add MMX and SSE versions of ff_snow_inner_add_yblock

Patch by Robert Edele < yartrebo AH earthlink POIS net >
Original Thread:
Date: Mar 22, 2006 3:24 AM
Subject: [Ffmpeg-devel] [PATCH] snow mmx + sse2 part 5

Originally committed as revision 5197 to svn://svn.ffmpeg.org/ffmpeg/trunk

2 files changed, 297 insertions, 0 deletions

diff --git a/libavcodec/i386/dsputil_mmx.c b/libavcodec/i386/dsputil_mmx.c
index 72e3bd1937..daf7d1add9 100644
--- a/libavcodec/i386/dsputil_mmx.c
+++ b/libavcodec/i386/dsputil_mmx.c
@@ -2569,6 +2569,10 @@ extern void ff_snow_horizontal_compose97i_sse2(DWTELEM *b, int width);
 extern void ff_snow_horizontal_compose97i_mmx(DWTELEM *b, int width);
 extern void ff_snow_vertical_compose97i_sse2(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width);
 extern void ff_snow_vertical_compose97i_mmx(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTELEM *b3, DWTELEM *b4, DWTELEM *b5, int width);
+extern void ff_snow_inner_add_yblock_sse2(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
+                           int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
+extern void ff_snow_inner_add_yblock_mmx(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
+                          int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
 #endif
 
 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
@@ -2962,10 +2966,12 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
         if(mm_flags & MM_SSE2){
             c->horizontal_compose97i = ff_snow_horizontal_compose97i_sse2;
             c->vertical_compose97i = ff_snow_vertical_compose97i_sse2;
+            c->inner_add_yblock = ff_snow_inner_add_yblock_sse2;
         }
         else{
             c->horizontal_compose97i = ff_snow_horizontal_compose97i_mmx;
             c->vertical_compose97i = ff_snow_vertical_compose97i_mmx;
+            c->inner_add_yblock = ff_snow_inner_add_yblock_mmx;
         }
 #endif
     }
diff --git a/libavcodec/i386/snowdsp_mmx.c b/libavcodec/i386/snowdsp_mmx.c
index a66b9fb5fc..b37fd9196f 100644
--- a/libavcodec/i386/snowdsp_mmx.c
+++ b/libavcodec/i386/snowdsp_mmx.c
@@ -653,3 +653,294 @@ void ff_snow_vertical_compose97i_mmx(DWTELEM *b0, DWTELEM *b1, DWTELEM *b2, DWTE
         "m"(b0),"m"(b1),"m"(b2),"m"(b3),"m"(b4),"m"(b5):
         "%"REG_a"","%"REG_b"","%"REG_c"");
 }
+
+#define snow_inner_add_yblock_sse2_header \
+    DWTELEM * * dst_array = sb->line + src_y;\
+    asm volatile(\
+             "mov  %6, %%"REG_c"             \n\t"\
+             "mov  %5, %%"REG_b"             \n\t"\
+             "mov  %3, %%"REG_S"             \n\t"\
+             "pxor %%xmm7, %%xmm7            \n\t" /* 0 */\
+             "pcmpeqd %%xmm3, %%xmm3         \n\t"\
+             "pslld $31, %%xmm3              \n\t"\
+             "psrld $24, %%xmm3              \n\t" /* FRAC_BITS >> 1 */\
+             "1:                             \n\t"\
+             "mov %1, %%"REG_D"              \n\t"\
+             "mov (%%"REG_D"), %%"REG_D"     \n\t"\
+             "add %2, %%"REG_D"              \n\t"
+
+#define snow_inner_add_yblock_sse2_start_8(out_reg1, out_reg2, ptr_offset, s_offset)\
+             "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
+             "movq (%%"REG_d"), %%"out_reg1" \n\t"\
+             "movq (%%"REG_d", %%"REG_c"), %%"out_reg2" \n\t"\
