freedreno/a6xx: Split tile loads and clears

This will give better visibility in perfetto, and prepares for the next
commit where we could have per-subpass clears.

While we are at it, start adopting vulkan terms for tile load/store.  No
need to be pointlessly different.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22895>

src/gallium/drivers/freedreno/a2xx/fd2_gmem.c

@@ -88,7 +88,7 @@ static void
 emit_gmem2mem_surf(struct fd_batch *batch, uint32_t base,
                    struct pipe_surface *psurf)
 {
-   struct fd_ringbuffer *ring = batch->tile_fini;
+   struct fd_ringbuffer *ring = batch->tile_store;
    struct fd_resource *rsc = fd_resource(psurf->texture);
    uint32_t offset =
       fd_resource_offset(rsc, psurf->u.tex.level, psurf->u.tex.first_layer);
@@ -141,9 +141,9 @@ prepare_tile_fini_ib(struct fd_batch *batch) assert_dt
    struct pipe_framebuffer_state *pfb = &batch->framebuffer;
    struct fd_ringbuffer *ring;
 
-   batch->tile_fini =
+   batch->tile_store =
       fd_submit_new_ringbuffer(batch->submit, 0x1000, FD_RINGBUFFER_STREAMING);
-   ring = batch->tile_fini;
+   ring = batch->tile_store;
 
    fd2_emit_vertex_bufs(ring, 0x9c,
                         (struct fd2_vertex_buf[]){
@@ -223,7 +223,7 @@ prepare_tile_fini_ib(struct fd_batch *batch) assert_dt
 static void
 fd2_emit_tile_gmem2mem(struct fd_batch *batch, const struct fd_tile *tile)
 {
-   fd2_emit_ib(batch->gmem, batch->tile_fini);
+   fd2_emit_ib(batch->gmem, batch->tile_store);
 }
 
 /* transfer from system memory to gmem */

src/gallium/drivers/freedreno/a6xx/fd6_gmem.cc

@@ -848,9 +848,9 @@ emit_msaa(struct fd_ringbuffer *ring, unsigned nr)
    OUT_RING(ring, A6XX_RB_BLIT_GMEM_MSAA_CNTL_SAMPLES(samples));
 }
 
-static void prepare_tile_setup_ib(struct fd_batch *batch);
+static void prepare_tile_setup(struct fd_batch *batch);
 template <chip CHIP>
-static void prepare_tile_fini_ib(struct fd_batch *batch);
+static void prepare_tile_fini(struct fd_batch *batch);
 
 /* before first tile */
 template <chip CHIP>
@@ -874,8 +874,8 @@ fd6_emit_tile_init(struct fd_batch *batch) assert_dt
 
    fd6_cache_inv(batch, ring);
 
-   prepare_tile_setup_ib(batch);
-   prepare_tile_fini_ib<CHIP>(batch);
+   prepare_tile_setup(batch);
+   prepare_tile_fini<CHIP>(batch);
 
    OUT_PKT7(ring, CP_SKIP_IB2_ENABLE_GLOBAL, 1);
    OUT_RING(ring, 0x0);
@@ -1316,18 +1316,23 @@ emit_restore_blits(struct fd_batch *batch, struct fd_ringbuffer *ring)
 }
 
 static void
-prepare_tile_setup_ib(struct fd_batch *batch)
+prepare_tile_setup(struct fd_batch *batch)
 {
-   if (!(batch->restore || batch->fast_cleared))
-      return;
+   if (batch->restore) {
+      batch->tile_loads =
+         fd_submit_new_ringbuffer(batch->submit, 0x1000, FD_RINGBUFFER_STREAMING);
 
-   batch->tile_setup =
-      fd_submit_new_ringbuffer(batch->submit, 0x1000, FD_RINGBUFFER_STREAMING);
+      set_blit_scissor(batch, batch->tile_loads);
+      emit_restore_blits(batch, batch->tile_loads);
+   }
 
-   set_blit_scissor(batch, batch->tile_setup);
+   if (batch->fast_cleared) {
+      batch->tile_clears =
+         fd_submit_new_ringbuffer(batch->submit, 0x1000, FD_RINGBUFFER_STREAMING);
 
