From 75b7dfd0d4189a162b359c1ff3182fbd9a1b31a6 Mon Sep 17 00:00:00 2001
From: Prakalp Srivastava <psrivas2@illinois.edu>
Date: Tue, 8 Mar 2016 01:04:32 -0600
Subject: [PATCH] (1) Generating correct addrspace for global pointers in
opencl kernels (2) Fixed a bug in sgemm visc_sh version. Wrong grid size was
used earlier, resulting in the bug
---
.../Target/NVPTX/MCTargetDesc/NVPTXBaseInfo.h | 4 ++--
.../DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp | 18 +++++++++++++++++-
.../benchmarks/sgemm/src/visc_sh/main.cc | 5 ++---
3 files changed, 21 insertions(+), 6 deletions(-)
diff --git a/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXBaseInfo.h b/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXBaseInfo.h
index f42f93f6c1..edf4a80b11 100644
--- a/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXBaseInfo.h
+++ b/llvm/lib/Target/NVPTX/MCTargetDesc/NVPTXBaseInfo.h
@@ -20,8 +20,8 @@
namespace llvm {
enum AddressSpace {
- ADDRESS_SPACE_GENERIC = 1,
- ADDRESS_SPACE_GLOBAL = 0,
+ ADDRESS_SPACE_GENERIC = 0,
+ ADDRESS_SPACE_GLOBAL = 1,
ADDRESS_SPACE_SHARED = 3,
ADDRESS_SPACE_CONST = 4,
ADDRESS_SPACE_LOCAL = 5,
diff --git a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
index 942b79f686..c614d2e4ce 100644
--- a/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
+++ b/llvm/lib/Transforms/DFG2LLVM_NVPTX/DFG2LLVM_NVPTX.cpp
@@ -904,7 +904,6 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
transformFunctionToVoid(F_nvptx);
-
//FIXME: For now, assume only one allocation node
kernel->AllocationNode = NULL;
@@ -940,11 +939,28 @@ void CGT_NVPTX::codeGen(DFLeafNode* N) {
}
std::sort(SharedMemArgs.begin(), SharedMemArgs.end());
+ // All pointer args which are not shared memory pointers have to be moved to
+ // global address space
+ unsigned argIndex = 0;
+ std::vector<unsigned> GlobalMemArgs;
+ for(auto& Arg: F_nvptx->getArgumentList()) {
+ if (Arg.getType()->isPointerTy()) {
+ // If the arguement is already chosen for shared memory arguemnt list, skip.
+ // Else put it in Global memory arguement list
+ if(std::count(SharedMemArgs.begin(), SharedMemArgs.end(), argIndex) == 0) {
+ GlobalMemArgs.push_back(argIndex);
+ }
+ }
+ argIndex++;
+ }
+ std::sort(GlobalMemArgs.begin(), GlobalMemArgs.end());
+
/* At this point, we assume that chescks for the fact that SharedMemArgs only
contains pointer arguments to GLOBAL_ADDRSPACE have been performed by the
analysis pass */
changeArgAddrspace(F_nvptx, SharedMemArgs, SHARED_ADDRSPACE);
+ changeArgAddrspace(F_nvptx, GlobalMemArgs, GLOBAL_ADDRSPACE);
// Go through all the instructions
diff --git a/llvm/test/VISC/parboil/benchmarks/sgemm/src/visc_sh/main.cc b/llvm/test/VISC/parboil/benchmarks/sgemm/src/visc_sh/main.cc
index c40ce36c4b..0e994d478b 100644
--- a/llvm/test/VISC/parboil/benchmarks/sgemm/src/visc_sh/main.cc
+++ b/llvm/test/VISC/parboil/benchmarks/sgemm/src/visc_sh/main.cc
@@ -254,8 +254,8 @@ __attribute__((noinline)) void basicSgemm( char transa, char transb, int m, int
int block_x = TILE_N;
int block_y = TILE_TB_HEIGHT;
- int grid_x = m*TILE_N/TILE_M;
- int grid_y = n*TILE_TB_HEIGHT/TILE_N;
+ int grid_x = m/TILE_M;
+ int grid_y = n/TILE_N;
// Pack data in struct
RootIn* args = (RootIn*) malloc(sizeof(RootIn));
packData(args,
@@ -289,7 +289,6 @@ int main (int argc, char *argv[]) {
int matBrow, matBcol;
std::vector<float> matA, matBT;
-
/* Read command line. Expect 3 inputs: A, B and B^T
in column-major layout*/
params = pb_ReadParameters(&argc, argv);
--
GitLab