freedreno: add support for hw queries
authorRob Clark <robclark@freedesktop.org>
Sat, 10 May 2014 17:45:54 +0000 (13:45 -0400)
committerRob Clark <robclark@freedesktop.org>
Tue, 13 May 2014 22:33:19 +0000 (18:33 -0400)
Real GPU queries need some infrastructure to track samples per tile and
accumulate the results.  But fortunately this can be shared across GPU
generation.

See:
https://github.com/freedreno/freedreno/wiki/Queries#hardware-queries

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
src/gallium/drivers/freedreno/freedreno_query_hw.c [new file with mode: 0644]
src/gallium/drivers/freedreno/freedreno_query_hw.h [new file with mode: 0644]
src/gallium/drivers/freedreno/freedreno_resource.c
src/gallium/drivers/freedreno/freedreno_util.h

index 3621abddb2843fc14c9788640bb825476e67b1ec..afb4c35ba75575f18af247b20c67b4507f57e2ab 100644 (file)
@@ -3,6 +3,7 @@ C_SOURCES := \
        freedreno_lowering.c \
        freedreno_program.c \
        freedreno_query.c \
+       freedreno_query_hw.c \
        freedreno_query_sw.c \
        freedreno_fence.c \
        freedreno_resource.c \
index a8fe3111c3db0b5868df56a2169b95662ac012cd..496a422709937d2584eb9c4786fbafee2019cd2f 100644 (file)
@@ -34,6 +34,7 @@
 #include "freedreno_state.h"
 #include "freedreno_gmem.h"
 #include "freedreno_query.h"
+#include "freedreno_query_hw.h"
 #include "freedreno_util.h"
 
 static struct fd_ringbuffer *next_rb(struct fd_context *ctx)
@@ -145,6 +146,7 @@ fd_context_destroy(struct pipe_context *pctx)
        DBG("");
 
        fd_prog_fini(pctx);
+       fd_hw_query_fini(pctx);
 
        util_slab_destroy(&ctx->transfer_pool);
 
@@ -221,6 +223,7 @@ fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
        fd_query_context_init(pctx);
        fd_texture_init(pctx);
        fd_state_init(pctx);
+       fd_hw_query_init(pctx);
 
        ctx->blitter = util_blitter_create(pctx);
        if (!ctx->blitter)
index 2e7fdedd9ae45a7129e1b2ee8122fbc48938170f..46984823427d578ea5b37a75ae8243cf15d57b26 100644 (file)
@@ -33,6 +33,7 @@
 #include "pipe/p_context.h"
 #include "indices/u_primconvert.h"
 #include "util/u_blitter.h"
+#include "util/u_double_list.h"
 #include "util/u_slab.h"
 #include "util/u_string.h"
 
@@ -82,16 +83,80 @@ struct fd_vertex_stateobj {
        unsigned num_elements;
 };
 
+/* Bitmask of stages in rendering that a particular query query is
+ * active.  Queries will be automatically started/stopped (generating
+ * additional fd_hw_sample_period's) on entrance/exit from stages that
+ * are applicable to the query.
+ *
+ * NOTE: set the stage to NULL at end of IB to ensure no query is still
+ * active.  Things aren't going to work out the way you want if a query
+ * is active across IB's (or between tile IB and draw IB)
+ */
+enum fd_render_stage {
+       FD_STAGE_NULL     = 0x00,
+       FD_STAGE_DRAW     = 0x01,
+       FD_STAGE_CLEAR    = 0x02,
+       /* TODO before queries which include MEM2GMEM or GMEM2MEM will
+        * work we will need to call fd_hw_query_prepare() from somewhere
+        * appropriate so that queries in the tiling IB get backed with
+        * memory to write results to.
+        */
+       FD_STAGE_MEM2GMEM = 0x04,
+       FD_STAGE_GMEM2MEM = 0x08,
+       /* used for driver internal draws (ie. util_blitter_blit()): */
+       FD_STAGE_BLIT     = 0x10,
+};
+
+#define MAX_HW_SAMPLE_PROVIDERS 4
+struct fd_hw_sample_provider;
+struct fd_hw_sample;
+
 struct fd_context {
        struct pipe_context base;
 
        struct fd_device *dev;
        struct fd_screen *screen;
+
        struct blitter_context *blitter;
        struct primconvert_context *primconvert;
 
