static void *
svga_create_gs_state(struct pipe_context *pipe,
const struct pipe_shader_state *templ)
{
struct svga_context *svga = svga_context(pipe);
struct svga_geometry_shader *gs = CALLOC_STRUCT(svga_geometry_shader);
if (!gs)
return NULL;
SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_CREATEGS);
gs->base.tokens = tgsi_dup_tokens(templ->tokens);
/* Collect basic info that we'll need later:
*/
tgsi_scan_shader(gs->base.tokens, &gs->base.info);
gs->draw_shader = draw_create_geometry_shader(svga->swtnl.draw, templ);
gs->base.id = svga->debug.shader_id++;
gs->generic_outputs = svga_get_generic_outputs_mask(&gs->base.info);
/* check for any stream output declarations */
if (templ->stream_output.num_outputs) {
gs->base.stream_output = svga_create_stream_output(svga, &gs->base,
&templ->stream_output);
}
SVGA_STATS_TIME_POP(svga_sws(svga));
return gs;
}
static boolean
svga_fence_finish(struct pipe_screen *screen,
struct pipe_context *ctx,
struct pipe_fence_handle *fence,
uint64_t timeout)
{
struct svga_winsys_screen *sws = svga_screen(screen)->sws;
boolean retVal;
SVGA_STATS_TIME_PUSH(sws, SVGA_STATS_TIME_FENCEFINISH);
if (!timeout) {
retVal = sws->fence_signalled(sws, fence, 0) == 0;
}
else {
SVGA_DBG(DEBUG_DMA|DEBUG_PERF, "%s fence_ptr %p\n",
__FUNCTION__, fence);
retVal = sws->fence_finish(sws, fence, 0) == 0;
}
SVGA_STATS_TIME_POP(sws);
return retVal;
}
static enum pipe_error
retry_draw_range_elements( struct svga_context *svga,
struct pipe_resource *index_buffer,
unsigned index_size,
int index_bias,
unsigned min_index,
unsigned max_index,
enum pipe_prim_type prim,
unsigned start,
unsigned count,
unsigned start_instance,
unsigned instance_count,
boolean do_retry )
{
enum pipe_error ret = PIPE_OK;
SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_DRAWELEMENTS);
svga_hwtnl_set_fillmode(svga->hwtnl, svga->curr.rast->hw_fillmode);
ret = svga_update_state( svga, SVGA_STATE_HW_DRAW );
if (ret != PIPE_OK)
goto retry;
/** determine if flatshade is to be used after svga_update_state()
* in case the fragment shader is changed.
*/
svga_hwtnl_set_flatshade(svga->hwtnl,
svga->curr.rast->templ.flatshade ||
svga->state.hw_draw.fs->uses_flat_interp,
svga->curr.rast->templ.flatshade_first);
ret = svga_hwtnl_draw_range_elements( svga->hwtnl,
index_buffer, index_size, index_bias,
min_index, max_index,
prim, start, count,
start_instance, instance_count);
if (ret != PIPE_OK)
goto retry;
goto done;
retry:
svga_context_flush( svga, NULL );
if (do_retry)
{
ret = retry_draw_range_elements(svga,
index_buffer, index_size, index_bias,
min_index, max_index,
prim, start, count,
start_instance, instance_count, FALSE);
}
done:
SVGA_STATS_TIME_POP(svga_sws(svga));
return ret;
}
/**
* All drawing filters down into this function, either directly
* on the hardware path or after doing software vertex processing.
*/
enum pipe_error
svga_hwtnl_prim(struct svga_hwtnl *hwtnl,
const SVGA3dPrimitiveRange * range,
unsigned vcount,
unsigned min_index,
unsigned max_index, struct pipe_resource *ib,
unsigned start_instance, unsigned instance_count)
{
enum pipe_error ret = PIPE_OK;
SVGA_STATS_TIME_PUSH(svga_sws(hwtnl->svga), SVGA_STATS_TIME_HWTNLPRIM);
if (svga_have_vgpu10(hwtnl->svga)) {
/* draw immediately */
ret = draw_vgpu10(hwtnl, range, vcount, min_index, max_index, ib,
start_instance, instance_count);
if (ret != PIPE_OK) {
svga_context_flush(hwtnl->svga, NULL);
ret = draw_vgpu10(hwtnl, range, vcount, min_index, max_index, ib,
start_instance, instance_count);
assert(ret == PIPE_OK);
}
}
else {
/* batch up drawing commands */
#ifdef DEBUG
check_draw_params(hwtnl, range, min_index, max_index, ib);
assert(start_instance == 0);
assert(instance_count <= 1);
#else
(void) check_draw_params;
#endif
if (hwtnl->cmd.prim_count + 1 >= QSZ) {
ret = svga_hwtnl_flush(hwtnl);
if (ret != PIPE_OK)
goto done;
}
/* min/max indices are relative to bias */
hwtnl->cmd.min_index[hwtnl->cmd.prim_count] = min_index;
hwtnl->cmd.max_index[hwtnl->cmd.prim_count] = max_index;
hwtnl->cmd.prim[hwtnl->cmd.prim_count] = *range;
hwtnl->cmd.prim[hwtnl->cmd.prim_count].indexBias += hwtnl->index_bias;
pipe_resource_reference(&hwtnl->cmd.prim_ib[hwtnl->cmd.prim_count], ib);
hwtnl->cmd.prim_count++;
}
done:
SVGA_STATS_TIME_POP(svga_screen(hwtnl->svga->pipe.screen)->sws);
return ret;
}
/**
* Fallback to the copy region utility which uses map/memcpy for the copy
*/
static void
copy_region_fallback(struct svga_context *svga,
struct pipe_resource *dst_tex, unsigned dst_level,
unsigned dstx, unsigned dsty, unsigned dstz,
struct pipe_resource *src_tex, unsigned src_level,
const struct pipe_box *src_box)
{
struct svga_winsys_screen *sws = svga_screen(svga->pipe.screen)->sws;
SVGA_STATS_TIME_PUSH(sws, SVGA_STATS_TIME_COPYREGIONFALLBACK);
util_resource_copy_region(&svga->pipe, dst_tex, dst_level, dstx,
dsty, dstz, src_tex, src_level, src_box);
SVGA_STATS_TIME_POP(sws);
(void) sws;
}
/**
* Emit any pending drawing commands to the command buffer.
