radeonsi: precompute IA_MULTI_VGT_PARAM values into a table
authorMarek Olšák <marek.olsak@amd.com>
Wed, 25 Jan 2017 01:47:15 +0000 (02:47 +0100)
committerMarek Olšák <marek.olsak@amd.com>
Mon, 30 Jan 2017 12:27:14 +0000 (13:27 +0100)
The perf difference is very small: 0.99% -> 0.40% for the time spent
in si_get_ia_multi_vgt_param when si_draw_vbo is 20%. Pretty much nothing.

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
src/gallium/drivers/radeonsi/si_pipe.c
src/gallium/drivers/radeonsi/si_pipe.h
src/gallium/drivers/radeonsi/si_state.c
src/gallium/drivers/radeonsi/si_state.h
src/gallium/drivers/radeonsi/si_state_draw.c
src/gallium/drivers/radeonsi/si_state_shaders.c

index c1480d4bf277767986c8950d7c75fa15c7fae56d..d4d156b8a69a07d70939b2ffffe7744331c85d88 100644 (file)
@@ -234,6 +234,7 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen,
        si_init_all_descriptors(sctx);
        si_init_state_functions(sctx);
        si_init_shader_functions(sctx);
+       si_init_ia_multi_vgt_param_table(sctx);
 
        if (sctx->b.chip_class >= CIK)
                cik_init_sdma_functions(sctx);
index 18cd25c60c1fccc10ed5cba1b3eb4748937a957c..724d89e747589df35725734cba8a4d03d557c2bb 100644 (file)
@@ -204,6 +204,28 @@ struct si_shader_ctx_state {
        struct si_shader                *current;
 };
 
+#define SI_NUM_VGT_PARAM_KEY_BITS 12
+#define SI_NUM_VGT_PARAM_STATES (1 << SI_NUM_VGT_PARAM_KEY_BITS)
+
+/* The IA_MULTI_VGT_PARAM key used to index the table of precomputed values.
+ * Some fields are set by state-change calls, most are set by draw_vbo.
+ */
+union si_vgt_param_key {
+       struct {
+               unsigned prim:4;
+               unsigned uses_instancing:1;
+               unsigned multi_instances_smaller_than_primgroup:1;
+               unsigned primitive_restart:1;
+               unsigned count_from_stream_output:1;
+               unsigned line_stipple_enabled:1;
+               unsigned uses_tess:1;
+               unsigned tcs_tes_uses_prim_id:1;
+               unsigned uses_gs:1;
+               unsigned _pad:32 - SI_NUM_VGT_PARAM_KEY_BITS;
+       } u;
+       uint32_t index;
+};
+
 struct si_context {
        struct r600_common_context      b;
        struct blitter_context          *blitter;
@@ -355,6 +377,10 @@ struct si_context {
 
        /* Other state */
        bool need_check_render_feedback;
+
+       /* Precomputed IA_MULTI_VGT_PARAM */
+       union si_vgt_param_key  ia_multi_vgt_param_key;
+       unsigned                ia_multi_vgt_param[SI_NUM_VGT_PARAM_STATES];
 };
 
 /* cik_sdma.c */
index 359058b99a285a4b4549ba9be744a65fe5e22828..de300764e3e4a72e6a7b60991642cba9bae8d89c 100644 (file)
@@ -918,6 +918,8 @@ static void si_bind_rs_state(struct pipe_context *ctx, void *state)
        si_update_poly_offset_state(sctx);
 
        si_mark_atom_dirty(sctx, &sctx->clip_regs);
+       sctx->ia_multi_vgt_param_key.u.line_stipple_enabled =
+               rs->line_stipple_enable;
        sctx->do_update_shaders = true;
 }
 
index bdcfb5b2a3f727d394f56c0e3a07c021e86fe92d..19880c5cfdd57c088022c106bbf1522a79088f65 100644 (file)
@@ -355,6 +355,7 @@ void si_destroy_shader_cache(struct si_screen *sscreen);
 void si_init_shader_selector_async(void *job, int thread_index);
 
