Merge f1aaf0777a into 0ab5bf1746

benchmark/gemm.c

@@ -237,7 +237,7 @@ int main(int argc, char *argv[]){
     timeg = time1/loops;
     fprintf(stderr,
 	    " %10.2f MFlops %10.6f sec\n",
-	    COMPSIZE * COMPSIZE * 2. * (double)m * (double)m * (double)n / timeg * 1.e-6, time1);
+	    COMPSIZE * COMPSIZE * (2.*(double)k+2.) * (double)m * (double)n / timeg * 1.e-6, time1);
     
   }
 

kernel/generic/ztrmm_ltcopy_8.c

@@ -801,32 +801,20 @@ int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, BLASLONG posX, BLASLON
 
     i = (m & 1);
     if (i > 0) {
-      if (X > posY) {
-	/* a01 += 2;
-	a02 += 2; */
-	b += 4;
-      } else
-	if (X < posY) {
-	  b[  0] = *(a01 +  0);
-	  b[  1] = *(a01 +  1);
-	  b[  2] = *(a01 +  2);
-	  b[  3] = *(a01 +  3);
-
-	  /* a01 += lda;
-	  a02 += lda; */
-	  b += 4;
-	} else {
 #ifdef UNIT
-	  b[  0] = ONE;
-	  b[  1] = ZERO;
-#else
-	  b[  0] = *(a01 +  0);
-	  b[  1] = *(a01 +  1);
+      if (X < posY) {
 #endif
-	  b[  2] = *(a01 +  2);
-	  b[  3] = *(a01 +  3);
-	  b += 4;
-	}
+	b[  0] = *(a01 +  0);
+	b[  1] = *(a01 +  1);
+#ifdef UNIT
+      } else {
+	b[  0] = ONE;
+	b[  1] = ZERO;
+      }
+#endif
+      b[  2] = *(a01 +  2);
+      b[  3] = *(a01 +  3);
+      b += 4;
     }
     posY += 2;
   }

kernel/x86_64/dtrmm_kernel_4x8_haswell.c

@@ -301,7 +301,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
 	        dtrmm_kernel_4x8( temp, &alpha , ptrba, ptrbb, C0, C1, C2, C3, C4, C5, C6, C7);
 
 		ptrba = ptrba + temp * 4;
-		ptrbb = ptrbb + temp * 8;
+		// ptrbb = ptrbb + temp * 8;
 
 /*
 		for (k=0; k<temp; k++)
@@ -446,7 +446,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
                                   temp - 8;  // number of values in B
 
                     ptrba += temp*4; // number of values in A
-		    ptrbb += temp*8; // number of values in B
+		    // ptrbb += temp*8; // number of values in B
                 }
 #ifdef LEFT
 		off += 4; // number of values in A
@@ -709,14 +709,14 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
 		off += 1; // number of values in A
 #endif
 
-		C0 = C0+1;
+		/* C0 = C0+1;
 		C1 = C1+1;
 		C2 = C2+1;
 		C3 = C3+1;
 		C4 = C4+1;
 		C5 = C5+1;
 		C6 = C6+1;
-		C7 = C7+1;
+		C7 = C7+1; */
 
 	}
 
@@ -862,7 +862,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
                                   temp - 4;  // number of values in B */
 
                     ptrba += temp*4; // number of values in A
-		    ptrbb += temp*4; // number of values in B
+		    // ptrbb += temp*4; // number of values in B
                 }
 #ifdef LEFT
 		off += 4; // number of values in A
@@ -1049,10 +1049,10 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
 		off += 1; // number of values in A
 #endif
 
-		C0 = C0+1;
+		/* C0 = C0+1;
 		C1 = C1+1;
 		C2 = C2+1;
-		C3 = C3+1;
+		C3 = C3+1; */
 
 	}
 
@@ -1311,8 +1311,8 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
 		off += 1; // number of values in A
 #endif
 
