Refs #532. Improve gemv paralel with small m and large n case.

Splite the matrix and reduction.

Author: xianyi

driver/level2/gemv_thread.c

@@ -62,6 +62,11 @@
 #endif
 #endif
 
+#ifndef TRANSA
+#define Y_DUMMY_NUM 1024
+static FLOAT y_dummy[Y_DUMMY_NUM];
+#endif
+
 static int gemv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *dummy1, FLOAT *buffer, BLASLONG pos){
 
   FLOAT *a, *x, *y;
@@ -99,10 +104,15 @@ static int gemv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
     a += n_from * lda  * COMPSIZE;
 #ifdef TRANSA
     y += n_from * incy * COMPSIZE;
+#else
+    //for split matrix row (n) direction and vector x of gemv_n
+    x += n_from * incx * COMPSIZE;
+    //store partial result for every thread
+    y += (m_to - m_from) * 1 * COMPSIZE * pos;
 #endif
   }
 
-  //  fprintf(stderr, "M_From = %d  M_To = %d  N_From = %d  N_To = %d\n", m_from, m_to, n_from, n_to);
+  //fprintf(stderr, "M_From = %d  M_To = %d  N_From = %d  N_To = %d POS=%d\n", m_from, m_to, n_from, n_to, pos);
 
   GEMV(m_to - m_from, n_to - n_from, 0,
        *((FLOAT *)args -> alpha + 0),
@@ -126,6 +136,10 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *alpha, FLOAT *a, BLASLONG lda, FLOAT *x
 
   BLASLONG width, i, num_cpu;
 
+#ifndef TRANSA
+  int split_x=0;
+#endif
+
 #ifdef SMP
 #ifndef COMPLEX
 #ifdef XDOUBLE
@@ -198,6 +212,58 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *alpha, FLOAT *a, BLASLONG lda, FLOAT *x
     i -= width;
   }
 
+#ifndef TRANSA
+  //try to split matrix on row direction and x.
+  //Then, reduction.
+  if (num_cpu < nthreads) {
+
+    //too small to split or bigger than the y_dummy buffer.
+    double MN = (double) m * (double) n;
+    if ( MN <= (24.0 * 24.0  * (double) (GEMM_MULTITHREAD_THRESHOLD*GEMM_MULTITHREAD_THRESHOLD))
+	 || m*COMPSIZE*nthreads > Y_DUMMY_NUM)
+      goto Outer;
+
+    num_cpu  = 0;
+    range[0] = 0;
+
+    memset(y_dummy, 0, sizeof(FLOAT) * m * COMPSIZE * nthreads);
+
+    args.ldc = 1;
+    args.c = (void *)y_dummy;
+
+    //split on row (n) and x
+    i=n;
+    split_x=1;
+    while (i > 0){
+
+      width  = blas_quickdivide(i + nthreads - num_cpu - 1, nthreads - num_cpu);
+      if (width < 4) width = 4;
+      if (i < width) width = i;
+
+      range[num_cpu + 1] = range[num_cpu] + width;
+
+      queue[num_cpu].mode    = mode;
+      queue[num_cpu].routine = gemv_kernel;
+      queue[num_cpu].args    = &args;
+
+      queue[num_cpu].position = num_cpu;
+
+      queue[num_cpu].range_m = NULL;
+      queue[num_cpu].range_n = &range[num_cpu];
+
+      queue[num_cpu].sa      = NULL;
+      queue[num_cpu].sb      = NULL;
+      queue[num_cpu].next    = &queue[num_cpu + 1];
+
+      num_cpu ++;
+      i -= width;
+    }
+
+  }
+
+  Outer:
+#endif
+
   if (num_cpu) {
     queue[0].sa = NULL;
     queue[0].sb = buffer;
@@ -206,5 +272,21 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *alpha, FLOAT *a, BLASLONG lda, FLOAT *x
     exec_blas(num_cpu, queue);
   }
 
+#ifndef TRANSA
+  if(split_x==1){
+    //reduction
+    for(i=0; i<num_cpu; i++){
+
+      int j;
+      for(j=0; j<m; j++){
+	y[j*incy*COMPSIZE] +=y_dummy[i*m*COMPSIZE + j*COMPSIZE];
+#ifdef COMPLEX
+	y[j*incy*COMPSIZE+1] +=y_dummy[i*m*COMPSIZE + j*COMPSIZE+1];
+#endif
+      }
+    }
+  }
+#endif
+
   return 0;
 }