static void
fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
{
struct fd_resource *rsc = fd_resource(prsc);
if (pending(rsc, FD_PENDING_WRITE | FD_PENDING_READ))
fd_context_render(pctx);
}
static void
fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
{
struct fd_resource *rsc = fd_resource(prsc);
if (rsc->dirty)
fd_context_render(pctx);
}
static void fd_resource_transfer_flush_region(struct pipe_context *pctx,
struct pipe_transfer *ptrans,
const struct pipe_box *box)
{
struct fd_context *ctx = fd_context(pctx);
struct fd_resource *rsc = fd_resource(ptrans->resource);
if (rsc->dirty)
fd_context_render(pctx);
if (rsc->timestamp) {
fd_pipe_wait(ctx->screen->pipe, rsc->timestamp);
rsc->timestamp = 0;
}
}
static void
fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fence,
unsigned flags)
{
DBG("fence=%p", fence);
#if 0
if (fence) {
fd_fence_ref(ctx->screen->fence.current,
(struct fd_fence **)fence);
}
#endif
fd_context_render(pctx);
}
static void
fd_set_framebuffer_state(struct pipe_context *pctx,
const struct pipe_framebuffer_state *framebuffer)
{
struct fd_context *ctx = fd_context(pctx);
struct pipe_framebuffer_state *cso = &ctx->framebuffer;
unsigned i;
DBG("%d: cbufs[0]=%p, zsbuf=%p", ctx->needs_flush,
cso->cbufs[0], cso->zsbuf);
fd_context_render(pctx);
for (i = 0; i < framebuffer->nr_cbufs; i++)
pipe_surface_reference(&cso->cbufs[i], framebuffer->cbufs[i]);
for (; i < ctx->framebuffer.nr_cbufs; i++)
pipe_surface_reference(&cso->cbufs[i], NULL);
cso->nr_cbufs = framebuffer->nr_cbufs;
cso->width = framebuffer->width;
cso->height = framebuffer->height;
pipe_surface_reference(&cso->zsbuf, framebuffer->zsbuf);
ctx->dirty |= FD_DIRTY_FRAMEBUFFER;
/* also need to reset the scissor.. mesa/gl state tracker
* does this for us, but u_blitter doesn't and other
* state trackers might not..
*/
ctx->scissor.minx = 0;
ctx->scissor.miny = 0;
ctx->scissor.maxx = cso->width;
ctx->scissor.maxy = cso->height;
ctx->dirty |= FD_DIRTY_SCISSOR;
}
static void *
fd_resource_transfer_map(struct pipe_context *pctx,
struct pipe_resource *prsc,
unsigned level, unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **pptrans)
{
struct fd_context *ctx = fd_context(pctx);
struct fd_resource *rsc = fd_resource(prsc);
struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
struct fd_transfer *trans;
struct pipe_transfer *ptrans;
enum pipe_format format = prsc->format;
uint32_t op = 0;
uint32_t offset;
char *buf;
int ret = 0;
DBG("prsc=%p, level=%u, usage=%x, box=%dx%d+%d,%d", prsc, level, usage,
box->width, box->height, box->x, box->y);
ptrans = util_slab_alloc(&ctx->transfer_pool);
if (!ptrans)
return NULL;
/* util_slab_alloc() doesn't zero: */
trans = fd_transfer(ptrans);
memset(trans, 0, sizeof(*trans));
pipe_resource_reference(&ptrans->resource, prsc);
ptrans->level = level;
ptrans->usage = usage;
ptrans->box = *box;
ptrans->stride = util_format_get_nblocksx(format, slice->pitch) * rsc->cpp;
ptrans->layer_stride = rsc->layer_first ? rsc->layer_size : slice->size0;
if (usage & PIPE_TRANSFER_READ)
op |= DRM_FREEDRENO_PREP_READ;
if (usage & PIPE_TRANSFER_WRITE)
op |= DRM_FREEDRENO_PREP_WRITE;
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
realloc_bo(rsc, fd_bo_size(rsc->bo));
if (rsc->stencil)
realloc_bo(rsc->stencil, fd_bo_size(rsc->stencil->bo));
fd_invalidate_resource(ctx, prsc);
} else if ((usage & PIPE_TRANSFER_WRITE) &&
prsc->target == PIPE_BUFFER &&
!util_ranges_intersect(&rsc->valid_buffer_range,
box->x, box->x + box->width)) {
/* We are trying to write to a previously uninitialized range. No need
* to wait.
