Skip to content

Conformed fp8 code to changes in rccl fp8-nccl-sync branch. #113

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
corey-derochie-amd wants to merge 1 commit into
base: develop
Choose a base branch
from
+19 −19
Draft

Conformed fp8 code to changes in rccl fp8-nccl-sync branch. #113

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions src/all_reduce.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,7 +65,7 @@ testResult_t AllReduceRunTest(struct threadArgs* args, int root, ncclDataType_t
ncclRedOp_t *run_ops;
const char **run_typenames, **run_opnames;
int type_count, op_count;
if((type == ncclFp8E4M3 || type == ncclFp8E5M2) && op == ncclProd)
if((type == ncclFloat8e4m3 || type == ncclFloat8e5m2) && op == ncclProd)
return testSuccess;

if ((int)type != -1) {
Expand All @@ -90,7 +90,7 @@ testResult_t AllReduceRunTest(struct threadArgs* args, int root, ncclDataType_t

for (int i=0; i<type_count; i++) {
for (int j=0; j<op_count; j++) {
if((i == ncclFp8E4M3 || i == ncclFp8E5M2) && j == ncclProd)
if((i == ncclFloat8e4m3 || i == ncclFloat8e5m2) && j == ncclProd)
continue;
TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], run_ops[j], run_opnames[j], -1));
}
Expand Down
6 changes: 3 additions & 3 deletions src/common.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,7 +36,7 @@ size_t cache_bytes = 192 * 1024 * 1024; // Use 192MB
, ncclBfloat16
#endif
#if RCCL_FLOAT8 == 1
, ncclFp8E4M3, ncclFp8E5M2
, ncclFloat8e4m3, ncclFloat8e5m2
#endif
};
const char *test_typenames[ncclNumTypes] = {
Expand Down Expand Up @@ -557,8 +557,8 @@ testResult_t startColl(struct threadArgs* args, ncclDataType_t type, ncclRedOp_t
case ncclBfloat16: bf16 = ncclVerifiablePremulScalar<hip_bfloat16>(rank); break;
#endif
#if defined(RCCL_FLOAT8)
case ncclFp8E4M3: fp8_e4m3 = ncclVerifiablePremulScalar<rccl_float8>(rank); break;
case ncclFp8E5M2: fp8_e5m2 = ncclVerifiablePremulScalar<rccl_bfloat8>(rank); break;
case ncclFloat8e4m3: fp8_e4m3 = ncclVerifiablePremulScalar<rccl_float8>(rank); break;
case ncclFloat8e5m2: fp8_e5m2 = ncclVerifiablePremulScalar<rccl_bfloat8>(rank); break;
#endif
case ncclNumTypes: break;
}
Expand Down
4 changes: 2 additions & 2 deletions src/common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,8 +250,8 @@ static size_t wordSize(ncclDataType_t type) {
//case ncclInt8:
case ncclUint8:
#if NCCL_MAJOR >= 2 && RCCL_FLOAT8 == 1
case ncclFp8E4M3:
case ncclFp8E5M2:
case ncclFloat8e4m3:
case ncclFloat8e5m2:
#endif
#endif
return 1;
Expand Down
24 changes: 12 additions & 12 deletions verifiable/verifiable.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -890,8 +890,8 @@ void prepareInput1(
case ncclBfloat16: CASE_TY(hip_bfloat16)
#endif
#if HAVE_ncclfp8
case ncclFp8E4M3: CASE_TY(rccl_float8)
case ncclFp8E5M2: CASE_TY(rccl_bfloat8)
case ncclFloat8e4m3: CASE_TY(rccl_float8)
case ncclFloat8e5m2: CASE_TY(rccl_bfloat8)
#endif
case ncclFloat32: CASE_TY(float)
case ncclFloat64: CASE_TY(double)
Expand Down Expand Up @@ -970,8 +970,8 @@ void prepareExpected1(
case ncclBfloat16: CASE_TY(hip_bfloat16)
#endif
#if HAVE_ncclfp8
case ncclFp8E4M3: CASE_TY(rccl_float8)
case ncclFp8E5M2: CASE_TY(rccl_bfloat8)
case ncclFloat8e4m3: CASE_TY(rccl_float8)
case ncclFloat8e5m2: CASE_TY(rccl_bfloat8)
#endif
case ncclFloat32: CASE_TY(float)
case ncclFloat64: CASE_TY(double)
Expand Down Expand Up @@ -1044,8 +1044,8 @@ __host__ __device__ unsigned calcSumFloatTolerance(int rank_n, int elt_ty) {
break;
#endif
#if HAVE_ncclfp8
case ncclFp8E4M3:
case ncclFp8E5M2:
case ncclFloat8e4m3:
case ncclFloat8e5m2:
power = .91f;
coef = .66f;
break;
Expand Down Expand Up @@ -1175,8 +1175,8 @@ void ncclVerifiableVerify(
floating |= elt_ty == ncclBfloat16;
#endif
#if HAVE_ncclfp8
floating |= elt_ty == ncclFp8E4M3;
floating |= elt_ty == ncclFp8E5M2;
floating |= elt_ty == ncclFloat8e4m3;
floating |= elt_ty == ncclFloat8e5m2;
#endif

unsigned tolerance = 0;
Expand Down Expand Up @@ -1207,8 +1207,8 @@ void ncclVerifiableVerify(
case ncclBfloat16: CASE_TY(hip_bfloat16, uint16_t)
#endif
#if HAVE_ncclfp8
case ncclFp8E4M3: CASE_TY(rccl_float8, uint8_t)
case ncclFp8E5M2: CASE_TY(rccl_bfloat8, uint8_t)
case ncclFloat8e4m3: CASE_TY(rccl_float8, uint8_t)
case ncclFloat8e5m2: CASE_TY(rccl_bfloat8, uint8_t)
#endif
case ncclFloat32: CASE_TY(float, uint32_t)
case ncclFloat64: CASE_TY(double, uint64_t)
Expand Down Expand Up @@ -1278,8 +1278,8 @@ __global__ void sweep() {
sweep1<hip_bfloat16>(ncclBfloat16, "bfloat16");
#endif
#if HAVE_ncclfp8
sweep1<rccl_float8>(ncclFp8E4M3, "fp8_e4m3");
sweep1<rccl_bfloat8>(ncclFp8E5M2, "fp8_e5m2");
sweep1<rccl_float8>(ncclFloat8e4m3, "fp8_e4m3");
sweep1<rccl_bfloat8>(ncclFloat8e5m2, "fp8_e5m2");
#endif
sweep1<float>(ncclFloat32, "float");
sweep1<double>(ncclFloat64, "double");
Expand Down