Skip to content

GitLab

Unverified Commit a24ed87e authored by Chris Austen's avatar Chris Austen Committed by GitHub
Browse files

Merge branch 'develop' into optimize_jenkinsfile

parents 6481cd69 a09dc502
Hide whitespace changes
Inline Side-by-side
Showing with 222 additions and 367 deletions
src/targets/gpu/kernels/include/migraphx/kernels/tensor_view.hpp
View file @ a24ed87e
......@@ -27,6 +27,7 @@
#include <migraphx/kernels/shape.hpp>
#include <migraphx/kernels/debug.hpp>
#include <migraphx/kernels/iota_iterator.hpp>
#include <migraphx/kernels/float8.hpp>
namespace migraphx {
......
src/targets/gpu/kernels/include/migraphx/kernels/type_traits.hpp
View file @ a24ed87e
......@@ -251,7 +251,7 @@ constexpr T numeric_max()
}
template <class T>
constexpr T numeric_lowest()
constexpr auto numeric_lowest() -> decltype(numeric_max<T>())
{
if constexpr(is_integral<T>{})
{
......
src/targets/gpu/kernels/include/migraphx/kernels/vec.hpp
View file @ a24ed87e
......@@ -207,7 +207,7 @@ struct implicit_conversion_op
template <class U>
constexpr operator U() const
{
return x;
return static_cast<U>(x);
}
};
......
src/targets/gpu/lowering.cpp
View file @ a24ed87e
......@@ -61,9 +61,8 @@ struct miopen_apply
const lowering* pass = nullptr;
std::unordered_map<std::string, std::function<instruction_ref(instruction_ref)>> apply_map{};
instruction_ref last{};
bool offload_copy = false;
bool int8_x4_format = true;
bool compute_fp32 = false;
bool offload_copy = false;
bool compute_fp32 = false;
context& get_context() const
{
......@@ -84,10 +83,8 @@ struct miopen_apply
assert(mod != nullptr);
assert(pass != nullptr);
auto& ctx = get_context();
int8_x4_format = get_int8_x4_format(ctx);
compute_fp32 = get_compute_fp32_flag();
offload_copy = (mod == mpm->get_root_module()) ? pass->offload_copy : false;
compute_fp32 = get_compute_fp32_flag();
offload_copy = (mod == mpm->get_root_module()) ? pass->offload_copy : false;
add_generic_op("contiguous");
add_extend_op("argmax");
......@@ -231,18 +228,15 @@ struct miopen_apply
assert(refs.size() == 2);
auto output = insert_allocation(ins, ins->get_shape());
refs.push_back(output);
return mod->replace_instruction(
ins, rocblas_gemm<Op>{Op{}, 1, 0, int8_x4_format, compute_fp32}, refs);
return mod->replace_instruction(ins, rocblas_gemm<Op>{Op{}, 1, 0, compute_fp32}, refs);
});
}
void add_convolution_op(const std::string& name)
{
apply_map.emplace(name, [=](instruction_ref ins) {
operation conv = make_op(
"gpu::" + name,
{{"op", ins->get_operator().to_value()}, {"int8_x4_format", int8_x4_format}});
auto output = insert_allocation(ins, ins->get_shape());
operation conv = make_op("gpu::" + name, {{"op", ins->get_operator().to_value()}});
auto output = insert_allocation(ins, ins->get_shape());
return mod->replace_instruction(ins,
make_op("gpu::miopen_op", {{"op", to_value(conv)}}),
......
