From 41a048fa4bb9ddefd4e4acff2207251bb3ddbf06 Mon Sep 17 00:00:00 2001
From: Cyril SIX <cyril.six@kalray.eu>
Date: Wed, 18 Apr 2018 16:44:43 +0200
Subject: MPPA - Added divide & conqueer test matmul

---
 test/mppa/mmult/mmult.c | 88 +++++++++++++++++++++++++++++++++++++++++++++++--
 test/mppa/mmult/mmult.h |  1 +
 2 files changed, 87 insertions(+), 2 deletions(-)

(limited to 'test')

diff --git a/test/mppa/mmult/mmult.c b/test/mppa/mmult/mmult.c
index 16dcf34c..c9e7ad5e 100644
--- a/test/mppa/mmult/mmult.c
+++ b/test/mppa/mmult/mmult.c
@@ -31,8 +31,91 @@ void mmult_col(uint64_t C[][SIZE], uint64_t A[][SIZE], uint64_t B[][SIZE]){
                 C[i][j] += A[i][k] * B[k][j];
 }
 
+typedef struct mblock {
+    int imin, imax, jmin, jmax;
+    uint64_t *mat;
+} mblock;
+
+#define MAT_XY(mat, x, y) (mat)[(x)*SIZE + (y)]
+#define MAT_IJ(block, i, j) MAT_XY((block)->mat, (block)->imin + (i), block->jmin + (j))
+
+int strassen_mul(mblock *C, const mblock *A, const mblock *B){
+    const int size = C->imax - C->imin;
+
+    for (int i = 0 ; i < size ; i++)
+        for (int j = 0 ; j < size ; j++)
+            for (int k = 0 ; k < size ; k++)
+                MAT_IJ(C, i, j) += MAT_IJ(A, i, k) * MAT_IJ(B, k, j);
+}
+
+#define BLOCK_X_MID(block) ((block)->imin + (block)->imax) / 2
+#define BLOCK_Y_MID(block) ((block)->jmin + (block)->jmax) / 2
+
+#define MAKE_MBLOCK(newb, block, I, J) \
+    mblock newb = {.mat=(block)->mat};\
+    if ((I) == 0){\
+        newb.imin = (block)->imin;\
+        newb.imax = BLOCK_X_MID((block));\
+    } else {\
+        newb.imin = BLOCK_X_MID((block));\
+        newb.imax = (block)->imax;\
+    } if ((J) == 0){\
+        newb.jmin = (block)->jmin;\
+        newb.jmax = BLOCK_Y_MID((block));\
+    } else {\
+        newb.jmin = BLOCK_Y_MID((block));\
+        newb.jmax = (block)->jmax;\
+    }
+
+int strassen_part(mblock *C, const mblock *A, const mblock *B);
+
+void strassen_wrap(mblock *C      , char IC, char JC,
+                   const mblock *A, char IA, char JA,
+                   const mblock *B, char IB, char JB){
+    MAKE_MBLOCK(Cb, C, IC, JC);
+    MAKE_MBLOCK(Ab, A, IA, JA);
+    MAKE_MBLOCK(Bb, B, IB, JB);
+
+    strassen_part(&Cb, &Ab, &Bb);
+}
+
+
+int strassen_part(mblock *C, const mblock *A, const mblock *B){
+    const int size = C->imax - C->imin;
+
+    if (size % 2 == 1)
+        strassen_mul(C, A, B);
+    else{
+        /* C_00 = A_00 B_00 + A_01 B_10 */
+        strassen_wrap(C, 0, 0, A, 0, 0, B, 0, 0);
+        strassen_wrap(C, 0, 0, A, 0, 1, B, 1, 0);
+
+        /* C_10 = A_10 B_00 + A_11 B_10 */
+        strassen_wrap(C, 1, 0, A, 1, 0, B, 0, 0);
+        strassen_wrap(C, 1, 0, A, 1, 1, B, 1, 0);
+
+        /* C_01 = A_00 B_01 + A_01 B_11 */
+        strassen_wrap(C, 0, 1, A, 0, 0, B, 0, 1);
+        strassen_wrap(C, 0, 1, A, 0, 1, B, 1, 1);
+
+        /* C_11 = A_10 B_01 + A_11 B_11 */
+        strassen_wrap(C, 1, 1, A, 1, 0, B, 0, 1);
+        strassen_wrap(C, 1, 1, A, 1, 1, B, 1, 1);
+    }
+    
+}
+
+void mmult_strassen(uint64_t C[][SIZE], uint64_t A[][SIZE], uint64_t B[][SIZE]){
+    mblock Cb = {.mat = (uint64_t *) C, .imin = 0, .imax = SIZE, .jmin = 0, .jmax = SIZE};
+    mblock Ab = {.mat = (uint64_t *) A , .imin = 0, .imax = SIZE, .jmin = 0, .jmax = SIZE};
+    mblock Bb = {.mat = (uint64_t *) B , .imin = 0, .imax = SIZE, .jmin = 0, .jmax = SIZE};
+
+    strassen_part(&Cb, &Ab, &Bb);
+}
+
 #ifdef __UNIT_TEST_MMULT__
-static uint64_t C1[SIZE][SIZE], C2[SIZE][SIZE], A[SIZE][SIZE], B[SIZE][SIZE];
+static uint64_t C1[SIZE][SIZE], C2[SIZE][SIZE], C3[SIZE][SIZE];
+static uint64_t A[SIZE][SIZE], B[SIZE][SIZE];
 
 int main(void){
     srand(42);
@@ -45,10 +128,11 @@ int main(void){
 
     mmult_row(C1, A, B);
     mmult_col(C2, A, B);
+    mmult_strassen(C3, A, B);
 
     for (int i = 0 ; i < SIZE ; i++)
         for (int j = 0 ; j < SIZE ; j++)
-            if (C1[i][j] != C2[i][j])
+            if (!(C1[i][j] == C2[i][j] && C1[i][j] == C3[i][j]))
                 return -1;
 
     return 0;
diff --git a/test/mppa/mmult/mmult.h b/test/mppa/mmult/mmult.h
index 50c04afd..3721784a 100644
--- a/test/mppa/mmult/mmult.h
+++ b/test/mppa/mmult/mmult.h
@@ -5,5 +5,6 @@
 
 void mmult_row(uint64_t *A, const uint64_t *B, const uint64_t *C);
 void mmult_column(uint64_t *A, const uint64_t *B, const uint64_t *C);
+void mmult_strassen(uint64_t *A, const uint64_t *B, const uint64_t *C);
 
 #endif /* __MMULT_H__ */
-- 
cgit