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

Enable fp8 support (#1101)

Browse Source 
* initial checkin

* resolve cr comments

* resolve the build issue

* fix the data correctless issue

* update fp8 header file and update the unit test for fp8 support

* remove fp16 from fp8 headers

* fix ut issue and catch up the latest code from develop

* udate according to cr comments

* update ut according to cr comments

* update num floats for each SumPostDiv from 4 to 6

* update fp8 header file name

* fix the typo

[ROCm/rccl commit: 6777e65c1d]
This commit is contained in:
Andy li
2024-03-09 07:17:53 +08:00
committed by GitHub
parent 2354601589
commit e373bd44bf
29 changed files with 1243 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/test/AllGatherTests.cpp
+1 -1
View File
@@ -32,7 +32,7 @@ namespace RcclUnitTesting
// Configuration
std::vector<ncclFunc_t> const funcTypes = {ncclCollAllGather};
std::vector<ncclDataType_t> const dataTypes = {ncclBfloat16, ncclFloat64};
std::vector<ncclDataType_t> const dataTypes = {ncclBfloat16, ncclFloat64, ncclFp8E4M3, ncclFp8E5M2};
std::vector<ncclRedOp_t> const redOps = {ncclSum};
std::vector<int> const roots = {0};
std::vector<int> const numElements = {586};
projects/rccl/test/AllReduceTests.cpp
+2 -2
View File
@@ -32,7 +32,7 @@ namespace RcclUnitTesting
// Configuration
std::vector<ncclFunc_t> const funcTypes = {ncclCollAllReduce};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat16, ncclFloat64};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat16, ncclFloat64, ncclFp8E4M3, ncclFp8E5M2};
std::vector<ncclRedOp_t> const redOps = {ncclMin};
std::vector<int> const roots = {0};
std::vector<int> const numElements = {12888};
@@ -70,7 +70,7 @@ namespace RcclUnitTesting
// Configuration
std::vector<ncclFunc_t> const funcTypes = {ncclCollAllReduce};
std::vector<ncclDataType_t> const dataTypes = {ncclInt32};
std::vector<ncclDataType_t> const dataTypes = {ncclInt32, ncclFp8E4M3, ncclFp8E5M2};
std::vector<ncclRedOp_t> const redOps = {ncclMax};
std::vector<int> const roots = {0};
std::vector<int> const numElements = {393216, 12888, 384};
projects/rccl/test/AllToAllTests.cpp
+1 -1
View File
@@ -35,7 +35,7 @@ namespace RcclUnitTesting
// Configuration
std::vector<ncclFunc_t> const funcTypes = {ncclCollAllToAll};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat64, ncclBfloat16};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat64, ncclBfloat16, ncclFp8E4M3, ncclFp8E5M2};
std::vector<ncclRedOp_t> const redOps = {ncclSum};
std::vector<int> const roots = {0};
std::vector<int> const numElements = {5685};
projects/rccl/test/BroadcastTests.cpp
+1 -1
View File
@@ -32,7 +32,7 @@ namespace RcclUnitTesting
// Configuration
std::vector<ncclFunc_t> const funcTypes = {ncclCollBroadcast};
std::vector<ncclDataType_t> const dataTypes = {ncclBfloat16, ncclFloat64};
std::vector<ncclDataType_t> const dataTypes = {ncclBfloat16, ncclFloat64, ncclFp8E4M3, ncclFp8E5M2};
std::vector<ncclRedOp_t> const redOps = {ncclSum};
std::vector<int> const roots = {0};
std::vector<int> const numElements = {586};
projects/rccl/test/ReduceScatterTests.cpp
+1 -1
View File
@@ -32,7 +32,7 @@ namespace RcclUnitTesting
// Configuration
std::vector<ncclFunc_t> const funcTypes = {ncclCollReduceScatter};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat64, ncclBfloat16};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat64, ncclBfloat16, ncclFp8E4M3, ncclFp8E5M2};