+             "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
+             "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
+             "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\
+             "movq "s_offset"+16(%%"REG_S"), %%xmm4 \n\t"\
+             "punpcklbw %%xmm7, %%xmm0       \n\t"\
+             "punpcklbw %%xmm7, %%xmm4       \n\t"\
+             "pmullw %%xmm0, %%"out_reg1"    \n\t"\
+             "pmullw %%xmm4, %%"out_reg2"    \n\t"
+
+#define snow_inner_add_yblock_sse2_start_16(out_reg1, out_reg2, ptr_offset, s_offset)\
+             "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
+             "movq (%%"REG_d"), %%"out_reg1" \n\t"\
+             "movq 8(%%"REG_d"), %%"out_reg2" \n\t"\
+             "punpcklbw %%xmm7, %%"out_reg1" \n\t"\
+             "punpcklbw %%xmm7, %%"out_reg2" \n\t"\
+             "movq "s_offset"(%%"REG_S"), %%xmm0 \n\t"\
+             "movq "s_offset"+8(%%"REG_S"), %%xmm4 \n\t"\
+             "punpcklbw %%xmm7, %%xmm0       \n\t"\
+             "punpcklbw %%xmm7, %%xmm4       \n\t"\
+             "pmullw %%xmm0, %%"out_reg1"    \n\t"\
+             "pmullw %%xmm4, %%"out_reg2"    \n\t"
+
+#define snow_inner_add_yblock_sse2_accum_8(ptr_offset, s_offset) \
+             snow_inner_add_yblock_sse2_start_8("xmm2", "xmm6", ptr_offset, s_offset)\
+             "paddusw %%xmm2, %%xmm1         \n\t"\
+             "paddusw %%xmm6, %%xmm5         \n\t"
+
+#define snow_inner_add_yblock_sse2_accum_16(ptr_offset, s_offset) \
+             snow_inner_add_yblock_sse2_start_16("xmm2", "xmm6", ptr_offset, s_offset)\
+             "paddusw %%xmm2, %%xmm1         \n\t"\
+             "paddusw %%xmm6, %%xmm5         \n\t"
+
+#define snow_inner_add_yblock_sse2_end_common1\
+             "add $32, %%"REG_S"             \n\t"\
+             "add %%"REG_c", %0              \n\t"\
+             "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\
+             "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\
+             "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\
+             "add %%"REG_c", (%%"REG_a")     \n\t"
+
+#define snow_inner_add_yblock_sse2_end_common2\
+             "jnz 1b                         \n\t"\
+             :"+m"(dst8),"+m"(dst_array)\
+             :\
+             "rm"((long)(src_x<<2)),"m"(obmc),"a"(block),"m"((long)b_h),"rm"((long)src_stride):\
+             "%"REG_b"","%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
+
+#define snow_inner_add_yblock_sse2_end_8\
+             "sal $1, %%"REG_c"              \n\t"\
+             "add $"PTR_SIZE"*2, %1          \n\t"\
+             snow_inner_add_yblock_sse2_end_common1\
+             "sar $1, %%"REG_c"              \n\t"\
+             "sub $2, %%"REG_b"              \n\t"\
+             snow_inner_add_yblock_sse2_end_common2
+
+#define snow_inner_add_yblock_sse2_end_16\
+             "add $"PTR_SIZE"*1, %1          \n\t"\
+             snow_inner_add_yblock_sse2_end_common1\
+             "dec %%"REG_b"                  \n\t"\
+             snow_inner_add_yblock_sse2_end_common2
+
+static void inner_add_yblock_bw_8_obmc_16_bh_even_sse2(uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h,
+                      int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){
+snow_inner_add_yblock_sse2_header
+snow_inner_add_yblock_sse2_start_8("xmm1", "xmm5", "3", "0")