+       /* slab for pipe_transfer allocations: */
        struct util_slab_mempool transfer_pool;
 
+       /* slabs for fd_hw_sample and fd_hw_sample_period allocations: */
+       struct util_slab_mempool sample_pool;
+       struct util_slab_mempool sample_period_pool;
+
+       /* next sample offset.. incremented for each sample in the batch/
+        * submit, reset to zero on next submit.
+        */
+       uint32_t next_sample_offset;
+
+       /* sample-providers for hw queries: */
+       const struct fd_hw_sample_provider *sample_providers[MAX_HW_SAMPLE_PROVIDERS];
+
+       /* cached samples (in case multiple queries need to reference
+        * the same sample snapshot)
+        */
+       struct fd_hw_sample *sample_cache[MAX_HW_SAMPLE_PROVIDERS];
+
+       /* tracking for current stage, to know when to start/stop
+        * any active queries:
+        */
+       enum fd_render_stage stage;
+
+       /* list of active queries: */
+       struct list_head active_queries;
+
+       /* list of queries that are not active, but were active in the
+        * current submit:
+        */
+       struct list_head current_queries;
+
+       /* current query result bo and tile stride: */
+       struct fd_bo *query_bo;
+       uint32_t query_tile_stride;
+
        /* table with PIPE_PRIM_MAX entries mapping PIPE_PRIM_x to
         * DI_PT_x value to use for draw initiator.  There are some
         * slight differences between generation:
@@ -258,10 +323,6 @@ struct fd_context {
        void (*draw)(struct fd_context *pctx, const struct pipe_draw_info *info);
        void (*clear)(struct fd_context *ctx, unsigned buffers,
                        const union pipe_color_union *color, double depth, unsigned stencil);
-
-       /* queries: */
-       struct fd_query * (*create_query)(struct fd_context *ctx,
-                       unsigned query_type);
 };
 
 static INLINE struct fd_context *
index 11bb8d8333ddec3b943f9c1f0fa762ae594612b0..1289bb4b8f2a71cadf0356330375a21c6bb4f85f 100644 (file)
@@ -36,6 +36,7 @@
 #include "freedreno_context.h"
 #include "freedreno_state.h"
 #include "freedreno_resource.h"
+#include "freedreno_query_hw.h"
 #include "freedreno_util.h"
 
 
@@ -156,6 +157,7 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
        /* and any buffers used, need to be resolved: */
        ctx->resolve |= buffers;
 
+       fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_DRAW);
        ctx->draw(ctx, info);
 }
 
@@ -188,6 +190,8 @@ fd_clear(struct pipe_context *pctx, unsigned buffers,
                util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
                util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 
+       fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_CLEAR);
+
        ctx->clear(ctx, buffers, color, depth, stencil);
 
        ctx->dirty |= FD_DIRTY_ZSA |
index 2d4de442452ef88ea61f90876907560dd954d5af..861ebf5675e96ff2a1a0a949abc514c150efd1cb 100644 (file)
@@ -35,6 +35,7 @@
 #include "freedreno_gmem.h"
 #include "freedreno_context.h"
 #include "freedreno_resource.h"
+#include "freedreno_query_hw.h"
 #include "freedreno_util.h"
 
 /*
@@ -273,17 +274,24 @@ render_tiles(struct fd_context *ctx)
 
                ctx->emit_tile_prep(ctx, tile);
 
-               if (ctx->restore)
+               if (ctx->restore) {
+                       fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_MEM2GMEM);
                        ctx->emit_tile_mem2gmem(ctx, tile);
+                       fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
+               }
 
                ctx->emit_tile_renderprep(ctx, tile);
 
+               fd_hw_query_prepare_tile(ctx, i, ctx->ring);
+
                /* emit IB to drawcmds: */
                OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
                fd_reset_wfi(ctx);
 
                /* emit gmem2mem to transfer tile back to system memory: */
+               fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_GMEM2MEM);
                ctx->emit_tile_gmem2mem(ctx, tile);
+               fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
        }
 }
 
