cell: implement fencing for texture buffers

If we delete a texture, we need to keep the underlying tiled data buffer
around until any rendering that references it has completed.
Keep a list of buffers referenced by a rendering batch.  Unref/free them when
the associated batch's fence is executed/signalled.
This commit is contained in:
Brian Paul 2008-10-22 10:34:13 -06:00
parent 0ae4728eb4
commit 70dd4379d2
13 changed files with 367 additions and 19 deletions

View file

@ -102,6 +102,8 @@
#define CELL_CMD_STATE_RASTERIZER 22
#define CELL_CMD_VS_EXECUTE 23
#define CELL_CMD_FLUSH_BUFFER_RANGE 24
#define CELL_CMD_FENCE 25
/** Command/batch buffers */
#define CELL_NUM_BUFFERS 4
@ -123,6 +125,29 @@
#define SPU_MAX_FRAGMENT_OPS_INSTS 64
#define CELL_FENCE_IDLE 0
#define CELL_FENCE_EMITTED 1
#define CELL_FENCE_SIGNALLED 2
struct cell_fence
{
/** There's a 16-byte status qword per SPU */
volatile uint status[CELL_MAX_SPUS][4];
};
/**
* Fence command sent to SPUs. In response, the SPUs will write
* CELL_FENCE_STATUS_SIGNALLED back to the fence status word in main memory.
*/
struct cell_command_fence
{
uint64_t opcode; /**< CELL_CMD_FENCE */
struct cell_fence *fence;
};
/**
* Command to specify per-fragment operations state and generated code.
* Note that the dsa, blend, blend_color fields are really only needed

View file

@ -24,6 +24,7 @@ SOURCES = \
cell_clear.c \
cell_context.c \
cell_draw_arrays.c \
cell_fence.c \
cell_flush.c \
cell_gen_fragment.c \
cell_gen_fp.c \

View file

@ -28,6 +28,7 @@
#include "cell_context.h"
#include "cell_batch.h"
#include "cell_fence.h"
#include "cell_spu.h"
@ -63,6 +64,10 @@ cell_get_empty_buffer(struct cell_context *cell)
printf("PPU: ALLOC BUFFER %u, %u tries\n", buf, tries);
*/
prev_buffer = buf;
/* release tex buffer associated w/ prev use of this batch buf */
cell_free_fenced_buffers(cell, &cell->fenced_buffers[buf]);
return buf;
}
}
@ -84,6 +89,26 @@ cell_get_empty_buffer(struct cell_context *cell)
}
/**
* Append a fence command to the current batch buffer.
* Note that we're sure there's always room for this because of the
* adjusted size check in cell_batch_free_space().
*/
static void
emit_fence(struct cell_context *cell)
{
const uint batch = cell->cur_batch;
const uint size = cell->buffer_size[batch];
struct cell_command_fence *fence_cmd;
ASSERT(size + sizeof(struct cell_command_fence) <= CELL_BUFFER_SIZE);
fence_cmd = (struct cell_command_fence *) (cell->buffer[batch] + size);
fence_cmd->opcode = CELL_CMD_FENCE;
fence_cmd->fence = &cell->fenced_buffers[batch].fence;
}
/**
* Flush the current batch buffer to the SPUs.
* An empty buffer will be found and set as the new current batch buffer
@ -102,6 +127,12 @@ cell_batch_flush(struct cell_context *cell)
if (size == 0)
return;
/* Before we use this batch buffer, make sure any fenced texture buffers
* are released.
*/
if (cell->fenced_buffers[batch].head)
emit_fence(cell);
flushing = TRUE;
assert(batch < CELL_NUM_BUFFERS);
@ -142,6 +173,7 @@ uint
cell_batch_free_space(const struct cell_context *cell)
{
uint free = CELL_BUFFER_SIZE - cell->buffer_size[cell->cur_batch];
free -= sizeof(struct cell_command_fence);
return free;
}