+snow_inner_add_yblock_sse2_accum_8("2", "8")
+snow_inner_add_yblock_sse2_accum_8("1", "128")
+snow_inner_add_yblock_sse2_accum_8("0", "136")
+
+             "mov %0, %%"REG_d"              \n\t"
+             "movdqa (%%"REG_D"), %%xmm0     \n\t"
+             "movdqa %%xmm1, %%xmm2          \n\t"
+
+             "punpckhwd %%xmm7, %%xmm1       \n\t"
+             "punpcklwd %%xmm7, %%xmm2       \n\t"
+             "paddd %%xmm2, %%xmm0           \n\t"
+             "movdqa 16(%%"REG_D"), %%xmm2   \n\t"
+             "paddd %%xmm1, %%xmm2           \n\t"
+             "paddd %%xmm3, %%xmm0           \n\t"
+             "paddd %%xmm3, %%xmm2           \n\t"
+
+             "mov %1, %%"REG_D"              \n\t"
+             "mov "PTR_SIZE"(%%"REG_D"), %%"REG_D";\n\t"
+             "add %2, %%"REG_D"              \n\t"
+
+             "movdqa (%%"REG_D"), %%xmm4     \n\t"
+             "movdqa %%xmm5, %%xmm6          \n\t"
+             "punpckhwd %%xmm7, %%xmm5       \n\t"
+             "punpcklwd %%xmm7, %%xmm6       \n\t"
+             "paddd %%xmm6, %%xmm4           \n\t"
+             "movdqa 16(%%"REG_D"), %%xmm6   \n\t"
+             "paddd %%xmm5, %%xmm6           \n\t"
+             "paddd %%xmm3, %%xmm4           \n\t"
+             "paddd %%xmm3, %%xmm6           \n\t"
+
+             "psrad $8, %%xmm0               \n\t" /* FRAC_BITS. */
+             "psrad $8, %%xmm2               \n\t" /* FRAC_BITS. */
+             "packssdw %%xmm2, %%xmm0        \n\t"
+             "packuswb %%xmm7, %%xmm0        \n\t"
+             "movq %%xmm0, (%%"REG_d")       \n\t"
+
+             "psrad $8, %%xmm4               \n\t" /* FRAC_BITS. */
+             "psrad $8, %%xmm6               \n\t" /* FRAC_BITS. */
+             "packssdw %%xmm6, %%xmm4        \n\t"
+             "packuswb %%xmm7, %%xmm4        \n\t"
+             "movq %%xmm4, (%%"REG_d",%%"REG_c");\n\t"
+snow_inner_add_yblock_sse2_end_8
+}
+
+static void inner_add_yblock_bw_16_obmc_32_sse2(uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h,
+                      int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){
+snow_inner_add_yblock_sse2_header
+snow_inner_add_yblock_sse2_start_16("xmm1", "xmm5", "3", "0")
+snow_inner_add_yblock_sse2_accum_16("2", "16")
+snow_inner_add_yblock_sse2_accum_16("1", "512")
+snow_inner_add_yblock_sse2_accum_16("0", "528")
+
+             "mov %0, %%"REG_d"              \n\t"
+             "movdqa %%xmm1, %%xmm0          \n\t"
+             "movdqa %%xmm5, %%xmm4          \n\t"
+             "punpcklwd %%xmm7, %%xmm0       \n\t"
+             "paddd (%%"REG_D"), %%xmm0      \n\t"
+             "punpckhwd %%xmm7, %%xmm1       \n\t"
+             "paddd 16(%%"REG_D"), %%xmm1    \n\t"
+             "punpcklwd %%xmm7, %%xmm4       \n\t"
+             "paddd 32(%%"REG_D"), %%xmm4    \n\t"
+             "punpckhwd %%xmm7, %%xmm5       \n\t"
+             "paddd 48(%%"REG_D"), %%xmm5    \n\t"
+             "paddd %%xmm3, %%xmm0           \n\t"
+             "paddd %%xmm3, %%xmm1           \n\t"
+             "paddd %%xmm3, %%xmm4           \n\t"
+             "paddd %%xmm3, %%xmm5           \n\t"
+             "psrad $8, %%xmm0               \n\t" /* FRAC_BITS. */
+             "psrad $8, %%xmm1               \n\t" /* FRAC_BITS. */
+             "psrad $8, %%xmm4               \n\t" /* FRAC_BITS. */
