From 7bb1206ab916bac96c029de850faf8d3bb1a08a2 Mon Sep 17 00:00:00 2001 From: Siarhei Volkau Date: Thu, 8 Jun 2023 13:42:18 +0300 Subject: [PATCH] target/mips/mxu: Add Q8MAC Q8MACSU instructions MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit These instructions do parallel quad 8-bit multiply and accumulate. They are close to existing Q8MUL Q8MULSU so the generation function modified to support all of them. Also the patch fixes decoding of Q8MULSU according to tests on hardware. Signed-off-by: Siarhei Volkau Message-Id: <20230608104222.1520143-30-lis8215@gmail.com> Signed-off-by: Philippe Mathieu-Daudé --- target/mips/tcg/mxu_translate.c | 128 +++++++++++++++++++++----------- 1 file changed, 86 insertions(+), 42 deletions(-) diff --git a/target/mips/tcg/mxu_translate.c b/target/mips/tcg/mxu_translate.c index f8c3fdb103..6a82849ef7 100644 --- a/target/mips/tcg/mxu_translate.c +++ b/target/mips/tcg/mxu_translate.c @@ -326,7 +326,7 @@ * ├─ 110111 ─ OPC_MXU_Q16SAR * │ 23..22 * ├─ 111000 ─ OPC_MXU__POOL19 ─┬─ 00 ─ OPC_MXU_Q8MUL - * │ └─ 01 ─ OPC_MXU_Q8MULSU + * │ └─ 10 ─ OPC_MXU_Q8MULSU * │ * │ 20..18 * ├─ 111001 ─ OPC_MXU__POOL20 ─┬─ 000 ─ OPC_MXU_Q8MOVZ @@ -404,6 +404,7 @@ enum { OPC_MXU_Q16SAR = 0x37, OPC_MXU__POOL19 = 0x38, OPC_MXU__POOL20 = 0x39, + OPC_MXU__POOL21 = 0x3A, }; @@ -539,7 +540,7 @@ enum { */ enum { OPC_MXU_Q8MUL = 0x00, - OPC_MXU_Q8MULSU = 0x01, + OPC_MXU_Q8MULSU = 0x02, }; /* @@ -554,6 +555,15 @@ enum { OPC_MXU_S32MOVN = 0x05, }; +/* + * MXU pool 21 + */ +enum { + OPC_MXU_Q8MAC = 0x00, + OPC_MXU_Q8MACSU = 0x02, +}; + + /* MXU accumulate add/subtract 1-bit pattern 'aptn1' */ #define MXU_APTN1_A 0 #define MXU_APTN1_S 1 @@ -1328,13 +1338,17 @@ static void gen_mxu_s16mad(DisasContext *ctx) } /* - * Q8MUL XRa, XRb, XRc, XRd - Parallel unsigned 8 bit pattern multiply - * Q8MULSU XRa, XRb, XRc, XRd - Parallel signed 8 bit pattern multiply + * Q8MUL XRa, XRb, XRc, XRd - Parallel quad unsigned 8 bit multiply + * Q8MULSU XRa, XRb, XRc, XRd - Parallel quad signed 8 bit multiply + * Q8MAC XRa, XRb, XRc, XRd - Parallel quad unsigned 8 bit multiply + * and accumulate + * Q8MACSU XRa, XRb, XRc, XRd - Parallel quad signed 8 bit multiply + * and accumulate */ -static void gen_mxu_q8mul_q8mulsu(DisasContext *ctx) +static void gen_mxu_q8mul_mac(DisasContext *ctx, bool su, bool mac) { TCGv t0, t1, t2, t3, t4, t5, t6, t7; - uint32_t XRa, XRb, XRc, XRd, sel; + uint32_t XRa, XRb, XRc, XRd, aptn2; t0 = tcg_temp_new(); t1 = tcg_temp_new(); @@ -1349,54 +1363,60 @@ static void gen_mxu_q8mul_q8mulsu(DisasContext *ctx) XRb = extract32(ctx->opcode, 10, 4); XRc = extract32(ctx->opcode, 14, 4); XRd = extract32(ctx->opcode, 18, 4); - sel = extract32(ctx->opcode, 22, 2); + aptn2 = extract32(ctx->opcode, 24, 2); gen_load_mxu_gpr(t3, XRb); gen_load_mxu_gpr(t7, XRc); - if (sel == 0x2) { - /* Q8MULSU */ - tcg_gen_ext8s_tl(t0, t3); - tcg_gen_shri_tl(t3, t3, 8); - tcg_gen_ext8s_tl(t1, t3); - tcg_gen_shri_tl(t3, t3, 8); - tcg_gen_ext8s_tl(t2, t3); - tcg_gen_shri_tl(t3, t3, 8); - tcg_gen_ext8s_tl(t3, t3); + if (su) { + /* Q8MULSU / Q8MACSU */ + tcg_gen_sextract_tl(t0, t3, 0, 