 /* si_state_draw.c */
+void si_init_ia_multi_vgt_param_table(struct si_context *sctx);
 void si_emit_cache_flush(struct si_context *sctx);
 void si_ce_pre_draw_synchronization(struct si_context *sctx);
 void si_ce_post_draw_synchronization(struct si_context *sctx);
index 5313a4d6db4cc602b91f3277c9270c4273687bb8..c82ed3a131273e97576a6bbddaf3106c54e001b0 100644 (file)
@@ -261,13 +261,11 @@ static unsigned si_num_prims_for_vertices(const struct pipe_draw_info *info)
        }
 }
 
-static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
-                                         const struct pipe_draw_info *info,
-                                         unsigned num_patches)
+static unsigned
+si_get_init_multi_vgt_param(struct si_screen *sscreen,
+                           union si_vgt_param_key *key)
 {
-       struct si_state_rasterizer *rs = sctx->queued.named.rasterizer;
-       unsigned prim = info->mode;
-       unsigned primgroup_size = 128; /* recommended without a GS */
+       STATIC_ASSERT(sizeof(union si_vgt_param_key) == 4);
        unsigned max_primgroup_in_wave = 2;
 
        /* SWITCH_ON_EOP(0) is always preferable. */
@@ -277,35 +275,28 @@ static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
        bool partial_vs_wave = false;
        bool partial_es_wave = false;
 
-       if (sctx->gs_shader.cso)
-               primgroup_size = 64; /* recommended with a GS */
-
-       if (sctx->tes_shader.cso) {
-               /* primgroup_size must be set to a multiple of NUM_PATCHES */
-               primgroup_size = num_patches;
-
+       if (key->u.uses_tess) {
                /* SWITCH_ON_EOI must be set if PrimID is used. */
-               if ((sctx->tcs_shader.cso && sctx->tcs_shader.cso->info.uses_primid) ||
-                   sctx->tes_shader.cso->info.uses_primid)
+               if (key->u.tcs_tes_uses_prim_id)
                        ia_switch_on_eoi = true;
 
                /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */
-               if ((sctx->b.family == CHIP_TAHITI ||
-                    sctx->b.family == CHIP_PITCAIRN ||
-                    sctx->b.family == CHIP_BONAIRE) &&
-                   sctx->gs_shader.cso)
+               if ((sscreen->b.family == CHIP_TAHITI ||
+                    sscreen->b.family == CHIP_PITCAIRN ||
+                    sscreen->b.family == CHIP_BONAIRE) &&
+                   key->u.uses_gs)
                        partial_vs_wave = true;
 
                /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */
-               if (sctx->screen->has_distributed_tess) {
-                       if (sctx->gs_shader.cso) {
+               if (sscreen->has_distributed_tess) {
+                       if (key->u.uses_gs) {
                                partial_es_wave = true;
 
                                /* GPU hang workaround. */
-                               if (sctx->b.family == CHIP_TONGA ||
-                                   sctx->b.family == CHIP_FIJI ||
-                                   sctx->b.family == CHIP_POLARIS10 ||
-                                   sctx->b.family == CHIP_POLARIS11)
+                               if (sscreen->b.family == CHIP_TONGA ||
+                                   sscreen->b.family == CHIP_FIJI ||
+                                   sscreen->b.family == CHIP_POLARIS10 ||
+                                   sscreen->b.family == CHIP_POLARIS11)
                                        partial_vs_wave = true;
                        } else {
                                partial_vs_wave = true;
@@ -314,13 +305,13 @@ static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
        }
 
        /* This is a hardware requirement. */
-       if ((rs && rs->line_stipple_enable) ||
-           (sctx->b.screen->debug_flags & DBG_SWITCH_ON_EOP)) {
+       if (key->u.line_stipple_enabled ||
+           (sscreen->b.debug_flags & DBG_SWITCH_ON_EOP)) {
                ia_switch_on_eop = true;
                wd_switch_on_eop = true;
        }
 