+             "psrad $8, %%xmm5               \n\t" /* FRAC_BITS. */
+
+             "packssdw %%xmm1, %%xmm0        \n\t"
+             "packssdw %%xmm5, %%xmm4        \n\t"
+             "packuswb %%xmm4, %%xmm0        \n\t"
+
+             "movdqu %%xmm0, (%%"REG_d")       \n\t"
+
+snow_inner_add_yblock_sse2_end_16
+}
+
+#define snow_inner_add_yblock_mmx_header \
+    DWTELEM * * dst_array = sb->line + src_y;\
+    asm volatile(\
+             "mov  %6, %%"REG_c"             \n\t"\
+             "mov  %5, %%"REG_b"             \n\t"\
+             "mov  %3, %%"REG_S"             \n\t"\
+             "pxor %%mm7, %%mm7              \n\t" /* 0 */\
+             "pcmpeqd %%mm3, %%mm3           \n\t"\
+             "pslld $31, %%mm3               \n\t"\
+             "psrld $24, %%mm3               \n\t" /* FRAC_BITS >> 1 */\
+             "1:                             \n\t"\
+             "mov %1, %%"REG_D"              \n\t"\
+             "mov (%%"REG_D"), %%"REG_D"     \n\t"\
+             "add %2, %%"REG_D"              \n\t"
+
+#define snow_inner_add_yblock_mmx_start(out_reg1, out_reg2, ptr_offset, s_offset, d_offset)\
+             "mov "PTR_SIZE"*"ptr_offset"(%%"REG_a"), %%"REG_d"; \n\t"\
+             "movd "d_offset"(%%"REG_d"), %%"out_reg1" \n\t"\
+             "movd "d_offset"+4(%%"REG_d"), %%"out_reg2" \n\t"\
+             "punpcklbw %%mm7, %%"out_reg1" \n\t"\
+             "punpcklbw %%mm7, %%"out_reg2" \n\t"\
+             "movd "s_offset"(%%"REG_S"), %%mm0 \n\t"\
+             "movd "s_offset"+4(%%"REG_S"), %%mm4 \n\t"\
+             "punpcklbw %%mm7, %%mm0       \n\t"\
+             "punpcklbw %%mm7, %%mm4       \n\t"\
+             "pmullw %%mm0, %%"out_reg1"    \n\t"\
+             "pmullw %%mm4, %%"out_reg2"    \n\t"
+
+#define snow_inner_add_yblock_mmx_accum(ptr_offset, s_offset, d_offset) \
+             snow_inner_add_yblock_mmx_start("mm2", "mm6", ptr_offset, s_offset, d_offset)\
+             "paddusw %%mm2, %%mm1         \n\t"\
+             "paddusw %%mm6, %%mm5         \n\t"
+
+#define snow_inner_add_yblock_mmx_mix(read_offset, write_offset)\
+             "mov %0, %%"REG_d"              \n\t"\
+             "movq %%mm1, %%mm0              \n\t"\
+             "movq %%mm5, %%mm4              \n\t"\
+             "punpcklwd %%mm7, %%mm0         \n\t"\
+             "paddd "read_offset"(%%"REG_D"), %%mm0 \n\t"\
+             "punpckhwd %%mm7, %%mm1         \n\t"\
+             "paddd "read_offset"+8(%%"REG_D"), %%mm1 \n\t"\
+             "punpcklwd %%mm7, %%mm4         \n\t"\
+             "paddd "read_offset"+16(%%"REG_D"), %%mm4 \n\t"\
+             "punpckhwd %%mm7, %%mm5         \n\t"\
+             "paddd "read_offset"+24(%%"REG_D"), %%mm5 \n\t"\
+             "paddd %%mm3, %%mm0             \n\t"\
+             "paddd %%mm3, %%mm1             \n\t"\
+             "paddd %%mm3, %%mm4             \n\t"\
+             "paddd %%mm3, %%mm5             \n\t"\
+             "psrad $8, %%mm0                \n\t"\
+             "psrad $8, %%mm1                \n\t"\
+             "psrad $8, %%mm4                \n\t"\
+             "psrad $8, %%mm5                \n\t"\
+\
+             "packssdw %%mm1, %%mm0          \n\t"\
+             "packssdw %%mm5, %%mm4          \n\t"\
+             "packuswb %%mm4, %%mm0          \n\t"\
+             "movq %%mm0, "write_offset"(%%"REG_d") \n\t"
