[Mesa-dev] [PATCH 2/2] amd/common/vi+: enable SMEM loads with GLC=1

[Bas Nieuwenhuizen]

Reviewed-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>

for both.


On Thu, Jan 10, 2019 at 11:26 PM Nicolai Hähnle <nhaehnle at gmail.com> wrote:
>
> From: Nicolai Hähnle <nicolai.haehnle at amd.com>
>
> Only on LLVM 8.0+, which supports the new intrinsic.
> ---
>  src/amd/common/ac_llvm_build.c | 10 +++++++---
>  1 file changed, 7 insertions(+), 3 deletions(-)
>
> diff --git a/src/amd/common/ac_llvm_build.c b/src/amd/common/ac_llvm_build.c
> index 4d7f15901e3..6aa96ee86d4 100644
> --- a/src/amd/common/ac_llvm_build.c
> +++ b/src/amd/common/ac_llvm_build.c
> @@ -1223,36 +1223,40 @@ ac_build_buffer_load(struct ac_llvm_context *ctx,
>                      unsigned slc,
>                      bool can_speculate,
>                      bool allow_smem)
>  {
>         LLVMValueRef offset = LLVMConstInt(ctx->i32, inst_offset, 0);
>         if (voffset)
>                 offset = LLVMBuildAdd(ctx->builder, offset, voffset, "");
>         if (soffset)
>                 offset = LLVMBuildAdd(ctx->builder, offset, soffset, "");
>
> -       /* TODO: VI and later generations can use SMEM with GLC=1.*/
> -       if (allow_smem && !glc && !slc) {
> +       if (allow_smem && !slc &&
> +           (!glc || (HAVE_LLVM >= 0x0800 && ctx->chip_class >= VI))) {
>                 assert(vindex == NULL);
>
>                 LLVMValueRef result[8];
>
>                 for (int i = 0; i < num_channels; i++) {
>                         if (i) {
>                                 offset = LLVMBuildAdd(ctx->builder, offset,
>                                                       LLVMConstInt(ctx->i32, 4, 0), "");
>                         }
>                         const char *intrname =
>                                 HAVE_LLVM >= 0x0800 ? "llvm.amdgcn.s.buffer.load.f32"
>                                                     : "llvm.SI.load.const";
>                         unsigned num_args = HAVE_LLVM >= 0x0800 ? 3 : 2;
> -                       LLVMValueRef args[3] = {rsrc, offset, ctx->i32_0};
> +                       LLVMValueRef args[3] = {
> +                               rsrc,
> +                               offset,
> +                               glc ? ctx->i32_1 : ctx->i32_0,
> +                       };
>                         result[i] = ac_build_intrinsic(ctx, intrname,
>                                                        ctx->f32, args, num_args,
>                                                        AC_FUNC_ATTR_READNONE |
>                                                        (HAVE_LLVM < 0x0800 ? AC_FUNC_ATTR_LEGACY : 0));
>                 }
>                 if (num_channels == 1)
>                         return result[0];
>
>                 if (num_channels == 3)
>                         result[num_channels++] = LLVMGetUndef(ctx->f32);
> --
> 2.19.1
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev