/**
* Allocates a block of space in the batchbuffer for indirect state.
*
* We don't want to allocate separate BOs for every bit of indirect
* state in the driver. It means overallocating by a significant
* margin (4096 bytes, even if the object is just a 20-byte surface
* state), and more buffers to walk and count for aperture size checking.
*
* However, due to the restrictions imposed by the aperture size
* checking performance hacks, we can't have the batch point at a
* separate indirect state buffer, because once the batch points at
* it, no more relocations can be added to it. So, we sneak these
* buffers in at the top of the batchbuffer.
*/
void *
brw_state_batch(struct brw_context *brw,
int size,
int alignment,
uint32_t *out_offset)
{
struct intel_batchbuffer *batch = &brw->batch;
uint32_t offset;
assert(size < batch->bo->size);
offset = ROUND_DOWN_TO(batch->state_batch_offset - size, alignment);
/* If allocating from the top would wrap below the batchbuffer, or
* if the batch's used space (plus the reserved pad) collides with our
* space, then flush and try again.
*/
if (batch->state_batch_offset < size ||
offset < 4 * USED_BATCH(*batch) + batch->reserved_space) {
intel_batchbuffer_flush(brw);
offset = ROUND_DOWN_TO(batch->state_batch_offset - size, alignment);
}
batch->state_batch_offset = offset;
if (unlikely(INTEL_DEBUG & DEBUG_BATCH)) {
_mesa_hash_table_insert(batch->state_batch_sizes,
(void *) (uintptr_t) offset,
(void *) (uintptr_t) size);
}
*out_offset = offset;
return batch->map + (offset>>2);
}
/**
* Allocates a block of space in the batchbuffer for indirect state.
*
* We don't want to allocate separate BOs for every bit of indirect
* state in the driver. It means overallocating by a significant
* margin (4096 bytes, even if the object is just a 20-byte surface
* state), and more buffers to walk and count for aperture size checking.
*
* However, due to the restrictions inposed by the aperture size
* checking performance hacks, we can't have the batch point at a
* separate indirect state buffer, because once the batch points at
* it, no more relocations can be added to it. So, we sneak these
* buffers in at the top of the batchbuffer.
*/
void *
brw_state_batch(struct brw_context *brw,
enum state_struct_type type,
int size,
int alignment,
uint32_t *out_offset)
{
struct intel_batchbuffer *batch = &brw->intel.batch;
uint32_t offset;
assert(size < batch->bo->size);
offset = ROUND_DOWN_TO(batch->state_batch_offset - size, alignment);
/* If allocating from the top would wrap below the batchbuffer, or
* if the batch's used space (plus the reserved pad) collides with our
* space, then flush and try again.
*/
if (batch->state_batch_offset < size ||
offset < 4*batch->used + batch->reserved_space) {
intel_batchbuffer_flush(&brw->intel);
offset = ROUND_DOWN_TO(batch->state_batch_offset - size, alignment);
}
batch->state_batch_offset = offset;
if (unlikely(INTEL_DEBUG & (DEBUG_BATCH | DEBUG_AUB)))
brw_track_state_batch(brw, type, offset, size);
*out_offset = offset;
return batch->map + (offset>>2);
}
/**
* Allocates a block of space in the batchbuffer for indirect state.
*
* We don't want to allocate separate BOs for every bit of indirect
* state in the driver. It means overallocating by a significant
* margin (4096 bytes, even if the object is just a 20-byte surface
* state), and more buffers to walk and count for aperture size checking.
*
* However, due to the restrictions inposed by the aperture size
* checking performance hacks, we can't have the batch point at a
* separate indirect state buffer, because once the batch points at
* it, no more relocations can be added to it. So, we sneak these
* buffers in at the top of the batchbuffer.
*/
void *
brw_state_batch(struct brw_context *brw,
int size,
int alignment,
drm_intel_bo **out_bo,
uint32_t *out_offset)
{
struct intel_batchbuffer *batch = brw->intel.batch;
uint32_t offset;
assert(size < batch->buf->size);
offset = ROUND_DOWN_TO(batch->state_batch_offset - size, alignment);
/* If allocating from the top would wrap below the batchbuffer, or
* if the batch's used space (plus the reserved pad) collides with our
* space, then flush and try again.
*/
if (batch->state_batch_offset < size ||
offset < batch->ptr - batch->map + batch->reserved_space) {
intel_batchbuffer_flush(batch);
offset = ROUND_DOWN_TO(batch->state_batch_offset - size, alignment);
}
batch->state_batch_offset = offset;
if (*out_bo != batch->buf) {
drm_intel_bo_unreference(*out_bo);
drm_intel_bo_reference(batch->buf);
*out_bo = batch->buf;
}
*out_offset = offset;
return batch->map + offset;
}
/**
* When the GS is not in use, we assign the entire URB space to the VS. When
* the GS is in use, we split the URB space evenly between the VS and the GS.
* This is not ideal, but it's simple.
*
* URB size / 2 URB size / 2
* _____________-______________ _____________-______________
* / \ / \
* +-------------------------------------------------------------+
* | Vertex Shader Entries | Geometry Shader Entries |
* +-------------------------------------------------------------+
*
* Sandybridge GT1 has 32kB of URB space, while GT2 has 64kB.
* (See the Sandybridge PRM, Volume 2, Part 1, Section 1.4.7: 3DSTATE_URB.)
*/
void
gen6_upload_urb(struct brw_context *brw, unsigned vs_size,
bool gs_present, unsigned gs_size)
{
int nr_vs_entries, nr_gs_entries;
int total_urb_size = brw->urb.size * 1024; /* in bytes */
const struct gen_device_info *devinfo = &brw->screen->devinfo;
/* Calculate how many entries fit in each stage's section of the URB */
if (gs_present) {
nr_vs_entries = (total_urb_size/2) / (vs_size * 128);
nr_gs_entries = (total_urb_size/2) / (gs_size * 128);
} else {
nr_vs_entries = total_urb_size / (vs_size * 128);
nr_gs_entries = 0;
}
/* Then clamp to the maximum allowed by the hardware */
if (nr_vs_entries > devinfo->urb.max_vs_entries)
nr_vs_entries = devinfo->urb.max_vs_entries;
if (nr_gs_entries > devinfo->urb.max_gs_entries)
nr_gs_entries = devinfo->urb.max_gs_entries;
/* Finally, both must be a multiple of 4 (see 3DSTATE_URB in the PRM). */
brw->urb.nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, 4);
brw->urb.nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, 4);
assert(brw->urb.nr_vs_entries >= devinfo->urb.min_vs_entries);
assert(brw->urb.nr_vs_entries % 4 == 0);
assert(brw->urb.nr_gs_entries % 4 == 0);
assert(vs_size <= 5);
assert(gs_size <= 5);
BEGIN_BATCH(3);
OUT_BATCH(_3DSTATE_URB << 16 | (3 - 2));
OUT_BATCH(((vs_size - 1) << GEN6_URB_VS_SIZE_SHIFT) |
((brw->urb.nr_vs_entries) << GEN6_URB_VS_ENTRIES_SHIFT));
OUT_BATCH(((gs_size - 1) << GEN6_URB_GS_SIZE_SHIFT) |
((brw->urb.nr_gs_entries) << GEN6_URB_GS_ENTRIES_SHIFT));
ADVANCE_BATCH();
/* From the PRM Volume 2 part 1, section 1.4.7:
*
* Because of a urb corruption caused by allocating a previous gsunit’s
* urb entry to vsunit software is required to send a "GS NULL
* Fence"(Send URB fence with VS URB size == 1 and GS URB size == 0) plus
* a dummy DRAW call before any case where VS will be taking over GS URB
* space.
*
* It is not clear exactly what this means ("URB fence" is a command that
* doesn't exist on Gen6). So for now we just do a full pipeline flush as
* a workaround.
*/
if (brw->urb.gs_present && !gs_present)
brw_emit_mi_flush(brw);
brw->urb.gs_present = gs_present;
}
/**
* Stencil is mapped as Y-tiled render target and the dimensions need to be
* adjusted in order for the Y-tiled rectangle to cover the entire linear
* memory space of the original W-tiled rectangle.
