diff options
author | Marek Olšák <[email protected]> | 2017-03-14 19:35:28 +0100 |
---|---|---|
committer | Marek Olšák <[email protected]> | 2017-04-28 21:47:35 +0200 |
commit | 0d6d25475d7c3ddc04349afee8e11b67d5ecf726 (patch) | |
tree | 5265e6b3425a546f2a2d11a4c3aa969e2e55be54 /src | |
parent | a84a6feac926beca0a5a6fdabf83350a509cdcf8 (diff) |
radeonsi/gfx9: set EXEC for non-mono merged shaders, add a barrier between them
Reviewed-by: Nicolai Hähnle <[email protected]>
Diffstat (limited to 'src')
-rw-r--r-- | src/gallium/drivers/radeonsi/si_shader.c | 43 |
1 files changed, 41 insertions, 2 deletions
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index d96d55917bd..126f7dc7e59 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -7079,9 +7079,22 @@ static void si_count_scratch_private_memory(struct si_shader_context *ctx) } } +static void si_init_exec_from_input(struct si_shader_context *ctx, + unsigned param, unsigned bitoffset) +{ + LLVMValueRef args[] = { + LLVMGetParam(ctx->main_fn, param), + LLVMConstInt(ctx->i32, bitoffset, 0), + }; + lp_build_intrinsic(ctx->gallivm.builder, + "llvm.amdgcn.init.exec.from.input", + ctx->voidt, args, 2, LP_FUNC_ATTR_CONVERGENT); +} + static bool si_compile_tgsi_main(struct si_shader_context *ctx, - struct si_shader *shader) + bool is_monolithic) { + struct si_shader *shader = ctx->shader; struct si_shader_selector *sel = shader->selector; struct lp_build_tgsi_context *bld_base = &ctx->bld_base; @@ -7127,6 +7140,29 @@ static bool si_compile_tgsi_main(struct si_shader_context *ctx, create_function(ctx); preload_ring_buffers(ctx); + /* For GFX9 merged shaders: + * - Set EXEC. If the prolog is present, set EXEC there instead. + * - Add a barrier before the second shader. + * + * The same thing for monolithic shaders is done in + * si_build_wrapper_function. + */ + if (ctx->screen->b.chip_class >= GFX9 && !is_monolithic) { + if (sel->info.num_instructions > 1 && /* not empty shader */ + (shader->key.as_es || shader->key.as_ls) && + (ctx->type == PIPE_SHADER_TESS_EVAL || + (ctx->type == PIPE_SHADER_VERTEX && + !sel->vs_needs_prolog))) { + si_init_exec_from_input(ctx, + ctx->param_merged_wave_info, 0); + } else if (ctx->type == PIPE_SHADER_TESS_CTRL || + ctx->type == PIPE_SHADER_GEOMETRY) { + si_init_exec_from_input(ctx, + ctx->param_merged_wave_info, 8); + si_llvm_emit_barrier(NULL, bld_base, NULL); + } + } + if (ctx->type == PIPE_SHADER_GEOMETRY) { int i; for (i = 0; i < 4; i++) { @@ -7642,7 +7678,7 @@ int si_compile_tgsi_shader(struct si_screen *sscreen, ctx.load_system_value = declare_system_value; - if (!si_compile_tgsi_main(&ctx, shader)) { + if (!si_compile_tgsi_main(&ctx, is_monolithic)) { si_llvm_dispose(&ctx); return -1; } @@ -7992,6 +8028,9 @@ static void si_build_vs_prolog_function(struct si_shader_context *ctx, num_params, last_sgpr); func = ctx->main_fn; + if (key->vs_prolog.num_merged_next_stage_vgprs) + si_init_exec_from_input(ctx, 3, 0); + /* Copy inputs to outputs. This should be no-op, as the registers match, * but it will prevent the compiler from overwriting them unintentionally. */ |