diff --git a/llvm/test/VISC/parboil/benchmarks/sgemm/src/visc_sh/Makefile b/llvm/test/VISC/parboil/benchmarks/sgemm/src/visc_sh/Makefile
index f74ee8921a534b6963ba06d089398114571d070b..96acb24ed36b2cef49d3882580831d67e7659114 100644
--- a/llvm/test/VISC/parboil/benchmarks/sgemm/src/visc_sh/Makefile
+++ b/llvm/test/VISC/parboil/benchmarks/sgemm/src/visc_sh/Makefile
@@ -6,3 +6,4 @@ VISC_OBJS=main.visc.ll
APP_CUDALDFLAGS=-lm -lstdc++
APP_CFLAGS=-ffast-math -O3
APP_CXXFLAGS=-ffast-math -O3
+APP_OPTFLAGS=-unroll-threshold=250 -loop-unroll -scalarrepl
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 25f67c31f21e5eea68966fd69d56805852ad5f5b..6906f419524463786795073320a06060351686d1 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
@@ -40,23 +40,27 @@ extern char* readFile(const char*);
}
typedef struct __attribute__((__packed__)) {
- float *A; size_t bytesA;
- int lda;
- float *B; size_t bytesB;
- int ldb;
- float *C; size_t bytesC;
- int ldc;
- int k;
- float alpha;
- float beta;
- int block_x;
- int block_y;
- int grid_x;
- int grid_y;
-} RootIn;
+ float *A;
+ size_t bytesA;
+ int lda;
+ float *B;
+ size_t bytesB;
+ int ldb;
+ float *C;
+ size_t bytesC;
+ int ldc;
+ int k;
+ float alpha;
+ float beta;
+ int block_x;
+ int block_y;
+ int grid_x;
+ int grid_y;
+}
+RootIn;
void packData(RootIn* args,
- float *A, size_t bytesA,
+ float *A, size_t bytesA,
int lda,
float *B, size_t bytesB,
int ldb,
@@ -69,35 +73,35 @@ void packData(RootIn* args,
int block_y,
int grid_x,
int grid_y) {
- args->A = A;
- args->bytesA = bytesA;
- args->lda = lda;
- args->B = B;
- args->bytesB = bytesB;
- args->ldb = ldb;
- args->C = C;
- args->bytesC = bytesC;
- args->ldc = ldc;
- args->k = k;
- args->alpha = alpha;
- args->beta = beta;
- args->block_x = block_x;
- args->block_y = block_y;
- args->grid_x = grid_x;
- args->grid_y = grid_y;
+ args->A = A;
+ args->bytesA = bytesA;
+ args->lda = lda;
+ args->B = B;
+ args->bytesB = bytesB;
+ args->ldb = ldb;
+ args->C = C;
+ args->bytesC = bytesC;
+ args->ldc = ldc;
+ args->k = k;
+ args->alpha = alpha;
+ args->beta = beta;
+ args->block_x = block_x;
+ args->block_y = block_y;
+ args->grid_x = grid_x;
+ args->grid_y = grid_y;
}
// TODO: decide between dynamic vs static allocation. Merely a convension - will
// be translated.
