hrx-rs 0.10.1

Owned native AMD GPU and NPU execution with Loom compilation
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52

diff --git a/libamdf/src/gpu/umd/kfd/memory.c b/libamdf/src/gpu/umd/kfd/memory.c
index 708f33440..aa96bb767 100644
--- a/libamdf/src/gpu/umd/kfd/memory.c
+++ b/libamdf/src/gpu/umd/kfd/memory.c
@@ -69,7 +69,9 @@ static void amdf_gpu_kfd_memory_plan(
       plan->native_flags |= KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC;
     }
   } else {
-    plan->native_flags |= KFD_IOC_ALLOC_MEM_FLAGS_COHERENT;
+    if ((plan->flags & AMDF_MEMORY_FLAG_HOST_COHERENT) != 0) {
+      plan->native_flags |= KFD_IOC_ALLOC_MEM_FLAGS_COHERENT;
+    }
     if ((profile->roles & AMDF_MEMORY_PROFILE_ROLE_REGISTER) != 0) {
       const uintptr_t pointer = (uintptr_t)create_info->registered_host_pointer;
       plan->host_byte_offset = pointer & (device->page_size - 1);
@@ -160,7 +162,9 @@ static amdf_memory_host_description_t amdf_gpu_umd_memory_describe_host(
     amdf_memory_flags_t flags) {
   (void)external_memory_type;
   const amdf_gpu_umd_device_t* device = data;
-  const bool write_back = (flags & AMDF_MEMORY_FLAG_HOST_COHERENT) != 0;
+  // KFD GTT mappings are CPU write-back even when GPU caches are enabled.
+  // Only local VRAM mappings use write-combined CPU access.
+  const bool write_back = (flags & AMDF_MEMORY_FLAG_DEVICE_LOCAL) == 0;
   const uint32_t line_size = write_back ? device->cache_line_size : 0;
   const amdf_cache_transition_t flush = {
       .kind = write_back ? AMDF_CACHE_TRANSITION_KIND_RANGE
@@ -348,7 +352,7 @@ amdf_status_t amdf_gpu_umd_host_mapping_cache_control(
     uint64_t memory_byte_offset, uint64_t byte_length) {
   const amdf_gpu_umd_memory_t* memory = (amdf_gpu_umd_memory_t*)mapping;
   (void)operation;
-  if ((memory->flags & AMDF_MEMORY_FLAG_HOST_COHERENT) != 0) {
+  if ((memory->flags & AMDF_MEMORY_FLAG_DEVICE_LOCAL) == 0) {
     // GPU snooping does not make this backing coherent with other consumers.
     // Explicit CPU operations cover the requested bytes even on coherent GTT.
     amdf_linux_host_cache_transfer(
diff --git a/libamdf/src/gpu/umd/kfd/memory_profile.c b/libamdf/src/gpu/umd/kfd/memory_profile.c
index 2c1143e46..65861604f 100644
--- a/libamdf/src/gpu/umd/kfd/memory_profile.c
+++ b/libamdf/src/gpu/umd/kfd/memory_profile.c
@@ -139,9 +139,10 @@ amdf_status_t amdf_gpu_kfd_query_memory_profile(
                     AMDF_MEMORY_PROFILE_ROLE_HOST_MAP;
     profile.guaranteed_flags =
         AMDF_MEMORY_FLAG_HOST_VISIBLE | AMDF_MEMORY_FLAG_SHAREABLE |
-        AMDF_MEMORY_FLAG_HOST_COHERENT | AMDF_MEMORY_FLAG_DEVICE_ADDRESS;
+        AMDF_MEMORY_FLAG_DEVICE_ADDRESS;
     profile.supported_flags =
-        profile.guaranteed_flags | AMDF_MEMORY_FLAG_QUEUE_STORAGE;
+        profile.guaranteed_flags | AMDF_MEMORY_FLAG_QUEUE_STORAGE |
+        AMDF_MEMORY_FLAG_HOST_COHERENT;
     profile.guaranteed_device_access = AMDF_MEMORY_ACCESS_READ;
     profile.supported_device_access = AMDF_MEMORY_ACCESS_READ |
                                       AMDF_MEMORY_ACCESS_WRITE |