[igt-dev] [PATCH i-g-t v3 1/1] tests/gem_blits: Add no-reloc capability
Zbigniew Kempczyński
zbigniew.kempczynski at intel.com
Wed Nov 17 12:53:33 UTC 2021
On Wed, Nov 17, 2021 at 01:11:29PM +0100, Kamil Konieczny wrote:
> Add no-relocation mode for GPU gens without relocations. In WC
> mode on discrete dg1 we need to use device_coherent mmap.
>
> Signed-off-by: Kamil Konieczny <kamil.konieczny at linux.intel.com>
>
> ---
> v3: moved flag set EXEC_OBJECT_WRITE inside no-reloc block as it
> is alraedy added if relocations were to apply, removed var name
^-- typo
> change, removed checks for ahnd != 0 for put functions (Zbyszek
> review) and change commit message
Maybe instead of long story:
v3: set EXEC_OBJECT_WRITE for no-reloc path only (Zbigniew)
> v2: Address Zbyszek Kempczynski review: removed has_relocs as it
> is sufficient to test if ahnd variable is non-zero, removed if-s
> blocks around relocations to make patch smaller for review.
v2: remove unnecessary variable rename and use allocator handle
as conditional to diverge reloc and no-reloc paths (Zbigniew)
> ---
> tests/i915/gem_blits.c | 81 +++++++++++++++++++++++++++++++++++-------
> 1 file changed, 69 insertions(+), 12 deletions(-)
>
> diff --git a/tests/i915/gem_blits.c b/tests/i915/gem_blits.c
> index 21dcee68..72dc3ac4 100644
> --- a/tests/i915/gem_blits.c
> +++ b/tests/i915/gem_blits.c
> @@ -38,6 +38,7 @@ struct device {
> int gen;
> int pciid;
> int llc;
> + uint64_t ahnd; /* ahnd != 0 if no-relocs */
> };
>
> struct buffer {
> @@ -119,8 +120,10 @@ static struct buffer *buffer_create(const struct device *device,
> buffer->size = ALIGN(buffer->stride * height, 4096);
> buffer->handle = gem_create(device->fd, buffer->size);
> buffer->caching = device->llc;
> -
> - buffer->gtt_offset = buffer->handle * buffer->size;
> + if (device->ahnd)
> + buffer->gtt_offset = get_offset(device->ahnd, buffer->handle, buffer->size, 0);
> + else
> + buffer->gtt_offset = buffer->handle * buffer->size;
>
> for (int y = 0; y < height; y++) {
> uint32_t *row = buffer->model + y * width;
> @@ -160,20 +163,34 @@ static void buffer_set_tiling(const struct device *device,
> execbuf.buffer_count = ARRAY_SIZE(obj);
> if (device->gen >= 6)
> execbuf.flags = I915_EXEC_BLT;
> + if (device->ahnd)
> + execbuf.flags |= I915_EXEC_NO_RELOC;
>
> memset(obj, 0, sizeof(obj));
> obj[0].handle = gem_create(device->fd, size);
> if (__gem_set_tiling(device->fd, obj[0].handle, tiling, stride) == 0)
> obj[0].flags = EXEC_OBJECT_NEEDS_FENCE;
> + if (device->ahnd) {
> + obj[0].flags |= EXEC_OBJECT_PINNED | EXEC_OBJECT_WRITE;
> + obj[0].offset = get_offset(device->ahnd, obj[0].handle, size, 0);
> + }
>
> obj[1].handle = buffer->handle;
> obj[1].offset = buffer->gtt_offset;
> if (buffer->fenced)
> obj[1].flags = EXEC_OBJECT_NEEDS_FENCE;
> + if (device->ahnd)
> + obj[1].flags |= EXEC_OBJECT_PINNED;
>
> obj[2].handle = gem_create(device->fd, 4096);
> - obj[2].relocs_ptr = to_user_pointer(memset(reloc, 0, sizeof(reloc)));
> - obj[2].relocation_count = 2;
> + if (device->ahnd) {
> + obj[2].offset = get_offset(device->ahnd, obj[2].handle, 4096, 0);
> + obj[2].flags |= EXEC_OBJECT_PINNED;
> + } else {
> + obj[2].relocs_ptr = to_user_pointer(memset(reloc, 0, sizeof(reloc)));
> + obj[2].relocation_count = 2;
> + }
> +
> batch = gem_mmap__cpu(device->fd, obj[2].handle, 0, 4096, PROT_WRITE);
>
> i = 0;
> @@ -247,6 +264,8 @@ static void buffer_set_tiling(const struct device *device,
> gem_execbuf(device->fd, &execbuf);
>
> gem_close(device->fd, obj[2].handle);
> + put_offset(device->ahnd, obj[2].offset);
> +
It seems you're getting offsets twice (for obj[0] and obj[2]) but putting
is only for obj[2]. At the moment reloc doesn't keep object offset tracking
so now there's no-op here, but this may change in the future. Then we would
have a offset leak.
