[Pixman] [PATCH 2/6] mmx: compile on MIPS for Loongson MMI optimizations
Matt Turner
mattst88 at gmail.com
Fri Feb 24 15:18:03 PST 2012
Signed-off-by: Matt Turner <mattst88 at gmail.com>
---
configure.ac | 45 +++++++++
pixman/Makefile.am | 12 +++
pixman/loongson-mmintrin.h | 218 ++++++++++++++++++++++++++++++++++++++++++++
pixman/pixman-cpu.c | 4 +-
pixman/pixman-mmx.c | 40 ++++++++-
pixman/pixman-private.h | 2 +-
6 files changed, 316 insertions(+), 5 deletions(-)
create mode 100644 pixman/loongson-mmintrin.h
diff --git a/configure.ac b/configure.ac
index cca4d16..c053496 100644
--- a/configure.ac
+++ b/configure.ac
@@ -271,6 +271,51 @@ PIXMAN_CHECK_CFLAG([-xldscope=hidden], [dnl
])
dnl ===========================================================================
+dnl Check for Loongson Multimedia Instructions
+
+have_loongson_mmi=no
+AC_MSG_CHECKING(whether to use Loongson MMI)
+
+AC_COMPILE_IFELSE([AC_LANG_SOURCE([[
+#ifndef __mips_loongson_vector_rev
+#error "Loongson Multimedia Instructions are only available on Loongson"
+#endif
+#if defined(__GNUC__) && (__GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ < 4))
+#error "Need GCC >= 4.4 for Loongson MMI compilation"
+#endif
+#include "pixman/loongson-mmintrin.h"
+int main () {
+ union {
+ __m64 v;
+ char c[8];
+ } a = { .c = {1, 2, 3, 4, 5, 6, 7, 8} };
+ int b = 4;
+ __m64 c = _mm_srli_pi16 (a.v, b);
+ return 0;
+}]])], have_loongson_mmi=yes)
+
+
+AC_ARG_ENABLE(loongson,
+ [AC_HELP_STRING([--disable-loongson],
+ [disable Loongson fast paths])],
+ [enable_loongson=$enableval], [enable_loongson=auto])
+
+if test $enable_loongson = no ; then
+ have_loongson_mmi=disabled
+fi
+
+if test $have_loongson_mmi = yes ; then
+ AC_DEFINE(USE_LOONGSON_MMI, 1, [use Loongson Multimedia Instructions])
+fi
+
+AC_MSG_RESULT($have_loongson_mmi)
+if test $enable_loongson = yes && test $have_loongson_mmi = no ; then
+ AC_MSG_ERROR([Loongson MMI not detected])
+fi
+
+AM_CONDITIONAL(USE_LOONGSON_MMI, test $have_loongson_mmi = yes)
+
+dnl ===========================================================================
dnl Check for MMX
if test "x$MMX_CFLAGS" = "x" ; then
diff --git a/pixman/Makefile.am b/pixman/Makefile.am
index 286b7cf..7fe3a80 100644
--- a/pixman/Makefile.am
+++ b/pixman/Makefile.am
@@ -102,5 +102,17 @@ libpixman_1_la_LIBADD += libpixman-iwmmxt.la
ASM_CFLAGS_IWMMXT=$(IWMMXT_CFLAGS)
endif
+# loongson code
+if USE_LOONGSON_MMI
+noinst_LTLIBRARIES += libpixman-loongson-simd.la
+libpixman_loongson_simd_la_SOURCES = pixman-mmx.c
+libpixman_loongson_simd_la_CFLAGS = $(DEP_CFLAGS) $(LS_CFLAGS)
+libpixman_loongson_simd_la_LIBADD = $(DEP_LIBS)
+libpixman_1_la_LDFLAGS += $(LS_LDFLAGS)
+libpixman_1_la_LIBADD += libpixman-loongson-simd.la
+
+ASM_CFLAGS_ls=$(LS_CFLAGS)
+endif
+
.c.s : $(libpixmaninclude_HEADERS) $(BUILT_SOURCES)
$(CC) $(CFLAGS) $(ASM_CFLAGS_$(@:pixman-%.s=%)) $(ASM_CFLAGS_$(@:pixman-arm-%.s=arm_%)) -DHAVE_CONFIG_H -I$(srcdir) -I$(builddir) -I$(top_builddir) -S -o $@ $<
diff --git a/pixman/loongson-mmintrin.h b/pixman/loongson-mmintrin.h
new file mode 100644
index 0000000..44d30f5
--- /dev/null
+++ b/pixman/loongson-mmintrin.h
@@ -0,0 +1,218 @@
+/* The gcc-provided loongson intrinsic functions are way too fucking broken
+ * to be of any use, otherwise I'd use them.