src/targets/gpu/mlir.cpp
View file @ a24ed87e
......@@ -37,7 +37,7 @@
#include <mlir-c/Pass.h>
#include <mlir-c/Support.h>
#include <mutex>
#if !defined(MLIR_MIGRAPHX_DIALECT_API_VERSION) || MLIR_MIGRAPHX_DIALECT_API_VERSION != 3
#if !defined(MLIR_MIGRAPHX_DIALECT_API_VERSION) || MLIR_MIGRAPHX_DIALECT_API_VERSION != 4
#warning "Incompatible version of rocMLIR library used, disabling"
// Only undefine when not using cppcheck
#ifndef CPPCHECK
......@@ -73,6 +73,7 @@ namespace gpu {
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TRACE_MLIR);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNE_EXHAUSTIVE);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNE_LIMIT);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNING_DB);
MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_MLIR_TUNING_CFG);
......@@ -318,31 +319,30 @@ struct mlir_program
return result;
}
MlirType make_tensor(const shape& s) const
MlirType make_mlir_shaped(const shape& s) const
{
if(not s.standard())
MIGRAPHX_THROW("MLIR expects all tensors to be in standard shape");
if(s.dynamic())
MIGRAPHX_THROW("MLIR does not support dynamic shapes");
std::vector<int64_t> lens(s.lens().begin(), s.lens().end());
return mlirRankedTensorTypeGet(
lens.size(), lens.data(), make_type(s.type()), mlirAttributeGetNull());
std::vector<int64_t> strides(s.strides().begin(), s.strides().end());
return rocmlirMIXRShapedTypeGet(
lens.size(), lens.data(), strides.data(), make_type(s.type()));
}
template <class Range>
std::vector<MlirType> make_tensors(const Range& r)
std::vector<MlirType> make_mlir_shapeds(const Range& r)
{
std::vector<MlirType> result;
std::transform(r.begin(), r.end(), std::back_inserter(result), [&](const auto& s) {
return make_tensor(s);
return make_mlir_shaped(s);
});
return result;
}
MlirType make_function_type(const std::vector<shape>& inputs, const std::vector<shape>& outputs)
{
auto in = make_tensors(inputs);
auto out = make_tensors(outputs);
auto in = make_mlir_shapeds(inputs);
auto out = make_mlir_shapeds(outputs);
return mlirFunctionTypeGet(ctx.get(), in.size(), in.data(), out.size(), out.data());
}
......@@ -504,11 +504,7 @@ struct mlir_program
mlir_operation_state& add_results(const std::vector<shape>& outputs)
{
std::vector<shape> reshaped(outputs.size());
std::transform(outputs.begin(), outputs.end(), reshaped.begin(), [](const shape& r) {
return shape{r.type(), r.lens()};
});
auto x = prog->make_tensors(reshaped);
auto x = prog->make_mlir_shapeds(outputs);
if(not x.empty())
{
mlirOperationStateAddResults(&op_state, x.size(), x.data());
......@@ -581,7 +577,7 @@ struct mlir_program
std::vector<shape> outputs = m.get_output_shapes();
std::vector<MlirLocation> arg_locs(inputs.size(), location);
auto body_inputs = make_tensors(inputs);
auto body_inputs = make_mlir_shapeds(inputs);
mlir_region region = mlirRegionCreate();
mlir_block fbody = mlirBlockCreate(body_inputs.size(), body_inputs.data(), arg_locs.data());
MlirBlock result = fbody.get();
......@@ -607,7 +603,7 @@ struct mlir_program
return "func.return";
if(ins->name() == "@literal")
{
return "tosa.const";
return "migraphx.literal";
}
return "migraphx." + ins->name();
}
......@@ -666,7 +662,8 @@ struct mlir_program
if(ins->name() == "@literal")
{
literal r = ins->get_literal();
MlirType tensor_type = make_tensor(ins->get_shape());
MlirType shaped_type = make_mlir_shaped(ins->get_shape());
MlirType tensor_type = rocmlirMIXRShapedTypeAsTensor(shaped_type);
MlirAttribute mlir_value_attr =
mlirDenseElementsAttrRawBufferGet(tensor_type, r.get_shape().bytes(), r.data());
ops.add_attributes({{"value", mlir_value_attr}});
......@@ -796,7 +793,9 @@ struct mlir_program
if(enabled(MIGRAPHX_MLIR_TUNE_EXHAUSTIVE{}))
tuning_mode = RocmlirTuningParamSetKindExhaustive;
mlir_tuning_space params{mlirRockTuningSpaceCreate(mmodule.get(), tuning_mode)};
for(auto i : range(mlirRockTuningGetNumParams(params.get())))
const auto limit =
value_of(MIGRAPHX_MLIR_TUNE_LIMIT{}, std::numeric_limits<std::size_t>::max());
for(auto i : range(std::min<std::size_t>(limit, mlirRockTuningGetNumParams(params.get()))))
{
mlir_tuning_param param{mlirRockTuningParamCreate()};
if(not mlirRockTuningParamGet(params.get(), i, param.get()))
......@@ -942,35 +941,7 @@ void adjust_param_shapes(module& m, const std::vector<shape>& inputs)
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); });
auto new_param = m.add_parameter(name + ".0", input);
m.replace_instruction(param, new_param);
m.remove_instruction(param);
}
......@@ -1032,6 +1003,15 @@ tuning_config get_tuning_config_mlir(const context& migraphx_ctx,
mlir_program mp;
mp.set_gpu_properties(migraphx_ctx);
mp.parse(m);
const bool trace = enabled(MIGRAPHX_TRACE_MLIR{});
static std::mutex mutex;
if(trace)
{
const std::lock_guard<std::mutex> lock(mutex);
auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
std::cout << mlir_print(&mlirOperationPrint, mod_op) << std::endl;
}
return mp.get_tuning_config(exhaustive);
}
......