8); + tcg_gen_sextract_tl(t1, t3, 8, 8); + tcg_gen_sextract_tl(t2, t3, 16, 8); + tcg_gen_sextract_tl(t3, t3, 24, 8); } else { - /* Q8MUL */ - tcg_gen_ext8u_tl(t0, t3); - tcg_gen_shri_tl(t3, t3, 8); - tcg_gen_ext8u_tl(t1, t3); - tcg_gen_shri_tl(t3, t3, 8); - tcg_gen_ext8u_tl(t2, t3); - tcg_gen_shri_tl(t3, t3, 8); - tcg_gen_ext8u_tl(t3, t3); + /* Q8MUL / Q8MAC */ + tcg_gen_extract_tl(t0, t3, 0, 8); + tcg_gen_extract_tl(t1, t3, 8, 8); + tcg_gen_extract_tl(t2, t3, 16, 8); + tcg_gen_extract_tl(t3, t3, 24, 8); } - tcg_gen_ext8u_tl(t4, t7); - tcg_gen_shri_tl(t7, t7, 8); - tcg_gen_ext8u_tl(t5, t7); - tcg_gen_shri_tl(t7, t7, 8); - tcg_gen_ext8u_tl(t6, t7); - tcg_gen_shri_tl(t7, t7, 8); - tcg_gen_ext8u_tl(t7, t7); + tcg_gen_extract_tl(t4, t7, 0, 8); + tcg_gen_extract_tl(t5, t7, 8, 8); + tcg_gen_extract_tl(t6, t7, 16, 8); + tcg_gen_extract_tl(t7, t7, 24, 8); tcg_gen_mul_tl(t0, t0, t4); tcg_gen_mul_tl(t1, t1, t5); tcg_gen_mul_tl(t2, t2, t6); tcg_gen_mul_tl(t3, t3, t7); - tcg_gen_andi_tl(t0, t0, 0xFFFF); - tcg_gen_andi_tl(t1, t1, 0xFFFF); - tcg_gen_andi_tl(t2, t2, 0xFFFF); - tcg_gen_andi_tl(t3, t3, 0xFFFF); + if (mac) { + gen_load_mxu_gpr(t4, XRd); + gen_load_mxu_gpr(t5, XRa); + tcg_gen_extract_tl(t6, t4, 0, 16); + tcg_gen_extract_tl(t7, t4, 16, 16); + if (aptn2 & 1) { + tcg_gen_sub_tl(t0, t6, t0); + tcg_gen_sub_tl(t1, t7, t1); + } else { + tcg_gen_add_tl(t0, t6, t0); + tcg_gen_add_tl(t1, t7, t1); + } + tcg_gen_extract_tl(t6, t5, 0, 16); + tcg_gen_extract_tl(t7, t5, 16, 16); + if (aptn2 & 2) { + tcg_gen_sub_tl(t2, t6, t2); + tcg_gen_sub_tl(t3, t7, t3); + } else { + tcg_gen_add_tl(t2, t6, t2); + tcg_gen_add_tl(t3, t7, t3); + } + } - tcg_gen_shli_tl(t1, t1, 16); - tcg_gen_shli_tl(t3, t3, 16); - - tcg_gen_or_tl(t0, t0, t1); - tcg_gen_or_tl(t1, t2, t3); + tcg_gen_deposit_tl(t0, t0, t1, 16, 16); + tcg_gen_deposit_tl(t1, t2, t3, 16, 16); gen_store_mxu_gpr(t0, XRd); gen_store_mxu_gpr(t1, XRa); @@ -4548,12 +4568,14 @@ static void decode_opc_mxu__pool18(DisasContext *ctx) static void decode_opc_mxu__pool19(DisasContext *ctx) { - uint32_t opcode = extract32(ctx->opcode, 22, 2); + uint32_t opcode = extract32(ctx->opcode, 22, 4); switch (opcode) { case OPC_MXU_Q8MUL: + gen_mxu_q8mul_mac(ctx, false, false); + break; case OPC_MXU_Q8MULSU: - gen_mxu_q8mul_q8mulsu(ctx); + gen_mxu_q8mul_mac(ctx, true, false); break; default: MIPS_INVAL("decode_opc_mxu"); @@ -4592,6 +4614,25 @@ static void decode_opc_mxu__pool20(DisasContext *ctx) } } +static void decode_opc_mxu__pool21(DisasContext *ctx) +{ + uint32_t opcode = extract32(ctx->opcode, 22, 2); + + switch (opcode) { + case OPC_MXU_Q8MAC: + gen_mxu_q8mul_mac(ctx, false, true); + break; + case OPC_MXU_Q8MACSU: + gen_mxu_q8mul_mac(ctx, true, true); + break; + default: + MIPS_INVAL("decode_opc_mxu"); + gen_reserved_instruction(ctx); + break; + } +} + + bool decode_ase_mxu(DisasContext *ctx, uint32_t insn) { uint32_t opcode = extract32(insn, 0, 6); @@ -4755,6 +4796,9 @@ bool decode_ase_mxu(DisasContext *ctx, uint32_t insn) case OPC_MXU__POOL20: decode_opc_mxu__pool20(ctx); break; + case OPC_MXU__POOL21: + decode_opc_mxu__pool21(ctx); + break; default: return false; }