freedreno: add basic query support
authorRob Clark <robclark@freedesktop.org>
Wed, 8 Jan 2014 02:39:13 +0000 (21:39 -0500)
committerRob Clark <robclark@freedesktop.org>
Wed, 8 Jan 2014 21:30:18 +0000 (16:30 -0500)
Add for now some simple/basic query support (ie. things not actually
requiring the GPU).  Might change around a bit when I actually add
GPU queries, but for now this enables some useful performance info
in the GALLIUM_HUD.  For example:

  GALLIUM_HUD=fps+batches+batches-sysmem+batches-gmem+restores,draw-calls

The driver specific specific queries are:

  + draw-calls
  + batches - number of batches per second, sum of batches-sysmem
    plus batches-gmem
  + batches-gmem - render a set of tiles in GMEM, for each tile
    (optionally) system mem -> gmem (restore), plus N draws,
    plus gmem -> system mem (resolve) per second
  + batches-sysmem - N draws to system memory (GMEM bypass) per
    second
  + restores - number of GMEM batches that required restore per
    second

Ideally for GMEM rendering, you want batches-gmem to equal fps.  If
the app is doing something that triggers multiple passes (ie. requires
extra round trip gmem <-> system memory) then the # of batches per
second will go up relative to fps.

Signed-off-by: Rob Clark <robclark@freedesktop.org>
src/gallium/drivers/freedreno/Makefile.sources
src/gallium/drivers/freedreno/freedreno_context.c
src/gallium/drivers/freedreno/freedreno_context.h
src/gallium/drivers/freedreno/freedreno_draw.c
src/gallium/drivers/freedreno/freedreno_gmem.c
src/gallium/drivers/freedreno/freedreno_query.c [new file with mode: 0644]
src/gallium/drivers/freedreno/freedreno_query.h [new file with mode: 0644]
src/gallium/drivers/freedreno/freedreno_screen.c

index e54bff0369dc4021c6547e5d4180d8be25ca81e9..092b09f80423db33b003889384f35f942d3b21af 100644 (file)
@@ -1,5 +1,6 @@
 C_SOURCES := \
        freedreno_util.c \
+       freedreno_query.c \
        freedreno_fence.c \
        freedreno_resource.c \
        freedreno_surface.c \
index 23f6a67734d1d5949d6cf61c86df87566e0b9078..370c8f69da6a6dc123012002bdc226d19efe246b 100644 (file)
@@ -32,6 +32,7 @@
 #include "freedreno_texture.h"
 #include "freedreno_state.h"
 #include "freedreno_gmem.h"
+#include "freedreno_query.h"
 #include "freedreno_util.h"
 
 static struct fd_ringbuffer *next_rb(struct fd_context *ctx)
@@ -212,6 +213,7 @@ fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
 
        fd_draw_init(pctx);
        fd_resource_context_init(pctx);
+       fd_query_context_init(pctx);
        fd_texture_init(pctx);
        fd_state_init(pctx);
 
index 5373de6c6383053718bb09a1ce352dac02d7a653..0871efc53c759f03477481b791e72984cf81f8a7 100644 (file)
@@ -134,7 +134,15 @@ struct fd_context {
                FD_GMEM_BLEND_ENABLED        = 0x10,
                FD_GMEM_LOGICOP_ENABLED      = 0x20,
        } gmem_reason;
-       unsigned num_draws;
+       unsigned num_draws;   /* number of draws in current batch */
+
+       /* Stats/counters:
+        */
+       struct {
+               uint64_t prims_emitted;
+               uint64_t draw_calls;
+               uint64_t batch_total, batch_sysmem, batch_gmem, batch_restore;
+       } stats;
 
        /* we can't really sanely deal with wraparound point in ringbuffer
         * and because of the way tiling works we can't really flush at
index d80f35656143eeb5d0853633a23b249d8ec85838..4382bf0d44e1142395265b1e3f80f7b9cab01b28 100644 (file)
@@ -142,6 +142,10 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 
        ctx->num_draws++;
 
+       ctx->stats.draw_calls++;
+       ctx->stats.prims_emitted +=
+               u_reduced_prims_for_vertices(info->mode, info->count);
+
        /* any buffers that haven't been cleared, we need to restore: */
        ctx->restore |= buffers & (FD_BUFFER_ALL & ~ctx->cleared);
        /* and any buffers used, need to be resolved: */
index 0270538a3d0e080f321ffb591f4642c3c499c4d3..6a55aa4c133baaf4821e79b43f650fe2f4afe91b 100644 (file)
@@ -260,6 +260,9 @@ render_tiles(struct fd_context *ctx)
 
        ctx->emit_tile_init(ctx);
 