+
+#define snow_inner_add_yblock_mmx_end(s_step)\
+             "add $"s_step", %%"REG_S"             \n\t"\
+             "add %%"REG_c", "PTR_SIZE"*3(%%"REG_a");\n\t"\
+             "add %%"REG_c", "PTR_SIZE"*2(%%"REG_a");\n\t"\
+             "add %%"REG_c", "PTR_SIZE"*1(%%"REG_a");\n\t"\
+             "add %%"REG_c", (%%"REG_a")     \n\t"\
+             "add $"PTR_SIZE"*1, %1          \n\t"\
+             "add %%"REG_c", %0              \n\t"\
+             "dec %%"REG_b"                  \n\t"\
+             "jnz 1b                         \n\t"\
+             :"+m"(dst8),"+m"(dst_array)\
+             :\
+             "rm"((long)(src_x<<2)),"m"(obmc),"a"(block),"m"((long)b_h),"rm"((long)src_stride):\
+             "%"REG_b"","%"REG_c"","%"REG_S"","%"REG_D"","%"REG_d"");
+
+static void inner_add_yblock_bw_8_obmc_16_mmx(uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h,
+                      int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){
+snow_inner_add_yblock_mmx_header
+snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
+snow_inner_add_yblock_mmx_accum("2", "8", "0")
+snow_inner_add_yblock_mmx_accum("1", "128", "0")
+snow_inner_add_yblock_mmx_accum("0", "136", "0")
+snow_inner_add_yblock_mmx_mix("0", "0")
+snow_inner_add_yblock_mmx_end("16")
+}
+
+static void inner_add_yblock_bw_16_obmc_32_mmx(uint8_t *obmc, const long obmc_stride, uint8_t * * block, int b_w, long b_h,
+                      int src_x, int src_y, long src_stride, slice_buffer * sb, int add, uint8_t * dst8){
+snow_inner_add_yblock_mmx_header
+snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "0", "0")
+snow_inner_add_yblock_mmx_accum("2", "16", "0")
+snow_inner_add_yblock_mmx_accum("1", "512", "0")
+snow_inner_add_yblock_mmx_accum("0", "528", "0")
+snow_inner_add_yblock_mmx_mix("0", "0")
+
+snow_inner_add_yblock_mmx_start("mm1", "mm5", "3", "8", "8")
+snow_inner_add_yblock_mmx_accum("2", "24", "8")
+snow_inner_add_yblock_mmx_accum("1", "520", "8")
+snow_inner_add_yblock_mmx_accum("0", "536", "8")
+snow_inner_add_yblock_mmx_mix("32", "8")
+snow_inner_add_yblock_mmx_end("32")
+}
+
+void ff_snow_inner_add_yblock_sse2(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
+                           int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
+
+    if (b_w == 16)
+        inner_add_yblock_bw_16_obmc_32_sse2(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
+    else if (b_w == 8 && obmc_stride == 16) {
+        if (!(b_h & 1))
+            inner_add_yblock_bw_8_obmc_16_bh_even_sse2(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
+        else
+            inner_add_yblock_bw_8_obmc_16_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
+    } else
+         ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
+}
+
+void ff_snow_inner_add_yblock_mmx(uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
+                          int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8){
+    if (b_w == 16)
+        inner_add_yblock_bw_16_obmc_32_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
+    else if (b_w == 8 && obmc_stride == 16)
+        inner_add_yblock_bw_8_obmc_16_mmx(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
+    else
+        ff_snow_inner_add_yblock(obmc, obmc_stride, block, b_w, b_h, src_x,src_y, src_stride, sb, add, dst8);
+}