Jaydeep Patel
289 rader
11 KiB
C++
289 rader
11 KiB
C++
/*
|
|
Copyright (c) 2023 Advanced Micro Devices, Inc. All rights reserved.
|
|
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
of this software and associated documentation files (the "Software"), to deal
|
|
in the Software without restriction, including without limitation the rights
|
|
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
copies of the Software, and to permit persons to whom the Software is
|
|
furnished to do so, subject to the following conditions:
|
|
The above copyright notice and this permission notice shall be included in
|
|
all copies or substantial portions of the Software.
|
|
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANNTY OF ANY KIND, EXPRESS OR
|
|
IMPLIED, INNCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
FITNNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANNY CLAIM, DAMAGES OR OTHER
|
|
LIABILITY, WHETHER INN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
OUT OF OR INN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
|
|
THE SOFTWARE.
|
|
*/
|
|
#pragma once
|
|
|
|
#include <hip_test_common.hh>
|
|
#include <resource_guards.hh>
|
|
#include <utils.hh>
|
|
|
|
namespace {
|
|
constexpr auto wait_ms = 500;
|
|
} // anonymous namespace
|
|
|
|
template <typename T> __global__ void kernel_500ms(T* host_res, int clk_rate) {
|
|
int tid = threadIdx.x + blockIdx.x * blockDim.x;
|
|
host_res[tid] = tid + 1;
|
|
__threadfence_system();
|
|
// expecting that the data is getting flushed to host here!
|
|
uint64_t start = clock64() / clk_rate, cur;
|
|
if (clk_rate > 1) {
|
|
do {
|
|
cur = clock64() / clk_rate - start;
|
|
} while (cur < wait_ms);
|
|
} else {
|
|
do {
|
|
cur = clock64() / start;
|
|
} while (cur < wait_ms);
|
|
}
|
|
}
|
|
|
|
template <typename T> __global__ void kernel_500ms_gfx11(T* host_res, int clk_rate) {
|
|
#if HT_AMD
|
|
int tid = threadIdx.x + blockIdx.x * blockDim.x;
|
|
host_res[tid] = tid + 1;
|
|
__threadfence_system();
|
|
// expecting that the data is getting flushed to host here!
|
|
uint64_t start = wall_clock64() / clk_rate, cur;
|
|
if (clk_rate > 1) {
|
|
do {
|
|
cur = wall_clock64() / clk_rate - start;
|
|
} while (cur < wait_ms);
|
|
} else {
|
|
do {
|
|
cur = wall_clock64() / start;
|
|
} while (cur < wait_ms);
|
|
}
|
|
#endif
|
|
}
|
|
|
|
template <typename F> void MallocMemPoolAsync_OneAlloc(F malloc_func, const MemPools mempool_type) {
|
|
int device_id = 0;
|
|
HIP_CHECK(hipSetDevice(device_id));
|
|
|
|
int mem_pool_support = 0;
|
|
HIP_CHECK(hipDeviceGetAttribute(&mem_pool_support, hipDeviceAttributeMemoryPoolsSupported, 0));
|
|
if (!mem_pool_support) {
|
|
SUCCEED("Runtime doesn't support Memory Pool. Skip the test case.");
|
|
return;
|
|
}
|
|
|
|
const auto allocation_size = GENERATE(kPageSize / 2, kPageSize, kPageSize * 2);
|
|
LinearAllocGuard<int> host_alloc(LinearAllocs::hipHostMalloc, allocation_size);
|
|
MemPoolGuard mempool(mempool_type, device_id);
|
|
|
|
int* alloc_mem;
|
|
StreamGuard stream(Streams::created);
|
|
|
|
HIP_CHECK(malloc_func(reinterpret_cast<void**>(&alloc_mem), allocation_size, mempool.mempool(),
|
|
stream.stream()));
|
|
|
|
int blocks = 16;
|
|
int clk_rate;
|
|
hipMemPoolAttr attr;
|
|
if (IsGfx11()) {
|
|
HIP_CHECK(hipDeviceGetAttribute(&clk_rate, hipDeviceAttributeWallClockRate, 0));
|
|
kernel_500ms_gfx11<<<32, blocks, 0, stream.stream()>>>(alloc_mem, clk_rate);
|
|
} else {
|
|
HIP_CHECK(hipDeviceGetAttribute(&clk_rate, hipDeviceAttributeClockRate, 0));
|
|
|
|
kernel_500ms<<<32, blocks, 0, stream.stream()>>>(alloc_mem, clk_rate);
|
|
}
|
|
|
|
const auto element_count = allocation_size / sizeof(int);
|
|
constexpr auto thread_count = 1024;
|
|
const auto block_count = element_count / thread_count + 1;
|
|
