[Intel-gfx] [PATCH v2 08/15] drm/i915/ttm Add a generic TTM memcpy move for page-based iomem
The internal ttm_bo_util memcpy uses vmap functionality, and while it probably might be possible to use it for copying in- and out of sglist represented io memory, using io_mem_reserve() / io_mem_free() callbacks, that would cause problems with fault(). Instead, implement a method mapping page-by-page using kmap_local() semantics. As an additional benefit we then avoid the occasional global TLB flushes of vmap() and consuming vmap space, elimination of a critical point of failure and with a slight change of semantics we could also push the memcpy out async for testing and async driver develpment purposes. Pushing out async can be done since there is no memory allocation going on that could violate the dma_fence lockdep rules. For copies from iomem, use the WC prefetching memcpy variant for additional speed. Note that drivers that don't want to use struct io_mapping but relies on memremap functionality, and that don't want to use scatterlists for VRAM may well define specialized (hopefully reusable) iterators for their particular environment. Cc: Christian König Signed-off-by: Thomas Hellström --- v2: - Move new TTM exports to a separate commit. (Reported by Christian König) - Avoid having the iterator init functions inline. (Reported by Jani Nikula) - Remove a stray comment. --- drivers/gpu/drm/i915/Makefile | 1 + .../gpu/drm/i915/gem/i915_gem_ttm_bo_util.c | 194 ++ .../gpu/drm/i915/gem/i915_gem_ttm_bo_util.h | 107 ++ 3 files changed, 302 insertions(+) create mode 100644 drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.c create mode 100644 drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.h diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile index cb8823570996..958ccc1edfed 100644 --- a/drivers/gpu/drm/i915/Makefile +++ b/drivers/gpu/drm/i915/Makefile @@ -155,6 +155,7 @@ gem-y += \ gem/i915_gem_stolen.o \ gem/i915_gem_throttle.o \ gem/i915_gem_tiling.o \ + gem/i915_gem_ttm_bo_util.o \ gem/i915_gem_userptr.o \ gem/i915_gem_wait.o \ gem/i915_gemfs.o diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.c new file mode 100644 index ..5f347a85bf44 --- /dev/null +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.c @@ -0,0 +1,194 @@ +// SPDX-License-Identifier: MIT +/* + * Copyright © 2021 Intel Corporation + */ + +/** + * DOC: Usage and intentions. + * + * This file contains functionality that we might want to move into + * ttm_bo_util.c if there is a common interest. + * Currently a kmap_local only memcpy with support for page-based iomem regions, + * and fast memcpy from write-combined memory. + */ + +#include +#include +#include +#include + +#include "i915_memcpy.h" + +#include "gem/i915_gem_ttm_bo_util.h" + +static void i915_ttm_kmap_iter_tt_kmap_local(struct i915_ttm_kmap_iter *iter, +struct dma_buf_map *dmap, +pgoff_t i) +{ + struct i915_ttm_kmap_iter_tt *iter_tt = + container_of(iter, typeof(*iter_tt), base); + + dma_buf_map_set_vaddr(dmap, kmap_local_page(iter_tt->tt->pages[i])); +} + +static void i915_ttm_kmap_iter_iomap_kmap_local(struct i915_ttm_kmap_iter *iter, + struct dma_buf_map *dmap, + pgoff_t i) +{ + struct i915_ttm_kmap_iter_iomap *iter_io = + container_of(iter, typeof(*iter_io), base); + void __iomem *addr; + +retry: + while (i >= iter_io->cache.end) { + iter_io->cache.sg = iter_io->cache.sg ? + sg_next(iter_io->cache.sg) : iter_io->st->sgl; + iter_io->cache.i = iter_io->cache.end; + iter_io->cache.end += sg_dma_len(iter_io->cache.sg) >> + PAGE_SHIFT; + iter_io->cache.offs = sg_dma_address(iter_io->cache.sg) - + iter_io->start; + } + + if (i < iter_io->cache.i) { + iter_io->cache.end = 0; + iter_io->cache.sg = NULL; + goto retry; + } + + addr = io_mapping_map_local_wc(iter_io->iomap, iter_io->cache.offs + + (((resource_size_t)i - iter_io->cache.i) + << PAGE_SHIFT)); + dma_buf_map_set_vaddr_iomem(dmap, addr); +} + +static const struct i915_ttm_kmap_iter_ops i915_ttm_kmap_iter_tt_ops = { + .kmap_local = i915_ttm_kmap_iter_tt_kmap_local +}; + +static const struct i915_ttm_kmap_iter_ops i915_ttm_kmap_iter_io_ops = { + .kmap_local = i915_ttm_kmap_iter_iomap_kmap_local +}; + +static void kunmap_local_dma_buf_map(struct dma_buf_map *map) +{ + if (map->is_iomem) + io_mapping_unmap_local(map->vaddr_iomem); + else + kunmap_local(map->vaddr); +} + +/** +