View file

@ -47,6 +47,7 @@
#include "cell_clear.h"
#include "cell_context.h"
#include "cell_draw_arrays.h"
#include "cell_fence.h"
#include "cell_flush.h"
#include "cell_state.h"
#include "cell_surface.h"
@ -104,6 +105,7 @@ cell_create_context(struct pipe_screen *screen,
struct cell_winsys *cws)
{
struct cell_context *cell;
uint i;
/* some fields need to be 16-byte aligned, so align the whole object */
cell = (struct cell_context*) align_malloc(sizeof(struct cell_context), 16);
@ -151,6 +153,10 @@ cell_create_context(struct pipe_screen *screen,
cell_debug_flags,
0 );
for (i = 0; i < CELL_NUM_BUFFERS; i++)
cell_fence_init(&cell->fenced_buffers[i].fence);
/*
* SPU stuff
*/

View file

@ -81,6 +81,19 @@ struct cell_fragment_ops_key
};
struct cell_buffer_node;
/**
* Fenced buffer list. List of buffers which can be unreferenced after
* the fence has been executed/signalled.
*/
struct cell_buffer_list
{
struct cell_fence fence;
struct cell_buffer_node *head;
};
/**
* Per-context state, subclass of pipe_context.
*/
@ -154,6 +167,14 @@ struct cell_context
uint buffer_status[CELL_MAX_SPUS][CELL_NUM_BUFFERS][4] ALIGN16_ATTRIB;
/** Associated with each command/batch buffer is a list of pipe_buffers
* that are fenced. When the last command in a buffer is executed, the
* fence will be signalled, indicating that any pipe_buffers preceeding
* that fence can be unreferenced (and probably freed).
*/
struct cell_buffer_list fenced_buffers[CELL_NUM_BUFFERS];
struct spe_function attrib_fetch;
unsigned attrib_fetch_offsets[PIPE_MAX_ATTRIBS];

View file

@ -0,0 +1,158 @@
/**************************************************************************
*
* Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas.
* All Rights Reserved.
*
* 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, sub license, 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 NON-INFRINGEMENT.
* IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS 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.
*
**************************************************************************/
#include <unistd.h>
#include "util/u_memory.h"
#include "pipe/p_inlines.h"
#include "cell_context.h"
#include "cell_batch.h"
#include "cell_fence.h"
#include "cell_texture.h"
void
cell_fence_init(struct cell_fence *fence)
{
uint i;
for (i = 0; i < CELL_MAX_SPUS; i++) {
fence->status[i][0] = CELL_FENCE_IDLE;
}
}
boolean
cell_fence_signalled(const struct cell_context *cell,
const struct cell_fence *fence)
{
uint i;
for (i = 0; i < cell->num_spus; i++) {
//ASSERT(fence->status[i][0] != CELL_FENCE_IDLE);
if (fence->status[i][0] == CELL_FENCE_EMITTED)
return FALSE;
}
return TRUE;
}
void
cell_fence_finish(const struct cell_context *cell,
const struct cell_fence *fence)
{
while (!cell_fence_signalled(cell, fence)) {
usleep(10);
}
}
struct cell_buffer_node
{
struct pipe_buffer *buffer;
struct cell_buffer_node *next;
};
static void
cell_add_buffer_to_list(struct cell_context *cell,
struct cell_buffer_list *list,
struct pipe_buffer *buffer)
{
struct pipe_screen *ps = cell->pipe.screen;
struct cell_buffer_node *node = CALLOC_STRUCT(cell_buffer_node);
/* create new list node which references the buffer, insert at head */
if (node) {
pipe_buffer_reference(ps, &node->buffer, buffer);
node->next = list->head;
list->head = node;
}
}
/**
* Wait for completion of the given fence, then unreference any buffers
* on the list.
* This typically unrefs/frees texture buffers after any rendering which uses
* them has completed.
*/
void
cell_free_fenced_buffers(struct cell_context *cell,
struct cell_buffer_list *list)
{
if (list->head) {
struct pipe_screen *ps = cell->pipe.screen;
struct cell_buffer_node *node;
cell_fence_finish(cell, &list->fence);
/* traverse the list, unreferencing buffers, freeing nodes */
node = list->head;
while (node) {
struct cell_buffer_node *next = node->next;
assert(node->buffer);
pipe_buffer_unmap(ps, node->buffer);
#if 0
printf("Unref buffer %p\n", node->buffer);
if (node->buffer->refcount == 1)
printf(" Delete!\n");
#endif
pipe_buffer_reference(ps, &node->buffer, NULL);
FREE(node);
node = next;
}
list->head = NULL;
}
}
/**
* This should be called for each render command.
* Any texture buffers that are current bound will be added to a fenced
* list to be freed later when the fence is executed/signalled.
*/
void
cell_add_fenced_textures(struct cell_context *cell)
{
struct cell_buffer_list *list = &cell->fenced_buffers[cell->cur_batch];
uint i;
for (i = 0; i < cell->num_textures; i++) {
struct cell_texture *ct = cell->texture[i];
if (ct) {
uint level;
for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) {
if (ct->tiled_buffer[level]) {
#if 0
printf("Adding texture %p buffer %p to list\n",
ct, ct->tiled_buffer[level]);
#endif
cell_add_buffer_to_list(cell, list, ct->tiled_buffer[level]);
}
}
}
}
}