void Allocation(int block_x, int block_y) {
- // Memory shared between threadblocks
- //int bl_x = TILE_N;
- //int bl_y = TILE_TB_HEIGHT;
- //void* shB = __visc__malloc(bl_x*bl_y*sizeof(float));
- void* shB = __visc__malloc(block_x*block_y*sizeof(float));
-
- //__visc__return(shB, bl_x*bl_y*sizeof(float));
- __visc__return(shB, block_x*block_y*sizeof(float));
+ // Memory shared between threadblocks
+ //int bl_x = TILE_N;
+ //int bl_y = TILE_TB_HEIGHT;
+ //void* shB = __visc__malloc(bl_x*bl_y*sizeof(float));
+ void* shB = __visc__malloc(block_x*block_y*sizeof(float));
+
+ //__visc__return(shB, bl_x*bl_y*sizeof(float));
+ __visc__return(shB, block_x*block_y*sizeof(float));
}
@@ -120,85 +124,36 @@ void SgemmLeaf( float* A, size_t bytesA, int lda, float* B, size_t bytesB, int l
float c[TILE_N];
for (int i=0; i < TILE_N; i++)
- c[i] = 0.0f;
- //float c0 = 0;
- //float c1 = 0;
- //float c2 = 0;
- //float c3 = 0;
- //float c4 = 0;
- //float c5 = 0;
- //float c6 = 0;
- //float c7 = 0;
- //float c8 = 0;
- //float c9 = 0;
- //float c10 = 0;
- //float c11 = 0;
- //float c12 = 0;
- //float c13 = 0;
- //float c14 = 0;
- //float c15 = 0;
-
+ c[i] = 0.0f;
+
int mid = ly*dimx+lx;
int m = gx * TILE_M + mid;
int n = gy * TILE_N + lx;
for (int i = 0; i < k; i+=TILE_TB_HEIGHT) {
- float a;
- //shB[ly][lx] = B[n+(i+ly)*ldb];
- shB[ly*dimx+lx] = B[n+(i+ly)*ldb];
-
- __visc__barrier();
- for (int j = 0; j < TILE_TB_HEIGHT; j++) {
- a = A[m + (i+j)*lda];
- for (int kk = 0; kk < TILE_N; kk++) {
- //c[kk] += a * shB[j][kk];
- c[kk] += a * shB[j*dimx+kk];
- }
- //c0 += a * shB[j*dimx+0];
- //c1 += a * shB[j*dimx+1];
- //c2 += a * shB[j*dimx+2];
- //c3 += a * shB[j*dimx+3];
- //c4 += a * shB[j*dimx+4];
- //c5 += a * shB[j*dimx+5];
- //c6 += a * shB[j*dimx+6];
- //c7 += a * shB[j*dimx+7];
- //c8 += a * shB[j*dimx+8];
- //c9 += a * shB[j*dimx+9];
- //c10 += a * shB[j*dimx+10];
- //c11 += a * shB[j*dimx+11];
- //c12 += a * shB[j*dimx+12];
- //c13 += a * shB[j*dimx+13];
- //c14 += a * shB[j*dimx+14];
- //c15 += a * shB[j*dimx+15];
-
- }
- __visc__barrier();
+ float a;
+ //shB[ly][lx] = B[n+(i+ly)*ldb];
+ shB[ly*dimx+lx] = B[n+(i+ly)*ldb];
+
+ __visc__barrier();
+ for (int j = 0; j < TILE_TB_HEIGHT; j++) {
+ a = A[m + (i+j)*lda];
+ for (int kk = 0; kk < TILE_N; kk++) {
+ //c[kk] += a * shB[j][kk];
+ c[kk] += a * shB[j*dimx+kk];
+ }
+ }
+ __visc__barrier();
}
int t = ldc * gy * TILE_N + m;
for (int i = 0; i < TILE_N; i++) {
- C[t+i*ldc] = C[t+i*ldc] * beta + alpha * c[i];
+ C[t+i*ldc] = C[t+i*ldc] * beta + alpha * c[i];
}
- //C[t+0*ldc] = C[t+0*ldc] * beta + alpha * c0;
- //C[t+1*ldc] = C[t+1*ldc] * beta + alpha * c1;
- //C[t+2*ldc] = C[t+2*ldc] * beta + alpha * c2;
- //C[t+3*ldc] = C[t+3*ldc] * beta + alpha * c3;
