[Mesa-dev] [PATCH 05/10] radeonsi: do only 1 big CE dump at end of IBs and one reload in the preamble
Marek Olšák
maraeo at gmail.com
Wed May 17 19:38:47 UTC 2017
From: Marek Olšák <marek.olsak at amd.com>
A later commit will only upload descriptors used by shaders, so we won't do
full dumps anymore, so the only way to have a complete mirror of CE RAM
in memory is to do a separate dump after the last draw call.
---
src/gallium/drivers/radeonsi/si_descriptors.c | 56 ++++++++++++---------------
src/gallium/drivers/radeonsi/si_hw_context.c | 8 +++-
src/gallium/drivers/radeonsi/si_pipe.c | 1 +
src/gallium/drivers/radeonsi/si_pipe.h | 3 ++
src/gallium/drivers/radeonsi/si_state.h | 6 +--
5 files changed, 37 insertions(+), 37 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index 91cc9a6..38e4ae1 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -142,74 +142,69 @@ static bool si_ce_upload(struct si_context *sctx, unsigned ce_offset, unsigned s
radeon_emit(sctx->ce_ib, va);
radeon_emit(sctx->ce_ib, va >> 32);
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, *out_buf,
RADEON_USAGE_READWRITE, RADEON_PRIO_DESCRIPTORS);
sctx->ce_need_synchronization = true;
return true;
}
-static void si_ce_reinitialize_descriptors(struct si_context *sctx,
- struct si_descriptors *desc)
+void si_ce_save_all_descriptors_at_ib_end(struct si_context* sctx)
{
- if (desc->buffer) {
- struct r600_resource *buffer = (struct r600_resource*)desc->buffer;
- unsigned list_size = desc->num_elements * desc->element_dw_size * 4;
- uint64_t va = buffer->gpu_address + desc->buffer_offset;
- struct radeon_winsys_cs *ib = sctx->ce_preamble_ib;
-
- if (!ib)
- ib = sctx->ce_ib;
+ bool success = si_ce_upload(sctx, 0, sctx->total_ce_ram_allocated,
+ &sctx->ce_ram_saved_offset,
+ &sctx->ce_ram_saved_buffer);
+ (void)success;
+ assert(success);
+}
- list_size = align(list_size, 32);
+void si_ce_restore_all_descriptors_at_ib_start(struct si_context *sctx)
+{
+ if (!sctx->ce_ram_saved_buffer)
+ return;
- radeon_emit(ib, PKT3(PKT3_LOAD_CONST_RAM, 3, 0));
- radeon_emit(ib, va);
- radeon_emit(ib, va >> 32);
- radeon_emit(ib, list_size / 4);
- radeon_emit(ib, desc->ce_offset);
+ struct radeon_winsys_cs *ib = sctx->ce_preamble_ib;
+ if (!ib)
+ ib = sctx->ce_ib;
- radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, desc->buffer,
- RADEON_USAGE_READ, RADEON_PRIO_DESCRIPTORS);
- }
- desc->ce_ram_dirty = false;
-}
+ uint64_t va = sctx->ce_ram_saved_buffer->gpu_address +
+ sctx->ce_ram_saved_offset;
-void si_ce_reinitialize_all_descriptors(struct si_context *sctx)
-{
- int i;
+ radeon_emit(ib, PKT3(PKT3_LOAD_CONST_RAM, 3, 0));
+ radeon_emit(ib, va);
+ radeon_emit(ib, va >> 32);
+ radeon_emit(ib, sctx->total_ce_ram_allocated / 4);
+ radeon_emit(ib, 0);
- for (i = 0; i < SI_NUM_DESCS; ++i)
- si_ce_reinitialize_descriptors(sctx, &sctx->descriptors[i]);
+ radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx,
+ sctx->ce_ram_saved_buffer,
+ RADEON_USAGE_READ, RADEON_PRIO_DESCRIPTORS);
}
void si_ce_enable_loads(struct radeon_winsys_cs *ib)
{
radeon_emit(ib, PKT3(PKT3_CONTEXT_CONTROL, 1, 0));
radeon_emit(ib, CONTEXT_CONTROL_LOAD_ENABLE(1) |
CONTEXT_CONTROL_LOAD_CE_RAM(1));
radeon_emit(ib, CONTEXT_CONTROL_SHADOW_ENABLE(1));
}
static bool si_upload_descriptors(struct si_context *sctx,
struct si_descriptors *desc,
struct r600_atom * atom)
{
unsigned list_size = desc->num_elements * desc->element_dw_size * 4;
if (sctx->ce_ib && desc->uses_ce) {
uint32_t const* list = (uint32_t const*)desc->list;
- if (desc->ce_ram_dirty)
- si_ce_reinitialize_descriptors(sctx, desc);
-
while(desc->dirty_mask) {
