[primitives,codec] guard SSE code with platform

SSE optimized code might be used in multiarch/universal builds. So not only guard with WITH_SSE2 but also with architecture defines from winpr/platform.h
2024-06-25 09:40:09 +02:00 · 2024-06-25 09:40:09 +02:00 · 896ea3c445
commit 896ea3c445
parent 55258cd9eb
15 changed files with 87 additions and 78 deletions
--- a/libfreerdp/codec/sse/nsc_sse2.c
+++ b/libfreerdp/codec/sse/nsc_sse2.c
@ -17,6 +17,7 @@
 * limitations under the License.
 */

+#include <winpr/platform.h>
 #include <freerdp/config.h>
 #include <freerdp/log.h>

@ -26,6 +27,12 @@
 #define TAG FREERDP_TAG("codec.nsc.sse2")

 #if defined(WITH_SSE2)
+#if defined(_M_IX86) || defined(_M_AMD64) || defined(_M_IA64) || defined(_M_IX86_AMD64)
+#define SSE2_ENABLED
+#endif
+#endif
+
+#if defined(SSE2_ENABLED)
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@ -381,7 +388,7 @@ static BOOL nsc_encode_sse2(NSC_CONTEXT* context, const BYTE* data, UINT32 scanl

 void nsc_init_sse2(NSC_CONTEXT* context)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	if (!IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
 		return;

--- a/libfreerdp/codec/sse/rfx_sse2.c
+++ b/libfreerdp/codec/sse/rfx_sse2.c
@ -18,6 +18,7 @@
 * limitations under the License.
 */

+#include <winpr/platform.h>
 #include <freerdp/config.h>
 #include <freerdp/log.h>

@ -27,6 +28,12 @@
 #define TAG FREERDP_TAG("codec.rfx.sse2")

 #if defined(WITH_SSE2)
+#if defined(_M_IX86) || defined(_M_AMD64) || defined(_M_IA64) || defined(_M_IX86_AMD64)
+#define SSE2_ENABLED
+#endif
+#endif
+
+#if defined(SSE2_ENABLED)
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@ -485,7 +492,7 @@ static void rfx_dwt_2d_encode_sse2(INT16* WINPR_RESTRICT buffer, INT16* WINPR_RE

 void rfx_init_sse2(RFX_CONTEXT* context)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	if (!IsProcessorFeaturePresent(PF_XMMI64_INSTRUCTIONS_AVAILABLE))
 		return;

--- a/libfreerdp/primitives/prim_internal.h
+++ b/libfreerdp/primitives/prim_internal.h
@ -17,6 +17,7 @@
 #ifndef FREERDP_LIB_PRIM_INTERNAL_H
 #define FREERDP_LIB_PRIM_INTERNAL_H

+#include <winpr/platform.h>
 #include <freerdp/config.h>

 #include <freerdp/primitives.h>
@ -33,15 +34,21 @@
 #endif
 #endif

-#if defined(WITH_SSE2) || defined(WITH_NEON) || defined(WITH_OPENCL)
+#if defined(WITH_SSE2)
+#if defined(_M_IX86) || defined(_M_AMD64) || defined(_M_IA64) || defined(_M_IX86_AMD64)
+#define SSE2_ENABLED
+#endif
+#endif
+
+#if defined(SSE2_ENABLED) || defined(WITH_NEON) || defined(WITH_OPENCL)
 #define HAVE_OPTIMIZED_PRIMITIVES 1
 #endif

-#if defined(WITH_SSE2) || defined(WITH_NEON)
+#if defined(SSE2_ENABLED) || defined(WITH_NEON)
 #define HAVE_CPU_OPTIMIZED_PRIMITIVES 1
 #endif

-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 /* Use lddqu for unaligned; load for 16-byte aligned. */
 #define LOAD_SI128(_ptr_)                                                       \
 	(((const ULONG_PTR)(_ptr_)&0x0f) ? _mm_lddqu_si128((const __m128i*)(_ptr_)) \
--- a/libfreerdp/primitives/sse/prim_YCoCg_ssse3.c
+++ b/libfreerdp/primitives/sse/prim_YCoCg_ssse3.c
@ -25,15 +25,13 @@

 #include "prim_YCoCg.h"

-#ifdef WITH_SSE2
-#include <emmintrin.h>
-#include <tmmintrin.h>
-#endif
-
 #include "prim_internal.h"
 #include "prim_templates.h"

-#ifdef WITH_SSE2
+#if defined(SSE2_ENABLED)
+#include <emmintrin.h>
+#include <tmmintrin.h>
+
 static primitives_t* generic = NULL;