Re: [Intel-gfx] [PATCH v2 08/15] drm/i915/ttm Add a generic TTM memcpy move for page-based iomem
Am 18.05.21 um 10:26 schrieb Thomas Hellström: The internal ttm_bo_util memcpy uses vmap functionality, and while it probably might be possible to use it for copying in- and out of sglist represented io memory, using io_mem_reserve() / io_mem_free() callbacks, that would cause problems with fault(). Instead, implement a method mapping page-by-page using kmap_local() semantics. As an additional benefit we then avoid the occasional global TLB flushes of vmap() and consuming vmap space, elimination of a critical point of failure and with a slight change of semantics we could also push the memcpy out async for testing and async driver develpment purposes. Pushing out async can be done since there is no memory allocation going on that could violate the dma_fence lockdep rules. For copies from iomem, use the WC prefetching memcpy variant for additional speed. Note that drivers that don't want to use struct io_mapping but relies on memremap functionality, and that don't want to use scatterlists for VRAM may well define specialized (hopefully reusable) iterators for their particular environment. In general yes please since I have that as TODO for TTM for a very long time. But I would prefer to fix the implementation in TTM instead and give it proper cursor handling. Amdgpu is also using page based iomem and we are having similar workarounds in place there as well. I think it makes sense to unify this inside TTM and remove the old memcpy util function when done. Regards, Christian. Cc: Christian König Signed-off-by: Thomas Hellström --- v2: - Move new TTM exports to a separate commit. (Reported by Christian König) - Avoid having the iterator init functions inline. (Reported by Jani Nikula) - Remove a stray comment. --- drivers/gpu/drm/i915/Makefile | 1 + .../gpu/drm/i915/gem/i915_gem_ttm_bo_util.c | 194 ++ .../gpu/drm/i915/gem/i915_gem_ttm_bo_util.h | 107 ++ 3 files changed, 302 insertions(+) create mode 100644 drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.c create mode 100644 drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.h diff --git a/drivers/gpu/drm/i915/Makefile b/drivers/gpu/drm/i915/Makefile index cb8823570996..958ccc1edfed 100644 --- a/drivers/gpu/drm/i915/Makefile +++ b/drivers/gpu/drm/i915/Makefile @@ -155,6 +155,7 @@ gem-y += \ gem/i915_gem_stolen.o \ gem/i915_gem_throttle.o \ gem/i915_gem_tiling.o \ + gem/i915_gem_ttm_bo_util.o \ gem/i915_gem_userptr.o \ gem/i915_gem_wait.o \ gem/i915_gemfs.o diff --git a/drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.c b/drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.c new file mode 100644 index ..5f347a85bf44 --- /dev/null +++ b/drivers/gpu/drm/i915/gem/i915_gem_ttm_bo_util.c @@ -0,0 +1,194 @@ +// SPDX-License-Identifier: MIT +/* + * Copyright © 2021 Intel Corporation + */ + +/** + * DOC: Usage and intentions. + * + * This file contains functionality that we might want to move into + * ttm_bo_util.c if there is a common interest. + * Currently a kmap_local only memcpy with support for page-based iomem regions, + * and fast memcpy from write-combined memory. + */ + +#include +#include +#include +#include + +#include "i915_memcpy.h" + +#include "gem/i915_gem_ttm_bo_util.h" + +static void i915_ttm_kmap_iter_tt_kmap_local(struct i915_ttm_kmap_iter *iter, +struct dma_buf_map *dmap, +pgoff_t i) +{ + struct i915_ttm_kmap_iter_tt *iter_tt = + container_of(iter, typeof(*iter_tt), base); + + dma_buf_map_set_vaddr(dmap, kmap_local_page(iter_tt->tt->pages[i])); +} + +static void i915_ttm_kmap_iter_iomap_kmap_local(struct i915_ttm_kmap_iter *iter, + struct dma_buf_map *dmap, + pgoff_t i) +{ + struct i915_ttm_kmap_iter_iomap *iter_io = + container_of(iter, typeof(*iter_io), base); + void __iomem *addr; + +retry: + while (i >= iter_io->cache.end) { + iter_io->cache.sg = iter_io->cache.sg ? + sg_next(iter_io->cache.sg) : iter_io->st->sgl; + iter_io->cache.i = iter_io->cache.end; + iter_io->cache.end += sg_dma_len(iter_io->cache.sg) >> + PAGE_SHIFT; + iter_io->cache.offs = sg_dma_address(iter_io->cache.sg) - + iter_io->start; + } + + if (i < iter_io->cache.i) { + iter_io->cache.end = 0; + iter_io->cache.sg = NULL; + goto retry; + } + + addr = io_mapping_map_local_wc(iter_io->iomap, iter_io->cache.offs + + (((resource_size_t)i - iter_io->cache.i) + << PAGE_SHIFT)); + dma_buf_map_set_vaddr_iomem(dmap, addr); +} + +static const struct i915
