Skip to content

GitLab

"...composable_kernel_rocm.git" did not exist on "63eee2d9991b08ca286f6895dd8f90da12a62da3"
Commit e0b473b6 authored by aska-0096's avatar aska-0096
Browse files

Update warpshuffle

parent 4ee40bcc
Hide whitespace changes
Inline Side-by-side
Showing with 153 additions and 113 deletions
example/66_complex_contraction_bilinear/run_complex_contraction_bilinear_example.inc 100755 → 100644
View file @ e0b473b6
...@@ -127,44 +127,47 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -127,44 +127,47 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
switch(init_method) switch(init_method)
{ {
case 0: break; case 0: break;
case 1: case 1:
a_ms_ks_re.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}); a_ms_ks_re.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_ns_ks_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}); b_ns_ks_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}); d_ms_ns_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
a_ms_ks_img.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}); a_ms_ks_img.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_ns_ks_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}); b_ns_ks_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}); d_ms_ns_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break; break;
default: default:
a_ms_ks_re.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}); a_ms_ks_re.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_ns_ks_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}); b_ns_ks_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}); d_ms_ns_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
a_ms_ks_img.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}); a_ms_ks_img.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_ns_ks_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}); b_ns_ks_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}); d_ms_ns_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break; break;
} }
DeviceMem a_device_buf_re(sizeof(ADataType) * a_ms_ks_re.mDesc.GetElementSpaceSize()); DeviceMem a_device_buf_re(sizeof(ADataType) * a_ms_ks_re.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf_re(sizeof(BDataType) * b_ns_ks_re.mDesc.GetElementSpaceSize()); DeviceMem b_device_buf_re(sizeof(BDataType) * b_ns_ks_re.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf_re(sizeof(DDataType) * d_ms_ns_re.mDesc.GetElementSpaceSize()); DeviceMem d_device_buf_re(sizeof(DDataType) * d_ms_ns_re.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_re(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize()); DeviceMem e_device_buf_re(sizeof(EDataType) *
e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
DeviceMem a_device_buf_img(sizeof(ADataType) * a_ms_ks_img.mDesc.GetElementSpaceSize()); DeviceMem a_device_buf_img(sizeof(ADataType) * a_ms_ks_img.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf_img(sizeof(BDataType) * b_ns_ks_img.mDesc.GetElementSpaceSize()); DeviceMem b_device_buf_img(sizeof(BDataType) * b_ns_ks_img.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf_img(sizeof(DDataType) * d_ms_ns_img.mDesc.GetElementSpaceSize()); DeviceMem d_device_buf_img(sizeof(DDataType) * d_ms_ns_img.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_img(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize()); DeviceMem e_device_buf_img(sizeof(EDataType) *
e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
// Intermediate Value For E Real and Img // Intermediate Value For E Real and Img
DeviceMem e_device_buf_re1(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize()); DeviceMem e_device_buf_re1(sizeof(EDataType) *
DeviceMem e_device_buf_img1(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize()); e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf_img1(sizeof(EDataType) *
e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
a_device_buf_re.ToDevice(a_ms_ks_re.mData.data()); a_device_buf_re.ToDevice(a_ms_ks_re.mData.data());
b_device_buf_re.ToDevice(b_ns_ks_re.mData.data()); b_device_buf_re.ToDevice(b_ns_ks_re.mData.data());
...@@ -181,7 +184,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -181,7 +184,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
// set zero for intermediate values // set zero for intermediate values
