Added support for logging gemm input values.

Added BLIS specific extension to AOCL DTL, in this
added support to print the input matrix sizes from BLIS
library.

AMD Internal: [CPUPL-806]

Change-Id: I80ed779d65f9b1c48466137fc2f05629fa2fb561

aocl_dtl/aocldtl.c

@@ -229,19 +229,13 @@ void DTL_Trace(
 		if ((ui8LogLevel >= AOCL_DTL_LEVEL_TRACE_1) &&
 			(ui8LogLevel <= AOCL_DTL_LEVEL_TRACE_8))
 		{
-			fprintf(pOutFile, "%d ", (ui8LogLevel - AOCL_DTL_LEVEL_TRACE_1)+1);
 			/* this loop is for formating the output log file */
 			for (i = 0; i < (ui8LogLevel - AOCL_DTL_LEVEL_TRACE_1); i++)
 			{
 				/* print tabs in the output file */
 				fprintf(pOutFile, "\t");
 			}
 		}
-		else
-		{
-			/* For non call traces we will just start the line with astrix */
-			fprintf(pOutFile, "* \t");
-		}
 
         switch (ui8LogType)
         {

aocl_dtl/aocldtl_blis.c

@@ -0,0 +1,46 @@
+/*===================================================================
+ * File Name :  aocldtl_blis.c
+ * 
+ * Description : BLIS library specific debug helpes.
+ *
+ * Copyright (C) 2020, Advanced Micro Devices, Inc
+ * 
+ *==================================================================*/
+
+
+#include "blis.h"
+
+#if AOCL_DTL_LOG_ENABLE
+void AOCL_DTL_log_gemm_sizes(int8 loglevel,
+							 obj_t* alpha,
+							 obj_t* a,
+							 obj_t* b,
+							 obj_t* beta,
+							 obj_t* c,
+							 const char* filename,
+							 const char* function_name,
+							 int line)
+{
+	char buffer[256];
+	gint_t m = bli_obj_length( c );
+	gint_t n = bli_obj_width( c );
+	gint_t k = bli_obj_length( b );
+	guint_t csa = bli_obj_col_stride( a );
+	guint_t csb = bli_obj_col_stride( b );
+	guint_t csc = bli_obj_col_stride( c );
+	guint_t rsa = bli_obj_row_stride( a );
+	guint_t rsb = bli_obj_row_stride( b );
+	guint_t rsc = bli_obj_row_stride( c );
+	const num_t dt_exec = bli_obj_dt( c );
+	float* alpha_cast = bli_obj_buffer_for_1x1( dt_exec, alpha );
+	float* beta_cast  = bli_obj_buffer_for_1x1( dt_exec, beta );
+
+	sprintf(buffer, "%ld %ld %ld %lu %lu %lu %lu %lu %lu %f %f",
+				 m, k, n,
+				 csa, csb, csc,
+				 rsa, rsb, rsc,
+				 *alpha_cast, *beta_cast);
+
+	DTL_Trace(loglevel, TRACE_TYPE_LOG, function_name, function_name, line, buffer);
+}
+#endif

aocl_dtl/aocldtl_blis.h

@@ -0,0 +1,34 @@
+/*===================================================================
+ * File Name :  aocldtl_blis.h
+ * 
+ * Description : BLIS library specific debug helpes.
+ *
+ * Copyright (C) 2020, Advanced Micro Devices, Inc
+ * 
+ *==================================================================*/
+
+
+#ifndef __AOCLDTL_BLIS_H
+#define __AOCLDTL_BLIS_H
+
+#include "blis.h"
+
+#if AOCL_DTL_LOG_ENABLE
+void AOCL_DTL_log_gemm_sizes(int8 loglevel,
+							 obj_t* alpha,
+							 obj_t* a,
+							 obj_t* b,
+							 obj_t* beta,
+							 obj_t* c,
+							 const char* filename,
+							 const char* functionn_name,
+							 int line);
+
+#define AOCL_DTL_LOG_GEMM_INPUTS(loglevel, alpha, a, b, beta, c)	\
+    AOCL_DTL_log_gemm_sizes(loglevel, alpha, a, b, beta, c, __FILE__, __FUNCTION__, __LINE__);
+#else
+#define AOCL_DTL_LOG_GEMM_INPUTS(loglevel, alpha, a, b, beta, c)
+#endif
+
+#endif
+

