projects / platform / upstream / mesa.git / commitdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: 52c8504)
radeonsi: limit loop unrolling for LLVM < 13
authorPierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Thu, 3 Feb 2022 08:35:07 +0000 (09:35 +0100)
committerPierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Thu, 3 Feb 2022 09:34:12 +0000 (10:34 +0100)
Without this change LLVM 12 hits this error:

"""
LLVM ERROR: Error while trying to spill SGPR0_SGPR1 from class SReg_64:
Cannot scavenge register without an emergency spill slot!
"""

when running glcts KHR-GL46.arrays_of_arrays_gl.AtomicUsage test.

Fixes: 9ff086052ab ("radeonsi: unroll loops of up to 128 iterations")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14848>

src/gallium/drivers/radeonsi/si_get.c patch | blob | history

diff --git a/src/gallium/drivers/radeonsi/si_get.c b/src/gallium/drivers/radeonsi/si_get.c
index 795461a..985daa1 100644 (file)
--- a/src/gallium/drivers/radeonsi/si_get.c
+++ b/src/gallium/drivers/radeonsi/si_get.c
@@ -1072,8 +1072,8 @@ void si_init_screen_get_functions(struct si_screen *sscreen)
       .has_udot_4x8 = sscreen->info.has_accelerated_dot_product,
       .has_dot_2x16 = sscreen->info.has_accelerated_dot_product,
       .optimize_sample_mask_in = true,
-      .max_unroll_iterations = 128,
-      .max_unroll_iterations_aggressive = 128,
+      .max_unroll_iterations = LLVM_VERSION_MAJOR >= 13 ? 128 : 32,
+      .max_unroll_iterations_aggressive = LLVM_VERSION_MAJOR >= 13 ? 128 : 32,
       .use_interpolated_input_intrinsics = true,
       .lower_uniforms_to_ubo = true,
       .support_16bit_alu = sscreen->options.fp16,
Domain: Graphics System / GL;