*/
static void
adjust_tiling(struct blit_dims *dims, int num_samples)
{
const unsigned x_align = 8, y_align = num_samples > 2 ? 8 : 4;
dims->dst_x0 = ROUND_DOWN_TO(dims->dst_x0, x_align) * 2;
dims->dst_y0 = ROUND_DOWN_TO(dims->dst_y0, y_align) / 2;
dims->dst_x1 = ALIGN(dims->dst_x1, x_align) * 2;
dims->dst_y1 = ALIGN(dims->dst_y1, y_align) / 2;
}
/**
* Samples in stencil buffer are interleaved, and unfortunately the data port
* does not support it as render target. Therefore the surface is set up as
* single sampled and the program handles the interleaving.
* In case of single sampled stencil, the render buffer is adjusted with
* twice the base level height in order for the program to be able to write
* any mip-level. (Used to set the drawing rectangle for the hw).
*/
static void
adjust_msaa(struct blit_dims *dims, int num_samples)
{
if (num_samples == 2) {
dims->dst_x0 *= 2;
dims->dst_x1 *= 2;
} else if (num_samples) {
const int x_num_samples = num_samples / 2;
dims->dst_x0 = ROUND_DOWN_TO(dims->dst_x0 * x_num_samples, num_samples);
dims->dst_y0 = ROUND_DOWN_TO(dims->dst_y0 * 2, 4);
dims->dst_x1 = ALIGN(dims->dst_x1 * x_num_samples, num_samples);
dims->dst_y1 = ALIGN(dims->dst_y1 * 2, 4);
}
}
static void
gen7_upload_urb(struct brw_context *brw)
{
struct intel_context *intel = &brw->intel;
const int push_size_kB = intel->is_haswell && intel->gt == 3 ? 32 : 16;
/* Total space for entries is URB size - 16kB for push constants */
int handle_region_size = (brw->urb.size - push_size_kB) * 1024; /* bytes */
/* CACHE_NEW_VS_PROG */
unsigned vs_size = MAX2(brw->vs.prog_data->urb_entry_size, 1);
int nr_vs_entries = handle_region_size / (vs_size * 64);
if (nr_vs_entries > brw->urb.max_vs_entries)
nr_vs_entries = brw->urb.max_vs_entries;
/* According to volume 2a, nr_vs_entries must be a multiple of 8. */
brw->urb.nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, 8);
/* URB Starting Addresses are specified in multiples of 8kB. */
brw->urb.vs_start = push_size_kB / 8; /* skip over push constants */
assert(brw->urb.nr_vs_entries % 8 == 0);
assert(brw->urb.nr_gs_entries % 8 == 0);
/* GS requirement */
assert(!brw->gs.prog_active);
gen7_emit_vs_workaround_flush(intel);
gen7_emit_urb_state(brw, brw->urb.nr_vs_entries, vs_size, brw->urb.vs_start);
}
/* We don't have a memmove-type blit like some other hardware, so we'll do a
* rectangular blit covering a large space, then emit 1-scanline blit at the
* end to cover the last if we need.
*/
void
intel_emit_linear_blit(struct intel_context *intel,
drm_intel_bo *dst_bo,
unsigned int dst_offset,
drm_intel_bo *src_bo,
unsigned int src_offset,
unsigned int size)
{
struct gl_context *ctx = &intel->ctx;
GLuint pitch, height;
bool ok;
/* The pitch given to the GPU must be DWORD aligned, and
* we want width to match pitch. Max width is (1 << 15 - 1),
* rounding that down to the nearest DWORD is 1 << 15 - 4
*/
pitch = ROUND_DOWN_TO(MIN2(size, (1 << 15) - 1), 4);
height = (pitch == 0) ? 1 : size / pitch;
ok = intelEmitCopyBlit(intel, 1,
pitch, src_bo, src_offset, I915_TILING_NONE,
pitch, dst_bo, dst_offset, I915_TILING_NONE,
0, 0, /* src x/y */
0, 0, /* dst x/y */
pitch, height, /* w, h */
GL_COPY);
if (!ok)
_mesa_problem(ctx, "Failed to linear blit %dx%d\n", pitch, height);
src_offset += pitch * height;
dst_offset += pitch * height;
size -= pitch * height;
assert (size < (1 << 15));
pitch = ALIGN(size, 4);
if (size != 0) {
ok = intelEmitCopyBlit(intel, 1,
pitch, src_bo, src_offset, I915_TILING_NONE,
pitch, dst_bo, dst_offset, I915_TILING_NONE,
0, 0, /* src x/y */
0, 0, /* dst x/y */
size, 1, /* w, h */
GL_COPY);
if (!ok)
_mesa_problem(ctx, "Failed to linear blit %dx%d\n", size, 1);
}
}
static void
get_fast_clear_rect(struct brw_context *brw, struct gl_framebuffer *fb,
struct intel_renderbuffer *irb, struct rect *rect)
{
unsigned int x_align, y_align;
unsigned int x_scaledown, y_scaledown;
if (irb->mt->msaa_layout == INTEL_MSAA_LAYOUT_NONE) {
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "Fast Color Clear" bullet (p327):
*
* Clear pass must have a clear rectangle that must follow
* alignment rules in terms of pixels and lines as shown in the
* table below. Further, the clear-rectangle height and width
* must be multiple of the following dimensions. If the height
* and width of the render target being cleared do not meet these
* requirements, an MCS buffer can be created such that it
* follows the requirement and covers the RT.
*
* The alignment size in the table that follows is related to the
* alignment size returned by intel_get_non_msrt_mcs_alignment(), but
* with X alignment multiplied by 16 and Y alignment multiplied by 32.
*/
intel_get_non_msrt_mcs_alignment(brw, irb->mt, &x_align, &y_align);
x_align *= 16;
y_align *= 32;
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "Fast Color Clear" bullet (p327):
*
* In order to optimize the performance MCS buffer (when bound to
* 1X RT) clear similarly to MCS buffer clear for MSRT case,
* clear rect is required to be scaled by the following factors
* in the horizontal and vertical directions:
*
* The X and Y scale down factors in the table that follows are each
* equal to half the alignment value computed above.
*/
x_scaledown = x_align / 2;
y_scaledown = y_align / 2;
/* From BSpec: 3D-Media-GPGPU Engine > 3D Pipeline > Pixel > Pixel
* Backend > MCS Buffer for Render Target(s) [DevIVB+] > Table "Color
* Clear of Non-MultiSampled Render Target Restrictions":
*
* Clear rectangle must be aligned to two times the number of
* pixels in the table shown below due to 16x16 hashing across the
* slice.
*/
x_align *= 2;
y_align *= 2;
} else {
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "MSAA Compression" bullet (p326):
*
* Clear pass for this case requires that scaled down primitive
* is sent down with upper left co-ordinate to coincide with
* actual rectangle being cleared. For MSAA, clear rectangle’s
* height and width need to as show in the following table in
* terms of (width,height) of the RT.
*
* MSAA Width of Clear Rect Height of Clear Rect
* 4X Ceil(1/8*width) Ceil(1/2*height)
* 8X Ceil(1/2*width) Ceil(1/2*height)
*
* The text "with upper left co-ordinate to coincide with actual
* rectangle being cleared" is a little confusing--it seems to imply
* that to clear a rectangle from (x,y) to (x+w,y+h), one needs to
* feed the pipeline using the rectangle (x,y) to
* (x+Ceil(w/N),y+Ceil(h/2)), where N is either 2 or 8 depending on
* the number of samples. Experiments indicate that this is not
* quite correct; actually, what the hardware appears to do is to
* align whatever rectangle is sent down the pipeline to the nearest
* multiple of 2x2 blocks, and then scale it up by a factor of N
* horizontally and 2 vertically. So the resulting alignment is 4
* vertically and either 4 or 16 horizontally, and the scaledown
* factor is 2 vertically and either 2 or 8 horizontally.
*/
switch (irb->mt->num_samples) {
case 2:
case 4:
x_scaledown = 8;
break;
case 8:
x_scaledown = 2;
break;
default:
unreachable("Unexpected sample count for fast clear");
}
y_scaledown = 2;
x_align = x_scaledown * 2;
y_align = y_scaledown * 2;
}
rect->x0 = fb->_Xmin;
rect->x1 = fb->_Xmax;
if (fb->Name != 0) {
rect->y0 = fb->_Ymin;
rect->y1 = fb->_Ymax;
} else {
rect->y0 = fb->Height - fb->_Ymax;
rect->y1 = fb->Height - fb->_Ymin;
}
rect->x0 = ROUND_DOWN_TO(rect->x0, x_align) / x_scaledown;
rect->y0 = ROUND_DOWN_TO(rect->y0, y_align) / y_scaledown;
rect->x1 = ALIGN(rect->x1, x_align) / x_scaledown;
rect->y1 = ALIGN(rect->y1, y_align) / y_scaledown;
}
/*
* Allocate memory from the given pool. Grow the pool if needed and if
* possible.