* When we receive VGPU9 drawing commands we accumulate them and don't
* immediately emit them into the command buffer.
* This function needs to be called before we change state that could
* effect those pending draws.
*/
enum pipe_error
svga_hwtnl_flush(struct svga_hwtnl *hwtnl)
{
enum pipe_error ret = PIPE_OK;
SVGA_STATS_TIME_PUSH(svga_sws(hwtnl->svga), SVGA_STATS_TIME_HWTNLFLUSH);
if (!svga_have_vgpu10(hwtnl->svga) && hwtnl->cmd.prim_count) {
/* we only queue up primitive for VGPU9 */
ret = draw_vgpu9(hwtnl);
}
SVGA_STATS_TIME_POP(svga_screen(hwtnl->svga->pipe.screen)->sws);
return ret;
}
static enum pipe_error
retry_draw_arrays( struct svga_context *svga,
enum pipe_prim_type prim, unsigned start, unsigned count,
unsigned start_instance, unsigned instance_count,
boolean do_retry )
{
enum pipe_error ret;
SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_DRAWARRAYS);
svga_hwtnl_set_fillmode(svga->hwtnl, svga->curr.rast->hw_fillmode);
ret = svga_update_state( svga, SVGA_STATE_HW_DRAW );
if (ret != PIPE_OK)
goto retry;
/** determine if flatshade is to be used after svga_update_state()
* in case the fragment shader is changed.
*/
svga_hwtnl_set_flatshade(svga->hwtnl,
svga->curr.rast->templ.flatshade ||
svga->state.hw_draw.fs->uses_flat_interp,
svga->curr.rast->templ.flatshade_first);
ret = svga_hwtnl_draw_arrays(svga->hwtnl, prim, start, count,
start_instance, instance_count);
if (ret != PIPE_OK)
goto retry;
goto done;
retry:
if (ret == PIPE_ERROR_OUT_OF_MEMORY && do_retry)
{
svga_context_flush( svga, NULL );
ret = retry_draw_arrays(svga, prim, start, count,
start_instance, instance_count,
FALSE);
}
done:
SVGA_STATS_TIME_POP(svga_sws(svga));
return ret;
}
static void
svga_buffer_transfer_unmap(struct pipe_context *pipe,
struct pipe_transfer *transfer)
{
struct svga_screen *ss = svga_screen(pipe->screen);
struct svga_context *svga = svga_context(pipe);
struct svga_buffer *sbuf = svga_buffer(transfer->resource);
SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_BUFFERTRANSFERUNMAP);
mtx_lock(&ss->swc_mutex);
assert(sbuf->map.count);
if (sbuf->map.count) {
--sbuf->map.count;
}
if (svga_buffer_has_hw_storage(sbuf)) {
/* Note: we may wind up flushing here and unmapping other buffers
* which leads to recursively locking ss->swc_mutex.
*/
svga_buffer_hw_storage_unmap(svga, sbuf);
}
if (transfer->usage & PIPE_TRANSFER_WRITE) {
if (!(transfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT)) {
/*
* Mapped range not flushed explicitly, so flush the whole buffer,
* and tell the host to discard the contents when processing the DMA
* command.
*/
SVGA_DBG(DEBUG_DMA, "flushing the whole buffer\n");
sbuf->dma.flags.discard = TRUE;
svga_buffer_add_range(sbuf, 0, sbuf->b.b.width0);
}
}
mtx_unlock(&ss->swc_mutex);
FREE(transfer);
SVGA_STATS_TIME_POP(svga_sws(svga));
}
enum pipe_error
svga_update_state(struct svga_context *svga, unsigned max_level)
{
struct svga_screen *screen = svga_screen(svga->pipe.screen);
enum pipe_error ret = PIPE_OK;
unsigned i;
SVGA_STATS_TIME_PUSH(screen->sws, SVGA_STATS_TIME_UPDATESTATE);
/* Check for updates to bound textures. This can't be done in an
* atom as there is no flag which could provoke this test, and we
* cannot create one.
*/
if (svga->state.texture_timestamp != screen->texture_timestamp) {
svga->state.texture_timestamp = screen->texture_timestamp;
svga->dirty |= SVGA_NEW_TEXTURE;
}
for (i = 0; i <= max_level; i++) {
svga->dirty |= svga->state.dirty[i];
if (svga->dirty) {
ret = update_state( svga,
state_levels[i],
&svga->dirty );
if (ret != PIPE_OK)
goto done;
svga->state.dirty[i] = 0;
}
}
for (; i < SVGA_STATE_MAX; i++)
svga->state.dirty[i] |= svga->dirty;
svga->dirty = 0;
svga->hud.num_validations++;
done:
SVGA_STATS_TIME_POP(screen->sws);
return ret;
}
static void *
svga_create_vs_state(struct pipe_context *pipe,
const struct pipe_shader_state *templ)
{
struct svga_context *svga = svga_context(pipe);
struct svga_vertex_shader *vs = CALLOC_STRUCT(svga_vertex_shader);
if (!vs)
return NULL;
SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_CREATEVS);
/* substitute a debug shader?