Re: [Intel-gfx] [PATCH v2 08/15] drm/i915/ttm Add a generic TTM memcpy move for page-based iomem
On 5/18/21 1:55 PM, Christian König wrote: Am 18.05.21 um 10:26 schrieb Thomas Hellström: The internal ttm_bo_util memcpy uses vmap functionality, and while it probably might be possible to use it for copying in- and out of sglist represented io memory, using io_mem_reserve() / io_mem_free() callbacks, that would cause problems with fault(). Instead, implement a method mapping page-by-page using kmap_local() semantics. As an additional benefit we then avoid the occasional global TLB flushes of vmap() and consuming vmap space, elimination of a critical point of failure and with a slight change of semantics we could also push the memcpy out async for testing and async driver develpment purposes. Pushing out async can be done since there is no memory allocation going on that could violate the dma_fence lockdep rules. For copies from iomem, use the WC prefetching memcpy variant for additional speed. Note that drivers that don't want to use struct io_mapping but relies on memremap functionality, and that don't want to use scatterlists for VRAM may well define specialized (hopefully reusable) iterators for their particular environment. In general yes please since I have that as TODO for TTM for a very long time. But I would prefer to fix the implementation in TTM instead and give it proper cursor handling. Amdgpu is also using page based iomem and we are having similar workarounds in place there as well. I think it makes sense to unify this inside TTM and remove the old memcpy util function when done. Regards, Christian. Christian, I was thinking when we replace the bo.mem with a pointer (and perhaps have a driver callback to allocate the bo->mem, we could perhaps embed a struct ttm_kmap_iter and use it for all mapping in one way or another). That would mean perhaps land this is i915 now and sort out the unification once the struct ttm_resource, struct ttm_buffer_object separation has landed? /Thomas ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH v2 08/15] drm/i915/ttm Add a generic TTM memcpy move for page-based iomem
Am 18.05.21 um 14:04 schrieb Thomas Hellström: On 5/18/21 1:55 PM, Christian König wrote: Am 18.05.21 um 10:26 schrieb Thomas Hellström: The internal ttm_bo_util memcpy uses vmap functionality, and while it probably might be possible to use it for copying in- and out of sglist represented io memory, using io_mem_reserve() / io_mem_free() callbacks, that would cause problems with fault(). Instead, implement a method mapping page-by-page using kmap_local() semantics. As an additional benefit we then avoid the occasional global TLB flushes of vmap() and consuming vmap space, elimination of a critical point of failure and with a slight change of semantics we could also push the memcpy out async for testing and async driver develpment purposes. Pushing out async can be done since there is no memory allocation going on that could violate the dma_fence lockdep rules. For copies from iomem, use the WC prefetching memcpy variant for additional speed. Note that drivers that don't want to use struct io_mapping but relies on memremap functionality, and that don't want to use scatterlists for VRAM may well define specialized (hopefully reusable) iterators for their particular environment. In general yes please since I have that as TODO for TTM for a very long time. But I would prefer to fix the implementation in TTM instead and give it proper cursor handling. Amdgpu is also using page based iomem and we are having similar workarounds in place there as well. I think it makes sense to unify this inside TTM and remove the old memcpy util function when done. Regards, Christian. Christian, I was thinking when we replace the bo.mem with a pointer (and perhaps have a driver callback to allocate the bo->mem, we could perhaps embed a struct ttm_kmap_iter and use it for all mapping in one way or another). That would mean perhaps land this is i915 now and sort out the unification once the struct ttm_resource, struct ttm_buffer_object separation has landed? That stuff is ready, reviewed and I'm just waiting for some amdgpu changes to land in drm-misc-next to push it. But yes in general an iterator for the resource object sounds like the right plan to me as well. Christian. /Thomas ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH v2 08/15] drm/i915/ttm Add a generic TTM memcpy move for page-based iomem