Be aware of offset shuffling so put appropriate offset here :)
> gem_close(device->fd, obj[1].handle);
>
> buffer->gtt_offset = obj[0].offset;
> @@ -292,19 +311,34 @@ static bool blit_to_linear(const struct device *device,
> execbuf.buffer_count = ARRAY_SIZE(obj);
> if (device->gen >= 6)
> execbuf.flags = I915_EXEC_BLT;
> + if (device->ahnd)
> + execbuf.flags |= I915_EXEC_NO_RELOC;
>
> memset(obj, 0, sizeof(obj));
> if (__gem_userptr(device->fd, linear, buffer->size, 0, 0, &obj[0].handle))
> return false;
>
> + if (device->ahnd) {
> + obj[0].flags |= EXEC_OBJECT_PINNED | EXEC_OBJECT_WRITE;
> + obj[0].offset = get_offset(device->ahnd, obj[0].handle, buffer->size, 0);
> + }
> +
> obj[1].handle = buffer->handle;
> obj[1].offset = buffer->gtt_offset;
> obj[1].flags = EXEC_OBJECT_NEEDS_FENCE;
> + if (device->ahnd)
> + obj[1].flags |= EXEC_OBJECT_PINNED;
>
> memset(reloc, 0, sizeof(reloc));
> obj[2].handle = gem_create(device->fd, 4096);
> - obj[2].relocs_ptr = to_user_pointer(reloc);
> - obj[2].relocation_count = ARRAY_SIZE(reloc);
> + if (device->ahnd) {
> + obj[2].flags |= EXEC_OBJECT_PINNED;
> + obj[2].offset = get_offset(device->ahnd, obj[2].handle, 4096, 0);
> + } else {
> + obj[2].relocs_ptr = to_user_pointer(reloc);
> + obj[2].relocation_count = ARRAY_SIZE(reloc);
> + }
> +
> batch = gem_mmap__cpu(device->fd, obj[2].handle, 0, 4096, PROT_WRITE);
>
> if (buffer->tiling >= I915_TILING_Y) {
> @@ -368,6 +402,7 @@ static bool blit_to_linear(const struct device *device,
>
> gem_execbuf(device->fd, &execbuf);
> gem_close(device->fd, obj[2].handle);
> + put_offset(device->ahnd, obj[2].offset);
>
> gem_sync(device->fd, obj[0].handle);
> gem_close(device->fd, obj[0].handle);
Same as above.
--
Zbigniew
> @@ -399,7 +434,8 @@ static void *download(const struct device *device,
> break;
>
> case WC:
> - if (!gem_mmap__has_wc(device->fd) || buffer->tiling)
> + if (!(gem_mmap__has_wc(device->fd) || gem_mmap__has_device_coherent(device->fd))
> + || buffer->tiling)
> mode = GTT;
> break;
>
> @@ -425,9 +461,12 @@ static void *download(const struct device *device,
> break;
>
> case WC:
> - src = gem_mmap__wc(device->fd, buffer->handle,
> - 0, buffer->size,
> - PROT_READ);
> + src = __gem_mmap__wc(device->fd, buffer->handle,
> + 0, buffer->size,
> + PROT_READ);
> + if (!src)
> + src = gem_mmap__device_coherent(device->fd, buffer->handle, 0,
> + buffer->size, PROT_READ);
>
> gem_set_domain(device->fd, buffer->handle,
> I915_GEM_DOMAIN_WC, 0);
> @@ -490,6 +529,7 @@ static void buffer_free(const struct device *device, struct buffer *buffer)
> {
> igt_assert(buffer_check(device, buffer, GTT));
> gem_close(device->fd, buffer->handle);
> + put_offset(device->ahnd, buffer->gtt_offset);
> free(buffer);
> }
>
> @@ -604,22 +644,33 @@ blit(const struct device *device,
> execbuf.buffer_count = ARRAY_SIZE(obj);
> if (device->gen >= 6)
> execbuf.flags = I915_EXEC_BLT;
> + if (device->ahnd)
> + execbuf.flags |= I915_EXEC_NO_RELOC;
>
> memset(obj, 0, sizeof(obj));
> obj[0].handle = dst->handle;
> obj[0].offset = dst->gtt_offset;
> if (dst->tiling)
> obj[0].flags = EXEC_OBJECT_NEEDS_FENCE;
> + if (device->ahnd)
> + obj[0].flags |= EXEC_OBJECT_PINNED | EXEC_OBJECT_WRITE;
>
> obj[1].handle = src->handle;
> obj[1].offset = src->gtt_offset;
> if (src->tiling)
> obj[1].flags = EXEC_OBJECT_NEEDS_FENCE;
> + if (device->ahnd)
> + obj[1].flags |= EXEC_OBJECT_PINNED;
>
> memset(reloc, 0, sizeof(reloc));
> obj[2].handle = gem_create(device->fd, 4096);
> - obj[2].relocs_ptr = to_user_pointer(reloc);
> - obj[2].relocation_count = ARRAY_SIZE(reloc);
> + if (device->ahnd) {
> + obj[2].offset = get_offset(device->ahnd, obj[2].handle, 4096, 0);
> + obj[2].flags |= EXEC_OBJECT_PINNED;
> + } else {
> + obj[2].relocs_ptr = to_user_pointer(reloc);
> + obj[2].relocation_count = ARRAY_SIZE(reloc);
> + }
> batch = gem_mmap__cpu(device->fd, obj[2].handle, 0, 4096, PROT_WRITE);
>
> if ((src->tiling | dst->tiling) >= I915_TILING_Y) {
> @@ -691,6 +742,7 @@ blit(const struct device *device,
>
> gem_execbuf(device->fd, &execbuf);
> gem_close(device->fd, obj[2].handle);
> + put_offset(device->ahnd, obj[2].offset);
>
> dst->gtt_offset = obj[0].offset;
> src->gtt_offset = obj[1].offset;
> @@ -733,6 +785,7 @@ igt_main
> device.pciid = intel_get_drm_devid(device.fd);
> device.gen = intel_gen(device.pciid);
> device.llc = gem_has_llc(device.fd);
> + device.ahnd = get_reloc_ahnd(device.fd, 0);
> }
>
> igt_subtest("basic") {
> @@ -794,4 +847,8 @@ igt_main
> }
> }
> }
> +
> + igt_fixture {
> + put_ahnd(device.ahnd);
> + }
> }
> --
> 2.32.0
>
More information about the igt-dev
mailing list