eps = info.add_literal(migraphx::literal{migraphx::shape{x_type}, {epsilon}});
auto scale_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[1]);
auto bias_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[2]);
auto mean_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[3]);
auto var_unsqueeze =
info.add_instruction(migraphx::make_op("unsqueeze", {{"axes", {1, 2}}}), args[4]);
auto numer = info.add_broadcastable_binary_op("sub", args[0], mean_unsqueeze);
auto var_eps = info.add_broadcastable_binary_op("add", var_unsqueeze, eps);
auto denom = info.add_broadcastable_binary_op("pow", var_eps, rt);
auto div0 = info.add_broadcastable_binary_op("div", numer, denom);
auto r0 = info.add_broadcastable_binary_op("mul", div0, scale_unsqueeze);
return info.add_broadcastable_binary_op("add", r0, bias_unsqueeze);
}
};
......
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