LLVMIntPredicate pred, LLVMValueRef src0,
LLVMValueRef src1)
{
+ LLVMTypeRef src0_type = LLVMTypeOf(src0);
+ LLVMTypeRef src1_type = LLVMTypeOf(src1);
+
+ if (LLVMGetTypeKind(src0_type) == LLVMPointerTypeKind &&
+ LLVMGetTypeKind(src1_type) != LLVMPointerTypeKind) {
+ src1 = LLVMBuildIntToPtr(ctx->builder, src1, src0_type, "");
+ } else if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
+ LLVMGetTypeKind(src0_type) != LLVMPointerTypeKind) {
+ src0 = LLVMBuildIntToPtr(ctx->builder, src0, src1_type, "");
+ }
+
LLVMValueRef result = LLVMBuildICmp(ctx->builder, pred, src0, src1, "");
return LLVMBuildSelect(ctx->builder, result,
LLVMConstInt(ctx->i32, 0xFFFFFFFF, false),
LLVMTypeRef result_type,
LLVMValueRef src0)
{
- char name[64];
+ char name[64], type[64];
LLVMValueRef params[] = {
ac_to_float(ctx, src0),
};
- ASSERTED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
- ac_get_elem_bits(ctx, result_type));
+ ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
+ ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
assert(length < sizeof(name));
return ac_build_intrinsic(ctx, name, result_type, params, 1, AC_FUNC_ATTR_READNONE);
}
LLVMTypeRef result_type,
LLVMValueRef src0, LLVMValueRef src1)
{
- char name[64];
+ char name[64], type[64];
LLVMValueRef params[] = {
ac_to_float(ctx, src0),
ac_to_float(ctx, src1),
};
- ASSERTED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
- ac_get_elem_bits(ctx, result_type));
+ ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
+ ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
assert(length < sizeof(name));
return ac_build_intrinsic(ctx, name, result_type, params, 2, AC_FUNC_ATTR_READNONE);
}
LLVMTypeRef result_type,
LLVMValueRef src0, LLVMValueRef src1, LLVMValueRef src2)
{
- char name[64];
+ char name[64], type[64];
LLVMValueRef params[] = {
ac_to_float(ctx, src0),
ac_to_float(ctx, src1),
ac_to_float(ctx, src2),
};
- ASSERTED const int length = snprintf(name, sizeof(name), "%s.f%d", intrin,
- ac_get_elem_bits(ctx, result_type));
+ ac_build_type_name_for_intr(LLVMTypeOf(params[0]), type, sizeof(type));
+ ASSERTED const int length = snprintf(name, sizeof(name), "%s.%s", intrin, type);
assert(length < sizeof(name));
return ac_build_intrinsic(ctx, name, result_type, params, 3, AC_FUNC_ATTR_READNONE);
}
LLVMTypeRef src1_type = LLVMTypeOf(src1);
LLVMTypeRef src2_type = LLVMTypeOf(src2);
- assert(LLVMGetTypeKind(LLVMTypeOf(src0)) != LLVMFixedVectorTypeKind);
+ assert(LLVMGetTypeKind(LLVMTypeOf(src0)) != LLVMVectorTypeKind);
if (LLVMGetTypeKind(src1_type) == LLVMPointerTypeKind &&
LLVMGetTypeKind(src2_type) != LLVMPointerTypeKind) {
result = LLVMBuildFMul(ctx->ac.builder, src[0], src[1], "");
break;
case nir_op_frcp:
- result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rcp",
- ac_to_float_type(&ctx->ac, def_type), src[0]);
+ /* For doubles, we need precise division to pass GLCTS. */
+ if (ctx->ac.float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL &&
+ ac_get_type_size(def_type) == 8) {
+ result = LLVMBuildFDiv(ctx->ac.builder, ctx->ac.f64_1,
+ ac_to_float(&ctx->ac, src[0]), "");
+ } else {
+ result = emit_intrin_1f_param(&ctx->ac, "llvm.amdgcn.rcp",
+ ac_to_float_type(&ctx->ac, def_type), src[0]);
+ }
break;
case nir_op_iand:
result = LLVMBuildAnd(ctx->ac.builder, src[0], src[1], "");
result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
break;
case nir_op_f2f16_rtz:
+ case nir_op_f2f16:
+ case nir_op_f2fmp:
src[0] = ac_to_float(&ctx->ac, src[0]);
- if (LLVMTypeOf(src[0]) == ctx->ac.f64)
- src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
- LLVMValueRef param[2] = { src[0], ctx->ac.f32_0 };
- result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
- result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
+
+ /* For OpenGL, we want fast packing with v_cvt_pkrtz_f16, but if we use it,
+ * all f32->f16 conversions have to round towards zero, because both scalar
+ * and vec2 down-conversions have to round equally.
+ */
+ if (ctx->ac.float_mode == AC_FLOAT_MODE_DEFAULT_OPENGL ||
+ instr->op == nir_op_f2f16_rtz) {
+ src[0] = ac_to_float(&ctx->ac, src[0]);
+
+ if (LLVMTypeOf(src[0]) == ctx->ac.f64)
+ src[0] = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ctx->ac.f32, "");
+
+ /* Fast path conversion. This only works if NIR is vectorized
+ * to vec2 16.