@@ -292,6 +300,8 @@ render_sysmem(struct fd_context *ctx)
 {
        ctx->emit_sysmem_prep(ctx);
 
+       fd_hw_query_prepare_tile(ctx, 0, ctx->ring);
+
        /* emit IB to drawcmds: */
        OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
        fd_reset_wfi(ctx);
@@ -314,6 +324,11 @@ fd_gmem_render_tiles(struct pipe_context *pctx)
                }
        }
 
+       /* close out the draw cmds by making sure any active queries are
+        * paused:
+        */
+       fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
+
        /* mark the end of the clear/draw cmds before emitting per-tile cmds: */
        fd_ringmarker_mark(ctx->draw_end);
        fd_ringmarker_mark(ctx->binning_end);
@@ -326,6 +341,7 @@ fd_gmem_render_tiles(struct pipe_context *pctx)
                DBG("rendering sysmem (%s/%s)",
                        util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
                        util_format_short_name(pipe_surface_format(pfb->zsbuf)));
+               fd_hw_query_prepare(ctx, 1);
                render_sysmem(ctx);
                ctx->stats.batch_sysmem++;
        } else {
@@ -334,6 +350,7 @@ fd_gmem_render_tiles(struct pipe_context *pctx)
                DBG("rendering %dx%d tiles (%s/%s)", gmem->nbins_x, gmem->nbins_y,
                        util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
                        util_format_short_name(pipe_surface_format(pfb->zsbuf)));
+               fd_hw_query_prepare(ctx, gmem->nbins_x * gmem->nbins_y);
                render_tiles(ctx);
                ctx->stats.batch_gmem++;
        }
index fb9908b39eeb460ddb2c85ff5dd15d5cb8fdadb2..8753a4b02c98f3611bb018885e571a3ff019c829 100644 (file)
@@ -31,6 +31,7 @@
 
 #include "freedreno_query.h"
 #include "freedreno_query_sw.h"
+#include "freedreno_query_hw.h"
 #include "freedreno_context.h"
 #include "freedreno_util.h"
 
@@ -45,8 +46,8 @@ fd_create_query(struct pipe_context *pctx, unsigned query_type)
        struct fd_query *q;
 
        q = fd_sw_create_query(ctx, query_type);
-       if (ctx->create_query && !q)
-               q = ctx->create_query(ctx, query_type);
+       if (!q)
+               q = fd_hw_create_query(ctx, query_type);
 