 /* ------------------------------------------------------------------------- */
@ -436,12 +434,12 @@ static pstatus_t ssse3_YCoCgRToRGB_8u_AC4R(const BYTE* WINPR_RESTRICT pSrc, INT3
 	}
 }

-#endif /* WITH_SSE2 */
+#endif

 /* ------------------------------------------------------------------------- */
 void primitives_init_YCoCg_ssse3(primitives_t* WINPR_RESTRICT prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	generic = primitives_get_generic();
 	primitives_init_YCoCg(prims);

--- a/libfreerdp/primitives/sse/prim_YUV_ssse3.c
+++ b/libfreerdp/primitives/sse/prim_YUV_ssse3.c
@ -31,7 +31,7 @@
 #include "prim_internal.h"
 #include "prim_YUV.h"

-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 #include <emmintrin.h>
 #include <tmmintrin.h>

@ -1498,7 +1498,7 @@ static pstatus_t ssse3_YUV420CombineToYUV444(avc444_frame_type type,

 void primitives_init_YUV_ssse3(primitives_t* WINPR_RESTRICT prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	generic = primitives_get_generic();
 	primitives_init_YUV(prims);

--- a/libfreerdp/primitives/sse/prim_add_sse3.c
+++ b/libfreerdp/primitives/sse/prim_add_sse3.c
@ -22,15 +22,13 @@

 #include "prim_add.h"

-#ifdef WITH_SSE2
-#include <emmintrin.h>
-#include <pmmintrin.h>
-#endif /* WITH_SSE2 */
-
 #include "prim_internal.h"
 #include "prim_templates.h"

-#ifdef WITH_SSE2
+#if defined(SSE2_ENABLED)
+#include <emmintrin.h>
+#include <pmmintrin.h>
+
 static primitives_t* generic = NULL;

 /* ------------------------------------------------------------------------- */
@ -172,13 +170,12 @@ static pstatus_t sse3_add_16s_inplace(INT16* WINPR_RESTRICT pSrcDst1,

 	return PRIMITIVES_SUCCESS;
 }
-
 #endif

 /* ------------------------------------------------------------------------- */
 void primitives_init_add_sse3(primitives_t* WINPR_RESTRICT prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	generic = primitives_get_generic();
 	primitives_init_add(prims);

--- a/libfreerdp/primitives/sse/prim_alphaComp_sse3.c
+++ b/libfreerdp/primitives/sse/prim_alphaComp_sse3.c
@ -28,15 +28,13 @@

 #include "prim_alphaComp.h"

-#ifdef WITH_SSE2
-#include <emmintrin.h>
-#include <pmmintrin.h>
-#endif /* WITH_SSE2 */
-
 #include "prim_internal.h"

 /* ------------------------------------------------------------------------- */
-#ifdef WITH_SSE2
+#if defined(SSE2_ENABLED)
+#include <emmintrin.h>
+#include <pmmintrin.h>
+
 static primitives_t* generic = NULL;

 static pstatus_t sse2_alphaComp_argb(const BYTE* WINPR_RESTRICT pSrc1, UINT32 src1Step,
@ -211,7 +209,7 @@ static pstatus_t sse2_alphaComp_argb(const BYTE* WINPR_RESTRICT pSrc1, UINT32 sr
 /* ------------------------------------------------------------------------- */
 void primitives_init_alphaComp_sse3(primitives_t* WINPR_RESTRICT prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	generic = primitives_get_generic();
 	primitives_init_alphaComp(prims);

--- a/libfreerdp/primitives/sse/prim_andor_sse3.c
+++ b/libfreerdp/primitives/sse/prim_andor_sse3.c
@ -21,27 +21,26 @@

 #include "prim_andor.h"

-#ifdef WITH_SSE2
-#include <emmintrin.h>
-#include <pmmintrin.h>
-#endif /* WITH_SSE2 */
-
 #include "prim_internal.h"
 #include "prim_templates.h"

-#ifdef WITH_SSE2
+#if defined(SSE2_ENABLED)
+#include <emmintrin.h>
+#include <pmmintrin.h>
+
 static primitives_t* generic = NULL;

 /* ------------------------------------------------------------------------- */
 SSE3_SCD_PRE_ROUTINE(sse3_andC_32u, UINT32, generic->andC_32u, _mm_and_si128,
                     *dptr++ = *sptr++ & val)
 SSE3_SCD_PRE_ROUTINE(sse3_orC_32u, UINT32, generic->orC_32u, _mm_or_si128, *dptr++ = *sptr++ | val)
+
 #endif

