AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Address post-staging issues in #1809 (#1894)

Browse Source 
Fixes SWDEV-223910 and SWDEV-223663

[ROCm/clr commit: cea5489f00]
This commit is contained in:
Alex Voicu
2020-02-27 12:51:12 +02:00
committed by GitHub
parent d805279060
commit 6a37ef97c4
1 changed files with 48 additions and 47 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/clr/hipamd/src/hip_memory.cpp
+44 -43
View File
@@ -139,6 +139,8 @@ namespace {
return r; return r;
}()}; }()};
constexpr std::uint32_t is_cpu_owned{UINT32_MAX};
inline inline
hsa_amd_pointer_info_t info(const void* p) hsa_amd_pointer_info_t info(const void* p)
{ {
@@ -148,14 +150,14 @@ namespace {
const_cast<void*>(p), &r, nullptr, nullptr, nullptr), const_cast<void*>(p), &r, nullptr, nullptr, nullptr),
__FILE__, __func__, __LINE__); __FILE__, __func__, __LINE__);
if (is_large_BAR) r.size = UINT32_MAX; if (type(r.agentOwner) == HSA_DEVICE_TYPE_CPU) r.size = is_cpu_owned;
else if (type(r.agentOwner) == HSA_DEVICE_TYPE_CPU) r.size = INT32_MAX;
return r; return r;
} }
constexpr size_t staging_sz{4 * 1024 * 1024}; // 2 Pages. constexpr size_t staging_sz{4 * 1024 * 1024}; // 2 Pages.
constexpr size_t max_std_memcpy_sz{8 * 1024}; // 8 KiB. constexpr size_t max_h2d_std_memcpy_sz{8 * 1024}; // 8 KiB.
constexpr size_t max_d2h_std_memcpy_sz{64}; // 1 cacheline.
thread_local const std::unique_ptr<void, void (*)(void *)> staging_buffer{ thread_local const std::unique_ptr<void, void (*)(void *)> staging_buffer{
[]() { []() {
@@ -224,14 +226,20 @@ void do_std_memcpy(
inline inline
void d2h_copy(void* __restrict dst, const void* __restrict src, size_t n, void d2h_copy(void* __restrict dst, const void* __restrict src, size_t n,
hsa_amd_pointer_info_t si) { hsa_amd_pointer_info_t si) {
if (si.size == INT32_MAX) return do_std_memcpy(dst, src, n); const auto di{info(dst)};
if (si.size == UINT32_MAX && n <= max_std_memcpy_sz) { const auto is_locked{di.type == HSA_EXT_POINTER_TYPE_LOCKED};
if (!is_locked && si.size == is_cpu_owned) {
return do_std_memcpy(dst, src, n); return do_std_memcpy(dst, src, n);
} }
if (!is_locked && is_large_BAR && n <= max_d2h_std_memcpy_sz) {
return do_std_memcpy(dst, src, n);
}
if (di.type == HSA_EXT_POINTER_TYPE_HSA) {
return do_copy(dst, src, n, si.agentOwner, si.agentOwner);
}
const auto di{info(dst)}; if (is_locked) {
if (di.type == HSA_EXT_POINTER_TYPE_LOCKED) {
dst = static_cast<char*>(di.agentBaseAddress) + dst = static_cast<char*>(di.agentBaseAddress) +
(static_cast<char*>(dst) - (static_cast<char*>(dst) -
static_cast<char*>(di.hostBaseAddress)); static_cast<char*>(di.hostBaseAddress));
@@ -247,7 +255,7 @@ void d2h_copy(void* __restrict dst, const void* __restrict src, size_t n,
throwing_result_check(hsa_amd_memory_lock(dst, n, &si.agentOwner, 1, throwing_result_check(hsa_amd_memory_lock(dst, n, &si.agentOwner, 1,
const_cast<void**>(&dst)), const_cast<void**>(&dst)),
__FILE__, __func__, __LINE__); __FILE__, __func__, __LINE__);
do_copy(dst, src, n, si.agentOwner, si.agentOwner); do_copy(dst, src, n, si.agentOwner, si.agentOwner);
} }
@@ -256,14 +264,20 @@ void d2h_copy(void* __restrict dst, const void* __restrict src, size_t n,
inline inline
void h2d_copy(void* __restrict dst, const void* __restrict src, size_t n, void h2d_copy(void* __restrict dst, const void* __restrict src, size_t n,
hsa_amd_pointer_info_t di) { hsa_amd_pointer_info_t di) {
if (di.size == INT32_MAX) return do_std_memcpy(dst, src, n); const auto si{info(const_cast<void*>(src))};
if (di.size == UINT32_MAX && n <= max_std_memcpy_sz) { const auto is_locked{si.type == HSA_EXT_POINTER_TYPE_LOCKED};
if (!is_locked && di.size == is_cpu_owned) {
return do_std_memcpy(dst, src, n); return do_std_memcpy(dst, src, n);
} }
if (!is_locked && is_large_BAR && n <= max_h2d_std_memcpy_sz) {
