From 8ec60d3031ca14d2429a2eb9dfc8df2c9fd21245 Mon Sep 17 00:00:00 2001 From: =?utf8?q?Nicolai=20H=C3=A4hnle?= Date: Tue, 7 May 2019 23:27:24 +0200 Subject: [PATCH] radeonsi/gfx10: add as_ngg shader key bit Also add the shader main part NGG variant, so that in principle we can switch between legacy in NGG modes. Acked-by: Bas Nieuwenhuizen --- src/gallium/drivers/radeonsi/si_shader.c | 5 ++++ src/gallium/drivers/radeonsi/si_shader.h | 7 ++++- src/gallium/drivers/radeonsi/si_state_shaders.c | 40 +++++++++++++++++++------ 3 files changed, 42 insertions(+), 10 deletions(-) diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c index 9346398..ba79af3 100644 --- a/src/gallium/drivers/radeonsi/si_shader.c +++ b/src/gallium/drivers/radeonsi/si_shader.c @@ -6080,6 +6080,7 @@ static void si_get_vs_prolog_key(const struct tgsi_shader_info *info, key->vs_prolog.last_input = MAX2(1, info->num_inputs) - 1; key->vs_prolog.as_ls = shader_out->key.as_ls; key->vs_prolog.as_es = shader_out->key.as_es; + key->vs_prolog.as_ngg = shader_out->key.as_ngg; if (shader_out->selector->type == PIPE_SHADER_TESS_CTRL) { key->vs_prolog.as_ls = 1; @@ -7046,6 +7047,7 @@ si_get_shader_part(struct si_screen *sscreen, case PIPE_SHADER_VERTEX: shader.key.as_ls = key->vs_prolog.as_ls; shader.key.as_es = key->vs_prolog.as_es; + shader.key.as_ngg = key->vs_prolog.as_ngg; break; case PIPE_SHADER_TESS_CTRL: assert(!prolog); @@ -7974,6 +7976,9 @@ bool si_shader_create(struct si_screen *sscreen, struct ac_llvm_compiler *compil /* LS, ES, VS are compiled on demand if the main part hasn't been * compiled for that stage. * + * GS are compiled on demand if the main part hasn't been compiled + * for the chosen NGG-ness. + * * Vertex shaders are compiled on demand when a vertex fetch * workaround must be applied. */ diff --git a/src/gallium/drivers/radeonsi/si_shader.h b/src/gallium/drivers/radeonsi/si_shader.h index 43d1b0d..f85a61b 100644 --- a/src/gallium/drivers/radeonsi/si_shader.h +++ b/src/gallium/drivers/radeonsi/si_shader.h @@ -331,6 +331,7 @@ struct si_shader_selector { struct si_shader *main_shader_part; struct si_shader *main_shader_part_ls; /* as_ls is set in the key */ struct si_shader *main_shader_part_es; /* as_es is set in the key */ + struct si_shader *main_shader_part_ngg; /* as_ngg is set in the key */ struct si_shader *gs_copy_shader; @@ -469,6 +470,7 @@ union si_shader_part_key { unsigned last_input:4; unsigned as_ls:1; unsigned as_es:1; + unsigned as_ngg:1; /* Prologs for monolithic shaders shouldn't set EXEC. */ unsigned is_monolithic:1; } vs_prolog; @@ -524,11 +526,12 @@ struct si_shader_key { } ps; } part; - /* These two are initially set according to the NEXT_SHADER property, + /* These three are initially set according to the NEXT_SHADER property, * or guessed if the property doesn't seem correct. */ unsigned as_es:1; /* export shader, which precedes GS */ unsigned as_ls:1; /* local shader, which precedes TCS */ + unsigned as_ngg:1; /* VS, TES, or GS compiled as NGG primitive shader */ /* Flags for monolithic compilation only. */ struct { @@ -753,6 +756,8 @@ si_get_main_shader_part(struct si_shader_selector *sel, return &sel->main_shader_part_ls; if (key->as_es) return &sel->main_shader_part_es; + if (key->as_ngg) + return &sel->main_shader_part_ngg; return &sel->main_shader_part; } diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c index 297be8e..e428b29 100644 --- a/src/gallium/drivers/radeonsi/si_state_shaders.c +++ b/src/gallium/drivers/radeonsi/si_state_shaders.c @@ -1454,6 +1454,7 @@ static void si_shader_selector_key_hw_vs(struct si_context *sctx, /* Compute the key for the hw shader variant */ static inline void si_shader_selector_key(struct pipe_context *ctx, struct si_shader_selector *sel, + union si_vgt_stages_key stages_key, struct si_shader_key *key) { struct si_context *sctx = (struct si_context *)ctx; @@ -1469,6 +1470,7 @@ static inline void si_shader_selector_key(struct pipe_context *ctx, else if (sctx->gs_shader.cso) key->as_es = 1; else { + key->as_ngg = stages_key.u.ngg; si_shader_selector_key_hw_vs(sctx, sel, key); if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid) @@ -1510,6 +1512,7 @@ static inline void si_shader_selector_key(struct pipe_context *ctx, if (sctx->gs_shader.cso) key->as_es = 1; else { + key->as_ngg = stages_key.u.ngg; si_shader_selector_key_hw_vs(sctx, sel, key); if (sctx->ps_shader.cso && sctx->ps_shader.cso->info.uses_primid) @@ -1527,6 +1530,8 @@ static inline void si_shader_selector_key(struct pipe_context *ctx, key->part.gs.prolog.gfx9_prev_is_vs = 1; } + key->as_ngg = stages_key.u.ngg; + /* Merged ES-GS can have unbalanced wave usage. * * ES threads are per-vertex, while GS threads are @@ -1765,6 +1770,7 @@ static bool si_check_missing_main_part(struct si_screen *sscreen, main_part->selector = sel; main_part->key.as_es = key->as_es; main_part->key.as_ls = key->as_ls; + main_part->key.as_ngg = key->as_ngg; main_part->is_monolithic = false; if (si_compile_tgsi_shader(sscreen, compiler_state->compiler, @@ -1898,12 +1904,15 @@ current_not_ready: */ if (!is_pure_monolithic && !key->opt.vs_as_prim_discard_cs) { - bool ok; + bool ok = true; /* Make sure the main shader part is present. This is needed * for shaders that can be compiled as VS, LS, or ES, and only * one of them is compiled at creation. * + * It is also needed for GS, which can be compiled as non-NGG + * and NGG. + * * For merged shaders, check that the starting shader's main * part is present. */ @@ -1922,10 +1931,13 @@ current_not_ready: previous_stage_sel, compiler_state, &shader1_key); mtx_unlock(&previous_stage_sel->mutex); - } else { + } + + if (ok) { ok = si_check_missing_main_part(sscreen, sel, compiler_state, key); } + if (!ok) { FREE(shader); mtx_unlock(&sel->mutex); @@ -2008,12 +2020,13 @@ current_not_ready: static int si_shader_select(struct pipe_context *ctx, struct si_shader_ctx_state *state, + union si_vgt_stages_key stages_key, struct si_compiler_ctx_state *compiler_state) { struct si_context *sctx = (struct si_context *)ctx; struct si_shader_key key; - si_shader_selector_key(ctx, state->cso, &key); + si_shader_selector_key(ctx, state->cso, stages_key, &key); return si_shader_select_with_key(sctx->screen, state, compiler_state, &key, -1, false); } @@ -2095,6 +2108,13 @@ static void si_init_shader_selector_async(void *job, int thread_index) si_parse_next_shader_property(&sel->info, sel->so.num_outputs != 0, &shader->key); + if (sscreen->info.chip_class >= GFX10 && + !sscreen->options.disable_ngg && + (((sel->type == PIPE_SHADER_VERTEX || + sel->type == PIPE_SHADER_TESS_EVAL) && + !shader->key.as_ls && !shader->key.as_es) || + sel->type == PIPE_SHADER_GEOMETRY)) + shader->key.as_ngg = 1; if (sel->tokens || sel->nir) ir_binary = si_get_ir_binary(sel); @@ -2808,6 +2828,8 @@ void si_destroy_shader_selector(struct si_context *sctx, si_delete_shader(sctx, sel->main_shader_part_ls); if (sel->main_shader_part_es) si_delete_shader(sctx, sel->main_shader_part_es); + if (sel->main_shader_part_ngg) + si_delete_shader(sctx, sel->main_shader_part_ngg); if (sel->gs_copy_shader) si_delete_shader(sctx, sel->gs_copy_shader); @@ -3435,7 +3457,7 @@ bool si_update_shaders(struct si_context *sctx) } if (sctx->tcs_shader.cso) { - r = si_shader_select(ctx, &sctx->tcs_shader, + r = si_shader_select(ctx, &sctx->tcs_shader, key, &compiler_state); if (r) return false; @@ -3449,7 +3471,7 @@ bool si_update_shaders(struct si_context *sctx) } r = si_shader_select(ctx, &sctx->fixed_func_tcs_shader, - &compiler_state); + key, &compiler_state); if (r) return false; si_pm4_bind_state(sctx, hs, @@ -3457,7 +3479,7 @@ bool si_update_shaders(struct si_context *sctx) } if (!sctx->gs_shader.cso || sctx->chip_class <= GFX8) { - r = si_shader_select(ctx, &sctx->tes_shader, &compiler_state); + r = si_shader_select(ctx, &sctx->tes_shader, key, &compiler_state); if (r) return false; @@ -3479,7 +3501,7 @@ bool si_update_shaders(struct si_context *sctx) /* Update GS. */ if (sctx->gs_shader.cso) { - r = si_shader_select(ctx, &sctx->gs_shader, &compiler_state); + r = si_shader_select(ctx, &sctx->gs_shader, key, &compiler_state); if (r) return false; si_pm4_bind_state(sctx, gs, sctx->gs_shader.current->pm4); @@ -3501,7 +3523,7 @@ bool si_update_shaders(struct si_context *sctx) /* Update VS. */ if ((!key.u.tess && !key.u.gs) || sctx->chip_class <= GFX8) { - r = si_shader_select(ctx, &sctx->vs_shader, &compiler_state); + r = si_shader_select(ctx, &sctx->vs_shader, key, &compiler_state); if (r) return false; @@ -3528,7 +3550,7 @@ bool si_update_shaders(struct si_context *sctx) if (sctx->ps_shader.cso) { unsigned db_shader_control; - r = si_shader_select(ctx, &sctx->ps_shader, &compiler_state); + r = si_shader_select(ctx, &sctx->ps_shader, key, &compiler_state); if (r) return false; si_pm4_bind_state(sctx, ps, sctx->ps_shader.current->pm4); -- 2.7.4