- //C[t+4*ldc] = C[t+4*ldc] * beta + alpha * c4;
- //C[t+5*ldc] = C[t+5*ldc] * beta + alpha * c5;
- //C[t+6*ldc] = C[t+6*ldc] * beta + alpha * c6;
- //C[t+7*ldc] = C[t+7*ldc] * beta + alpha * c7;
- //C[t+8*ldc] = C[t+8*ldc] * beta + alpha * c8;
- //C[t+9*ldc] = C[t+9*ldc] * beta + alpha * c9;
- //C[t+10*ldc] = C[t+10*ldc] * beta + alpha * c10;
- //C[t+11*ldc] = C[t+11*ldc] * beta + alpha * c11;
- //C[t+12*ldc] = C[t+12*ldc] * beta + alpha * c12;
- //C[t+13*ldc] = C[t+13*ldc] * beta + alpha * c13;
- //C[t+14*ldc] = C[t+14*ldc] * beta + alpha * c14;
- //C[t+15*ldc] = C[t+15*ldc] * beta + alpha * c15;
}
// Thread block node for sgemm - Creates allocation node and leaf (thread) node
-void SgemmTB(float *A, size_t bytesA,
+void SgemmTB(float *A, size_t bytesA,
int lda,
float *B, size_t bytesB,
int ldb,
@@ -209,38 +164,38 @@ void SgemmTB(float *A, size_t bytesA,
float beta,
int block_x,
int block_y) {
- __visc__hint(visc::CPU_TARGET);
- __visc__attributes(3, A, B, C, 1, C);
- void* AllocationNode = __visc__createNode(Allocation);
- void* SgemmLeafNode = __visc__createNode2D(SgemmLeaf, block_x, block_y);
-
- // Bind edges
- __visc__bindIn(SgemmLeafNode, 0, 0, 0); // Bind A
- __visc__bindIn(SgemmLeafNode, 1, 1, 0); // Bind bytesA
- __visc__bindIn(SgemmLeafNode, 2, 2, 0); // Bind lda
- __visc__bindIn(SgemmLeafNode, 3, 3, 0); // Bind B
- __visc__bindIn(SgemmLeafNode, 4, 4, 0); // Bind bytesB
- __visc__bindIn(SgemmLeafNode, 5, 5, 0); // Bind ldb
- __visc__bindIn(SgemmLeafNode, 6, 6, 0); // Bind C
- __visc__bindIn(SgemmLeafNode, 7, 7, 0); // Bind bytesC
- __visc__bindIn(SgemmLeafNode, 8, 8, 0); // Bind ldc
- __visc__bindIn(SgemmLeafNode, 9, 9, 0); // Bind k
- __visc__bindIn(SgemmLeafNode, 10, 10, 0); // Bind alpha
- __visc__bindIn(SgemmLeafNode, 11, 11, 0); // Bind beta
-
- __visc__bindIn(AllocationNode, 12, 0, 0); // Bind block_x
- __visc__bindIn(AllocationNode, 13, 1, 0); // Bind block_y
-
-
- // Create Edges between AllocationNode and BFSLeafNodeNode
- __visc__edge(AllocationNode, SgemmLeafNode, 0, 12, 0); // Edge local_B
- __visc__edge(AllocationNode, SgemmLeafNode, 1, 13, 0); // Edge bytes_local_B
-
- //TODO: bindOut : for now with out attribute
+ __visc__hint(visc::CPU_TARGET);
+ __visc__attributes(3, A, B, C, 1, C);
+ void* AllocationNode = __visc__createNode(Allocation);
+ void* SgemmLeafNode = __visc__createNode2D(SgemmLeaf, block_x, block_y);
+
+ // Bind edges
+ __visc__bindIn(SgemmLeafNode, 0, 0, 0); // Bind A
+ __visc__bindIn(SgemmLeafNode, 1, 1, 0); // Bind bytesA
+ __visc__bindIn(SgemmLeafNode, 2, 2, 0); // Bind lda
+ __visc__bindIn(SgemmLeafNode, 3, 3, 0); // Bind B
+ __visc__bindIn(SgemmLeafNode, 4, 4, 0); // Bind bytesB
