A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 15:45 UTC
[Bugfix][Kernel] Implement acquire/release polyfill for Pascal (#8776)
sasha0552 committed
Sep 25, 2024 at 04:26 UTC
b4522474a32b6e0bf5573a9b6a6830cb787dfb63
2 files changed
+18
csrc/custom_all_reduce.cuh
+11
@@ -131,15 +131,26 @@ DINLINE O downcast(array_t<float, O::size> val) {
131
}
132
133
static DINLINE void st_flag_release(FlagType* flag_addr, FlagType flag) {
134
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
135
asm volatile("st.release.sys.global.u32 [%1], %0;" ::"r"(flag),
136
"l"(flag_addr));
137
+#else
138
+ asm volatile("membar.sys; st.volatile.global.u32 [%1], %0;" ::"r"(flag),
139
+ "l"(flag_addr));
140
+#endif
141
}
142
143
static DINLINE FlagType ld_flag_acquire(FlagType* flag_addr) {
144
FlagType flag;
145
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
146
asm volatile("ld.acquire.sys.global.u32 %0, [%1];"
147
: "=r"(flag)
148
: "l"(flag_addr));
149
+#else
150
+ asm volatile("ld.volatile.global.u32 %0, [%1]; membar.gl;"
151
+ : "=r"(flag)
152
+ : "l"(flag_addr));
153
+#endif
154
return flag;
155
}
156
csrc/custom_all_reduce_test.cu
+7
@@ -44,7 +44,14 @@
44
} while (0)
45
46
__global__ void dummy_kernel() {
47
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 700
48
for (int i = 0; i < 100; i++) __nanosleep(1000000); // 100ms
49
+#else
50
+ for (int i = 0; i < 100; i++) {
51
+ long long int start = clock64();
52
+ while (clock64() - start < 150000000); // approximately 98.4ms on P40
53
+ }
54
+#endif
55
}
56
57
template <typename T>