r600g: split INVAL_READ_CACHES into vertex, tex, and const cache flags

also flushing any cache in evergreen_emit_cs_shader seems to be superfluous
(we don't flush caches when changing the other shaders either)

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
This commit is contained in:
Marek Olšák 2013-06-30 17:01:24 +02:00
parent 098316211c
commit 1b40398d02
5 changed files with 52 additions and 27 deletions

View file

@ -159,7 +159,7 @@ static void evergreen_cs_set_vertex_buffer(
/* The vertex instructions in the compute shaders use the texture cache,
* so we need to invalidate it. */
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES;
rctx->flags |= R600_CONTEXT_INV_VERTEX_CACHE;
state->enabled_mask |= 1 << vb_index;
state->dirty_mask |= 1 << vb_index;
state->atom.dirty = true;
@ -470,7 +470,9 @@ static void compute_emit_cs(struct r600_context *ctx, const uint *block_layout,
/* XXX evergreen_flush_emit() hardcodes the CP_COHER_SIZE to 0xffffffff
*/
ctx->flags |= R600_CONTEXT_INVAL_READ_CACHES;
ctx->flags |= R600_CONTEXT_INV_CONST_CACHE |
R600_CONTEXT_INV_VERTEX_CACHE |
R600_CONTEXT_INV_TEX_CACHE;
r600_flush_emit(ctx);
#if 0
@ -519,8 +521,6 @@ void evergreen_emit_cs_shader(
r600_write_value(cs, PKT3C(PKT3_NOP, 0, 0));
r600_write_value(cs, r600_context_bo_reloc(rctx, &rctx->rings.gfx,
kernel->code_bo, RADEON_USAGE_READ));
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES;
}
static void evergreen_launch_grid(

View file

@ -123,7 +123,9 @@ void evergreen_cp_dma_clear_buffer(struct r600_context *rctx,
/* We flush the caches, because we might read from or write
* to resources which are bound right now. */
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES |
rctx->flags |= R600_CONTEXT_INV_CONST_CACHE |
R600_CONTEXT_INV_VERTEX_CACHE |
R600_CONTEXT_INV_TEX_CACHE |
R600_CONTEXT_FLUSH_AND_INV |
R600_CONTEXT_FLUSH_AND_INV_CB |
R600_CONTEXT_FLUSH_AND_INV_DB |
@ -168,7 +170,9 @@ void evergreen_cp_dma_clear_buffer(struct r600_context *rctx,
}
/* Invalidate the read caches. */
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES;
rctx->flags |= R600_CONTEXT_INV_CONST_CACHE |
R600_CONTEXT_INV_VERTEX_CACHE |
R600_CONTEXT_INV_TEX_CACHE;
util_range_add(&r600_resource(dst)->valid_buffer_range, offset,
offset + size);

View file

@ -220,6 +220,14 @@ void r600_flush_emit(struct r600_context *rctx)
(rctx->flags & R600_CONTEXT_FLUSH_AND_INV_DB_META)) {
cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0);
/* Set FULL_CACHE_ENA for DB META flushes on r7xx and later.
*
* This hack predates use of FLUSH_AND_INV_DB_META, so it's
* unclear whether it's still needed or even whether it has
* any effect.
*/
cp_coher_cntl |= S_0085F0_FULL_CACHE_ENA(1);
}
if (rctx->flags & R600_CONTEXT_FLUSH_AND_INV) {
@ -227,11 +235,15 @@ void r600_flush_emit(struct r600_context *rctx)
cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT) | EVENT_INDEX(0);
}
if (rctx->flags & R600_CONTEXT_INVAL_READ_CACHES) {
cp_coher_cntl |= S_0085F0_VC_ACTION_ENA(1) |
S_0085F0_TC_ACTION_ENA(1) |
S_0085F0_SH_ACTION_ENA(1) |
S_0085F0_FULL_CACHE_ENA(1);
if (rctx->flags & R600_CONTEXT_INV_CONST_CACHE) {
cp_coher_cntl |= S_0085F0_SH_ACTION_ENA(1);
}
if (rctx->flags & R600_CONTEXT_INV_VERTEX_CACHE) {
cp_coher_cntl |= rctx->has_vertex_cache ? S_0085F0_VC_ACTION_ENA(1)
: S_0085F0_TC_ACTION_ENA(1);
}
if (rctx->flags & R600_CONTEXT_INV_TEX_CACHE) {
cp_coher_cntl |= S_0085F0_TC_ACTION_ENA(1);
}
if (rctx->flags & R600_CONTEXT_FLUSH_AND_INV_DB) {
@ -616,7 +628,9 @@ void r600_cp_dma_copy_buffer(struct r600_context *rctx,
/* We flush the caches, because we might read from or write
* to resources which are bound right now. */
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES |
rctx->flags |= R600_CONTEXT_INV_CONST_CACHE |
R600_CONTEXT_INV_VERTEX_CACHE |
R600_CONTEXT_INV_TEX_CACHE |
R600_CONTEXT_FLUSH_AND_INV |
R600_CONTEXT_FLUSH_AND_INV_CB |
R600_CONTEXT_FLUSH_AND_INV_DB |
@ -666,7 +680,9 @@ void r600_cp_dma_copy_buffer(struct r600_context *rctx,
}
/* Invalidate the read caches. */
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES;
rctx->flags |= R600_CONTEXT_INV_CONST_CACHE |
R600_CONTEXT_INV_VERTEX_CACHE |
R600_CONTEXT_INV_TEX_CACHE;
util_range_add(&r600_resource(dst)->valid_buffer_range, dst_offset,
dst_offset + size);