*/
vs->base.tokens = tgsi_dup_tokens(substitute_vs(svga->debug.shader_id,
templ->tokens));
/* Collect basic info that we'll need later:
*/
tgsi_scan_shader(vs->base.tokens, &vs->base.info);
{
/* Need to do construct a new template in case we substitued a
* debug shader.
*/
struct pipe_shader_state tmp2 = *templ;
tmp2.tokens = vs->base.tokens;
vs->draw_shader = draw_create_vertex_shader(svga->swtnl.draw, &tmp2);
}
vs->base.id = svga->debug.shader_id++;
vs->generic_outputs = svga_get_generic_outputs_mask(&vs->base.info);
/* check for any stream output declarations */
if (templ->stream_output.num_outputs) {
vs->base.stream_output = svga_create_stream_output(svga, &vs->base,
&templ->stream_output);
}
SVGA_STATS_TIME_POP(svga_sws(svga));
return vs;
}
static void
svga_texture_transfer_unmap(struct pipe_context *pipe,
struct pipe_transfer *transfer)
{
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 = svga_transfer(transfer);
struct svga_texture *tex = svga_texture(transfer->resource);
SVGA_STATS_TIME_PUSH(sws, SVGA_STATS_TIME_TEXTRANSFERUNMAP);
if (!st->use_direct_map) {
svga_texture_transfer_unmap_dma(svga, st);
}
else if (st->upload.buf) {
svga_texture_transfer_unmap_upload(svga, st);
}
else {
svga_texture_transfer_unmap_direct(svga, st);
}
if (st->base.usage & PIPE_TRANSFER_WRITE) {
svga->hud.num_resource_updates++;
/* Mark the texture level as dirty */
ss->texture_timestamp++;
svga_age_texture_view(tex, transfer->level);
if (transfer->resource->target == PIPE_TEXTURE_CUBE)
svga_define_texture_level(tex, st->slice, transfer->level);
else
svga_define_texture_level(tex, 0, transfer->level);
}
pipe_resource_reference(&st->base.resource, NULL);
FREE(st);
SVGA_STATS_TIME_POP(sws);
(void) sws;
}
/**
* Request a transfer map to the texture resource
*/
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_winsys_screen *sws = svga_screen(pipe->screen)->sws;
struct svga_texture *tex = svga_texture(texture);
struct svga_transfer *st;
struct svga_winsys_surface *surf = tex->handle;
boolean use_direct_map = svga_have_gb_objects(svga) &&
!svga_have_gb_dma(svga);
void *map = NULL;
int64_t begin = svga_get_time(svga);
SVGA_STATS_TIME_PUSH(sws, SVGA_STATS_TIME_TEXTRANSFERMAP);
if (!surf)
goto done;
/* 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
goto done;
}
st = CALLOC_STRUCT(svga_transfer);
if (!st)
goto done;
st->base.level = level;
st->base.usage = usage;
st->base.box = *box;
/* The modified transfer map box with the array index removed from z.
* The array index is specified in slice.
*/
st->box.x = box->x;
st->box.y = box->y;
st->box.z = box->z;
st->box.w = box->width;
st->box.h = box->height;
st->box.d = box->depth;
switch (tex->b.b.target) {
case PIPE_TEXTURE_CUBE:
st->slice = st->base.box.z;
st->box.z = 0; /* so we don't apply double offsets below */
break;
case PIPE_TEXTURE_1D_ARRAY:
case PIPE_TEXTURE_2D_ARRAY:
case PIPE_TEXTURE_CUBE_ARRAY:
st->slice = st->base.box.z;
st->box.z = 0; /* so we don't apply double offsets below */
/* Force direct map for transfering multiple slices */
if (st->base.box.depth > 1)
use_direct_map = svga_have_gb_objects(svga);
break;
default:
st->slice = 0;
break;
}
/* Force direct map for multisample surface */
if (texture->nr_samples > 1) {
assert(svga_have_gb_objects(svga));
assert(sws->have_sm4_1);
use_direct_map = TRUE;
}
st->use_direct_map = use_direct_map;
pipe_resource_reference(&st->base.resource, texture);
/* If this is the first time mapping to the surface in this
* command buffer, clear the dirty masks of this surface.
*/
if (sws->surface_is_flushed(sws, surf)) {
svga_clear_texture_dirty(tex);
}
if (!use_direct_map) {
/* upload to the DMA buffer */
map = svga_texture_transfer_map_dma(svga, st);
}
else {
boolean can_use_upload = tex->can_use_upload &&
!(st->base.usage & PIPE_TRANSFER_READ);
boolean was_rendered_to =
svga_was_texture_rendered_to(svga_texture(texture),
st->slice, st->base.level);
/* If the texture was already rendered to and upload buffer
* is supported, then we will use upload buffer to
* avoid the need to read back the texture content; otherwise,
* we'll first try to map directly to the GB surface, if it is blocked,
* then we'll try the upload buffer.
*/
if (was_rendered_to && can_use_upload) {
map = svga_texture_transfer_map_upload(svga, st);
}
else {
unsigned orig_usage = st->base.usage;
/* First try directly map to the GB surface */
if (can_use_upload)
st->base.usage |= PIPE_TRANSFER_DONTBLOCK;
map = svga_texture_transfer_map_direct(svga, st);
st->base.usage = orig_usage;
if (!map && can_use_upload) {
/* if direct map with DONTBLOCK fails, then try upload to the
* texture upload buffer.
*/
map = svga_texture_transfer_map_upload(svga, st);
}
}
/* If upload fails, then try direct map again without forcing it
* to DONTBLOCK.