e_device_buf_re1.SetZero(); e_device_buf_re1.SetZero();
e_device_buf_img1.SetZero(); e_device_buf_img1.SetZero();
auto a_element_op = AElementOp{}; auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{}; auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta}; auto cde_element_op = CDEElementOp{alpha, beta};
...@@ -189,23 +192,24 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -189,23 +192,24 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
// device operation // device operation
// For real Intermediate Value re_1 // For real Intermediate Value re_1
auto op = DeviceOpInstance{}; auto op = DeviceOpInstance{};
auto invoker = op.MakeInvoker(); auto invoker = op.MakeInvoker();
auto argument_re1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(), auto argument_re1 =
b_device_buf_re.GetDeviceBuffer(), op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf_re.GetDeviceBuffer()}, b_device_buf_re.GetDeviceBuffer(),
e_device_buf_re1.GetDeviceBuffer(), std::array<const void*, 1>{d_device_buf_re.GetDeviceBuffer()},
a_ms_ks_lengths, e_device_buf_re1.GetDeviceBuffer(),
a_ms_ks_strides, a_ms_ks_lengths,
b_ns_ks_lengths, a_ms_ks_strides,
b_ns_ks_strides, b_ns_ks_lengths,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths}, b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides}, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
e_ms_ns_lengths, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_strides, e_ms_ns_lengths,
a_element_op, e_ms_ns_strides,
b_element_op, a_element_op,
cde_element_op); b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument_re1)) if(!op.IsSupportedArgument(argument_re1))
{ {
...@@ -216,7 +220,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -216,7 +220,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
float ave_time_re1 = invoker.Run(argument_re1, StreamConfig{nullptr, time_kernel}); float ave_time_re1 = invoker.Run(argument_re1, StreamConfig{nullptr, time_kernel});
alpha = -1.f; alpha = -1.f;
beta = 1.f; beta = 1.f;
...@@ -228,21 +231,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -228,21 +231,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
// For real Intermediate Value re_2 // For real Intermediate Value re_2
// auto op = DeviceOpInstance{}; // auto op = DeviceOpInstance{};
// auto invoker = op.MakeInvoker(); // auto invoker = op.MakeInvoker();
auto argument_re2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(), auto argument_re2 =
b_device_buf_img.GetDeviceBuffer(), op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
std::array<const void*, 1>{e_device_buf_re1.GetDeviceBuffer()}, b_device_buf_img.GetDeviceBuffer(),
e_device_buf_re.GetDeviceBuffer(), std::array<const void*, 1>{e_device_buf_re1.GetDeviceBuffer()},
a_ms_ks_lengths, e_device_buf_re.GetDeviceBuffer(),
a_ms_ks_strides, a_ms_ks_lengths,
b_ns_ks_lengths, a_ms_ks_strides,
b_ns_ks_strides, b_ns_ks_lengths,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths}, b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides}, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
e_ms_ns_lengths, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_strides, e_ms_ns_lengths,
a_element_op, e_ms_ns_strides,
b_element_op, a_element_op,
cde_element_op); b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument_re2)) if(!op.IsSupportedArgument(argument_re2))
{ {
...@@ -253,7 +257,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -253,7 +257,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
float ave_time_re2 = invoker.Run(argument_re2, StreamConfig{nullptr, time_kernel}); float ave_time_re2 = invoker.Run(argument_re2, StreamConfig{nullptr, time_kernel});
alpha = 1.f; alpha = 1.f;
beta = 1.f; beta = 1.f;
...@@ -261,22 +264,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -261,22 +264,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
b_element_op = BElementOp{}; b_element_op = BElementOp{};
cde_element_op = CDEElementOp{alpha, beta}; cde_element_op = CDEElementOp{alpha, beta};
auto argument_img1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(), auto argument_img1 =
b_device_buf_img.GetDeviceBuffer(), op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf_img.GetDeviceBuffer()}, b_device_buf_img.GetDeviceBuffer(),
