Skip to content

GitLab

Commit 43777959 authored by aska-0096's avatar aska-0096
Browse files

Fix errors in 

1. example, fmha
2. gridwise pipeline
3. deviceop, fmha, change some containers from vector to array
parent 83d926dc
Hide whitespace changes
Inline Side-by-side
Showing with 126 additions and 188 deletions
example/32_batched_gemm_scale_softmax_gemm/run_batched_gemm_scale_softmax_gemm_permute.inc
View file @ 43777959
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[]) int run(int argc, char* argv[])
{ {
...@@ -117,41 +117,6 @@ int run(int argc, char* argv[]) ...@@ -117,41 +117,6 @@ int run(int argc, char* argv[])
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{}); b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{}); b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
break; break;
case 4: // A, B0, B1 1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 5: // Rand: b1 b0; unit: a
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 6: // Rand: a b0 ; unit: B1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 7: // Rand: a b1 ; unit: b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
case 8: // Rand: a ; unit: b0 b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 9: // Rand: b0 ; unit: a b1
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_1<B1DataType>{});
break;
case 10: // Rand: b1 ; unit: a b0
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
break;
default: default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{}); a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{}); b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
...@@ -175,166 +140,139 @@ int run(int argc, char* argv[]) ...@@ -175,166 +140,139 @@ int run(int argc, char* argv[])
auto c_element_op = CElementOp{}; auto c_element_op = CElementOp{};
// do GEMM // do GEMM
float best_perf = .0;
float best_time = .0;
int not_pass = 0;
std::string best_kernel = "";
printf("Verification: %s\n", do_verification ? "ON" : "OFF");
// TODO ANT: replace array with vector? // TODO ANT: replace array with vector?
ck::static_for<0, std::tuple_size_v<DeviceMHAFactory>, 1>{}([&](auto i) -> void { auto gemm = DeviceGemmInstance{};
const auto device_conv_mha_instance = std::get<i>(DeviceMHAFactory{}); auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(
using DeviceMHAInstance = ck::remove_cvref_t<decltype(device_conv_mha_instance)>; static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
auto gemm = DeviceMHAInstance{}; static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
auto invoker = gemm.MakeInvoker(); static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
auto argument = gemm.MakeArgument(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()), static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()), {}, // std::array<void*, 1> p_acc0_biases;
static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()), {}, // std::array<void*, 1> p_acc1_biases;
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()), a_gs_ms_ks_lengths,
M, a_gs_ms_ks_strides,
N, b0_gs_ns_ks_lengths,
K, b0_gs_ns_ks_strides,
O, b1_gs_os_ns_lengths,
G0, b1_gs_os_ns_strides,
G1, c_gs_ms_os_lengths,
alpha, c_gs_ms_os_strides,
input_permute, {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_lengths},
output_permute); {}, // std::array<std::vector<ck::index_t>, 1>{acc0_biases_gs_ms_ns_strides},
{}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_lengths},
if(!gemm.IsSupportedArgument(argument)) {}, // std::array<std::vector<ck::index_t>, 1>{acc1_biases_gs_ms_os_strides},
{ a_element_op,
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl; b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op);
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
// return 0; return 0;
} }
ck::index_t BatchCount = G0 * G1; ck::index_t BatchCount = G0 * G1;
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount; std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N + std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) * sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O) *
BatchCount; BatchCount;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time; float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time; float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< " GB/s, " << gemm.GetTypeString() << std::endl; << gemm.GetTypeString() << std::endl;
if(tflops > best_perf)
{ if(do_verification)
best_perf = tflops; {
best_time = ave_time * 1000; c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
best_kernel = gemm.GetTypeString();
} Tensor<ADataType> a_g_m_k({BatchCount, M, K});
if(do_verification) Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n({BatchCount, M, N}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
// gemm 0
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// masking
const auto mask = DeviceGemmInstance::C0MatrixMask(N);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
// gemm1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(
a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
{ {
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data()); rtol = 1e-2;
atol = 1e-2;
Tensor<ADataType> a_g_m_k({BatchCount, M, K});
Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
Tensor<Acc0DataType> acc0_g_m_n({BatchCount, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n({BatchCount, M, N}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
// gemm 0
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// masking
const auto mask = typename DeviceMHAInstance::C0MatrixMask(N);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
ref_softmax_invoker.Run(ref_softmax_argument);
// gemm1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n,
b1_g_n_o,
c_g_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<ADataType, ck::bhalf_t> && std::is_same_v<B0DataType, ck::bhalf_t> &&
std::is_same_v<B1DataType, ck::bhalf_t> && std::is_same_v<CDataType, ck::bhalf_t>)
{
rtol = 1e-2;
atol = 1e-2;
}
bool this_run_verification = ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData,
"Error: Incorrect results!",
rtol,
atol);
printf("Verification: %s, Pass: %s\n",
do_verification ? "ON" : "OFF",
this_run_verification ? "YES" : "NO");
if(!this_run_verification)
{
not_pass = 1;
printf("%d th MHA instance verification Failed \n", i.value);
}
} }
});
std::cout << "---------------------------------------------------------------------------------" return ck::utils::check_err(c_gs_ms_os_device_result.mData,
"-----------" c_gs_ms_os_host_result.mData,
<< std::endl; "Error: Incorrect results!",
std::cout << "Problem Size: BatchCount: " << G0 << ", HeadNum: " << G1 << ", M: " << M rtol,
<< ", N: " << N << ", K: " << K << ", O: " << O << std::endl; atol)
std::cout << "---------------------------------------------------------------------------------" ? 0
"-----------" : 1;
<< std::endl; }
std::cout << "Best kernel: " << best_kernel << " , " << best_perf << " TFlops , " << best_time
<< " us" << std::endl; return 0;
std::cout << "---------------------------------------------------------------------------------"
"-----------"
<< std::endl;
return not_pass;
} }
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