nagai/kernel
3
0
Fork 0
forked from mirrors/linux
Code Pull requests Activity

drm/i915/cmdparser: Accelerate copies from WC memory

Browse source 
If we need to use clflush to prepare our batch for reads from memory, we
can bypass the cache instead by using non-temporal copies.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Matthew Auld <matthew.william.auld@gmail.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20160818161718.27187-39-chris@chris-wilson.co.uk
This commit is contained in:
Chris Wilson 2016-08-18 17:17:18 +01:00
parent 76ff480ec9
commit 52a42cec4b
1 changed files with 39 additions and 23 deletions
repo.diff.show_diff_stats Download patch file Download diff file Expand all files Collapse all files

62
drivers/gpu/drm/i915/i915_cmd_parser.c
View file

@ -965,8 +965,7 @@ static u32 *copy_batch(struct drm_i915_gem_object *dst_obj,
{
unsigned int src_needs_clflush;
unsigned int dst_needs_clflush;
void *dst, *ptr;
int offset, n;
void *dst, *src;
int ret;
ret = i915_gem_obj_prepare_shmem_read(src_obj, &src_needs_clflush);
@ -983,31 +982,48 @@ static u32 *copy_batch(struct drm_i915_gem_object *dst_obj,
if (IS_ERR(dst))
goto unpin_dst;
ptr = dst;
offset = offset_in_page(batch_start_offset);
src = ERR_PTR(-ENODEV);
if (src_needs_clflush &&
i915_memcpy_from_wc((void *)(uintptr_t)batch_start_offset, 0, 0)) {
src = i915_gem_object_pin_map(src_obj, I915_MAP_WC);
if (!IS_ERR(src)) {
i915_memcpy_from_wc(dst,
src + batch_start_offset,
ALIGN(batch_len, 16));
i915_gem_object_unpin_map(src_obj);
}
}
if (IS_ERR(src)) {
void *ptr;
int offset, n;
/* We can avoid clflushing partial cachelines before the write if we
* only every write full cache-lines. Since we know that both the
* source and destination are in multiples of PAGE_SIZE, we can simply
* round up to the next cacheline. We don't care about copying too much
* here as we only validate up to the end of the batch.
*/
if (dst_needs_clflush & CLFLUSH_BEFORE)
batch_len = roundup(batch_len, boot_cpu_data.x86_clflush_size);
offset = offset_in_page(batch_start_offset);
for (n = batch_start_offset >> PAGE_SHIFT; batch_len; n++) {
int len = min_t(int, batch_len, PAGE_SIZE - offset);
void *vaddr;
/* We can avoid clflushing partial cachelines before the write
* if we only every write full cache-lines. Since we know that
* both the source and destination are in multiples of
* PAGE_SIZE, we can simply round up to the next cacheline.
* We don't care about copying too much here as we only
* validate up to the end of the batch.
*/
if (dst_needs_clflush & CLFLUSH_BEFORE)
batch_len = roundup(batch_len,
boot_cpu_data.x86_clflush_size);
vaddr = kmap_atomic(i915_gem_object_get_page(src_obj, n));
if (src_needs_clflush)
drm_clflush_virt_range(vaddr + offset, len);
memcpy(ptr, vaddr + offset, len);
kunmap_atomic(vaddr);
ptr = dst;
for (n = batch_start_offset >> PAGE_SHIFT; batch_len; n++) {
int len = min_t(int, batch_len, PAGE_SIZE - offset);
ptr += len;
batch_len -= len;
offset = 0;
src = kmap_atomic(i915_gem_object_get_page(src_obj, n));
if (src_needs_clflush)
drm_clflush_virt_range(src + offset, len);
memcpy(ptr, src + offset, len);
kunmap_atomic(src);
ptr += len;
batch_len -= len;
offset = 0;
}
}
/* dst_obj is returned with vmap pinned */