        return (struct pipe_query *) q;
 }
diff --git a/src/gallium/drivers/freedreno/freedreno_query_hw.c b/src/gallium/drivers/freedreno/freedreno_query_hw.c
new file mode 100644 (file)
index 0000000..38bd3de
--- /dev/null
@@ -0,0 +1,465 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2014 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_memory.h"
+#include "util/u_inlines.h"
+
+#include "freedreno_query_hw.h"
+#include "freedreno_context.h"
+#include "freedreno_util.h"
+
+struct fd_hw_sample_period {
+       struct fd_hw_sample *start, *end;
+       struct list_head list;
+};
+
+/* maps query_type to sample provider idx: */
+static int pidx(unsigned query_type)
+{
+       switch (query_type) {
+       case PIPE_QUERY_OCCLUSION_COUNTER:
+               return 0;
+       case PIPE_QUERY_OCCLUSION_PREDICATE:
+               return 1;
+       default:
+               return -1;
+       }
+}
+
+static struct fd_hw_sample *
+get_sample(struct fd_context *ctx, struct fd_ringbuffer *ring,
+               unsigned query_type)
+{
+       struct fd_hw_sample *samp = NULL;
+       int idx = pidx(query_type);
+
+       if (!ctx->sample_cache[idx]) {
+               ctx->sample_cache[idx] =
+                       ctx->sample_providers[idx]->get_sample(ctx, ring);
+       }
+
+       fd_hw_sample_reference(ctx, &samp, ctx->sample_cache[idx]);
+
+       return samp;
+}
+
+static void
+clear_sample_cache(struct fd_context *ctx)
+{
+       int i;
+
+       for (i = 0; i < ARRAY_SIZE(ctx->sample_cache); i++)
+               fd_hw_sample_reference(ctx, &ctx->sample_cache[i], NULL);
+}
+
+static bool
+is_active(struct fd_hw_query *hq, enum fd_render_stage stage)
+{
+       return !!(hq->provider->active & stage);
+}
+
+
+static void
+resume_query(struct fd_context *ctx, struct fd_hw_query *hq,
+               struct fd_ringbuffer *ring)
+{
+       assert(!hq->period);
+       hq->period = util_slab_alloc(&ctx->sample_period_pool);
+       list_inithead(&hq->period->list);
+       hq->period->start = get_sample(ctx, ring, hq->base.type);
+       /* NOTE: util_slab_alloc() does not zero out the buffer: */
+       hq->period->end = NULL;
+}
+
+static void
+pause_query(struct fd_context *ctx, struct fd_hw_query *hq,
+               struct fd_ringbuffer *ring)
+{
+       assert(hq->period && !hq->period->end);
+       hq->period->end = get_sample(ctx, ring, hq->base.type);
+       list_addtail(&hq->period->list, &hq->current_periods);
+       hq->period = NULL;
+}
+
+static void
+destroy_periods(struct fd_context *ctx, struct list_head *list)
+{
+       struct fd_hw_sample_period *period, *s;
+       LIST_FOR_EACH_ENTRY_SAFE(period, s, list, list) {
+               fd_hw_sample_reference(ctx, &period->start, NULL);
+               fd_hw_sample_reference(ctx, &period->end, NULL);
+               list_del(&period->list);
+               util_slab_free(&ctx->sample_period_pool, period);
+       }
+}
+
+static void
+fd_hw_destroy_query(struct fd_context *ctx, struct fd_query *q)
+{
+       struct fd_hw_query *hq = fd_hw_query(q);
+
+       destroy_periods(ctx, &hq->periods);
+       destroy_periods(ctx, &hq->current_periods);
+       list_del(&hq->list);
+
+       free(hq);
+}
+
+static void
+fd_hw_begin_query(struct fd_context *ctx, struct fd_query *q)
+{
+       struct fd_hw_query *hq = fd_hw_query(q);
+       if (q->active)
+               return;
+
+       /* begin_query() should clear previous results: */
+       destroy_periods(ctx, &hq->periods);
+
+       if (is_active(hq, ctx->stage))
+               resume_query(ctx, hq, ctx->ring);
+
+       q->active = true;
+
+       /* add to active list: */
+       list_del(&hq->list);
+       list_addtail(&hq->list, &ctx->active_queries);
+}
+
+static void
+fd_hw_end_query(struct fd_context *ctx, struct fd_query *q)
+{
+       struct fd_hw_query *hq = fd_hw_query(q);
+       if (!q->active)
+               return;
+       if (is_active(hq, ctx->stage))
+               pause_query(ctx, hq, ctx->ring);
+       q->active = false;
+       /* move to current list: */
+       list_del(&hq->list);
+       list_addtail(&hq->list, &ctx->current_queries);
+}
+
+/* helper to get ptr to specified sample: */
+static void * sampptr(struct fd_hw_sample *samp, uint32_t n, void *ptr)
+{
+       return ((char *)ptr) + (samp->tile_stride * n) + samp->offset;
+}
+
+static boolean
+fd_hw_get_query_result(struct fd_context *ctx, struct fd_query *q,
+               boolean wait, union pipe_query_result *result)
+{
+       struct fd_hw_query *hq = fd_hw_query(q);
+       const struct fd_hw_sample_provider *p = hq->provider;
+       struct fd_hw_sample_period *period;
+
+       if (q->active)
+               return false;
+
+       /* if the app tries to read back the query result before the
+        * back is submitted, that forces us to flush so that there
