static struct pipe_query *svga_create_query( struct pipe_context *pipe,
unsigned query_type )
{
struct svga_screen *svgascreen = svga_screen(pipe->screen);
struct svga_winsys_screen *sws = svgascreen->sws;
struct svga_query *sq;
SVGA_DBG(DEBUG_QUERY, "%s\n", __FUNCTION__);
sq = CALLOC_STRUCT(svga_query);
if (!sq)
goto no_sq;
sq->type = SVGA3D_QUERYTYPE_OCCLUSION;
sq->hwbuf = svga_winsys_buffer_create(svgascreen,
1,
SVGA_BUFFER_USAGE_PINNED,
sizeof *sq->queryResult);
if(!sq->hwbuf)
goto no_hwbuf;
sq->queryResult = (SVGA3dQueryResult *)sws->buffer_map(sws,
sq->hwbuf,
PIPE_BUFFER_USAGE_CPU_WRITE);
if(!sq->queryResult)
goto no_query_result;
sq->queryResult->totalSize = sizeof *sq->queryResult;
sq->queryResult->state = SVGA3D_QUERYSTATE_NEW;
/*
* We request the buffer to be pinned and assume it is always mapped.
*
* The reason is that we don't want to wait for fences when checking the
* query status.
*/
sws->buffer_unmap(sws, sq->hwbuf);
return &sq->base;
no_query_result:
sws->buffer_destroy(sws, sq->hwbuf);
no_hwbuf:
FREE(sq);
no_sq:
return NULL;
}
/**
* Allocate DMA'ble storage for the buffer.
*
* Called before mapping a buffer.
*/
static INLINE enum pipe_error
svga_buffer_create_hw_storage(struct svga_screen *ss,
struct svga_buffer *sbuf)
{
assert(!sbuf->user);
if(!sbuf->hwbuf) {
unsigned alignment = sbuf->base.alignment;
unsigned usage = 0;
unsigned size = sbuf->base.size;
sbuf->hwbuf = svga_winsys_buffer_create(ss, alignment, usage, size);
if(!sbuf->hwbuf)
return PIPE_ERROR_OUT_OF_MEMORY;
assert(!sbuf->dma.pending);
}
return PIPE_OK;
}
static void *
svga_texture_transfer_map(struct pipe_context *pipe,
struct pipe_resource *texture,
unsigned level,
unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **ptransfer)
{
struct svga_context *svga = svga_context(pipe);
struct svga_screen *ss = svga_screen(pipe->screen);
struct svga_winsys_screen *sws = ss->sws;
struct svga_texture *tex = svga_texture(texture);
struct svga_transfer *st;
unsigned nblocksx, nblocksy;
boolean use_direct_map = svga_have_gb_objects(svga) &&
!svga_have_gb_dma(svga);
unsigned d;
void *returnVal;
int64_t begin = os_time_get();
/* We can't map texture storage directly unless we have GB objects */
if (usage & PIPE_TRANSFER_MAP_DIRECTLY) {
if (svga_have_gb_objects(svga))
use_direct_map = TRUE;
else
return NULL;
}
st = CALLOC_STRUCT(svga_transfer);
if (!st)
return NULL;
{
unsigned w, h;
if (use_direct_map) {
/* we'll directly access the guest-backed surface */
w = u_minify(texture->width0, level);
h = u_minify(texture->height0, level);
d = u_minify(texture->depth0, level);
}
else {
/* we'll put the data into a tightly packed buffer */
w = box->width;
h = box->height;
d = box->depth;
}
nblocksx = util_format_get_nblocksx(texture->format, w);
nblocksy = util_format_get_nblocksy(texture->format, h);
}
pipe_resource_reference(&st->base.resource, texture);
st->base.level = level;
st->base.usage = usage;
st->base.box = *box;
st->base.stride = nblocksx*util_format_get_blocksize(texture->format);
st->base.layer_stride = st->base.stride * nblocksy;
switch (tex->b.b.target) {
case PIPE_TEXTURE_CUBE:
case PIPE_TEXTURE_2D_ARRAY:
case PIPE_TEXTURE_1D_ARRAY:
st->slice = st->base.box.z;
