As we do not maintain the contents of the batch buffers as we reuse and may move the objects between cycles, the requirements for I915_EXEC_NORELOC are not met. Rely on natural reloc skipping instead, so long as the ioctl overhead is less than the GPU, it should not have any impact on the incoherency detection rates.
Closes: https://gitlab.freedesktop.org/drm/intel/issues/1203 Signed-off-by: Chris Wilson <ch...@chris-wilson.co.uk> --- tests/i915/gem_exec_flush.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/tests/i915/gem_exec_flush.c b/tests/i915/gem_exec_flush.c index 9b6f2ed19..513cde364 100644 --- a/tests/i915/gem_exec_flush.c +++ b/tests/i915/gem_exec_flush.c @@ -163,7 +163,7 @@ static void run(int fd, unsigned ring, int nchild, int timeout, memset(&execbuf, 0, sizeof(execbuf)); execbuf.buffers_ptr = to_user_pointer(obj); execbuf.buffer_count = 3; - execbuf.flags = ring | (1 << 11) | (1<<12); + execbuf.flags = ring | (1 << 12); if (gen < 6) execbuf.flags |= I915_EXEC_SECURE; @@ -251,7 +251,7 @@ static void run(int fd, unsigned ring, int nchild, int timeout, i = 16 * (idx % 64) + (idx / 64); obj[1].relocs_ptr = to_user_pointer(&reloc0[i]); obj[2].relocs_ptr = to_user_pointer(&reloc1[i]); - execbuf.batch_start_offset = 64*i; + execbuf.batch_start_offset = 64 * i; overwrite: if ((flags & BEFORE) && @@ -396,7 +396,7 @@ static void batch(int fd, unsigned ring, int nchild, int timeout, memset(&execbuf, 0, sizeof(execbuf)); execbuf.buffers_ptr = to_user_pointer(obj); execbuf.buffer_count = 2; - execbuf.flags = ring | (1 << 11) | (1<<12); + execbuf.flags = ring | (1 << 12); if (gen < 6) execbuf.flags |= I915_EXEC_SECURE; -- 2.25.0 _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