*/
if (!map) {
map = svga_texture_transfer_map_direct(svga, st);
}
}
if (!map) {
FREE(st);
}
else {
*ptransfer = &st->base;
svga->hud.num_textures_mapped++;
if (usage & PIPE_TRANSFER_WRITE) {
/* record texture upload for HUD */
svga->hud.num_bytes_uploaded +=
st->base.layer_stride * st->box.d;
/* mark this texture level as dirty */
svga_set_texture_dirty(tex, st->slice, level);
}
}
done:
svga->hud.map_buffer_time += (svga_get_time(svga) - begin);
SVGA_STATS_TIME_POP(sws);
(void) sws;
return map;
}
/**
* Parse TGSI shader and translate to SVGA/DX9 serialized
* representation.
*
* In this function SVGA shader is emitted to an in-memory buffer that
* can be dynamically grown. Once we've finished and know how large
* it is, it will be copied to a hardware buffer for upload.
*/
struct svga_shader_variant *
svga_tgsi_vgpu9_translate(struct svga_context *svga,
const struct svga_shader *shader,
const struct svga_compile_key *key,
enum pipe_shader_type unit)
{
struct svga_shader_variant *variant = NULL;
struct svga_shader_emitter emit;
SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_TGSIVGPU9TRANSLATE);
memset(&emit, 0, sizeof(emit));
emit.size = 1024;
emit.buf = MALLOC(emit.size);
if (emit.buf == NULL) {
goto fail;
}
emit.ptr = emit.buf;
emit.unit = unit;
emit.key = *key;
tgsi_scan_shader(shader->tokens, &emit.info);
emit.imm_start = emit.info.file_max[TGSI_FILE_CONSTANT] + 1;
if (unit == PIPE_SHADER_FRAGMENT)
emit.imm_start += key->num_unnormalized_coords;
if (unit == PIPE_SHADER_VERTEX) {
emit.imm_start += key->vs.need_prescale ? 2 : 0;
}
emit.nr_hw_float_const =
(emit.imm_start + emit.info.file_max[TGSI_FILE_IMMEDIATE] + 1);
emit.nr_hw_temp = emit.info.file_max[TGSI_FILE_TEMPORARY] + 1;
if (emit.nr_hw_temp >= SVGA3D_TEMPREG_MAX) {
debug_printf("svga: too many temporary registers (%u)\n",
emit.nr_hw_temp);
goto fail;
}
if (emit.info.indirect_files & (1 << TGSI_FILE_TEMPORARY)) {
debug_printf(
"svga: indirect indexing of temporary registers is not supported.\n");
goto fail;
}
emit.in_main_func = TRUE;
if (!svga_shader_emit_header(&emit)) {
debug_printf("svga: emit header failed\n");
goto fail;
}
if (!svga_shader_emit_instructions(&emit, shader->tokens)) {
debug_printf("svga: emit instructions failed\n");
goto fail;
}
variant = svga_new_shader_variant(svga);
if (!variant)
goto fail;
variant->shader = shader;
variant->tokens = (const unsigned *) emit.buf;
variant->nr_tokens = (emit.ptr - emit.buf) / sizeof(unsigned);
memcpy(&variant->key, key, sizeof(*key));
variant->id = UTIL_BITMASK_INVALID_INDEX;
variant->pstipple_sampler_unit = emit.pstipple_sampler_unit;
/* If there was exactly one write to a fragment shader output register
* and it came from a constant buffer, we know all fragments will have
* the same color (except for blending).
*/
variant->constant_color_output =
emit.constant_color_output && emit.num_output_writes == 1;
#if 0
if (!svga_shader_verify(variant->tokens, variant->nr_tokens) ||
SVGA_DEBUG & DEBUG_TGSI) {
debug_printf("#####################################\n");
debug_printf("Shader %u below\n", shader->id);
tgsi_dump(shader->tokens, 0);
if (SVGA_DEBUG & DEBUG_TGSI) {
debug_printf("Shader %u compiled below\n", shader->id);
svga_shader_dump(variant->tokens, variant->nr_tokens, FALSE);
}
debug_printf("#####################################\n");
}
#endif
goto done;
fail:
FREE(variant);
if (emit.buf != err_buf)
FREE(emit.buf);
variant = NULL;
done:
SVGA_STATS_TIME_POP(svga_sws(svga));
return variant;
}
/**
* Create a buffer transfer.
*
* Unlike texture DMAs (which are written immediately to the command buffer and
* therefore inherently serialized with other context operations), for buffers
* we try to coalesce multiple range mappings (i.e, multiple calls to this
* function) into a single DMA command, for better efficiency in command
* processing. This means we need to exercise extra care here to ensure that
* the end result is exactly the same as if one DMA was used for every mapped
* range.
*/
static void *
svga_buffer_transfer_map(struct pipe_context *pipe,
struct pipe_resource *resource,
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_buffer *sbuf = svga_buffer(resource);
struct pipe_transfer *transfer;
uint8_t *map = NULL;
int64_t begin = svga_get_time(svga);
SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_BUFFERTRANSFERMAP);
assert(box->y == 0);
assert(box->z == 0);
assert(box->height == 1);
assert(box->depth == 1);
transfer = MALLOC_STRUCT(pipe_transfer);
if (!transfer) {
goto done;
}
transfer->resource = resource;
transfer->level = level;
transfer->usage = usage;
transfer->box = *box;
transfer->stride = 0;
transfer->layer_stride = 0;
if (usage & PIPE_TRANSFER_WRITE) {
/* If we write to the buffer for any reason, free any saved translated
* vertices.
*/
pipe_resource_reference(&sbuf->translated_indices.buffer, NULL);
}
if ((usage & PIPE_TRANSFER_READ) && sbuf->dirty) {
enum pipe_error ret;
/* Host-side buffers can only be dirtied with vgpu10 features
* (streamout and buffer copy).
