+ "testl $1, %0 \n\t"
+ " jz 1f \n\t"
+ "movq (%1), %%mm0 \n\t"
+ "movq (%2), %%mm1 \n\t"
+ "add %4, %1 \n\t"
+ "add $8, %2 \n\t"
+ PAVGB(%%mm0, %%mm1, %%mm4, %%mm6)
+ "movq %%mm4, (%3) \n\t"
+ "add %5, %3 \n\t"
+ "decl %0 \n\t"
+ ".balign 8 \n\t"
+ "1: \n\t"
+ "movq (%1), %%mm0 \n\t"
+ "movq (%2), %%mm1 \n\t"
+ "add %4, %1 \n\t"
+ "movq (%1), %%mm2 \n\t"
+ "movq 8(%2), %%mm3 \n\t"
+ "add %4, %1 \n\t"
+ PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
+ "movq %%mm4, (%3) \n\t"
+ "add %5, %3 \n\t"
+ "movq %%mm5, (%3) \n\t"
+ "add %5, %3 \n\t"
+ "movq (%1), %%mm0 \n\t"
+ "movq 16(%2), %%mm1 \n\t"
+ "add %4, %1 \n\t"
+ "movq (%1), %%mm2 \n\t"
+ "movq 24(%2), %%mm3 \n\t"
+ "add %4, %1 \n\t"
+ "add $32, %2 \n\t"
+ PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
+ "movq %%mm4, (%3) \n\t"
+ "add %5, %3 \n\t"
+ "movq %%mm5, (%3) \n\t"
+ "add %5, %3 \n\t"
+ "subl $4, %0 \n\t"
+ "jnz 1b \n\t"
+#ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used
+ :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
+#else
+ :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
+#endif
+ :"S"((long)src1Stride), "D"((long)dstStride)
+ :"memory");
+}
+
+static void DEF(put, pixels16_x2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+ MOVQ_BFE(mm6);
+ __asm __volatile(
+ "lea (%3, %3), %%"REG_a" \n\t"
+ ".balign 8 \n\t"
+ "1: \n\t"
+ "movq (%1), %%mm0 \n\t"
+ "movq 1(%1), %%mm1 \n\t"
+ "movq (%1, %3), %%mm2 \n\t"
+ "movq 1(%1, %3), %%mm3 \n\t"
+ PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
+ "movq %%mm4, (%2) \n\t"
+ "movq %%mm5, (%2, %3) \n\t"
+ "movq 8(%1), %%mm0 \n\t"
+ "movq 9(%1), %%mm1 \n\t"
+ "movq 8(%1, %3), %%mm2 \n\t"
+ "movq 9(%1, %3), %%mm3 \n\t"
+ PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
+ "movq %%mm4, 8(%2) \n\t"
+ "movq %%mm5, 8(%2, %3) \n\t"
+ "add %%"REG_a", %1 \n\t"
+ "add %%"REG_a", %2 \n\t"
+ "movq (%1), %%mm0 \n\t"
+ "movq 1(%1), %%mm1 \n\t"
+ "movq (%1, %3), %%mm2 \n\t"
+ "movq 1(%1, %3), %%mm3 \n\t"
+ PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
+ "movq %%mm4, (%2) \n\t"
+ "movq %%mm5, (%2, %3) \n\t"
+ "movq 8(%1), %%mm0 \n\t"
+ "movq 9(%1), %%mm1 \n\t"
+ "movq 8(%1, %3), %%mm2 \n\t"
+ "movq 9(%1, %3), %%mm3 \n\t"
+ PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
+ "movq %%mm4, 8(%2) \n\t"
+ "movq %%mm5, 8(%2, %3) \n\t"
+ "add %%"REG_a", %1 \n\t"
+ "add %%"REG_a", %2 \n\t"
+ "subl $4, %0 \n\t"
+ "jnz 1b \n\t"
+ :"+g"(h), "+S"(pixels), "+D"(block)
+ :"r"((long)line_size)
+ :REG_a, "memory");
+}
+
+static void attribute_unused DEF(put, pixels16_l2)(uint8_t *dst, uint8_t *src1, uint8_t *src2, int dstStride, int src1Stride, int h)
+{
+ MOVQ_BFE(mm6);
+ __asm __volatile(
+ "testl $1, %0 \n\t"
+ " jz 1f \n\t"
+ "movq (%1), %%mm0 \n\t"
+ "movq (%2), %%mm1 \n\t"
+ "movq 8(%1), %%mm2 \n\t"
+ "movq 8(%2), %%mm3 \n\t"
+ "add %4, %1 \n\t"
+ "add $16, %2 \n\t"
+ PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
+ "movq %%mm4, (%3) \n\t"
+ "movq %%mm5, 8(%3) \n\t"
+ "add %5, %3 \n\t"
+ "decl %0 \n\t"
+ ".balign 8 \n\t"
+ "1: \n\t"
+ "movq (%1), %%mm0 \n\t"
+ "movq (%2), %%mm1 \n\t"
+ "movq 8(%1), %%mm2 \n\t"
+ "movq 8(%2), %%mm3 \n\t"
+ "add %4, %1 \n\t"
+ PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
+ "movq %%mm4, (%3) \n\t"
+ "movq %%mm5, 8(%3) \n\t"
+ "add %5, %3 \n\t"
+ "movq (%1), %%mm0 \n\t"
+ "movq 16(%2), %%mm1 \n\t"
+ "movq 8(%1), %%mm2 \n\t"
+ "movq 24(%2), %%mm3 \n\t"
+ "add %4, %1 \n\t"
+ PAVGBP(%%mm0, %%mm1, %%mm4, %%mm2, %%mm3, %%mm5)
+ "movq %%mm4, (%3) \n\t"
+ "movq %%mm5, 8(%3) \n\t"
+ "add %5, %3 \n\t"
+ "add $32, %2 \n\t"
+ "subl $2, %0 \n\t"
+ "jnz 1b \n\t"
+#ifdef PIC //Note "+bm" and "+mb" are buggy too (with gcc 3.2.2 at least) and cant be used
+ :"+m"(h), "+a"(src1), "+c"(src2), "+d"(dst)
+#else
+ :"+b"(h), "+a"(src1), "+c"(src2), "+d"(dst)
+#endif
+ :"S"((long)src1Stride), "D"((long)dstStride)
+ :"memory");
+}
+
+static void DEF(put, pixels8_y2)(uint8_t *block, const uint8_t *pixels, int line_size, int h)
+{
+ MOVQ_BFE(mm6);
+ __asm __volatile(
+ "lea (%3, %3), %%"REG_a" \n\t"