+        * are actually results to wait for:
+        */
+       if (!LIST_IS_EMPTY(&hq->list)) {
+               DBG("reading query result forces flush!");
+               ctx->needs_flush = true;
+               fd_context_render(&ctx->base);
+       }
+
+       util_query_clear_result(result, q->type);
+
+       if (LIST_IS_EMPTY(&hq->periods))
+               return true;
+
+       assert(LIST_IS_EMPTY(&hq->list));
+       assert(LIST_IS_EMPTY(&hq->current_periods));
+       assert(!hq->period);
+
+       if (LIST_IS_EMPTY(&hq->periods))
+               return true;
+
+       /* if !wait, then check the last sample (the one most likely to
+        * not be ready yet) and bail if it is not ready:
+        */
+       if (!wait) {
+               int ret;
+
+               period = LIST_ENTRY(struct fd_hw_sample_period,
+                               hq->periods.prev, list);
+
+               ret = fd_bo_cpu_prep(period->end->bo, ctx->screen->pipe,
+                               DRM_FREEDRENO_PREP_READ | DRM_FREEDRENO_PREP_NOSYNC);
+               if (ret)
+                       return false;
+
+               fd_bo_cpu_fini(period->end->bo);
+       }
+
+       /* sum the result across all sample periods: */
+       LIST_FOR_EACH_ENTRY(period, &hq->periods, list) {
+               struct fd_hw_sample *start = period->start;
+               struct fd_hw_sample *end = period->end;
+               unsigned i;
+
+               /* start and end samples should be from same batch: */
+               assert(start->bo == end->bo);
+               assert(start->num_tiles == end->num_tiles);
+
+               for (i = 0; i < start->num_tiles; i++) {
+                       void *ptr;
+
+                       fd_bo_cpu_prep(start->bo, ctx->screen->pipe,
+                                       DRM_FREEDRENO_PREP_READ);
+
+                       ptr = fd_bo_map(start->bo);
+
+                       p->accumulate_result(ctx, sampptr(period->start, i, ptr),
+                                       sampptr(period->end, i, ptr), result);
+
+                       fd_bo_cpu_fini(start->bo);
+               }
+       }
+
+       return true;
+}
+
+static const struct fd_query_funcs hw_query_funcs = {
+               .destroy_query    = fd_hw_destroy_query,
+               .begin_query      = fd_hw_begin_query,
+               .end_query        = fd_hw_end_query,
+               .get_query_result = fd_hw_get_query_result,
+};
+
+struct fd_query *
+fd_hw_create_query(struct fd_context *ctx, unsigned query_type)
+{
+       struct fd_hw_query *hq;
+       struct fd_query *q;
+       int idx = pidx(query_type);
+
+       if ((idx < 0) || !ctx->sample_providers[idx])
+               return NULL;
+
+       hq = CALLOC_STRUCT(fd_hw_query);
+       if (!hq)
+               return NULL;
+
+       hq->provider = ctx->sample_providers[idx];
+
+       list_inithead(&hq->periods);
+       list_inithead(&hq->current_periods);
+       list_inithead(&hq->list);
+
+       q = &hq->base;
+       q->funcs = &hw_query_funcs;
+       q->type = query_type;
+
+       return q;
+}
+
+struct fd_hw_sample *
+fd_hw_sample_init(struct fd_context *ctx, uint32_t size)
+{
+       struct fd_hw_sample *samp = util_slab_alloc(&ctx->sample_pool);
+       pipe_reference_init(&samp->reference, 1);
+       samp->size = size;
+       samp->offset = ctx->next_sample_offset;
+       /* NOTE: util_slab_alloc() does not zero out the buffer: */
+       samp->bo = NULL;
+       samp->num_tiles = 0;
+       samp->tile_stride = 0;
+       ctx->next_sample_offset += size;
+       return samp;
+}
+
+void
+__fd_hw_sample_destroy(struct fd_context *ctx, struct fd_hw_sample *samp)
+{
+       if (samp->bo)
+               fd_bo_del(samp->bo);
+       util_slab_free(&ctx->sample_pool, samp);
+}
+
+static void
+prepare_sample(struct fd_hw_sample *samp, struct fd_bo *bo,
+               uint32_t num_tiles, uint32_t tile_stride)
+{
+       if (samp->bo) {
+               assert(samp->bo == bo);
+               assert(samp->num_tiles == num_tiles);
+               assert(samp->tile_stride == tile_stride);
+               return;
+       }
+       samp->bo = bo;
+       samp->num_tiles = num_tiles;
+       samp->tile_stride = tile_stride;
+}
+
+static void
+prepare_query(struct fd_hw_query *hq, struct fd_bo *bo,
+               uint32_t num_tiles, uint32_t tile_stride)
+{
+       struct fd_hw_sample_period *period, *s;
+
+       /* prepare all the samples in the query: */
+       LIST_FOR_EACH_ENTRY_SAFE(period, s, &hq->current_periods, list) {
+               prepare_sample(period->start, bo, num_tiles, tile_stride);
+               prepare_sample(period->end, bo, num_tiles, tile_stride);
+
+               /* move from current_periods list to periods list: */
+               list_del(&period->list);