*/
assert(svga_have_vgpu10(svga));
if (!sbuf->user) {
(void) svga_buffer_handle(svga, resource, sbuf->bind_flags);
}
if (sbuf->dma.pending) {
svga_buffer_upload_flush(svga, sbuf);
svga_context_finish(svga);
}
assert(sbuf->handle);
ret = SVGA3D_vgpu10_ReadbackSubResource(svga->swc, sbuf->handle, 0);
if (ret != PIPE_OK) {
svga_context_flush(svga, NULL);
ret = SVGA3D_vgpu10_ReadbackSubResource(svga->swc, sbuf->handle, 0);
assert(ret == PIPE_OK);
}
svga->hud.num_readbacks++;
svga_context_finish(svga);
sbuf->dirty = FALSE;
}
if (usage & PIPE_TRANSFER_WRITE) {
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
/*
* Flush any pending primitives, finish writing any pending DMA
* commands, and tell the host to discard the buffer contents on
* the next DMA operation.
*/
svga_hwtnl_flush_buffer(svga, resource);
if (sbuf->dma.pending) {
svga_buffer_upload_flush(svga, sbuf);
/*
* Instead of flushing the context command buffer, simply discard
* the current hwbuf, and start a new one.
* With GB objects, the map operation takes care of this
* if passed the PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE flag,
* and the old backing store is busy.
*/
if (!svga_have_gb_objects(svga))
svga_buffer_destroy_hw_storage(ss, sbuf);
}
sbuf->map.num_ranges = 0;
sbuf->dma.flags.discard = TRUE;
}
if (usage & PIPE_TRANSFER_UNSYNCHRONIZED) {
if (!sbuf->map.num_ranges) {
/*
* No pending ranges to upload so far, so we can tell the host to
* not synchronize on the next DMA command.
*/
sbuf->dma.flags.unsynchronized = TRUE;
}
} else {
/*
* Synchronizing, so flush any pending primitives, finish writing any
* pending DMA command, and ensure the next DMA will be done in order.
*/
svga_hwtnl_flush_buffer(svga, resource);
if (sbuf->dma.pending) {
svga_buffer_upload_flush(svga, sbuf);
if (svga_buffer_has_hw_storage(sbuf)) {
/*
* We have a pending DMA upload from a hardware buffer, therefore
* we need to ensure that the host finishes processing that DMA
* command before the state tracker can start overwriting the
* hardware buffer.
*
* XXX: This could be avoided by tying the hardware buffer to
* the transfer (just as done with textures), which would allow
* overlapping DMAs commands to be queued on the same context
* buffer. However, due to the likelihood of software vertex
* processing, it is more convenient to hold on to the hardware
* buffer, allowing to quickly access the contents from the CPU
* without having to do a DMA download from the host.
*/
if (usage & PIPE_TRANSFER_DONTBLOCK) {
/*
* Flushing the command buffer here will most likely cause
* the map of the hwbuf below to block, so preemptively
* return NULL here if DONTBLOCK is set to prevent unnecessary
* command buffer flushes.
*/
FREE(transfer);
goto done;
}
svga_context_flush(svga, NULL);
}
}
sbuf->dma.flags.unsynchronized = FALSE;
}
}
if (!sbuf->swbuf && !svga_buffer_has_hw_storage(sbuf)) {
if (svga_buffer_create_hw_storage(ss, sbuf, sbuf->bind_flags) != PIPE_OK) {
/*
* We can't create a hardware buffer big enough, so create a malloc
* buffer instead.
*/
if (0) {
debug_printf("%s: failed to allocate %u KB of DMA, "
"splitting DMA transfers\n",
__FUNCTION__,
(sbuf->b.b.width0 + 1023)/1024);
}
sbuf->swbuf = align_malloc(sbuf->b.b.width0, 16);
if (!sbuf->swbuf) {
FREE(transfer);
goto done;
}
}
}
if (sbuf->swbuf) {
/* User/malloc buffer */
map = sbuf->swbuf;
}
else if (svga_buffer_has_hw_storage(sbuf)) {
boolean retry;
map = svga_buffer_hw_storage_map(svga, sbuf, transfer->usage, &retry);
if (map == NULL && retry) {
/*
* At this point, svga_buffer_get_transfer() has already
* hit the DISCARD_WHOLE_RESOURCE path and flushed HWTNL
* for this buffer.
*/
svga_context_flush(svga, NULL);
map = svga_buffer_hw_storage_map(svga, sbuf, transfer->usage, &retry);
}
}
else {
map = NULL;
}
if (map) {
++sbuf->map.count;
map += transfer->box.x;
*ptransfer = transfer;
} else {
FREE(transfer);
}
svga->hud.map_buffer_time += (svga_get_time(svga) - begin);
done:
SVGA_STATS_TIME_POP(svga_sws(svga));
return map;
}
static void
svga_draw_vbo(struct pipe_context *pipe, const struct pipe_draw_info *info)
{
struct svga_context *svga = svga_context( pipe );
unsigned reduced_prim = u_reduced_prim( info->mode );
unsigned count = info->count;
enum pipe_error ret = 0;
boolean needed_swtnl;
SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_DRAWVBO);
svga->hud.num_draw_calls++; /* for SVGA_QUERY_NUM_DRAW_CALLS */
if (u_reduced_prim(info->mode) == PIPE_PRIM_TRIANGLES &&
svga->curr.rast->templ.cull_face == PIPE_FACE_FRONT_AND_BACK)
goto done;
/*
* Mark currently bound target surfaces as dirty
* doesn't really matter if it is done before drawing.
*
* TODO If we ever normaly return something other then
* true we should not mark it as dirty then.
