diff options
| author | Samuel Pitoiset <[email protected]> | 2020-03-23 12:02:15 +0100 |
|---|---|---|
| committer | Marge Bot <[email protected]> | 2020-03-23 14:19:03 +0000 |
| commit | 7ac8bb33cd6025f805a390e7647506e932f4db0d (patch) | |
| tree | 4207815186c10d6e4352e2f05e5ccc8b4d9d4aa6 /src/amd/llvm | |
| parent | 2a70a1d69d3151e6c95111a297e715e887692ce3 (diff) | |
radv/llvm: fix subgroup shuffle for chips without bpermute
bpermute only exists on GFX8+ and only with Wave32 on GFX10. Instead
we have to use readlane with a waterfall loop to defeat the LLVM
backend.
This fixes DOOM Eternal which requires subgroup shuffle.
Cc: <[email protected]>
Signed-off-by: Samuel Pitoiset <[email protected]>
Reviewed-by: Bas Nieuwenhuizen <[email protected]>
Tested-by: Marge Bot <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4284>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4284>
Diffstat (limited to 'src/amd/llvm')
| -rw-r--r-- | src/amd/llvm/ac_nir_to_llvm.c | 29 |
1 files changed, 27 insertions, 2 deletions
diff --git a/src/amd/llvm/ac_nir_to_llvm.c b/src/amd/llvm/ac_nir_to_llvm.c index 49627990163..871c6abc17f 100644 --- a/src/amd/llvm/ac_nir_to_llvm.c +++ b/src/amd/llvm/ac_nir_to_llvm.c @@ -3950,8 +3950,33 @@ static void visit_intrinsic(struct ac_nir_context *ctx, break; } case nir_intrinsic_shuffle: - result = ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]), - get_src(ctx, instr->src[1])); + if (ctx->ac.chip_class == GFX8 || + ctx->ac.chip_class == GFX9 || + (ctx->ac.chip_class == GFX10 && ctx->ac.wave_size == 32)) { + result = ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]), + get_src(ctx, instr->src[1])); + } else { + LLVMValueRef src = get_src(ctx, instr->src[0]); + LLVMValueRef index = get_src(ctx, instr->src[1]); + LLVMTypeRef type = LLVMTypeOf(src); + struct waterfall_context wctx; + LLVMValueRef index_val; + + index_val = enter_waterfall(ctx, &wctx, index, true); + + src = LLVMBuildZExt(ctx->ac.builder, src, + ctx->ac.i32, ""); + + result = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.readlane", + ctx->ac.i32, + (LLVMValueRef []) { src, index_val }, 2, + AC_FUNC_ATTR_READNONE | + AC_FUNC_ATTR_CONVERGENT); + + result = LLVMBuildTrunc(ctx->ac.builder, result, type, ""); + + result = exit_waterfall(ctx, &wctx, result); + } break; case nir_intrinsic_reduce: result = ac_build_reduce(&ctx->ac, |