*/
} else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
/* If the GPU is writing to the resource, or if it is reading from the
* resource and we're trying to write to it, flush the renders.
*/
if (((ptrans->usage & PIPE_TRANSFER_WRITE) &&
pending(rsc, FD_PENDING_READ | FD_PENDING_WRITE)) ||
pending(rsc, FD_PENDING_WRITE))
fd_context_render(pctx);
/* The GPU keeps track of how the various bo's are being used, and
* will wait if necessary for the proper operation to have
* completed.
*/
ret = fd_bo_cpu_prep(rsc->bo, ctx->screen->pipe, op);
if (ret)
goto fail;
}
buf = fd_bo_map(rsc->bo);
if (!buf)
goto fail;
offset = slice->offset +
box->y / util_format_get_blockheight(format) * ptrans->stride +
box->x / util_format_get_blockwidth(format) * rsc->cpp +
fd_resource_layer_offset(rsc, slice, box->z);
if (prsc->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT ||
prsc->format == PIPE_FORMAT_X32_S8X24_UINT) {
assert(trans->base.box.depth == 1);
trans->base.stride = trans->base.box.width * rsc->cpp * 2;
trans->staging = malloc(trans->base.stride * trans->base.box.height);
if (!trans->staging)
goto fail;
/* if we're not discarding the whole range (or resource), we must copy
* the real data in.
*/
if (!(usage & (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
PIPE_TRANSFER_DISCARD_RANGE))) {
struct fd_resource_slice *sslice =
fd_resource_slice(rsc->stencil, level);
void *sbuf = fd_bo_map(rsc->stencil->bo);
if (!sbuf)
goto fail;
float *depth = (float *)(buf + slice->offset +
fd_resource_layer_offset(rsc, slice, box->z) +
box->y * slice->pitch * 4 + box->x * 4);
uint8_t *stencil = sbuf + sslice->offset +
fd_resource_layer_offset(rsc->stencil, sslice, box->z) +
box->y * sslice->pitch + box->x;
if (format != PIPE_FORMAT_X32_S8X24_UINT)
util_format_z32_float_s8x24_uint_pack_z_float(
trans->staging, trans->base.stride,
depth, slice->pitch * 4,
box->width, box->height);
util_format_z32_float_s8x24_uint_pack_s_8uint(
trans->staging, trans->base.stride,
stencil, sslice->pitch,
box->width, box->height);
}
buf = trans->staging;
offset = 0;
} else if (rsc->internal_format != format &&
util_format_description(format)->layout == UTIL_FORMAT_LAYOUT_RGTC) {
assert(trans->base.box.depth == 1);
trans->base.stride = util_format_get_stride(
format, trans->base.box.width);
trans->staging = malloc(
util_format_get_2d_size(format, trans->base.stride,
trans->base.box.height));
if (!trans->staging)
goto fail;
/* if we're not discarding the whole range (or resource), we must copy
* the real data in.
*/
if (!(usage & (PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE |
PIPE_TRANSFER_DISCARD_RANGE))) {
uint8_t *rgba8 = (uint8_t *)buf + slice->offset +
fd_resource_layer_offset(rsc, slice, box->z) +
box->y * slice->pitch * rsc->cpp + box->x * rsc->cpp;
switch (format) {
case PIPE_FORMAT_RGTC1_UNORM:
case PIPE_FORMAT_RGTC1_SNORM:
case PIPE_FORMAT_LATC1_UNORM:
case PIPE_FORMAT_LATC1_SNORM:
util_format_rgtc1_unorm_pack_rgba_8unorm(
trans->staging, trans->base.stride,
rgba8, slice->pitch * rsc->cpp,
box->width, box->height);
break;
case PIPE_FORMAT_RGTC2_UNORM:
case PIPE_FORMAT_RGTC2_SNORM:
case PIPE_FORMAT_LATC2_UNORM:
case PIPE_FORMAT_LATC2_SNORM:
util_format_rgtc2_unorm_pack_rgba_8unorm(
trans->staging, trans->base.stride,
rgba8, slice->pitch * rsc->cpp,
box->width, box->height);
break;
default:
assert(!"Unexpected format");
break;
}
}
buf = trans->staging;
offset = 0;
}
*pptrans = ptrans;
return buf + offset;
fail:
fd_resource_transfer_unmap(pctx, ptrans);
return NULL;
}
static void
fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
{
struct fd_context *ctx = fd_context(pctx);
struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
unsigned i, prims, buffers = 0;
/* if we supported transform feedback, we'd have to disable this: */
if (((scissor->maxx - scissor->minx) *
(scissor->maxy - scissor->miny)) == 0) {
return;
}
/* TODO: push down the region versions into the tiles */
if (!fd_render_condition_check(pctx))
return;
/* emulate unsupported primitives: */
if (!fd_supported_prim(ctx, info->mode)) {
if (ctx->streamout.num_targets > 0)
debug_error("stream-out with emulated prims");
util_primconvert_save_index_buffer(ctx->primconvert, &ctx->indexbuf);
util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
util_primconvert_draw_vbo(ctx->primconvert, info);
return;
}
ctx->needs_flush = true;
/*
* Figure out the buffers/features we need:
*/
if (fd_depth_enabled(ctx)) {
buffers |= FD_BUFFER_DEPTH;
resource_written(ctx, pfb->zsbuf->texture);
ctx->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
}
if (fd_stencil_enabled(ctx)) {
buffers |= FD_BUFFER_STENCIL;
resource_written(ctx, pfb->zsbuf->texture);
ctx->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
}
if (fd_logicop_enabled(ctx))
ctx->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
for (i = 0; i < pfb->nr_cbufs; i++) {
struct pipe_resource *surf;
if (!pfb->cbufs[i])
continue;
surf = pfb->cbufs[i]->texture;
resource_written(ctx, surf);
buffers |= PIPE_CLEAR_COLOR0 << i;
if (surf->nr_samples > 1)
ctx->gmem_reason |= FD_GMEM_MSAA_ENABLED;
if (fd_blend_enabled(ctx, i))
ctx->gmem_reason |= FD_GMEM_BLEND_ENABLED;
}
/* Skip over buffer 0, that is sent along with the command stream */
for (i = 1; i < PIPE_MAX_CONSTANT_BUFFERS; i++) {
resource_read(ctx, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
resource_read(ctx, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
}
/* Mark VBOs as being read */
for (i = 0; i < ctx->vtx.vertexbuf.count; i++) {
assert(!ctx->vtx.vertexbuf.vb[i].user_buffer);
resource_read(ctx, ctx->vtx.vertexbuf.vb[i].buffer);
}
/* Mark index buffer as being read */
resource_read(ctx, ctx->indexbuf.buffer);
/* Mark textures as being read */
for (i = 0; i < ctx->verttex.num_textures; i++)
if (ctx->verttex.textures[i])
resource_read(ctx, ctx->verttex.textures[i]->texture);
for (i = 0; i < ctx->fragtex.num_textures; i++)
if (ctx->fragtex.textures[i])
resource_read(ctx, ctx->fragtex.textures[i]->texture);
/* Mark streamout buffers as being written.. */
for (i = 0; i < ctx->streamout.num_targets; i++)
if (ctx->streamout.targets[i])
resource_written(ctx, ctx->streamout.targets[i]->buffer);
ctx->num_draws++;
prims = u_reduced_prims_for_vertices(info->mode, info->count);
ctx->stats.draw_calls++;
/* TODO prims_emitted should be clipped when the stream-out buffer is
* not large enough. See max_tf_vtx().. probably need to move that
* into common code. Although a bit more annoying since a2xx doesn't
* use ir3 so no common way to get at the pipe_stream_output_info
* which is needed for this calculation.
*/
if (ctx->streamout.num_targets > 0)
ctx->stats.prims_emitted += prims;
ctx->stats.prims_generated += prims;
/* any buffers that haven't been cleared yet, we need to restore: */
ctx->restore |= buffers & (FD_BUFFER_ALL & ~ctx->cleared);
/* and any buffers used, need to be resolved: */
ctx->resolve |= buffers;
DBG("%x num_draws=%u (%s/%s)", buffers, ctx->num_draws,
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_DRAW);
ctx->draw_vbo(ctx, info);
for (i = 0; i < ctx->streamout.num_targets; i++)
ctx->streamout.offsets[i] += info->count;
if (fd_mesa_debug & FD_DBG_DDRAW)
ctx->dirty = 0xffffffff;
/* if an app (or, well, piglit test) does many thousands of draws
* without flush (or anything which implicitly flushes, like
* changing render targets), we can exceed the ringbuffer size.
* Since we don't currently have a sane way to wrapparound, and
* we use the same buffer for both draw and tiling commands, for
* now we need to do this hack and trigger flush if we are running
* low on remaining space for cmds:
*/
if (((ctx->ring->cur - ctx->ring->start) >
(ctx->ring->size/4 - FD_TILING_COMMANDS_DWORDS)) ||
(fd_mesa_debug & FD_DBG_FLUSH))
fd_context_render(pctx);
}
static void *
fd_resource_transfer_map(struct pipe_context *pctx,
struct pipe_resource *prsc,
unsigned level, unsigned usage,
const struct pipe_box *box,
struct pipe_transfer **pptrans)
{
struct fd_context *ctx = fd_context(pctx);
struct fd_resource *rsc = fd_resource(prsc);
struct fd_resource_slice *slice = fd_resource_slice(rsc, level);
struct pipe_transfer *ptrans;
enum pipe_format format = prsc->format;
uint32_t op = 0;
uint32_t offset;
char *buf;
int ret = 0;
DBG("prsc=%p, level=%u, usage=%x", prsc, level, usage);
ptrans = util_slab_alloc(&ctx->transfer_pool);
if (!ptrans)
return NULL;
/* util_slab_alloc() doesn't zero: */
memset(ptrans, 0, sizeof(*ptrans));
pipe_resource_reference(&ptrans->resource, prsc);
ptrans->level = level;
ptrans->usage = usage;
ptrans->box = *box;
ptrans->stride = slice->pitch * rsc->cpp;
ptrans->layer_stride = slice->size0;
if (usage & PIPE_TRANSFER_READ)
op |= DRM_FREEDRENO_PREP_READ;
if (usage & PIPE_TRANSFER_WRITE)
op |= DRM_FREEDRENO_PREP_WRITE;
if (usage & PIPE_TRANSFER_DISCARD_WHOLE_RESOURCE) {
realloc_bo(rsc, fd_bo_size(rsc->bo));
fd_invalidate_resource(ctx, prsc);
} else if ((usage & PIPE_TRANSFER_WRITE) &&
prsc->target == PIPE_BUFFER &&
!util_ranges_intersect(&rsc->valid_buffer_range,
box->x, box->x + box->width)) {
/* We are trying to write to a previously uninitialized range. No need
* to wait.
*/
} else if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
/* If the GPU is writing to the resource, or if it is reading from the
* resource and we're trying to write to it, flush the renders.
