derek.holloway
7502bf64a4
CodeQL / Analyze (actions) (push) Canceled after 0s
CodeQL / Analyze (python) (push) Canceled after 0s
pre-commit / pre-commit (push) Canceled after 0s
Multi-Arch CI / setup (push) Canceled after 0s
Publish build_manylinux_x86_64 images / publish_build_manylinux_x86_64 (push) Canceled after 0s
Setup (Multi-Arch) / setup (push) Canceled after 0s
Multi-Arch CI / Linux::${{ fromJSON(needs.setup.outputs.linux_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubi10 image / publish_no_rocm_image_ubi10 (push) Canceled after 0s
Multi-Arch CI / Windows::${{ fromJSON(needs.setup.outputs.windows_build_config || '{}').build_variant_label || 'skip' }} (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Multi-Arch CI / Manifest Diff (push) Canceled after 0s
TheRock Manifest Diff Report / Generate Manifest Diff Report (push) Canceled after 0s
Multi-Arch CI / CI Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_ocl_rt images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04_rocgdb images / publish_no_rocm_image_ubuntu24_04_rocgdb (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_ocl_rt (push) Canceled after 0s
Publish TheRock Docker image / build-and-push-image (push) Canceled after 0s
Unit Tests / Unit Tests :: ubuntu-24.04 (push) Canceled after 0s
Unit Tests / Unit Tests :: windows-2022 (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_media (push) Canceled after 0s
Unit Tests / Unit Tests Summary (push) Canceled after 0s
Publish no_rocm_image_ubuntu24_04 images / publish_no_rocm_image_ubuntu24_04_openmpi (push) Canceled after 0s
4490 lines
171 KiB
C
4490 lines
171 KiB
C
// This file was automatically generated. Do not edit.
|
|
|
|
// We use a special strategy to force the generation of the PTX. This is mainly
|
|
// a fight against dead-code-elimination in the NVVM layer.
|
|
//
|
|
// The reason we need this strategy is because certain older versions of ptxas
|
|
// segfault when a non-sensical sequence of PTX is generated. So instead, we try
|
|
// to force the instantiation and compilation to PTX of all the overloads of the
|
|
// PTX wrapping functions.
|
|
//
|
|
// We do this by writing a function pointer of each overload to the kernel
|
|
// parameter `fn_ptr`.
|
|
//
|
|
// Because `fn_ptr` is possibly visible outside this translation unit, the
|
|
// compiler must compile all the functions which are stored.
|
|
|
|
__global__ void test_ld(void** fn_ptr)
|
|
{
|
|
#if __cccl_ptx_isa >= 100
|
|
NV_IF_TARGET(NV_PROVIDES_SM_50,
|
|
(
|
|
// ld.global.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld));));
|
|
#endif // __cccl_ptx_isa >= 100
|
|
|
|
#if __cccl_ptx_isa >= 100
|
|
NV_IF_TARGET(NV_PROVIDES_SM_50,
|
|
(
|
|
// ld.global.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld));));
|
|
#endif // __cccl_ptx_isa >= 100
|
|
|
|
#if __cccl_ptx_isa >= 100
|
|
NV_IF_TARGET(NV_PROVIDES_SM_50,
|
|
(
|
|
// ld.global.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld));));
|
|
#endif // __cccl_ptx_isa >= 100
|
|
|
|
#if __cccl_ptx_isa >= 100
|
|
NV_IF_TARGET(NV_PROVIDES_SM_50,
|
|
(
|
|
// ld.global.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld));));
|
|
#endif // __cccl_ptx_isa >= 100
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(cuda::ptx::ld_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(cuda::ptx::ld_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(cuda::ptx::ld_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(cuda::ptx::ld_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_normal.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_normal.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_normal.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_normal.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_normal.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_normal.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_unchanged.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_unchanged.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_unchanged.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_unchanged.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_unchanged.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_unchanged.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_first.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_first.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_first.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_first.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_first.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_first.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_first.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_last.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_last.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_last.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_last.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::evict_last.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::evict_last.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::evict_last.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::no_allocate.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::no_allocate.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::no_allocate.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::no_allocate.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.L1::no_allocate.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.L1::no_allocate.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 100
|
|
NV_IF_TARGET(NV_PROVIDES_SM_50,
|
|
(
|
|
// ld.global.nc.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc));));
|
|
#endif // __cccl_ptx_isa >= 100
|
|
|
|
#if __cccl_ptx_isa >= 100
|
|
NV_IF_TARGET(NV_PROVIDES_SM_50,
|
|
(
|
|
// ld.global.nc.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_nc));));
|
|
#endif // __cccl_ptx_isa >= 100
|
|
|
|
#if __cccl_ptx_isa >= 100
|
|
NV_IF_TARGET(NV_PROVIDES_SM_50,
|
|
(
|
|
// ld.global.nc.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_nc));));
|
|
#endif // __cccl_ptx_isa >= 100
|
|
|
|
#if __cccl_ptx_isa >= 100
|
|
NV_IF_TARGET(NV_PROVIDES_SM_50,
|
|
(
|
|
// ld.global.nc.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_nc));));
|
|
#endif // __cccl_ptx_isa >= 100
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_nc));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_nc_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_nc_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_nc_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_nc_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_nc_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_nc_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_nc_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_nc_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_nc_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_nc_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_nc_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_nc_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_nc_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_nc_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_nc_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_nc_L1_evict_normal));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_normal.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_normal_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_nc_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_nc_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_nc_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_unchanged.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_unchanged_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_first.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_first.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_nc_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_first.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_nc_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_first.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_nc_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_first.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_nc_L1_evict_first));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_first.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_first_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_last.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_last.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_nc_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_last.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_nc_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_last.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_nc_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::evict_last.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_nc_L1_evict_last));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::evict_last.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_evict_last_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(cuda::ptx::ld_nc_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(cuda::ptx::ld_nc_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(cuda::ptx::ld_nc_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_70,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(cuda::ptx::ld_nc_L1_no_allocate));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::64B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(
|
|
static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(cuda::ptx::ld_nc_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::64B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::64B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::64B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::64B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::128B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::128B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::128B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::128B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_75,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::128B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::256B.b8 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::256B.b16 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::256B.b32 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::256B.b64 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::256B.b128 dest, [addr];
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::64B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::64B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::64B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::64B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::64B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_64B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::128B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::128B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::128B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::128B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::128B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_128B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::256B.b8 dest, [addr], cache_policy;
|
|
* fn_ptr++ = reinterpret_cast<void*>(static_cast<int8_t (*)(cuda::ptx::space_global_t, const int8_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::256B.b16 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int16_t (*)(cuda::ptx::space_global_t, const int16_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::256B.b32 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int32_t (*)(cuda::ptx::space_global_t, const int32_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 740
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::256B.b64 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<int64_t (*)(cuda::ptx::space_global_t, const int64_t*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 740
|
|
|
|
#if __cccl_ptx_isa >= 830
|
|
NV_IF_TARGET(
|
|
NV_PROVIDES_SM_80,
|
|
(
|
|
// ld.global.nc.L1::no_allocate.L2::cache_hint.L2::256B.b128 dest, [addr], cache_policy;
|
|
* fn_ptr++ =
|
|
reinterpret_cast<void*>(static_cast<longlong2 (*)(cuda::ptx::space_global_t, const longlong2*, uint64_t)>(
|
|
cuda::ptx::ld_nc_L1_no_allocate_L2_cache_hint_L2_256B));));
|
|
#endif // __cccl_ptx_isa >= 830
|
|
}
|