-       if (sctx->b.chip_class >= CIK) {
+       if (sscreen->b.chip_class >= CIK) {
                /* WD_SWITCH_ON_EOP has no effect on GPUs with less than
                 * 4 shader engines. Set 1 to pass the assertion below.
                 * The other cases are hardware requirements.
@@ -328,24 +319,24 @@ static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
                 * Polaris supports primitive restart with WD_SWITCH_ON_EOP=0
                 * for points, line strips, and tri strips.
                 */
-               if (sctx->b.screen->info.max_se < 4 ||
-                   prim == PIPE_PRIM_POLYGON ||
-                   prim == PIPE_PRIM_LINE_LOOP ||
-                   prim == PIPE_PRIM_TRIANGLE_FAN ||
-                   prim == PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY ||
-                   (info->primitive_restart &&
-                    (sctx->b.family < CHIP_POLARIS10 ||
-                     (prim != PIPE_PRIM_POINTS &&
-                      prim != PIPE_PRIM_LINE_STRIP &&
-                      prim != PIPE_PRIM_TRIANGLE_STRIP))) ||
-                   info->count_from_stream_output)
+               if (sscreen->b.info.max_se < 4 ||
+                   key->u.prim == PIPE_PRIM_POLYGON ||
+                   key->u.prim == PIPE_PRIM_LINE_LOOP ||
+                   key->u.prim == PIPE_PRIM_TRIANGLE_FAN ||
+                   key->u.prim == PIPE_PRIM_TRIANGLE_STRIP_ADJACENCY ||
+                   (key->u.primitive_restart &&
+                    (sscreen->b.family < CHIP_POLARIS10 ||
+                     (key->u.prim != PIPE_PRIM_POINTS &&
+                      key->u.prim != PIPE_PRIM_LINE_STRIP &&
+                      key->u.prim != PIPE_PRIM_TRIANGLE_STRIP))) ||
+                   key->u.count_from_stream_output)
                        wd_switch_on_eop = true;
 
                /* Hawaii hangs if instancing is enabled and WD_SWITCH_ON_EOP is 0.
                 * We don't know that for indirect drawing, so treat it as
                 * always problematic. */
-               if (sctx->b.family == CHIP_HAWAII &&
-                   (info->indirect || info->instance_count > 1))
+               if (sscreen->b.family == CHIP_HAWAII &&
+                   key->u.uses_instancing)
                        wd_switch_on_eop = true;
 
                /* Performance recommendation for 4 SE Gfx7-8 parts if
@@ -353,42 +344,27 @@ static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
                 * Assume indirect draws always use small instances.
                 * This is needed for good VS wave utilization.
                 */
-               if (sctx->b.chip_class <= VI &&
-                   sctx->b.screen->info.max_se >= 4 &&
-                   (info->indirect ||
-                    (info->instance_count > 1 &&
-                     si_num_prims_for_vertices(info) < primgroup_size)))
+               if (sscreen->b.chip_class <= VI &&
+                   sscreen->b.info.max_se == 4 &&
+                   key->u.multi_instances_smaller_than_primgroup)
                        wd_switch_on_eop = true;
 
                /* Required on CIK and later. */
-               if (sctx->b.screen->info.max_se > 2 && !wd_switch_on_eop)
+               if (sscreen->b.info.max_se > 2 && !wd_switch_on_eop)
                        ia_switch_on_eoi = true;
 
                /* Required by Hawaii and, for some special cases, by VI. */
                if (ia_switch_on_eoi &&
-                   (sctx->b.family == CHIP_HAWAII ||
-                    (sctx->b.chip_class == VI &&
-                     (sctx->gs_shader.cso || max_primgroup_in_wave != 2))))
+                   (sscreen->b.family == CHIP_HAWAII ||
+                    (sscreen->b.chip_class == VI &&
+                     (key->u.uses_gs || max_primgroup_in_wave != 2))))
                        partial_vs_wave = true;
 
                /* Instancing bug on Bonaire. */
-               if (sctx->b.family == CHIP_BONAIRE && ia_switch_on_eoi &&
-                   (info->indirect || info->instance_count > 1))
+               if (sscreen->b.family == CHIP_BONAIRE && ia_switch_on_eoi &&
+                   key->u.uses_instancing)
                        partial_vs_wave = true;
 