std::vector<ncclRedOp_t> const redOps = {ncclMax};
std::vector<int> const roots = {0};
std::vector<int> const numElements = {1048576};
projects/rccl/test/ReduceTests.cpp
+2 -2
View File
@@ -32,7 +32,7 @@ namespace RcclUnitTesting
// Configuration
std::vector<ncclFunc_t> const funcTypes = {ncclCollReduce};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat16, ncclFloat64};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat16, ncclFloat64, ncclFp8E4M3, ncclFp8E5M2};
std::vector<ncclRedOp_t> const redOps = {ncclMin};
std::vector<int> const roots = {0};
std::vector<int> const numElements = {393216};
@@ -70,7 +70,7 @@ namespace RcclUnitTesting
// Configuration
std::vector<ncclFunc_t> const funcTypes = {ncclCollReduce};
std::vector<ncclDataType_t> const dataTypes = {ncclBfloat16};
std::vector<ncclDataType_t> const dataTypes = {ncclBfloat16, ncclFp8E4M3, ncclFp8E5M2};
std::vector<ncclRedOp_t> const redOps = {ncclMax};
std::vector<int> const roots = {0};
std::vector<int> const numElements = {393216};
projects/rccl/test/ScatterTests.cpp
+1 -1
View File
@@ -32,7 +32,7 @@ namespace RcclUnitTesting
// Configuration
std::vector<ncclFunc_t> const funcTypes = {ncclCollScatter};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat64, ncclBfloat16};
std::vector<ncclDataType_t> const dataTypes = {ncclFloat64, ncclBfloat16, ncclFp8E4M3, ncclFp8E5M2};
std::vector<ncclRedOp_t> const redOps = {ncclSum};
std::vector<int> const roots = {1};
std::vector<int> const numElements = {24658};
projects/rccl/test/common/CollectiveArgs.cpp
+2
View File
@@ -199,8 +199,10 @@ namespace RcclUnitTesting
case ncclUint32: ss << scalarsPerRank.U4[this->globalRank]; break;
case ncclInt64: ss << scalarsPerRank.I8[this->globalRank]; break;
case ncclUint64: ss << scalarsPerRank.U8[this->globalRank]; break;
case ncclFp8E4M3: ss << scalarsPerRank.F1[this->globalRank]; break;
case ncclFloat32: ss << scalarsPerRank.F4[this->globalRank]; break;
case ncclFloat64: ss << scalarsPerRank.F8[this->globalRank]; break;
case ncclFp8E5M2: ss << scalarsPerRank.B1[this->globalRank]; break;
case ncclBfloat16: ss << scalarsPerRank.B2[this->globalRank]; break;
default: ss << "(UNKNOWN)";
}
projects/rccl/test/common/CollectiveArgs.hpp
+3 -1
View File
@@ -53,7 +53,9 @@ namespace RcclUnitTesting
"ncclFloat16",
"ncclFloat32",
"ncclFloat64",
"ncclBfloat16"
"ncclBfloat16",
"ncclFp8E4M3",
"ncclFp8E5M2"
};
char const ncclRedOpNames[ncclNumOps][32] =
projects/rccl/test/common/EnvVars.cpp
+2
View File
@@ -114,6 +114,8 @@ namespace RcclUnitTesting
dataTypes.push_back(ncclFloat32);
dataTypes.push_back(ncclFloat64);
dataTypes.push_back(ncclBfloat16);
dataTypes.push_back(ncclFp8E4M3);
dataTypes.push_back(ncclFp8E5M2);
}
// Build list of possible # GPU ranks based on env vars
projects/rccl/test/common/PtrUnion.cpp
+31 -11
View File
@@ -14,6 +14,8 @@ namespace RcclUnitTesting
{
case ncclInt8: return 1;
case ncclUint8: return 1;
case ncclFp8E4M3:return 1;
case ncclFp8E5M2:return 1;
case ncclInt32: return 4;
case ncclUint32: return 4;
case ncclInt64: return 8;
@@ -160,9 +162,11 @@ namespace RcclUnitTesting
case ncclUint32: U4[idx] = valueI; break;
case ncclInt64: I8[idx] = valueI; break;
case ncclUint64: U8[idx] = valueI; break;
case ncclFp8E4M3: F1[idx] = rccl_float8(valueF); break;
case ncclFloat16: F2[idx] = __float2half(static_cast<float>(valueF)); break;
case ncclFloat32: F4[idx] = valueF; break;