*/
static unsigned long
AllocFromPool(ScrnInfoPtr pScrn, I830MemRange *result, I830MemPool *pool,
unsigned long size, unsigned long alignment, int flags)
{
I830Ptr pI830 = I830PTR(pScrn);
unsigned long needed, start, end;
Bool dryrun = ((flags & ALLOCATE_DRY_RUN) != 0);
if (!result || !pool || !size)
return 0;
/* Calculate how much space is needed. */
if (alignment <= GTT_PAGE_SIZE)
needed = size;
else {
if (flags & ALLOCATE_AT_BOTTOM) {
start = ROUND_TO(pool->Free.Start, alignment);
if (flags & ALIGN_BOTH_ENDS)
end = ROUND_TO(start + size, alignment);
else
end = start + size;
needed = end - pool->Free.Start;
} else { /* allocate at top */
if (flags & ALIGN_BOTH_ENDS)
end = ROUND_DOWN_TO(pool->Free.End, alignment);
else
end = pool->Free.End;
start = ROUND_DOWN_TO(end - size, alignment);
needed = pool->Free.End - start;
}
}
if (needed > pool->Free.Size) {
unsigned long extra;
/* See if the pool can be grown. */
if (pI830->StolenOnly && !dryrun)
return 0;
extra = needed - pool->Free.Size;
extra = ROUND_TO_PAGE(extra);
if (extra > pI830->FreeMemory) {
if (dryrun)
pI830->FreeMemory = extra;
else
return 0;
}
if (!dryrun && (extra > pI830->MemoryAperture.Size))
return 0;
pool->Free.Size += extra;
pool->Free.End += extra;
pool->Total.Size += extra;
pool->Total.End += extra;
pI830->FreeMemory -= extra;
pI830->MemoryAperture.Start += extra;
pI830->MemoryAperture.Size -= extra;
}
if (flags & ALLOCATE_AT_BOTTOM) {
result->Start = ROUND_TO(pool->Free.Start, alignment);
pool->Free.Start += needed;
result->End = pool->Free.Start;
} else {
result->Start = ROUND_DOWN_TO(pool->Free.End - size, alignment) -
pool->Total.End;
result->End = pool->Free.End - pool->Total.End;
pool->Free.End -= needed;
}
pool->Free.Size = pool->Free.End - pool->Free.Start;
result->Size = result->End - result->Start;
result->Pool = pool;
result->Alignment = alignment;
return needed;
}
Bool
I830Allocate2DMemory(ScrnInfoPtr pScrn, const int flags)
{
I830Ptr pI830 = I830PTR(pScrn);
unsigned long size, alloced;
Bool dryrun = ((flags & ALLOCATE_DRY_RUN) != 0);
int verbosity = dryrun ? 4 : 1;
const char *s = dryrun ? "[dryrun] " : "";
Bool tileable;
int align, alignflags;
DPRINTF(PFX, "I830Allocate2DMemory: inital is %s\n",
BOOLTOSTRING(flags & ALLOC_INITIAL));
if (!pI830->StolenOnly &&
(!xf86AgpGARTSupported() || !xf86AcquireGART(pScrn->scrnIndex))) {
if (!dryrun) {
xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
"AGP GART support is either not available or cannot "
"be used.\n"
"\tMake sure your kernel has agpgart support or has the\n"
"\tagpgart module loaded.\n");
}
return FALSE;
}
/*
* The I830 is slightly different from the I830/I815, it has no
* dcache and it has stolen memory by default in its gtt. All
* additional memory must go after it.
*/
DPRINTF(PFX,
"size == %luk (%lu bytes == pScrn->videoRam)\n"
"pI830->StolenSize == %luk (%lu bytes)\n",
pScrn->videoRam, pScrn->videoRam * 1024,
pI830->StolenPool.Free.Size / 1024,
pI830->StolenPool.Free.Size);
if (flags & ALLOC_INITIAL) {
unsigned long minspace, avail, lineSize;
int cacheLines, maxCacheLines;
if (pI830->NeedRingBufferLow)
AllocateRingBuffer(pScrn, flags | FORCE_LOW);
/* Clear everything first. */
memset(&(pI830->FbMemBox), 0, sizeof(pI830->FbMemBox));
memset(&(pI830->FrontBuffer), 0, sizeof(pI830->FrontBuffer));
pI830->FrontBuffer.Key = -1;
pI830->FbMemBox.x1 = 0;
pI830->FbMemBox.x2 = pScrn->displayWidth;
pI830->FbMemBox.y1 = 0;
pI830->FbMemBox.y2 = pScrn->virtualY;
/*
* Calculate how much framebuffer memory to allocate. For the
* initial allocation, calculate a reasonable minimum. This is
* enough for the virtual screen size, plus some pixmap cache
* space.
*/
lineSize = pScrn->displayWidth * pI830->cpp;
minspace = lineSize * pScrn->virtualY;
avail = pScrn->videoRam * 1024;
maxCacheLines = (avail - minspace) / lineSize;
/* This shouldn't happen. */
if (maxCacheLines < 0) {
xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
"Internal Error: "
"maxCacheLines < 0 in I830Allocate2DMemory()\n");
maxCacheLines = 0;
}
if (maxCacheLines > (MAX_DISPLAY_HEIGHT - pScrn->virtualY))
maxCacheLines = MAX_DISPLAY_HEIGHT - pScrn->virtualY;
if (pI830->CacheLines >= 0) {
cacheLines = pI830->CacheLines;
} else {
#if 1
/* Make sure there is enough for two DVD sized YUV buffers */
cacheLines = (pScrn->depth == 24) ? 256 : 384;
if (pScrn->displayWidth <= 1024)
cacheLines *= 2;
#else
/*
* Make sure there is enough for two DVD sized YUV buffers.
* Make that 1.5MB, which is around what was allocated with
* the old algorithm
*/
cacheLines = (MB(1) + KB(512)) / pI830->cpp / pScrn->displayWidth;
#endif
}
if (cacheLines > maxCacheLines)
cacheLines = maxCacheLines;
pI830->FbMemBox.y2 += cacheLines;
xf86DrvMsgVerb(pScrn->scrnIndex, X_INFO, verbosity,
"%sAllocating at least %d scanlines for pixmap cache\n",
s, cacheLines);
tileable = !(flags & ALLOC_NO_TILING) && pI830->allowPageFlip &&
IsTileable(pScrn->displayWidth * pI830->cpp);
if (tileable) {
align = KB(512);
alignflags = ALIGN_BOTH_ENDS;
} else {
align = KB(64);
alignflags = 0;
}
size = lineSize * (pScrn->virtualY + cacheLines);
size = ROUND_TO_PAGE(size);
xf86DrvMsgVerb(pScrn->scrnIndex, X_INFO, verbosity,
"%sInitial framebuffer allocation size: %d kByte\n", s,
size / 1024);
alloced = I830AllocVidMem(pScrn, &(pI830->FrontBuffer),
&(pI830->StolenPool), size, align,
flags | alignflags |
FROM_ANYWHERE | ALLOCATE_AT_BOTTOM);
if (alloced < size) {
if (!dryrun) {
xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
"Failed to allocate framebuffer.\n");
}
return FALSE;
}
} else {
unsigned long lineSize;
unsigned long extra = 0;
unsigned long maxFb = 0;
/*
* XXX Need to "free" up any 3D allocations if the DRI ended up
* and make them available for 2D. The best way to do this would
* be position all of those regions contiguously at the end of the
* StolenPool.
*/
extra = GetFreeSpace(pScrn);
if (extra == 0)
return TRUE;
maxFb = pI830->FrontBuffer.Size + extra;
lineSize = pScrn->displayWidth * pI830->cpp;
maxFb = ROUND_DOWN_TO(maxFb, lineSize);
if (maxFb > lineSize * MAX_DISPLAY_HEIGHT)
maxFb = lineSize * MAX_DISPLAY_HEIGHT;
if (maxFb > pI830->FrontBuffer.Size) {
unsigned long oldsize;
/*
* Sanity check -- the fb should be the last thing allocated at
* the bottom of the stolen pool.
