From: Jerome Glisse <jgli...@redhat.com>
Flushing and synchronization only need to happen at begining
and end of cs, and after each draw packet if necessary. This
patch is especialy needed for hyperz/htile feature.
v2: Separate evergreen and r6xx/r7xx flushing/syncing allow
easier specialization of each functions. Fix r6xx/r7xx
regression.
Signed-off-by: Jerome Glisse <jgli...@redhat.com>
---
src/gallium/drivers/r600/evergreen_compute.c | 23 +--
.../drivers/r600/evergreen_compute_internal.c | 4 +-
src/gallium/drivers/r600/evergreen_hw_context.c | 110 ++++++++++-
src/gallium/drivers/r600/evergreen_state.c | 14 +-
src/gallium/drivers/r600/evergreend.h | 3 +-
src/gallium/drivers/r600/r600.h | 19 +-
src/gallium/drivers/r600/r600_buffer.c | 2 +-
src/gallium/drivers/r600/r600_hw_context.c | 203 ++++++++++++--------
src/gallium/drivers/r600/r600_hw_context_priv.h | 3 +-
src/gallium/drivers/r600/r600_pipe.c | 2 -
src/gallium/drivers/r600/r600_pipe.h | 6 +-
src/gallium/drivers/r600/r600_state.c | 23 +--
src/gallium/drivers/r600/r600_state_common.c | 68 ++-----
13 files changed, 297 insertions(+), 183 deletions(-)
diff --git a/src/gallium/drivers/r600/evergreen_compute.c
b/src/gallium/drivers/r600/evergreen_compute.c
index 947a328..37c3395 100644
--- a/src/gallium/drivers/r600/evergreen_compute.c
+++ b/src/gallium/drivers/r600/evergreen_compute.c
@@ -96,7 +96,7 @@ static void evergreen_cs_set_vertex_buffer(
vb->buffer = buffer;
vb->user_buffer = NULL;
- r600_inval_vertex_cache(rctx);
+ rctx->flags |= R600_CONTEXT_VTX_FLUSH;
state->dirty_mask |= 1 << vb_index;
r600_atom_dirty(rctx, &state->atom);
}
@@ -208,8 +208,7 @@ static void evergreen_bind_compute_state(struct
pipe_context *ctx_, void *state)
res->usage = RADEON_USAGE_READ;
res->coher_bo_size = ctx->cs_shader->bc.ndw*4;
- r600_inval_shader_cache(ctx);
-
+ ctx->flags |= R600_CONTEXT_SH_FLUSH;
}
/* The kernel parameters are stored a vtx buffer (ID=0), besides the explicit
@@ -364,8 +363,11 @@ static void compute_emit_cs(struct r600_context *ctx)
*/
r600_emit_atom(ctx, &ctx->start_compute_cs_cmd.atom);
+ ctx->flags |= R600_CONTEXT_CB_FLUSH;
+ r600_flush_emit(ctx);
+
/* Emit cb_state */
- cb_state = ctx->states[R600_PIPE_STATE_FRAMEBUFFER];
+ cb_state = ctx->states[R600_PIPE_STATE_FRAMEBUFFER];
r600_context_pipe_state_emit(ctx, cb_state,
RADEON_CP_PACKET3_COMPUTE_MODE);
/* Emit vertex buffer state */
@@ -405,15 +407,8 @@ static void compute_emit_cs(struct r600_context *ctx)
}
}
- /* r600_flush_framebuffer() updates the cb_flush_flags and then
- * calls r600_emit_atom() on the ctx->surface_sync_cmd.atom, which emits
- * a SURFACE_SYNC packet via r600_emit_surface_sync().
- *
- * XXX r600_emit_surface_sync() hardcodes the CP_COHER_SIZE to
- * 0xffffffff, so we will need to add a field to struct
- * r600_surface_sync_cmd if we want to manually set this value.
- */
- r600_flush_framebuffer(ctx, true /* Flush now */);
+ ctx->flags |= R600_CONTEXT_CB_FLUSH;
+ r600_flush_emit(ctx);
#if 0
COMPUTE_DBG("cdw: %i\n", cs->cdw);
@@ -460,6 +455,8 @@ static void evergreen_launch_grid(
evergreen_set_lds(ctx->cs_shader, 0, 0, num_waves);
evergreen_compute_upload_input(ctx_, block_layout, grid_layout, input);
evergreen_direct_dispatch(ctx_, block_layout, grid_layout);
+ /* set draw pending so flush function know we mean business */
+ ctx->flags |= R600_CONTEXT_DRAW_PENDING;
compute_emit_cs(ctx);
}
diff --git a/src/gallium/drivers/r600/evergreen_compute_internal.c
b/src/gallium/drivers/r600/evergreen_compute_internal.c
index 1d11bab..8bb6426 100644
--- a/src/gallium/drivers/r600/evergreen_compute_internal.c