*/
if (rsc->dirty ||
((ptrans->usage & PIPE_TRANSFER_WRITE) && rsc->reading))
fd_context_render(pctx);
/* The GPU keeps track of how the various bo's are being used, and
* will wait if necessary for the proper operation to have
* completed.
*/
ret = fd_bo_cpu_prep(rsc->bo, ctx->screen->pipe, op);
if (ret)
goto fail;
}
buf = fd_bo_map(rsc->bo);
if (!buf) {
fd_resource_transfer_unmap(pctx, ptrans);
return NULL;
}
*pptrans = ptrans;
if (rsc->layer_first) {
offset = slice->offset +
box->y / util_format_get_blockheight(format) * ptrans->stride +
box->x / util_format_get_blockwidth(format) * rsc->cpp +
box->z * rsc->layer_size;
} else {
offset = slice->offset +
box->y / util_format_get_blockheight(format) * ptrans->stride +
box->x / util_format_get_blockwidth(format) * rsc->cpp +
box->z * slice->size0;
}
return buf + offset;
fail:
fd_resource_transfer_unmap(pctx, ptrans);
return NULL;
}
static void
fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
{
struct fd_context *ctx = fd_context(pctx);
struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
unsigned i, buffers = 0;
/* if we supported transform feedback, we'd have to disable this: */
if (((scissor->maxx - scissor->minx) *
(scissor->maxy - scissor->miny)) == 0) {
return;
}
/* emulate unsupported primitives: */
if (!fd_supported_prim(ctx, info->mode)) {
util_primconvert_save_index_buffer(ctx->primconvert, &ctx->indexbuf);
util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
util_primconvert_draw_vbo(ctx->primconvert, info);
return;
}
ctx->needs_flush = true;
/*
* Figure out the buffers/features we need:
*/
if (fd_depth_enabled(ctx)) {
buffers |= FD_BUFFER_DEPTH;
fd_resource(pfb->zsbuf->texture)->dirty = true;
ctx->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
}
if (fd_stencil_enabled(ctx)) {
buffers |= FD_BUFFER_STENCIL;
fd_resource(pfb->zsbuf->texture)->dirty = true;
ctx->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
}
if (fd_logicop_enabled(ctx))
ctx->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
for (i = 0; i < pfb->nr_cbufs; i++) {
struct pipe_resource *surf;
if (!pfb->cbufs[i])
continue;
surf = pfb->cbufs[i]->texture;
fd_resource(surf)->dirty = true;
buffers |= FD_BUFFER_COLOR;
if (surf->nr_samples > 1)
ctx->gmem_reason |= FD_GMEM_MSAA_ENABLED;
if (fd_blend_enabled(ctx, i))
ctx->gmem_reason |= FD_GMEM_BLEND_ENABLED;
}
ctx->num_draws++;
ctx->stats.draw_calls++;
ctx->stats.prims_emitted +=
u_reduced_prims_for_vertices(info->mode, info->count);
/* any buffers that haven't been cleared yet, we need to restore: */
ctx->restore |= buffers & (FD_BUFFER_ALL & ~ctx->cleared);
/* and any buffers used, need to be resolved: */
ctx->resolve |= buffers;
DBG("%x num_draws=%u (%s/%s)", buffers, ctx->num_draws,
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_DRAW);
ctx->draw_vbo(ctx, info);
/* if an app (or, well, piglit test) does many thousands of draws
* without flush (or anything which implicitly flushes, like
* changing render targets), we can exceed the ringbuffer size.
* Since we don't currently have a sane way to wrapparound, and
* we use the same buffer for both draw and tiling commands, for
* now we need to do this hack and trigger flush if we are running
* low on remaining space for cmds:
*/
if (((ctx->ring->cur - ctx->ring->start) >
(ctx->ring->size/4 - FD_TILING_COMMANDS_DWORDS)) ||
(fd_mesa_debug & FD_DBG_FLUSH))
fd_context_render(pctx);
}