+               list_addtail(&period->list, &hq->periods);
+       }
+}
+
+static void
+prepare_queries(struct fd_context *ctx, struct fd_bo *bo,
+               uint32_t num_tiles, uint32_t tile_stride,
+               struct list_head *list, bool remove)
+{
+       struct fd_hw_query *hq, *s;
+       LIST_FOR_EACH_ENTRY_SAFE(hq, s, list, list) {
+               prepare_query(hq, bo, num_tiles, tile_stride);
+               if (remove)
+                       list_delinit(&hq->list);
+       }
+}
+
+/* called from gmem code once total storage requirements are known (ie.
+ * number of samples times number of tiles)
+ */
+void
+fd_hw_query_prepare(struct fd_context *ctx, uint32_t num_tiles)
+{
+       uint32_t tile_stride = ctx->next_sample_offset;
+       struct fd_bo *bo;
+
+       if (ctx->query_bo)
+               fd_bo_del(ctx->query_bo);
+
+       if (tile_stride > 0) {
+               bo = fd_bo_new(ctx->dev, tile_stride * num_tiles,
+                               DRM_FREEDRENO_GEM_CACHE_WCOMBINE |
+                               DRM_FREEDRENO_GEM_TYPE_KMEM);
+       } else {
+               bo = NULL;
+       }
+
+       ctx->query_bo = bo;
+       ctx->query_tile_stride = tile_stride;
+
+       prepare_queries(ctx, bo, num_tiles, tile_stride,
+                       &ctx->active_queries, false);
+       prepare_queries(ctx, bo, num_tiles, tile_stride,
+                       &ctx->current_queries, true);
+
+       /* reset things for next batch: */
+       ctx->next_sample_offset = 0;
+}
+
+void
+fd_hw_query_prepare_tile(struct fd_context *ctx, uint32_t n,
+               struct fd_ringbuffer *ring)
+{
+       uint32_t tile_stride = ctx->query_tile_stride;
+       uint32_t offset = tile_stride * n;
+
+       /* bail if no queries: */
+       if (tile_stride == 0)
+               return;
+
+       fd_wfi(ctx, ring);
+       OUT_PKT0 (ring, HW_QUERY_BASE_REG, 1);
+       OUT_RELOCW(ring, ctx->query_bo, offset, 0, 0);
+}
+
+void
+fd_hw_query_set_stage(struct fd_context *ctx, struct fd_ringbuffer *ring,
+               enum fd_render_stage stage)
+{
+       /* special case: internal blits (like mipmap level generation)
+        * go through normal draw path (via util_blitter_blit()).. but
+        * we need to ignore the FD_STAGE_DRAW which will be set, so we
+        * don't enable queries which should be paused during internal
+        * blits:
+        */
+       if ((ctx->stage == FD_STAGE_BLIT) &&
+                       (stage != FD_STAGE_NULL))
+               return;
+
+       if (stage != ctx->stage) {
+               struct fd_hw_query *hq;
+               LIST_FOR_EACH_ENTRY(hq, &ctx->active_queries, list) {
+                       bool was_active = is_active(hq, ctx->stage);
+                       bool now_active = is_active(hq, stage);
+
+                       if (now_active && !was_active)
+                               resume_query(ctx, hq, ring);
+                       else if (was_active && !now_active)
+                               pause_query(ctx, hq, ring);
+               }
+       }
+       clear_sample_cache(ctx);
+       ctx->stage = stage;
+}
+
+void
+fd_hw_query_register_provider(struct pipe_context *pctx,
+               const struct fd_hw_sample_provider *provider)
+{
+       struct fd_context *ctx = fd_context(pctx);
+       int idx = pidx(provider->query_type);
+
+       assert((0 <= idx) && (idx < MAX_HW_SAMPLE_PROVIDERS));
+       assert(!ctx->sample_providers[idx]);
+
+       ctx->sample_providers[idx] = provider;
+}
+
+void
+fd_hw_query_init(struct pipe_context *pctx)
+{
+       struct fd_context *ctx = fd_context(pctx);
+
+       util_slab_create(&ctx->sample_pool, sizeof(struct fd_hw_sample),
+                       16, UTIL_SLAB_SINGLETHREADED);
+       util_slab_create(&ctx->sample_period_pool, sizeof(struct fd_hw_sample_period),
+                       16, UTIL_SLAB_SINGLETHREADED);
+       list_inithead(&ctx->active_queries);
+       list_inithead(&ctx->current_queries);
+}
+
+void
+fd_hw_query_fini(struct pipe_context *pctx)
+{
+       struct fd_context *ctx = fd_context(pctx);
+
+       util_slab_destroy(&ctx->sample_pool);
+       util_slab_destroy(&ctx->sample_period_pool);
+}
diff --git a/src/gallium/drivers/freedreno/freedreno_query_hw.h b/src/gallium/drivers/freedreno/freedreno_query_hw.h
new file mode 100644 (file)
index 0000000..62baa3a
--- /dev/null
@@ -0,0 +1,164 @@
+/* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
+
+/*
+ * Copyright (C) 2014 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_HW_H_
+#define FREEDRENO_QUERY_HW_H_
+
+#include "util/u_double_list.h"
+
+#include "freedreno_query.h"
+#include "freedreno_context.h"
+
+
+/*
+ * HW Queries:
+ *
+ * See: https://github.com/freedreno/freedreno/wiki/Queries#hardware-queries
+ *