st->base.box.z = 0; /* so we don't apply double offsets below */
break;
default:
st->slice = 0;
break;
}
if (usage & PIPE_TRANSFER_WRITE) {
/* record texture upload for HUD */
svga->hud.num_bytes_uploaded +=
nblocksx * nblocksy * d * util_format_get_blocksize(texture->format);
}
if (!use_direct_map) {
/* Use a DMA buffer */
st->hw_nblocksy = nblocksy;
st->hwbuf = svga_winsys_buffer_create(svga, 1, 0,
st->hw_nblocksy * st->base.stride * d);
while(!st->hwbuf && (st->hw_nblocksy /= 2)) {
st->hwbuf = svga_winsys_buffer_create(svga, 1, 0,
st->hw_nblocksy * st->base.stride * d);
}
if (!st->hwbuf) {
FREE(st);
return NULL;
}
if (st->hw_nblocksy < nblocksy) {
/* We couldn't allocate a hardware buffer big enough for the transfer,
* so allocate regular malloc memory instead */
if (0) {
debug_printf("%s: failed to allocate %u KB of DMA, "
"splitting into %u x %u KB DMA transfers\n",
__FUNCTION__,
(nblocksy*st->base.stride + 1023)/1024,
(nblocksy + st->hw_nblocksy - 1)/st->hw_nblocksy,
(st->hw_nblocksy*st->base.stride + 1023)/1024);
}
st->swbuf = MALLOC(nblocksy * st->base.stride * d);
if (!st->swbuf) {
sws->buffer_destroy(sws, st->hwbuf);
FREE(st);
return NULL;
}
}
if (usage & PIPE_TRANSFER_READ) {
SVGA3dSurfaceDMAFlags flags;
memset(&flags, 0, sizeof flags);
svga_transfer_dma(svga, st, SVGA3D_READ_HOST_VRAM, flags);
}
} else {
struct pipe_transfer *transfer = &st->base;
struct svga_winsys_surface *surf = tex->handle;
if (!surf) {
FREE(st);
return NULL;
}
if (need_tex_readback(transfer)) {
enum pipe_error ret;
svga_surfaces_flush(svga);
if (svga_have_vgpu10(svga)) {
ret = readback_image_vgpu10(svga, surf, st->slice, transfer->level,
tex->b.b.last_level + 1);
} else {
ret = readback_image_vgpu9(svga, surf, st->slice, transfer->level);
}
assert(ret == PIPE_OK);
(void) ret;
svga_context_flush(svga, NULL);
/*
* Note: if PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE were specified
* we could potentially clear the flag for all faces/layers/mips.
*/
svga_clear_texture_rendered_to(tex, st->slice, transfer->level);
}
else {
assert(transfer->usage & PIPE_TRANSFER_WRITE);
if ((transfer->usage & PIPE_TRANSFER_UNSYNCHRONIZED) == 0) {
svga_surfaces_flush(svga);
if (!sws->surface_is_flushed(sws, surf))
svga_context_flush(svga, NULL);
}
}
}
st->use_direct_map = use_direct_map;
*ptransfer = &st->base;
/*
* Begin mapping code
*/
if (st->swbuf) {
returnVal = st->swbuf;
}
else if (!st->use_direct_map) {
returnVal = sws->buffer_map(sws, st->hwbuf, usage);
}
else {
SVGA3dSize baseLevelSize;
struct svga_texture *tex = svga_texture(texture);
struct svga_winsys_surface *surf = tex->handle;
uint8_t *map;
boolean retry;
unsigned offset, mip_width, mip_height;
unsigned xoffset = st->base.box.x;
unsigned yoffset = st->base.box.y;
unsigned zoffset = st->base.box.z;
map = svga->swc->surface_map(svga->swc, surf, usage, &retry);
if (map == NULL && retry) {
/*
* At this point, the svga_surfaces_flush() should already have
* called in svga_texture_get_transfer().
*/
svga_context_flush(svga, NULL);
map = svga->swc->surface_map(svga->swc, surf, usage, &retry);
}
/*
* Make sure we return NULL if the map fails
*/
if (!map) {
FREE(st);
return map;
}
/**
* Compute the offset to the specific texture slice in the buffer.
