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

Add hipify steps prior to build (#62)

Browse Source 
* Add hipify steps prior to build
This commit is contained in:
Bertan Dogancay authored and GitHub committed 2024-03-05 09:47:18 -07:00
1 parent b9b73002da
commit 88cf7dbf45
20 files changed
+295 -204
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

verifiable/inexact_regress.cu
+5 -5
View File
@@ -24,11 +24,11 @@
#include <cstdio>
#include <cstdint>
#include <hip/hip_bfloat16.h>
#include <hip/hip_fp16.h>
#include <cuda_fp16.h>
using std::uint64_t;
using std::uint32_t;
using bfloat16 = hip_bfloat16;
using bfloat16 = __nv_bfloat16;
template<typename T>
struct float_traits;
@@ -182,14 +182,14 @@ __global__ void kernel() {
int main() {
std::printf("type=float:\n");
kernel<float><<<1,32>>>();
hipDeviceSynchronize();
cudaDeviceSynchronize();
std::printf("\ntype=half:\n");
kernel<half><<<1,32>>>();
hipDeviceSynchronize();
cudaDeviceSynchronize();
std::printf("\ntype=bfloat16:\n");
kernel<bfloat16><<<1,32>>>();
hipDeviceSynchronize();
cudaDeviceSynchronize();
return 0;
}
verifiable/verifiable.cu
+21 -21
View File
@@ -8,8 +8,8 @@
//#pragma nv_diag_suppress declared_but_not_referenced
#include "verifiable.h"
#include <hip/hip_runtime.h>
#include <hip/hip_fp16.h>
#include <cuda_runtime.h>
#include <cuda_fp16.h>
#include <hip/hip_bfloat16.h>
#include "rccl/rccl.h"
@@ -91,7 +91,7 @@ template<>
struct IsIntegral<__half>: std::false_type {};
#if RCCL_BFLOAT16 == 1
template<>
struct IsIntegral<hip_bfloat16>: std::false_type {};
struct IsIntegral<__nv_bfloat16>: std::false_type {};
#endif
}
@@ -126,7 +126,7 @@ namespace {
}
#if RCCL_BFLOAT16 == 1
template<>
__host__ __device__ hip_bfloat16 castTo<hip_bfloat16>(float x) {
__host__ __device__ __nv_bfloat16 castTo<__nv_bfloat16>(float x) {
return hip_bfloat16(x);
}
#endif
@@ -153,7 +153,7 @@ struct ReduceSum {
return __float2half(__half2float(a) + __half2float(b));
}
#if RCCL_BFLOAT16 == 1
__host__ __device__ hip_bfloat16 operator()(hip_bfloat16 a, hip_bfloat16 b) const {
__host__ __device__ __nv_bfloat16 operator()(__nv_bfloat16 a, __nv_bfloat16 b) const {
return hip_bfloat16(static_cast<float>(a) + static_cast<float>(b));
}
#endif
@@ -169,7 +169,7 @@ struct ReduceProd {
return __float2half(__half2float(a) * __half2float(b));
}
#if RCCL_BFLOAT16 == 1
__host__ __device__ hip_bfloat16 operator()(hip_bfloat16 a, hip_bfloat16 b) const {
__host__ __device__ __nv_bfloat16 operator()(__nv_bfloat16 a, __nv_bfloat16 b) const {
return hip_bfloat16(static_cast<float>(a) * static_cast<float>(b));
}
#endif
@@ -185,7 +185,7 @@ struct ReduceMin {
return __half2float(a) < __half2float(b) ? a : b;
}
#if RCCL_BFLOAT16 == 1
__host__ __device__ hip_bfloat16 operator()(hip_bfloat16 a, hip_bfloat16 b) const {
__host__ __device__ __nv_bfloat16 operator()(__nv_bfloat16 a, __nv_bfloat16 b) const {
return static_cast<float>(a) < static_cast<float>(b) ? a : b;
}
#endif
@@ -201,7 +201,7 @@ struct ReduceMax {
return __half2float(a) > __half2float(b) ? a : b;
}
#if RCCL_BFLOAT16 == 1
__host__ __device__ hip_bfloat16 operator()(hip_bfloat16 a, hip_bfloat16 b) const {
__host__ __device__ __nv_bfloat16 operator()(__nv_bfloat16 a, __nv_bfloat16 b) const {
return static_cast<float>(a) > static_cast<float>(b) ? a : b;
}
#endif
@@ -280,7 +280,7 @@ struct FloatLayout<__half> {
};
#if RCCL_BFLOAT16 == 1
template<>
struct FloatLayout<hip_bfloat16> {
struct FloatLayout<__nv_bfloat16> {
static constexpr int exponent_bits = 8, mantissa_bits = 7;
static constexpr int exponent_bias = (1<<(exponent_bits-1))-1;
};
@@ -801,7 +801,7 @@ __global__ void prepareInput2(
template<typename ReduceOp>
void prepareInput1(
void *elts, intptr_t elt_n, int elt_ty, ReduceOp op, int rank_n, int rank_me,
uint64_t seed, intptr_t elt_ix0, hipStream_t stream
uint64_t seed, intptr_t elt_ix0, cudaStream_t stream
) {