frame/3/bli_l3_oapi.c

@@ -58,6 +58,7 @@ void PASTEMAC(opname,EX_SUF) \
 	bli_init_once(); \
 \
 	BLIS_OAPI_EX_DECLS \
+	AOCL_DTL_LOG_GEMM_INPUTS(AOCL_DTL_LEVEL_TRACE_2, alpha, a, b, beta, c); \
 \
 	/* If C has a zero dimension, return early.	*/	\
 	if ( bli_obj_has_zero_dim( c ) ) {\

frame/3/bli_l3_sup.c

@@ -45,22 +45,31 @@ err_t bli_gemmsup
        rntm_t* rntm
      )
 {
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_2);
+	AOCL_DTL_LOG_GEMM_INPUTS(AOCL_DTL_LEVEL_TRACE_2, alpha, a, b, beta, c);
+
 	// Return early if small matrix handling is disabled at configure-time.
 	#ifdef BLIS_DISABLE_SUP_HANDLING
+	AOCL_DTL_TRACE_EXIT_ERR(AOCL_DTL_LEVEL_TRACE_2, "SUP is Disabled.");
 	return BLIS_FAILURE;
 	#endif
 
 	// Return early if this is a mixed-datatype computation.
 	if ( bli_obj_dt( c ) != bli_obj_dt( a ) ||
 	     bli_obj_dt( c ) != bli_obj_dt( b ) ||
-	     bli_obj_comp_prec( c ) != bli_obj_prec( c ) ) return BLIS_FAILURE;
+	     bli_obj_comp_prec( c ) != bli_obj_prec( c ) ) {
+		AOCL_DTL_TRACE_EXIT_ERR(AOCL_DTL_LEVEL_TRACE_2, "SUP doesn't support Mixed datatypes.");
+		return BLIS_FAILURE;
+	}
 
 
 	const stor3_t stor_id = bli_obj_stor3_from_strides( c, a, b );
 
 	/*General stride is not yet supported in sup*/
-	if(BLIS_XXX==stor_id)
+	if(BLIS_XXX==stor_id) {
+		AOCL_DTL_TRACE_EXIT_ERR(AOCL_DTL_LEVEL_TRACE_2, "SUP doesn't support general stride.");
 		return BLIS_FAILURE;
+	}
 
 	const dim_t m  = bli_obj_length( c );
 	const dim_t n  = bli_obj_width( c );
@@ -74,6 +83,7 @@ err_t bli_gemmsup
 	|| ((transb == BLIS_CONJ_NO_TRANSPOSE) || (transb == BLIS_CONJ_TRANSPOSE))
 	)){
 		//printf(" gemmsup: Returning with for un-supported storage types and conjugate property in cgemmsup \n");
+		AOCL_DTL_TRACE_EXIT_ERR(AOCL_DTL_LEVEL_TRACE_2, "SUP - Unsuppported storage type for cgemm");
 		return BLIS_FAILURE;
 	}
 
@@ -84,6 +94,7 @@ err_t bli_gemmsup
 	|| ((transb == BLIS_CONJ_NO_TRANSPOSE) || (transb == BLIS_CONJ_TRANSPOSE))
 	)){
 		//printf(" gemmsup: Returning with for un-supported storage types and conjugate property in zgemmsup \n");
+		AOCL_DTL_TRACE_EXIT_ERR(AOCL_DTL_LEVEL_TRACE_2, "SUP - Unsuppported storage type for zgemm.");
 		return BLIS_FAILURE;
 	}
 
@@ -103,16 +114,20 @@ err_t bli_gemmsup
 
 		// Pass in m and n reversed, which simulates a transposition of the
 		// entire operation pursuant to the microkernel storage preference.
-		if ( !bli_cntx_l3_sup_thresh_is_met( dt, n, m, k, cntx ) )
+		if ( !bli_cntx_l3_sup_thresh_is_met( dt, n, m, k, cntx ) ) {
+			AOCL_DTL_TRACE_EXIT_ERR(AOCL_DTL_LEVEL_TRACE_2, "SUP - Traspostion results in unsupported storage for matrix C.");
 			return BLIS_FAILURE;
+		}
 	}
 	else // ukr_prefers_storage_of( c, ... )
 	{
 		const num_t dt = bli_obj_dt( c );
 		const dim_t k  = bli_obj_width_after_trans( a );
 