View file

@ -64,16 +64,21 @@
#define R600_ERR(fmt, args...) \
fprintf(stderr, "EE %s:%d %s - "fmt, __FILE__, __LINE__, __func__, ##args)
#define R600_CONTEXT_INVAL_READ_CACHES (1 << 0)
#define R600_CONTEXT_STREAMOUT_FLUSH (1 << 1)
#define R600_CONTEXT_WAIT_3D_IDLE (1 << 2)
#define R600_CONTEXT_WAIT_CP_DMA_IDLE (1 << 3)
#define R600_CONTEXT_FLUSH_AND_INV (1 << 4)
#define R600_CONTEXT_FLUSH_AND_INV_CB_META (1 << 5)
#define R600_CONTEXT_PS_PARTIAL_FLUSH (1 << 6)
#define R600_CONTEXT_FLUSH_AND_INV_DB_META (1 << 7)
#define R600_CONTEXT_FLUSH_AND_INV_DB (1 << 8)
#define R600_CONTEXT_FLUSH_AND_INV_CB (1 << 9)
/* read caches */
#define R600_CONTEXT_INV_VERTEX_CACHE (1 << 0)
#define R600_CONTEXT_INV_TEX_CACHE (1 << 1)
#define R600_CONTEXT_INV_CONST_CACHE (1 << 2)
/* read-write caches */
#define R600_CONTEXT_STREAMOUT_FLUSH (1 << 8)
#define R600_CONTEXT_FLUSH_AND_INV (1 << 9)
#define R600_CONTEXT_FLUSH_AND_INV_CB_META (1 << 10)
#define R600_CONTEXT_FLUSH_AND_INV_DB_META (1 << 11)
#define R600_CONTEXT_FLUSH_AND_INV_DB (1 << 12)
#define R600_CONTEXT_FLUSH_AND_INV_CB (1 << 13)
/* engine synchronization */
#define R600_CONTEXT_PS_PARTIAL_FLUSH (1 << 16)
#define R600_CONTEXT_WAIT_3D_IDLE (1 << 17)
#define R600_CONTEXT_WAIT_CP_DMA_IDLE (1 << 18)
#define R600_QUERY_DRAW_CALLS (PIPE_QUERY_DRIVER_SPECIFIC + 0)
#define R600_QUERY_REQUESTED_VRAM (PIPE_QUERY_DRIVER_SPECIFIC + 1)

View file

@ -89,7 +89,7 @@ static void r600_texture_barrier(struct pipe_context *ctx)
{
struct r600_context *rctx = (struct r600_context *)ctx;
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES |
rctx->flags |= R600_CONTEXT_INV_TEX_CACHE |
R600_CONTEXT_FLUSH_AND_INV_CB |
R600_CONTEXT_FLUSH_AND_INV |
R600_CONTEXT_WAIT_3D_IDLE;
@ -500,7 +500,7 @@ static void r600_set_index_buffer(struct pipe_context *ctx,
void r600_vertex_buffers_dirty(struct r600_context *rctx)
{
if (rctx->vertex_buffer_state.dirty_mask) {
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES;
rctx->flags |= R600_CONTEXT_INV_VERTEX_CACHE;
rctx->vertex_buffer_state.atom.num_dw = (rctx->chip_class >= EVERGREEN ? 12 : 11) *
util_bitcount(rctx->vertex_buffer_state.dirty_mask);
rctx->vertex_buffer_state.atom.dirty = true;
@ -557,7 +557,7 @@ void r600_sampler_views_dirty(struct r600_context *rctx,
struct r600_samplerview_state *state)
{
if (state->dirty_mask) {
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES;
rctx->flags |= R600_CONTEXT_INV_TEX_CACHE;
state->atom.num_dw = (rctx->chip_class >= EVERGREEN ? 14 : 13) *
util_bitcount(state->dirty_mask);
state->atom.dirty = true;
@ -912,7 +912,7 @@ static void r600_delete_vs_state(struct pipe_context *ctx, void *state)
void r600_constant_buffers_dirty(struct r600_context *rctx, struct r600_constbuf_state *state)
{
if (state->dirty_mask) {
rctx->flags |= R600_CONTEXT_INVAL_READ_CACHES;
rctx->flags |= R600_CONTEXT_INV_CONST_CACHE;
state->atom.num_dw = rctx->chip_class >= EVERGREEN ? util_bitcount(state->dirty_mask)*20
: util_bitcount(state->dirty_mask)*19;
state->atom.dirty = true;