case ncclFloat64: F8[idx] = valueF; break;
case ncclFp8E5M2: B1[idx] = rccl_bfloat8(valueF); break;
case ncclBfloat16: B2[idx] = rccl_bfloat16(static_cast<float>(valueF)); break;
default:
ERROR("Unsupported datatype\n");
@@ -181,9 +185,11 @@ namespace RcclUnitTesting
case ncclUint32: valueI = U4[idx]; break;
case ncclInt64: valueI = I8[idx]; break;
case ncclUint64: valueI = U8[idx]; break;
case ncclFp8E4M3: valueF = float(F1[idx]); break;
case ncclFloat16: valueF = __half2float(F2[idx]); break;
case ncclFloat32: valueF = F4[idx]; break;
case ncclFloat64: valueF = F8[idx]; break;
case ncclFp8E5M2: valueF = float(B1[idx]); break;
case ncclBfloat16: valueF = B2[idx]; break;
default:
ERROR("Unsupported datatype\n");
@@ -211,9 +217,11 @@ namespace RcclUnitTesting
case ncclUint32: U4[idx] *= scalarsPerRank.U4[rank]; break;
case ncclInt64: I8[idx] *= scalarsPerRank.I8[rank]; break;
case ncclUint64: U8[idx] *= scalarsPerRank.U8[rank]; break;
case ncclFp8E4M3: F1[idx] = rccl_float8(F1[idx] * scalarsPerRank.F1[rank]); break;
case ncclFloat16: F2[idx] = __float2half(__half2float(F2[idx]) * __half2float(scalarsPerRank.F2[rank])); break;
case ncclFloat32: F4[idx] *= scalarsPerRank.F4[rank]; break;
case ncclFloat64: F8[idx] *= scalarsPerRank.F8[rank]; break;
case ncclFp8E5M2: B1[idx] = rccl_bfloat8(B1[idx] * scalarsPerRank.B1[rank]); break;
case ncclBfloat16: B2[idx] *= scalarsPerRank.B2[rank]; break;
default:
ERROR("Unsupported datatype\n");
@@ -244,9 +252,11 @@ namespace RcclUnitTesting
case ncclUint32: U4[idx] = ReduceOp(op, U4[idx], inputCpu.U4[idx]); break;
case ncclInt64: I8[idx] = ReduceOp(op, I8[idx], inputCpu.I8[idx]); break;
case ncclUint64: U8[idx] = ReduceOp(op, U8[idx], inputCpu.U8[idx]); break;
case ncclFp8E4M3: F1[idx] = rccl_float8(ReduceOp(op, float(F1[idx]), float(inputCpu.F1[idx]))); break;
case ncclFloat16: F2[idx] = __float2half(ReduceOp(op, __half2float(F2[idx]), __half2float(inputCpu.F2[idx]))); break;
case ncclFloat32: F4[idx] = ReduceOp(op, F4[idx], inputCpu.F4[idx]); break;
case ncclFloat64: F8[idx] = ReduceOp(op, F8[idx], inputCpu.F8[idx]); break;
case ncclFp8E5M2: B1[idx] = rccl_bfloat8(ReduceOp(op, float(B1[idx]), float(inputCpu.B1[idx]))); break;
case ncclBfloat16: B2[idx] = ReduceOp(op, B2[idx], inputCpu.B2[idx]); break;
default:
ERROR("Unsupported datatype\n");
@@ -271,10 +281,12 @@ namespace RcclUnitTesting
case ncclUint32: U4[idx] /= divisor; break;
case ncclInt64: I8[idx] /= divisor; break;
case ncclUint64: U8[idx] /= divisor; break;
case ncclFloat16: F2[idx] = __float2half(__half2float(F2[idx])/divisor); break;
case ncclFp8E4M3: F1[idx] = (rccl_float8((float)(F1[idx]) / divisor)); break;
case ncclFloat16: F2[idx] = __float2half(__half2float(F2[idx])/divisor); break;
case ncclFloat32: F4[idx] /= divisor; break;
case ncclFloat64: F8[idx] /= divisor; break;
case ncclBfloat16: B2[idx] = (rccl_bfloat16((float)(B2[idx]) / divisor)); break;
case ncclFp8E5M2: B1[idx] = (rccl_bfloat8((float)(B1[idx]) / divisor)); break;
case ncclBfloat16: B2[idx] = (rccl_bfloat16((float)(B2[idx]) / divisor)); break;
default:
ERROR("Unsupported datatype\n");
return TEST_FAIL;
@@ -295,15 +307,17 @@ namespace RcclUnitTesting
{
switch (dataType)
{
case ncclInt8: isMatch = (I1[idx] == expected.I1[idx]); break;
case ncclUint8: isMatch = (U1[idx] == expected.U1[idx]); break;
case ncclInt32: isMatch = (I4[idx] == expected.I4[idx]); break;
case ncclUint32: isMatch = (U4[idx] == expected.U4[idx]); break;