*/
baseLevelSize.width = tex->b.b.width0;
baseLevelSize.height = tex->b.b.height0;
baseLevelSize.depth = tex->b.b.depth0;
offset = svga3dsurface_get_image_offset(tex->key.format, baseLevelSize,
tex->b.b.last_level + 1, /* numMips */
st->slice, level);
if (level > 0) {
assert(offset > 0);
}
mip_width = u_minify(tex->b.b.width0, level);
mip_height = u_minify(tex->b.b.height0, level);
offset += svga3dsurface_get_pixel_offset(tex->key.format,
mip_width, mip_height,
xoffset, yoffset, zoffset);
returnVal = (void *) (map + offset);
}
svga->hud.map_buffer_time += (os_time_get() - begin);
svga->hud.num_resources_mapped++;
return returnVal;
}
/**
* Use DMA for the transfer request
*/
static void *
svga_texture_transfer_map_dma(struct svga_context *svga,
struct svga_transfer *st)
{
struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
struct pipe_resource *texture = st->base.resource;
unsigned nblocksx, nblocksy;
unsigned d;
unsigned usage = st->base.usage;
/* we'll put the data into a tightly packed buffer */
nblocksx = util_format_get_nblocksx(texture->format, st->box.w);
nblocksy = util_format_get_nblocksy(texture->format, st->box.h);
d = st->box.d;
st->base.stride = nblocksx*util_format_get_blocksize(texture->format);
st->base.layer_stride = st->base.stride * nblocksy;
st->hw_nblocksy = nblocksy;
st->hwbuf = svga_winsys_buffer_create(svga, 1, 0,
st->hw_nblocksy * st->base.stride * d);
while (!st->hwbuf && (st->hw_nblocksy /= 2)) {
st->hwbuf =
svga_winsys_buffer_create(svga, 1, 0,
st->hw_nblocksy * st->base.stride * d);
}
if (!st->hwbuf)
return NULL;
if (st->hw_nblocksy < nblocksy) {
/* We couldn't allocate a hardware buffer big enough for the transfer,
* so allocate regular malloc memory instead
*/
if (0) {
debug_printf("%s: failed to allocate %u KB of DMA, "
"splitting into %u x %u KB DMA transfers\n",
__FUNCTION__,
(nblocksy * st->base.stride + 1023) / 1024,
(nblocksy + st->hw_nblocksy - 1) / st->hw_nblocksy,
(st->hw_nblocksy * st->base.stride + 1023) / 1024);
}
st->swbuf = MALLOC(nblocksy * st->base.stride * d);
if (!st->swbuf) {
sws->buffer_destroy(sws, st->hwbuf);
return NULL;
}
}
if (usage & PIPE_TRANSFER_READ) {
SVGA3dSurfaceDMAFlags flags;
memset(&flags, 0, sizeof flags);
svga_transfer_dma(svga, st, SVGA3D_READ_HOST_VRAM, flags);
}
if (st->swbuf) {
return st->swbuf;
}
else {
return sws->buffer_map(sws, st->hwbuf, usage);
}
}
/* XXX: Still implementing this as if it was a screen function, but
* can now modify it to queue transfers on the context.