-		if ( !bli_cntx_l3_sup_thresh_is_met( dt, m, n, k, cntx ) )
+		if ( !bli_cntx_l3_sup_thresh_is_met( dt, m, n, k, cntx ) ) {
+			AOCL_DTL_TRACE_EXIT_ERR(AOCL_DTL_LEVEL_TRACE_2, "SUP - Unsupported storage for matrix C.");
 			return BLIS_FAILURE;
+		}
 	}
 
 	// Initialize a local runtime with global settings if necessary. Note
@@ -158,6 +173,8 @@ printf( "dims: %d %d %d (threshs: %d %d %d)\n",
 	  cntx,
 	  rntm
 	);
+
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_2);
 }
 
 

frame/3/bli_l3_sup_int.c

@@ -46,6 +46,9 @@ err_t bli_gemmsup_int
        thrinfo_t* thread
      )
 {
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_4);
+	AOCL_DTL_LOG_GEMM_INPUTS(AOCL_DTL_LEVEL_TRACE_4, alpha, a, b, beta, c);
+
 #if 0
 	//bli_gemmsup_ref_var2
 	//bli_gemmsup_ref_var1
@@ -80,7 +83,10 @@ err_t bli_gemmsup_int
 
 	// Don't use the small/unpacked implementation if one of the matrices
 	// uses general stride.
-	if ( stor_id == BLIS_XXX ) return BLIS_FAILURE;
+	if ( stor_id == BLIS_XXX ) {
+		AOCL_DTL_TRACE_EXIT_ERR(AOCL_DTL_LEVEL_TRACE_4, "SUP doesn't support general stide.");
+		return BLIS_FAILURE;
+	}
 
 	const bool_t  is_rrr_rrc_rcr_crr = ( stor_id == BLIS_RRR ||
 	                                     stor_id == BLIS_RRC ||
@@ -237,6 +243,7 @@ err_t bli_gemmsup_int
 	}
 
 	// Return success so that the caller knows that we computed the solution.
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_4)
 	return BLIS_SUCCESS;
 }
 

frame/3/bli_l3_sup_packm_a.c

@@ -51,7 +51,8 @@ void PASTEMAC(ch,opname) \
        thrinfo_t* restrict thread  \
      ) \
 { \
-	/* Inspect whether we are going to be packing matrix A. */ \
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5); \
+	/* Inspect whether we are going to be packing matrix A. */	\
 	if ( will_pack == FALSE ) \
 	{ \
 	} \
@@ -164,6 +165,7 @@ void PASTEMAC(ch,opname) \
 			} \
 		} \
 	} \
+    AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5);\
 }
 
 INSERT_GENTFUNC_BASIC0( packm_sup_init_mem_a )
@@ -180,6 +182,7 @@ void PASTEMAC(ch,opname) \
        thrinfo_t* restrict thread  \
      ) \
 { \
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5);\
 	/* Inspect whether we previously packed matrix A. */ \
 	if ( did_pack == FALSE ) \
 	{ \
@@ -202,6 +205,7 @@ void PASTEMAC(ch,opname) \
 			} \
 		} \
 	} \
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5);\
 }
 
 INSERT_GENTFUNC_BASIC0( packm_sup_finalize_mem_a )
@@ -228,6 +232,7 @@ void PASTEMAC(ch,opname) \
        thrinfo_t* restrict thread  \
      ) \
 { \
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5);\
 	/* Inspect whether we are going to be packing matrix A. */ \
 	if ( will_pack == FALSE ) \
 	{ \
@@ -297,6 +302,7 @@ void PASTEMAC(ch,opname) \
 		   broker. */ \
 		*p = bli_mem_buffer( mem ); \
 	} \
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5);\
 }
 