int begin, count;
u_bit_scan_consecutive_range64(&desc->dirty_mask, &begin,
&count);
begin *= desc->element_dw_size;
count *= desc->element_dw_size;
radeon_emit(sctx->ce_ib,
PKT3(PKT3_WRITE_CONST_RAM, count, 0));
@@ -240,22 +235,20 @@ static bool si_upload_descriptors(struct si_context *sctx,
if (atom)
si_mark_atom_dirty(sctx, atom);
return true;
}
static void
si_descriptors_begin_new_cs(struct si_context *sctx, struct si_descriptors *desc)
{
- desc->ce_ram_dirty = true;
-
if (!desc->buffer)
return;
radeon_add_to_buffer_list(&sctx->b, &sctx->b.gfx, desc->buffer,
RADEON_USAGE_READ, RADEON_PRIO_DESCRIPTORS);
}
/* SAMPLER VIEWS */
static unsigned
@@ -2037,20 +2030,21 @@ void si_init_all_descriptors(struct si_context *sctx)
SI_NUM_RW_BUFFERS, SI_SGPR_RW_BUFFERS,
/* The second set of usage/priority is used by
* const buffers in RW buffer slots. */
RADEON_USAGE_READWRITE, RADEON_USAGE_READ,
RADEON_PRIO_SHADER_RINGS, RADEON_PRIO_CONST_BUFFER,
&ce_offset);
si_init_descriptors(&sctx->vertex_buffers, SI_SGPR_VERTEX_BUFFERS,
4, SI_NUM_VERTEX_BUFFERS, NULL);
sctx->descriptors_dirty = u_bit_consecutive(0, SI_NUM_DESCS);
+ sctx->total_ce_ram_allocated = ce_offset;
if (sctx->b.chip_class >= GFX9)
assert(ce_offset <= 4096);
else
assert(ce_offset <= 32768);
/* Set pipe_context functions. */
sctx->b.b.bind_sampler_states = si_bind_sampler_states;
sctx->b.b.set_shader_images = si_set_shader_images;
sctx->b.b.set_constant_buffer = si_pipe_set_constant_buffer;
diff --git a/src/gallium/drivers/radeonsi/si_hw_context.c b/src/gallium/drivers/radeonsi/si_hw_context.c
index e15f6a9..5e97d56 100644
--- a/src/gallium/drivers/radeonsi/si_hw_context.c
+++ b/src/gallium/drivers/radeonsi/si_hw_context.c
@@ -116,20 +116,24 @@ void si_context_gfx_flush(void *context, unsigned flags,
* This code is only needed when the driver flushes the GFX IB
* internally, and it never asks for a fence handle.
*/
if (radeon_emitted(ctx->b.dma.cs, 0)) {
assert(fence == NULL); /* internal flushes only */
ctx->b.dma.flush(ctx, flags, NULL);
}
ctx->gfx_flush_in_progress = true;
+ /* This CE dump should be done in parallel with the last draw. */
+ if (ctx->ce_ib)
+ si_ce_save_all_descriptors_at_ib_end(ctx);
+
r600_preflush_suspend_features(&ctx->b);
ctx->b.flags |= SI_CONTEXT_CS_PARTIAL_FLUSH |
SI_CONTEXT_PS_PARTIAL_FLUSH;
/* DRM 3.1.0 doesn't flush TC for VI correctly. */
if (ctx->b.chip_class == VI && ctx->b.screen->info.drm_minor <= 1)
ctx->b.flags |= SI_CONTEXT_INV_GLOBAL_L2 |
SI_CONTEXT_INV_VMEM_L1;
@@ -200,22 +204,22 @@ void si_begin_new_cs(struct si_context *ctx)
/* The CS initialization should be emitted before everything else. */
si_pm4_emit(ctx, ctx->init_config);
if (ctx->init_config_gs_rings)
si_pm4_emit(ctx, ctx->init_config_gs_rings);
if (ctx->ce_preamble_ib)
si_ce_enable_loads(ctx->ce_preamble_ib);
else if (ctx->ce_ib)
si_ce_enable_loads(ctx->ce_ib);
- if (ctx->ce_preamble_ib)
- si_ce_reinitialize_all_descriptors(ctx);
+ if (ctx->ce_ib)
+ si_ce_restore_all_descriptors_at_ib_start(ctx);
if (ctx->b.chip_class >= CIK)
si_mark_atom_dirty(ctx, &ctx->prefetch_L2);
ctx->framebuffer.dirty_cbufs = (1 << 8) - 1;
ctx->framebuffer.dirty_zsbuf = true;
si_mark_atom_dirty(ctx, &ctx->framebuffer.atom);
si_mark_atom_dirty(ctx, &ctx->clip_regs);
si_mark_atom_dirty(ctx, &ctx->clip_state.atom);
diff --git a/src/gallium/drivers/radeonsi/si_pipe.c b/src/gallium/drivers/radeonsi/si_pipe.c
index dd962e0..eaa3348 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.c
+++ b/src/gallium/drivers/radeonsi/si_pipe.c
@@ -48,20 +48,21 @@ static void si_destroy_context(struct pipe_context *context)
* properly.