*/
if (pI830->StolenPool.Free.Start != pI830->FrontBuffer.End) {
xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
"Internal error in I830Allocate2DMemory():\n\t"
"Framebuffer isn't the last allocation at the bottom"
" of StolenPool\n\t(%x != %x).\n",
pI830->FrontBuffer.End,
pI830->StolenPool.Free.Start);
return FALSE;
}
/*
* XXX Maybe should have a "Free" function. This should be
* the only place where a region is resized, and we know that
* the fb is always at the bottom of the aperture/stolen pool,
* and is the only region that is allocated bottom-up.
* Allowing for more general realloction would require a smarter
* allocation system.
*/
oldsize = pI830->FrontBuffer.Size;
pI830->StolenPool.Free.Size += pI830->FrontBuffer.Size;
pI830->StolenPool.Free.Start -= pI830->FrontBuffer.Size;
xf86DrvMsgVerb(pScrn->scrnIndex, X_INFO, verbosity,
"%sUpdated framebuffer allocation size from %d "
"to %d kByte\n", s, oldsize / 1024, maxFb / 1024);
xf86DrvMsgVerb(pScrn->scrnIndex, X_INFO, verbosity,
"%sUpdated pixmap cache from %d scanlines to %d "
"scanlines\n", s,
oldsize / lineSize - pScrn->virtualY,
maxFb / lineSize - pScrn->virtualY);
pI830->FbMemBox.y2 = maxFb / lineSize;
tileable = !(flags & ALLOC_NO_TILING) && pI830->allowPageFlip &&
IsTileable(pScrn->displayWidth * pI830->cpp);
if (tileable) {
align = KB(512);
alignflags = ALIGN_BOTH_ENDS;
} else {
align = KB(64);
alignflags = 0;
}
alloced = I830AllocVidMem(pScrn, &(pI830->FrontBuffer),
&(pI830->StolenPool), maxFb, align,
flags | alignflags |
FROM_ANYWHERE | ALLOCATE_AT_BOTTOM);
if (alloced < maxFb) {
if (!dryrun) {
xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
"Failed to re-allocate framebuffer\n");
}
return FALSE;
}
}
return TRUE;
}
#if REMAP_RESERVED
/*
* Allocate a dummy page to pass when attempting to rebind the
* pre-allocated region.
*/
if (!dryrun) {
memset(&(pI830->Dummy), 0, sizeof(pI830->Dummy));
pI830->Dummy.Key =
xf86AllocateGARTMemory(pScrn->scrnIndex, size, 0, NULL);
pI830->Dummy.Offset = 0;
}
#endif
/* Clear cursor info */
memset(&(pI830->CursorMem), 0, sizeof(pI830->CursorMem));
pI830->CursorMem.Key = -1;
if (!pI830->SWCursor) {
int cursFlags = 0;
/*
* Mouse cursor -- The i810-i830 need a physical address in system
* memory from which to upload the cursor. We get this from
* the agpgart module using a special memory type.
*/
size = HWCURSOR_SIZE;
cursFlags = FROM_ANYWHERE | ALLOCATE_AT_TOP;
if (pI830->CursorNeedsPhysical)
cursFlags |= NEED_PHYSICAL_ADDR;
alloced = I830AllocVidMem(pScrn, &(pI830->CursorMem),
&(pI830->StolenPool), size,
GTT_PAGE_SIZE, flags | cursFlags);
if (alloced < size) {
if (!dryrun) {
xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
"Failed to allocate HW cursor space.\n");
}
} else {
xf86DrvMsgVerb(pScrn->scrnIndex, X_INFO, verbosity,
"%sAllocated %d kB for HW cursor at 0x%x", s,
alloced / 1024, pI830->CursorMem.Start);
if (pI830->CursorNeedsPhysical)
xf86ErrorFVerb(verbosity, " (0x%08x)", pI830->CursorMem.Physical);
xf86ErrorFVerb(verbosity, "\n");
}
}
#ifdef I830_XV
AllocateOverlay(pScrn, flags);
#endif
if (!pI830->NeedRingBufferLow)
AllocateRingBuffer(pScrn, flags);
/* Clear scratch info */
memset(&(pI830->Scratch), 0, sizeof(pI830->Scratch));
pI830->Scratch.Key = -1;
if (!pI830->noAccel) {
size = MAX_SCRATCH_BUFFER_SIZE;
alloced = I830AllocVidMem(pScrn, &(pI830->Scratch), &(pI830->StolenPool),
size, GTT_PAGE_SIZE,
flags | FROM_ANYWHERE | ALLOCATE_AT_TOP);
if (alloced < size) {
size = MIN_SCRATCH_BUFFER_SIZE;
alloced = I830AllocVidMem(pScrn, &(pI830->Scratch),
&(pI830->StolenPool), size,
GTT_PAGE_SIZE,
flags | FROM_ANYWHERE | ALLOCATE_AT_TOP);
}
if (alloced < size) {
if (!dryrun) {
xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
"Failed to allocate scratch buffer space\n");
}
return FALSE;
}
xf86DrvMsgVerb(pScrn->scrnIndex, X_INFO, verbosity,
"%sAllocated %d kB for the scratch buffer at 0x%x\n", s,
alloced / 1024, pI830->Scratch.Start);
}
return TRUE;
}
static unsigned long
AllocFromAGP(ScrnInfoPtr pScrn, I830MemRange *result, unsigned long size,
unsigned long alignment, int flags)
{
I830Ptr pI830 = I830PTR(pScrn);
unsigned long start, end;
unsigned long newApStart, newApEnd;
Bool dryrun = ((flags & ALLOCATE_DRY_RUN) != 0);
if (!result || !size)
return 0;
if ((flags & ALLOCATE_AT_BOTTOM) && pI830->StolenMemory.Size != 0) {
xf86DrvMsg(pScrn->scrnIndex, X_ERROR,
"AllocFromAGP(): can't allocate from "
"bottom when there is stolen memory\n");
return 0;
}
if (size > pI830->FreeMemory) {
if (dryrun)
pI830->FreeMemory = size;
else
return 0;
}
/* Calculate offset */
if (flags & ALLOCATE_AT_BOTTOM) {
start = ROUND_TO(pI830->MemoryAperture.Start, alignment);
if (flags & ALIGN_BOTH_ENDS)
end = ROUND_TO(start + size, alignment);
else
end = start + size;
newApStart = end;
newApEnd = pI830->MemoryAperture.End;
} else {
if (flags & ALIGN_BOTH_ENDS)
end = ROUND_DOWN_TO(pI830->MemoryAperture.End, alignment);
else
end = pI830->MemoryAperture.End;
start = ROUND_DOWN_TO(end - size, alignment);
newApStart = pI830->MemoryAperture.Start;
newApEnd = start;
}
if (!dryrun) {
if (newApStart > newApEnd)
return 0;
if (flags & NEED_PHYSICAL_ADDR) {
result->Key = xf86AllocateGARTMemory(pScrn->scrnIndex, size, 2,
&(result->Physical));
} else {
result->Key = xf86AllocateGARTMemory(pScrn->scrnIndex, size, 0, NULL);
}
if (result->Key == -1)
return 0;
}
pI830->allocatedMemory += size;
pI830->MemoryAperture.Start = newApStart;
pI830->MemoryAperture.End = newApEnd;
pI830->MemoryAperture.Size = newApEnd - newApStart;
pI830->FreeMemory -= size;
result->Start = start;
result->End = start + size;
result->Size = size;
result->Offset = start;
result->Alignment = alignment;
result->Pool = NULL;
return size;
}
brw_blorp_clear_params::brw_blorp_clear_params(struct brw_context *brw,
struct gl_framebuffer *fb,
struct gl_renderbuffer *rb,
GLubyte *color_mask,
bool partial_clear,
unsigned layer)
{
struct gl_context *ctx = &brw->ctx;
struct intel_renderbuffer *irb = intel_renderbuffer(rb);
dst.set(brw, irb->mt, irb->mt_level, layer, true);
/* Override the surface format according to the context's sRGB rules. */
mesa_format format = _mesa_get_render_format(ctx, irb->mt->format);
dst.brw_surfaceformat = brw->render_target_format[format];
x0 = fb->_Xmin;
x1 = fb->_Xmax;
if (rb->Name != 0) {
y0 = fb->_Ymin;
y1 = fb->_Ymax;
} else {
y0 = rb->Height - fb->_Ymax;
y1 = rb->Height - fb->_Ymin;
}
float *push_consts = (float *)&wm_push_consts;
push_consts[0] = ctx->Color.ClearColor.f[0];
push_consts[1] = ctx->Color.ClearColor.f[1];
push_consts[2] = ctx->Color.ClearColor.f[2];
push_consts[3] = ctx->Color.ClearColor.f[3];
use_wm_prog = true;
memset(&wm_prog_key, 0, sizeof(wm_prog_key));
wm_prog_key.use_simd16_replicated_data = true;
/* From the SNB PRM (Vol4_Part1):
*
* "Replicated data (Message Type = 111) is only supported when
* accessing tiled memory. Using this Message Type to access linear
* (untiled) memory is UNDEFINED."