On 5/18/21 2:09 PM, Christian König wrote: Am 18.05.21 um 14:04 schrieb Thomas Hellström: On 5/18/21 1:55 PM, Christian König wrote: Am 18.05.21 um 10:26 schrieb Thomas Hellström: The internal ttm_bo_util memcpy uses vmap functionality, and while it probably might be possible to use it for copying in- and out of sglist represented io memory, using io_mem_reserve() / io_mem_free() callbacks, that would cause problems with fault(). Instead, implement a method mapping page-by-page using kmap_local() semantics. As an additional benefit we then avoid the occasional global TLB flushes of vmap() and consuming vmap space, elimination of a critical point of failure and with a slight change of semantics we could also push the memcpy out async for testing and async driver develpment purposes. Pushing out async can be done since there is no memory allocation going on that could violate the dma_fence lockdep rules. For copies from iomem, use the WC prefetching memcpy variant for additional speed. Note that drivers that don't want to use struct io_mapping but relies on memremap functionality, and that don't want to use scatterlists for VRAM may well define specialized (hopefully reusable) iterators for their particular environment. In general yes please since I have that as TODO for TTM for a very long time. But I would prefer to fix the implementation in TTM instead and give it proper cursor handling. Amdgpu is also using page based iomem and we are having similar workarounds in place there as well. I think it makes sense to unify this inside TTM and remove the old memcpy util function when done. Regards, Christian. Christian, I was thinking when we replace the bo.mem with a pointer (and perhaps have a driver callback to allocate the bo->mem, we could perhaps embed a struct ttm_kmap_iter and use it for all mapping in one way or another). That would mean perhaps land this is i915 now and sort out the unification once the struct ttm_resource, struct ttm_buffer_object separation has landed? That stuff is ready, reviewed and I'm just waiting for some amdgpu changes to land in drm-misc-next to push it. But yes in general an iterator for the resource object sounds like the right plan to me as well. Christian. OK, so then are you OK with landing this in i915 for now? That would also ofc mean the export you NAK'd but strictly for this memcpy use until we merge it with TTM? /Thomas /Thomas ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH v2 08/15] drm/i915/ttm Add a generic TTM memcpy move for page-based iomem
Am 18.05.21 um 14:52 schrieb Thomas Hellström: On 5/18/21 2:09 PM, Christian König wrote: Am 18.05.21 um 14:04 schrieb Thomas Hellström: On 5/18/21 1:55 PM, Christian König wrote: Am 18.05.21 um 10:26 schrieb Thomas Hellström: The internal ttm_bo_util memcpy uses vmap functionality, and while it probably might be possible to use it for copying in- and out of sglist represented io memory, using io_mem_reserve() / io_mem_free() callbacks, that would cause problems with fault(). Instead, implement a method mapping page-by-page using kmap_local() semantics. As an additional benefit we then avoid the occasional global TLB flushes of vmap() and consuming vmap space, elimination of a critical point of failure and with a slight change of semantics we could also push the memcpy out async for testing and async driver develpment purposes. Pushing out async can be done since there is no memory allocation going on that could violate the dma_fence lockdep rules. For copies from iomem, use the WC prefetching memcpy variant for additional speed. Note that drivers that don't want to use struct io_mapping