-               /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
-                * The hw doc says all multi-SE chips are affected, but Vulkan
-                * only applies it to Hawaii. Do what Vulkan does.
-                */
-               if (sctx->b.family == CHIP_HAWAII &&
-                   sctx->gs_shader.cso &&
-                   ia_switch_on_eoi &&
-                   (info->indirect ||
-                    (info->instance_count > 1 &&
-                     si_num_prims_for_vertices(info) <= 1)))
-                       sctx->b.flags |= SI_CONTEXT_VGT_FLUSH;
-
-
                /* If the WD switch is false, the IA switch must be false too. */
                assert(wd_switch_on_eop || !ia_switch_on_eop);
        }
@@ -397,20 +373,92 @@ static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
        if (ia_switch_on_eoi)
                partial_es_wave = true;
 
-       /* GS requirement. */
-       if (SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
-               partial_es_wave = true;
-
        return S_028AA8_SWITCH_ON_EOP(ia_switch_on_eop) |
                S_028AA8_SWITCH_ON_EOI(ia_switch_on_eoi) |
                S_028AA8_PARTIAL_VS_WAVE_ON(partial_vs_wave) |
                S_028AA8_PARTIAL_ES_WAVE_ON(partial_es_wave) |
-               S_028AA8_PRIMGROUP_SIZE(primgroup_size - 1) |
-               S_028AA8_WD_SWITCH_ON_EOP(sctx->b.chip_class >= CIK ? wd_switch_on_eop : 0) |
-               S_028AA8_MAX_PRIMGRP_IN_WAVE(sctx->b.chip_class >= VI ?
+               S_028AA8_WD_SWITCH_ON_EOP(sscreen->b.chip_class >= CIK ? wd_switch_on_eop : 0) |
+               S_028AA8_MAX_PRIMGRP_IN_WAVE(sscreen->b.chip_class >= VI ?
                                             max_primgroup_in_wave : 0);
 }
 
+void si_init_ia_multi_vgt_param_table(struct si_context *sctx)
+{
+       for (int prim = 0; prim <= R600_PRIM_RECTANGLE_LIST; prim++)
+       for (int uses_instancing = 0; uses_instancing < 2; uses_instancing++)
+       for (int multi_instances = 0; multi_instances < 2; multi_instances++)
+       for (int primitive_restart = 0; primitive_restart < 2; primitive_restart++)
+       for (int count_from_so = 0; count_from_so < 2; count_from_so++)
+       for (int line_stipple = 0; line_stipple < 2; line_stipple++)
+       for (int uses_tess = 0; uses_tess < 2; uses_tess++)
+       for (int tess_uses_primid = 0; tess_uses_primid < 2; tess_uses_primid++)
+       for (int uses_gs = 0; uses_gs < 2; uses_gs++) {
+               union si_vgt_param_key key;
+
+               key.index = 0;
+               key.u.prim = prim;
+               key.u.uses_instancing = uses_instancing;
+               key.u.multi_instances_smaller_than_primgroup = multi_instances;
+               key.u.primitive_restart = primitive_restart;
+               key.u.count_from_stream_output = count_from_so;
+               key.u.line_stipple_enabled = line_stipple;
+               key.u.uses_tess = uses_tess;
+               key.u.tcs_tes_uses_prim_id = tess_uses_primid;
+               key.u.uses_gs = uses_gs;
+
+               sctx->ia_multi_vgt_param[key.index] =
+                       si_get_init_multi_vgt_param(sctx->screen, &key);
+       }
+}
+
+static unsigned si_get_ia_multi_vgt_param(struct si_context *sctx,
+                                         const struct pipe_draw_info *info,
+                                         unsigned num_patches)
+{
+       union si_vgt_param_key key = sctx->ia_multi_vgt_param_key;
+       unsigned primgroup_size;
+       unsigned ia_multi_vgt_param;
+
+       if (sctx->tes_shader.cso) {
+               primgroup_size = num_patches; /* must be a multiple of NUM_PATCHES */
+       } else if (sctx->gs_shader.cso) {
+               primgroup_size = 64; /* recommended with a GS */
+       } else {
+               primgroup_size = 128; /* recommended without a GS and tess */
+       }
+
+       key.u.prim = info->mode;
+       key.u.uses_instancing = info->indirect || info->instance_count > 1;
+       key.u.multi_instances_smaller_than_primgroup =
+               info->indirect ||
+               (info->instance_count > 1 &&
+                si_num_prims_for_vertices(info) < primgroup_size);
+       key.u.primitive_restart = info->primitive_restart;
+       key.u.count_from_stream_output = info->count_from_stream_output != NULL;
+
+       ia_multi_vgt_param = sctx->ia_multi_vgt_param[key.index] |
+                            S_028AA8_PRIMGROUP_SIZE(primgroup_size - 1);
+
+       if (sctx->gs_shader.cso) {
+               /* GS requirement. */
+               if (SI_GS_PER_ES / primgroup_size >= sctx->screen->gs_table_depth - 3)
+                       ia_multi_vgt_param |= S_028AA8_PARTIAL_ES_WAVE_ON(1);
+
+               /* GS hw bug with single-primitive instances and SWITCH_ON_EOI.
+                * The hw doc says all multi-SE chips are affected, but Vulkan
+                * only applies it to Hawaii. Do what Vulkan does.
+                */
+               if (sctx->b.family == CHIP_HAWAII &&
+                   G_028AA8_SWITCH_ON_EOI(ia_multi_vgt_param) &&
+                   (info->indirect ||
+                    (info->instance_count > 1 &&
+                     si_num_prims_for_vertices(info) <= 1)))
+                       sctx->b.flags |= SI_CONTEXT_VGT_FLUSH;
+       }
+
+       return ia_multi_vgt_param;
+}
+
 static void si_emit_scratch_reloc(struct si_context *sctx)
 {
        struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
index 94fffdb4263c5b97ddb5b9b1474cfd83cf69349e..ebd243545d4e935f853d1294f49c1dc144d51ceb 100644 (file)
@@ -1720,6 +1720,7 @@ static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
 