*/
if (irb->mt->tiling == I915_TILING_NONE)
wm_prog_key.use_simd16_replicated_data = false;
/* Constant color writes ignore everyting in blend and color calculator
* state. This is not documented.
*/
for (int i = 0; i < 4; i++) {
if (_mesa_format_has_color_component(irb->mt->format, i) &&
!color_mask[i]) {
color_write_disable[i] = true;
wm_prog_key.use_simd16_replicated_data = false;
}
}
/* If we can do this as a fast color clear, do so.
*
* Note that the condition "!partial_clear" means we only try to do full
* buffer clears using fast color clear logic. This is necessary because
* the fast color clear alignment requirements mean that we typically have
* to clear a larger rectangle than (x0, y0) to (x1, y1). Restricting fast
* color clears to the full-buffer condition guarantees that the extra
* memory locations that get written to are outside the image boundary (and
* hence irrelevant). Note that the rectangle alignment requirements are
* never larger than the size of a tile, so there is no danger of
* overflowing beyond the memory belonging to the region.
*/
if (irb->mt->fast_clear_state != INTEL_FAST_CLEAR_STATE_NO_MCS &&
!partial_clear && wm_prog_key.use_simd16_replicated_data &&
is_color_fast_clear_compatible(brw, format, &ctx->Color.ClearColor)) {
memset(push_consts, 0xff, 4*sizeof(float));
fast_clear_op = GEN7_FAST_CLEAR_OP_FAST_CLEAR;
/* Figure out what the clear rectangle needs to be aligned to, and how
* much it needs to be scaled down.
*/
unsigned x_align, y_align, x_scaledown, y_scaledown;
if (irb->mt->msaa_layout == INTEL_MSAA_LAYOUT_NONE) {
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "Fast Color Clear" bullet (p327):
*
* Clear pass must have a clear rectangle that must follow
* alignment rules in terms of pixels and lines as shown in the
* table below. Further, the clear-rectangle height and width
* must be multiple of the following dimensions. If the height
* and width of the render target being cleared do not meet these
* requirements, an MCS buffer can be created such that it
* follows the requirement and covers the RT.
*
* The alignment size in the table that follows is related to the
* alignment size returned by intel_get_non_msrt_mcs_alignment(), but
* with X alignment multiplied by 16 and Y alignment multiplied by 32.
*/
intel_get_non_msrt_mcs_alignment(brw, irb->mt, &x_align, &y_align);
x_align *= 16;
y_align *= 32;
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "Fast Color Clear" bullet (p327):
*
* In order to optimize the performance MCS buffer (when bound to
* 1X RT) clear similarly to MCS buffer clear for MSRT case,
* clear rect is required to be scaled by the following factors
* in the horizontal and vertical directions:
*
* The X and Y scale down factors in the table that follows are each
* equal to half the alignment value computed above.
*/
x_scaledown = x_align / 2;
y_scaledown = y_align / 2;
/* From BSpec: 3D-Media-GPGPU Engine > 3D Pipeline > Pixel > Pixel
* Backend > MCS Buffer for Render Target(s) [DevIVB+] > Table "Color
* Clear of Non-MultiSampled Render Target Restrictions":
*
* Clear rectangle must be aligned to two times the number of
* pixels in the table shown below due to 16x16 hashing across the
* slice.
*/
x_align *= 2;
y_align *= 2;
} else {
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "MSAA Compression" bullet (p326):
*
* Clear pass for this case requires that scaled down primitive
* is sent down with upper left co-ordinate to coincide with
* actual rectangle being cleared. For MSAA, clear rectangle’s
* height and width need to as show in the following table in
* terms of (width,height) of the RT.
*
* MSAA Width of Clear Rect Height of Clear Rect
* 4X Ceil(1/8*width) Ceil(1/2*height)
* 8X Ceil(1/2*width) Ceil(1/2*height)
*
* The text "with upper left co-ordinate to coincide with actual
* rectangle being cleared" is a little confusing--it seems to imply
* that to clear a rectangle from (x,y) to (x+w,y+h), one needs to
* feed the pipeline using the rectangle (x,y) to
* (x+Ceil(w/N),y+Ceil(h/2)), where N is either 2 or 8 depending on
* the number of samples. Experiments indicate that this is not
* quite correct; actually, what the hardware appears to do is to
* align whatever rectangle is sent down the pipeline to the nearest
* multiple of 2x2 blocks, and then scale it up by a factor of N
* horizontally and 2 vertically. So the resulting alignment is 4
* vertically and either 4 or 16 horizontally, and the scaledown
* factor is 2 vertically and either 2 or 8 horizontally.
*/
switch (irb->mt->num_samples) {
case 4:
x_scaledown = 8;
break;
case 8:
x_scaledown = 2;
break;
default:
assert(!"Unexpected sample count for fast clear");
break;
}
y_scaledown = 2;
x_align = x_scaledown * 2;
y_align = y_scaledown * 2;
}
/* Do the alignment and scaledown. */
x0 = ROUND_DOWN_TO(x0, x_align) / x_scaledown;
y0 = ROUND_DOWN_TO(y0, y_align) / y_scaledown;
x1 = ALIGN(x1, x_align) / x_scaledown;
y1 = ALIGN(y1, y_align) / y_scaledown;
}
}
brw_blorp_clear_params::brw_blorp_clear_params(struct brw_context *brw,
struct gl_framebuffer *fb,
struct gl_renderbuffer *rb,
GLubyte *color_mask,
bool partial_clear)
{
struct gl_context *ctx = &brw->ctx;
struct intel_renderbuffer *irb = intel_renderbuffer(rb);
dst.set(brw, irb->mt, irb->mt_level, irb->mt_layer);
/* Override the surface format according to the context's sRGB rules. */
gl_format format = _mesa_get_render_format(ctx, irb->mt->format);
dst.brw_surfaceformat = brw->render_target_format[format];
x0 = fb->_Xmin;
x1 = fb->_Xmax;
if (rb->Name != 0) {
y0 = fb->_Ymin;
y1 = fb->_Ymax;
} else {
y0 = rb->Height - fb->_Ymax;
y1 = rb->Height - fb->_Ymin;
}
float *push_consts = (float *)&wm_push_consts;
push_consts[0] = ctx->Color.ClearColor.f[0];
push_consts[1] = ctx->Color.ClearColor.f[1];
push_consts[2] = ctx->Color.ClearColor.f[2];
push_consts[3] = ctx->Color.ClearColor.f[3];
use_wm_prog = true;
memset(&wm_prog_key, 0, sizeof(wm_prog_key));
wm_prog_key.use_simd16_replicated_data = true;
/* From the SNB PRM (Vol4_Part1):
*
* "Replicated data (Message Type = 111) is only supported when
* accessing tiled memory. Using this Message Type to access linear
* (untiled) memory is UNDEFINED."
*/
if (irb->mt->region->tiling == I915_TILING_NONE)
wm_prog_key.use_simd16_replicated_data = false;
/* Constant color writes ignore everyting in blend and color calculator
* state. This is not documented.
*/
for (int i = 0; i < 4; i++) {
if (!color_mask[i]) {
color_write_disable[i] = true;
wm_prog_key.use_simd16_replicated_data = false;
}
}
/* If we can do this as a fast color clear, do so. */
if (irb->mt->mcs_state != INTEL_MCS_STATE_NONE && !partial_clear &&
wm_prog_key.use_simd16_replicated_data &&
is_color_fast_clear_compatible(brw, format, &ctx->Color.ClearColor)) {
memset(push_consts, 0xff, 4*sizeof(float));
fast_clear_op = GEN7_FAST_CLEAR_OP_FAST_CLEAR;
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "Fast Color Clear" bullet (p327):
*
* Clear pass must have a clear rectangle that must follow alignment
* rules in terms of pixels and lines as shown in the table
* below. Further, the clear-rectangle height and width must be
* multiple of the following dimensions. If the height and width of
* the render target being cleared do not meet these requirements,
* an MCS buffer can be created such that it follows the requirement
* and covers the RT.