 INSERT_GENTFUNC_BASIC0( packm_sup_init_a )
@@ -335,6 +341,7 @@ void PASTEMAC(ch,opname) \
 	dim_t  k_max; \
 	dim_t  pd_p; \
 \
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5);\
 	/* Prepare the packing destination buffer. If packing is not requested,
 	   this function will reduce to a no-op. */ \
 	PASTEMAC(ch,packm_sup_init_mem_a) \
@@ -424,6 +431,7 @@ void PASTEMAC(ch,opname) \
 		/* Barrier so that packing is done before computation. */ \
 		bli_thread_barrier( thread ); \
 	} \
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5);\
 }
 
 INSERT_GENTFUNC_BASIC0( packm_sup_a )

frame/3/bli_l3_sup_packm_b.c

@@ -51,6 +51,7 @@ void PASTEMAC(ch,opname) \
        thrinfo_t* restrict thread  \
      ) \
 { \
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5);\
 	/* Inspect whether we are going to be packing matrix B. */ \
 	if ( will_pack == FALSE ) \
 	{ \
@@ -164,6 +165,7 @@ void PASTEMAC(ch,opname) \
 			} \
 		} \
 	} \
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5);\
 }
 
 INSERT_GENTFUNC_BASIC0( packm_sup_init_mem_b )
@@ -180,6 +182,7 @@ void PASTEMAC(ch,opname) \
        thrinfo_t* restrict thread  \
      ) \
 { \
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5); \
 	/* Inspect whether we previously packed matrix A. */ \
 	if ( did_pack == FALSE ) \
 	{ \
@@ -202,6 +205,7 @@ void PASTEMAC(ch,opname) \
 			} \
 		} \
 	} \
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5); \
 }
 
 INSERT_GENTFUNC_BASIC0( packm_sup_finalize_mem_b )
@@ -228,6 +232,7 @@ void PASTEMAC(ch,opname) \
        thrinfo_t* restrict thread  \
      ) \
 { \
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5);\
 	/* Inspect whether we are going to be packing matrix B. */ \
 	if ( will_pack == FALSE ) \
 	{ \
@@ -297,6 +302,7 @@ void PASTEMAC(ch,opname) \
 		   broker. */ \
 		*p = bli_mem_buffer( mem ); \
 	} \
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5);\
 }
 
 INSERT_GENTFUNC_BASIC0( packm_sup_init_b )
@@ -330,6 +336,7 @@ void PASTEMAC(ch,opname) \
        thrinfo_t* restrict thread  \
      ) \
 { \
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5); \
 	pack_t schema; \
 	dim_t  k_max; \
 	dim_t  n_max; \
@@ -424,6 +431,7 @@ void PASTEMAC(ch,opname) \
 		/* Barrier so that packing is done before computation. */ \
 		bli_thread_barrier( thread ); \
 	} \
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5); \
 }
 
 INSERT_GENTFUNC_BASIC0( packm_sup_b )

frame/3/bli_l3_sup_var1n2m.c

@@ -76,6 +76,7 @@ void bli_gemmsup_ref_var1n
        thrinfo_t* thread
      )
 {
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5);
 #if 0
 	obj_t at, bt;
 
@@ -227,6 +228,7 @@ void bli_gemmsup_ref_var1n
 		  thread
 		);
 	}
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5);
 }
 
 
@@ -724,6 +726,7 @@ void bli_gemmsup_ref_var2m
        thrinfo_t* thread
      )
 {
+	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_5);
 #if 0
 	obj_t at, bt;
 
@@ -875,6 +878,7 @@ void bli_gemmsup_ref_var2m
 		  thread
 		);
 	}
+	AOCL_DTL_TRACE_EXIT(AOCL_DTL_LEVEL_TRACE_5);
 }
 
 

frame/3/gemm/bli_gemm_front.c

@@ -48,6 +48,7 @@ void bli_gemm_front
      )
 {
 	AOCL_DTL_TRACE_ENTRY(AOCL_DTL_LEVEL_TRACE_3);
+	AOCL_DTL_LOG_GEMM_INPUTS(AOCL_DTL_LEVEL_TRACE_3, alpha, a, b, beta, c);
 	bli_init_once();
 
 	obj_t   a_local;