@@ -314,8 +314,8 @@ typedef struct __tile1024i_str {
314314/// \param stride
315315/// The stride between the rows' data to be loaded in memory.
316316__DEFAULT_FN_ATTRS_TILE
317- static void __tile_loadd (__tile1024i * dst , const void * base ,
318- __SIZE_TYPE__ stride ) {
317+ static __inline__ void __tile_loadd (__tile1024i * dst , const void * base ,
318+ __SIZE_TYPE__ stride ) {
319319 dst -> tile = _tile_loadd_internal (dst -> row , dst -> col , base , stride );
320320}
321321
@@ -335,8 +335,8 @@ static void __tile_loadd(__tile1024i *dst, const void *base,
335335/// \param stride
336336/// The stride between the rows' data to be loaded in memory.
337337__DEFAULT_FN_ATTRS_TILE
338- static void __tile_stream_loadd (__tile1024i * dst , const void * base ,
339- __SIZE_TYPE__ stride ) {
338+ static __inline__ void __tile_stream_loadd (__tile1024i * dst , const void * base ,
339+ __SIZE_TYPE__ stride ) {
340340 dst -> tile = _tile_loaddt1_internal (dst -> row , dst -> col , base , stride );
341341}
342342
@@ -357,8 +357,8 @@ static void __tile_stream_loadd(__tile1024i *dst, const void *base,
357357/// \param src1
358358/// The 2nd source tile. Max size is 1024 Bytes.
359359__DEFAULT_FN_ATTRS_INT8
360- static void __tile_dpbssd (__tile1024i * dst , __tile1024i src0 ,
361- __tile1024i src1 ) {
360+ static __inline__ void __tile_dpbssd (__tile1024i * dst , __tile1024i src0 ,
361+ __tile1024i src1 ) {
362362 dst -> tile = _tile_dpbssd_internal (src0 .row , src1 .col , src0 .col , dst -> tile ,
363363 src0 .tile , src1 .tile );
364364}
@@ -380,8 +380,8 @@ static void __tile_dpbssd(__tile1024i *dst, __tile1024i src0,
380380/// \param src1
381381/// The 2nd source tile. Max size is 1024 Bytes.
382382__DEFAULT_FN_ATTRS_INT8
383- static void __tile_dpbsud (__tile1024i * dst , __tile1024i src0 ,
384- __tile1024i src1 ) {
383+ static __inline__ void __tile_dpbsud (__tile1024i * dst , __tile1024i src0 ,
384+ __tile1024i src1 ) {
385385 dst -> tile = _tile_dpbsud_internal (src0 .row , src1 .col , src0 .col , dst -> tile ,
386386 src0 .tile , src1 .tile );
387387}
@@ -403,8 +403,8 @@ static void __tile_dpbsud(__tile1024i *dst, __tile1024i src0,
403403/// \param src1
404404/// The 2nd source tile. Max size is 1024 Bytes.
405405__DEFAULT_FN_ATTRS_INT8
406- static void __tile_dpbusd (__tile1024i * dst , __tile1024i src0 ,
407- __tile1024i src1 ) {
406+ static __inline__ void __tile_dpbusd (__tile1024i * dst , __tile1024i src0 ,
407+ __tile1024i src1 ) {
408408 dst -> tile = _tile_dpbusd_internal (src0 .row , src1 .col , src0 .col , dst -> tile ,
409409 src0 .tile , src1 .tile );
410410}
@@ -426,8 +426,8 @@ static void __tile_dpbusd(__tile1024i *dst, __tile1024i src0,
426426/// \param src1
427427/// The 2nd source tile. Max size is 1024 Bytes.
428428__DEFAULT_FN_ATTRS_INT8
429- static void __tile_dpbuud (__tile1024i * dst , __tile1024i src0 ,
430- __tile1024i src1 ) {
429+ static __inline__ void __tile_dpbuud (__tile1024i * dst , __tile1024i src0 ,
430+ __tile1024i src1 ) {
431431 dst -> tile = _tile_dpbuud_internal (src0 .row , src1 .col , src0 .col , dst -> tile ,
432432 src0 .tile , src1 .tile );
433433}
@@ -446,7 +446,8 @@ static void __tile_dpbuud(__tile1024i *dst, __tile1024i src0,
446446/// \param stride
447447/// The stride between the rows' data to be stored in memory.
448448__DEFAULT_FN_ATTRS_TILE
449- static void __tile_stored (void * base , __SIZE_TYPE__ stride , __tile1024i src ) {
449+ static __inline__ void __tile_stored (void * base , __SIZE_TYPE__ stride ,
450+ __tile1024i src ) {
450451 _tile_stored_internal (src .row , src .col , base , stride , src .tile );
451452}
452453
@@ -459,7 +460,7 @@ static void __tile_stored(void *base, __SIZE_TYPE__ stride, __tile1024i src) {
459460/// \param dst
460461/// The destination tile to be zero. Max size is 1024 Bytes.
461462__DEFAULT_FN_ATTRS_TILE
462- static void __tile_zero (__tile1024i * dst ) {
463+ static __inline__ void __tile_zero (__tile1024i * dst ) {
463464 dst -> tile = __builtin_ia32_tilezero_internal (dst -> row , dst -> col );
464465}
465466
@@ -479,8 +480,8 @@ static void __tile_zero(__tile1024i *dst) {
479480/// \param src1
480481/// The 2nd source tile. Max size is 1024 Bytes.
481482__DEFAULT_FN_ATTRS_BF16
482- static void __tile_dpbf16ps (__tile1024i * dst , __tile1024i src0 ,
483- __tile1024i src1 ) {
483+ static __inline__ void __tile_dpbf16ps (__tile1024i * dst , __tile1024i src0 ,
484+ __tile1024i src1 ) {
484485 dst -> tile = _tile_dpbf16ps_internal (src0 .row , src1 .col , src0 .col , dst -> tile ,
485486 src0 .tile , src1 .tile );
486487}
0 commit comments