constexpr int expected_value = 17;
|
|
VectorSet<<<block_count, thread_count, 0, stream.stream()>>>(alloc_mem, expected_value,
|
|
element_count);
|
|
|
|
HIP_CHECK(hipMemcpyAsync(host_alloc.host_ptr(), alloc_mem, allocation_size, hipMemcpyDeviceToHost,
|
|
stream.stream()));
|
|
|
|
HIP_CHECK(hipFreeAsync(reinterpret_cast<void*>(alloc_mem), stream.stream()));
|
|
|
|
attr = hipMemPoolAttrReservedMemCurrent;
|
|
std::uint64_t res_before_sync = 0;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &res_before_sync));
|
|
HIP_CHECK(hipStreamSynchronize(stream.stream()));
|
|
|
|
std::uint64_t res_after_sync = 0;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &res_after_sync));
|
|
// Sync must release memory to OS
|
|
REQUIRE(res_after_sync <= res_before_sync);
|
|
|
|
std::uint64_t used_mem = 10;
|
|
attr = hipMemPoolAttrUsedMemCurrent;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &used_mem));
|
|
REQUIRE(0 == used_mem);
|
|
|
|
ArrayFindIfNot(host_alloc.host_ptr(), expected_value, element_count);
|
|
}
|
|
|
|
template <typename F>
|
|
void MallocMemPoolAsync_TwoAllocs(F malloc_func, const MemPools mempool_type) {
|
|
int device_id = 0;
|
|
HIP_CHECK(hipSetDevice(device_id));
|
|
|
|
int mem_pool_support = 0;
|
|
HIP_CHECK(hipDeviceGetAttribute(&mem_pool_support, hipDeviceAttributeMemoryPoolsSupported, 0));
|
|
if (!mem_pool_support) {
|
|
SUCCEED("Runtime doesn't support Memory Pool. Skip the test case.");
|
|
return;
|
|
}
|
|
|
|
const auto allocation_size = GENERATE(kPageSize / 2, kPageSize, kPageSize * 2);
|
|
LinearAllocGuard<int> host_alloc(LinearAllocs::hipHostMalloc, allocation_size);
|
|
MemPoolGuard mempool(mempool_type, device_id);
|
|
|
|
int* alloc_mem1;
|
|
int* alloc_mem2;
|
|
StreamGuard stream(Streams::created);
|
|
|
|
HIP_CHECK(malloc_func(reinterpret_cast<void**>(&alloc_mem1), allocation_size, mempool.mempool(),
|
|
stream.stream()));
|
|
HIP_CHECK(malloc_func(reinterpret_cast<void**>(&alloc_mem2), allocation_size, mempool.mempool(),
|
|
stream.stream()));
|
|
|
|
int blocks = 16;
|
|
int clk_rate;
|
|
hipMemPoolAttr attr;
|
|
if (IsGfx11()) {
|
|
HIP_CHECK(hipDeviceGetAttribute(&clk_rate, hipDeviceAttributeWallClockRate, 0));
|
|
kernel_500ms_gfx11<<<32, blocks, 0, stream.stream()>>>(alloc_mem1, clk_rate);
|
|
} else {
|
|
HIP_CHECK(hipDeviceGetAttribute(&clk_rate, hipDeviceAttributeClockRate, 0));
|
|
|
|
kernel_500ms<<<32, blocks, 0, stream.stream()>>>(alloc_mem1, clk_rate);
|
|
}
|
|
|
|
const auto element_count = allocation_size / sizeof(int);
|
|
constexpr auto thread_count = 1024;
|
|
const auto block_count = element_count / thread_count + 1;
|
|
constexpr int expected_value = 17;
|
|
VectorSet<<<block_count, thread_count, 0, stream.stream()>>>(alloc_mem1, expected_value,
|
|
element_count);
|
|
HIP_CHECK(hipGetLastError());
|
|
|
|
HIP_CHECK(hipMemcpyAsync(alloc_mem2, alloc_mem1, allocation_size, hipMemcpyDeviceToDevice,
|
|
stream.stream()));
|
|
|
|
HIP_CHECK(hipMemcpyAsync(host_alloc.host_ptr(), alloc_mem2, allocation_size,
|
|
hipMemcpyDeviceToHost, stream.stream()));
|
|
|
|
HIP_CHECK(hipFreeAsync(reinterpret_cast<void*>(alloc_mem1), stream.stream()));
|
|
|
|
attr = hipMemPoolAttrReservedMemCurrent;
|
|
std::uint64_t res_before_sync = 0;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &res_before_sync));
|
|
HIP_CHECK(hipStreamSynchronize(stream.stream()));
|
|
|
|
std::uint64_t res_after_sync = 0;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &res_after_sync));
|
|
// Sync must release memory to OS
|
|
REQUIRE(res_after_sync <= res_before_sync);
|
|
|
|
std::uint64_t used_mem = 0;
|
|
attr = hipMemPoolAttrUsedMemCurrent;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &used_mem));
|
|
// Make sure the current usage query works - just second buffer is left
|
|
REQUIRE(allocation_size == used_mem);