src/targets/gpu/pack_int8_args.cpp deleted 100644 → 0
View file @ 6481cd69
/*
* 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 <iterator>
#include <migraphx/gpu/pack_int8_args.hpp>
#include <migraphx/gpu/int8_gemm_pack.hpp>
#include <migraphx/gpu/int8_conv_pack.hpp>
#include <migraphx/gpu/hip.hpp>
#include <migraphx/instruction.hpp>
#include <migraphx/instruction_ref.hpp>
#include <migraphx/program.hpp>
#include <migraphx/iterator_for.hpp>
#include <migraphx/make_op.hpp>
#include <migraphx/permutation.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
static instruction_ref pad_ins(module& m, instruction_ref ins, int offset)
{
auto s = ins->get_shape();
auto lens = s.lens();
auto k = lens[lens.size() + offset];
auto pad_k = (k + 3) / 4 * 4;
auto pad_lens = lens;
pad_lens[lens.size() + offset] = pad_k;
auto ret_ins = ins;
if(pad_k != k)
{
std::vector<int64_t> pad_dims(lens.size() * 2, 0);
pad_dims[lens.size() + offset] = pad_k - k;
shape ps{s.type(), pad_lens};
auto ins_out =
m.insert_instruction(ins, make_op("hip::allocate", {{"shape", to_value(ps)}}));
auto pad = make_op("pad", {{"pads", pad_dims}});
ret_ins =
m.insert_instruction(std::next(ins), make_op("gpu::pad", pad.to_value()), ins, ins_out);
}
return ret_ins;
}
static std::vector<instruction_ref> pad_inputs(module& m, instruction_ref ins)
{
std::vector<instruction_ref> ret_inputs;
auto inputs = ins->inputs();
auto in0 = inputs.at(0);
auto sa = in0->get_shape();
bool transa = sa.transposed();
if(transa)
{
auto perm = find_permutation(sa);
auto val = in0->get_operator().to_value();
if(val.contains("dims"))
{
int offset = static_cast<int>(perm.back()) - static_cast<int>(perm.size());
auto t_in = in0->inputs().front();
auto p_in = pad_ins(m, t_in, offset);
auto dims = val.at("dims").to_vector<int64_t>();
auto r_in =
m.insert_instruction(ins, make_op("transpose", {{"permutation", dims}}), p_in);
ret_inputs.push_back(r_in);
}
else
{
shape cs{in0->get_shape().type(), in0->get_shape().lens()};
auto con_out =
m.insert_instruction(ins, make_op("hip::allocate", {{"shape", to_value(cs)}}));
auto cin0 = m.insert_instruction(ins, make_op("gpu::contiguous"), in0, con_out);
ret_inputs.push_back(pad_ins(m, cin0, -1));
}
}
else
{
ret_inputs.push_back(pad_ins(m, in0, -1));
}
auto in1 = inputs.at(1);
auto sb = in1->get_shape();
bool transb = sb.transposed();
if(transb)
{
auto perm = find_permutation(sb);
auto val = in1->get_operator().to_value();
if(val.contains("dims"))
{
int offset = static_cast<int>(perm[perm.size() - 2]) - static_cast<int>(perm.size());
auto t_in = in1->inputs().front();
auto p_in = pad_ins(m, t_in, offset);
auto dims = val.at("dims").to_vector<int64_t>();
auto r_in =
m.insert_instruction(ins, make_op("transpose", {{"permutation", dims}}), p_in);
ret_inputs.push_back(r_in);
}
else
{
shape cs{in1->get_shape().type(), in1->get_shape().lens()};
auto con_out =
m.insert_instruction(ins, make_op("hip::allocate", {{"shape", to_value(cs)}}));
auto cin1 = m.insert_instruction(ins, make_op("gpu::contiguous"), in1, con_out);
ret_inputs.push_back(pad_ins(m, cin1, -2));
}
}
else
{
ret_inputs.push_back(pad_ins(m, in1, -2));
}
std::copy(inputs.begin() + 2, inputs.end(), std::back_inserter(ret_inputs));
return ret_inputs;
}
void pack_int8_args::apply(module& m) const
{
for(auto ins : iterator_for(m))
{
if(ins->name() == "gpu::quant_gemm")
{
auto val = ins->get_operator().to_value();
assert(val.contains("int8_x4_format"));
if(not val.at("int8_x4_format").to<bool>())
{
continue;
}
auto inputs = ins->inputs();
auto lens = inputs.at(0)->get_shape().lens();
// gemm need the k to be multiple of 4, so need packing that dimension
auto old_inputs = inputs;
if((lens.back() % 4) != 0)
{
inputs = pad_inputs(m, ins);
}
bool transa = inputs[0]->get_shape().transposed();
bool transb = inputs[1]->get_shape().transposed();
if(not transb)