+       if (ctx->restore)
+               ctx->stats.batch_restore++;
+
        for (i = 0; i < (gmem->nbins_x * gmem->nbins_y); i++) {
                struct fd_tile *tile = &ctx->tile[i];
 
@@ -311,11 +314,14 @@ fd_gmem_render_tiles(struct pipe_context *pctx)
        fd_ringmarker_mark(ctx->draw_end);
        fd_ringmarker_mark(ctx->binning_end);
 
+       ctx->stats.batch_total++;
+
        if (sysmem) {
                DBG("rendering sysmem (%s/%s)",
                        util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
                        util_format_short_name(pipe_surface_format(pfb->zsbuf)));
                render_sysmem(ctx);
+               ctx->stats.batch_sysmem++;
        } else {
                struct fd_gmem_stateobj *gmem = &ctx->gmem;
                calculate_tiles(ctx);
@@ -323,6 +329,7 @@ fd_gmem_render_tiles(struct pipe_context *pctx)
                        util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
                        util_format_short_name(pipe_surface_format(pfb->zsbuf)));
                render_tiles(ctx);
+               ctx->stats.batch_gmem++;
        }
 
        /* GPU executes starting from tile cmds, which IB back to draw cmds: */
diff --git a/src/gallium/drivers/freedreno/freedreno_query.c b/src/gallium/drivers/freedreno/freedreno_query.c
new file mode 100644 (file)
index 0000000..35bd5d7
--- /dev/null
@@ -0,0 +1,213 @@
+/* -*- mode: C; c-file-style: "k&r"; ttxab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * 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:
+ *    Rob Clark <robclark@freedesktop.org>
+ */
+
+#include "pipe/p_state.h"
+#include "util/u_string.h"
+#include "util/u_memory.h"
+#include "util/u_inlines.h"
+#include "os/os_time.h"
+
+#include "freedreno_query.h"
+#include "freedreno_context.h"
+#include "freedreno_util.h"
+
+#define FD_QUERY_DRAW_CALLS      (PIPE_QUERY_DRIVER_SPECIFIC + 0)
+#define FD_QUERY_BATCH_TOTAL     (PIPE_QUERY_DRIVER_SPECIFIC + 1)  /* total # of batches (submits) */
+#define FD_QUERY_BATCH_SYSMEM    (PIPE_QUERY_DRIVER_SPECIFIC + 2)  /* batches using system memory (GMEM bypass) */
+#define FD_QUERY_BATCH_GMEM      (PIPE_QUERY_DRIVER_SPECIFIC + 3)  /* batches using GMEM */
+#define FD_QUERY_BATCH_RESTORE   (PIPE_QUERY_DRIVER_SPECIFIC + 4)  /* batches requiring GMEM restore */
+
+/* Currently just simple cpu query's supported.. probably need
+ * to refactor this a bit when I'm eventually ready to add gpu
+ * queries:
+ */
+struct fd_query {
+       int type;
+       /* storage for the collected data */
+       union pipe_query_result data;
+       bool active;
+       uint64_t begin_value, end_value;
+       uint64_t begin_time, end_time;
+};
+
+static inline struct fd_query *
+fd_query(struct pipe_query *pq)
+{
+       return (struct fd_query *)pq;
+}
+
+static struct pipe_query *
+fd_create_query(struct pipe_context *pctx, unsigned query_type)
+{
+       struct fd_query *q;
+
+       switch (query_type) {
+       case PIPE_QUERY_PRIMITIVES_GENERATED:
+       case PIPE_QUERY_PRIMITIVES_EMITTED:
+       case FD_QUERY_DRAW_CALLS:
+       case FD_QUERY_BATCH_TOTAL:
+       case FD_QUERY_BATCH_SYSMEM:
+       case FD_QUERY_BATCH_GMEM:
+       case FD_QUERY_BATCH_RESTORE:
+               break;
+       default:
+               return NULL;
+       }
+
+       q = CALLOC_STRUCT(fd_query);
+       if (!q)
+               return NULL;
+
+       q->type = query_type;
+
+       return (struct pipe_query *) q;
+}
+
+static void
+fd_destroy_query(struct pipe_context *pctx, struct pipe_query *pq)
+{
+       struct fd_query *q = fd_query(pq);
+       free(q);
+}
+
+static uint64_t
+read_counter(struct pipe_context *pctx, int type)
+{
+       struct fd_context *ctx = fd_context(pctx);
+       switch (type) {
+       case PIPE_QUERY_PRIMITIVES_GENERATED:
+               /* for now same thing as _PRIMITIVES_EMITTED */
+       case PIPE_QUERY_PRIMITIVES_EMITTED:
+               return ctx->stats.prims_emitted;
+       case FD_QUERY_DRAW_CALLS:
+               return ctx->stats.draw_calls;
+       case FD_QUERY_BATCH_TOTAL:
+               return ctx->stats.batch_total;
+       case FD_QUERY_BATCH_SYSMEM:
+               return ctx->stats.batch_sysmem;
+       case FD_QUERY_BATCH_GMEM:
+               return ctx->stats.batch_gmem;
+       case FD_QUERY_BATCH_RESTORE:
+               return ctx->stats.batch_restore;
+       }
+       return 0;
+}
+
+static bool
+is_rate_query(struct fd_query *q)
+{
+       switch (q->type) {
+       case FD_QUERY_BATCH_TOTAL:
+       case FD_QUERY_BATCH_SYSMEM:
+       case FD_QUERY_BATCH_GMEM:
+       case FD_QUERY_BATCH_RESTORE:
+               return true;
+       default:
+               return false;
+       }
+}
+
+static void
+fd_begin_query(struct pipe_context *pctx, struct pipe_query *pq)
+{
+       struct fd_query *q = fd_query(pq);
+       q->active = true;
+       q->begin_value = read_counter(pctx, q->type);
+       if (is_rate_query(q))
+               q->begin_time = os_time_get();
+}
+
+static void
+fd_end_query(struct pipe_context *pctx, struct pipe_query *pq)
+{
+       struct fd_query *q = fd_query(pq);
+       q->active = false;
+       q->end_value = read_counter(pctx, q->type);
+       if (is_rate_query(q))
+               q->end_time = os_time_get();
+}
+
+static boolean
+fd_get_query_result(struct pipe_context *pctx, struct pipe_query *pq,
+               boolean wait, union pipe_query_result *result)
+{
+       struct fd_query *q = fd_query(pq);
+
+       if (q->active)
+               return false;
+
+       util_query_clear_result(result, q->type);
+
+       result->u64 = q->end_value - q->begin_value;
+
+       if (is_rate_query(q)) {
+               double fps = (result->u64 * 1000000) /
+                               (double)(q->end_time - q->begin_time);
+               result->u64 = (uint64_t)fps;
+       }
+
+       return true;
+}
+
+static int
+fd_get_driver_query_info(struct pipe_screen *pscreen,
+               unsigned index, struct pipe_driver_query_info *info)
+{
+       struct pipe_driver_query_info list[] = {
+                       {"draw-calls", FD_QUERY_DRAW_CALLS, 0},
+                       {"batches", FD_QUERY_BATCH_TOTAL, 0},
+                       {"batches-sysmem", FD_QUERY_BATCH_SYSMEM, 0},
+                       {"batches-gmem", FD_QUERY_BATCH_GMEM, 0},
+                       {"restores", FD_QUERY_BATCH_RESTORE, 0},
+       };
+
+       if (!info)
+               return ARRAY_SIZE(list);
+
+       if (index >= ARRAY_SIZE(list))
+               return 0;
+
+       *info = list[index];
+       return 1;
+}
+
+void
+fd_query_screen_init(struct pipe_screen *pscreen)
+{
+       pscreen->get_driver_query_info = fd_get_driver_query_info;
+}
+
+void
+fd_query_context_init(struct pipe_context *pctx)
+{
+       pctx->create_query = fd_create_query;
+       pctx->destroy_query = fd_destroy_query;
+       pctx->begin_query = fd_begin_query;
+       pctx->end_query = fd_end_query;
+       pctx->get_query_result = fd_get_query_result;
+}
diff --git a/src/gallium/drivers/freedreno/freedreno_query.h b/src/gallium/drivers/freedreno/freedreno_query.h
new file mode 100644 (file)
index 0000000..8bcbba2
--- /dev/null
@@ -0,0 +1,37 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * 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:
+ *    Rob Clark <robclark@freedesktop.org>
+ */
+
+#ifndef FREEDRENO_QUERY_H_
+#define FREEDRENO_QUERY_H_
+
+#include "pipe/p_context.h"
+
+void fd_query_screen_init(struct pipe_screen *pscreen);
+void fd_query_context_init(struct pipe_context *pctx);
+
+#endif /* FREEDRENO_QUERY_H_ */
index 28a09166acdaca2ce087e30fb498d30dada40d1f..10c5e186228a7dcfd4d3c2bcb367c192255de214 100644 (file)
@@ -47,6 +47,7 @@
 #include "freedreno_screen.h"
 #include "freedreno_resource.h"
 #include "freedreno_fence.h"
+#include "freedreno_query.h"
 #include "freedreno_util.h"
 
 #include "fd2_screen.h"
@@ -457,6 +458,7 @@ fd_screen_create(struct fd_device *dev)
        pscreen->get_shader_param = fd_screen_get_shader_param;
 
        fd_resource_screen_init(pscreen);
+       fd_query_screen_init(pscreen);
 
        pscreen->get_name = fd_screen_get_name;
        pscreen->get_vendor = fd_screen_get_vendor;