WIP on adding MMX support for some of the HR codec primitives

4 files changed, 124 insertions, 3 deletions

diff --git a/configure.ac b/configure.ac
index 1208025..70232c3 100644
--- a/configure.ac
+++ b/configure.ac
@@ -45,6 +45,7 @@ AC_SUBST(SYMBOL_VISIBILITY)
 
 # Checks for programs.
 AC_PROG_CC
+AM_PROG_AS
 AC_PROG_LIBTOOL
 
 # Checks for libraries.
diff --git a/libgsmhr/Makefile.am b/libgsmhr/Makefile.am
index 07390f9..a3a5a83 100644
--- a/libgsmhr/Makefile.am
+++ b/libgsmhr/Makefile.am
@@ -4,7 +4,7 @@ AM_CFLAGS = -fPIC -Wall ${SYMBOL_VISIBILITY}
 LIBVERSION=0:0:0
 
 REFSRC_PATH=refsrc
-REFSRC_SRC=refsrc/dtx.c refsrc/globdefs.c refsrc/host.c refsrc/mathhalf.c refsrc/sp_enc.c refsrc/sp_rom.c refsrc/vad.c refsrc/err_conc.c refsrc/homing.c refsrc/mathdp31.c refsrc/sp_dec.c refsrc/sp_frm.c refsrc/sp_sfrm.c
+REFSRC_SRC=refsrc/dtx.c refsrc/globdefs.c refsrc/host.c refsrc/mathhalf.c refsrc/sp_enc.c refsrc/sp_rom.c refsrc/vad.c refsrc/err_conc.c refsrc/homing.c refsrc/mathdp31.c refsrc/sp_dec.c refsrc/sp_frm.c refsrc/sp_sfrm.c asm.s mmx.c
 
 ${REFSRC_PATH}/.downloaded:
 	./fetch_sources.py "${REFSRC_PATH}"
@@ -19,5 +19,5 @@ ${REFSRC_PATH}/dtx.c: ${REFSRC_PATH}/.downloaded
 lib_LTLIBRARIES = libgsmhr.la
 libgsmhr_la_SOURCES = $(REFSRC_SRC) libgsmhr.c
 
-clean-local:
-	-rm -rf ${REFSRC_PATH}
+#clean-local:
+#	-rm -rf ${REFSRC_PATH}
diff --git a/libgsmhr/asm.s b/libgsmhr/asm.s
new file mode 100644
index 0000000..d1d491e
--- /dev/null
+++ b/libgsmhr/asm.s
@@ -0,0 +1,49 @@
+
+.globl sat_adds32b
+.type   sat_adds32b,@function
+sat_adds32b:
+	mov    %edi, %eax
+	shr    $0x1f, %edi
+	add    $0x7fffffff, %edi
+	add    %esi, %eax
+	cmovo  %edi, %eax
+	retq
+.size	sat_adds32b, .-sat_adds32b
+
+.globl sat_subs32b
+.type   sat_subs32b,@function
+sat_subs32b:
+	mov    %edi, %eax
+	shr    $0x1f, %edi
+	add    $0x7fffffff, %edi
+	sub    %esi, %eax
+	cmovo  %edi, %eax
+	retq
+.size	sat_subs32b, .-sat_subs32b
+
+.globl sat_divs32b
+.type   sat_divs32b,@function
+sat_divs32b:
+	mov    %edi, %eax
+	lea    0x1(%rsi), %edx
+	add    $0x80000000, %edi
+	or     %edx, %edi
+	cdq
+	neg    %edi
+	sbb    $-1, %eax
+	idiv   %esi
+	retq
+.size	sat_divs32b, .-sat_divs32b
+
+.globl sat_muls32b
+.type   sat_muls32b,@function
+sat_muls32b:
+	mov    %edi, %eax
+	xor    %esi, %edi
+	shr    $0x1f, %edi
+	add    $0x7fffffff, %edi
+	imul   %esi
+	cmovc  %edi, %eax
+	retq
+.size	sat_muls32b, .-sat_muls32b
+
diff --git a/libgsmhr/mmx.c b/libgsmhr/mmx.c
new file mode 100644
index 0000000..8e4fad2
--- /dev/null
+++ b/libgsmhr/mmx.c
@@ -0,0 +1,71 @@
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <stdint.h>
+
+typedef int16_t v4hi __attribute__ ((vector_size (8)));
+typedef int32_t v2si __attribute__ ((vector_size (8)));
+
+union v4_s16 {
+	v4hi	v;
+	int16_t	s16[4];
+};
+
+union v2_s32 {
+	v2si	v;
+	int32_t s32[2];
+};
+
+extern int32_t sat_adds32b(int32_t a, int32_t b);
+
+/* multiply-accumulate of 8 16-bit values */
+int32_t mmx_mac_unsat(int16_t *_x, int16_t *_y)
+{
+	register union v4_s16 *x = (union v4_s16 *)_x;
+	union v4_s16 *y = (union v4_s16 *)_y;
+	union v2_s32 im1, im2, r;
+
+	/* im1.s32[0] = x->s16[0] * y->s16[0] + x->s16[1] * y->s16[1]
+	 * im1.s32[1] = x->s16[2] * y->s16[2] + x->s16[3] * y->s16[3] */
+	im1.v = __builtin_ia32_pmaddwd(x->v, y->v);
+	/* results are saturated */
+
+	/* im2.s32[0] = x->s16[4] * y->s16[4] + x->s16[5] * y->s16[5]
+	 * im2.s32[1] = x->s16[6] * y->s16[6]+ x->s16[7] * y->s16[7] */
+	im2.v = __builtin_ia32_pmaddwd((x+1)->v, (y+1)->v);
+	/* results are saturated */
+
+	/* im1.s32[0] = im1.s32[0] + im2.s32[0]
+	 * im1.s32[1] = im1.s32[1] + im2.s32[1] */
+	/* FIXME: overflow in addition could happen */
+	im1.v = __builtin_ia32_paddd(im1.v, im2.v);
+
+	/* FIXME: overflow in addition could happen */
+	return (im1.s32[0] + im1.s32[1]) << 1;
+}
+
+/* multiply-accumulate of 8 16-bit values */
+int32_t mmx_mac_sat(int16_t *_x, int16_t *_y)
+{
+	register union v4_s16 *x = (union v4_s16 *)_x;
+	union v4_s16 *y = (union v4_s16 *)_y;
+	union v2_s32 im1, im2, r;
+
+	/* im1.s32[0] = x->s16[0] * y->s16[0] + x->s16[1] * y->s16[1]
+	 * im1.s32[1] = x->s16[2] * y->s16[2] + x->s16[3] * y->s16[3] */
+	im1.v = __builtin_ia32_pmaddwd(x->v, y->v);
+	/* results are saturated */
+
+	/* im2.s32[0] = x->s16[4] * y->s16[4] + x->s16[5] * y->s16[5]
+	 * im2.s32[1] = x->s16[6] * y->s16[6]+ x->s16[7] * y->s16[7] */
+	im2.v = __builtin_ia32_pmaddwd((x+1)->v, (y+1)->v);
+	/* results are saturated */
+
+	/* im1.s32[0] = im1.s32[0] + im2.s32[0]
+	 * im1.s32[1] = im1.s32[1] + im2.s32[1] */
+	im1.s32[0] = sat_adds32b(im1.s32[0], im2.s32[0]);
+	im1.s32[1] = sat_adds32b(im1.s32[1], im2.s32[1]);
+
+	/* FIXME: overflow in addition could happen */
+	return sat_adds32b(im1.s32[0], im1.s32[1]) << 1;
+}