int block_n = std::min<intptr_t>(32, (elt_n + 4*512-1)/(4*512));
#define CASE_TY(T) prepareInput2<<<block_n, 512, 0, stream>>>((T*)elts, elt_n, op, rank_n, rank_me, seed, elt_ix0); break;
@@ -814,7 +814,7 @@ void prepareInput1(
case ncclUint64: CASE_TY(uint64_t)
case ncclFloat16: CASE_TY(__half)
#if HAVE_ncclBfloat16
case ncclBfloat16: CASE_TY(hip_bfloat16)
case ncclBfloat16: CASE_TY(__nv_bfloat16)
#endif
case ncclFloat32: CASE_TY(float)
case ncclFloat64: CASE_TY(double)
@@ -826,7 +826,7 @@ void prepareInput1(
void ncclVerifiablePrepareInput(
void *elts, intptr_t elt_n, int elt_ty, int red_op, int rank_n, int rank_me,
uint64_t seed, intptr_t elt_ix0, hipStream_t stream
uint64_t seed, intptr_t elt_ix0, cudaStream_t stream
) {
#define CASE_OP(op) \
if(rank_n == 1) \
@@ -877,7 +877,7 @@ __global__ void prepareExpected2(
template<typename ReduceOp>
void prepareExpected1(
void *elts, intptr_t elt_n, int elt_ty, ReduceOp op, int rank_n,
uint64_t seed, intptr_t elt_ix0, hipStream_t stream
uint64_t seed, intptr_t elt_ix0, cudaStream_t stream
) {
int block_n = std::min<intptr_t>(32, (elt_n + 4*512-1)/(4*512));
#define CASE_TY(T) prepareExpected2<<<block_n, 512, 0, stream>>>((T*)elts, elt_n, op, rank_n, seed, elt_ix0); break;
@@ -890,7 +890,7 @@ void prepareExpected1(
case ncclUint64: CASE_TY(uint64_t)
case ncclFloat16: CASE_TY(__half)
#if HAVE_ncclBfloat16
case ncclBfloat16: CASE_TY(hip_bfloat16)
case ncclBfloat16: CASE_TY(__nv_bfloat16)
#endif
case ncclFloat32: CASE_TY(float)
case ncclFloat64: CASE_TY(double)
@@ -902,7 +902,7 @@ void prepareExpected1(
void ncclVerifiablePrepareExpected(
void *elts, intptr_t elt_n, int elt_ty, int red_op, int rank_n,
uint64_t seed, intptr_t elt_ix0, hipStream_t stream
uint64_t seed, intptr_t elt_ix0, cudaStream_t stream
) {
#define CASE_OP(op) \
if(rank_n == 1) \
@@ -1051,7 +1051,7 @@ __global__ void verifyInline2(
template<typename T, typename Uint>
void verifyInline1(
T const *results, intptr_t elt_n, int red_op, int rank_n, uint64_t seed, intptr_t elt_ix0,
unsigned tolerance, int64_t *bad_elt_n, hipStream_t stream, int block_n
unsigned tolerance, int64_t *bad_elt_n, cudaStream_t stream, int block_n
) {
#define CASE_OP(op) \
if(rank_n == 1) \
@@ -1080,7 +1080,7 @@ void verifyInline1(
void ncclVerifiableVerify(
void const *results, void const *expected, intptr_t elt_n, int elt_ty,
int red_op, int rank_n, uint64_t seed, intptr_t elt_ix0,
int64_t *bad_elt_n, hipStream_t stream
int64_t *bad_elt_n, cudaStream_t stream
) {
bool floating = elt_ty == ncclFloat16 || elt_ty == ncclFloat32 || elt_ty == ncclFloat64;
#if HAVE_ncclBfloat16
@@ -1112,7 +1112,7 @@ void ncclVerifiableVerify(
case ncclUint64: CASE_TY(uint64_t, uint64_t)
case ncclFloat16: CASE_TY(__half, uint16_t)
#if HAVE_ncclBfloat16
case ncclBfloat16: CASE_TY(hip_bfloat16, uint16_t)
case ncclBfloat16: CASE_TY(__nv_bfloat16, uint16_t)
#endif
case ncclFloat32: CASE_TY(float, uint32_t)
case ncclFloat64: CASE_TY(double, uint64_t)
@@ -1179,7 +1179,7 @@ __global__ void sweep() {
sweep1<uint64_t>(ncclUint64, "uint64");
sweep1<__half>(ncclFloat16, "half");
#if HAVE_ncclBfloat16
sweep1<hip_bfloat16>(ncclBfloat16, "bfloat16");
sweep1<__nv_bfloat16>(ncclBfloat16, "bfloat16");
#endif
sweep1<float>(ncclFloat32, "float");
sweep1<double>(ncclFloat64, "double");
@@ -1187,9 +1187,9 @@ __global__ void sweep() {
int main(int arg_n, char **args) {
std::cerr<<"You are hoping to see no output beyond this line."<<std::endl;
hipSetDevice(0);
cudaSetDevice(0);
sweep<<<1,512>>>();
hipDeviceSynchronize();
cudaDeviceSynchronize();
return 0;
}
#endif
verifiable/verifiable.h
+4 -4
View File
@@ -8,7 +8,7 @@
#ifndef _d41d8cd98f00b204e9800998ecf8427e
#define _d41d8cd98f00b204e9800998ecf8427e
#include <hip/hip_runtime.h>
#include <cuda_runtime.h>
#include <stdint.h>
@@ -43,13 +43,13 @@ __host__ __device__ T ncclVerifiablePremulScalar(int rank_me) {
// Enqueue kernel to generate data which is to be reduced.