-   emit_restore_blits(batch, batch->tile_setup);
-   emit_clears(batch, batch->tile_setup);
+      set_blit_scissor(batch, batch->tile_clears);
+      emit_clears(batch, batch->tile_clears);
+   }
 }
 
 /*
@@ -1342,10 +1347,16 @@ fd6_emit_tile_mem2gmem(struct fd_batch *batch, const struct fd_tile *tile)
 static void
 fd6_emit_tile_renderprep(struct fd_batch *batch, const struct fd_tile *tile)
 {
-   if (batch->tile_setup) {
-      trace_start_clear_restore(&batch->trace, batch->gmem, batch->fast_cleared, batch->restore);
-      emit_conditional_ib(batch, tile, batch->tile_setup);
-      trace_end_clear_restore(&batch->trace, batch->gmem);
+   if (batch->tile_loads) {
+      trace_start_tile_loads(&batch->trace, batch->gmem, batch->restore);
+      emit_conditional_ib(batch, tile, batch->tile_loads);
+      trace_end_tile_loads(&batch->trace, batch->gmem);
+   }
+
+   if (batch->tile_clears) {
+      trace_start_clears(&batch->trace, batch->gmem, batch->fast_cleared);
+      emit_conditional_ib(batch, tile, batch->tile_clears);
+      trace_end_clears(&batch->trace, batch->gmem);
    }
 }
 
@@ -1465,15 +1476,16 @@ emit_resolve_blit(struct fd_batch *batch, struct fd_ringbuffer *ring,
 
 template <chip CHIP>
 static void
-prepare_tile_fini_ib(struct fd_batch *batch) assert_dt
+prepare_tile_fini(struct fd_batch *batch)
+   assert_dt
 {
    const struct fd_gmem_stateobj *gmem = batch->gmem_state;
    struct pipe_framebuffer_state *pfb = &batch->framebuffer;
    struct fd_ringbuffer *ring;
 
-   batch->tile_fini =
+   batch->tile_store =
       fd_submit_new_ringbuffer(batch->submit, 0x1000, FD_RINGBUFFER_STREAMING);
-   ring = batch->tile_fini;
+   ring = batch->tile_store;
 
    set_blit_scissor(batch, ring);
 
@@ -1540,10 +1552,10 @@ fd6_emit_tile_gmem2mem(struct fd_batch *batch, const struct fd_tile *tile)
    OUT_RING(ring, A6XX_CP_SET_MARKER_0_MODE(RM6_RESOLVE));
    emit_marker6(ring, 7);
 
-   if (batch->tile_fini) {
-      trace_start_resolve(&batch->trace, batch->gmem, batch->resolve);
-      emit_conditional_ib(batch, tile, batch->tile_fini);
-      trace_end_resolve(&batch->trace, batch->gmem);
+   if (batch->tile_store) {
+      trace_start_tile_stores(&batch->trace, batch->gmem, batch->resolve);
+      emit_conditional_ib(batch, tile, batch->tile_store);
+      trace_end_tile_stores(&batch->trace, batch->gmem);
    }
 }
 
@@ -1581,7 +1593,7 @@ emit_sysmem_clears(struct fd_batch *batch, struct fd_ringbuffer *ring) assert_dt
    struct pipe_box box2d;
    u_box_2d(0, 0, pfb->width, pfb->height, &box2d);
 
-   trace_start_clear_restore(&batch->trace, ring, buffers, 0);
+   trace_start_clears(&batch->trace, ring, buffers);
 
    if (buffers & PIPE_CLEAR_COLOR) {
       for (int i = 0; i < pfb->nr_cbufs; i++) {
@@ -1626,7 +1638,7 @@ emit_sysmem_clears(struct fd_batch *batch, struct fd_ringbuffer *ring) assert_dt
    fd6_event_write(batch, ring, PC_CCU_FLUSH_COLOR_TS, true);
    fd_wfi(batch, ring);
 