{
auto packed_b = m.insert_instruction(
ins, make_op("hip::allocate", {{"shape", to_value(inputs[1]->get_shape())}}));
auto output_b = m.insert_instruction(
ins, make_op("gpu::int8_gemm_pack_a"), {inputs[1], packed_b});
inputs[1] = output_b;
}
if(transa)
{
auto packed_a = m.insert_instruction(
ins, make_op("hip::allocate", {{"shape", to_value(inputs[0]->get_shape())}}));
auto output_a = m.insert_instruction(
ins, make_op("gpu::int8_gemm_pack_b"), {inputs[0], packed_a});
inputs[0] = output_a;
}
if(inputs != old_inputs)
{
m.replace_instruction(ins, ins->get_operator(), inputs);
}
}
else if(ins->name() == "gpu::quant_convolution")
{
auto val = ins->get_operator().to_value();
if(not val.at("int8_x4_format").to<bool>())
{
continue;
}
auto inputs = ins->inputs();
auto packed_x = m.insert_instruction(
ins,
make_op("hip::allocate",
{{"shape", to_value(pack_int8_shape(inputs[0]->get_shape()))}}));
auto output_x =
m.insert_instruction(ins, make_op("gpu::int8_conv_pack"), {inputs[0], packed_x});
instruction::replace_argument(ins, inputs[0], output_x);
auto packed_w = m.insert_instruction(
ins,
make_op("hip::allocate",
{{"shape", to_value(pack_int8_shape(inputs[1]->get_shape()))}}));
auto output_w =
m.insert_instruction(ins, make_op("gpu::int8_conv_pack"), {inputs[1], packed_w});
instruction::replace_argument(ins, inputs[1], output_w);
}
}
}
shape pack_int8_args::pack_int8_shape(const shape& s) const
{
if(s.type() != shape::int8_type)
{
MIGRAPHX_THROW("PACK_INT8_ARGS: 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/pad.cpp deleted 100644 → 0
View file @ 6481cd69
/*
* 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/pad.hpp>
#include <migraphx/gpu/context.hpp>
#include <migraphx/gpu/device/pad.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
namespace gpu {
shape hip_pad::compute_shape(std::vector<shape> inputs) const
{
inputs.pop_back();
check_shapes{inputs, *this}.has(1).standard();
return op.compute_shape(inputs);
}
argument hip_pad::compute(context& ctx, const shape&, const std::vector<argument>& args) const
{
return device::pad(ctx.get_stream().get(), args.back(), args.front(), op.value, op.pads);
}
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/prefuse_ops.cpp
View file @ a24ed87e
......@@ -28,7 +28,10 @@
#include <migraphx/register_op.hpp>
#include <migraphx/pass_manager.hpp>
#include <migraphx/dead_code_elimination.hpp>
#ifdef MIGRAPHX_USE_COMPOSABLEKERNEL
#include <migraphx/gpu/ck.hpp>
#endif
#include <migraphx/gpu/fuse_mlir.hpp>
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......@@ -128,26 +131,49 @@ struct pre_gemm_softmax_gemm : gemm_softmax_gemm
};
MIGRAPHX_REGISTER_OP(pre_gemm_softmax_gemm);
MIGRAPHX_PRED_MATCHER(is_ck_gemm, instruction_ref ins)
auto is_ck_gemm()
{
if(ins->name() != "dot")
return match::make_basic_pred_matcher([=](instruction_ref ins) {
#ifdef MIGRAPHX_USE_COMPOSABLEKERNEL
if(not enabled(MIGRAPHX_ENABLE_CK{}))
return false;
if(ins->name() != "dot")
return false;
if(not pre_gemm_softmax_gemm::is_ck_supported_type(ins->get_shape().type()))
return false;
return true;
#else
(void)ins;
return false;
if(not pre_gemm_softmax_gemm::is_ck_supported_type(ins->get_shape().type()))
return false;
return true;
#endif
});
}
auto is_mlir_gemm()
{
return match::make_basic_pred_matcher([=](instruction_ref ins) {
if(not mlir_attention_enabled())
return false;
if(ins->name() != "dot")
return false;
return std::all_of(ins->inputs().begin(), ins->inputs().end(), [&](auto i) {
return pre_gemm_softmax_gemm::is_mlir_supported_type(i->get_shape().type());
});
});
}
struct find_gemm_softmax_gemm
{
auto matcher() const
{
auto gemm1 =
match::skip(match::name("contiguous"))(match::name("dot")(is_ck_gemm().bind("gemm1")));
auto gemm1 = match::skip(match::name("contiguous"))(
match::name("dot")(match::any_of(is_ck_gemm(), is_mlir_gemm()).bind("gemm1")));
