radeonsi: wait before s_barrier in TCS epilog to fix LLVM 15
Only LGKM is needed here. Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16304>
This commit is contained in:
@@ -219,7 +219,6 @@ LLVMValueRef si_insert_input_ret_float(struct si_shader_context *ctx, LLVMValueR
|
||||
LLVMValueRef si_insert_input_ptr(struct si_shader_context *ctx, LLVMValueRef ret,
|
||||
struct ac_arg param, unsigned return_index);
|
||||
LLVMValueRef si_prolog_get_internal_bindings(struct si_shader_context *ctx);
|
||||
void si_llvm_emit_barrier(struct si_shader_context *ctx);
|
||||
void si_llvm_declare_esgs_ring(struct si_shader_context *ctx);
|
||||
LLVMValueRef si_unpack_param(struct si_shader_context *ctx, struct ac_arg param, unsigned rshift,
|
||||
unsigned bitwidth);
|
||||
|
@@ -319,20 +319,6 @@ LLVMValueRef si_prolog_get_internal_bindings(struct si_shader_context *ctx)
|
||||
return list;
|
||||
}
|
||||
|
||||
void si_llvm_emit_barrier(struct si_shader_context *ctx)
|
||||
{
|
||||
/* GFX6 only (thanks to a hw bug workaround):
|
||||
* The real barrier instruction isn’t needed, because an entire patch
|
||||
* always fits into a single wave.
|
||||
*/
|
||||
if (ctx->screen->info.chip_class == GFX6 && ctx->stage == MESA_SHADER_TESS_CTRL) {
|
||||
ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM | AC_WAIT_VLOAD | AC_WAIT_VSTORE);
|
||||
return;
|
||||
}
|
||||
|
||||
ac_build_s_barrier(&ctx->ac);
|
||||
}
|
||||
|
||||
/* Ensure that the esgs ring is declared.
|
||||
*
|
||||
* We declare it with 64KB alignment as a hint that the
|
||||
|
@@ -683,8 +683,15 @@ static void si_write_tess_factors(struct si_shader_context *ctx, LLVMValueRef re
|
||||
unsigned stride, outer_comps, inner_comps, i, offset;
|
||||
|
||||
/* Add a barrier before loading tess factors from LDS. */
|
||||
if (!shader->key.ge.part.tcs.epilog.invoc0_tess_factors_are_def)
|
||||
si_llvm_emit_barrier(ctx);
|
||||
if (!shader->key.ge.part.tcs.epilog.invoc0_tess_factors_are_def) {
|
||||
ac_build_waitcnt(&ctx->ac, AC_WAIT_LGKM);
|
||||
|
||||
/* GFX6 only: s_barrier isn’t needed in TCS because an entire patch always fits into
|
||||
* a single wave due to a bug workaround disallowing multi-wave HS workgroups.
|
||||
*/
|
||||
if (ctx->screen->info.chip_class != GFX6)
|
||||
ac_build_s_barrier(&ctx->ac);
|
||||
}
|
||||
|
||||
/* Do this only for invocation 0, because the tess levels are per-patch,
|
||||
* not per-vertex.
|
||||
|
Reference in New Issue
Block a user