*/
static struct pipe_transfer *
svga_texture_get_transfer(struct pipe_context *pipe,
struct pipe_resource *texture,
unsigned level,
unsigned usage,
const struct pipe_box *box)
{
struct svga_context *svga = svga_context(pipe);
struct svga_screen *ss = svga_screen(pipe->screen);
struct svga_winsys_screen *sws = ss->sws;
struct svga_transfer *st;
unsigned nblocksx = util_format_get_nblocksx(texture->format, box->width);
unsigned nblocksy = util_format_get_nblocksy(texture->format, box->height);
/* We can't map texture storage directly */
if (usage & PIPE_TRANSFER_MAP_DIRECTLY)
return NULL;
assert(box->depth == 1);
st = CALLOC_STRUCT(svga_transfer);
if (!st)
return NULL;
pipe_resource_reference(&st->base.resource, texture);
st->base.level = level;
st->base.usage = usage;
st->base.box = *box;
st->base.stride = nblocksx*util_format_get_blocksize(texture->format);
st->base.layer_stride = 0;
st->hw_nblocksy = nblocksy;
st->hwbuf = svga_winsys_buffer_create(svga,
1,
0,
st->hw_nblocksy*st->base.stride);
while(!st->hwbuf && (st->hw_nblocksy /= 2)) {
st->hwbuf = svga_winsys_buffer_create(svga,
1,
0,
st->hw_nblocksy*st->base.stride);
}
if(!st->hwbuf)
goto no_hwbuf;
if(st->hw_nblocksy < nblocksy) {
/* We couldn't allocate a hardware buffer big enough for the transfer,
* so allocate regular malloc memory instead */
if (0) {
debug_printf("%s: failed to allocate %u KB of DMA, "
"splitting into %u x %u KB DMA transfers\n",
__FUNCTION__,
(nblocksy*st->base.stride + 1023)/1024,
(nblocksy + st->hw_nblocksy - 1)/st->hw_nblocksy,
(st->hw_nblocksy*st->base.stride + 1023)/1024);
}
st->swbuf = MALLOC(nblocksy*st->base.stride);
if(!st->swbuf)
goto no_swbuf;
}
if (usage & PIPE_TRANSFER_READ) {
SVGA3dSurfaceDMAFlags flags;
memset(&flags, 0, sizeof flags);
svga_transfer_dma(svga, st, SVGA3D_READ_HOST_VRAM, flags);
}
return &st->base;
no_swbuf:
sws->buffer_destroy(sws, st->hwbuf);
no_hwbuf:
FREE(st);
return NULL;
}
/**
* Upload the buffer to the host in a piecewise fashion.
*
* Used when the buffer is too big to fit in the GMR aperture.
*/
static INLINE enum pipe_error
svga_buffer_upload_piecewise(struct svga_screen *ss,
struct svga_context *svga,
struct svga_buffer *sbuf)
{
struct svga_winsys_screen *sws = ss->sws;
const unsigned alignment = sizeof(void *);
const unsigned usage = 0;
unsigned i;
assert(sbuf->map.num_ranges);
assert(!sbuf->dma.pending);
SVGA_DBG(DEBUG_DMA, "dma to sid %p\n", sbuf->handle);
for (i = 0; i < sbuf->map.num_ranges; ++i) {
struct svga_buffer_range *range = &sbuf->map.ranges[i];
unsigned offset = range->start;
unsigned size = range->end - range->start;
while (offset < range->end) {
struct svga_winsys_buffer *hwbuf;
uint8_t *map;
enum pipe_error ret;
if (offset + size > range->end)
size = range->end - offset;
hwbuf = svga_winsys_buffer_create(ss, alignment, usage, size);
while (!hwbuf) {
size /= 2;
if (!size)
return PIPE_ERROR_OUT_OF_MEMORY;
hwbuf = svga_winsys_buffer_create(ss, alignment, usage, size);
}
SVGA_DBG(DEBUG_DMA, " bytes %u - %u\n",
offset, offset + size);
map = sws->buffer_map(sws, hwbuf,
PIPE_BUFFER_USAGE_CPU_WRITE |
PIPE_BUFFER_USAGE_DISCARD);
assert(map);
if (map) {
memcpy(map, sbuf->swbuf, size);
sws->buffer_unmap(sws, hwbuf);
}
ret = SVGA3D_BufferDMA(svga->swc,
hwbuf, sbuf->handle,
SVGA3D_WRITE_HOST_VRAM,
size, 0, offset, sbuf->dma.flags);
if(ret != PIPE_OK) {
svga_context_flush(svga, NULL);
ret = SVGA3D_BufferDMA(svga->swc,
hwbuf, sbuf->handle,
SVGA3D_WRITE_HOST_VRAM,
size, 0, offset, sbuf->dma.flags);
assert(ret == PIPE_OK);
}
sbuf->dma.flags.discard = FALSE;
sws->buffer_destroy(sws, hwbuf);
offset += size;
}
}
sbuf->map.num_ranges = 0;
return PIPE_OK;
}