e_device_buf_img1.GetDeviceBuffer(), std::array<const void*, 1>{d_device_buf_img.GetDeviceBuffer()},
a_ms_ks_lengths, e_device_buf_img1.GetDeviceBuffer(),
a_ms_ks_strides, a_ms_ks_lengths,
b_ns_ks_lengths, a_ms_ks_strides,
b_ns_ks_strides, b_ns_ks_lengths,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths}, b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides}, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
e_ms_ns_lengths, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_strides, e_ms_ns_lengths,
a_element_op, e_ms_ns_strides,
b_element_op, a_element_op,
cde_element_op); b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument_img1)) if(!op.IsSupportedArgument(argument_img1))
{ {
...@@ -290,23 +293,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -290,23 +293,22 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
alpha = 1.f; alpha = 1.f;
beta = 1.f; beta = 1.f;
auto argument_img2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(), auto argument_img2 =
b_device_buf_re.GetDeviceBuffer(), op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
std::array<const void*, 1>{e_device_buf_img1.GetDeviceBuffer()}, b_device_buf_re.GetDeviceBuffer(),
e_device_buf_img.GetDeviceBuffer(), std::array<const void*, 1>{e_device_buf_img1.GetDeviceBuffer()},
a_ms_ks_lengths, e_device_buf_img.GetDeviceBuffer(),
a_ms_ks_strides, a_ms_ks_lengths,
b_ns_ks_lengths, a_ms_ks_strides,
b_ns_ks_strides, b_ns_ks_lengths,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths}, b_ns_ks_strides,
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides}, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
e_ms_ns_lengths, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_strides, e_ms_ns_lengths,
a_element_op, e_ms_ns_strides,
b_element_op, a_element_op,
cde_element_op); b_element_op,
cde_element_op);
if(!op.IsSupportedArgument(argument_img2)) if(!op.IsSupportedArgument(argument_img2))
{ {
...@@ -317,7 +319,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -317,7 +319,6 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
float ave_time_img2 = invoker.Run(argument_img2, StreamConfig{nullptr, time_kernel}); float ave_time_img2 = invoker.Run(argument_img2, StreamConfig{nullptr, time_kernel});
ck::index_t M = ck::index_t M =
ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{}); ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
...@@ -331,9 +332,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -331,9 +332,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(DDataType) * M * N + sizeof(EDataType) * M * N * 2; sizeof(DDataType) * M * N + sizeof(EDataType) * M * N * 2;
float ave_time = ave_time_img2 + ave_time_img1 + ave_time_re2 + ave_time_re1 ; float ave_time = ave_time_img2 + ave_time_img1 + ave_time_re2 + ave_time_re1;
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 << " GB/s, " std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
...@@ -343,7 +344,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -343,7 +344,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
e_device_buf_img.FromDevice(e_ms_ns_device_result_img.mData.data()); e_device_buf_img.FromDevice(e_ms_ns_device_result_img.mData.data());
auto isRealOk = 0; auto isRealOk = 0;
auto isImgOk = 0; auto isImgOk = 0;
if(do_verification) if(do_verification)
{ {
...@@ -366,17 +367,16 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -366,17 +367,16 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
auto ref_op = ReferenceOpInstance{}; auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_op.MakeInvoker(); auto ref_invoker = ref_op.MakeInvoker();
auto ref_argument_re = auto ref_argument_re = ref_op.MakeArgument(
ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_re, c_ms_ns_host_result_re, a_element_op, b_element_op); a_ms_ks_re, b_ns_ks_re, c_ms_ns_host_result_re, a_element_op, b_element_op);
ref_invoker.Run(ref_argument_re); ref_invoker.Run(ref_argument_re);
alpha = 1.f; alpha = 1.f;
beta = 1.f; beta = 1.f;
cde_element_op = CDEElementOp{alpha, beta}; cde_element_op = CDEElementOp{alpha, beta};
for(size_t m0 = 0; m0 < e_ms_ns_host_result_re.mDesc.GetLengths()[0]; ++m0) for(size_t m0 = 0; m0 < e_ms_ns_host_result_re.mDesc.GetLengths()[0]; ++m0)
{ {