+ *
+ * - The hardware instructions are very similar to MMX or iwMMXt. Certainly
+ * close enough that they could have implemented the _mm_*-style intrinsic
+ * interface and had a ton of optimized code available to them. Instead they
+ * implemented something much, much worse.
+ *
+ * - pshuf takes a dead first argument, causing extra instructions to be
+ * generated.
+ *
+ * - There are no 64-bit shift or logical intrinsics, which means you have
+ * to implement them with inline assembly, but this is a nightmare because
+ * gcc doesn't understand that the integer vector datatypes are actually in
+ * floating-point registers, so you end up with braindead code like
+ *
+ * punpcklwd $f9,$f9,$f5
+ * dmtc1 v0,$f8
+ * punpcklwd $f19,$f19,$f5
+ * dmfc1 t9,$f9
+ * dmtc1 v0,$f9
+ * dmtc1 t9,$f20
+ * dmfc1 s0,$f19
+ * punpcklbh $f20,$f20,$f2
+ *
+ * where crap just gets copied back and forth between integer and floating-
+ * point registers ad nauseum.
+ *
+ * Instead of trying to workaround the problems from these crap intrinsics, I
+ * just implement the _mm_* intrinsics needed for pixman-mmx.c using inline
+ * assembly.
+ */
+
+#include <stdint.h>
+
+/* vectors are stored in 64-bit floating-point registers */
+typedef double __m64;
+/* having a 32-bit datatype allows us to use 32-bit loads in places like load8888 */
+typedef float __m32;
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_setzero_si64 (void)
+{
+ return 0.0;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_adds_pu16 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("paddush %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_adds_pu8 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("paddusb %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_and_si64 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("and %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_empty (void)
+{
+
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_mulhi_pu16 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("pmulhuh %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_mullo_pi16 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("pmullh %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_or_si64 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("or %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_packs_pu16 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("packushb %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_shuffle_pi16 (__m64 __m, int64_t __n)
+{
+ __m64 ret;
+ asm("pshufh %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m), "f" (*(__m64 *)&__n)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_slli_si64 (__m64 __m, int64_t __count)
+{
+ __m64 ret;
+ asm("dsll %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m), "f" (*(__m64 *)&__count)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_srli_pi16 (__m64 __m, int64_t __count)
+{
+ __m64 ret;
+ asm("psrlh %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m), "f" (*(__m64 *)&__count)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_srli_si64 (__m64 __m, int64_t __count)
+{
+ __m64 ret;
+ asm("dsrl %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m), "f" (*(__m64 *)&__count)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_unpackhi_pi8 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("punpckhbh %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_unpacklo_pi8 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("punpcklbh %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+/* Since punpcklbh doesn't care about the high 32-bits, we use the __m32 datatype which
+ * allows load8888 to use 32-bit loads */
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_unpacklo_pi8_f (__m32 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("punpcklbh %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
+
+extern __inline __m64 __attribute__((__gnu_inline__, __always_inline__, __artificial__))
+_mm_xor_si64 (__m64 __m1, __m64 __m2)
+{
+ __m64 ret;
+ asm("xor %0, %1, %2\n\t"
+ : "=f" (ret)
+ : "f" (__m1), "f" (__m2)
+ );
+ return ret;
+}
diff --git a/pixman/pixman-cpu.c b/pixman/pixman-cpu.c
index 92942b2..1fc9faa 100644
--- a/pixman/pixman-cpu.c
+++ b/pixman/pixman-cpu.c
@@ -690,7 +690,9 @@ _pixman_choose_implementation (void)
if (pixman_have_arm_iwmmxt ())
imp = _pixman_implementation_create_mmx (imp);
#endif
-
+#ifdef USE_LOONGSON_MMI
+ imp = _pixman_implementation_create_mmx (imp);
+#endif
#ifdef USE_ARM_NEON
if (pixman_have_arm_neon ())
imp = _pixman_implementation_create_arm_neon (imp);
diff --git a/pixman/pixman-mmx.c b/pixman/pixman-mmx.c
index 86af00d..2ca3000 100644
--- a/pixman/pixman-mmx.c
+++ b/pixman/pixman-mmx.c
@@ -33,9 +33,13 @@
#include <config.h>
#endif
-#if defined USE_X86_MMX || defined USE_ARM_IWMMXT
+#if defined USE_X86_MMX || defined USE_ARM_IWMMXT || defined USE_LOONGSON_MMI
+#ifdef USE_LOONGSON_MMI
+#include <loongson-mmintrin.h>
+#else
#include <mmintrin.h>
+#endif
#include "pixman-private.h"
#include "pixman-combine32.h"
@@ -114,11 +118,14 @@ _mm_shuffle_pi16 (__m64 __A, int8_t const __N)
* uint64_t and __m64 values, then define USE_CVT_INTRINSICS.
