[Mesa-dev] [PATCH 3/7] radeonsi: merge si_set_streamout_targets with si_common_set_streamout_targets
Marek Olšák
maraeo at gmail.com
Sat Oct 7 22:47:02 UTC 2017
From: Marek Olšák <marek.olsak at amd.com>
---
src/gallium/drivers/radeonsi/si_descriptors.c | 114 ---------------------
src/gallium/drivers/radeonsi/si_state.h | 4 -
src/gallium/drivers/radeonsi/si_state_streamout.c | 117 ++++++++++++++++++++--
3 files changed, 109 insertions(+), 126 deletions(-)
diff --git a/src/gallium/drivers/radeonsi/si_descriptors.c b/src/gallium/drivers/radeonsi/si_descriptors.c
index dd1f1e9..cda7d94 100644
--- a/src/gallium/drivers/radeonsi/si_descriptors.c
+++ b/src/gallium/drivers/radeonsi/si_descriptors.c
@@ -1356,133 +1356,20 @@ void si_set_ring_buffer(struct pipe_context *ctx, uint slot,
buffers->enabled_mask |= 1u << slot;
} else {
/* Clear the descriptor. */
memset(descs->list + slot*4, 0, sizeof(uint32_t) * 4);
buffers->enabled_mask &= ~(1u << slot);
}
sctx->descriptors_dirty |= 1u << SI_DESCS_RW_BUFFERS;
}
-/* STREAMOUT BUFFERS */
-
-static void si_set_streamout_targets(struct pipe_context *ctx,
- unsigned num_targets,
- struct pipe_stream_output_target **targets,
- const unsigned *offsets)
-{
- struct si_context *sctx = (struct si_context *)ctx;
- struct si_buffer_resources *buffers = &sctx->rw_buffers;
- struct si_descriptors *descs = &sctx->descriptors[SI_DESCS_RW_BUFFERS];
- unsigned old_num_targets = sctx->streamout.num_targets;
- unsigned i, bufidx;
-
- /* We are going to unbind the buffers. Mark which caches need to be flushed. */
- if (sctx->streamout.num_targets && sctx->streamout.begin_emitted) {
- /* Since streamout uses vector writes which go through TC L2
- * and most other clients can use TC L2 as well, we don't need
- * to flush it.
- *
- * The only cases which requires flushing it is VGT DMA index
- * fetching (on <= CIK) and indirect draw data, which are rare
- * cases. Thus, flag the TC L2 dirtiness in the resource and
- * handle it at draw call time.
- */
- for (i = 0; i < sctx->streamout.num_targets; i++)
- if (sctx->streamout.targets[i])
- r600_resource(sctx->streamout.targets[i]->b.buffer)->TC_L2_dirty = true;
-
- /* Invalidate the scalar cache in case a streamout buffer is
- * going to be used as a constant buffer.
- *
- * Invalidate TC L1, because streamout bypasses it (done by
- * setting GLC=1 in the store instruction), but it can contain
- * outdated data of streamout buffers.
- *
- * VS_PARTIAL_FLUSH is required if the buffers are going to be
- * used as an input immediately.
- */
- sctx->b.flags |= SI_CONTEXT_INV_SMEM_L1 |
- SI_CONTEXT_INV_VMEM_L1 |
- SI_CONTEXT_VS_PARTIAL_FLUSH;
- }
-
- /* All readers of the streamout targets need to be finished before we can
- * start writing to the targets.
- */
- if (num_targets)
- sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
- SI_CONTEXT_CS_PARTIAL_FLUSH;
-
- /* Streamout buffers must be bound in 2 places:
- * 1) in VGT by setting the VGT_STRMOUT registers
- * 2) as shader resources
- */
-
- /* Set the VGT regs. */
- si_common_set_streamout_targets(ctx, num_targets, targets, offsets);
-
- /* Set the shader resources.*/
- for (i = 0; i < num_targets; i++) {
- bufidx = SI_VS_STREAMOUT_BUF0 + i;
-
- if (targets[i]) {
- struct pipe_resource *buffer = targets[i]->buffer;
- uint64_t va = r600_resource(buffer)->gpu_address;
-
- /* Set the descriptor.
- *
- * On VI, the format must be non-INVALID, otherwise
- * the buffer will be considered not bound and store
- * instructions will be no-ops.