*/
svga_mark_surfaces_dirty(svga_context(pipe));
if (svga->curr.reduced_prim != reduced_prim) {
svga->curr.reduced_prim = reduced_prim;
svga->dirty |= SVGA_NEW_REDUCED_PRIMITIVE;
}
if (need_fallback_prim_restart(svga, info)) {
enum pipe_error r;
r = util_draw_vbo_without_prim_restart(pipe, &svga->curr.ib, info);
assert(r == PIPE_OK);
(void) r;
goto done;
}
if (!u_trim_pipe_prim( info->mode, &count ))
goto done;
needed_swtnl = svga->state.sw.need_swtnl;
svga_update_state_retry( svga, SVGA_STATE_NEED_SWTNL );
if (svga->state.sw.need_swtnl) {
svga->hud.num_fallbacks++; /* for SVGA_QUERY_NUM_FALLBACKS */
if (!needed_swtnl) {
/*
* We're switching from HW to SW TNL. SW TNL will require mapping all
* currently bound vertex buffers, some of which may already be
* referenced in the current command buffer as result of previous HW
* TNL. So flush now, to prevent the context to flush while a referred
* vertex buffer is mapped.
*/
svga_context_flush(svga, NULL);
}
/* Avoid leaking the previous hwtnl bias to swtnl */
svga_hwtnl_set_index_bias( svga->hwtnl, 0 );
ret = svga_swtnl_draw_vbo( svga, info );
}
else {
if (info->indexed && svga->curr.ib.buffer) {
unsigned offset;
assert(svga->curr.ib.offset % svga->curr.ib.index_size == 0);
offset = svga->curr.ib.offset / svga->curr.ib.index_size;
ret = retry_draw_range_elements( svga,
svga->curr.ib.buffer,
svga->curr.ib.index_size,
info->index_bias,
info->min_index,
info->max_index,
info->mode,
info->start + offset,
count,
info->start_instance,
info->instance_count,
TRUE );
}
else {
ret = retry_draw_arrays(svga, info->mode, info->start, count,
info->start_instance, info->instance_count,
TRUE);
}
}
/* XXX: Silence warnings, do something sensible here? */
(void)ret;
if (SVGA_DEBUG & DEBUG_FLUSH) {
svga_hwtnl_flush_retry( svga );
svga_context_flush(svga, NULL);
}
done:
SVGA_STATS_TIME_POP(svga_sws(svga));
;
}
/**
* Define a vgpu10 sampler state.
*/
static void
define_sampler_state_object(struct svga_context *svga,
struct svga_sampler_state *ss,
const struct pipe_sampler_state *ps)
{
uint8_t max_aniso = (uint8_t) 255; /* XXX fix me */
boolean anisotropic;
uint8 compare_func;
SVGA3dFilter filter;
SVGA3dRGBAFloat bcolor;
unsigned try;
float min_lod, max_lod;
assert(svga_have_vgpu10(svga));
anisotropic = ss->aniso_level > 1.0f;
filter = translate_filter_mode(ps->min_mip_filter,
ps->min_img_filter,
ps->mag_img_filter,
anisotropic,
ss->compare_mode);
compare_func = translate_comparison_func(ss->compare_func);
COPY_4V(bcolor.value, ps->border_color.f);
assert(ps->min_lod <= ps->max_lod);
if (ps->min_mip_filter == PIPE_TEX_MIPFILTER_NONE) {
/* just use the base level image */
min_lod = max_lod = 0.0f;
}
else {
min_lod = ps->min_lod;
max_lod = ps->max_lod;
}
/* If shadow comparisons are enabled, create two sampler states: one
* with the given shadow compare mode, another with shadow comparison off.
* We need the later because in some cases, we have to do the shadow
* compare in the shader. So, we don't want to do it twice.
*/
STATIC_ASSERT(PIPE_TEX_COMPARE_NONE == 0);
STATIC_ASSERT(PIPE_TEX_COMPARE_R_TO_TEXTURE == 1);
ss->id[1] = SVGA3D_INVALID_ID;
unsigned i;
for (i = 0; i <= ss->compare_mode; i++) {
ss->id[i] = util_bitmask_add(svga->sampler_object_id_bm);
/* Loop in case command buffer is full and we need to flush and retry */
for (try = 0; try < 2; try++) {
enum pipe_error ret =
SVGA3D_vgpu10_DefineSamplerState(svga->swc,
ss->id[i],
filter,
ss->addressu,
ss->addressv,
ss->addressw,
ss->lod_bias, /* float */
max_aniso,
compare_func,
bcolor,
min_lod, /* float */
max_lod); /* float */
if (ret == PIPE_OK)
break;
svga_context_flush(svga, NULL);
}
/* turn off the shadow compare option for second iteration */
filter &= ~SVGA3D_FILTER_COMPARE;
}
}
static void *
svga_create_sampler_state(struct pipe_context *pipe,
const struct pipe_sampler_state *sampler)
{
struct svga_context *svga = svga_context(pipe);
struct svga_sampler_state *cso = CALLOC_STRUCT( svga_sampler_state );
if (!cso)
return NULL;
cso->mipfilter = translate_mip_filter(sampler->min_mip_filter);
cso->magfilter = translate_img_filter( sampler->mag_img_filter );
cso->minfilter = translate_img_filter( sampler->min_img_filter );
cso->aniso_level = MAX2( sampler->max_anisotropy, 1 );
if (sampler->max_anisotropy)
cso->magfilter = cso->minfilter = SVGA3D_TEX_FILTER_ANISOTROPIC;