*
* The alignment size in the table that follows is related to the
* alignment size returned by intel_get_non_msrt_mcs_alignment(), but
* with X alignment multiplied by 16 and Y alignment multiplied by 32.
*/
unsigned x_align, y_align;
intel_get_non_msrt_mcs_alignment(brw, irb->mt, &x_align, &y_align);
x_align *= 16;
y_align *= 32;
/* From BSpec: 3D-Media-GPGPU Engine > 3D Pipeline > Pixel > Pixel
* Backend > MCS Buffer for Render Target(s) [DevIVB+] > Table "Color
* Clear of Non-MultiSampled Render Target Restrictions":
*
* Clear rectangle must be aligned to two times the number of pixels in
* the table shown below due to 16x16 hashing across the slice.
*/
x0 = ROUND_DOWN_TO(x0, 2 * x_align);
y0 = ROUND_DOWN_TO(y0, 2 * y_align);
x1 = ALIGN(x1, 2 * x_align);
y1 = ALIGN(y1, 2 * y_align);
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "Fast Color Clear" bullet (p327):
*
* In order to optimize the performance MCS buffer (when bound to 1X
* RT) clear similarly to MCS buffer clear for MSRT case, clear rect
* is required to be scaled by the following factors in the
* horizontal and vertical directions:
*
* The X and Y scale down factors in the table that follows are each
* equal to half the alignment value computed above.
*/
unsigned x_scaledown = x_align / 2;
unsigned y_scaledown = y_align / 2;
x0 /= x_scaledown;
y0 /= y_scaledown;
x1 /= x_scaledown;
y1 /= y_scaledown;
}
}
static void
gen7_upload_urb(struct brw_context *brw)
{
const int push_size_kB =
(brw->gen >= 8 || (brw->is_haswell && brw->gt == 3)) ? 32 : 16;
/* BRW_NEW_VS_PROG_DATA */
unsigned vs_size = MAX2(brw->vs.prog_data->base.urb_entry_size, 1);
unsigned vs_entry_size_bytes = vs_size * 64;
/* BRW_NEW_GEOMETRY_PROGRAM, BRW_NEW_GS_PROG_DATA */
bool gs_present = brw->geometry_program;
unsigned gs_size = gs_present ? brw->gs.prog_data->base.urb_entry_size : 1;
unsigned gs_entry_size_bytes = gs_size * 64;
/* If we're just switching between programs with the same URB requirements,
* skip the rest of the logic.
*/
if (!(brw->ctx.NewDriverState & BRW_NEW_CONTEXT) &&
brw->urb.vsize == vs_size &&
brw->urb.gs_present == gs_present &&
brw->urb.gsize == gs_size) {
return;
}
brw->urb.vsize = vs_size;
brw->urb.gs_present = gs_present;
brw->urb.gsize = gs_size;
/* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
*
* VS Number of URB Entries must be divisible by 8 if the VS URB Entry
* Allocation Size is less than 9 512-bit URB entries.
*
* Similar text exists for GS.
*/
unsigned vs_granularity = (vs_size < 9) ? 8 : 1;
unsigned gs_granularity = (gs_size < 9) ? 8 : 1;
/* URB allocations must be done in 8k chunks. */
unsigned chunk_size_bytes = 8192;
/* Determine the size of the URB in chunks.
*/
unsigned urb_chunks = brw->urb.size * 1024 / chunk_size_bytes;
/* Reserve space for push constants */
unsigned push_constant_bytes = 1024 * push_size_kB;
unsigned push_constant_chunks =
push_constant_bytes / chunk_size_bytes;
/* Initially, assign each stage the minimum amount of URB space it needs,
* and make a note of how much additional space it "wants" (the amount of
* additional space it could actually make use of).
*/
/* VS has a lower limit on the number of URB entries */
unsigned vs_chunks =
ALIGN(brw->urb.min_vs_entries * vs_entry_size_bytes, chunk_size_bytes) /
chunk_size_bytes;
unsigned vs_wants =
ALIGN(brw->urb.max_vs_entries * vs_entry_size_bytes,
chunk_size_bytes) / chunk_size_bytes - vs_chunks;
unsigned gs_chunks = 0;
unsigned gs_wants = 0;
if (gs_present) {
/* There are two constraints on the minimum amount of URB space we can
* allocate:
*
* (1) We need room for at least 2 URB entries, since we always operate
* the GS in DUAL_OBJECT mode.
*
* (2) We can't allocate less than nr_gs_entries_granularity.
*/
gs_chunks = ALIGN(MAX2(gs_granularity, 2) * gs_entry_size_bytes,
chunk_size_bytes) / chunk_size_bytes;
gs_wants =
ALIGN(brw->urb.max_gs_entries * gs_entry_size_bytes,
chunk_size_bytes) / chunk_size_bytes - gs_chunks;
}
/* There should always be enough URB space to satisfy the minimum
* requirements of each stage.
*/
unsigned total_needs = push_constant_chunks + vs_chunks + gs_chunks;
assert(total_needs <= urb_chunks);
/* Mete out remaining space (if any) in proportion to "wants". */
unsigned total_wants = vs_wants + gs_wants;
unsigned remaining_space = urb_chunks - total_needs;
if (remaining_space > total_wants)
remaining_space = total_wants;
if (remaining_space > 0) {
unsigned vs_additional = (unsigned)
roundf(vs_wants * (((float) remaining_space) / total_wants));
vs_chunks += vs_additional;
remaining_space -= vs_additional;
gs_chunks += remaining_space;
}
/* Sanity check that we haven't over-allocated. */
assert(push_constant_chunks + vs_chunks + gs_chunks <= urb_chunks);
/* Finally, compute the number of entries that can fit in the space
* allocated to each stage.
*/
unsigned nr_vs_entries = vs_chunks * chunk_size_bytes / vs_entry_size_bytes;
unsigned nr_gs_entries = gs_chunks * chunk_size_bytes / gs_entry_size_bytes;
/* Since we rounded up when computing *_wants, this may be slightly more
* than the maximum allowed amount, so correct for that.
*/
nr_vs_entries = MIN2(nr_vs_entries, brw->urb.max_vs_entries);
nr_gs_entries = MIN2(nr_gs_entries, brw->urb.max_gs_entries);
/* Ensure that we program a multiple of the granularity. */
nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, vs_granularity);
nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, gs_granularity);
/* Finally, sanity check to make sure we have at least the minimum number
* of entries needed for each stage.
*/
assert(nr_vs_entries >= brw->urb.min_vs_entries);
if (gs_present)
assert(nr_gs_entries >= 2);
/* Gen7 doesn't actually use brw->urb.nr_{vs,gs}_entries, but it seems
* better to put reasonable data in there rather than leave them
* uninitialized.
*/
brw->urb.nr_vs_entries = nr_vs_entries;
brw->urb.nr_gs_entries = nr_gs_entries;
/* Lay out the URB in the following order:
* - push constants
* - VS
* - GS
*/
brw->urb.vs_start = push_constant_chunks;
brw->urb.gs_start = push_constant_chunks + vs_chunks;
if (brw->gen == 7 && !brw->is_haswell && !brw->is_baytrail)
gen7_emit_vs_workaround_flush(brw);
gen7_emit_urb_state(brw,
brw->urb.nr_vs_entries, vs_size, brw->urb.vs_start,
brw->urb.nr_gs_entries, gs_size, brw->urb.gs_start);
}
/**
* When the GS is not in use, we assign the entire URB space to the VS. When
* the GS is in use, we split the URB space evenly between the VS and the GS.
* This is not ideal, but it's simple.
*
* URB size / 2 URB size / 2
* _____________-______________ _____________-______________
* / \ / \
* +-------------------------------------------------------------+
* | Vertex Shader Entries | Geometry Shader Entries |
* +-------------------------------------------------------------+
*
* Sandybridge GT1 has 32kB of URB space, while GT2 has 64kB.
* (See the Sandybridge PRM, Volume 2, Part 1, Section 1.4.7: 3DSTATE_URB.)
*/
static void
gen6_upload_urb( struct brw_context *brw )
{
struct intel_context *intel = &brw->intel;
int nr_vs_entries, nr_gs_entries;
int total_urb_size = brw->urb.size * 1024; /* in bytes */
/* CACHE_NEW_VS_PROG */
brw->urb.vs_size = MAX2(brw->vs.prog_data->urb_entry_size, 1);
/* We use the same VUE layout for VS outputs and GS outputs (as it's what
* the SF and Clipper expect), so we can simply make the GS URB entry size
* the same as for the VS. This may technically be too large in cases
* where we have few vertex attributes and a lot of varyings, since the VS
* size is determined by the larger of the two. For now, it's safe.