- */
- uint32_t *desc = descs->list + bufidx*4;
- desc[0] = va;
- desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32);
- desc[2] = 0xffffffff;
- desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
- S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
- S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
- S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
- S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
-
- /* Set the resource. */
- pipe_resource_reference(&buffers->buffers[bufidx],
- buffer);
- radeon_add_to_buffer_list_check_mem(&sctx->b, &sctx->b.gfx,
- (struct r600_resource*)buffer,
- buffers->shader_usage,
- RADEON_PRIO_SHADER_RW_BUFFER,
- true);
- r600_resource(buffer)->bind_history |= PIPE_BIND_STREAM_OUTPUT;
-
- buffers->enabled_mask |= 1u << bufidx;
- } else {
- /* Clear the descriptor and unset the resource. */
- memset(descs->list + bufidx*4, 0,
- sizeof(uint32_t) * 4);
- pipe_resource_reference(&buffers->buffers[bufidx],
- NULL);
- buffers->enabled_mask &= ~(1u << bufidx);
- }
- }
- for (; i < old_num_targets; i++) {
- bufidx = SI_VS_STREAMOUT_BUF0 + i;
- /* Clear the descriptor and unset the resource. */
- memset(descs->list + bufidx*4, 0, sizeof(uint32_t) * 4);
- pipe_resource_reference(&buffers->buffers[bufidx], NULL);
- buffers->enabled_mask &= ~(1u << bufidx);
- }
-
- sctx->descriptors_dirty |= 1u << SI_DESCS_RW_BUFFERS;
-}
-
static void si_desc_reset_buffer_offset(struct pipe_context *ctx,
uint32_t *desc, uint64_t old_buf_va,
struct pipe_resource *new_buf)
{
/* Retrieve the buffer offset from the descriptor. */
uint64_t old_desc_va =
desc[0] | ((uint64_t)G_008F04_BASE_ADDRESS_HI(desc[1]) << 32);
assert(old_buf_va <= old_desc_va);
uint64_t offset_within_buffer = old_desc_va - old_buf_va;
@@ -2687,21 +2574,20 @@ void si_init_all_descriptors(struct si_context *sctx)
sctx->descriptors_dirty = u_bit_consecutive(0, SI_NUM_DESCS);
/* 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;
sctx->b.b.set_polygon_stipple = si_set_polygon_stipple;
sctx->b.b.set_shader_buffers = si_set_shader_buffers;
sctx->b.b.set_sampler_views = si_set_sampler_views;
- sctx->b.b.set_stream_output_targets = si_set_streamout_targets;
sctx->b.b.create_texture_handle = si_create_texture_handle;
sctx->b.b.delete_texture_handle = si_delete_texture_handle;
sctx->b.b.make_texture_handle_resident = si_make_texture_handle_resident;
sctx->b.b.create_image_handle = si_create_image_handle;
sctx->b.b.delete_image_handle = si_delete_image_handle;
sctx->b.b.make_image_handle_resident = si_make_image_handle_resident;
sctx->b.invalidate_buffer = si_invalidate_buffer;
sctx->b.rebind_buffer = si_rebind_buffer;
/* Shader user data. */
diff --git a/src/gallium/drivers/radeonsi/si_state.h b/src/gallium/drivers/radeonsi/si_state.h
index 9d29878..a686d0d 100644
--- a/src/gallium/drivers/radeonsi/si_state.h
+++ b/src/gallium/drivers/radeonsi/si_state.h
@@ -418,24 +418,20 @@ void si_draw_rectangle(struct blitter_context *blitter,
void *vertex_elements_cso,
blitter_get_vs_func get_vs,
int x1, int y1, int x2, int y2,
float depth, unsigned num_instances,
enum blitter_attrib_type type,
const union blitter_attrib *attrib);
void si_trace_emit(struct si_context *sctx);
/* si_state_streamout.c */
void si_streamout_buffers_dirty(struct si_context *sctx);
-void si_common_set_streamout_targets(struct pipe_context *ctx,