auto mul = match::name("mul")(
match::nargs(2), match::either_arg(0, 1)(match::is_constant().bind("scale"), gemm1));
auto softmax = match::name("softmax")(match::arg(0)(mul)).bind("softmax");
return match::name("dot")(is_ck_gemm().bind("gemm2"))(match::arg(0)(softmax));
return match::name("dot")(match::any_of(is_ck_gemm(), is_mlir_gemm()).bind("gemm2"))(
match::arg(0)(softmax));
}
void apply(module_pass_manager& mpm, const match::matcher_result& r) const
......@@ -182,8 +208,7 @@ void prefuse_ops::apply(module_pass_manager& mpm) const
match::find_matches(mpm.get_module(), find_layernorm{});
mpm.run_pass(dead_code_elimination{});
match::find_matches(mpm.get_module(), find_add_layernorm{});
if(enabled(MIGRAPHX_ENABLE_CK{}))
match::find_matches(mpm, find_gemm_softmax_gemm{});
match::find_matches(mpm, find_gemm_softmax_gemm{});
}
} // namespace gpu
......
src/targets/gpu/rocblas.cpp
View file @ a24ed87e
......@@ -53,19 +53,16 @@ bool get_compute_fp32_flag()
return (starts_with(device_name, "gfx9") and device_name >= "gfx908");
}
bool get_int8_x4_format(context& ctx)
bool rocblas_fp8_available()
{
#if ROCBLAS_VERSION_MAJOR >= 3
(void)(ctx);
#ifndef MIGRAPHX_USE_ROCBLAS_FP8_API
return false;
#else
// int8x4 packed format is only available starting from rocblas-v2.38 and it is deprecated in
// v3.0 and will be removed in v4.0
rocblas_gemm_flags flag;
rocblas_query_int8_layout_flag(ctx.get_stream().get_rocblas(), &flag);
return flag == rocblas_gemm_flags_pack_int8x4;
const auto device_name = trim(split_string(get_device_name(), ':').front());
return (starts_with(device_name, "gfx9") and device_name >= "gfx940");
#endif
}
} // namespace gpu
} // namespace MIGRAPHX_INLINE_NS
} // namespace migraphx
src/targets/gpu/target.cpp
View file @ a24ed87e
......@@ -63,7 +63,6 @@
#include <migraphx/gpu/fuse_ops.hpp>
#include <migraphx/gpu/prefuse_ops.hpp>
#include <migraphx/gpu/lowering.hpp>
#include <migraphx/gpu/pack_int8_args.hpp>
#include <migraphx/gpu/schedule_model.hpp>
#include <migraphx/gpu/sync_device.hpp>
#include <migraphx/gpu/target.hpp>
......@@ -99,12 +98,28 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
ctx.set_exhaustive_tune_flag(options.exhaustive_tune);
std::set<shape::type_t> unsupported_types(shape::types().begin(), shape::types().end());
unsupported_types.erase(shape::type_t::float_type);
unsupported_types.erase(shape::type_t::fp8e4m3fnuz_type);
unsupported_types.erase(shape::type_t::half_type);
unsupported_types.erase(shape::type_t::bool_type);
unsupported_types.erase(shape::type_t::int8_type);
unsupported_types.erase(shape::type_t::uint8_type);
unsupported_types.erase(shape::type_t::int32_type);
unsupported_types.erase(shape::type_t::tuple_type);
std::set<std::string> unsupported_fp8_ops = {};
if(not gpu::rocblas_fp8_available())
{
unsupported_fp8_ops.insert("dot");
}
// add all device kernels
unsupported_fp8_ops.insert("logsoftmax");
unsupported_fp8_ops.insert("nonzero");
unsupported_fp8_ops.insert("prefix_scan_sum");
unsupported_fp8_ops.insert("scatter_none");
unsupported_fp8_ops.insert("topk");
unsupported_fp8_ops.insert("rnn_var_sl_shift_output");
unsupported_fp8_ops.insert("multinomial");
unsupported_fp8_ops.insert("argmax");
unsupported_fp8_ops.insert("argmin");
// clang-format off
return
{
......@@ -136,6 +151,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
prefuse_ops{},
dead_code_elimination{},
auto_contiguous{},
eliminate_data_type{{migraphx::shape::fp8e4m3fnuz_type}, shape::float_type, unsupported_fp8_ops},
dead_code_elimination{},
optimize_module{},
fuse_pointwise{},
dead_code_elimination{},
......@@ -154,7 +171,6 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
dead_code_elimination{},
compile_miopen{&gctx},
dead_code_elimination{},
pack_int8_args{},
dead_code_elimination{},
fuse_ops{&ctx, options.fast_math},
dead_code_elimination{},
......