for(size_t m1 = 0; m1 < e_ms_ns_host_result_re.mDesc.GetLengths()[1]; ++m1) for(size_t m1 = 0; m1 < e_ms_ns_host_result_re.mDesc.GetLengths()[1]; ++m1)
...@@ -395,11 +395,11 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -395,11 +395,11 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
alpha = 1.f; alpha = 1.f;
beta = -1.f; beta = -1.f;
cde_element_op = CDEElementOp{alpha, beta}; cde_element_op = CDEElementOp{alpha, beta};
auto ref_argument_re1 = auto ref_argument_re1 = ref_op.MakeArgument(
ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_img, c_ms_ns_host_result_re1, a_element_op, b_element_op); a_ms_ks_img, b_ns_ks_img, c_ms_ns_host_result_re1, a_element_op, b_element_op);
ref_invoker.Run(ref_argument_re1); ref_invoker.Run(ref_argument_re1);
...@@ -419,23 +419,20 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -419,23 +419,20 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
} }
} }
isRealOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1; isRealOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
// Img Part Verification // Img Part Verification
Tensor<CShuffleDataType> c_ms_ns_host_result_img(e_ms_ns_lengths, e_ms_ns_strides); Tensor<CShuffleDataType> c_ms_ns_host_result_img(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<CShuffleDataType> c_ms_ns_host_result_img1(e_ms_ns_lengths, e_ms_ns_strides); Tensor<CShuffleDataType> c_ms_ns_host_result_img1(e_ms_ns_lengths, e_ms_ns_strides);
auto ref_argument_img = auto ref_argument_img = ref_op.MakeArgument(
ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_img, c_ms_ns_host_result_img, a_element_op, b_element_op); a_ms_ks_re, b_ns_ks_img, c_ms_ns_host_result_img, a_element_op, b_element_op);
ref_invoker.Run(ref_argument_img); ref_invoker.Run(ref_argument_img);
alpha = 1.f; alpha = 1.f;
beta = 1.f; beta = 1.f;
cde_element_op = CDEElementOp{alpha, beta}; cde_element_op = CDEElementOp{alpha, beta};
for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0) for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
...@@ -454,9 +451,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -454,9 +451,9 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
} }
} }
auto ref_argument_img1 = auto ref_argument_img1 = ref_op.MakeArgument(
ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_re, c_ms_ns_host_result_img1, a_element_op, b_element_op); a_ms_ks_img, b_ns_ks_re, c_ms_ns_host_result_img1, a_element_op, b_element_op);
ref_invoker.Run(ref_argument_img1); ref_invoker.Run(ref_argument_img1);
for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0) for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
...@@ -475,7 +472,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[]) ...@@ -475,7 +472,7 @@ int run_complex_contraction_bilinear_example(int argc, char* argv[])
} }
} }
isImgOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1; isImgOk = ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
return (isRealOk && isImgOk); return (isRealOk && isImgOk);
} }
......
include/ck_tile/core/arch/utility.hpp
View file @ e0b473b6
...@@ -59,4 +59,47 @@ CK_TILE_DEVICE T warp_shuffle_down(const T& v_local, uint32_t lane_delta) ...@@ -59,4 +59,47 @@ CK_TILE_DEVICE T warp_shuffle_down(const T& v_local, uint32_t lane_delta)
#endif #endif
} }
template <typename T>
CK_TILE_DEVICE T warp_shuffle(const T& v_local, uint32_t src_lane)
{
#if 0
return __shfl(v_local, src_lane);
#elif 1
if constexpr(sizeof(int32_t) > sizeof(T))
{
union packet
{
int32_t x;
T v;
};
packet p;
p.v = v_local;
packet p_remote;
p_remote.x = __builtin_amdgcn_ds_bpermute(src_lane << 2, bit_cast<int32_t>(p));
return p_remote.v;
}
else if constexpr(sizeof(int32_t) == sizeof(T))
{
const int32_t v_remote_tmp =
__builtin_amdgcn_ds_bpermute(src_lane << 2, bit_cast<int32_t>(v_local));
return bit_cast<T>(v_remote_tmp);
}
else
{
static_assert(sizeof(T) % sizeof(int32_t) == 0, "wrong!");
constexpr index_t elm = sizeof(T) / sizeof(int32_t);
using vector_type = thread_buffer<int32_t, elm>;
auto vs = bit_cast<vector_type>(v_local);
auto vs_remote = vector_type{};
static_for<0, elm, 1>{}([&](auto i_e) {
int32_t tmp = __builtin_amdgcn_ds_bpermute(src_lane << 2, bit_cast<int32_t>(vs[i_e]));
vs_remote(i_e) = tmp;
});
return bit_cast<T>(vs_remote);
}
#endif
}
} // namespace ck_tile } // namespace ck_tile
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