- unsigned num_targets,
- struct pipe_stream_output_target **targets,
- const unsigned *offset);
void si_emit_streamout_end(struct si_context *sctx);
void si_update_prims_generated_query_state(struct si_context *sctx,
unsigned type, int diff);
void si_init_streamout_functions(struct si_context *sctx);
static inline unsigned
si_tile_mode_index(struct r600_texture *rtex, unsigned level, bool stencil)
{
if (stencil)
diff --git a/src/gallium/drivers/radeonsi/si_state_streamout.c b/src/gallium/drivers/radeonsi/si_state_streamout.c
index 92c5c59..35fbcea 100644
--- a/src/gallium/drivers/radeonsi/si_state_streamout.c
+++ b/src/gallium/drivers/radeonsi/si_state_streamout.c
@@ -19,20 +19,21 @@
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
* Authors: Marek Olšák <maraeo at gmail.com>
*
*/
#include "si_pipe.h"
#include "si_state.h"
+#include "sid.h"
#include "radeon/r600_cs.h"
#include "util/u_memory.h"
static void si_set_streamout_enable(struct si_context *sctx, bool enable);
static inline void si_so_target_reference(struct si_streamout_target **dst,
struct pipe_stream_output_target *src)
{
pipe_so_target_reference((struct pipe_stream_output_target**)dst, src);
@@ -83,60 +84,159 @@ static void si_so_target_destroy(struct pipe_context *ctx,
void si_streamout_buffers_dirty(struct si_context *sctx)
{
if (!sctx->streamout.enabled_mask)
return;
si_mark_atom_dirty(sctx, &sctx->streamout.begin_atom);
si_set_streamout_enable(sctx, true);
}
-void si_common_set_streamout_targets(struct pipe_context *ctx,
+static void si_set_streamout_targets(struct pipe_context *ctx,
unsigned num_targets,
struct pipe_stream_output_target **targets,
const unsigned *offsets)
{
struct si_context *sctx = (struct si_context *)ctx;
- unsigned i;
- unsigned enabled_mask = 0, append_bitmask = 0;
+ struct si_buffer_resources *buffers = &sctx->rw_buffers;
+ struct si_descriptors *descs = &sctx->descriptors[SI_DESCS_RW_BUFFERS];
+ unsigned old_num_targets = sctx->streamout.num_targets;
+ unsigned i, bufidx;
- /* Stop streamout. */
+ /* We are going to unbind the buffers. Mark which caches need to be flushed. */
if (sctx->streamout.num_targets && sctx->streamout.begin_emitted) {
- si_emit_streamout_end(sctx);
+ /* Since streamout uses vector writes which go through TC L2
+ * and most other clients can use TC L2 as well, we don't need
+ * to flush it.
+ *
+ * The only cases which requires flushing it is VGT DMA index
+ * fetching (on <= CIK) and indirect draw data, which are rare
+ * cases. Thus, flag the TC L2 dirtiness in the resource and
+ * handle it at draw call time.
+ */
+ for (i = 0; i < sctx->streamout.num_targets; i++)
+ if (sctx->streamout.targets[i])
+ r600_resource(sctx->streamout.targets[i]->b.buffer)->TC_L2_dirty = true;
+
+ /* Invalidate the scalar cache in case a streamout buffer is
+ * going to be used as a constant buffer.
+ *
+ * Invalidate TC L1, because streamout bypasses it (done by
+ * setting GLC=1 in the store instruction), but it can contain
+ * outdated data of streamout buffers.
+ *
+ * VS_PARTIAL_FLUSH is required if the buffers are going to be
+ * used as an input immediately.
+ */
+ sctx->b.flags |= SI_CONTEXT_INV_SMEM_L1 |
+ SI_CONTEXT_INV_VMEM_L1 |
+ SI_CONTEXT_VS_PARTIAL_FLUSH;
}
+ /* All readers of the streamout targets need to be finished before we can
+ * start writing to the targets.