 /* ------------------------------------------------------------------------- */
 void primitives_init_andor_sse3(primitives_t* WINPR_RESTRICT prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	generic = primitives_get_generic();
 	primitives_init_andor(prims);

--- a/libfreerdp/primitives/sse/prim_colors_sse2.c
+++ b/libfreerdp/primitives/sse/prim_colors_sse2.c
@ -25,14 +25,12 @@

 #include "prim_colors.h"

-#ifdef WITH_SSE2
-#include <emmintrin.h>
-#endif /* WITH_SSE2 */
-
 #include "prim_internal.h"
 #include "prim_templates.h"

-#ifdef WITH_SSE2
+#if defined(SSE2_ENABLED)
+#include <emmintrin.h>
+
 static primitives_t* generic = NULL;

 #ifdef __GNUC__
@ -1243,11 +1241,11 @@ sse2_RGBToRGB_16s8u_P3AC4R(const INT16* const WINPR_RESTRICT pSrc[3], /* 16-bit
 			return generic->RGBToRGB_16s8u_P3AC4R(pSrc, srcStep, pDst, dstStep, DstFormat, roi);
 	}
 }
-#endif /* WITH_SSE2 */
+#endif

 void primitives_init_colors_sse2(primitives_t* prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	generic = primitives_get_generic();
 	primitives_init_colors(prims);

--- a/libfreerdp/primitives/sse/prim_copy_avx2.c
+++ b/libfreerdp/primitives/sse/prim_copy_avx2.c
@ -30,7 +30,7 @@

 #define TAG FREERDP_TAG("primitives.copy")

-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 #include <emmintrin.h>
 #include <immintrin.h>

@ -265,13 +265,12 @@ static pstatus_t avx2_image_copy_no_overlap(BYTE* WINPR_RESTRICT pDstData, DWORD
 		                                          nXSrc, nYSrc, palette, srcVMultiplier, srcVOffset,
 		                                          dstVMultiplier, dstVOffset);
 }
-
 #endif

 /* ------------------------------------------------------------------------- */
 void primitives_init_copy_avx2(primitives_t* prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	if (IsProcessorFeaturePresent(PF_AVX2_INSTRUCTIONS_AVAILABLE))
 	{
 		WLog_VRB(PRIM_TAG, "AVX2 optimizations");
--- a/libfreerdp/primitives/sse/prim_copy_sse4_1.c
+++ b/libfreerdp/primitives/sse/prim_copy_sse4_1.c
@ -30,7 +30,7 @@

 #define TAG FREERDP_TAG("primitives.copy")

-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 #include <emmintrin.h>
 #include <immintrin.h>

@ -264,13 +264,12 @@ static pstatus_t sse_image_copy_no_overlap(BYTE* WINPR_RESTRICT pDstData, DWORD
 		                                         nXSrc, nYSrc, palette, srcVMultiplier, srcVOffset,
 		                                         dstVMultiplier, dstVOffset);
 }
-
 #endif

 /* ------------------------------------------------------------------------- */
 void primitives_init_copy_sse41(primitives_t* prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	if (IsProcessorFeaturePresent(PF_SSE4_1_INSTRUCTIONS_AVAILABLE))
 	{
 		WLog_VRB(PRIM_TAG, "SSE4.1 optimizations");
--- a/libfreerdp/primitives/sse/prim_set_sse2.c
+++ b/libfreerdp/primitives/sse/prim_set_sse2.c
@ -21,15 +21,13 @@
 #include <freerdp/primitives.h>
 #include <winpr/sysinfo.h>

-#ifdef WITH_SSE2
-#include <emmintrin.h>
-#endif /* WITH_SSE2 */
-
 #include "prim_internal.h"
 #include "prim_set.h"

 /* ========================================================================= */
-#ifdef WITH_SSE2
+#if defined(SSE2_ENABLED)
+#include <emmintrin.h>
+
 static primitives_t* generic = NULL;

 static pstatus_t sse2_set_8u(BYTE val, BYTE* WINPR_RESTRICT pDst, UINT32 len)
@ -113,10 +111,8 @@ static pstatus_t sse2_set_8u(BYTE val, BYTE* WINPR_RESTRICT pDst, UINT32 len)

 	return PRIMITIVES_SUCCESS;
 }
-#endif /* WITH_SSE2 */

 /* ------------------------------------------------------------------------- */
-#ifdef WITH_SSE2
 static pstatus_t sse2_set_32u(UINT32 val, UINT32* WINPR_RESTRICT pDst, UINT32 len)
 {
 	const primitives_t* prim = primitives_get_generic();
@ -214,12 +210,12 @@ static pstatus_t sse2_set_32s(INT32 val, INT32* WINPR_RESTRICT pDst, UINT32 len)
 	UINT32 uval = *((UINT32*)&val);
 	return sse2_set_32u(uval, (UINT32*)pDst, len);
 }
-#endif /* WITH_SSE2 */
+#endif