+++ b/src/gallium/drivers/r600/evergreen_compute_internal.c
@@ -559,7 +559,7 @@ void evergreen_set_tex_resource(
res->coher_bo_size = tmp->offset[0] +
util_format_get_blockwidth(tmp->real_format)*view->base.texture->width0*height*depth;
- r600_inval_texture_cache(pipe->ctx);
+ pipe->ctx->flags |= R600_CONTEXT_TEX_FLUSH;
evergreen_emit_force_reloc(res);
evergreen_emit_force_reloc(res);
@@ -618,7 +618,7 @@ void evergreen_set_const_cache(
res->usage = RADEON_USAGE_READ;
res->coher_bo_size = size;
- r600_inval_shader_cache(pipe->ctx);
+ pipe->ctx->flags |= R600_CONTEXT_SH_FLUSH;
}
struct r600_resource* r600_compute_buffer_alloc_vram(
diff --git a/src/gallium/drivers/r600/evergreen_hw_context.c
b/src/gallium/drivers/r600/evergreen_hw_context.c
index 546c884..cf4a225 100644
--- a/src/gallium/drivers/r600/evergreen_hw_context.c
+++ b/src/gallium/drivers/r600/evergreen_hw_context.c
@@ -761,8 +761,9 @@ static inline void
evergreen_context_pipe_state_set_sampler_border(struct r600_c
/* We have to flush the shaders before we change the border color
* registers, or previous draw commands that haven't completed yet
* will end up using the new border color. */
- if (dirty & R600_BLOCK_STATUS_DIRTY)
- r600_context_ps_partial_flush(ctx);
+ if (dirty & R600_BLOCK_STATUS_DIRTY) {
+ ctx->flags |= R600_CONTEXT_DRAW_FLUSH;
+ }
if (dirty)
r600_context_dirty_block(ctx, block, dirty, 4);
}
@@ -823,3 +824,108 @@ void evergreen_set_streamout_enable(struct r600_context
*ctx, unsigned buffer_en
cs->buf[cs->cdw++] = S_028B94_STREAMOUT_0_EN(0);
}
}
+
+void evergreen_flush_emit(struct r600_context *rctx)
+{
+ struct radeon_winsys_cs *cs = rctx->cs;
+ unsigned mask;
+
+ if (!(rctx->flags & R600_CONTEXT_DRAW_PENDING)) {
+ return;
+ }
+ /* for GPU without vertex cache flush the texture cache */
+ if (!rctx->has_vertex_cache) {
+ rctx->flags |= R600_CONTEXT_TEX_FLUSH;
+ }
+
+ if (rctx->flags & R600_CONTEXT_DRAW_FLUSH) {
+ rctx->flags &= ~R600_CONTEXT_DRAW_FLUSH;
+ cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
+ cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_PS_PARTIAL_FLUSH) |
EVENT_INDEX(4);
+ }
+
+ if (rctx->flags & R600_CONTEXT_DB_FLUSH) {
+ cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
+ cs->buf[cs->cdw++] = EVENT_TYPE(FLUSH_AND_INV_DB_META) |
EVENT_INDEX(0);
+ }
+
+ mask = R600_CONTEXT_CB_FLUSH |
+ R600_CONTEXT_DB_FLUSH |
+ R600_CONTEXT_SH_FLUSH |
+ R600_CONTEXT_TEX_FLUSH |
+ R600_CONTEXT_VTX_FLUSH |
+ R600_CONTEXT_STREAMOUT_FLUSH;
+ if (rctx->flags & mask) {
+ /* anything left (cb, vtx, shader, streamout) can be flushed
+ * in a more generic way
+ */
+ unsigned flags = 0;
+
+ flags |= (rctx->flags & R600_CONTEXT_CB_FLUSH) ?
S_0085F0_CB_ACTION_ENA(1) |
+
S_0085F0_CB0_DEST_BASE_ENA(1) |
+
S_0085F0_CB1_DEST_BASE_ENA(1) |
+
S_0085F0_CB2_DEST_BASE_ENA(1) |
+
S_0085F0_CB3_DEST_BASE_ENA(1) |
+
S_0085F0_CB4_DEST_BASE_ENA(1) |
+
S_0085F0_CB5_DEST_BASE_ENA(1) |
+
S_0085F0_CB6_DEST_BASE_ENA(1) |
+
S_0085F0_CB7_DEST_BASE_ENA(1) |
+
S_0085F0_CB8_DEST_BASE_ENA(1) |
+
S_0085F0_CB9_DEST_BASE_ENA(1) |
+
S_0085F0_CB10_DEST_BASE_ENA(1) |
+
S_0085F0_CB11_DEST_BASE_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_DB_FLUSH) ?
S_0085F0_DB_ACTION_ENA(1) |
+
S_0085F0_DB_DEST_BASE_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_SH_FLUSH) ?
S_0085F0_SH_ACTION_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_TEX_FLUSH) ?
S_0085F0_TC_ACTION_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_VTX_FLUSH) ?
S_0085F0_VC_ACTION_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_STREAMOUT_FLUSH) ?