+ */
+ if (num_targets)
+ sctx->b.flags |= SI_CONTEXT_PS_PARTIAL_FLUSH |
+ SI_CONTEXT_CS_PARTIAL_FLUSH;
+
+ /* Streamout buffers must be bound in 2 places:
+ * 1) in VGT by setting the VGT_STRMOUT registers
+ * 2) as shader resources
+ */
+
+ /* Stop streamout. */
+ if (sctx->streamout.num_targets && sctx->streamout.begin_emitted)
+ si_emit_streamout_end(sctx);
+
/* Set the new targets. */
+ unsigned enabled_mask = 0, append_bitmask = 0;
for (i = 0; i < num_targets; i++) {
si_so_target_reference(&sctx->streamout.targets[i], targets[i]);
if (!targets[i])
continue;
r600_context_add_resource_size(ctx, targets[i]->buffer);
enabled_mask |= 1 << i;
+
if (offsets[i] == ((unsigned)-1))
append_bitmask |= 1 << i;
}
- for (; i < sctx->streamout.num_targets; i++) {
+
+ for (; i < sctx->streamout.num_targets; i++)
si_so_target_reference(&sctx->streamout.targets[i], NULL);
- }
sctx->streamout.enabled_mask = enabled_mask;
-
sctx->streamout.num_targets = num_targets;
sctx->streamout.append_bitmask = append_bitmask;
+ /* Update dirty state bits. */
if (num_targets) {
si_streamout_buffers_dirty(sctx);
} else {
si_set_atom_dirty(sctx, &sctx->streamout.begin_atom, false);
si_set_streamout_enable(sctx, false);
}
+
+ /* Set the shader resources.*/
+ for (i = 0; i < num_targets; i++) {
+ bufidx = SI_VS_STREAMOUT_BUF0 + i;
+
+ if (targets[i]) {
+ struct pipe_resource *buffer = targets[i]->buffer;
+ uint64_t va = r600_resource(buffer)->gpu_address;
+
+ /* Set the descriptor.
+ *
+ * On VI, the format must be non-INVALID, otherwise
+ * the buffer will be considered not bound and store
+ * instructions will be no-ops.
+ */
+ uint32_t *desc = descs->list + bufidx*4;
+ desc[0] = va;
+ desc[1] = S_008F04_BASE_ADDRESS_HI(va >> 32);
+ desc[2] = 0xffffffff;
+ desc[3] = S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+ S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+ S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+ S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+ S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32);
+
+ /* Set the resource. */
+ pipe_resource_reference(&buffers->buffers[bufidx],
+ buffer);
+ radeon_add_to_buffer_list_check_mem(&sctx->b, &sctx->b.gfx,
+ (struct r600_resource*)buffer,
+ buffers->shader_usage,
+ RADEON_PRIO_SHADER_RW_BUFFER,
+ true);
+ r600_resource(buffer)->bind_history |= PIPE_BIND_STREAM_OUTPUT;
+
+ buffers->enabled_mask |= 1u << bufidx;
+ } else {
+ /* Clear the descriptor and unset the resource. */
+ memset(descs->list + bufidx*4, 0,
+ sizeof(uint32_t) * 4);
+ pipe_resource_reference(&buffers->buffers[bufidx],
+ NULL);
+ buffers->enabled_mask &= ~(1u << bufidx);
+ }
+ }
+ for (; i < old_num_targets; i++) {
+ bufidx = SI_VS_STREAMOUT_BUF0 + i;
+ /* Clear the descriptor and unset the resource. */
+ memset(descs->list + bufidx*4, 0, sizeof(uint32_t) * 4);
+ pipe_resource_reference(&buffers->buffers[bufidx], NULL);
+ buffers->enabled_mask &= ~(1u << bufidx);
+ }
+
+ sctx->descriptors_dirty |= 1u << SI_DESCS_RW_BUFFERS;
}
static void si_flush_vgt_streamout(struct si_context *sctx)
{
struct radeon_winsys_cs *cs = sctx->b.gfx.cs;
unsigned reg_strmout_cntl;
/* The register is at different places on different ASICs. */
if (sctx->b.chip_class >= CIK) {
reg_strmout_cntl = R_0300FC_CP_STRMOUT_CNTL;
@@ -306,13 +406,14 @@ void si_update_prims_generated_query_state(struct si_context *sctx,
if (old_strmout_en != si_get_strmout_en(sctx))
si_mark_atom_dirty(sctx, &sctx->streamout.enable_atom);
}
}
void si_init_streamout_functions(struct si_context *sctx)
{
sctx->b.b.create_stream_output_target = si_create_so_target;
sctx->b.b.stream_output_target_destroy = si_so_target_destroy;
+ sctx->b.b.set_stream_output_targets = si_set_streamout_targets;
sctx->streamout.begin_atom.emit = si_emit_streamout_begin;
sctx->streamout.enable_atom.emit = si_emit_streamout_enable;
}
--
2.7.4
More information about the mesa-dev
mailing list