+ * Hardware queries will be specific to gpu generation, but they need
+ * some common infrastructure for triggering start/stop samples at
+ * various points (for example, to exclude mem2gmem/gmem2mem or clear)
+ * as well as per tile tracking.
+ *
+ * NOTE: in at least some cases hw writes sample values to memory addr
+ * specified in some register.  So we don't really have the option to
+ * just sample the same counter multiple times for multiple different
+ * queries with the same query_type.  So we cache per sample provider
+ * the most recent sample since the last draw.  This way multiple
+ * sample periods for multiple queries can reference the same sample.
+ *
+ * fd_hw_sample_provider:
+ *   - one per query type, registered/implemented by gpu generation
+ *     specific code
+ *   - can construct fd_hw_samples on demand
+ *   - most recent sample (since last draw) cached so multiple
+ *     different queries can ref the same sample
+ *
+ * fd_hw_sample:
+ *   - abstracts one snapshot of counter value(s) across N tiles
+ *   - backing object not allocated until submit time when number
+ *     of samples and number of tiles is known
+ *
+ * fd_hw_sample_period:
+ *   - consists of start and stop sample
+ *   - a query accumulates a list of sample periods
+ *   - the query result is the sum of the sample periods
+ */
+
+struct fd_hw_sample_provider {
+       unsigned query_type;
+
+       /* stages applicable to the query type: */
+       enum fd_render_stage active;
+
+       /* when a new sample is required, emit appropriate cmdstream
+        * and return a sample object:
+        */
+       struct fd_hw_sample *(*get_sample)(struct fd_context *ctx,
+                       struct fd_ringbuffer *ring);
+
+       /* accumulate the results from specified sample period: */
+       void (*accumulate_result)(struct fd_context *ctx,
+                       const void *start, const void *end,
+                       union pipe_query_result *result);
+};
+
+struct fd_hw_sample {
+       struct pipe_reference reference;  /* keep this first */
+
+       /* offset and size of the sample are know at the time the
+        * sample is constructed.
+        */
+       uint32_t size;
+       uint32_t offset;
+
+       /* backing object, offset/stride/etc are determined not when
+        * the sample is constructed, but when the batch is submitted.
+        * This way we can defer allocation until total # of requested
+        * samples, and total # of tiles, is known.
+        */
+       struct fd_bo *bo;
+       uint32_t num_tiles;
+       uint32_t tile_stride;
+};
+
+struct fd_hw_sample_period;
+
+struct fd_hw_query {
+       struct fd_query base;
+
+       const struct fd_hw_sample_provider *provider;
+
+       /* list of fd_hw_sample_period in previous submits: */
+       struct list_head periods;
+
+       /* list of fd_hw_sample_period's in current submit: */
+       struct list_head current_periods;
+
+       /* if active and not paused, the current sample period (not
+        * yet added to current_periods):
+        */
+       struct fd_hw_sample_period *period;
+
+       struct list_head list;  /* list-node in ctx->active_queries */
+};
+
+static inline struct fd_hw_query *
+fd_hw_query(struct fd_query *q)
+{
+       return (struct fd_hw_query *)q;
+}
+
+struct fd_query * fd_hw_create_query(struct fd_context *ctx, unsigned query_type);
+/* helper for sample providers: */
+struct fd_hw_sample * fd_hw_sample_init(struct fd_context *ctx, uint32_t size);
+/* don't call directly, use fd_hw_sample_reference() */
+void __fd_hw_sample_destroy(struct fd_context *ctx, struct fd_hw_sample *samp);
+void fd_hw_query_prepare(struct fd_context *ctx, uint32_t num_tiles);
+void fd_hw_query_prepare_tile(struct fd_context *ctx, uint32_t n,
+               struct fd_ringbuffer *ring);
+void fd_hw_query_set_stage(struct fd_context *ctx,
+               struct fd_ringbuffer *ring, enum fd_render_stage stage);
+void fd_hw_query_register_provider(struct pipe_context *pctx,
+               const struct fd_hw_sample_provider *provider);
+void fd_hw_query_init(struct pipe_context *pctx);
+void fd_hw_query_fini(struct pipe_context *pctx);
+
+static inline void
+fd_hw_sample_reference(struct fd_context *ctx,
+               struct fd_hw_sample **ptr, struct fd_hw_sample *samp)
+{
+       struct fd_hw_sample *old_samp = *ptr;
+
+       if (pipe_reference(&(*ptr)->reference, &samp->reference))
+               __fd_hw_sample_destroy(ctx, old_samp);
+       if (ptr)
+               *ptr = samp;
+}
+
+#endif /* FREEDRENO_QUERY_HW_H_ */
index bd8c6cb9907ff0b60bfdd5b6ac22f99c6d10c160..289f3653e12eb10939ba402498411558db630de9 100644 (file)
@@ -36,6 +36,7 @@
 #include "freedreno_screen.h"
 #include "freedreno_surface.h"
 #include "freedreno_context.h"
+#include "freedreno_query_hw.h"
 #include "freedreno_util.h"
 