but relies on memremap functionality, and that don't want to use scatterlists for VRAM may well define specialized (hopefully reusable) iterators for their particular environment. In general yes please since I have that as TODO for TTM for a very long time. But I would prefer to fix the implementation in TTM instead and give it proper cursor handling. Amdgpu is also using page based iomem and we are having similar workarounds in place there as well. I think it makes sense to unify this inside TTM and remove the old memcpy util function when done. Regards, Christian. Christian, I was thinking when we replace the bo.mem with a pointer (and perhaps have a driver callback to allocate the bo->mem, we could perhaps embed a struct ttm_kmap_iter and use it for all mapping in one way or another). That would mean perhaps land this is i915 now and sort out the unification once the struct ttm_resource, struct ttm_buffer_object separation has landed? That stuff is ready, reviewed and I'm just waiting for some amdgpu changes to land in drm-misc-next to push it. But yes in general an iterator for the resource object sounds like the right plan to me as well. Christian. OK, so then are you OK with landing this in i915 for now? That would also ofc mean the export you NAK'd but strictly for this memcpy use until we merge it with TTM? Well you can of course prototype that in i915, but I really don't want to export the TT functions upstream. Can we cleanly move that functionality into TTM instead? Christian. /Thomas /Thomas ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH v2 08/15] drm/i915/ttm Add a generic TTM memcpy move for page-based iomem
On 5/18/21 3:08 PM, Christian König wrote: Am 18.05.21 um 14:52 schrieb Thomas Hellström: On 5/18/21 2:09 PM, Christian König wrote: Am 18.05.21 um 14:04 schrieb Thomas Hellström: On 5/18/21 1:55 PM, Christian König wrote: Am 18.05.21 um 10:26 schrieb Thomas Hellström: The internal ttm_bo_util memcpy uses vmap functionality, and while it probably might be possible to use it for copying in- and out of sglist represented io memory, using io_mem_reserve() / io_mem_free() callbacks, that would cause problems with fault(). Instead, implement a method mapping page-by-page using kmap_local() semantics. As an additional benefit we then avoid the occasional global TLB flushes of vmap() and consuming vmap space, elimination of a critical point of failure and with a slight change of semantics we could also push the memcpy out async for testing and async driver develpment purposes. Pushing out async can be done since there is no memory allocation going on that could violate the dma_fence lockdep rules. For copies from iomem, use the WC prefetching memcpy variant for additional speed. Note that drivers that don't want to use struct io_mapping but relies on memremap functionality, and that don't want to use scatterlists for VRAM may well define specialized (hopefully reusable) iterators for their particular environment. In general yes please since I have that as TODO for TTM for a very long time. But I would prefer to fix the implementation in TTM instead and give it proper cursor handling. Amdgpu is also using page based iomem and we are having similar workarounds in place there as well. I think it makes sense to unify this inside TTM and remove the old memcpy util function when done. Regards, Christian. Christian, I was thinking when we replace the bo.mem with a pointer (and perhaps have a driver callback to allocate the bo->mem, we could perhaps embed a struct ttm_kmap_iter and use it for all mapping in one way or another). That would mean perhaps land this is i915 now and sort out the unification once the struct ttm_resource, struct ttm_buffer_object separation has landed? That stuff is ready, reviewed and I'm just waiting for some amdgpu changes to land in drm-misc-next to push it. But yes in general an iterator for the resource object sounds like the right plan to me as well. Christian. OK, so then are you OK with landing this in i915 for now? That would also ofc mean the export you NAK'd but strictly for this memcpy use until we merge it with TTM? Well you can of course prototype that in i915, but I really don't want to export the TT functions upstream. I understand, I once had the same thoughts trying to avoid that as far as possible, so this function was actually then added to the ttm_bo interface, (hence the awkward naming) as a helper for drivers implementing move(), essentially a very special case of ttm_bo_move_accel_cleanup(), but anyway, see below: Can we cleanly move that functionality into TTM instead? I'll take a look at that, but I think we'd initially be having iterators mimicing the current move_memcpy() for the linear iomem !WC cases, hope that's OK. /Thomas Christian. /Thomas /Thomas ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
