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 17, 00:30 UTC
Revert "[Kernel] Marlin_24: Ensure the mma.sp instruction is using the ::ordered_metadata modifier (introduced with PTX 8.5)" (#5149)
Simon Mo committed
May 31, 2024 at 00:00 UTC
e9d3aa04f6e55e2bb540f0810da97ddd0deebb13
1 file changed
+4
-8
csrc/quantization/marlin/sparse/common/mma.h
+4
-8
@@ -32,8 +32,7 @@ __device__ inline void mma_sp(const FragB& a_frag0, const FragB& a_frag1,
32
float* c = reinterpret_cast<float*>(&frag_c);
33
if (psel == 0) {
34
asm volatile(
35
- "mma.sp::ordered_metadata.sync.aligned.m16n8k32.row.col.f32.f16.f16."
36
- "f32 "
35
+ "mma.sp.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
36
"{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
37
"{%12,%13,%14,%15}, %16, 0x0;\n"
38
: "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3])
@@ -41,8 +40,7 @@ __device__ inline void mma_sp(const FragB& a_frag0, const FragB& a_frag1,
40
"r"(b[4]), "r"(b[6]), "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3]),
41
"r"(e[0]));
42
asm volatile(
44
- "mma.sp::ordered_metadata.sync.aligned.m16n8k32.row.col.f32.f16.f16."
45
- "f32 "
43
+ "mma.sp.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
44
"{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
45
"{%12,%13,%14,%15}, %16, 0x0;\n"
46
: "=f"(c[4]), "=f"(c[5]), "=f"(c[6]), "=f"(c[7])
@@ -51,8 +49,7 @@ __device__ inline void mma_sp(const FragB& a_frag0, const FragB& a_frag1,
49
"r"(e[0]));
50
} else {
51
asm volatile(
54
- "mma.sp::ordered_metadata.sync.aligned.m16n8k32.row.col.f32.f16.f16."
55
- "f32 "
52
+ "mma.sp.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
53
"{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
54
"{%12,%13,%14,%15}, %16, 0x1;\n"
55
: "=f"(c[0]), "=f"(c[1]), "=f"(c[2]), "=f"(c[3])
@@ -60,8 +57,7 @@ __device__ inline void mma_sp(const FragB& a_frag0, const FragB& a_frag1,
57
"r"(b[4]), "r"(b[6]), "f"(c[0]), "f"(c[1]), "f"(c[2]), "f"(c[3]),
58
"r"(e[0]));
59
asm volatile(
63
- "mma.sp::ordered_metadata.sync.aligned.m16n8k32.row.col.f32.f16.f16."
64
- "f32 "
60
+ "mma.sp.sync.aligned.m16n8k32.row.col.f32.f16.f16.f32 "
61
"{%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9, %10,%11}, "
62
"{%12,%13,%14,%15}, %16, 0x1;\n"
63
: "=f"(c[4]), "=f"(c[5]), "=f"(c[6]), "=f"(c[7])