+ __visc__bindIn(SgemmLeafNode, 5, 5, 0); // Bind ldb
+ __visc__bindIn(SgemmLeafNode, 6, 6, 0); // Bind C
+ __visc__bindIn(SgemmLeafNode, 7, 7, 0); // Bind bytesC
+ __visc__bindIn(SgemmLeafNode, 8, 8, 0); // Bind ldc
+ __visc__bindIn(SgemmLeafNode, 9, 9, 0); // Bind k
+ __visc__bindIn(SgemmLeafNode, 10, 10, 0); // Bind alpha
+ __visc__bindIn(SgemmLeafNode, 11, 11, 0); // Bind beta
+
+ __visc__bindIn(AllocationNode, 12, 0, 0); // Bind block_x
+ __visc__bindIn(AllocationNode, 13, 1, 0); // Bind block_y
+
+
+ // Create Edges between AllocationNode and BFSLeafNodeNode
+ __visc__edge(AllocationNode, SgemmLeafNode, 0, 12, 0); // Edge local_B
+ __visc__edge(AllocationNode, SgemmLeafNode, 1, 13, 0); // Edge bytes_local_B
+
+ //TODO: bindOut : for now with out attribute
}
// Root node for sgemm - Creates thread block node
-void SgemmRoot(float *A, size_t bytesA,
+void SgemmRoot(float *A, size_t bytesA,
int lda,
float *B, size_t bytesB,
int ldb,
@@ -253,27 +208,27 @@ void SgemmRoot(float *A, size_t bytesA,
int block_y,
int grid_x,
int grid_y) {
- __visc__hint(visc::CPU_TARGET);
- __visc__attributes(3, A, B, C, 1, C);
- void* SgemmTBNode = __visc__createNode2D(SgemmTB, grid_x, grid_y);
-
- // Bind edges
- __visc__bindIn(SgemmTBNode, 0, 0, 0); // Bind A
- __visc__bindIn(SgemmTBNode, 1, 1, 0); // Bind bytesA
- __visc__bindIn(SgemmTBNode, 2, 2, 0); // Bind lda
- __visc__bindIn(SgemmTBNode, 3, 3, 0); // Bind B
- __visc__bindIn(SgemmTBNode, 4, 4, 0); // Bind bytesB
- __visc__bindIn(SgemmTBNode, 5, 5, 0); // Bind ldb
- __visc__bindIn(SgemmTBNode, 6, 6, 0); // Bind C
- __visc__bindIn(SgemmTBNode, 7, 7, 0); // Bind bytesC
- __visc__bindIn(SgemmTBNode, 8, 8, 0); // Bind ldc
- __visc__bindIn(SgemmTBNode, 9, 9, 0); // Bind k
- __visc__bindIn(SgemmTBNode, 10, 10, 0); // Bind alpha
- __visc__bindIn(SgemmTBNode, 11, 11, 0); // Bind beta
- __visc__bindIn(SgemmTBNode, 12, 12, 0); // Bind block_x
- __visc__bindIn(SgemmTBNode, 13, 13, 0); // Bind block_y
-
- //TODO: bindOut : for now with out attribute
+ __visc__hint(visc::CPU_TARGET);
+ __visc__attributes(3, A, B, C, 1, C);
+ void* SgemmTBNode = __visc__createNode2D(SgemmTB, grid_x, grid_y);
+
+ // Bind edges
+ __visc__bindIn(SgemmTBNode, 0, 0, 0); // Bind A
+ __visc__bindIn(SgemmTBNode, 1, 1, 0); // Bind bytesA
+ __visc__bindIn(SgemmTBNode, 2, 2, 0); // Bind lda
+ __visc__bindIn(SgemmTBNode, 3, 3, 0); // Bind B
+ __visc__bindIn(SgemmTBNode, 4, 4, 0); // Bind bytesB
+ __visc__bindIn(SgemmTBNode, 5, 5, 0); // Bind ldb
+ __visc__bindIn(SgemmTBNode, 6, 6, 0); // Bind C
+ __visc__bindIn(SgemmTBNode, 7, 7, 0); // Bind bytesC
+ __visc__bindIn(SgemmTBNode, 8, 8, 0); // Bind ldc
+ __visc__bindIn(SgemmTBNode, 9, 9, 0); // Bind k