S_0085F0_DEST_BASE_0_ENA(1) |
+
S_0085F0_SMX_ACTION_ENA(1) |
+
S_0085F0_SO0_DEST_BASE_ENA(1) |
+
S_0085F0_SO1_DEST_BASE_ENA(1) |
+
S_0085F0_SO2_DEST_BASE_ENA(1) |
+
S_0085F0_SO3_DEST_BASE_ENA(1) : 0;
+ rctx->flags &= ~mask;
+ cs->buf[cs->cdw++] = PKT3(PKT3_SURFACE_SYNC, 3, 0);
+ cs->buf[cs->cdw++] = flags; /* CP_COHER_CNTL */
+ cs->buf[cs->cdw++] = 0xffffffff; /* CP_COHER_SIZE */
+ cs->buf[cs->cdw++] = 0; /* CP_COHER_BASE */
+ cs->buf[cs->cdw++] = 0x0000000A; /* POLL_INTERVAL */
+ }
+
+ if (rctx->flags & R600_CONTEXT_FLUSH_AND_INV) {
+ rctx->flags &= ~R600_CONTEXT_FLUSH_AND_INV;
+
+ cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
+ cs->buf[cs->cdw++] =
EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT) | EVENT_INDEX(0);
+ cs->buf[cs->cdw++] = PKT3(PKT3_SURFACE_SYNC, 3, 0);
+ /* CP_COHER_CNTL */
+ cs->buf[cs->cdw++] = S_0085F0_SO0_DEST_BASE_ENA(1) |
+ S_0085F0_SO1_DEST_BASE_ENA(1) |
+ S_0085F0_SO2_DEST_BASE_ENA(1) |
+ S_0085F0_SO3_DEST_BASE_ENA(1) |
+ S_0085F0_CB0_DEST_BASE_ENA(1) |
+ S_0085F0_CB1_DEST_BASE_ENA(1) |
+ S_0085F0_CB2_DEST_BASE_ENA(1) |
+ S_0085F0_CB3_DEST_BASE_ENA(1) |
+ S_0085F0_CB4_DEST_BASE_ENA(1) |
+ S_0085F0_CB5_DEST_BASE_ENA(1) |
+ S_0085F0_CB6_DEST_BASE_ENA(1) |
+ S_0085F0_CB7_DEST_BASE_ENA(1) |
+ S_0085F0_DB_DEST_BASE_ENA(1) |
+ S_0085F0_CB8_DEST_BASE_ENA(1) |
+ S_0085F0_CB9_DEST_BASE_ENA(1) |
+ S_0085F0_CB10_DEST_BASE_ENA(1) |
+ S_0085F0_CB11_DEST_BASE_ENA(1) |
+ S_0085F0_TC_ACTION_ENA(1) |
+ S_0085F0_VC_ACTION_ENA(1) |
+ S_0085F0_SH_ACTION_ENA(1) |
+ S_0085F0_SMX_ACTION_ENA(1);
+ cs->buf[cs->cdw++] = 0xffffffff; /* CP_COHER_SIZE */
+ cs->buf[cs->cdw++] = 0; /* CP_COHER_BASE */
+ cs->buf[cs->cdw++] = 0x0000000A; /* POLL_INTERVAL */
+ }
+
+ /* everything is properly flush */
+ rctx->flags &= R600_CONTEXT_DRAW_PENDING;
+}
diff --git a/src/gallium/drivers/r600/evergreen_state.c
b/src/gallium/drivers/r600/evergreen_state.c
index 214d76b..8e3eb95 100644
--- a/src/gallium/drivers/r600/evergreen_state.c
+++ b/src/gallium/drivers/r600/evergreen_state.c
@@ -1160,8 +1160,9 @@ static void evergreen_bind_ps_sampler(struct pipe_context
*ctx, unsigned count,
struct r600_context *rctx = (struct r600_context *)ctx;
struct r600_pipe_state **rstates = (struct r600_pipe_state **)states;
- if (count)
- r600_inval_texture_cache(rctx);
+ if (count) {
+ rctx->flags |= R600_CONTEXT_TEX_FLUSH;
+ }
memcpy(rctx->ps_samplers.samplers, states, sizeof(void*) * count);
rctx->ps_samplers.n_samplers = count;
@@ -1176,8 +1177,9 @@ static void evergreen_bind_vs_sampler(struct pipe_context
*ctx, unsigned count,
struct r600_context *rctx = (struct r600_context *)ctx;
struct r600_pipe_state **rstates = (struct r600_pipe_state **)states;
- if (count)
- r600_inval_texture_cache(rctx);
+ if (count) {
+ rctx->flags |= R600_CONTEXT_TEX_FLUSH;
+ }
for (int i = 0; i < count; i++) {
evergreen_context_pipe_state_set_vs_sampler(rctx, rstates[i],
i);
@@ -1674,6 +1676,7 @@ static void evergreen_db(struct r600_context *rctx,
struct r600_pipe_state *rsta
if (rtex->hyperz) {
uint64_t htile_offset =
rtex->hyperz->surface.level[level].offset;
+ rctx->flags |= R600_CONTEXT_FLUSH_AND_INV;
rctx->db_misc_state.hyperz = true;
rctx->db_misc_state.db_htile_surface_mask = 0xffffffff;
r600_atom_dirty(rctx, &rctx->db_misc_state.atom);
@@ -1709,7 +1712,7 @@ static void evergreen_set_framebuffer_state(struct
pipe_context *ctx,
if (rstate == NULL)
return;
- r600_flush_framebuffer(rctx, false);
+ rctx->flags |= R600_CONTEXT_CB_FLUSH | R600_CONTEXT_FLUSH_AND_INV;
/* unreference old buffer and reference new one */
rstate->id = R600_PIPE_STATE_FRAMEBUFFER;
@@ -1729,6 +1732,7 @@ static void evergreen_set_framebuffer_state(struct
pipe_context *ctx,
}
if (state->zsbuf) {
+ rctx->flags |= R600_CONTEXT_DB_FLUSH;
evergreen_db(rctx, rstate, state);
}
diff --git a/src/gallium/drivers/r600/evergreend.h
b/src/gallium/drivers/r600/evergreend.h
index 1ac5944..672e698 100644
--- a/src/gallium/drivers/r600/evergreend.h
+++ b/src/gallium/drivers/r600/evergreend.h
@@ -46,7 +46,8 @@
#define EVENT_TYPE_PS_PARTIAL_FLUSH 0x10
#define EVENT_TYPE_ZPASS_DONE 0x15
#define EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT 0x16
-#define EVENT_TYPE_SO_VGTSTREAMOUT_FLUSH 0x1f
+#define EVENT_TYPE_SO_VGTSTREAMOUT_FLUSH 0x1f
