Merge pull request #3547 from martin-frbg/issue3540-2

More build fixes for CooperLake with BFLOAT16 and DYNAMIC_ARCH

Author: martin-frbg

kernel/CMakeLists.txt

@@ -678,7 +678,7 @@ endif ()
           set(SBGEMM_SMALL_K_B0_TN ../generic/gemm_small_matrix_kernel_tn.c)
       endif ()
       if (NOT DEFINED SBGEMM_SMALL_K_B0_TT)
-          set($SBGEMM_SMALL_K_B0_TT ../generic/gemm_small_matrix_kernel_tt.c)
+          set(SBGEMM_SMALL_K_B0_TT ../generic/gemm_small_matrix_kernel_tt.c)
       endif ()
 	    GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_M_PERMIT}" "" "gemm_small_matrix_permit" false "" "" false "BFLOAT16")
             GenerateNamedObjects("${KERNELDIR}/${SBGEMM_SMALL_K_NN}" "" "gemm_small_kernel_nn" false "" "" false "BFLOAT16")

kernel/setparam-ref.c

@@ -1824,6 +1824,13 @@ static void init_parameter(void) {
   fprintf(stderr, "L2 = %8d DGEMM_P  .. %d\n", l2, TABLE_NAME.dgemm_p);
 #endif
 
+#if BUILD_BFLOAT16==1
+  TABLE_NAME.sbgemm_r = (((BUFFER_SIZE -
+			       ((TABLE_NAME.sbgemm_p * TABLE_NAME.sbgemm_q *  4 + TABLE_NAME.offsetA
+				 + TABLE_NAME.align) & ~TABLE_NAME.align)
+			       ) / (TABLE_NAME.sbgemm_q *  4) - 15) & ~15);
+#endif
+
 #if BUILD_SINGLE==1
   TABLE_NAME.sgemm_r = (((BUFFER_SIZE -
 			       ((TABLE_NAME.sgemm_p * TABLE_NAME.sgemm_q *  4 + TABLE_NAME.offsetA

kernel/x86_64/sbgemm_microk_cooperlake_template.c

@@ -13,6 +13,8 @@
 #define ONE  1.e0f
 #define ZERO  0.e0f
 
+#define SHUFFLE_MAGIC_NO (const int) 0x39
+
 #undef STORE16_COMPLETE_RESULT
 #undef STORE16_MASK_COMPLETE_RESULT
 #undef SBGEMM_BLOCK_KERNEL_NN_32x8xK
@@ -356,7 +358,6 @@ void sbgemm_block_kernel_nn_32xNx32_one(BLASLONG m, BLASLONG n, BLASLONG k, floa
     bfloat16 * B_addr = B;
     float    * C_addr = C;
 
-    int  SHUFFLE_MAGIC_NO = 0x39;
     BLASLONG tag_k_32x = k & (~31);
 
 #ifndef ONE_ALPHA
@@ -465,7 +466,6 @@ void sbgemm_block_kernel_nn_16xNx32_one(BLASLONG m, BLASLONG n, BLASLONG k, floa
     bfloat16 * B_addr = B;
     float    * C_addr = C;
 
-    int  SHUFFLE_MAGIC_NO = 0x39;
     BLASLONG tag_k_32x = k & (~31);
 
 #ifndef ONE_ALPHA
@@ -1192,7 +1192,6 @@ void sbgemm_block_kernel_tn_32xNx32_one(BLASLONG m, BLASLONG n, BLASLONG k, floa
     bfloat16 * B_addr = B;
     float    * C_addr = C;
 
-    int  SHUFFLE_MAGIC_NO = 0x39;
     BLASLONG tag_k_32x = k & (~31);
 
 #ifndef ONE_ALPHA
@@ -1291,7 +1290,6 @@ void sbgemm_block_kernel_tn_16xNx32_one(BLASLONG m, BLASLONG n, BLASLONG k, floa
     bfloat16 * B_addr = B;
     float    * C_addr = C;
 
-    int  SHUFFLE_MAGIC_NO = 0x39;
     BLASLONG tag_k_32x = k & (~31);
 
 #ifndef ONE_ALPHA