*/
brw->urb.gs_size = brw->urb.vs_size;
/* Calculate how many entries fit in each stage's section of the URB */
if (brw->gs.prog_active) {
nr_vs_entries = (total_urb_size/2) / (brw->urb.vs_size * 128);
nr_gs_entries = (total_urb_size/2) / (brw->urb.gs_size * 128);
} else {
nr_vs_entries = total_urb_size / (brw->urb.vs_size * 128);
nr_gs_entries = 0;
}
/* Then clamp to the maximum allowed by the hardware */
if (nr_vs_entries > brw->urb.max_vs_entries)
nr_vs_entries = brw->urb.max_vs_entries;
if (nr_gs_entries > brw->urb.max_gs_entries)
nr_gs_entries = brw->urb.max_gs_entries;
/* Finally, both must be a multiple of 4 (see 3DSTATE_URB in the PRM). */
brw->urb.nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, 4);
brw->urb.nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, 4);
assert(brw->urb.nr_vs_entries >= 24);
assert(brw->urb.nr_vs_entries % 4 == 0);
assert(brw->urb.nr_gs_entries % 4 == 0);
assert(brw->urb.vs_size < 5);
assert(brw->urb.gs_size < 5);
BEGIN_BATCH(3);
OUT_BATCH(_3DSTATE_URB << 16 | (3 - 2));
OUT_BATCH(((brw->urb.vs_size - 1) << GEN6_URB_VS_SIZE_SHIFT) |
((brw->urb.nr_vs_entries) << GEN6_URB_VS_ENTRIES_SHIFT));
OUT_BATCH(((brw->urb.gs_size - 1) << GEN6_URB_GS_SIZE_SHIFT) |
((brw->urb.nr_gs_entries) << GEN6_URB_GS_ENTRIES_SHIFT));
ADVANCE_BATCH();
/* From the PRM Volume 2 part 1, section 1.4.7:
*
* Because of a urb corruption caused by allocating a previous gsunit’s
* urb entry to vsunit software is required to send a "GS NULL
* Fence"(Send URB fence with VS URB size == 1 and GS URB size == 0) plus
* a dummy DRAW call before any case where VS will be taking over GS URB
* space.
*
* It is not clear exactly what this means ("URB fence" is a command that
* doesn't exist on Gen6). So for now we just do a full pipeline flush as
* a workaround.
*/
if (brw->urb.gen6_gs_previously_active && !brw->gs.prog_active)
intel_batchbuffer_emit_mi_flush(intel);
brw->urb.gen6_gs_previously_active = brw->gs.prog_active;
}
/* The x0, y0, x1, and y1 parameters must already be populated with the render
* area of the framebuffer to be cleared.
*/
static void
get_fast_clear_rect(const struct isl_device *dev,
const struct isl_surf *aux_surf,
unsigned *x0, unsigned *y0,
unsigned *x1, unsigned *y1)
{
unsigned int x_align, y_align;
unsigned int x_scaledown, y_scaledown;
/* Only single sampled surfaces need to (and actually can) be resolved. */
if (aux_surf->usage == ISL_SURF_USAGE_CCS_BIT) {
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "Fast Color Clear" bullet (p327):
*
* Clear pass must have a clear rectangle that must follow
* alignment rules in terms of pixels and lines as shown in the
* table below. Further, the clear-rectangle height and width
* must be multiple of the following dimensions. If the height
* and width of the render target being cleared do not meet these
* requirements, an MCS buffer can be created such that it
* follows the requirement and covers the RT.
*
* The alignment size in the table that follows is related to the
* alignment size that is baked into the CCS surface format but with X
* alignment multiplied by 16 and Y alignment multiplied by 32.
*/
x_align = isl_format_get_layout(aux_surf->format)->bw;
y_align = isl_format_get_layout(aux_surf->format)->bh;
x_align *= 16;
/* SKL+ line alignment requirement for Y-tiled are half those of the prior
* generations.
*/
if (dev->info->gen >= 9)
y_align *= 16;
else
y_align *= 32;
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "Fast Color Clear" bullet (p327):
*
* In order to optimize the performance MCS buffer (when bound to
* 1X RT) clear similarly to MCS buffer clear for MSRT case,
* clear rect is required to be scaled by the following factors
* in the horizontal and vertical directions:
*
* The X and Y scale down factors in the table that follows are each
* equal to half the alignment value computed above.
*/
x_scaledown = x_align / 2;
y_scaledown = y_align / 2;
/* From BSpec: 3D-Media-GPGPU Engine > 3D Pipeline > Pixel > Pixel
* Backend > MCS Buffer for Render Target(s) [DevIVB+] > Table "Color
* Clear of Non-MultiSampled Render Target Restrictions":
*
* Clear rectangle must be aligned to two times the number of
* pixels in the table shown below due to 16x16 hashing across the
* slice.
*/
x_align *= 2;
y_align *= 2;
} else {
assert(aux_surf->usage == ISL_SURF_USAGE_MCS_BIT);
/* From the Ivy Bridge PRM, Vol2 Part1 11.7 "MCS Buffer for Render
* Target(s)", beneath the "MSAA Compression" bullet (p326):
*
* Clear pass for this case requires that scaled down primitive
* is sent down with upper left co-ordinate to coincide with
* actual rectangle being cleared. For MSAA, clear rectangle’s
* height and width need to as show in the following table in
* terms of (width,height) of the RT.
*
* MSAA Width of Clear Rect Height of Clear Rect
* 2X Ceil(1/8*width) Ceil(1/2*height)
* 4X Ceil(1/8*width) Ceil(1/2*height)
* 8X Ceil(1/2*width) Ceil(1/2*height)
* 16X width Ceil(1/2*height)
*
* The text "with upper left co-ordinate to coincide with actual
* rectangle being cleared" is a little confusing--it seems to imply
* that to clear a rectangle from (x,y) to (x+w,y+h), one needs to
* feed the pipeline using the rectangle (x,y) to
* (x+Ceil(w/N),y+Ceil(h/2)), where N is either 2 or 8 depending on
* the number of samples. Experiments indicate that this is not
* quite correct; actually, what the hardware appears to do is to
* align whatever rectangle is sent down the pipeline to the nearest
* multiple of 2x2 blocks, and then scale it up by a factor of N
* horizontally and 2 vertically. So the resulting alignment is 4
* vertically and either 4 or 16 horizontally, and the scaledown
* factor is 2 vertically and either 2 or 8 horizontally.
*/
switch (aux_surf->format) {
case ISL_FORMAT_MCS_2X:
case ISL_FORMAT_MCS_4X:
x_scaledown = 8;
break;
case ISL_FORMAT_MCS_8X:
x_scaledown = 2;
break;
case ISL_FORMAT_MCS_16X:
x_scaledown = 1;
break;
default:
unreachable("Unexpected MCS format for fast clear");
}
y_scaledown = 2;
x_align = x_scaledown * 2;
y_align = y_scaledown * 2;
}
*x0 = ROUND_DOWN_TO(*x0, x_align) / x_scaledown;
*y0 = ROUND_DOWN_TO(*y0, y_align) / y_scaledown;
*x1 = ALIGN(*x1, x_align) / x_scaledown;
*y1 = ALIGN(*y1, y_align) / y_scaledown;
}
/**
* When the GS is not in use, we assign the entire URB space to the VS. When
* the GS is in use, we split the URB space evenly between the VS and the GS.
* This is not ideal, but it's simple.
*
* URB size / 2 URB size / 2
* _____________-______________ _____________-______________
* / \ / \
* +-------------------------------------------------------------+
* | Vertex Shader Entries | Geometry Shader Entries |
* +-------------------------------------------------------------+
*
* Sandybridge GT1 has 32kB of URB space, while GT2 has 64kB.
* (See the Sandybridge PRM, Volume 2, Part 1, Section 1.4.7: 3DSTATE_URB.)
*/
static void
gen6_upload_urb( struct brw_context *brw )
{
int nr_vs_entries, nr_gs_entries;
int total_urb_size = brw->urb.size * 1024; /* in bytes */
bool gs_present = brw->ff_gs.prog_active || brw->geometry_program;
/* BRW_NEW_VS_PROG_DATA */
unsigned vs_size = MAX2(brw->vs.prog_data->base.urb_entry_size, 1);
/* Whe using GS to do transform feedback only we use the same VUE layout for
* VS outputs and GS outputs (as it's what the SF and Clipper expect), so we
* can simply make the GS URB entry size the same as for the VS. This may
* technically be too large in cases where we have few vertex attributes and
* a lot of varyings, since the VS size is determined by the larger of the
* two. For now, it's safe.