+#define FLUSH_AND_INV_DB_META 0x2c
#define EVENT_TYPE(x) ((x) << 0)
#define EVENT_INDEX(x) ((x) << 8)
diff --git a/src/gallium/drivers/r600/r600.h b/src/gallium/drivers/r600/r600.h
index 0ae7959..820b356 100644
--- a/src/gallium/drivers/r600/r600.h
+++ b/src/gallium/drivers/r600/r600.h
@@ -188,8 +188,17 @@ struct r600_so_target {
unsigned so_index;
};
-#define R600_CONTEXT_DRAW_PENDING (1 << 0)
-#define R600_CONTEXT_DST_CACHES_DIRTY (1 << 1)
+#define R600_CONTEXT_DRAW_PENDING (1 << 0)
+#define R600_CONTEXT_DRAW_FLUSH (1 << 1)
+#define R600_CONTEXT_CB_FLUSH (1 << 2)
+#define R600_CONTEXT_DB_FLUSH (1 << 3)
+#define R600_CONTEXT_SH_FLUSH (1 << 4)
+#define R600_CONTEXT_TEX_FLUSH (1 << 5)
+#define R600_CONTEXT_VTX_FLUSH (1 << 6)
+#define R600_CONTEXT_STREAMOUT_FLUSH (1 << 7)
+#define R600_CONTEXT_WAIT_IDLE (1 << 8)
+#define R600_CONTEXT_FLUSH_AND_INV (1 << 9)
+#define R600_CONTEXT_HTILE_ERRATA (1 << 10)
struct r600_context;
struct r600_screen;
@@ -207,10 +216,7 @@ void r600_context_flush(struct r600_context *ctx, unsigned
flags);
void r600_context_emit_fence(struct r600_context *ctx, struct r600_resource
*fence,
unsigned offset, unsigned value);
-void r600_inval_shader_cache(struct r600_context *ctx);
-void r600_inval_texture_cache(struct r600_context *ctx);
-void r600_inval_vertex_cache(struct r600_context *ctx);
-void r600_flush_framebuffer(struct r600_context *ctx, bool flush_now);
+void r600_flush_emit(struct r600_context *ctx);
void r600_context_streamout_begin(struct r600_context *ctx);
void r600_context_streamout_end(struct r600_context *ctx);
@@ -222,6 +228,7 @@ void r600_context_block_resource_emit_dirty(struct
r600_context *ctx, struct r60
int evergreen_context_init(struct r600_context *ctx);
void evergreen_context_pipe_state_set_ps_sampler(struct r600_context *ctx,
struct r600_pipe_state *state, unsigned id);
void evergreen_context_pipe_state_set_vs_sampler(struct r600_context *ctx,
struct r600_pipe_state *state, unsigned id);
+void evergreen_flush_emit(struct r600_context *rctx);
void _r600_pipe_state_add_reg_bo(struct r600_context *ctx,
struct r600_pipe_state *state,
diff --git a/src/gallium/drivers/r600/r600_buffer.c
b/src/gallium/drivers/r600/r600_buffer.c
index 8e2deb1..774b876 100644
--- a/src/gallium/drivers/r600/r600_buffer.c
+++ b/src/gallium/drivers/r600/r600_buffer.c
@@ -110,7 +110,7 @@ static void *r600_buffer_transfer_map(struct pipe_context
*pipe,
struct r600_vertexbuf_state * state =
&rctx->vertex_buffer_state;
state->dirty_mask |= 1 << i;
- r600_inval_vertex_cache(rctx);
+ rctx->flags |= R600_CONTEXT_VTX_FLUSH;
r600_atom_dirty(rctx, &state->atom);
}
}
diff --git a/src/gallium/drivers/r600/r600_hw_context.c
b/src/gallium/drivers/r600/r600_hw_context.c
index 4f2c03a..95c861a 100644
--- a/src/gallium/drivers/r600/r600_hw_context.c
+++ b/src/gallium/drivers/r600/r600_hw_context.c
@@ -114,19 +114,6 @@ err:
return;
}
-void r600_context_ps_partial_flush(struct r600_context *ctx)
-{
- struct radeon_winsys_cs *cs = ctx->cs;
-
- if (!(ctx->flags & R600_CONTEXT_DRAW_PENDING))
- return;
-
- cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
- cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_PS_PARTIAL_FLUSH) |
EVENT_INDEX(4);
-
- ctx->flags &= ~R600_CONTEXT_DRAW_PENDING;
-}
-
static void r600_init_block(struct r600_context *ctx,
struct r600_block *block,
const struct r600_reg *reg, int index, int nreg,
@@ -825,7 +812,7 @@ void r600_need_cs_space(struct r600_context *ctx, unsigned
num_dw,
}
/* Count in framebuffer cache flushes at the end of CS. */
- num_dw += 7; /* one SURFACE_SYNC and CACHE_FLUSH_AND_INV (r6xx-only) */
+ num_dw += 44; /* one SURFACE_SYNC and CACHE_FLUSH_AND_INV (r6xx-only) */
/* Save 16 dwords for the fence mechanism. */
num_dw += 16;
@@ -853,7 +840,7 @@ void r600_context_dirty_block(struct r600_context *ctx,
LIST_ADDTAIL(&block->list,&ctx->dirty);
if (block->flags & REG_FLAG_FLUSH_CHANGE) {
- r600_context_ps_partial_flush(ctx);
+ ctx->flags |= R600_CONTEXT_DRAW_FLUSH;
}
}
}
@@ -1085,8 +1072,9 @@ static inline void
r600_context_pipe_state_set_sampler_border(struct r600_contex
/* We have to flush the shaders before we change the border color
* registers, or previous draw commands that haven't completed yet