void ncclVerifiablePrepareInput(
void *elts, intptr_t elt_n, int elt_ty, int red_op, int rank_n, int rank_me,
uint64_t seed, intptr_t elt_ix0, hipStream_t stream
uint64_t seed, intptr_t elt_ix0, cudaStream_t stream
);
// Enqueue kernel to generate expected results of reduction.
void ncclVerifiablePrepareExpected(
void *elts, intptr_t elt_n, int elt_ty, int red_op, int rank_n,
uint64_t seed, intptr_t elt_ix0, hipStream_t stream
uint64_t seed, intptr_t elt_ix0, cudaStream_t stream
);
// Enqueue kernel to verify reduced data matches expectation. The number of
@@ -61,6 +61,6 @@ void ncclVerifiablePrepareExpected(
void ncclVerifiableVerify(
void const *results, void const *expected, intptr_t elt_n, int elt_ty,
int red_op, int rank_n, uint64_t seed, intptr_t elt_ix0,
int64_t *bad_elt_n, hipStream_t stream
int64_t *bad_elt_n, cudaStream_t stream
);
#endif
verifiable/verifiable.mk
+16 -6
View File
@@ -1,5 +1,5 @@
# Copyright (c) 2016-2022, NVIDIA CORPORATION. All rights reserved.
# Modifications Copyright (c) 2020-2022 Advanced Micro Devices, Inc. All rights reserved.
# Modifications Copyright (c) 2020-2024 Advanced Micro Devices, Inc. All rights reserved.
#
# See LICENSE.txt for license information
@@ -8,11 +8,21 @@
# TEST_VERIFIABLE_SRCDIR = <points to this directory>
# TEST_VERIFIABLE_BUILDDIR = <points to destination of .o file>
TEST_VERIFIABLE_HDRS = $(TEST_VERIFIABLE_SRCDIR)/verifiable.h
TEST_VERIFIABLE_OBJS = $(TEST_VERIFIABLE_BUILDDIR)/verifiable.o
TEST_VERIFIABLE_HDRS = $(TEST_VERIFIABLE_SRCDIR)/verifiable.h
TEST_VERIFIABLE_OBJS = $(TEST_VERIFIABLE_BUILDDIR)/verifiable.o
$(TEST_VERIFIABLE_BUILDDIR)/verifiable.o: $(TEST_VERIFIABLE_SRCDIR)/verifiable.cu $(TEST_VERIFY_REDUCE_HDRS)
${HIPIFY_DIR}/verifiable.cu.cpp: $(TEST_VERIFIABLE_SRCDIR)/verifiable.cu
@printf "Hipifying %-35s > %s\n" $< $@
@mkdir -p ${HIPIFY_DIR}
hipify-perl -quiet-warnings $< > $@
${HIPIFY_DIR}/verifiable.h: $(TEST_VERIFIABLE_SRCDIR)/verifiable.h
@printf "Hipifying %-35s > %s\n" $< $@
@mkdir -p ${HIPIFY_DIR}
hipify-perl -quiet-warnings $< > $@
$(TEST_VERIFIABLE_BUILDDIR)/verifiable.o: $(HIPIFY_DIR)/verifiable.cu.cpp $(HIPIFY_DIR)/verifiable.h
@printf "Compiling %s\n" $@
@mkdir -p $(TEST_VERIFIABLE_BUILDDIR)
echo " $(HIPCC) -o $@ $(HIPCUFLAGS) -c $(TEST_VERIFIABLE_SRCDIR)/verifiable.cu"
$(HIPCC) -o $@ $(HIPCUFLAGS) -c $(TEST_VERIFIABLE_SRCDIR)/verifiable.cu
echo " $(HIPCC) -o $@ $(HIPCUFLAGS) -c $<"
$(HIPCC) -o $@ $(HIPCUFLAGS) -c $<