View file

@ -0,0 +1,57 @@
/**************************************************************************
*
* Copyright 2008 Tungsten Graphics, Inc., Cedar Park, Texas.
* All Rights Reserved.
*
* 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, sub license, 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 NON-INFRINGEMENT.
* IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS 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.
*
**************************************************************************/
#ifndef CELL_FENCE_H
#define CELL_FENCE_H
extern void
cell_fence_init(struct cell_fence *fence);
extern boolean
cell_fence_signalled(const struct cell_context *cell,
const struct cell_fence *fence);
extern void
cell_fence_finish(const struct cell_context *cell,
const struct cell_fence *fence);
extern void
cell_free_fenced_buffers(struct cell_context *cell,
struct cell_buffer_list *list);
extern void
cell_add_fenced_textures(struct cell_context *cell);
#endif /* CELL_FENCE_H */

View file

@ -225,7 +225,7 @@ cell_emit_state(struct cell_context *cell)
if (cell->texture[i]) {
uint level;
for (level = 0; level < CELL_MAX_TEXTURE_LEVELS; level++) {
texture->start[level] = cell->texture[i]->tiled_data[level];
texture->start[level] = cell->texture[i]->tiled_mapped[level];
texture->width[level] = cell->texture[i]->base.width[level];
texture->height[level] = cell->texture[i]->base.height[level];
texture->depth[level] = cell->texture[i]->base.depth[level];

View file

@ -136,6 +136,9 @@ cell_texture_release(struct pipe_screen *screen,
__FUNCTION__, (void *) *pt, (*pt)->refcount - 1);
*/
if (--(*pt)->refcount <= 0) {
/* Delete this texture now.
* But note that the underlying pipe_buffer may linger...
*/
struct cell_texture *ct = cell_texture(*pt);
uint i;
@ -146,14 +149,12 @@ cell_texture_release(struct pipe_screen *screen,
pipe_buffer_reference(screen, &ct->buffer, NULL);
for (i = 0; i < CELL_MAX_TEXTURE_LEVELS; i++) {
if (ct->tiled_data[i]) {
/* XXX need to use a fenced buffer for tiled data so that
* it's properly freed after rendering has completed.
* Disabling this free() allows glDrawPixels to work for now.
*/
#if 0
align_free(ct->tiled_data[i]);
#endif
/* Unreference the tiled image buffer.
* It may not actually be deleted until a fence is hit.
*/
if (ct->tiled_buffer[i]) {
ct->tiled_mapped[i] = NULL;
winsys_buffer_reference(screen->winsys, &ct->tiled_buffer[i], NULL);
}
}
@ -234,12 +235,18 @@ cell_twiddle_texture(struct pipe_screen *screen,
int offset = bufWidth * bufHeight * 4 * surface->face;
uint *dst;
if (!ct->tiled_data[level]) {
ct->tiled_data[level] =
align_malloc(bufWidth * bufHeight * 4 * numFaces, 16);
if (!ct->tiled_buffer[level]) {
/* allocate buffer for tiled data now */
struct pipe_winsys *ws = screen->winsys;
uint bytes = bufWidth * bufHeight * 4 * numFaces;
ct->tiled_buffer[level] = ws->buffer_create(ws, 16,
PIPE_BUFFER_USAGE_PIXEL,
bytes);
/* and map it */
ct->tiled_mapped[level] = ws->buffer_map(ws, ct->tiled_buffer[level],
PIPE_BUFFER_USAGE_GPU_READ);
}
dst = (uint *) ((ubyte *) ct->tiled_data[level] + offset);
dst = (uint *) ((ubyte *) ct->tiled_mapped[level] + offset);
twiddle_image_uint(texWidth, texHeight, TILE_SIZE, dst,
surface->stride, src);

