Home
last modified time | relevance | path

Searched refs:blockDim (Results 1 – 10 of 10) sorted by relevance

/external/clang/test/SemaCUDA/
Dcuda-builtin-vars.cu21 out[i++] = blockDim.x; in kernel()
22 blockDim.x = 0; // expected-error {{no setter defined for property 'x'}} in kernel()
23 out[i++] = blockDim.y; in kernel()
24 blockDim.y = 0; // expected-error {{no setter defined for property 'y'}} in kernel()
25 out[i++] = blockDim.z; in kernel()
26 blockDim.z = 0; // expected-error {{no setter defined for property 'z'}} in kernel()
/external/eigen/unsupported/Eigen/CXX11/src/Tensor/
DTensorReductionCuda.h115 const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x; in ReductionInitKernel()
116 const Index num_threads = blockDim.x * gridDim.x; in ReductionInitKernel()
192 eigen_assert(blockDim.x == 1); in ReductionInitFullReduxKernelHalfFloat()
205 const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x; in ReductionInitKernelHalfFloat()
206 const Index num_threads = blockDim.x * gridDim.x; in ReductionInitKernelHalfFloat()
378 eigen_assert(blockDim.y == 1);
379 eigen_assert(blockDim.z == 1);
386 const Index input_col_blocks = divup<Index>(num_coeffs_to_reduce, blockDim.x * NumPerThread);
389 const Index num_threads = blockDim.x * gridDim.x;
390 const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x;
[all …]
DTensorConvolution.h584 const int first_plane = blockIdx.y * blockDim.y;
585 const int plane_stride = blockDim.y * gridDim.y;
592 for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) {
603 for (int i = threadIdx.x; i < num_x_output; i += blockDim.x) {
638 const int first_plane = blockIdx.z * blockDim.z;
639 const int plane_stride = blockDim.z * gridDim.z;
648 for (int j = threadIdx.y; j < num_y_input; j += blockDim.y) {
651 for (int i = threadIdx.x; i < num_x_input; i += blockDim.x) {
663 for (int j = threadIdx.y; j < num_y_output; j += blockDim.y) {
665 for (int i = threadIdx.x; i < num_x_output; i += blockDim.x) {
[all …]
DTensorRandom.h24 blockIdx.x * blockDim.x + threadIdx.x + in get_random_seed()
25 gridDim.x * blockDim.x * (blockIdx.y * blockDim.y + threadIdx.y); in get_random_seed()
DTensorExecutor.h239 const Index first_index = blockIdx.x * blockDim.x + threadIdx.x;
240 const Index step_size = blockDim.x * gridDim.x;
DTensorScan.h253 Index val = threadIdx.x + blockIdx.x * blockDim.x;
/external/clang/test/CodeGenCUDA/
Dcuda-builtin-vars.cu17 out[i++] = blockDim.x; // CHECK: call i32 @llvm.nvvm.read.ptx.sreg.ntid.x() in kernel()
18 out[i++] = blockDim.y; // CHECK: call i32 @llvm.nvvm.read.ptx.sreg.ntid.y() in kernel()
19 out[i++] = blockDim.z; // CHECK: call i32 @llvm.nvvm.read.ptx.sreg.ntid.z() in kernel()
/external/eigen/test/
Dcuda_common.h11 dim3 threadIdx, blockDim, blockIdx; variable
26 int i = threadIdx.x + blockIdx.x*blockDim.x; in run_on_cuda_meta_kernel()
/external/clang/lib/Headers/
Dcuda_builtin_vars.h114 __CUDA_BUILTIN_VAR __cuda_builtin_blockDim_t blockDim; variable
/external/llvm/docs/
DNVPTXUsage.rst230 ``blockDim`` ``@llvm.nvvm.read.ptx.sreg.ntid.*``
567 ``i32 @llvm.nvvm.read.ptx.sreg.ntid.{x,y,z}`` blockDim.{x,y,z}