|
|
|
|
attr = hipMemPoolAttrUsedMemHigh;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &used_mem));
|
|
// Make sure the high watermark usage works - both buffers must be reported
|
|
REQUIRE((2 * allocation_size) == used_mem);
|
|
|
|
HIP_CHECK(hipFreeAsync(reinterpret_cast<void*>(alloc_mem2), stream.stream()));
|
|
HIP_CHECK(hipStreamSynchronize(stream.stream()));
|
|
|
|
attr = hipMemPoolAttrUsedMemCurrent;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &used_mem));
|
|
// Make sure the current usage query works - none of the buffers are used
|
|
REQUIRE(0 == used_mem);
|
|
|
|
ArrayFindIfNot(host_alloc.host_ptr(), expected_value, element_count);
|
|
}
|
|
|
|
template <typename F> void MallocMemPoolAsync_Reuse(F malloc_func, const MemPools mempool_type) {
|
|
int device_id = 0;
|
|
HIP_CHECK(hipSetDevice(device_id));
|
|
|
|
int mem_pool_support = 0;
|
|
HIP_CHECK(hipDeviceGetAttribute(&mem_pool_support, hipDeviceAttributeMemoryPoolsSupported, 0));
|
|
if (!mem_pool_support) {
|
|
SUCCEED("Runtime doesn't support Memory Pool. Skip the test case.");
|
|
return;
|
|
}
|
|
|
|
MemPoolGuard mempool(mempool_type, device_id);
|
|
|
|
int *alloc_mem1, *alloc_mem2, *alloc_mem3;
|
|
StreamGuard stream(Streams::created);
|
|
|
|
size_t allocation_size1 = kPageSize * kPageSize * 2;
|
|
HIP_CHECK(malloc_func(reinterpret_cast<void**>(&alloc_mem1), allocation_size1, mempool.mempool(),
|
|
stream.stream()));
|
|
|
|
size_t allocation_size2 = kPageSize;
|
|
HIP_CHECK(malloc_func(reinterpret_cast<void**>(&alloc_mem3), allocation_size2, mempool.mempool(),
|
|
stream.stream()));
|
|
|
|
int blocks = 2;
|
|
int clk_rate;
|
|
|
|
if (IsGfx11()) {
|
|
HIP_CHECK(hipDeviceGetAttribute(&clk_rate, hipDeviceAttributeWallClockRate, 0));
|
|
kernel_500ms_gfx11<<<32, blocks, 0, stream.stream()>>>(alloc_mem1, clk_rate);
|
|
} else {
|
|
HIP_CHECK(hipDeviceGetAttribute(&clk_rate, hipDeviceAttributeClockRate, 0));
|
|
|
|
kernel_500ms<<<32, blocks, 0, stream.stream()>>>(alloc_mem1, clk_rate);
|
|
}
|
|
|
|
hipMemPoolAttr attr;
|
|
// Not a real free, since kernel isn't done
|
|
HIP_CHECK(hipFreeAsync(reinterpret_cast<void*>(alloc_mem1), stream.stream()));
|
|
|
|
HIP_CHECK(malloc_func(reinterpret_cast<void**>(&alloc_mem2), allocation_size1, mempool.mempool(),
|
|
stream.stream()));
|
|
// Runtime must reuse the pointer
|
|
REQUIRE(alloc_mem1 == alloc_mem2);
|
|
|
|
// Make a sync before the second kernel launch to make sure memory B isn't gone
|
|
HIP_CHECK(hipStreamSynchronize(stream.stream()));
|
|
|
|
// Second kernel launch with new memory
|
|
if (IsGfx11()) {
|
|
kernel_500ms_gfx11<<<32, blocks, 0, stream.stream()>>>(alloc_mem2, clk_rate);
|
|
} else {
|
|
kernel_500ms<<<32, blocks, 0, stream.stream()>>>(alloc_mem2, clk_rate);
|
|
}
|
|
|
|
HIP_CHECK(hipStreamSynchronize(stream.stream()));
|
|
|
|
attr = hipMemPoolAttrUsedMemCurrent;
|
|
std::uint64_t value64 = 0;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &value64));
|
|
// Make sure the current usage reports the both buffers
|
|
REQUIRE((allocation_size1 + allocation_size2) == value64);
|
|
|
|
attr = hipMemPoolAttrUsedMemHigh;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &value64));
|
|
// Make sure the high watermark usage works - the both buffers must be reported
|
|
REQUIRE((allocation_size1 + allocation_size2) == value64);
|
|
|
|
HIP_CHECK(hipFreeAsync(reinterpret_cast<void*>(alloc_mem2), stream.stream()));
|
|
attr = hipMemPoolAttrUsedMemCurrent;
|
|
HIP_CHECK(hipMemPoolGetAttribute(mempool.mempool(), attr, &value64));
|
|
// Make sure the current usage reports just one buffer, because the above free doesn't hold memory
|
|
REQUIRE(allocation_size2 == value64);
|
|
|
|
HIP_CHECK(hipFreeAsync(reinterpret_cast<void*>(alloc_mem3), stream.stream()));
|
|
}
|