+ __visc__bindIn(SgemmTBNode, 10, 10, 0); // Bind alpha
+ __visc__bindIn(SgemmTBNode, 11, 11, 0); // Bind beta
+ __visc__bindIn(SgemmTBNode, 12, 12, 0); // Bind block_x
+ __visc__bindIn(SgemmTBNode, 13, 13, 0); // Bind block_y
+
+ //TODO: bindOut : for now with out attribute
}
// Creates root node for sgemm
@@ -291,8 +246,8 @@ __attribute__((noinline)) void basicSgemm( char transa, char transb, int m, int
// In this code we assume the matrix sizes are multiple of tile size
if ((m%TILE_M) || (n%TILE_N)) {
- std::cerr << "unsupported size of matrix. m should be multiple of " << TILE_M
- << "; n should be multiple of " << TILE_N << std::endl;
+ std::cerr << "unsupported size of matrix. m should be multiple of " << TILE_M
+ << "; n should be multiple of " << TILE_N << std::endl;
return;
}
@@ -302,27 +257,27 @@ __attribute__((noinline)) void basicSgemm( char transa, char transb, int m, int
// unsigned sgemmDFG = __visc__node(mysgemmNT, 2, 2, db[0], db[1], dg[0]/db[0], dg[1]/db[1], 12, A, bytesA, lda, B, bytesB, ldb, C, bytesC, ldc, k, alpha, beta, 0);
- int block_x = TILE_N;
- int block_y = TILE_TB_HEIGHT;
- int grid_x = m/TILE_M;
- int grid_y = n/TILE_N;
- // Pack data in struct
- RootIn* args = (RootIn*) malloc(sizeof(RootIn));
- packData(args,
- A, bytesA,
- lda,
- B, bytesB,
- ldb,
- C, bytesC,
- ldc,
- k,
- alpha,
- beta,
- block_x,
- block_y,
- grid_x,
- grid_y
- );
+ int block_x = TILE_N;
+ int block_y = TILE_TB_HEIGHT;
+ int grid_x = m/TILE_M;
+ int grid_y = n/TILE_N;
+ // Pack data in struct
+ RootIn* args = (RootIn*) malloc(sizeof(RootIn));
+ packData(args,
+ A, bytesA,
+ lda,
+ B, bytesB,
+ ldb,
+ C, bytesC,
+ ldc,
+ k,
+ alpha,
+ beta,
+ block_x,
+ block_y,
+ grid_x,
+ grid_y
+ );
void* sgemmDFG = __visc__launch(0, SgemmRoot, (void*) args);
diff --git a/llvm/test/VISC/parboil/common/mk/visc.mk b/llvm/test/VISC/parboil/common/mk/visc.mk
index 6c71fa2b69086b581d60661fa036af525f05ddad..ce2d2c6ecb8293aaa6c5a2555727d9c892259d58 100644
--- a/llvm/test/VISC/parboil/common/mk/visc.mk
+++ b/llvm/test/VISC/parboil/common/mk/visc.mk
@@ -97,6 +97,7 @@ TEST_OBJS = $(call INBUILDDIR,$(VISC_OBJS))
PARBOIL_OBJS = $(call INBUILDDIR,parboil.ll)
KERNEL = $(TEST_OBJS).kernels.ll
ifneq ($(TARGET),x86)
+ KERNEL_OPT = $(BUILDDIR)/$(APP).kernels.opt.ll
KERNEL_LINKED = $(BUILDDIR)/$(APP).kernels.linked.ll
#KERNEL = $(TEST_OBJS).kernels.ll
PTX_ASSEMBLY = $(TEST_OBJS).nvptx.s
@@ -142,9 +143,12 @@ clean :
$(PTX_ASSEMBLY) : $(KERNEL_LINKED)
$(CC) $(KERNEL_GEN_FLAGS) -S $< -o $@
-$(KERNEL_LINKED) : $(KERNEL)
+$(KERNEL_LINKED) : $(KERNEL_OPT)
$(LLVM_LINK) $(LIBCLC_NVPTX_LIB) -S $< -o $@
+$(KERNEL_OPT) : $(KERNEL)
+ $(OPT) $(APP_OPTFLAGS) -S $< -o $@
+
$(BIN) : $(HOST_LINKED)
$(CXX) -O3 $(LDFLAGS) $< -o $@