*
* For user-provided GS the assumption above does not hold since the GS
* outputs can be different from the VS outputs.
*/
unsigned gs_size = vs_size;
if (brw->geometry_program) {
gs_size = brw->gs.prog_data->base.urb_entry_size;
assert(gs_size >= 1);
}
/* Calculate how many entries fit in each stage's section of the URB */
if (gs_present) {
nr_vs_entries = (total_urb_size/2) / (vs_size * 128);
nr_gs_entries = (total_urb_size/2) / (gs_size * 128);
} else {
nr_vs_entries = total_urb_size / (vs_size * 128);
nr_gs_entries = 0;
}
/* Then clamp to the maximum allowed by the hardware */
if (nr_vs_entries > brw->urb.max_vs_entries)
nr_vs_entries = brw->urb.max_vs_entries;
if (nr_gs_entries > brw->urb.max_gs_entries)
nr_gs_entries = brw->urb.max_gs_entries;
/* Finally, both must be a multiple of 4 (see 3DSTATE_URB in the PRM). */
brw->urb.nr_vs_entries = ROUND_DOWN_TO(nr_vs_entries, 4);
brw->urb.nr_gs_entries = ROUND_DOWN_TO(nr_gs_entries, 4);
assert(brw->urb.nr_vs_entries >= brw->urb.min_vs_entries);
assert(brw->urb.nr_vs_entries % 4 == 0);
assert(brw->urb.nr_gs_entries % 4 == 0);
assert(vs_size <= 5);
assert(gs_size <= 5);
BEGIN_BATCH(3);
OUT_BATCH(_3DSTATE_URB << 16 | (3 - 2));
OUT_BATCH(((vs_size - 1) << GEN6_URB_VS_SIZE_SHIFT) |
((brw->urb.nr_vs_entries) << GEN6_URB_VS_ENTRIES_SHIFT));
OUT_BATCH(((gs_size - 1) << GEN6_URB_GS_SIZE_SHIFT) |
((brw->urb.nr_gs_entries) << GEN6_URB_GS_ENTRIES_SHIFT));
ADVANCE_BATCH();
/* From the PRM Volume 2 part 1, section 1.4.7:
*
* Because of a urb corruption caused by allocating a previous gsunit’s
* urb entry to vsunit software is required to send a "GS NULL
* Fence"(Send URB fence with VS URB size == 1 and GS URB size == 0) plus
* a dummy DRAW call before any case where VS will be taking over GS URB
* space.
*
* It is not clear exactly what this means ("URB fence" is a command that
* doesn't exist on Gen6). So for now we just do a full pipeline flush as
* a workaround.
*/
if (brw->urb.gs_present && !gs_present)
brw_emit_mi_flush(brw);
brw->urb.gs_present = gs_present;
}
/**
* Decide how to partition the URB among the various stages.
*
* \param[in] push_constant_bytes - space allocate for push constants.
* \param[in] urb_size_bytes - total size of the URB (from L3 config).
* \param[in] tess_present - are tessellation shaders active?
* \param[in] gs_present - are geometry shaders active?
* \param[in] entry_size - the URB entry size (from the shader compiler)
* \param[out] entries - the number of URB entries for each stage
* \param[out] start - the starting offset for each stage
*/
void
gen_get_urb_config(const struct gen_device_info *devinfo,
unsigned push_constant_bytes, unsigned urb_size_bytes,
bool tess_present, bool gs_present,
const unsigned entry_size[4],
unsigned entries[4], unsigned start[4])
{
const bool active[4] = { true, tess_present, tess_present, gs_present };
/* URB allocations must be done in 8k chunks. */
const unsigned chunk_size_bytes = 8192;
const unsigned push_constant_chunks =
push_constant_bytes / chunk_size_bytes;
const unsigned urb_chunks = urb_size_bytes / chunk_size_bytes;
/* From p35 of the Ivy Bridge PRM (section 1.7.1: 3DSTATE_URB_GS):
*
* VS Number of URB Entries must be divisible by 8 if the VS URB Entry
* Allocation Size is less than 9 512-bit URB entries.
*
* Similar text exists for HS, DS and GS.
*/
unsigned granularity[4];
for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
granularity[i] = (entry_size[i] < 9) ? 8 : 1;
}
unsigned min_entries[4] = {
/* VS has a lower limit on the number of URB entries.
*
* From the Broadwell PRM, 3DSTATE_URB_VS instruction:
* "When tessellation is enabled, the VS Number of URB Entries must be
* greater than or equal to 192."
*/
[MESA_SHADER_VERTEX] = tess_present && devinfo->gen == 8 ?
192 : devinfo->urb.min_entries[MESA_SHADER_VERTEX],
/* There are two constraints on the minimum amount of URB space we can
* allocate:
*
* (1) We need room for at least 2 URB entries, since we always operate
* the GS in DUAL_OBJECT mode.
*
* (2) We can't allocate less than nr_gs_entries_granularity.
*/
[MESA_SHADER_GEOMETRY] = gs_present ? 2 : 0,
[MESA_SHADER_TESS_CTRL] = tess_present ? 1 : 0,
[MESA_SHADER_TESS_EVAL] = tess_present ?
devinfo->urb.min_entries[MESA_SHADER_TESS_EVAL] : 0,
};
/* Min VS Entries isn't a multiple of 8 on Cherryview/Broxton; round up.
* Round them all up.
*/
for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
min_entries[i] = ALIGN(min_entries[i], granularity[i]);
}
unsigned entry_size_bytes[4];
for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
entry_size_bytes[i] = 64 * entry_size[i];
}
/* Initially, assign each stage the minimum amount of URB space it needs,
* and make a note of how much additional space it "wants" (the amount of
* additional space it could actually make use of).
*/
unsigned chunks[4];
unsigned wants[4];
unsigned total_needs = push_constant_chunks;
unsigned total_wants = 0;
for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
if (active[i]) {
chunks[i] = DIV_ROUND_UP(min_entries[i] * entry_size_bytes[i],
chunk_size_bytes);
wants[i] =
DIV_ROUND_UP(devinfo->urb.max_entries[i] * entry_size_bytes[i],
chunk_size_bytes) - chunks[i];
} else {
chunks[i] = 0;
wants[i] = 0;
}
total_needs += chunks[i];
total_wants += wants[i];
}
assert(total_needs <= urb_chunks);
/* Mete out remaining space (if any) in proportion to "wants". */
unsigned remaining_space = MIN2(urb_chunks - total_needs, total_wants);
if (remaining_space > 0) {
for (int i = MESA_SHADER_VERTEX;
total_wants > 0 && i <= MESA_SHADER_TESS_EVAL; i++) {
unsigned additional = (unsigned)
roundf(wants[i] * (((float) remaining_space) / total_wants));
chunks[i] += additional;
remaining_space -= additional;
total_wants -= wants[i];
}
chunks[MESA_SHADER_GEOMETRY] += remaining_space;
}
/* Sanity check that we haven't over-allocated. */
unsigned total_chunks = push_constant_chunks;
for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
total_chunks += chunks[i];
}
assert(total_chunks <= urb_chunks);
/* Finally, compute the number of entries that can fit in the space
* allocated to each stage.
*/
for (int i = MESA_SHADER_VERTEX; i <= MESA_SHADER_GEOMETRY; i++) {
entries[i] = chunks[i] * chunk_size_bytes / entry_size_bytes[i];
/* Since we rounded up when computing wants[], this may be slightly
* more than the maximum allowed amount, so correct for that.
*/
entries[i] = MIN2(entries[i], devinfo->urb.max_entries[i]);
/* Ensure that we program a multiple of the granularity. */
entries[i] = ROUND_DOWN_TO(entries[i], granularity[i]);
/* Finally, sanity check to make sure we have at least the minimum
* number of entries needed for each stage.
*/
assert(entries[i] >= min_entries[i]);
}
/* Lay out the URB in pipeline order: push constants, VS, HS, DS, GS. */
start[0] = push_constant_chunks;
for (int i = MESA_SHADER_TESS_CTRL; i <= MESA_SHADER_GEOMETRY; i++) {
start[i] = start[i - 1] + chunks[i - 1];
}
}