cso->lod_bias = sampler->lod_bias;
cso->addressu = translate_wrap_mode(sampler->wrap_s);
cso->addressv = translate_wrap_mode(sampler->wrap_t);
cso->addressw = translate_wrap_mode(sampler->wrap_r);
cso->normalized_coords = sampler->normalized_coords;
cso->compare_mode = sampler->compare_mode;
cso->compare_func = sampler->compare_func;
{
uint32 r = float_to_ubyte(sampler->border_color.f[0]);
uint32 g = float_to_ubyte(sampler->border_color.f[1]);
uint32 b = float_to_ubyte(sampler->border_color.f[2]);
uint32 a = float_to_ubyte(sampler->border_color.f[3]);
cso->bordercolor = (a << 24) | (r << 16) | (g << 8) | b;
}
/* No SVGA3D support for:
* - min/max LOD clamping
*/
cso->min_lod = 0;
cso->view_min_lod = MAX2((int) (sampler->min_lod + 0.5), 0);
cso->view_max_lod = MAX2((int) (sampler->max_lod + 0.5), 0);
/* Use min_mipmap */
if (svga->debug.use_min_mipmap) {
if (cso->view_min_lod == cso->view_max_lod) {
cso->min_lod = cso->view_min_lod;
cso->view_min_lod = 0;
cso->view_max_lod = 1000; /* Just a high number */
cso->mipfilter = SVGA3D_TEX_FILTER_NONE;
}
}
if (svga_have_vgpu10(svga)) {
define_sampler_state_object(svga, cso, sampler);
}
SVGA_DBG(DEBUG_SAMPLERS,
"New sampler: min %u, view(min %u, max %u) lod, mipfilter %s\n",
cso->min_lod, cso->view_min_lod, cso->view_max_lod,
cso->mipfilter == SVGA3D_TEX_FILTER_NONE ? "SVGA3D_TEX_FILTER_NONE" : "SOMETHING");
svga->hud.num_sampler_objects++;
SVGA_STATS_COUNT_INC(svga_screen(svga->pipe.screen)->sws,
SVGA_STATS_COUNT_SAMPLER);
return cso;
}
static void
svga_bind_sampler_states(struct pipe_context *pipe,
enum pipe_shader_type shader,
unsigned start,
unsigned num,
void **samplers)
{
struct svga_context *svga = svga_context(pipe);
unsigned i;
boolean any_change = FALSE;
assert(shader < PIPE_SHADER_TYPES);
assert(start + num <= PIPE_MAX_SAMPLERS);
/* Pre-VGPU10 only supports FS textures */
if (!svga_have_vgpu10(svga) && shader != PIPE_SHADER_FRAGMENT)
return;
for (i = 0; i < num; i++) {
if (svga->curr.sampler[shader][start + i] != samplers[i])
any_change = TRUE;
svga->curr.sampler[shader][start + i] = samplers[i];
}
if (!any_change) {
return;
}
/* find highest non-null sampler[] entry */
{
unsigned j = MAX2(svga->curr.num_samplers[shader], start + num);
while (j > 0 && svga->curr.sampler[shader][j - 1] == NULL)
j--;
svga->curr.num_samplers[shader] = j;
}
svga->dirty |= SVGA_NEW_SAMPLER;
}
static void
svga_delete_sampler_state(struct pipe_context *pipe, void *sampler)
{
struct svga_sampler_state *ss = (struct svga_sampler_state *) sampler;
struct svga_context *svga = svga_context(pipe);
if (svga_have_vgpu10(svga)) {
unsigned i;
for (i = 0; i < 2; i++) {
enum pipe_error ret;
if (ss->id[i] != SVGA3D_INVALID_ID) {
svga_hwtnl_flush_retry(svga);
ret = SVGA3D_vgpu10_DestroySamplerState(svga->swc, ss->id[i]);
if (ret != PIPE_OK) {
svga_context_flush(svga, NULL);
ret = SVGA3D_vgpu10_DestroySamplerState(svga->swc, ss->id[i]);
}
util_bitmask_clear(svga->sampler_object_id_bm, ss->id[i]);
}
}
}
FREE(sampler);
svga->hud.num_sampler_objects--;
}
static struct pipe_sampler_view *
svga_create_sampler_view(struct pipe_context *pipe,
struct pipe_resource *texture,
const struct pipe_sampler_view *templ)
{
struct svga_context *svga = svga_context(pipe);
struct svga_pipe_sampler_view *sv = CALLOC_STRUCT(svga_pipe_sampler_view);
if (!sv) {
return NULL;
}
sv->base = *templ;
sv->base.reference.count = 1;
sv->base.texture = NULL;
pipe_resource_reference(&sv->base.texture, texture);
sv->base.context = pipe;
sv->id = SVGA3D_INVALID_ID;
svga->hud.num_samplerview_objects++;
SVGA_STATS_COUNT_INC(svga_screen(svga->pipe.screen)->sws,
SVGA_STATS_COUNT_SAMPLERVIEW);
return &sv->base;
}
static void
svga_sampler_view_destroy(struct pipe_context *pipe,
struct pipe_sampler_view *view)
{
struct svga_context *svga = svga_context(pipe);
struct svga_pipe_sampler_view *sv = svga_pipe_sampler_view(view);
if (svga_have_vgpu10(svga) && sv->id != SVGA3D_INVALID_ID) {
if (view->context != pipe) {
/* The SVGA3D device will generate an error (and on Linux, cause
* us to abort) if we try to destroy a shader resource view from
* a context other than the one it was created with. Skip the
* SVGA3D_vgpu10_DestroyShaderResourceView() and leak the sampler
* view for now. This should only sometimes happen when a shared
* texture is deleted.