View file

@ -48,7 +48,10 @@ struct cell_texture
struct pipe_buffer *buffer;
unsigned long buffer_size;
void *tiled_data[CELL_MAX_TEXTURE_LEVELS]; /* XXX this may be temporary */ /*ALIGN16*/
/** Texture data in tiled layout is held here */
struct pipe_buffer *tiled_buffer[CELL_MAX_TEXTURE_LEVELS];
/** Mapped, tiled texture data */
void *tiled_mapped[CELL_MAX_TEXTURE_LEVELS];
};

View file

@ -38,6 +38,7 @@
#include "cell_batch.h"
#include "cell_context.h"
#include "cell_fence.h"
#include "cell_flush.h"
#include "cell_spu.h"
#include "cell_vbuf.h"
@ -108,6 +109,11 @@ cell_vbuf_release_vertices(struct vbuf_render *vbr, void *vertices,
__FUNCTION__, cvbr->vertex_buf, vertices_used);
*/
/* Make sure texture buffers aren't released until we're done rendering
* with them.
*/
cell_add_fenced_textures(cell);
/* Tell SPUs they can release the vert buf */
if (cvbr->vertex_buf != ~0U) {
struct cell_command_release_verts *release

View file

@ -76,9 +76,10 @@ static void
release_buffer(uint buffer)
{
/* Evidently, using less than a 16-byte status doesn't work reliably */
static const uint status[4] ALIGN16_ATTRIB
= {CELL_BUFFER_STATUS_FREE, 0, 0, 0};
static const vector unsigned int status = {CELL_BUFFER_STATUS_FREE,
CELL_BUFFER_STATUS_FREE,
CELL_BUFFER_STATUS_FREE,
CELL_BUFFER_STATUS_FREE};
const uint index = 4 * (spu.init.id * CELL_NUM_BUFFERS + buffer);
uint *dst = spu.init.buffer_status + index;
@ -93,6 +94,29 @@ release_buffer(uint buffer)
}
/**
* Write CELL_FENCE_SIGNALLED back to the fence status qword in main memory.
* There's a qword of status per SPU.
*/
static void
cmd_fence(struct cell_command_fence *fence_cmd)
{
static const vector unsigned int status = {CELL_FENCE_SIGNALLED,
CELL_FENCE_SIGNALLED,
CELL_FENCE_SIGNALLED,
CELL_FENCE_SIGNALLED};
uint *dst = (uint *) fence_cmd->fence;
dst += 4 * spu.init.id; /* main store/memory address, not local store */
mfc_put((void *) &status, /* src in local memory */
(unsigned int) dst, /* dst in main memory */
sizeof(status), /* size */
TAG_FENCE, /* tag */
0, /* tid */
0 /* rid */);
}
static void
cmd_clear_surface(const struct cell_command_clear_surface *clear)
{
@ -637,6 +661,14 @@ cmd_batch(uint opcode)
cmd_finish();
pos += 1;
break;
case CELL_CMD_FENCE:
{
struct cell_command_fence *fence_cmd =
(struct cell_command_fence *) &buffer[pos];
cmd_fence(fence_cmd);
pos += sizeof(*fence_cmd) / 8;
}
break;
case CELL_CMD_RELEASE_VERTS:
{
struct cell_command_release_verts *release

View file

@ -210,7 +210,7 @@ extern struct spu_global spu;
#define TAG_DCACHE1 21
#define TAG_DCACHE2 22
#define TAG_DCACHE3 23
#define TAG_FENCE 24
static INLINE void