        sctx->gs_shader.cso = sel;
        sctx->gs_shader.current = sel ? sel->first_variant : NULL;
+       sctx->ia_multi_vgt_param_key.u.uses_gs = sel != NULL;
        sctx->do_update_shaders = true;
        si_mark_atom_dirty(sctx, &sctx->clip_regs);
        sctx->last_rast_prim = -1; /* reset this so that it gets updated */
@@ -1729,6 +1730,15 @@ static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
        r600_update_vs_writes_viewport_index(&sctx->b, si_get_vs_info(sctx));
 }
 
+static void si_update_tcs_tes_uses_prim_id(struct si_context *sctx)
+{
+       sctx->ia_multi_vgt_param_key.u.tcs_tes_uses_prim_id =
+               (sctx->tes_shader.cso &&
+                sctx->tes_shader.cso->info.uses_primid) ||
+               (sctx->tcs_shader.cso &&
+                sctx->tcs_shader.cso->info.uses_primid);
+}
+
 static void si_bind_tcs_shader(struct pipe_context *ctx, void *state)
 {
        struct si_context *sctx = (struct si_context *)ctx;
@@ -1740,6 +1750,7 @@ static void si_bind_tcs_shader(struct pipe_context *ctx, void *state)
 
        sctx->tcs_shader.cso = sel;
        sctx->tcs_shader.current = sel ? sel->first_variant : NULL;
+       si_update_tcs_tes_uses_prim_id(sctx);
        sctx->do_update_shaders = true;
 
        if (enable_changed)
@@ -1757,6 +1768,8 @@ static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
 
        sctx->tes_shader.cso = sel;
        sctx->tes_shader.current = sel ? sel->first_variant : NULL;
+       sctx->ia_multi_vgt_param_key.u.uses_tess = sel != NULL;
+       si_update_tcs_tes_uses_prim_id(sctx);
        sctx->do_update_shaders = true;
        si_mark_atom_dirty(sctx, &sctx->clip_regs);
        sctx->last_rast_prim = -1; /* reset this so that it gets updated */