 /* ------------------------------------------------------------------------- */
 void primitives_init_set_sse2(primitives_t* WINPR_RESTRICT prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	generic = primitives_get_generic();
 	primitives_init_set(prims);
 	/* Pick tuned versions if possible. */
--- a/libfreerdp/primitives/sse/prim_shift_sse3.c
+++ b/libfreerdp/primitives/sse/prim_shift_sse3.c
@ -21,15 +21,13 @@

 #include "prim_shift.h"

-#ifdef WITH_SSE2
-#include <emmintrin.h>
-#include <pmmintrin.h>
-#endif /* WITH_SSE2 */
-
 #include "prim_internal.h"
 #include "prim_templates.h"

-#ifdef WITH_SSE2
+#if defined(SSE2_ENABLED)
+#include <emmintrin.h>
+#include <pmmintrin.h>
+
 static primitives_t* generic = NULL;

 /* ------------------------------------------------------------------------- */
@ -135,7 +133,6 @@ static pstatus_t sse2_lShiftC_16s_inplace(INT16* WINPR_RESTRICT pSrcDst, UINT32

 	return PRIMITIVES_SUCCESS;
 }
-
 #endif

 /* Note: the IPP version will have to call ippLShiftC_16s or ippRShiftC_16s
@ -146,7 +143,7 @@ static pstatus_t sse2_lShiftC_16s_inplace(INT16* WINPR_RESTRICT pSrcDst, UINT32
 /* ------------------------------------------------------------------------- */
 void primitives_init_shift_sse3(primitives_t* WINPR_RESTRICT prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	generic = primitives_get_generic();
 	primitives_init_shift(prims);

--- a/libfreerdp/primitives/sse/prim_sign_ssse3.c
+++ b/libfreerdp/primitives/sse/prim_sign_ssse3.c
@ -21,14 +21,12 @@

 #include "prim_sign.h"

-#if defined(WITH_SSE2)
-#include <emmintrin.h>
-#include <tmmintrin.h>
-#endif
-
 #include "prim_internal.h"

-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
+#include <emmintrin.h>
+#include <tmmintrin.h>
+
 static primitives_t* generic = NULL;

 /* ------------------------------------------------------------------------- */
@ -166,12 +164,13 @@ static pstatus_t ssse3_sign_16s(const INT16* WINPR_RESTRICT pSrc, INT16* WINPR_R

 	return PRIMITIVES_SUCCESS;
 }
-#endif /* WITH_SSE2 */
+
+#endif /* SSE2_ENABLED */

 /* ------------------------------------------------------------------------- */
 void primitives_init_sign_ssse3(primitives_t* WINPR_RESTRICT prims)
 {
-#if defined(WITH_SSE2)
+#if defined(SSE2_ENABLED)
 	generic = primitives_get_generic();
 	primitives_init_sign(prims);
 	/* Pick tuned versions if possible. */
--- a/winpr/include/winpr/platform.h
+++ b/winpr/include/winpr/platform.h
@ -94,9 +94,10 @@ WINPR_PRAGMA_DIAG_IGNORED_RESERVED_ID_MACRO

 /* Intel x86 (_M_IX86) */

-#if defined(i386) || defined(__i386) || defined(__i386__) || defined(__i486__) ||   \
-    defined(__i586__) || defined(__i686__) || defined(__X86__) || defined(_X86_) || \
-    defined(__I86__) || defined(__IA32__) || defined(__THW_INTEL__) || defined(__INTEL__)
+#if defined(i386) || defined(__i386) || defined(__i386__) || defined(__i486__) ||            \
+    defined(__i586__) || defined(__i686__) || defined(__X86__) || defined(_X86_) ||          \
+    defined(__I86__) || defined(__IA32__) || defined(__THW_INTEL__) || defined(__INTEL__) || \
+    defined(_M_IX86)
 #ifndef _M_IX86
 #define _M_IX86 1
 #endif
@ -111,6 +112,13 @@ WINPR_PRAGMA_DIAG_IGNORED_RESERVED_ID_MACRO
 #endif
 #endif

+/* Intel ia64 */
+#if defined(__ia64) || defined(__ia64__) || defined(_M_IA64)
+#ifndef _M_IA64
+#define _M_IA64 1
+#endif
+#endif
+
 /* Intel x86 or AMD64 (_M_IX86_AMD64) */

 #if defined(_M_IX86) || defined(_M_AMD64)