[PATCH i-g-t 2/2] lib/amdgpu: add gfx12 shader blob for sync_dependency_test

[Jesse Zhang <Jesse.Zhang@amd.com>]

The shader blob used by amdgpu_sync_dependency_test() is a GFX11 (RDNA3)
assembly. GFX12 (RDNA4) re-encoded several of the SOP/FLAT opcodes it uses
(s_endpgm, s_cbranch_scc0 and the store), so on gfx1200 the GFX11 blob decodes
into different instructions: the shader neither terminates nor stores its
result, and the test hangs/fails.

Add a gfx1200 re-assembled copy of the same kernel and select it in
get_shader_bin() by family_id, falling back to the existing GFX11 blob for
everything else.

Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
---
 lib/amdgpu/shaders/amd_shaders.c | 35 ++++++++++++++++++++++++++++++--
 lib/amdgpu/shaders/amd_shaders.h |  3 ++-
 tests/amdgpu/amd_basic.c         |  3 ++-
 3 files changed, 37 insertions(+), 4 deletions(-)

diff --git a/lib/amdgpu/shaders/amd_shaders.c b/lib/amdgpu/shaders/amd_shaders.c
index adc4fc051..07489706e 100644
--- a/lib/amdgpu/shaders/amd_shaders.c
+++ b/lib/amdgpu/shaders/amd_shaders.c
@@ -109,12 +109,43 @@ uint32_t shader_bin[] = {
 	SWAP_32(0x000070e0), SWAP_32(0x00000080), SWAP_32(0x000081bf)
 };
 
+/*
+ * GFX12 (RDNA4) re-encoded several of the SOP/FLAT opcodes used above
+ * (s_endpgm, s_cbranch_scc0, the store), so the GFX11 blob decodes into
+ * different instructions and the shader neither terminates nor stores its
+ * result. This is the same kernel re-assembled for gfx1200:
+ *
+ *	s_mov_b32 s2, 0
+ * .Lloop:
+ *	s_add_co_i32 s2, s2, 1
+ *	s_cmp_gt_u32 s2, 0x0098967f
+ *	s_cbranch_scc0 .Lloop
+ *	v_mov_b32_e32 v0, 42
+ *	v_mov_b32_e32 v1, s0
+ *	v_mov_b32_e32 v2, s1
+ *	flat_store_b32 v[1:2], v0
+ *	s_endpgm
+ */
+static  const
+uint32_t shader_bin_gfx12[] = {
+	SWAP_32(0x800082be), SWAP_32(0x02810281), SWAP_32(0x02ff08bf), SWAP_32(0x7f969800),
+	SWAP_32(0xfcffa1bf), SWAP_32(0xaa02007e), SWAP_32(0x0002027e), SWAP_32(0x0102047e),
+	SWAP_32(0x7c8006ec), SWAP_32(0x00000000), SWAP_32(0x01000000), SWAP_32(0x0000b0bf)
+};
+
 const uint32_t *
-get_shader_bin(uint32_t *size_bytes, uint32_t *code_offset, uint32_t *data_offset)
+get_shader_bin(uint32_t *size_bytes, uint32_t *code_offset, uint32_t *data_offset,
+	       uint32_t family_id)
 {
-	*size_bytes = sizeof(shader_bin);
 	*code_offset =  CODE_OFFSET;
 	*data_offset = DATA_OFFSET;
+
+	if (family_id == AMDGPU_FAMILY_GC_12_0_0) {
+		*size_bytes = sizeof(shader_bin_gfx12);
+		return shader_bin_gfx12;
+	}
+
+	*size_bytes = sizeof(shader_bin);
 	return shader_bin;
 }
 
diff --git a/lib/amdgpu/shaders/amd_shaders.h b/lib/amdgpu/shaders/amd_shaders.h
index f7d4cab07..284c1b2f8 100644
--- a/lib/amdgpu/shaders/amd_shaders.h
+++ b/lib/amdgpu/shaders/amd_shaders.h
@@ -30,7 +30,8 @@
 #include "amdgpu/compute_utils/amd_shared_dispatch.h"
 
 const uint32_t *
-get_shader_bin(uint32_t *size_bytes, uint32_t *code_offset, uint32_t *data_offset);
+get_shader_bin(uint32_t *size_bytes, uint32_t *code_offset, uint32_t *data_offset,
+	       uint32_t family_id);
 
 int
 amdgpu_dispatch_load_cs_shader_hang_slow(uint32_t *ptr, uint32_t family_id);
diff --git a/tests/amdgpu/amd_basic.c b/tests/amdgpu/amd_basic.c
index 3ad023472..84a6d8618 100644
--- a/tests/amdgpu/amd_basic.c
+++ b/tests/amdgpu/amd_basic.c
@@ -553,7 +553,8 @@ amdgpu_sync_dependency_test(amdgpu_device_handle device_handle, bool user_queue)
 		igt_assert_eq(r, 0);
 	}
 
-	shader = get_shader_bin(&size_bytes, &code_offset, &data_offset);
+	shader = get_shader_bin(&size_bytes, &code_offset, &data_offset,
+				ip_block->funcs->family_id);
 
 	/* assign cmd buffer */
 	base->attach_buf(base, ib_result_cpu, const_size);
-- 
2.49.0