src/targets/ref/CMakeLists.txt
View file @ a24ed87e
......@@ -33,8 +33,9 @@ rocm_set_soversion(migraphx_ref ${MIGRAPHX_SO_VERSION})
find_path(BLAZE_INCLUDE blaze/Blaze.h)
rocm_clang_tidy_check(migraphx_ref)
target_link_libraries(migraphx_ref PRIVATE Threads::Threads)
target_link_libraries(migraphx_ref PUBLIC migraphx)
target_include_directories(migraphx_ref PRIVATE ${BLAZE_INCLUDE})
target_include_directories(migraphx_ref SYSTEM PRIVATE ${BLAZE_INCLUDE})
target_compile_definitions(migraphx_ref PRIVATE -DBLAZE_USE_CPP_THREADS)
migraphx_generate_export_header(migraphx_ref)
......
src/tf/CMakeLists.txt
View file @ a24ed87e
......@@ -38,7 +38,11 @@ protobuf_generate_cpp(
)
add_library(tf-proto STATIC ${PROTO_SRCS})
target_include_directories(tf-proto SYSTEM PUBLIC ${CMAKE_CURRENT_BINARY_DIR} ${PROTOBUF_INCLUDE_DIR})
target_compile_options(tf-proto PRIVATE -w)
if(MSVC)
target_compile_options(tf-proto PRIVATE /w)
else()
target_compile_options(tf-proto PRIVATE -w)
endif()
target_link_libraries(tf-proto PRIVATE ${PROTOBUF_LIBRARY})
set_target_properties(tf-proto PROPERTIES POSITION_INDEPENDENT_CODE On)
......@@ -49,7 +53,10 @@ target_include_directories(migraphx_tf PRIVATE include)
set_target_properties(migraphx_tf PROPERTIES EXPORT_NAME tf)
rocm_set_soversion(migraphx_tf ${MIGRAPHX_SO_VERSION})
rocm_clang_tidy_check(migraphx_tf)
target_link_libraries(migraphx_tf PRIVATE tf-proto "-Wl,--exclude-libs,ALL")
target_link_libraries(migraphx_tf PRIVATE tf-proto)
if(NOT WIN32)
target_link_libraries(migraphx_tf PRIVATE "-Wl,--exclude-libs,ALL")
endif()
target_link_libraries(migraphx_tf PUBLIC migraphx)
rocm_install_targets(
......
src/tmp_dir.cpp
View file @ a24ed87e
......@@ -31,8 +31,18 @@
#include <sstream>
#include <iostream>
#include <string>
#include <sys/types.h>
#ifdef _WIN32
// cppcheck-suppress definePrefix
#define WIN32_LEAN_AND_MEAN
#include <Windows.h>
#undef getpid
// cppcheck-suppress [definePrefix, defineUpperCase]
#define getpid _getpid
#else
#include <unistd.h>
#include <sys/types.h>
#endif
namespace migraphx {
inline namespace MIGRAPHX_INLINE_NS {
......
src/verify_args.cpp
View file @ a24ed87e
......@@ -88,7 +88,6 @@ bool verify_args(const std::string& name,
if(target_nan_idx >= 0)
std::cout << "Non finite number found in target at " << target_nan_idx << ": "
<< target[target_nan_idx] << std::endl;
std::cout << "MIGraphX verification passed successfully." << std::endl;
}
});
return passed;
......
