[Mesa-dev] [PATCH 6/7] ac: handle subgroup intrinsics

[Michael Schellenberger Costa]

HI Daniel,


Am 08.03.2018 um 18:10 schrieb Daniel Schürmann:
> ---
>   src/amd/common/ac_nir_to_llvm.c | 66 +++++++++++++++++++++++------------------
>   1 file changed, 37 insertions(+), 29 deletions(-)
>
> diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
> index 9b85069860..0f4cc32f15 100644
> --- a/src/amd/common/ac_nir_to_llvm.c
> +++ b/src/amd/common/ac_nir_to_llvm.c
> @@ -4363,36 +4363,12 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
>   		result = ac_build_ballot(&ctx->ac, get_src(ctx, instr->src[0]));
>   		break;
>   	case nir_intrinsic_read_invocation:
> -	case nir_intrinsic_read_first_invocation: {
> -		LLVMValueRef args[2];
> -
> -		/* Value */
> -		args[0] = get_src(ctx, instr->src[0]);
> -
> -		unsigned num_args;
> -		const char *intr_name;
> -		if (instr->intrinsic == nir_intrinsic_read_invocation) {
> -			num_args = 2;
> -			intr_name = "llvm.amdgcn.readlane";
> -
> -			/* Invocation */
> -			args[1] = get_src(ctx, instr->src[1]);
> -		} else {
> -			num_args = 1;
> -			intr_name = "llvm.amdgcn.readfirstlane";
> -		}
> -
> -		/* We currently have no other way to prevent LLVM from lifting the icmp
> -		 * calls to a dominating basic block.
> -		 */
> -		ac_build_optimization_barrier(&ctx->ac, &args[0]);
> -
> -		result = ac_build_intrinsic(&ctx->ac, intr_name,
> -					    ctx->ac.i32, args, num_args,
> -					    AC_FUNC_ATTR_READNONE |
> -					    AC_FUNC_ATTR_CONVERGENT);
> +		result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]),
> +				get_src(ctx, instr->src[1]));
> +		break;
> +	case nir_intrinsic_read_first_invocation:
> +		result = ac_build_readlane(&ctx->ac, get_src(ctx, instr->src[0]), NULL);
>   		break;
> -	}
>   	case nir_intrinsic_load_subgroup_invocation:
>   		result = ac_get_thread_id(&ctx->ac);
>   		break;
> @@ -4646,6 +4622,38 @@ static void visit_intrinsic(struct ac_nir_context *ctx,
>   		result = LLVMBuildSExt(ctx->ac.builder, tmp, ctx->ac.i32, "");
>   		break;
>   	}
> +	case nir_intrinsic_shuffle:
> +		result = ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]),
> +				get_src(ctx, instr->src[1]));
> +		break;
> +	case nir_intrinsic_reduce:
> +		result = ac_build_reduce(&ctx->ac,
> +				get_src(ctx, instr->src[0]),
> +				instr->const_index[0],
Identation here and below is off. Took me a second look to see 
instr->const_index[0] is just the argument.
--Michael
> +				instr->const_index[1]);
> +		break;
> +	case nir_intrinsic_inclusive_scan:
> +		result = ac_build_inclusive_scan(&ctx->ac,
> +				get_src(ctx, instr->src[0]),
> +				instr->const_index[0]);
> +		break;
> +	case nir_intrinsic_exclusive_scan:
> +		result = ac_build_exclusive_scan(&ctx->ac,
> +				get_src(ctx, instr->src[0]),
> +				instr->const_index[0]);
> +		break;
> +	case nir_intrinsic_quad_broadcast: {
> +		unsigned lane = instr->const_index[0];
> +		result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
> +				lane, lane, lane, lane);
> +		break;
> +	}
> +	case nir_intrinsic_quad_swap_horizontal:
> +	case nir_intrinsic_quad_swap_vertical:
> +	case nir_intrinsic_quad_swap_diagonal:
> +		result = ac_build_quad_swap(&ctx->ac, get_src(ctx, instr->src[0]),
> +				instr->const_index[0]);
> +		break;
>   	default:
>   		fprintf(stderr, "Unknown intrinsic: ");
>   		nir_print_instr(&instr->instr, stderr);