Franscobec/gpufetch
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

I guess we can rely on HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED

Browse Source
This commit is contained in:
Dr-Noob
2025-10-23 21:11:25 +02:00
parent 3b567b9381
commit 9b519828f4
2 changed files with 27 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/common/printer.cpp
View File

@@ -502,7 +502,7 @@ bool print_gpufetch_amd(struct gpu_info* gpu, STYLE s, struct color** cs, struct
setAttribute(art, ATTRIBUTE_FREQUENCY, max_frequency);
setAttribute(art, ATTRIBUTE_COMPUTE_UNITS, cus);
setAttribute(art, ATTRIBUTE_LDS_SIZE, lds_size);
// setAttribute(art, ATTRIBUTE_MEMORY, mem_size);
setAttribute(art, ATTRIBUTE_MEMORY, mem_size);
setAttribute(art, ATTRIBUTE_BUS_WIDTH, bus_width);
bool use_short = false;

35
src/hsa/hsa.cpp
View File

@@ -25,7 +25,7 @@ struct agent_info {
uint32_t compute_unit;
uint32_t bus_width;
uint32_t lds_size;
uint32_t global_size;
uint64_t global_size;
};
#define RET_IF_HSA_ERR(err) { \
@@ -43,32 +43,43 @@ struct agent_info {
} \
}
hsa_status_t region_callback(hsa_region_t region, void* data) {
hsa_region_segment_t segment;
hsa_status_t err = hsa_region_get_info(region, HSA_REGION_INFO_SEGMENT, &segment);
hsa_status_t memory_pool_callback(hsa_amd_memory_pool_t pool, void* data) {
hsa_amd_segment_t segment;
hsa_status_t err = hsa_amd_memory_pool_get_info(pool, HSA_AMD_MEMORY_POOL_INFO_SEGMENT, &segment);
RET_IF_HSA_ERR(err);
uint32_t *mems = reinterpret_cast<uint32_t *>(data);
uint64_t *mems = reinterpret_cast<uint64_t *>(data);
if (segment == HSA_REGION_SEGMENT_GROUP) {
// LDS memory
size_t size = 0;
err = hsa_region_get_info(region, HSA_REGION_INFO_SIZE, &size);
err = hsa_amd_memory_pool_get_info(pool, HSA_AMD_MEMORY_POOL_INFO_SIZE, &size);
RET_IF_HSA_ERR(err);
mems[0] = size;
}
else if (segment == HSA_REGION_SEGMENT_GLOBAL) {
// Global memory
// LDS memory
size_t size = 0;
uint32_t global_flags = 0;
err = hsa_region_get_info(region, HSA_REGION_INFO_SIZE, &size);
err = hsa_amd_memory_pool_get_info(pool, HSA_AMD_MEMORY_POOL_INFO_GLOBAL_FLAGS, &global_flags);
RET_IF_HSA_ERR(err);
if (global_flags & HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED) {
if (mems[1] != 0) {
printErr("Found HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_EXTENDED_SCOPE_FINE_GRAINED twice!");
return HSA_STATUS_ERROR;
}
uint64_t size = 0;
err = hsa_amd_memory_pool_get_info(pool, HSA_AMD_MEMORY_POOL_INFO_SIZE, &size);
RET_IF_HSA_ERR(err);
mems[1] = size;
}
}
return HSA_STATUS_SUCCESS;
}
@@ -99,8 +110,8 @@ hsa_status_t agent_callback(hsa_agent_t agent, void *data) {
err = hsa_agent_get_info(agent, (hsa_agent_info_t) HSA_AMD_AGENT_INFO_MEMORY_WIDTH, &info->bus_width);
RET_IF_HSA_ERR(err);
uint32_t mems[2];
err = hsa_agent_iterate_regions(agent, region_callback, &mems);
uint64_t mems[2] = {0, 0};
err = hsa_amd_agent_iterate_memory_pools(agent, memory_pool_callback, &mems);
RET_IF_HSA_ERR(err);
info->lds_size = mems[0];
@@ -125,6 +136,8 @@ struct memory* get_memory_info(struct gpu_info* gpu, struct agent_info info) {
mem->lds_size = info.lds_size;
mem->size_bytes = info.global_size;
printf("xxx: %ld %ld\n", mem->size_bytes, info.global_size);
return mem;
}