-   trace_end_clear_restore(&batch->trace, ring);
+   trace_end_clears(&batch->trace, ring);
 }
 
 template <chip CHIP>

src/gallium/drivers/freedreno/freedreno_batch.c

@@ -153,14 +153,19 @@ cleanup_submit(struct fd_batch *batch)
       batch->epilogue = NULL;
    }
 
-   if (batch->tile_setup) {
-      fd_ringbuffer_del(batch->tile_setup);
-      batch->tile_setup = NULL;
+   if (batch->tile_loads) {
+      fd_ringbuffer_del(batch->tile_loads);
+      batch->tile_loads = NULL;
    }
 
-   if (batch->tile_fini) {
-      fd_ringbuffer_del(batch->tile_fini);
-      batch->tile_fini = NULL;
+   if (batch->tile_clears) {
+      fd_ringbuffer_del(batch->tile_clears);
+      batch->tile_clears = NULL;
+   }
+
+   if (batch->tile_store) {
+      fd_ringbuffer_del(batch->tile_store);
+      batch->tile_store = NULL;
    }
 
    fd_submit_del(batch->submit);

src/gallium/drivers/freedreno/freedreno_batch.h

@@ -206,8 +206,9 @@ struct fd_batch {
    /** epilogue cmdstream (executed after all tiles): */
    struct fd_ringbuffer *epilogue;
 
-   struct fd_ringbuffer *tile_setup;
-   struct fd_ringbuffer *tile_fini;
+   struct fd_ringbuffer *tile_loads;
+   struct fd_ringbuffer *tile_clears;
+   struct fd_ringbuffer *tile_store;
 
    union pipe_color_union clear_color[MAX_RENDER_TARGETS];
    double clear_depth;

src/gallium/drivers/freedreno/freedreno_perfetto.cc

@@ -464,35 +464,51 @@ fd_end_compute(struct pipe_context *pctx, uint64_t ts_ns,
 }
 
 void
-fd_start_clear_restore(struct pipe_context *pctx, uint64_t ts_ns,
-                       const void *flush_data,
-                       const struct trace_start_clear_restore *payload)
+fd_start_clears(struct pipe_context *pctx, uint64_t ts_ns,
+                const void *flush_data,
+                const struct trace_start_clears *payload)
 {
-   stage_start(pctx, ts_ns, CLEAR_RESTORE_STAGE_ID);
+   stage_start(pctx, ts_ns, CLEAR_STAGE_ID);
 }
 
 void
-fd_end_clear_restore(struct pipe_context *pctx, uint64_t ts_ns,
+fd_end_clears(struct pipe_context *pctx, uint64_t ts_ns,
+              const void *flush_data,
+              const struct trace_end_clears *payload)
+{
+   stage_end(pctx, ts_ns, CLEAR_STAGE_ID);
+}
+
+void
+fd_start_tile_loads(struct pipe_context *pctx, uint64_t ts_ns,
+                    const void *flush_data,
+                    const struct trace_start_tile_loads *payload)
+{
+   stage_start(pctx, ts_ns, TILE_LOAD_STAGE_ID);
+}
+
+void
+fd_end_tile_loads(struct pipe_context *pctx, uint64_t ts_ns,
+                  const void *flush_data,
+                  const struct trace_end_tile_loads *payload)
+{
+   stage_end(pctx, ts_ns, TILE_LOAD_STAGE_ID);
+}
+
+void
+fd_start_tile_stores(struct pipe_context *pctx, uint64_t ts_ns,
                      const void *flush_data,
-                     const struct trace_end_clear_restore *payload)
+                     const struct trace_start_tile_stores *payload)
 {
-   stage_end(pctx, ts_ns, CLEAR_RESTORE_STAGE_ID);
+   stage_start(pctx, ts_ns, TILE_STORE_STAGE_ID);
 }
 