return do_std_memcpy(dst, src, n);
}
if (si.type == HSA_EXT_POINTER_TYPE_HSA) {
return do_copy(dst, src, n, di.agentOwner, di.agentOwner);
}
const auto si{info(const_cast<void*>(src))}; if (is_locked) {
if (si.type == HSA_EXT_POINTER_TYPE_LOCKED) {
src = static_cast<char*>(si.agentBaseAddress) + src = static_cast<char*>(si.agentBaseAddress) +
(static_cast<const char*>(src) - (static_cast<const char*>(src) -
static_cast<char*>(si.hostBaseAddress)); static_cast<char*>(si.hostBaseAddress));
@@ -280,7 +294,7 @@ void h2d_copy(void* __restrict dst, const void* __restrict src, size_t n,
throwing_result_check(hsa_amd_memory_lock(const_cast<void*>(src), n, throwing_result_check(hsa_amd_memory_lock(const_cast<void*>(src), n,
&di.agentOwner, 1, &di.agentOwner, 1,
const_cast<void**>(&src)), const_cast<void**>(&src)),
__FILE__, __func__, __LINE__); __FILE__, __func__, __LINE__);
do_copy(dst, src, n, di.agentOwner, di.agentOwner); do_copy(dst, src, n, di.agentOwner, di.agentOwner);
} }
@@ -289,36 +303,23 @@ void h2d_copy(void* __restrict dst, const void* __restrict src, size_t n,
inline inline
void generic_copy(void* __restrict dst, const void* __restrict src, size_t n, void generic_copy(void* __restrict dst, const void* __restrict src, size_t n,
hsa_amd_pointer_info_t di, hsa_amd_pointer_info_t si) { hsa_amd_pointer_info_t di, hsa_amd_pointer_info_t si) {
if (di.size == INT32_MAX && si.size == INT32_MAX) { if (di.size == is_cpu_owned && si.size == is_cpu_owned) {
return do_std_memcpy(dst, src, n);
}
if (di.size == UINT32_MAX && si.size == UINT32_MAX &&
n <= max_std_memcpy_sz) {
return do_std_memcpy(dst, src, n); return do_std_memcpy(dst, src, n);
} }
if (di.size == is_cpu_owned) return d2h_copy(dst, src, n, si);
if (si.size == is_cpu_owned) return h2d_copy(dst, src, n, di);
switch (type(si.agentOwner)) { throwing_result_check(hsa_amd_agents_allow_access(1u, &si.agentOwner,
case HSA_DEVICE_TYPE_GPU: nullptr,
if (type(di.agentOwner) == HSA_DEVICE_TYPE_GPU) { di.agentBaseAddress),
throwing_result_check( __FILE__, __func__, __LINE__);
hsa_amd_agents_allow_access(
1u, &si.agentOwner, nullptr, di.agentBaseAddress), return do_copy(dst, src, n, di.agentOwner, si.agentOwner);
__FILE__, __func__, __LINE__);
return do_copy(dst, src, n, di.agentOwner, si.agentOwner);
}
return d2h_copy(dst, src, n, si);
case HSA_DEVICE_TYPE_CPU:
if (type(di.agentOwner) == HSA_DEVICE_TYPE_CPU) {
return do_std_memcpy(dst, src, n);
}
return h2d_copy(dst, src, n, di);
default: throw std::runtime_error{"Unsupported copy type."};
}
} }
inline inline
void memcpy_impl(void* __restrict dst, const void* __restrict src, size_t n, void memcpy_impl(void* __restrict dst, const void* __restrict src, size_t n,
hipMemcpyKind k) noexcept { hipMemcpyKind k) {
switch (k) { switch (k) {
case hipMemcpyHostToHost: std::memcpy(dst, src, n); break; case hipMemcpyHostToHost: std::memcpy(dst, src, n); break;
case hipMemcpyHostToDevice: return h2d_copy(dst, src, n, info(dst)); case hipMemcpyHostToDevice: return h2d_copy(dst, src, n, info(dst));
@@ -326,10 +327,10 @@ void memcpy_impl(void* __restrict dst, const void* __restrict src, size_t n,
case hipMemcpyDeviceToDevice: { case hipMemcpyDeviceToDevice: {
const auto di{info(dst)}; const auto di{info(dst)};
const auto si{info(src)}; const auto si{info(src)};
throwing_result_check( throwing_result_check(hsa_amd_agents_allow_access(1u, &si.agentOwner,
hsa_amd_agents_allow_access( nullptr,
1u, &si.agentOwner, nullptr, di.agentBaseAddress), di.agentBaseAddress),
__FILE__, __func__, __LINE__); __FILE__, __func__, __LINE__);
return do_copy(dst, src, n, di.agentOwner, si.agentOwner); return do_copy(dst, src, n, di.agentOwner, si.agentOwner);
} }
default: return generic_copy(dst, src, n, info(dst), info(src)); default: return generic_copy(dst, src, n, info(dst), info(src));