Re: [Qemu-devel] [PATCH v5 2/2] target/mips: Optimize ILVEV.<B|H|W|D> MSA instructions

[Philippe Mathieu-Daudé]

Hi Mateja,

On 4/2/19 5:15 PM, Mateja Marjanovic wrote:
> From: Mateja Marjanovic <address@hidden>
> 
> Optimize set of MSA instructions ILVEV, using directly
> tcg registers and performing logic on them instead of
> using helpers.

Maybe you can still let this previous comment (if still valid):

  Performance measurement is done by executing the
  instructions large number of times on a computer
  with Intel Core i7-3770 CPU @ 3.40GHz×8.

> 
> In the following table, the first column is the performance
> before this patch. The second represents the performance,
> after converting from helpers to tcg, but without using
> tcg_gen_deposit function. The third one is the solution
> which is implemented in this patch.

You are describing the "no-deposit" which refers to a previous series
but won't be accessible in the git repository.

I think this table is useful in the cover of this series, and in the
commit message you should use || before || after || and drop the
"no-deposit" case.

> 
>  instr    ||   before    || no-deposit ||  with-deposit
> ========================================================
>  ilvev.b  ||  126.92 ms  ||  24.52 ms  ||  24.43 ms
>  ilvev.h  ||   93.67 ms  ||  23.92 ms  ||  23.86 ms
>  ilvev.w  ||  117.86 ms  ||  23.83 ms  ||  22.17 ms
>  ilvev.d  ||   45.49 ms  ||  19.74 ms  ||  19.71 ms
> 
> The solution with deposit is suggested by Richard Henderson.
> 

The gitdm parsable form is:

Suggested-by: Richard Henderson <address@hidden>

> Signed-off-by: Mateja Marjanovic <address@hidden>
> ---
>  target/mips/helper.h     |   1 -
>  target/mips/msa_helper.c |   9 ----
>  target/mips/translate.c  | 105 
> ++++++++++++++++++++++++++++++++++++++++++++++-
>  3 files changed, 104 insertions(+), 11 deletions(-)
> 
> diff --git a/target/mips/helper.h b/target/mips/helper.h
> index 02e16c7..82f6a40 100644
> --- a/target/mips/helper.h
> +++ b/target/mips/helper.h
> @@ -864,7 +864,6 @@ DEF_HELPER_5(msa_pckev_df, void, env, i32, i32, i32, i32)
>  DEF_HELPER_5(msa_pckod_df, void, env, i32, i32, i32, i32)
>  DEF_HELPER_5(msa_ilvl_df, void, env, i32, i32, i32, i32)
>  DEF_HELPER_5(msa_ilvr_df, void, env, i32, i32, i32, i32)
> -DEF_HELPER_5(msa_ilvev_df, void, env, i32, i32, i32, i32)
>  DEF_HELPER_5(msa_vshf_df, void, env, i32, i32, i32, i32)
>  DEF_HELPER_5(msa_srar_df, void, env, i32, i32, i32, i32)
>  DEF_HELPER_5(msa_srlr_df, void, env, i32, i32, i32, i32)
> diff --git a/target/mips/msa_helper.c b/target/mips/msa_helper.c
> index a7ea6aa..d5c3842 100644
> --- a/target/mips/msa_helper.c
> +++ b/target/mips/msa_helper.c
> @@ -1197,15 +1197,6 @@ MSA_FN_DF(ilvl_df)
>      } while (0)
>  MSA_FN_DF(ilvr_df)
>  #undef MSA_DO
> -
> -#define MSA_DO(DF)                      \
> -    do {                                \
> -        pwx->DF[2*i]   = pwt->DF[2*i];  \
> -        pwx->DF[2*i+1] = pws->DF[2*i];  \
> -    } while (0)
> -MSA_FN_DF(ilvev_df)
> -#undef MSA_DO
> -
>  #undef MSA_LOOP_COND
>  
>  #define MSA_LOOP_COND(DF) \
> diff --git a/target/mips/translate.c b/target/mips/translate.c
> index 04406d6..e26c6a6 100644
> --- a/target/mips/translate.c
> +++ b/target/mips/translate.c
> @@ -28974,6 +28974,94 @@ static inline void gen_ilvod_d(CPUMIPSState *env, 
> uint32_t wd,
>      tcg_gen_mov_i64(msa_wr_d[wd * 2 + 1], msa_wr_d[ws * 2 + 1]);
>  }
>  
> +/*
> + * [MSA] ILVEV.B wd, ws, wt
> + *
> + *   Vector Interleave Even (byte data elements)
> + *
> + */
> +static inline void gen_ilvev_b(CPUMIPSState *env, uint32_t wd,
> +                               uint32_t ws, uint32_t wt)
> +{
> +    TCGv_i64 t1 = tcg_temp_new_i64();
> +    TCGv_i64 t2 = tcg_temp_new_i64();
> +    const uint64_t mask = 0x00ff00ff00ff00ffULL;
> +
> +    tcg_gen_andi_i64(t1, msa_wr_d[wt * 2], mask);
> +    tcg_gen_andi_i64(t2, msa_wr_d[ws * 2], mask);
> +    tcg_gen_shli_i64(t2, t2, 8);
> +    tcg_gen_or_i64(msa_wr_d[wd * 2], t1, t2);
> +

Richard, is it cheaper to use another register to keep the constant mask
(here reused 4x)?