* will end up using the new border color. */
- if (dirty & R600_BLOCK_STATUS_DIRTY)
- r600_context_ps_partial_flush(ctx);
+ if (dirty & R600_BLOCK_STATUS_DIRTY) {
+ ctx->flags |= R600_CONTEXT_DRAW_FLUSH;
+ }
if (dirty)
r600_context_dirty_block(ctx, block, dirty, 3);
}
@@ -1200,54 +1188,116 @@ void r600_context_block_resource_emit_dirty(struct
r600_context *ctx, struct r60
LIST_DELINIT(&block->list);
}
-void r600_inval_shader_cache(struct r600_context *ctx)
+void r600_flush_emit(struct r600_context *rctx)
{
- ctx->surface_sync_cmd.flush_flags |= S_0085F0_SH_ACTION_ENA(1);
- r600_atom_dirty(ctx, &ctx->surface_sync_cmd.atom);
-}
-
-void r600_inval_texture_cache(struct r600_context *ctx)
-{
- ctx->surface_sync_cmd.flush_flags |= S_0085F0_TC_ACTION_ENA(1);
- r600_atom_dirty(ctx, &ctx->surface_sync_cmd.atom);
-}
-
-void r600_inval_vertex_cache(struct r600_context *ctx)
-{
- if (ctx->has_vertex_cache) {
- ctx->surface_sync_cmd.flush_flags |= S_0085F0_VC_ACTION_ENA(1);
- } else {
- /* Some GPUs don't have the vertex cache and must use the
texture cache instead. */
- ctx->surface_sync_cmd.flush_flags |= S_0085F0_TC_ACTION_ENA(1);
- }
- r600_atom_dirty(ctx, &ctx->surface_sync_cmd.atom);
-}
+ struct radeon_winsys_cs *cs = rctx->cs;
+ unsigned mask;
-void r600_flush_framebuffer(struct r600_context *ctx, bool flush_now)
-{
- if (!(ctx->flags & R600_CONTEXT_DST_CACHES_DIRTY))
+ if (!(rctx->flags & R600_CONTEXT_DRAW_PENDING)) {
return;
+ }
+ /* for GPU without vertex cache flush the texture cache */
+ if (!rctx->has_vertex_cache) {
+ rctx->flags |= R600_CONTEXT_TEX_FLUSH;
+ }
- ctx->surface_sync_cmd.flush_flags |=
- r600_get_cb_flush_flags(ctx) |
- (ctx->framebuffer.zsbuf ? S_0085F0_DB_ACTION_ENA(1) |
S_0085F0_DB_DEST_BASE_ENA(1) : 0);
+ if (rctx->flags & R600_CONTEXT_DRAW_FLUSH) {
+ rctx->flags &= ~R600_CONTEXT_DRAW_FLUSH;
+ cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
+ cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_PS_PARTIAL_FLUSH) |
EVENT_INDEX(4);
+ }
- if (flush_now) {
- r600_emit_atom(ctx, &ctx->surface_sync_cmd.atom);
- } else {
- r600_atom_dirty(ctx, &ctx->surface_sync_cmd.atom);
+ if (rctx->flags & R600_CONTEXT_WAIT_IDLE) {
+ /* wait for things to settle */
+ rctx->flags &= ~R600_CONTEXT_WAIT_IDLE;
+ cs->buf[cs->cdw++] = PKT3(PKT3_SET_CONFIG_REG, 1, 0);
+ cs->buf[cs->cdw++] = (R_008040_WAIT_UNTIL -
R600_CONFIG_REG_OFFSET) >> 2;
+ cs->buf[cs->cdw++] = S_008040_WAIT_3D_IDLE(1);
}
- /* Also add a complete cache flush to work around broken flushing on
R6xx. */
- if (ctx->chip_class == R600) {
- if (flush_now) {
- r600_emit_atom(ctx, &ctx->r6xx_flush_and_inv_cmd);
- } else {
- r600_atom_dirty(ctx, &ctx->r6xx_flush_and_inv_cmd);
+ if (rctx->flags & R600_CONTEXT_FLUSH_AND_INV) {
+ rctx->flags &= ~R600_CONTEXT_FLUSH_AND_INV;
+
+ cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
+ cs->buf[cs->cdw++] =
EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT) | EVENT_INDEX(0);
+
+ /* db flush are special due to errata with hyperz, we need to
+ * insert noop so cache have time to really flush
+ */
+ if (rctx->flags & R600_CONTEXT_HTILE_ERRATA) {
+ /* R600_CONTEXT_HTILE_ERRATA is persistant for whole cs
*/
+ cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 31, 0);
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
+ cs->buf[cs->cdw++] = 0xdeadcafe;
}
}
- ctx->flags &= ~R600_CONTEXT_DST_CACHES_DIRTY;
+ mask = R600_CONTEXT_CB_FLUSH |
+ R600_CONTEXT_DB_FLUSH |
+ R600_CONTEXT_SH_FLUSH |
+ R600_CONTEXT_TEX_FLUSH |
+ R600_CONTEXT_VTX_FLUSH |
+ R600_CONTEXT_STREAMOUT_FLUSH;
+ if (rctx->flags & mask) {
+ /* anything left (cb, vtx, shader, streamout) can be flushed
+ * in a more generic way
+ */
+ unsigned flags = 0;
+
+ flags |= S_0085F0_CB1_DEST_BASE_ENA(1) |
S_0085F0_DEST_BASE_0_ENA(1);
+ flags |= (rctx->flags & R600_CONTEXT_CB_FLUSH) ?
S_0085F0_CB_ACTION_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_DB_FLUSH) ?
S_0085F0_DB_ACTION_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_SH_FLUSH) ?
S_0085F0_SH_ACTION_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_TEX_FLUSH) ?
S_0085F0_TC_ACTION_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_VTX_FLUSH) ?
S_0085F0_VC_ACTION_ENA(1) : 0;
+ flags |= (rctx->flags & R600_CONTEXT_STREAMOUT_FLUSH) ?