case ncclInt64: isMatch = (I8[idx] == expected.I8[idx]); break;
case ncclUint64: isMatch = (U8[idx] == expected.U8[idx]); break;
case ncclFloat16: isMatch = (fabs(__half2float(F2[idx]) - __half2float(expected.F2[idx])) < 9e-2); break;
case ncclFloat32: isMatch = (fabs(F4[idx] - expected.F4[idx]) < 1e-5); break;
case ncclFloat64: isMatch = (fabs(F8[idx] - expected.F8[idx]) < 1e-12); break;
case ncclInt8: isMatch = (I1[idx] == expected.I1[idx]); break;
case ncclUint8: isMatch = (U1[idx] == expected.U1[idx]); break;
case ncclInt32: isMatch = (I4[idx] == expected.I4[idx]); break;
case ncclUint32: isMatch = (U4[idx] == expected.U4[idx]); break;
case ncclInt64: isMatch = (I8[idx] == expected.I8[idx]); break;
case ncclUint64: isMatch = (U8[idx] == expected.U8[idx]); break;
case ncclFp8E4M3: isMatch = (fabs(float(F1[idx]) - float(expected.F1[idx])) < 9e-2); break;
case ncclFloat16: isMatch = (fabs(__half2float(F2[idx]) - __half2float(expected.F2[idx])) < 9e-2); break;
case ncclFloat32: isMatch = (fabs(F4[idx] - expected.F4[idx]) < 1e-5); break;
case ncclFloat64: isMatch = (fabs(F8[idx] - expected.F8[idx]) < 1e-12); break;
case ncclFp8E5M2: isMatch = (fabs(float(B1[idx]) - float(expected.B1[idx])) < 9e-2); break;
case ncclBfloat16: isMatch = (fabs((float)B2[idx] - (float)expected.B2[idx]) < 9e-2); break;
default:
ERROR("Unsupported datatype\n");
@@ -328,12 +342,16 @@ namespace RcclUnitTesting
ERROR("Expected output: %ld. Actual output: %ld at index %lu\n", expected.I8[idx], I8[idx], idx); break;
case ncclUint64:
ERROR("Expected output: %lu. Actual output: %lu at index %lu\n", expected.U8[idx], U8[idx], idx); break;
case ncclFp8E4M3:
ERROR("Expected output: %f. Actual output: %f at index %lu\n", (float)expected.F1[idx], (float)F1[idx], idx); break;
case ncclFloat16:
ERROR("Expected output: %f. Actual output: %f at index %lu\n", __half2float(expected.F2[idx]), __half2float(F2[idx]), idx); break;
case ncclFloat32:
ERROR("Expected output: %f. Actual output: %f at index %lu\n", expected.F4[idx], F4[idx], idx); break;
case ncclFloat64:
ERROR("Expected output: %lf. Actual output: %lf at index %lu\n", expected.F8[idx], F8[idx], idx); break;
case ncclFp8E5M2:
ERROR("Expected output: %f. Actual output: %f at index %lu\n", (float)expected.B1[idx], (float)B1[idx], idx); break;
case ncclBfloat16:
ERROR("Expected output: %f. Actual output: %f at index %lu\n", (float)expected.B2[idx], (float)B2[idx], idx); break;
default:
@@ -358,9 +376,11 @@ namespace RcclUnitTesting
case ncclUint32: ss << U4[i]; break;
case ncclInt64: ss << I8[i]; break;
case ncclUint64: ss << U8[i]; break;
case ncclFp8E4M3: ss << (float)F1[i]; break;
case ncclFloat16: ss << __half2float(F2[i]); break;
case ncclFloat32: ss << F4[i]; break;
case ncclFloat64: ss << F8[i]; break;
case ncclFp8E5M2: ss << (float)B1[i]; break;
case ncclBfloat16: ss << (float)B2[i]; break;
default: break;
}
projects/rccl/test/common/PtrUnion.hpp
+3
View File
@@ -7,6 +7,7 @@
#pragma once
#include "ErrCode.hpp"
#include "rccl/rccl.h"
#include "rccl_float8.h"
#include "rccl_bfloat16.h"
#include "hip/hip_fp16.h"
@@ -43,8 +44,10 @@ namespace RcclUnitTesting
int64_t* I8; // ncclInt64
uint64_t* U8; // ncclUint64
__half* F2; // ncclFloat16
rccl_float8* F1; // ncclFp8E4M3
float* F4; // ncclFloat32
double* F8; // ncclFloat64
rccl_bfloat8* B1; // ncclFp8E5M2
rccl_bfloat16* B2; // ncclBfloat16
constexpr PtrUnion() : ptr(nullptr) {}