radv/llvm,aco: always split typed vertex buffer loads on GFX6 and GFX10+

To avoid any alignment issues that triggers memory violations and
eventually a GPU. This can happen if the stride (static or dynamic)
is unaligned and also if the VBO offset is aligned to scalar
(eg. stride is 8 and VBO offset is 2 for R16G16B16A16_SNORM).

The AMD Windows driver also always splits typed vertex fetches.

fossils-db (Sienna Cichlid):
Totals from 56508 (40.54% of 139391) affected shaders:
SGPRs: 2643545 -> 2664516 (+0.79%); split: -0.19%, +0.98%
VGPRs: 2007472 -> 1995408 (-0.60%); split: -0.74%, +0.13%
CodeSize: 70596372 -> 73913312 (+4.70%); split: -0.00%, +4.70%
MaxWaves: 772653 -> 774916 (+0.29%); split: +0.37%, -0.08%
Instrs: 14074162 -> 14567072 (+3.50%); split: -0.00%, +3.51%
Cycles: 69281276 -> 71253252 (+2.85%); split: -0.00%, +2.85%
VMEM: 22047039 -> 25554196 (+15.91%); split: +17.20%, -1.29%
SMEM: 4120370 -> 4360820 (+5.84%); split: +7.41%, -1.58%
VClause: 416913 -> 438361 (+5.14%); split: -1.86%, +7.01%
SClause: 536739 -> 542637 (+1.10%); split: -0.33%, +1.43%
Copies: 977194 -> 970015 (-0.73%); split: -2.43%, +1.69%
Branches: 241205 -> 241193 (-0.00%); split: -0.06%, +0.06%
PreVGPRs: 1505645 -> 1505379 (-0.02%)

This fixes GPU hangs with bin/draw-vertices from Piglit on GFX10+
with Zink.

Cc: mesa-stable
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/8363>

2 files changed, 16 insertions, 12 deletions

diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp
index b4044a8af20..191164727fb 100644
--- a/src/amd/compiler/aco_instruction_selection.cpp
+++ b/src/amd/compiler/aco_instruction_selection.cpp
@@ -4663,11 +4663,17 @@ void visit_load_interpolated_input(isel_context *ctx, nir_intrinsic_instr *instr
 bool check_vertex_fetch_size(isel_context *ctx, const ac_data_format_info *vtx_info,
                              unsigned offset, unsigned stride, unsigned channels)
 {
-   unsigned vertex_byte_size = vtx_info->chan_byte_size * channels;
    if (vtx_info->chan_byte_size != 4 && channels == 3)
       return false;
+
+   /* Always split typed vertex buffer loads on GFX6 and GFX10+ to avoid any
+    * alignment issues that triggers memory violations and eventually a GPU
+    * hang. This can happen if the stride (static or dynamic) is unaligned and
+    * also if the VBO offset is aligned to a scalar (eg. stride is 8 and VBO
+    * offset is 2 for R16G16B16A16_SNORM).
+    */
    return (ctx->options->chip_class >= GFX7 && ctx->options->chip_class <= GFX9) ||
-          (offset % vertex_byte_size == 0 && stride % vertex_byte_size == 0);
+          (channels == 1);
 }
 
 uint8_t get_fetch_data_format(isel_context *ctx, const ac_data_format_info *vtx_info,
diff --git a/src/amd/vulkan/radv_nir_to_llvm.c b/src/amd/vulkan/radv_nir_to_llvm.c
index bc8399ff47e..458586a28f2 100644
--- a/src/amd/vulkan/radv_nir_to_llvm.c
+++ b/src/amd/vulkan/radv_nir_to_llvm.c
@@ -1183,17 +1183,15 @@ handle_vs_input_decl(struct radv_shader_context *ctx,
 		t_offset = LLVMConstInt(ctx->ac.i32, attrib_binding, false);
 		t_list = ac_build_load_to_sgpr(&ctx->ac, t_list_ptr, t_offset);
 
-		/* Perform per-channel vertex fetch operations if unaligned
-		 * access are detected. Only GFX6 and GFX10 are affected.
+		/* Always split typed vertex buffer loads on GFX6 and GFX10+
+		 * to avoid any alignment issues that triggers memory
+		 * violations and eventually a GPU hang. This can happen if
+		 * the stride (static or dynamic) is unaligned and also if the
+		 * VBO offset is aligned to a scalar (eg. stride is 8 and VBO
+		 * offset is 2 for R16G16B16A16_SNORM).
 		 */
-		bool unaligned_vertex_fetches = false;
-		if ((ctx->ac.chip_class == GFX6 || ctx->ac.chip_class >= GFX10) &&
-		    vtx_info->chan_format != data_format &&
-		    ((attrib_offset % vtx_info->element_size) ||
-		     (attrib_stride % vtx_info->element_size)))
-			unaligned_vertex_fetches = true;
-
-		if (unaligned_vertex_fetches) {
+		if (ctx->ac.chip_class == GFX6 ||
+		    ctx->ac.chip_class >= GFX10) {
 			unsigned chan_format = vtx_info->chan_format;
 			LLVMValueRef values[4];