Such:

       TCGv_i64 mask = tcg_const_i64(0x00ff00ff00ff00ffULL);

       tcg_gen_and_i64(t1, msa_wr_d[wt * 2], mask);
       tcg_gen_and_i64(t2, msa_wr_d[ws * 2], mask);
       tcg_gen_shli_i64(t2, t2, 8);
       tcg_gen_or_i64(msa_wr_d[wd * 2], t1, t2);

> +    tcg_gen_andi_i64(t1, msa_wr_d[wt * 2 + 1], mask);
> +    tcg_gen_andi_i64(t2, msa_wr_d[ws * 2 + 1], mask);

Here use tcg_gen_and_i64() too.

> +    tcg_gen_shli_i64(t2, t2, 8);
> +    tcg_gen_or_i64(msa_wr_d[wd * 2 + 1], t1, t2);
> +

       tcg_temp_free_i64(mask);

> +    tcg_temp_free_i64(t1);
> +    tcg_temp_free_i64(t2);

Mateja: Can you test for perf easily?

> +}
> +
> +/*
> + * [MSA] ILVEV.H wd, ws, wt
> + *
> + *   Vector Interleave Even (halfword data elements)
> + *
> + */
> +static inline void gen_ilvev_h(CPUMIPSState *env, uint32_t wd,
> +                               uint32_t ws, uint32_t wt)
> +{
> +    TCGv_i64 t1 = tcg_temp_new_i64();
> +    TCGv_i64 t2 = tcg_temp_new_i64();
> +    const uint64_t mask = 0x0000ffff0000ffffULL;
> +
> +    tcg_gen_andi_i64(t1, msa_wr_d[wt * 2], mask);
> +    tcg_gen_andi_i64(t2, msa_wr_d[ws * 2], mask);
> +    tcg_gen_shli_i64(t2, t2, 16);
> +    tcg_gen_or_i64(msa_wr_d[wd * 2], t1, t2);
> +
> +    tcg_gen_andi_i64(t1, msa_wr_d[wt * 2 + 1], mask);
> +    tcg_gen_andi_i64(t2, msa_wr_d[ws * 2 + 1], mask);
> +    tcg_gen_shli_i64(t2, t2, 16);
> +    tcg_gen_or_i64(msa_wr_d[wd * 2 + 1], t1, t2);
> +
> +    tcg_temp_free_i64(t1);
> +    tcg_temp_free_i64(t2);

Very similiar to gen_ilvev_b(), changing the mask and shift. Maybe worth
a refactor.

> +}
> +
> +/*
> + * [MSA] ILVEV.W wd, ws, wt
> + *
> + *   Vector Interleave Even (word data elements)
> + *
> + */
> +static inline void gen_ilvev_w(CPUMIPSState *env, uint32_t wd,
> +                               uint32_t ws, uint32_t wt)
> +{
> +    TCGv_i64 t1 = tcg_temp_new_i64();
> +    const uint64_t mask = 0x00000000ffffffffULL;
> +
> +    tcg_gen_andi_i64(t1, msa_wr_d[wt * 2], mask);
> +    tcg_gen_deposit_i64(msa_wr_d[wd * 2], t1, msa_wr_d[ws * 2], 32, 32);
> +
> +    tcg_gen_andi_i64(t1, msa_wr_d[wt * 2 + 1], mask);
> +    tcg_gen_deposit_i64(msa_wr_d[wd * 2 + 1], t1, msa_wr_d[ws * 2 + 1], 32, 
> 32);
> +
> +    tcg_temp_free_i64(t1);
> +}
> +
> +/*
> + * [MSA] ILVEV.D wd, ws, wt
> + *
> + *   Vector Interleave Even (Double data elements)
> + *
> + */
> +static inline void gen_ilvev_d(CPUMIPSState *env, uint32_t wd,
> +                               uint32_t ws, uint32_t wt)
> +{
> +    tcg_gen_mov_i64(msa_wr_d[wd * 2 + 1], msa_wr_d[ws * 2]);
> +    tcg_gen_mov_i64(msa_wr_d[wd * 2], msa_wr_d[wt * 2]);
> +}
> +
>  static void gen_msa_3r(CPUMIPSState *env, DisasContext *ctx)
>  {
>  #define MASK_MSA_3R(op)    (MASK_MSA_MINOR(op) | (op & (0x7 << 23)))
> @@ -29130,7 +29218,22 @@ static void gen_msa_3r(CPUMIPSState *env, 
> DisasContext *ctx)
>          gen_helper_msa_mod_s_df(cpu_env, tdf, twd, tws, twt);
>          break;
>      case OPC_ILVEV_df:
> -        gen_helper_msa_ilvev_df(cpu_env, tdf, twd, tws, twt);
> +        switch (df) {
> +        case DF_BYTE:
> +            gen_ilvev_b(env, wd, ws, wt);
> +            break;
> +        case DF_HALF:
> +            gen_ilvev_h(env, wd, ws, wt);
> +            break;
> +        case DF_WORD:
> +            gen_ilvev_w(env, wd, ws, wt);
> +            break;
> +        case DF_DOUBLE:
> +            gen_ilvev_d(env, wd, ws, wt);
> +            break;
> +        default:
> +            assert(0);
> +        }
>          break;
>      case OPC_BINSR_df:
>          gen_helper_msa_binsr_df(cpu_env, tdf, twd, tws, twt);
> 

Thanks,

Phil.