-		C0 = C0+1;
-		C1 = C1+1;
+		/* C0 = C0+1;
+		C1 = C1+1; */
 
 	}
 
@@ -1532,7 +1532,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alpha,FLOAT* ba,FLOAT* bb,FL
 		off += 1; // number of values in A
 #endif
 
-		C0 = C0+1;
+		// C0 = C0+1;
 
 	}
 

lapack/getrf/getrf_parallel.c

@@ -124,13 +124,13 @@ static void inner_basic_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *ra
       min_jj = js + min_j - jjs;
       if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
-      if (0 && GEMM_UNROLL_N <= 8) {
+      /* if (0 && GEMM_UNROLL_N <= 8) {
 
 	LASWP_NCOPY(min_jj, off + 1, off + k,
 		    c + (- off + jjs * lda) * COMPSIZE, lda,
 		    ipiv, sbb + k * (jjs - js) * COMPSIZE);
 
-      } else {
+      } else { */
 
 	LASWP_PLUS(min_jj, off + 1, off + k, ZERO,
 #ifdef COMPLEX
@@ -140,7 +140,7 @@ static void inner_basic_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *ra
 
 	GEMM_ONCOPY (k, min_jj, c + jjs * lda * COMPSIZE, lda, sbb + (jjs - js) * k * COMPSIZE);
 
-      }
+      // }
 
       for (is = 0; is < k; is += GEMM_P) {
 	min_i = k - is;
@@ -251,14 +251,14 @@ static int inner_advanced_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *
       min_jj = MIN(n_to, xxx + div_n) - jjs;
       if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
-      if (0 && GEMM_UNROLL_N <= 8) {
+      /* if (0 && GEMM_UNROLL_N <= 8) {
 	printf("helllo\n");
 
 	LASWP_NCOPY(min_jj, off + 1, off + k,
 		    b + (- off + jjs * lda) * COMPSIZE, lda,
 		    ipiv, buffer[bufferside] + (jjs - xxx) * k * COMPSIZE);
 
-      } else {
+      } else { */
 
 	LASWP_PLUS(min_jj, off + 1, off + k, ZERO,
 #ifdef COMPLEX
@@ -268,7 +268,7 @@ static int inner_advanced_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *
 
 	GEMM_ONCOPY (k, min_jj, b + jjs * lda * COMPSIZE, lda,
 		     buffer[bufferside] + (jjs - xxx) * k * COMPSIZE);
-      }
+      // }
 
       for (is = 0; is < k; is += GEMM_P) {
 	min_i = k - is;

lapack/trtri/trtri_L_parallel.c

@@ -54,7 +54,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
   BLASLONG n, info;
   BLASLONG bk, i, blocking, start_i;
   int mode;
-  BLASLONG lda, range_N[2];
+  BLASLONG lda;//, range_N[2];
   blas_arg_t newarg;
   FLOAT *a;
   FLOAT alpha[2] = { ONE, ZERO};
@@ -100,8 +100,8 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
     bk = n - i;
     if (bk > blocking) bk = blocking;
 
-    range_N[0] = i;
-    range_N[1] = i + bk;
+    /* range_N[0] = i;
+    range_N[1] = i + bk; */
 
     newarg.lda = lda;
     newarg.ldb = lda;

lapack/trtri/trtri_U_parallel.c

@@ -54,7 +54,7 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
   BLASLONG n, info;
   BLASLONG bk, i, blocking;
   int mode;
-  BLASLONG lda, range_N[2];
+  BLASLONG lda; // , range_N[2];
   blas_arg_t newarg;
   FLOAT *a;
   FLOAT alpha[2] = { ONE, ZERO};
@@ -96,8 +96,8 @@ blasint CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa,
     bk = n - i;
     if (bk > blocking) bk = blocking;
 
-    range_N[0] = i;
-    range_N[1] = i + bk;
+    /* range_N[0] = i;
+    range_N[1] = i + bk; */
 
     newarg.lda = lda;
     newarg.ldb = lda;