*/
struct pipe_framebuffer_state fb = {};
context->set_framebuffer_state(context, &fb);
si_release_all_descriptors(sctx);
if (sctx->ce_suballocator)
u_suballocator_destroy(sctx->ce_suballocator);
+ r600_resource_reference(&sctx->ce_ram_saved_buffer, NULL);
pipe_resource_reference(&sctx->esgs_ring, NULL);
pipe_resource_reference(&sctx->gsvs_ring, NULL);
pipe_resource_reference(&sctx->tf_ring, NULL);
pipe_resource_reference(&sctx->tess_offchip_ring, NULL);
pipe_resource_reference(&sctx->null_const_buf.buffer, NULL);
r600_resource_reference(&sctx->border_color_buffer, NULL);
free(sctx->border_color_table);
r600_resource_reference(&sctx->scratch_buffer, NULL);
r600_resource_reference(&sctx->compute_scratch_buffer, NULL);
diff --git a/src/gallium/drivers/radeonsi/si_pipe.h b/src/gallium/drivers/radeonsi/si_pipe.h
index 449a802..13ec072 100644
--- a/src/gallium/drivers/radeonsi/si_pipe.h
+++ b/src/gallium/drivers/radeonsi/si_pipe.h
@@ -229,20 +229,23 @@ struct si_context {
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;
struct radeon_winsys_cs *ce_ib;
struct radeon_winsys_cs *ce_preamble_ib;
+ struct r600_resource *ce_ram_saved_buffer;
+ unsigned ce_ram_saved_offset;
+ unsigned total_ce_ram_allocated;
bool ce_need_synchronization;
struct u_suballocator *ce_suballocator;
struct si_shader_ctx_state fixed_func_tcs_shader;
LLVMTargetMachineRef tm; /* only non-threaded compilation */
bool gfx_flush_in_progress;
bool compute_is_busy;
/* Atoms (direct states). */
union si_state_atoms atoms;
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index c4ef903..9b506a8 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -225,23 +225,20 @@ struct si_descriptors {
unsigned buffer_offset;
/* Offset in CE RAM */
unsigned ce_offset;
/* elements of the list that are changed and need to be uploaded */
uint64_t dirty_mask;
/* Whether CE is used to upload this descriptor array. */
bool uses_ce;
- /* Whether the CE ram is dirty and needs to be reinitialized entirely
- * before we can do partial updates. */
- bool ce_ram_dirty;
/* The shader userdata offset within a shader where the 64-bit pointer to the descriptor
* array will be stored. */
unsigned shader_userdata_offset;
};
struct si_sampler_views {
struct pipe_sampler_view *views[SI_NUM_SAMPLERS];
struct si_sampler_state *sampler_states[SI_NUM_SAMPLERS];
@@ -275,21 +272,22 @@ struct si_buffer_resources {
#define si_pm4_delete_state(sctx, member, value) \
do { \
if ((sctx)->queued.named.member == (value)) { \
(sctx)->queued.named.member = NULL; \
} \
si_pm4_free_state(sctx, (struct si_pm4_state *)(value), \
si_pm4_block_idx(member)); \
} while(0)
/* si_descriptors.c */
-void si_ce_reinitialize_all_descriptors(struct si_context *sctx);
+void si_ce_save_all_descriptors_at_ib_end(struct si_context* sctx);
+void si_ce_restore_all_descriptors_at_ib_start(struct si_context *sctx);
void si_ce_enable_loads(struct radeon_winsys_cs *ib);
void si_set_mutable_tex_desc_fields(struct si_screen *sscreen,
struct r600_texture *tex,
const struct legacy_surf_level *base_level_info,
unsigned base_level, unsigned first_level,
unsigned block_width, bool is_stencil,
uint32_t *state);
void si_get_pipe_constant_buffer(struct si_context *sctx, uint shader,
uint slot, struct pipe_constant_buffer *cbuf);
void si_get_shader_buffers(struct si_context *sctx,
--
2.7.4
More information about the mesa-dev
mailing list