src/version.h.in
View file @ a24ed87e
......@@ -25,5 +25,5 @@
#define MIGRAPHX_VERSION_MAJOR @PROJECT_VERSION_MAJOR@
#define MIGRAPHX_VERSION_MINOR @PROJECT_VERSION_MINOR@
#define MIGRAPHX_VERSION_PATCH @PROJECT_VERSION_PATCH@
#define MIGRAPHX_VERSION_TWEAK @PROJECT_VERSION_TWEAK@
#define MIGRAPHX_VERSION_TWEAK "@PROJECT_VERSION_TWEAK@"
// clang-format on
test/CMakeLists.txt
View file @ a24ed87e
......@@ -25,7 +25,7 @@
cmake_policy(SET CMP0057 NEW)
find_package(Threads REQUIRED)
rocm_test_link_libraries(Threads::Threads migraphx migraphx_ref migraphx_onnx migraphx_tf)
rocm_test_link_libraries(Threads::Threads migraphx migraphx_onnx migraphx_tf)
rocm_test_include_directories(include)
set(MIGRAPHX_DISABLE_LARGE_BUFFER_TESTS Off CACHE BOOL "")
......@@ -146,7 +146,11 @@ endfunction()
function(test_headers PREFIX)
file(GLOB HEADERS CONFIGURE_DEPENDS ${ARGN})
if(NOT MIGRAPHX_USE_COMPOSABLEKERNEL)
list(REMOVE_ITEM HEADERS
${CMAKE_SOURCE_DIR}/src/targets/gpu/include/migraphx/gpu/ck.hpp)
endif()
list(REMOVE_ITEM HEADERS ${CMAKE_SOURCE_DIR}/src/include/migraphx/float8_impl.hpp)
foreach(HEADER ${HEADERS})
file(RELATIVE_PATH HEADER_REL ${CMAKE_SOURCE_DIR} ${HEADER})
string(MAKE_C_IDENTIFIER ${HEADER_REL} TEST_NAME)
......
test/api/CMakeLists.txt
View file @ a24ed87e
......@@ -30,6 +30,9 @@ function(add_api_test TEST_NAME TEST_SRC TEST_DIR)
add_test(NAME ${NAME} COMMAND $<TARGET_FILE:${NAME}> WORKING_DIRECTORY ${TEST_DIR})
add_dependencies(tests ${NAME})
add_dependencies(check ${NAME})
if(WIN32)
target_compile_definitions(${NAME} PRIVATE _CRT_SECURE_NO_WARNINGS)
endif()
endfunction()
# Workaround: C file dont work with clang-tidy right now, need a fix in rocm-cmake
......@@ -41,6 +44,9 @@ function(add_c_api_test TEST_NAME TEST_SRC TEST_DIR)
add_test(NAME ${NAME} COMMAND $<TARGET_FILE:${NAME}> WORKING_DIRECTORY ${TEST_DIR})
add_dependencies(tests ${NAME})
add_dependencies(check ${NAME})
if(WIN32)
target_compile_definitions(${NAME} PRIVATE _CRT_SECURE_NO_WARNINGS)
endif()
endfunction()
add_api_test(array_base test_array_base.cpp ${TEST_ONNX_DIR})
......@@ -57,10 +63,6 @@ add_api_test(custom_op test_custom_op.cpp ${TEST_ONNX_DIR})
add_api_test(tf_parser test_tf_parser.cpp ${TEST_TF_DIR})
# GPU-based tests
if(MIGRAPHX_ENABLE_GPU)
list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
find_package(hip)
add_api_test(gpu test_gpu.cpp ${TEST_ONNX_DIR})
target_link_libraries(test_api_gpu)
add_api_test(custom_op_gpu test_custom_op_gpu.cpp ${TEST_ONNX_DIR})
target_link_libraries(test_api_custom_op_gpu)
endif()
test/api/test_cpu.cpp
View file @ a24ed87e
......@@ -198,4 +198,29 @@ TEST_CASE(set_loop_default_iter_num)
EXPECT(out_shapes[1].lengths() == out_lens1);
}
TEST_CASE(set_loop_limit_iterations)
{
migraphx::onnx_options option;
option.set_default_loop_iterations(15);
option.set_limit_loop_iterations(10);
auto p = migraphx::parse_onnx("loop_default_test.onnx", option);
auto out_shapes = p.get_output_shapes();
std::vector<std::size_t> out_lens0 = {1};
EXPECT(out_shapes[0].lengths() == out_lens0);
std::vector<std::size_t> out_lens1 = {10, 1};
EXPECT(out_shapes[1].lengths() == out_lens1);
}
TEST_CASE(set_loop_limit_iterations2)
{
migraphx::onnx_options option;
option.set_limit_loop_iterations(10);
auto p = migraphx::parse_onnx("loop_test_implicit_tripcnt.onnx", option);
auto out_shapes = p.get_output_shapes();
std::vector<std::size_t> out_lens0 = {1};
EXPECT(out_shapes[0].lengths() == out_lens0);
std::vector<std::size_t> out_lens1 = {10, 1};
EXPECT(out_shapes[1].lengths() == out_lens1);
}
int main(int argc, const char* argv[]) { test::run(argc, argv); }