S_0085F0_DEST_BASE_0_ENA(1) |
+
S_0085F0_SO0_DEST_BASE_ENA(1) |
+
S_0085F0_SO1_DEST_BASE_ENA(1) |
+
S_0085F0_SO2_DEST_BASE_ENA(1) |
+
S_0085F0_SO3_DEST_BASE_ENA(1) |
+
S_0085F0_SMX_ACTION_ENA(1) : 0;
+ rctx->flags &= ~mask;
+ cs->buf[cs->cdw++] = PKT3(PKT3_SURFACE_SYNC, 3, 0);
+ cs->buf[cs->cdw++] = flags; /* CP_COHER_CNTL */
+ cs->buf[cs->cdw++] = 0xffffffff; /* CP_COHER_SIZE */
+ cs->buf[cs->cdw++] = 0; /* CP_COHER_BASE */
+ cs->buf[cs->cdw++] = 0x0000000A; /* POLL_INTERVAL */
+ }
+
+ /* everything is properly flush */
+ rctx->flags &= R600_CONTEXT_DRAW_PENDING |
+ /* R600_CONTEXT_HTILE_ERRATA is persistant for whole cs */
+ R600_CONTEXT_HTILE_ERRATA;
}
void r600_context_flush(struct r600_context *ctx, unsigned flags)
@@ -1276,10 +1326,13 @@ void r600_context_flush(struct r600_context *ctx,
unsigned flags)
streamout_suspended = true;
}
- r600_flush_framebuffer(ctx, true);
-
- /* partial flush is needed to avoid lockups on some chips with user
fences */
- r600_context_ps_partial_flush(ctx);
+ ctx->flags |= R600_CONTEXT_DRAW_FLUSH;
+ ctx->flags |= R600_CONTEXT_FLUSH_AND_INV;
+ if (ctx->chip_class >= EVERGREEN) {
+ evergreen_flush_emit(ctx);
+ } else {
+ r600_flush_emit(ctx);
+ }
/* old kernels and userspace don't set SX_MISC, so we must reset it to
0 here */
if (ctx->chip_class <= R700) {
@@ -1298,11 +1351,9 @@ void r600_context_flush(struct r600_context *ctx,
unsigned flags)
/* Begin a new CS. */
r600_emit_atom(ctx, &ctx->start_cs_cmd.atom);
- /* Invalidate caches. */
- r600_inval_vertex_cache(ctx);
- r600_inval_texture_cache(ctx);
- r600_inval_shader_cache(ctx);
- r600_flush_framebuffer(ctx, false);
+ ctx->flags |= R600_CONTEXT_SH_FLUSH |
+ R600_CONTEXT_TEX_FLUSH |
+ R600_CONTEXT_VTX_FLUSH;
/* Re-emit states. */
r600_atom_dirty(ctx, &ctx->cb_misc_state.atom);
@@ -1357,7 +1408,10 @@ void r600_context_emit_fence(struct r600_context *ctx,
struct r600_resource *fen
va = r600_resource_va(&ctx->screen->screen, (void*)fence_bo);
va = va + (offset << 2);
- r600_context_ps_partial_flush(ctx);
+ ctx->flags &= ~R600_CONTEXT_DRAW_FLUSH;
+ cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
+ cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_PS_PARTIAL_FLUSH) |
EVENT_INDEX(4);
+
cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE_EOP, 4, 0);
cs->buf[cs->cdw++] =
EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5);
cs->buf[cs->cdw++] = va & 0xFFFFFFFFUL; /* ADDRESS_LO */
@@ -1518,7 +1572,7 @@ void r600_context_streamout_end(struct r600_context *ctx)
{
struct radeon_winsys_cs *cs = ctx->cs;
struct r600_so_target **t = ctx->so_targets;
- unsigned i, flush_flags = 0;
+ unsigned i;
uint64_t va;
if (ctx->chip_class >= EVERGREEN) {
@@ -1545,7 +1599,6 @@ void r600_context_streamout_end(struct r600_context *ctx)
r600_context_bo_reloc(ctx, t[i]->filled_size,
RADEON_USAGE_WRITE);
- flush_flags |= S_0085F0_SO0_DEST_BASE_ENA(1) << i;
}
}
@@ -1554,23 +1607,11 @@ void r600_context_streamout_end(struct r600_context
*ctx)
} else {
r600_set_streamout_enable(ctx, 0);
}
-
+ ctx->flags |= R600_CONTEXT_STREAMOUT_FLUSH;
/* This is needed to fix cache flushes on r600. */
if (ctx->chip_class == R600) {
- if (ctx->family == CHIP_RV670 ||
- ctx->family == CHIP_RS780 ||
- ctx->family == CHIP_RS880) {
- flush_flags |= S_0085F0_DEST_BASE_0_ENA(1);
- }
-
- r600_atom_dirty(ctx, &ctx->r6xx_flush_and_inv_cmd);
+ ctx->flags |= R600_CONTEXT_FLUSH_AND_INV;
}
-
- /* Flush streamout caches. */
- ctx->surface_sync_cmd.flush_flags |=
- S_0085F0_SMX_ACTION_ENA(1) | flush_flags;
- r600_atom_dirty(ctx, &ctx->surface_sync_cmd.atom);
-
ctx->num_cs_dw_streamout_end = 0;
#if 0
diff --git a/src/gallium/drivers/r600/r600_hw_context_priv.h
b/src/gallium/drivers/r600/r600_hw_context_priv.h
index 037d5e3..6929336 100644
--- a/src/gallium/drivers/r600/r600_hw_context_priv.h
+++ b/src/gallium/drivers/r600/r600_hw_context_priv.h
@@ -28,7 +28,8 @@