 #include <errno.h>
@@ -401,7 +402,9 @@ render_blit(struct pipe_context *pctx, struct pipe_blit_info *info)
        util_blitter_save_fragment_sampler_views(ctx->blitter,
                        ctx->fragtex.num_textures, ctx->fragtex.textures);
 
+       fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_BLIT);
        util_blitter_blit(ctx->blitter, info);
+       fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
 
        return true;
 }
index 356f4165792f432f1744f665b241cece482ea8fd..b57702c54c8682ff4c983d8510aeed73c6245b0d 100644 (file)
@@ -223,11 +223,18 @@ OUT_IB(struct fd_ringbuffer *ring, struct fd_ringmarker *start,
        emit_marker(ring, 6);
 }
 
+/* CP_SCRATCH_REG4 is used to hold base address for query results: */
+#define HW_QUERY_BASE_REG REG_AXXX_CP_SCRATCH_REG4
+
 static inline void
 emit_marker(struct fd_ringbuffer *ring, int scratch_idx)
 {
        extern unsigned marker_cnt;
-       OUT_PKT0(ring, REG_AXXX_CP_SCRATCH_REG0 + scratch_idx, 1);
+       unsigned reg = REG_AXXX_CP_SCRATCH_REG0 + scratch_idx;
+       assert(reg != HW_QUERY_BASE_REG);
+       if (reg == HW_QUERY_BASE_REG)
+               return;
+       OUT_PKT0(ring, reg, 1);
        OUT_RING(ring, ++marker_cnt);
 }