test/api/test_gpu.cpp
View file @ a24ed87e
......@@ -317,4 +317,59 @@ TEST_CASE(loop_test)
}
}
TEST_CASE(loop_test_limit_max_iter)
{
auto run_prog = [&](int64_t limit_max_iterations) {
migraphx::onnx_options parse_options;
parse_options.set_limit_loop_iterations(limit_max_iterations);
auto p = migraphx::parse_onnx("loop_test_implicit_tripcnt.onnx", parse_options);
auto shapes_before = p.get_output_shapes();
migraphx::compile_options options;
options.set_offload_copy();
p.compile(migraphx::target("gpu"), options);
auto shapes_after = p.get_output_shapes();
CHECK(shapes_before.size() == 2);
CHECK(bool{shapes_before.front() == shapes_after.front()});
migraphx::program_parameters pp;
auto param_shapes = p.get_parameter_shapes();
auto aas = param_shapes["a"];
std::vector<float> xd = {1.0f};
pp.add("a", migraphx::argument(aas, xd.data()));
auto bbs = param_shapes["b"];
std::vector<float> yd = {2.0};
pp.add("b", migraphx::argument(bbs, yd.data()));
auto cs = param_shapes["keep_going_cond"];
bool cond = true;
pp.add("keep_going_cond", migraphx::argument(cs, &cond));
auto outputs = p.eval(pp);
auto output = outputs[0];
std::vector<std::vector<float>> ret;
ret.push_back(output.as_vector<float>());
output = outputs[1];
ret.push_back(output.as_vector<float>());
return ret;
};
{
auto result_vector = run_prog(5);
std::vector<float> gold0 = {2.0f};
EXPECT(result_vector.at(0) == gold0);
std::vector<float> gold1 = {-2, 4, 0, 0, 0};
EXPECT(result_vector.at(1) == gold1);
}
{
auto result_vector = run_prog(20);
std::vector<float> gold0 = {2.0f};
EXPECT(result_vector.at(0) == gold0);
std::vector<float> gold1 = {-2, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
EXPECT(result_vector.at(1) == gold1);
}
}
int main(int argc, const char* argv[]) { test::run(argc, argv); }
test/float_equal.cpp
View file @ a24ed87e
......@@ -22,6 +22,7 @@
* THE SOFTWARE.
*/
#include <migraphx/float_equal.hpp>
#include <migraphx/float8.hpp>
#include <migraphx/half.hpp>
#include "test.hpp"
......@@ -53,7 +54,7 @@ auto test_float_equal(T x, U y)
template <class T, class U>
void test_equality()
{
auto x1 = T(0.1);
auto x1 = T(0.125);
auto x2 = U(0.0);
auto x3 = U(1.0);
EXPECT(test_float_equal(x1, x1));
......@@ -71,8 +72,12 @@ void test_equality()
TEST_CASE_REGISTER(test_equality<double, float>);
TEST_CASE_REGISTER(test_equality<double, int>);
TEST_CASE_REGISTER(test_equality<double, migraphx::half>);
TEST_CASE_REGISTER(test_equality<double, migraphx::fp8::fp8e4m3fnuz>);
TEST_CASE_REGISTER(test_equality<float, int>);
TEST_CASE_REGISTER(test_equality<float, migraphx::fp8::fp8e4m3fnuz>);
TEST_CASE_REGISTER(test_equality<migraphx::half, int>);
TEST_CASE_REGISTER(test_equality<migraphx::half, migraphx::fp8::fp8e4m3fnuz>);
TEST_CASE_REGISTER(test_equality<migraphx::fp8::fp8e4m3fnuz, int>);
template <class T, class U>
void test_limits()
......@@ -110,8 +115,13 @@ void test_limits()
TEST_CASE_REGISTER(test_limits<double, float>);
TEST_CASE_REGISTER(test_limits<double, int>);
TEST_CASE_REGISTER(test_limits<double, migraphx::half>);
TEST_CASE_REGISTER(test_limits<double, migraphx::fp8::fp8e4m3fnuz>);
TEST_CASE_REGISTER(test_limits<float, int>);
TEST_CASE_REGISTER(test_limits<float, migraphx::fp8::fp8e4m3fnuz>);
TEST_CASE_REGISTER(test_limits<int, migraphx::half>);
TEST_CASE_REGISTER(test_limits<int, migraphx::fp8::fp8e4m3fnuz>);
TEST_CASE_REGISTER(test_limits<migraphx::fp8::fp8e4m3fnuz, migraphx::half>);
#ifndef _WIN32
// On Windows, types int and long have the same min and max values.
TEST_CASE_REGISTER(test_limits<long, int>);
......
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