target/mips/mxu: Add Q8SAD instruction

The instruction implements SAD (sum-absolute-difference) operation which
is used in motion estimation algorithms. The instruction handles four
8-bit data in parallel.

Signed-off-by: Siarhei Volkau <lis8215@gmail.com>
Message-Id: <20230608104222.1520143-34-lis8215@gmail.com>
Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
This commit is contained in:
Siarhei Volkau 2023-06-08 13:42:22 +03:00 committed by Philippe Mathieu-Daudé
parent 4b9680d338
commit 8aedfb64cd
1 changed files with 45 additions and 0 deletions

View File

@ -408,6 +408,7 @@ enum {
OPC_MXU_Q16SCOP = 0x3B,
OPC_MXU_Q8MADL = 0x3C,
OPC_MXU_S32SFL = 0x3D,
OPC_MXU_Q8SAD = 0x3E,
};
@ -4039,6 +4040,47 @@ static void gen_mxu_s32sfl(DisasContext *ctx)
gen_store_mxu_gpr(t3, XRd);
}
/*
* Q8SAD XRa, XRd, XRb, XRc
* Typical SAD opration for motion estimation.
*/
static void gen_mxu_q8sad(DisasContext *ctx)
{
uint32_t XRd, XRc, XRb, XRa;
XRd = extract32(ctx->opcode, 18, 4);
XRc = extract32(ctx->opcode, 14, 4);
XRb = extract32(ctx->opcode, 10, 4);
XRa = extract32(ctx->opcode, 6, 4);
TCGv t0 = tcg_temp_new();
TCGv t1 = tcg_temp_new();
TCGv t2 = tcg_temp_new();
TCGv t3 = tcg_temp_new();
TCGv t4 = tcg_temp_new();
TCGv t5 = tcg_temp_new();
gen_load_mxu_gpr(t2, XRb);
gen_load_mxu_gpr(t3, XRc);
gen_load_mxu_gpr(t5, XRd);
tcg_gen_movi_tl(t4, 0);
for (int i = 0; i < 4; i++) {
tcg_gen_andi_tl(t0, t2, 0xff);
tcg_gen_andi_tl(t1, t3, 0xff);
tcg_gen_sub_tl(t0, t0, t1);
tcg_gen_abs_tl(t0, t0);
tcg_gen_add_tl(t4, t4, t0);
if (i < 3) {
tcg_gen_shri_tl(t2, t2, 8);
tcg_gen_shri_tl(t3, t3, 8);
}
}
tcg_gen_add_tl(t5, t5, t4);
gen_store_mxu_gpr(t4, XRa);
gen_store_mxu_gpr(t5, XRd);
}
/*
* MXU instruction category: align
* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@ -5040,6 +5082,9 @@ bool decode_ase_mxu(DisasContext *ctx, uint32_t insn)
case OPC_MXU_S32SFL:
gen_mxu_s32sfl(ctx);
break;
case OPC_MXU_Q8SAD:
gen_mxu_q8sad(ctx);
break;
default:
return false;
}