#include "r600_pipe.h"
-#define R600_MAX_DRAW_CS_DWORDS 11
+/* we alsoe here account size needed for flushing */
+#define R600_MAX_DRAW_CS_DWORDS 64
/* these flags are used in register flags and added into block flags */
#define REG_FLAG_NEED_BO 1
diff --git a/src/gallium/drivers/r600/r600_pipe.c
b/src/gallium/drivers/r600/r600_pipe.c
index 9f20560..07a398f 100644
--- a/src/gallium/drivers/r600/r600_pipe.c
+++ b/src/gallium/drivers/r600/r600_pipe.c
@@ -231,8 +231,6 @@ static struct pipe_context *r600_create_context(struct
pipe_screen *screen, void
rctx->context.create_video_decoder = vl_create_decoder;
rctx->context.create_video_buffer = vl_video_buffer_create;
- r600_init_common_atoms(rctx);
-
switch (rctx->chip_class) {
case R600:
case R700:
diff --git a/src/gallium/drivers/r600/r600_pipe.h
b/src/gallium/drivers/r600/r600_pipe.h
index 5ff4d51..4add90c 100644
--- a/src/gallium/drivers/r600/r600_pipe.h
+++ b/src/gallium/drivers/r600/r600_pipe.h
@@ -341,9 +341,7 @@ struct r600_context {
struct r600_command_buffer start_cs_cmd; /* invariant state mostly
*/
/** Compute specific registers initializations. The start_cs_cmd atom
* must be emitted before start_compute_cs_cmd. */
- struct r600_command_buffer start_compute_cs_cmd;
- struct r600_surface_sync_cmd surface_sync_cmd;
- struct r600_atom r6xx_flush_and_inv_cmd;
+ struct r600_command_buffer start_compute_cs_cmd;
struct r600_cb_misc_state cb_misc_state;
struct r600_db_misc_state db_misc_state;
/** Vertex buffers for fetch shaders */
@@ -528,8 +526,6 @@ void r600_translate_index_buffer(struct r600_context *r600,
void r600_init_atom(struct r600_atom *atom,
void (*emit)(struct r600_context *ctx, struct r600_atom
*state),
unsigned num_dw, enum r600_atom_flags flags);
-void r600_init_common_atoms(struct r600_context *rctx);
-unsigned r600_get_cb_flush_flags(struct r600_context *rctx);
void r600_texture_barrier(struct pipe_context *ctx);
void r600_set_index_buffer(struct pipe_context *ctx,
const struct pipe_index_buffer *ib);
diff --git a/src/gallium/drivers/r600/r600_state.c
b/src/gallium/drivers/r600/r600_state.c
index 8925a23..3f43a9d 100644
--- a/src/gallium/drivers/r600/r600_state.c
+++ b/src/gallium/drivers/r600/r600_state.c
@@ -1133,8 +1133,9 @@ static void r600_set_sampler_views(struct r600_context
*rctx,
struct r600_pipe_sampler_view **rviews = (struct r600_pipe_sampler_view
**)views;
unsigned i;
- if (count)
- r600_inval_texture_cache(rctx);
+ if (count) {
+ rctx->flags |= R600_CONTEXT_TEX_FLUSH;
+ }
for (i = 0; i < count; i++) {
if (rviews[i]) {
@@ -1632,6 +1633,8 @@ static void r600_db(struct r600_context *rctx, struct
r600_pipe_state *rstate,
if (rtex->hyperz) {
uint64_t htile_offset =
rtex->hyperz->surface.level[level].offset;
+ rctx->flags |= R600_CONTEXT_FLUSH_AND_INV;
+ rctx->flags |= R600_CONTEXT_HTILE_ERRATA;
rctx->db_misc_state.hyperz = true;
rctx->db_misc_state.db_htile_surface_mask = 0xffffffff;
r600_atom_dirty(rctx, &rctx->db_misc_state.atom);
@@ -1676,7 +1679,10 @@ static void r600_set_framebuffer_state(struct
pipe_context *ctx,
if (rstate == NULL)
return;
- r600_flush_framebuffer(rctx, false);
+ /* the htile errata is also needed for cb should probably rename */
+ rctx->flags |= R600_CONTEXT_CB_FLUSH |
+ R600_CONTEXT_DRAW_FLUSH |
+ R600_CONTEXT_FLUSH_AND_INV;
/* unreference old buffer and reference new one */
rstate->id = R600_PIPE_STATE_FRAMEBUFFER;
@@ -1692,6 +1698,7 @@ static void r600_set_framebuffer_state(struct
pipe_context *ctx,
r600_cb(rctx, rstate, state, i);
}
if (state->zsbuf) {
+ rctx->flags |= R600_CONTEXT_DB_FLUSH;
r600_db(rctx, rstate, state);
}
@@ -1962,14 +1969,8 @@ void r600_adjust_gprs(struct r600_context *rctx)
unsigned tmp;
int diff;
- /* XXX: Following call moved from r600_bind_[ps|vs]_shader,
- * it seems eg+ doesn't need it, r6xx/7xx probably need it only for
- * adjusting the GPR allocation?
