[Mesa-dev] [PATCH 12/17] radeonsi: precompute IA_MULTI_VGT_PARAM values into a table
Marek Olšák
maraeo at gmail.com
Thu Jan 26 16:04:28 UTC 2017
From: Marek Olšák <marek.olsak at amd.com>
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.
---
src/gallium/drivers/radeonsi/si_pipe.c | 1 +
src/gallium/drivers/radeonsi/si_pipe.h | 26 ++++
src/gallium/drivers/radeonsi/si_state.c | 2 +
src/gallium/drivers/radeonsi/si_state.h | 1 +
src/gallium/drivers/radeonsi/si_state_draw.c | 192 +++++++++++++++---------
src/gallium/drivers/radeonsi/si_state_shaders.c | 13 ++
6 files changed, 163 insertions(+), 72 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index 5de0781..8c54997 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -227,20 +227,21 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen,
sctx->border_color_map =
ws->buffer_map(sctx->border_color_buffer->buf,
NULL, PIPE_TRANSFER_WRITE);
if (!sctx->border_color_map)
goto fail;
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);
else
si_init_dma_functions(sctx);
if (sscreen->b.debug_flags & DBG_FORCE_DMA)
sctx->b.b.resource_copy_region = sctx->b.dma_copy;
sctx->blitter = util_blitter_create(&sctx->b.b);
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 18cd25c..724d89e 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -197,20 +197,42 @@ struct si_sample_mask {
/* A shader state consists of the shader selector, which is a constant state
* object shared by multiple contexts and shouldn't be modified, and
* the current shader variant selected for this context.
*/
struct si_shader_ctx_state {
struct si_shader_selector *cso;
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;
void *custom_dsa_flush;
void *custom_blend_resolve;
void *custom_blend_decompress;
void *custom_blend_fastclear;
void *custom_blend_dcc_decompress;
struct si_screen *screen;
@@ -348,20 +370,24 @@ struct si_context {
bool is_debug;
struct radeon_saved_cs last_gfx;
struct r600_resource *last_trace_buf;
struct r600_resource *trace_buf;
unsigned trace_id;
uint64_t dmesg_timestamp;
unsigned apitrace_call_number;
/* 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 */
void cik_init_sdma_functions(struct si_context *sctx);
/* si_blit.c */
void si_init_blit_functions(struct si_context *sctx);
void si_decompress_graphics_textures(struct si_context *sctx);
void si_decompress_compute_textures(struct si_context *sctx);
void si_resource_copy_region(struct pipe_context *ctx,
diff --git a/src/gallium/drivers/radeonsi/si_state.c b/src/gallium/drivers/radeonsi/si_state.c
index 359058b..de30076 100644
--- a/src/gallium/drivers/radeonsi/si_state.c
+++ b/src/gallium/drivers/radeonsi/si_state.c
@@ -911,20 +911,22 @@ static void si_bind_rs_state(struct pipe_context *ctx, void *state)
sctx->framebuffer.nr_samples > 1)
si_mark_atom_dirty(sctx, &sctx->msaa_sample_locs.atom);
}
r600_viewport_set_rast_deps(&sctx->b, rs->scissor_enable, rs->clip_halfz);
si_pm4_bind_state(sctx, rasterizer, rs);
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;
}
static void si_delete_rs_state(struct pipe_context *ctx, void *state)
{
struct si_context *sctx = (struct si_context *)ctx;
if (sctx->queued.named.rasterizer == state)
si_pm4_bind_state(sctx, poly_offset, NULL);
si_pm4_delete_state(sctx, rasterizer, (struct si_state_rasterizer *)state);
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index bdcfb5b..19880c5 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -348,20 +348,21 @@ si_create_sampler_view_custom(struct pipe_context *ctx,
unsigned force_level);
/* si_state_shader.c */
bool si_update_shaders(struct si_context *sctx);
void si_init_shader_functions(struct si_context *sctx);
bool si_init_shader_cache(struct si_screen *sscreen);
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);
void si_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info *dinfo);
void si_trace_emit(struct si_context *sctx);
static inline unsigned
si_tile_mode_index(struct r600_texture *rtex, unsigned level, bool stencil)
{
diff --git a/src/gallium/drivers/radeonsi/si_state_draw.c b/src/gallium/drivers/radeonsi/si_state_draw.c
index 3ce19aa..1ce8ac8 100644
--- a/src/gallium/drivers/radeonsi/si_state_draw.c
+++ b/src/gallium/drivers/radeonsi/si_state_draw.c
@@ -254,170 +254,218 @@ static unsigned si_num_prims_for_vertices(const struct pipe_draw_info *info)
switch (info->mode) {
case PIPE_PRIM_PATCHES:
return info->count / info->vertices_per_patch;
case R600_PRIM_RECTANGLE_LIST:
return info->count / 3;
default:
return u_prims_for_vertices(info->mode, info->count);
}
}
-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. */
bool wd_switch_on_eop = false;
bool ia_switch_on_eop = false;
bool ia_switch_on_eoi = false;
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;
}
}
}
/* 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.
*
* 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
* instances are smaller than a primgroup.
* 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);
}
/* If SWITCH_ON_EOI is set, PARTIAL_ES_WAVE must be set too. */
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;
if (!sctx->emit_scratch_reloc)
return;
radeon_set_context_reg(cs, R_0286E8_SPI_TMPRING_SIZE,
sctx->spi_tmpring_size);
diff --git a/src/gallium/drivers/radeonsi/si_state_shaders.c b/src/gallium/drivers/radeonsi/si_state_shaders.c
index 94fffdb..ebd2435 100644
--- a/src/gallium/drivers/radeonsi/si_state_shaders.c
+++ b/src/gallium/drivers/radeonsi/si_state_shaders.c
@@ -1713,57 +1713,70 @@ static void si_bind_gs_shader(struct pipe_context *ctx, void *state)
{
struct si_context *sctx = (struct si_context *)ctx;
struct si_shader_selector *sel = state;
bool enable_changed = !!sctx->gs_shader.cso != !!sel;
if (sctx->gs_shader.cso == sel)
return;
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 */
if (enable_changed)
si_shader_change_notify(sctx);
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;
struct si_shader_selector *sel = state;
bool enable_changed = !!sctx->tcs_shader.cso != !!sel;
if (sctx->tcs_shader.cso == sel)
return;
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)
sctx->last_tcs = NULL; /* invalidate derived tess state */
}
static void si_bind_tes_shader(struct pipe_context *ctx, void *state)
{
struct si_context *sctx = (struct si_context *)ctx;
struct si_shader_selector *sel = state;
bool enable_changed = !!sctx->tes_shader.cso != !!sel;
if (sctx->tes_shader.cso == sel)
return;
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 */
if (enable_changed) {
si_shader_change_notify(sctx);
sctx->last_tes_sh_base = -1; /* invalidate derived tess state */
}
r600_update_vs_writes_viewport_index(&sctx->b, si_get_vs_info(sctx));
}
--
2.7.4
More information about the mesa-dev
mailing list