* If __m64 and uint64_t values can just be cast to each other directly,
* then define USE_M64_CASTS.
+ * If __m64 is a double datatype, then define USE_M64_DOUBLE.
*/
#ifdef _MSC_VER
# define M64_MEMBER m64_u64
#elif defined(__ICC)
# define USE_CVT_INTRINSICS
+#elif defined(USE_LOONGSON_MMI)
+# define USE_M64_DOUBLE
#elif defined(__GNUC__)
# define USE_M64_CASTS
#elif defined(__SUNPRO_C)
@@ -136,7 +143,7 @@ _mm_shuffle_pi16 (__m64 __A, int8_t const __N)
# endif
#endif
-#if defined(USE_M64_CASTS) || defined(USE_CVT_INTRINSICS)
+#if defined(USE_M64_CASTS) || defined(USE_CVT_INTRINSICS) || defined(USE_M64_DOUBLE)
typedef uint64_t mmxdatafield;
#else
typedef __m64 mmxdatafield;
@@ -188,6 +195,8 @@ static const mmx_data_t c =
# define MC(x) to_m64 (c.mmx_ ## x)
#elif defined(USE_M64_CASTS)
# define MC(x) ((__m64)c.mmx_ ## x)
+#elif defined(USE_M64_DOUBLE)
+# define MC(x) (*(__m64 *)&c.mmx_ ## x)
#else
# define MC(x) c.mmx_ ## x
#endif
@@ -202,6 +211,8 @@ to_m64 (uint64_t x)
res.M64_MEMBER = x;
return res;
+#elif defined USE_M64_DOUBLE
+ return *(__m64 *)&x;
#else /* USE_M64_CASTS */
return (__m64)x;
#endif
@@ -215,6 +226,8 @@ to_uint64 (__m64 x)
#elif defined M64_MEMBER /* __m64 is a struct, not an integral type */
uint64_t res = x.M64_MEMBER;
return res;
+#elif defined USE_M64_DOUBLE
+ return *(uint64_t *)&x;
#else /* USE_M64_CASTS */
return (uint64_t)x;
#endif
@@ -347,13 +360,26 @@ static __inline__ uint32_t ldl_u(uint32_t *p)
static force_inline __m64
load (const uint32_t *v)
{
+#ifdef USE_LOONGSON_MMI
+ __m64 ret;
+ asm("lwc1 %0, %1\n\t"
+ : "=f" (ret)
+ : "m" (*v)
+ );
+ return ret;
+#else
return _mm_cvtsi32_si64 (*v);
+#endif
}
static force_inline __m64
load8888 (const uint32_t *v)
{
+#ifdef USE_LOONGSON_MMI
+ return _mm_unpacklo_pi8_f (*(__m32 *)v, _mm_setzero_si64 ());
+#else
return _mm_unpacklo_pi8 (load (v), _mm_setzero_si64 ());
+#endif
}
static force_inline __m64
@@ -372,7 +398,15 @@ pack8888 (__m64 lo, __m64 hi)
static force_inline void
store (uint32_t *dest, __m64 v)
{
+#ifdef USE_LOONGSON_MMI
+ asm("swc1 %1, %0\n\t"
+ : "=m" (*dest)
+ : "f" (v)
+ : "memory"
+ );
+#else
*dest = _mm_cvtsi64_si32 (v);
+#endif
}
static force_inline void
@@ -3263,4 +3297,4 @@ _pixman_implementation_create_mmx (pixman_implementation_t *fallback)
return imp;
}
-#endif /* USE_X86_MMX || USE_ARM_IWMMXT */
+#endif /* USE_X86_MMX || USE_ARM_IWMMXT || USE_LOONGSON_MMI */
diff --git a/pixman/pixman-private.h b/pixman/pixman-private.h
index 8560385..31c3ee7 100644
--- a/pixman/pixman-private.h
+++ b/pixman/pixman-private.h
@@ -539,7 +539,7 @@ _pixman_implementation_create_fast_path (pixman_implementation_t *fallback);
pixman_implementation_t *
_pixman_implementation_create_noop (pixman_implementation_t *fallback);
-#if defined USE_X86_MMX || defined USE_ARM_IWMMXT
+#if defined USE_X86_MMX || defined USE_ARM_IWMMXT || defined USE_LOONGSON_MMI
pixman_implementation_t *
_pixman_implementation_create_mmx (pixman_implementation_t *fallback);
#endif
--
1.7.3.4
More information about the Pixman
mailing list