- * Do we need this if we aren't really changing config below? */
- r600_inval_shader_cache(rctx);
-
- if (rctx->ps_shader->current->shader.bc.ngpr > rctx->default_ps_gprs)
- {
+ rctx->flags |= R600_CONTEXT_SH_FLUSH;
+ if (rctx->ps_shader->current->shader.bc.ngpr > rctx->default_ps_gprs) {
diff = rctx->ps_shader->current->shader.bc.ngpr -
rctx->default_ps_gprs;
num_vs_gprs -= diff;
num_ps_gprs += diff;
diff --git a/src/gallium/drivers/r600/r600_state_common.c
b/src/gallium/drivers/r600/r600_state_common.c
index 3c42a44..c9fd362 100644
--- a/src/gallium/drivers/r600/r600_state_common.c
+++ b/src/gallium/drivers/r600/r600_state_common.c
@@ -56,27 +56,6 @@ void r600_release_command_buffer(struct r600_command_buffer
*cb)
FREE(cb->buf);
}
-static void r600_emit_surface_sync(struct r600_context *rctx, struct r600_atom
*atom)
-{
- struct radeon_winsys_cs *cs = rctx->cs;
- struct r600_surface_sync_cmd *a = (struct r600_surface_sync_cmd*)atom;
-
- cs->buf[cs->cdw++] = PKT3(PKT3_SURFACE_SYNC, 3, 0);
- cs->buf[cs->cdw++] = a->flush_flags; /* CP_COHER_CNTL */
- cs->buf[cs->cdw++] = 0xffffffff; /* CP_COHER_SIZE */
- cs->buf[cs->cdw++] = 0; /* CP_COHER_BASE */
- cs->buf[cs->cdw++] = 0x0000000A; /* POLL_INTERVAL */
-
- a->flush_flags = 0;
-}
-
-static void r600_emit_r6xx_flush_and_inv(struct r600_context *rctx, struct
r600_atom *atom)
-{
- struct radeon_winsys_cs *cs = rctx->cs;
- cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
- cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT) |
EVENT_INDEX(0);
-}
-
void r600_init_atom(struct r600_atom *atom,
void (*emit)(struct r600_context *ctx, struct r600_atom
*state),
unsigned num_dw, enum r600_atom_flags flags)
@@ -86,37 +65,11 @@ void r600_init_atom(struct r600_atom *atom,
atom->flags = flags;
}
-void r600_init_common_atoms(struct r600_context *rctx)
-{
- r600_init_atom(&rctx->surface_sync_cmd.atom, r600_emit_surface_sync,
5, EMIT_EARLY);
- r600_init_atom(&rctx->r6xx_flush_and_inv_cmd,
r600_emit_r6xx_flush_and_inv, 2, EMIT_EARLY);
-}
-
-unsigned r600_get_cb_flush_flags(struct r600_context *rctx)
-{
- unsigned flags = 0;
-
- if (rctx->framebuffer.nr_cbufs) {
- flags |= S_0085F0_CB_ACTION_ENA(1) |
- (((1 << rctx->framebuffer.nr_cbufs) - 1) <<
S_0085F0_CB0_DEST_BASE_ENA_SHIFT);
- }
-
- /* Workaround for broken flushing on some R6xx chipsets. */
- if (rctx->family == CHIP_RV670 ||
- rctx->family == CHIP_RS780 ||
- rctx->family == CHIP_RS880) {
- flags |= S_0085F0_CB1_DEST_BASE_ENA(1) |
- S_0085F0_DEST_BASE_0_ENA(1);
- }
- return flags;
-}
-
void r600_texture_barrier(struct pipe_context *ctx)
{
struct r600_context *rctx = (struct r600_context *)ctx;
- rctx->surface_sync_cmd.flush_flags |= S_0085F0_TC_ACTION_ENA(1) |
r600_get_cb_flush_flags(rctx);
- r600_atom_dirty(rctx, &rctx->surface_sync_cmd.atom);
+ rctx->flags |= R600_CONTEXT_DRAW_FLUSH | R600_CONTEXT_CB_FLUSH |
R600_CONTEXT_TEX_FLUSH;
}
static bool r600_conv_pipe_prim(unsigned pprim, unsigned *prim)
@@ -368,7 +321,7 @@ void r600_bind_vertex_elements(struct pipe_context *ctx,
void *state)
rctx->vertex_elements = v;
if (v) {
- r600_inval_shader_cache(rctx);
+ rctx->flags |= R600_CONTEXT_SH_FLUSH;
rctx->states[v->rstate.id] = &v->rstate;
r600_context_pipe_state_set(rctx, &v->rstate);
@@ -412,9 +365,9 @@ void r600_set_vertex_buffers(struct pipe_context *ctx,
unsigned count,
util_copy_vertex_buffers(rctx->vertex_buffer, &rctx->nr_vertex_buffers,
buffers, count);
- r600_inval_vertex_cache(rctx);
+ rctx->flags |= R600_CONTEXT_VTX_FLUSH;
state->atom.num_dw = (rctx->chip_class >= EVERGREEN ? 12 : 11) *
- rctx->nr_vertex_buffers;
+ rctx->nr_vertex_buffers;
for (i = 0 ; i < rctx->nr_vertex_buffers; i++) {
state->dirty_mask |= 1 << i;
}
@@ -523,6 +476,9 @@ static int r600_shader_select(struct pipe_context *ctx,
if (dirty)
*dirty = 1;
+ /* we are changing shader need a flush */
+ rctx->flags |= R600_CONTEXT_SH_FLUSH;
+
shader->next_variant = sel->current;
sel->current = shader;
@@ -667,7 +623,7 @@ static void r600_update_alpha_ref(struct r600_context *rctx)
void r600_constant_buffers_dirty(struct r600_context *rctx, struct
r600_constbuf_state *state)
{
- r600_inval_shader_cache(rctx);
+ rctx->flags |= R600_CONTEXT_SH_FLUSH;
state->atom.num_dw = rctx->chip_class >= EVERGREEN ?
util_bitcount(state->dirty_mask)*20
:
util_bitcount(state->dirty_mask)*19;
r600_atom_dirty(rctx, &state->atom);
@@ -893,6 +849,12 @@ void r600_draw_vbo(struct pipe_context *ctx, const struct
pipe_draw_info *dinfo)
r600_update_derived_state(rctx);
+ if (rctx->chip_class >= EVERGREEN) {
+ evergreen_flush_emit(rctx);
+ } else {
+ r600_flush_emit(rctx);
+ }
+
if (info.indexed) {
/* Initialize the index buffer struct. */
pipe_resource_reference(&ib.buffer, rctx->index_buffer.buffer);
@@ -999,7 +961,7 @@ void r600_draw_vbo(struct pipe_context *ctx, const struct
pipe_draw_info *dinfo)
(info.count_from_stream_output ?
S_0287F0_USE_OPAQUE(1) : 0);
}
- rctx->flags |= R600_CONTEXT_DST_CACHES_DIRTY |
R600_CONTEXT_DRAW_PENDING;
+ rctx->flags |= R600_CONTEXT_DRAW_PENDING;
/* Set the depth buffer as dirty. */
if (rctx->framebuffer.zsbuf) {
--
1.7.10.4
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/mesa-dev
