From: Dave Airlie Date: Wed, 1 Feb 2017 23:35:40 +0000 (+1000) Subject: radeonsi/ac: move get thread id to shared code. X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=c5f0a56aeba546dc6b9e62ba12bd77d552ac4f31;p=mesa.git radeonsi/ac: move get thread id to shared code. radv will use this. Reviewed-by: Bas Nieuwenhuizen Reviewed-by: Nicolai Hähnle Signed-off-by: Dave Airlie --- diff --git a/src/amd/common/ac_llvm_util.c b/src/amd/common/ac_llvm_util.c index bc309bd81c4..4c114585292 100644 --- a/src/amd/common/ac_llvm_util.c +++ b/src/amd/common/ac_llvm_util.c @@ -168,6 +168,9 @@ ac_llvm_context_init(struct ac_llvm_context *ctx, LLVMContextRef context) ctx->v4f32 = LLVMVectorType(ctx->f32, 4); ctx->v16i8 = LLVMVectorType(ctx->i8, 16); + ctx->range_md_kind = LLVMGetMDKindIDInContext(ctx->context, + "range", 5); + ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(ctx->context, "invariant.load", 14); @@ -803,3 +806,48 @@ ac_build_buffer_load(struct ac_llvm_context *ctx, ARRAY_SIZE(args), AC_FUNC_ATTR_READONLY); } } + +/** + * Set range metadata on an instruction. This can only be used on load and + * call instructions. If you know an instruction can only produce the values + * 0, 1, 2, you would do set_range_metadata(value, 0, 3); + * \p lo is the minimum value inclusive. + * \p hi is the maximum value exclusive. + */ +static void set_range_metadata(struct ac_llvm_context *ctx, + LLVMValueRef value, unsigned lo, unsigned hi) +{ + LLVMValueRef range_md, md_args[2]; + LLVMTypeRef type = LLVMTypeOf(value); + LLVMContextRef context = LLVMGetTypeContext(type); + + md_args[0] = LLVMConstInt(type, lo, false); + md_args[1] = LLVMConstInt(type, hi, false); + range_md = LLVMMDNodeInContext(context, md_args, 2); + LLVMSetMetadata(value, ctx->range_md_kind, range_md); +} + +LLVMValueRef +ac_get_thread_id(struct ac_llvm_context *ctx) +{ + LLVMValueRef tid; + + if (HAVE_LLVM < 0x0308) { + tid = ac_emit_llvm_intrinsic(ctx, "llvm.SI.tid", + ctx->i32, + NULL, 0, AC_FUNC_ATTR_READNONE); + } else { + LLVMValueRef tid_args[2]; + tid_args[0] = LLVMConstInt(ctx->i32, 0xffffffff, false); + tid_args[1] = LLVMConstInt(ctx->i32, 0, false); + tid_args[1] = ac_emit_llvm_intrinsic(ctx, + "llvm.amdgcn.mbcnt.lo", ctx->i32, + tid_args, 2, AC_FUNC_ATTR_READNONE); + + tid = ac_emit_llvm_intrinsic(ctx, "llvm.amdgcn.mbcnt.hi", + ctx->i32, tid_args, + 2, AC_FUNC_ATTR_READNONE); + } + set_range_metadata(ctx, tid, 0, 64); + return tid; +} diff --git a/src/amd/common/ac_llvm_util.h b/src/amd/common/ac_llvm_util.h index 7b08a269fcb..0244dc48b3d 100644 --- a/src/amd/common/ac_llvm_util.h +++ b/src/amd/common/ac_llvm_util.h @@ -57,7 +57,8 @@ struct ac_llvm_context { LLVMTypeRef v4i32; LLVMTypeRef v4f32; LLVMTypeRef v16i8; - + + unsigned range_md_kind; unsigned invariant_load_md_kind; unsigned uniform_md_kind; unsigned fpmath_md_kind; @@ -176,6 +177,9 @@ ac_build_buffer_load(struct ac_llvm_context *ctx, unsigned inst_offset, unsigned glc, unsigned slc); + +LLVMValueRef +ac_get_thread_id(struct ac_llvm_context *ctx); #ifdef __cplusplus } #endif diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index e4ed40e2581..e483fe4923a 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -1266,49 +1266,6 @@ static LLVMValueRef get_sample_id(struct si_shader_context *radeon_bld) SI_PARAM_ANCILLARY, 8, 4); } -/** - * Set range metadata on an instruction. This can only be used on load and - * call instructions. If you know an instruction can only produce the values - * 0, 1, 2, you would do set_range_metadata(value, 0, 3); - * \p lo is the minimum value inclusive. - * \p hi is the maximum value exclusive. - */ -static void set_range_metadata(struct si_shader_context *ctx, - LLVMValueRef value, unsigned lo, unsigned hi) -{ - LLVMValueRef range_md, md_args[2]; - LLVMTypeRef type = LLVMTypeOf(value); - LLVMContextRef context = LLVMGetTypeContext(type); - - md_args[0] = LLVMConstInt(type, lo, false); - md_args[1] = LLVMConstInt(type, hi, false); - range_md = LLVMMDNodeInContext(context, md_args, 2); - LLVMSetMetadata(value, ctx->range_md_kind, range_md); -} - -static LLVMValueRef get_thread_id(struct si_shader_context *ctx) -{ - struct gallivm_state *gallivm = &ctx->gallivm; - LLVMValueRef tid; - - if (HAVE_LLVM < 0x0308) { - tid = lp_build_intrinsic(gallivm->builder, "llvm.SI.tid", - ctx->i32, NULL, 0, LP_FUNC_ATTR_READNONE); - } else { - LLVMValueRef tid_args[2]; - tid_args[0] = lp_build_const_int32(gallivm, 0xffffffff); - tid_args[1] = lp_build_const_int32(gallivm, 0); - tid_args[1] = lp_build_intrinsic(gallivm->builder, - "llvm.amdgcn.mbcnt.lo", ctx->i32, - tid_args, 2, LP_FUNC_ATTR_READNONE); - - tid = lp_build_intrinsic(gallivm->builder, - "llvm.amdgcn.mbcnt.hi", ctx->i32, - tid_args, 2, LP_FUNC_ATTR_READNONE); - } - set_range_metadata(ctx, tid, 0, 64); - return tid; -} /** * Load a dword from a constant buffer. @@ -2069,7 +2026,7 @@ static void si_llvm_emit_streamout(struct si_shader_context *ctx, LLVMValueRef so_vtx_count = unpack_param(ctx, ctx->param_streamout_config, 16, 7); - LLVMValueRef tid = get_thread_id(ctx); + LLVMValueRef tid = ac_get_thread_id(&ctx->ac); /* can_emit = tid < so_vtx_count; */ LLVMValueRef can_emit = @@ -4806,7 +4763,7 @@ static void si_llvm_emit_ddxy( int idx; unsigned mask; - thread_id = get_thread_id(ctx); + thread_id = ac_get_thread_id(&ctx->ac); if (opcode == TGSI_OPCODE_DDX_FINE) mask = TID_MASK_LEFT; @@ -5237,14 +5194,6 @@ static void si_create_function(struct si_shader_context *ctx, } } -static void create_meta_data(struct si_shader_context *ctx) -{ - struct gallivm_state *gallivm = ctx->bld_base.base.gallivm; - - ctx->range_md_kind = LLVMGetMDKindIDInContext(gallivm->context, - "range", 5); -} - static void declare_streamout_params(struct si_shader_context *ctx, struct pipe_stream_output_info *so, LLVMTypeRef *params, LLVMTypeRef i32, @@ -6216,7 +6165,6 @@ si_generate_gs_copy_shader(struct si_screen *sscreen, builder = gallivm->builder; - create_meta_data(&ctx); create_function(&ctx); preload_ring_buffers(&ctx); @@ -6714,7 +6662,6 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, return false; } - create_meta_data(ctx); create_function(ctx); preload_ring_buffers(ctx);