 void
-fd_start_resolve(struct pipe_context *pctx, uint64_t ts_ns,
-                 const void *flush_data,
-                 const struct trace_start_resolve *payload)
+fd_end_tile_stores(struct pipe_context *pctx, uint64_t ts_ns,
+                   const void *flush_data,
+                   const struct trace_end_tile_stores *payload)
 {
-   stage_start(pctx, ts_ns, RESOLVE_STAGE_ID);
-}
-
-void
-fd_end_resolve(struct pipe_context *pctx, uint64_t ts_ns,
-               const void *flush_data,
-               const struct trace_end_resolve *payload)
-{
-   stage_end(pctx, ts_ns, RESOLVE_STAGE_ID);
+   stage_end(pctx, ts_ns, TILE_STORE_STAGE_ID);
 }
 
 void

src/gallium/drivers/freedreno/freedreno_perfetto.h

@@ -40,8 +40,9 @@ enum fd_stage_id {
    BYPASS_STAGE_ID,
    BLIT_STAGE_ID,
    COMPUTE_STAGE_ID,
-   CLEAR_RESTORE_STAGE_ID,
-   RESOLVE_STAGE_ID,
+   CLEAR_STAGE_ID,
+   TILE_LOAD_STAGE_ID,
+   TILE_STORE_STAGE_ID,
    STATE_RESTORE_STAGE_ID,
    VSC_OVERFLOW_STAGE_ID,
    PROLOGUE_STAGE_ID,
@@ -59,8 +60,9 @@ static const struct {
    [BYPASS_STAGE_ID]  = {"Render", "Rendering to system memory"},
    [BLIT_STAGE_ID]    = {"Blit", "Performing a Blit operation"},
    [COMPUTE_STAGE_ID] = {"Compute", "Compute job"},
-   [CLEAR_RESTORE_STAGE_ID] = {"Clear/Restore", "Clear (sysmem) or per-tile clear or restore (GMEM)"},
-   [RESOLVE_STAGE_ID] = {"Resolve", "Per tile resolve (GMEM to system memory"},
+   [CLEAR_STAGE_ID]   = {"Clear", "Clear (sysmem) or per-tile clear (GMEM)"},
+   [TILE_LOAD_STAGE_ID]  = {"Tile Load", "Per tile load (system memory to GMEM)"},
+   [TILE_STORE_STAGE_ID] = {"Tile Store", "Per tile store (GMEM to system memory)"},
    [STATE_RESTORE_STAGE_ID] = {"State Restore", "Setup at the beginning of new cmdstream buffer"},
    [VSC_OVERFLOW_STAGE_ID] = {"VSC Overflow Test", ""},
    [PROLOGUE_STAGE_ID] = {"Prologue", "Preemble cmdstream (executed once before first tile)"},

src/gallium/drivers/freedreno/freedreno_tracepoints.py

@@ -123,17 +123,20 @@ begin_end_tp('binning_ib')
 begin_end_tp('vsc_overflow_test')
 begin_end_tp('prologue')
 
-# For GMEM pass, where this could either be a clear or restore
-begin_end_tp('clear_restore',
-    args=[TracepointArg(type='uint16_t', var='fast_cleared', c_format='0x%x'),
-          TracepointArg(type='uint16_t', var='restore', c_format='0x%x')],
-    tp_print=['fast_cleared: 0x%x, restore=0x%x', '__entry->fast_cleared',
-              '__entry->restore'],
+# Either sysmem or gmem clears
+begin_end_tp('clears',
+    args=[TracepointArg(type='uint16_t', var='fast_cleared', c_format='0x%x')],
+    tp_print=['fast_cleared: 0x%x', '__entry->fast_cleared'],
 )
 
-begin_end_tp('resolve',
-    args=[TracepointArg(type='uint16_t', var='stored', c_format='0x%x')],
-    tp_print=['stored: 0x%x', '__entry->stored'],
+begin_end_tp('tile_loads',
+    args=[TracepointArg(type='uint16_t', var='load', c_format='0x%x')],
+    tp_print=['load=0x%x', '__entry->load'],
+)
+
+begin_end_tp('tile_stores',
+    args=[TracepointArg(type='uint16_t', var='store', c_format='0x%x')],
+    tp_print=['store: 0x%x', '__entry->store'],
 )
 
 singular_tp('start_tile',