*/
_debug_printf("context mismatch in %s\n", __func__);
}
else {
enum pipe_error ret;
svga_hwtnl_flush_retry(svga); /* XXX is this needed? */
ret = SVGA3D_vgpu10_DestroyShaderResourceView(svga->swc, sv->id);
if (ret != PIPE_OK) {
svga_context_flush(svga, NULL);
ret = SVGA3D_vgpu10_DestroyShaderResourceView(svga->swc, sv->id);
}
util_bitmask_clear(svga->sampler_view_id_bm, sv->id);
}
}
pipe_resource_reference(&sv->base.texture, NULL);
FREE(sv);
svga->hud.num_samplerview_objects--;
}
static void
svga_set_sampler_views(struct pipe_context *pipe,
enum pipe_shader_type shader,
unsigned start,
unsigned num,
struct pipe_sampler_view **views)
{
struct svga_context *svga = svga_context(pipe);
unsigned flag_1d = 0;
unsigned flag_srgb = 0;
uint i;
boolean any_change = FALSE;
assert(shader < PIPE_SHADER_TYPES);
assert(start + num <= ARRAY_SIZE(svga->curr.sampler_views[shader]));
/* Pre-VGPU10 only supports FS textures */
if (!svga_have_vgpu10(svga) && shader != PIPE_SHADER_FRAGMENT)
return;
SVGA_STATS_TIME_PUSH(svga_sws(svga), SVGA_STATS_TIME_SETSAMPLERVIEWS);
/* This bit of code works around a quirk in the CSO module.
* If start=num=0 it means all sampler views should be released.
* Note that the CSO module treats sampler views for fragment shaders
* differently than other shader types.
*/
if (start == 0 && num == 0 && svga->curr.num_sampler_views[shader] > 0) {
for (i = 0; i < svga->curr.num_sampler_views[shader]; i++) {
pipe_sampler_view_release(pipe, &svga->curr.sampler_views[shader][i]);
}
any_change = TRUE;
}
for (i = 0; i < num; i++) {
enum pipe_texture_target target;
if (svga->curr.sampler_views[shader][start + i] != views[i]) {
/* Note: we're using pipe_sampler_view_release() here to work around
* a possible crash when the old view belongs to another context that
* was already destroyed.
*/
pipe_sampler_view_release(pipe, &svga->curr.sampler_views[shader][start + i]);
pipe_sampler_view_reference(&svga->curr.sampler_views[shader][start + i],
views[i]);
any_change = TRUE;
}
if (!views[i])
continue;
if (util_format_is_srgb(views[i]->format))
flag_srgb |= 1 << (start + i);
target = views[i]->target;
if (target == PIPE_TEXTURE_1D) {
flag_1d |= 1 << (start + i);
} else if (target == PIPE_TEXTURE_RECT) {
/* If the size of the bound texture changes, we need to emit new
* const buffer values.
*/
svga->dirty |= SVGA_NEW_TEXTURE_CONSTS;
} else if (target == PIPE_BUFFER) {
/* If the size of the bound buffer changes, we need to emit new
* const buffer values.
*/
svga->dirty |= SVGA_NEW_TEXTURE_CONSTS;
}
}
if (!any_change) {
goto done;
}
/* find highest non-null sampler_views[] entry */
{
unsigned j = MAX2(svga->curr.num_sampler_views[shader], start + num);
while (j > 0 && svga->curr.sampler_views[shader][j - 1] == NULL)
j--;
svga->curr.num_sampler_views[shader] = j;
}
svga->dirty |= SVGA_NEW_TEXTURE_BINDING;
if (flag_srgb != svga->curr.tex_flags.flag_srgb ||
flag_1d != svga->curr.tex_flags.flag_1d) {
svga->dirty |= SVGA_NEW_TEXTURE_FLAGS;
svga->curr.tex_flags.flag_1d = flag_1d;
svga->curr.tex_flags.flag_srgb = flag_srgb;
}
/* Check if any of the sampler view resources collide with the framebuffer
* color buffers or depth stencil resource. If so, set the NEW_FRAME_BUFFER
* dirty bit so that emit_framebuffer can be invoked to create backed view
* for the conflicted surface view.
*/
if (svga_check_sampler_framebuffer_resource_collision(svga, shader)) {
svga->dirty |= SVGA_NEW_FRAME_BUFFER;
}
done:
SVGA_STATS_TIME_POP(svga_sws(svga));
}
/**
* Clean up sampler, sampler view state at context destruction time
*/
void
svga_cleanup_sampler_state(struct svga_context *svga)
{
enum pipe_shader_type shader;
for (shader = 0; shader <= PIPE_SHADER_GEOMETRY; shader++) {
unsigned i;
for (i = 0; i < svga->state.hw_draw.num_sampler_views[shader]; i++) {
pipe_sampler_view_release(&svga->pipe,
&svga->state.hw_draw.sampler_views[shader][i]);
}
}
/* free polygon stipple state */
if (svga->polygon_stipple.sampler) {
svga->pipe.delete_sampler_state(&svga->pipe, svga->polygon_stipple.sampler);
}
if (svga->polygon_stipple.sampler_view) {
svga->pipe.sampler_view_destroy(&svga->pipe,
&svga->polygon_stipple.sampler_view->base);
}
pipe_resource_reference(&svga->polygon_stipple.texture, NULL);
}
void
svga_init_sampler_functions( struct svga_context *svga )
{
svga->pipe.create_sampler_state = svga_create_sampler_state;
svga->pipe.bind_sampler_states = svga_bind_sampler_states;
svga->pipe.delete_sampler_state = svga_delete_sampler_state;
svga->pipe.set_sampler_views = svga_set_sampler_views;
svga->pipe.create_sampler_view = svga_create_sampler_view;
svga->pipe.sampler_view_destroy = svga_sampler_view_destroy;
}