return LLVMBuildOr(builder, div_mask, result, "");
}
+static LLVMValueRef
+do_quantize_to_f16(struct lp_build_nir_context *bld_base,
+ LLVMValueRef src)
+{
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef result;
+ result = LLVMBuildFPTrunc(builder, src, LLVMVectorType(LLVMHalfTypeInContext(gallivm->context), bld_base->base.type.length), "");
+ result = LLVMBuildFPExt(builder, result, bld_base->base.vec_type, "");
+ return result;
+}
+
static LLVMValueRef do_alu_action(struct lp_build_nir_context *bld_base,
nir_op op, unsigned src_bit_size[NIR_MAX_VEC_COMPONENTS], LLVMValueRef src[NIR_MAX_VEC_COMPONENTS])
{
case nir_op_fpow:
result = lp_build_pow(&bld_base->base, src[0], src[1]);
break;
+ case nir_op_fquantize2f16:
+ result = do_quantize_to_f16(bld_base, src[0]);
+ break;
case nir_op_frcp:
result = lp_build_rcp(get_flt_bld(bld_base, src_bit_size[0]), src[0]);
break;
uint32_t const_offset = 0;
LLVMValueRef offset = NULL;
- if (var->data.compact) {
+ if (var->data.compact && nir_src_is_const(instr->arr.index)) {
assert(instr->deref_type == nir_deref_type_array);
const_offset = nir_src_as_uint(instr->arr.index);
goto out;
void lp_build_opt_nir(struct nir_shader *nir)
{
bool progress;
+
+ static const struct nir_lower_tex_options lower_tex_options = {
+ .lower_tg4_offsets = true,
+ };
+ NIR_PASS_V(nir, nir_lower_tex, &lower_tex_options);
+ NIR_PASS_V(nir, nir_lower_frexp);
+
do {
progress = false;
NIR_PASS_V(nir, nir_opt_constant_folding);