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

added new gfx target (#171)

Browse Source
This commit is contained in:
akolliasAMD
2025-06-25 17:35:54 -06:00
committed by GitHub
parent 47bd7ec0d8
commit e2bae5131a
2 changed files with 12 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CMakeLists.txt
+3 -1
View File
@@ -109,7 +109,9 @@ if (NOT BUILD_TESTS_ONLY)
gfx90a:xnack-;
gfx90a:xnack+;
gfx942:xnack-;
gfx942:xnack+)
gfx942:xnack+;
gfx950:xnack-;
gfx950:xnack+)
if (BUILD_LOCAL_GPU_TARGET_ONLY)
message(STATUS "Building only for local GPU target")
src/assembly.hpp
+9 -9
View File
@@ -53,7 +53,7 @@ __device__ __forceinline__ int uncached_load_ubyte(uint8_t* src) {
: "=v"(ret)
: "v"(src));
#endif
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
asm volatile(
"global_load_ubyte %0 %1 off sc0 sc1 \n"
"s_waitcnt vmcnt(0)"
@@ -76,7 +76,7 @@ __device__ __forceinline__ void refresh_volatile_sbyte(volatile int *assigned_va
: "=v"(*assigned_value)
: "v"(read_value));
#endif
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
asm volatile(
"global_load_sbyte %0 %1 off sc0 sc1\n "
"s_waitcnt vmcnt(0)"
@@ -98,7 +98,7 @@ __device__ __forceinline__ void refresh_volatile_dwordx2(volatile uint64_t *assi
: "=v"(*assigned_value)
: "v"(read_value));
#endif
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
asm volatile(
"global_load_dwordx2 %0 %1 off sc0 sc1\n "
"s_waitcnt vmcnt(0)"
@@ -129,7 +129,7 @@ NOWARN(-Wdeprecated-volatile,
: "=v"(ret)
: "v"(src));
#endif
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
asm volatile(
"global_load_dword %0 %1 off sc0 sc1 \n"
"s_waitcnt vmcnt(0)"
@@ -149,7 +149,7 @@ NOWARN(-Wdeprecated-volatile,
: "=v"(ret)
: "v"(src));
#endif
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
asm volatile(
"global_load_dwordx2 %0 %1 off sc0 sc1 \n"
"s_waitcnt vmcnt(0)"
@@ -175,7 +175,7 @@ __device__ __forceinline__ void __roc_flush() {
// asm volatile("s_dcache_wb;");
// asm volatile("buffer_wbl2;");
#endif
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
// asm volatile("s_dcache_wb;");
// asm volatile("buffer_wbl2;");
#endif
@@ -194,7 +194,7 @@ __device__ __forceinline__ void store_asm(uint8_t* val, uint8_t* dst,
#if defined(__gfx90a__)
asm volatile("flat_store_short %0 %1 glc slc" : : "v"(dst), "v"(val16));
#endif
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
asm volatile("flat_store_short %0 %1 sc0 sc1" : : "v"(dst), "v"(val16));
#endif
break;
@@ -208,7 +208,7 @@ __device__ __forceinline__ void store_asm(uint8_t* val, uint8_t* dst,
#if defined(__gfx90a__)
asm volatile("flat_store_dword %0 %1 glc slc" : : "v"(dst), "v"(val32));
#endif
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
asm volatile("flat_store_dword %0 %1 sc0 sc1" : : "v"(dst), "v"(val32));
#endif
break;
@@ -222,7 +222,7 @@ __device__ __forceinline__ void store_asm(uint8_t* val, uint8_t* dst,
#if defined(__gfx90a__)
asm volatile("flat_store_dwordx2 %0 %1 glc slc" : : "v"(dst), "v"(val64));
#endif
#if defined(__gfx942__)
#if defined(__gfx942__) || defined(__gfx950__)
asm volatile("flat_store_dwordx2 %0 %1 sc0 sc1" : : "v"(dst), "v"(val64));
#endif
break;