void transpose(Param<T> out, CParam<T> in, const bool conjugate,
const bool is32multiple) {
static const std::string source(transpose_cuh, transpose_cuh_len);
// clang-format off
auto transpose = getKernel("cuda::transpose", source,
{
TemplateTypename<T>(),
TemplateArg(conjugate),
TemplateArg(is32multiple)
},
{
DefineValue(TILE_DIM),
DefineValue(THREADS_Y)
}
);
// clang-format on
dim3 threads(kernel::THREADS_X, kernel::THREADS_Y);
int blk_x = divup(in.dims[0], TILE_DIM);
int blk_y = divup(in.dims[1], TILE_DIM);
dim3 blocks(blk_x * in.dims[2], blk_y * in.dims[3]);
const int maxBlocksY =
cuda::getDeviceProp(getActiveDeviceId()).maxGridSize[1];
blocks.z = divup(blocks.y, maxBlocksY);
blocks.y = divup(blocks.y, blocks.z);
EnqueueArgs qArgs(blocks, threads, getActiveStream());
transpose(qArgs, out, in, blk_x, blk_y);
POST_LAUNCH_CHECK();
}
void DefineFloatConstant(Symbol p)
{
int align = p->ty->align;
p->aname = FormatName(".flt%d", FloatNum++);
Align(p);
Print("%s:\t", p->aname);
DefineValue(p->ty, p->val);
}
