rt->format = panfrost_mfbd_format(surf);
+ if (layer_stride)
+ rt->format.flags |= MALI_MFBD_FORMAT_MSAA | MALI_MFBD_FORMAT_LAYERED;
+
/* Now, we set the layout specific pieces */
if (rsrc->layout == MALI_TEXTURE_LINEAR) {
bool is_bifrost = dev->quirks & IS_BIFROST;
struct panfrost_resource *rsrc = pan_resource(surf->texture);
+ unsigned nr_samples = surf->nr_samples;
+
+ if (!nr_samples)
+ nr_samples = surf->texture->nr_samples;
+
+ nr_samples = MAX2(nr_samples, 1);
+
+ fbx->zs_samples = MALI_POSITIVE(nr_samples);
+
unsigned level = surf->u.tex.level;
unsigned first_layer = surf->u.tex.first_layer;
assert(surf->u.tex.last_layer == first_layer);
unsigned header_size = rsrc->slices[level].header_size;
- fb->mfbd_flags |= MALI_MFBD_EXTRA;
+ fb->mfbd_flags |= MALI_MFBD_EXTRA | MALI_MFBD_DEPTH_WRITE;
fbx->flags_hi |= MALI_EXTRA_PRESENT;
fbx->flags_lo |= MALI_EXTRA_ZS | 0x1; /* unknown */
int stride = rsrc->slices[level].stride;
- fb->mfbd_flags |= MALI_MFBD_EXTRA;
+ unsigned layer_stride = (nr_samples > 1) ? rsrc->slices[level].size0 : 0;
+
+ fb->mfbd_flags |= MALI_MFBD_EXTRA | MALI_MFBD_DEPTH_WRITE;
fbx->flags_hi |= MALI_EXTRA_PRESENT;
fbx->flags_lo |= MALI_EXTRA_ZS;
if (rsrc->layout == MALI_TEXTURE_LINEAR) {
fbx->zs_block = MALI_BLOCK_LINEAR;
fbx->ds_linear.depth_stride = stride / 16;
+ fbx->ds_linear.depth_layer_stride = layer_stride;
} else {
if (is_bifrost) {
fbx->zs_block = MALI_BLOCK_UNKNOWN;
- fbx->flags_hi |= 0x4400;
+ fbx->flags_hi |= 0x440;
fbx->flags_lo |= 0x1;
} else {
fbx->zs_block = MALI_BLOCK_TILED;
}
fbx->ds_linear.depth_stride = stride;
+ fbx->ds_linear.depth_layer_stride = layer_stride;
}
if (panfrost_is_z24s8_variant(surf->format)) {
fb->mfbd_flags ^= 0x100;
fb->mfbd_flags |= 0x200;
} else if (surf->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT) {
- fbx->flags_hi |= 0x400;
+ fbx->flags_hi |= 0x40;
fbx->flags_lo |= 0xA;
fb->mfbd_flags ^= 0x100;
fb->mfbd_flags |= 0x201;
struct panfrost_resource *stencil = rsrc->separate_stencil;
struct panfrost_slice stencil_slice = stencil->slices[level];
+ unsigned stencil_layer_stride = (nr_samples > 1) ? stencil_slice.size0 : 0;
fbx->ds_linear.stencil = panfrost_get_texture_address(stencil, level, first_layer, 0);
fbx->ds_linear.stencil_stride = stencil_slice.stride;
+ fbx->ds_linear.stencil_layer_stride = stencil_layer_stride;
}
} else {
sizeof(struct mali_render_target) * 4;
struct panfrost_transfer m_f_trans =
- panfrost_allocate_transient(batch, total_sz);
+ panfrost_pool_alloc(&batch->pool, total_sz);
/* Do the transfer */
#undef UPLOAD
+/* Determines the # of bytes per pixel we need to reserve for a given format in
+ * the tilebuffer (compared to 128-bit budget, etc). Usually the same as the
+ * bytes per pixel of the format itself, but there are some special cases I
+ * don't understand. */
+
+static unsigned
+pan_bytes_per_pixel_tib(enum pipe_format format)
+{
+ const struct util_format_description *desc =
+ util_format_description(format);
+
+ if (util_format_is_unorm8(desc) || format == PIPE_FORMAT_B5G6R5_UNORM)
+ return 4;
+
+ return desc->block.bits / 8;
+}
+
/* Determines whether a framebuffer uses too much tilebuffer space (requiring
* us to scale up the tile at a performance penalty). This is conservative but
* afaict you get 128-bits per pixel normally */
for (int cb = 0; cb < batch->key.nr_cbufs; ++cb) {
struct pipe_surface *surf = batch->key.cbufs[cb];
assert(surf);
- unsigned bpp = util_format_get_blocksize(surf->format);
- size += ALIGN_POT(bpp, 4);
+ size += pan_bytes_per_pixel_tib(surf->format);
}
return (size > 16);
unsigned rt_descriptors = MAX2(batch->key.nr_cbufs, 1);
fb.rt_count_1 = MALI_POSITIVE(rt_descriptors);
- fb.rt_count_2 = rt_descriptors;
fb.mfbd_flags = 0x100;
- /* TODO: MRT clear */
- panfrost_mfbd_clear(batch, &fb, &fbx, rts, fb.rt_count_2);
-
+ panfrost_mfbd_clear(batch, &fb, &fbx, rts, rt_descriptors);
/* Upload either the render target or a dummy GL_NONE target */
+ unsigned offset = 0;
+ bool is_large = pan_is_large_tib(batch);
+
for (int cb = 0; cb < rt_descriptors; ++cb) {
struct pipe_surface *surf = batch->key.cbufs[cb];
+ unsigned rt_offset = offset * 0x100;
- if (surf) {
- panfrost_mfbd_set_cbuf(&rts[cb], surf);
+ if (surf && ((batch->clear | batch->draws) & (PIPE_CLEAR_COLOR0 << cb))) {
+ unsigned nr_samples = surf->nr_samples;
+
+ if (!nr_samples)
+ nr_samples = surf->texture->nr_samples;
- /* What is this? Looks like some extension of the bpp
- * field. Maybe it establishes how much internal
- * tilebuffer space is reserved? */
+ if (nr_samples > 1)
+ batch->requirements |= PAN_REQ_MSAA;
- unsigned bpp = util_format_get_blocksize(surf->format);
- fb.rt_count_2 = MAX2(fb.rt_count_2, ALIGN_POT(bpp, 4) / 4);
+ panfrost_mfbd_set_cbuf(&rts[cb], surf);
+
+ offset += pan_bytes_per_pixel_tib(surf->format);
} else {
struct mali_rt_format null_rt = {
.unk1 = 0x4000000,
}
/* TODO: Break out the field */
- rts[cb].format.unk1 |= (cb * 0x400);
+ rts[cb].format.unk1 |= is_large ? (rt_offset / 4) : rt_offset;
}
- if (batch->key.zsbuf) {
+ fb.rt_count_2 = MAX2(DIV_ROUND_UP(offset, is_large ? 16 : 4), 1);
+
+ if (batch->key.zsbuf && ((batch->clear | batch->draws) & PIPE_CLEAR_DEPTHSTENCIL)) {
panfrost_mfbd_set_zsbuf(&fb, &fbx, batch->key.zsbuf);
}