[Mesa-dev] [PATCH] nir,ac/nir: fix cube_face_coord

[Samuel Pitoiset]

Reviewed-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>

On 4/12/19 12:15 PM, Rhys Perry wrote:
> Seems it was missing the "/ ma + 0.5" and the order was swapped.
>
> Fixes: a1a2a8dfda7b9cac7e ('nir: add AMD_gcn_shader extended instructions')
> Signed-off-by: Rhys Perry <pendingchaos02 at gmail.com>
> ---
>   src/amd/common/ac_nir_to_llvm.c | 11 +++++++++--
>   src/compiler/nir/nir_opcodes.py | 21 +++++++++++++++------
>   2 files changed, 24 insertions(+), 8 deletions(-)
>
> diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
> index 0c8891d26a0..12c4c21a8d9 100644
> --- a/src/amd/common/ac_nir_to_llvm.c
> +++ b/src/amd/common/ac_nir_to_llvm.c
> @@ -1081,10 +1081,17 @@ static void visit_alu(struct ac_nir_context *ctx, const nir_alu_instr *instr)
>   		LLVMValueRef in[3];
>   		for (unsigned chan = 0; chan < 3; chan++)
>   			in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
> -		results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
> +		results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
>   						ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
> -		results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
> +		results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
>   						ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
> +		LLVMValueRef ma = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubema",
> +						     ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
> +		results[0] = ac_build_fdiv(&ctx->ac, results[0], ma);
> +		results[1] = ac_build_fdiv(&ctx->ac, results[1], ma);
> +		LLVMValueRef offset = LLVMConstReal(ctx->ac.f32, 0.5);
> +		results[0] = LLVMBuildFAdd(ctx->ac.builder, results[0], offset, "");
> +		results[1] = LLVMBuildFAdd(ctx->ac.builder, results[1], offset, "");
>   		result = ac_build_gather_values(&ctx->ac, results, 2);
>   		break;
>   	}
> diff --git a/src/compiler/nir/nir_opcodes.py b/src/compiler/nir/nir_opcodes.py
> index 90f7aed0c0d..0f56dd9596c 100644
> --- a/src/compiler/nir/nir_opcodes.py
> +++ b/src/compiler/nir/nir_opcodes.py
> @@ -410,12 +410,21 @@ dst.x = dst.y = 0.0;
>   float absX = fabs(src0.x);
>   float absY = fabs(src0.y);
>   float absZ = fabs(src0.z);
> -if (src0.x >= 0 && absX >= absY && absX >= absZ) { dst.x = -src0.y; dst.y = -src0.z; }
> -if (src0.x < 0 && absX >= absY && absX >= absZ) { dst.x = -src0.y; dst.y = src0.z; }
> -if (src0.y >= 0 && absY >= absX && absY >= absZ) { dst.x = src0.z; dst.y = src0.x; }
> -if (src0.y < 0 && absY >= absX && absY >= absZ) { dst.x = -src0.z; dst.y = src0.x; }
> -if (src0.z >= 0 && absZ >= absX && absZ >= absY) { dst.x = -src0.y; dst.y = src0.x; }
> -if (src0.z < 0 && absZ >= absX && absZ >= absY) { dst.x = -src0.y; dst.y = -src0.x; }
> +
> +float ma = 0.0;
> +if (absX >= absY && absX >= absZ) { ma = 2 * src0.x; }
> +if (absY >= absX && absY >= absZ) { ma = 2 * src0.y; }
> +if (absZ >= absX && absZ >= absY) { ma = 2 * src0.z; }
> +
> +if (src0.x >= 0 && absX >= absY && absX >= absZ) { dst.x = -src0.z; dst.y = -src0.y; }
> +if (src0.x < 0 && absX >= absY && absX >= absZ) { dst.x = src0.z; dst.y = -src0.y; }
> +if (src0.y >= 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = src0.z; }
> +if (src0.y < 0 && absY >= absX && absY >= absZ) { dst.x = src0.x; dst.y = -src0.z; }
> +if (src0.z >= 0 && absZ >= absX && absZ >= absY) { dst.x = src0.x; dst.y = -src0.y; }
> +if (src0.z < 0 && absZ >= absX && absZ >= absY) { dst.x = -src0.x; dst.y = -src0.y; }
> +
> +dst.x = dst.x / ma + 0.5;
> +dst.y = dst.y / ma + 0.5;
>   """)
>   
>   unop_horiz("cube_face_index", 1, tfloat32, 3, tfloat32, """