Re: [Intel-gfx] [PATCH v2 08/15] drm/i915/ttm Add a generic TTM memcpy move for page-based iomem
Am 18.05.21 um 15:24 schrieb Thomas Hellström: On 5/18/21 3:08 PM, Christian König wrote: Am 18.05.21 um 14:52 schrieb Thomas Hellström: On 5/18/21 2:09 PM, Christian König wrote: Am 18.05.21 um 14:04 schrieb Thomas Hellström: On 5/18/21 1:55 PM, Christian König wrote: Am 18.05.21 um 10:26 schrieb Thomas Hellström: The internal ttm_bo_util memcpy uses vmap functionality, and while it probably might be possible to use it for copying in- and out of sglist represented io memory, using io_mem_reserve() / io_mem_free() callbacks, that would cause problems with fault(). Instead, implement a method mapping page-by-page using kmap_local() semantics. As an additional benefit we then avoid the occasional global TLB flushes of vmap() and consuming vmap space, elimination of a critical point of failure and with a slight change of semantics we could also push the memcpy out async for testing and async driver develpment purposes. Pushing out async can be done since there is no memory allocation going on that could violate the dma_fence lockdep rules. For copies from iomem, use the WC prefetching memcpy variant for additional speed. Note that drivers that don't want to use struct io_mapping but relies on memremap functionality, and that don't want to use scatterlists for VRAM may well define specialized (hopefully reusable) iterators for their particular environment. In general yes please since I have that as TODO for TTM for a very long time. But I would prefer to fix the implementation in TTM instead and give it proper cursor handling. Amdgpu is also using page based iomem and we are having similar workarounds in place there as well. I think it makes sense to unify this inside TTM and remove the old memcpy util function when done. Regards, Christian. Christian, I was thinking when we replace the bo.mem with a pointer (and perhaps have a driver callback to allocate the bo->mem, we could perhaps embed a struct ttm_kmap_iter and use it for all mapping in one way or another). That would mean perhaps land this is i915 now and sort out the unification once the struct ttm_resource, struct ttm_buffer_object separation has landed? That stuff is ready, reviewed and I'm just waiting for some amdgpu changes to land in drm-misc-next to push it. But yes in general an iterator for the resource object sounds like the right plan to me as well. Christian. OK, so then are you OK with landing this in i915 for now? That would also ofc mean the export you NAK'd but strictly for this memcpy use until we merge it with TTM? Well you can of course prototype that in i915, but I really don't want to export the TT functions upstream. I understand, I once had the same thoughts trying to avoid that as far as possible, so this function was actually then added to the ttm_bo interface, (hence the awkward naming) as a helper for drivers implementing move(), essentially a very special case of ttm_bo_move_accel_cleanup(), but anyway, see below: Can we cleanly move that functionality into TTM instead? I'll take a look at that, but I think we'd initially be having iterators mimicing the current move_memcpy() for the linear iomem !WC cases, hope that's OK. Yeah, that's peefectly fine with me. I can tackle cleaning up all drivers and move over to the new implementation when that is fully complete. As I said we already have the same problem in amdgpu and only solved it by avoiding memcpy all together. Christian. /Thomas Christian. /Thomas /Thomas ___ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx