[Mesa-dev] [PATCH] swr: [rasterizer jit] use signed integer representation for logic op

[Ilia Mirkin]

On Thu, Nov 24, 2016 at 6:11 PM, Ilia Mirkin <imirkin at alum.mit.edu> wrote:
> Instead of (incorrectly) biasing the snorm value to make it look like a
> unorm, just use signed integer math.
>
> This fixes arb_color_buffer_float-render GL_RGBA8_SNORM
>
> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
> ---
>  src/gallium/drivers/swr/rasterizer/jitter/blend_jit.cpp | 17 ++++++++++++-----
>  1 file changed, 12 insertions(+), 5 deletions(-)
>
> diff --git a/src/gallium/drivers/swr/rasterizer/jitter/blend_jit.cpp b/src/gallium/drivers/swr/rasterizer/jitter/blend_jit.cpp
> index ad809c4..339ca52 100644
> --- a/src/gallium/drivers/swr/rasterizer/jitter/blend_jit.cpp
> +++ b/src/gallium/drivers/swr/rasterizer/jitter/blend_jit.cpp
> @@ -692,9 +692,13 @@ struct BlendJit : public Builder
>                      dst[i] = BITCAST(dst[i], mSimdInt32Ty);
>                      break;
>                  case SWR_TYPE_SNORM:
> -                    src[i] = FADD(src[i], VIMMED1(0.5f));
> -                    dst[i] = FADD(dst[i], VIMMED1(0.5f));
> -                    /* fallthrough */
> +                    src[i] = FP_TO_SI(
> +                        FMUL(src[i], VIMMED1(scale[i])),
> +                        mSimdInt32Ty);
> +                    dst[i] = FP_TO_SI(
> +                        FMUL(dst[i], VIMMED1(scale[i])),
> +                        mSimdInt32Ty);
> +                    break;
>                  case SWR_TYPE_UNORM:
>                      src[i] = FP_TO_UI(
>                          FMUL(src[i], VIMMED1(scale[i])),
> @@ -728,11 +732,14 @@ struct BlendJit : public Builder
>                      result[i] = BITCAST(result[i], mSimdFP32Ty);
>                      break;
>                  case SWR_TYPE_SNORM:
> +                    result[i] = SHL(result[i], 32 - info.bpc[i]);
> +                    result[i] = ASHR(result[i], 32 - info.bpc[i]);

These two immediate arguments should probably have a C() around them.
I've fixed that up in my tree. Hopefully these will emit as VPSLLD and
VPSRAD. Not sure how to check that.

> +                    result[i] = FMUL(SI_TO_FP(result[i], mSimdFP32Ty),
> +                                     VIMMED1(1.0f / scale[i]));
> +                    break;
>                  case SWR_TYPE_UNORM:
>                      result[i] = FMUL(UI_TO_FP(result[i], mSimdFP32Ty),
>                                       VIMMED1(1.0f / scale[i]));
> -                    if (info.type[i] == SWR_TYPE_SNORM)
> -                        result[i] = FADD(result[i], VIMMED1(-0.5f));
>                      break;
>                  }
>
> --
> 2.7.3
>