+ */
+ if (LLVMTypeOf(src[0]) == ctx->ac.v2f32) {
+ LLVMValueRef args[] = {
+ ac_llvm_extract_elem(&ctx->ac, src[0], 0),
+ ac_llvm_extract_elem(&ctx->ac, src[0], 1),
+ };
+ result = ac_build_cvt_pkrtz_f16(&ctx->ac, args);
+ break;
+ }
+
+ assert(ac_get_llvm_num_components(src[0]) == 1);
+ LLVMValueRef param[2] = { src[0], LLVMGetUndef(ctx->ac.f32) };
+ result = ac_build_cvt_pkrtz_f16(&ctx->ac, param);
+ result = LLVMBuildExtractElement(ctx->ac.builder, result, ctx->ac.i32_0, "");
+ } else {
+ if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
+ result = LLVMBuildFPExt(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
+ else
+ result = LLVMBuildFPTrunc(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
+ }
break;
case nir_op_f2f16_rtne:
- case nir_op_f2f16:
case nir_op_f2f32:
case nir_op_f2f64:
src[0] = ac_to_float(&ctx->ac, src[0]);
break;
case nir_op_u2u8:
case nir_op_u2u16:
+ case nir_op_u2ump:
case nir_op_u2u32:
case nir_op_u2u64:
if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
break;
case nir_op_i2i8:
case nir_op_i2i16:
+ case nir_op_i2imp:
case nir_op_i2i32:
case nir_op_i2i64:
if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF) {
unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
+ assert(instr->dest.is_ssa);
return ac_build_buffer_load_format(&ctx->ac,
args->resource,
args->coords[0],
ctx->ac.i32_0,
util_last_bit(mask),
- 0, true);
+ 0, true,
+ instr->dest.ssa.bit_size == 16);
}
args->opcode = ac_image_sample;
break;
case nir_texop_tg4:
args->opcode = ac_image_gather4;
- args->level_zero = true;
+ if (!args->lod && !args->bias)
+ args->level_zero = true;
break;
case nir_texop_lod:
args->opcode = ac_image_get_lod;
if (instr->dest.ssa.bit_size == 8) {
unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
- LLVMTypeRef vec_type = LLVMVectorType(LLVMInt8TypeInContext(ctx->ac.context), 4 * load_dwords);
+ LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i8, 4 * load_dwords);
ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
LLVMValueRef params[3];
if (load_dwords > 1) {
- LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i32, 2), "");
+ LLVMValueRef res_vec = LLVMBuildBitCast(ctx->ac.builder, res, ctx->ac.v2i32, "");
params[0] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 1, false), "");
params[1] = LLVMBuildExtractElement(ctx->ac.builder, res_vec, LLVMConstInt(ctx->ac.i32, 0, false), "");
} else {
res = LLVMBuildTrunc(ctx->ac.builder, res, LLVMIntTypeInContext(ctx->ac.context, instr->dest.ssa.num_components * 8), "");
if (instr->dest.ssa.num_components > 1)
- res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(LLVMInt8TypeInContext(ctx->ac.context), instr->dest.ssa.num_components), "");
+ res = LLVMBuildBitCast(ctx->ac.builder, res, LLVMVectorType(ctx->ac.i8, instr->dest.ssa.num_components), "");
return res;
} else if (instr->dest.ssa.bit_size == 16) {
unsigned load_dwords = instr->dest.ssa.num_components / 2 + 1;
- LLVMTypeRef vec_type = LLVMVectorType(LLVMInt16TypeInContext(ctx->ac.context), 2 * load_dwords);
+ LLVMTypeRef vec_type = LLVMVectorType(ctx->ac.i16, 2 * load_dwords);
ptr = ac_cast_ptr(&ctx->ac, ptr, vec_type);
LLVMValueRef res = LLVMBuildLoad(ctx->ac.builder, ptr, "");
res = LLVMBuildBitCast(ctx->ac.builder, res, vec_type, "");
LLVMTypeRef dest_type = get_def_type(ctx, &instr->dest.ssa);
LLVMTypeRef src_component_type;
- if (LLVMGetTypeKind(dest_type) == LLVMFixedVectorTypeKind)
+ if (LLVMGetTypeKind(dest_type) == LLVMVectorTypeKind)
src_component_type = LLVMGetElementType(dest_type);
else
src_component_type = dest_type;
bool split_loads = ctx->ac.chip_class == GFX6 && elem_size_bytes < 4;
if (stride != natural_stride || split_loads) {
- if (LLVMGetTypeKind(result_type) == LLVMFixedVectorTypeKind)
+ if (LLVMGetTypeKind(result_type) == LLVMVectorTypeKind)
result_type = LLVMGetElementType(result_type);
LLVMTypeRef ptr_type = LLVMPointerType(result_type,
LLVMValueRef offset = LLVMConstInt(ctx->ac.i32, i * stride / natural_stride, 0);
values[i] = LLVMBuildLoad(ctx->ac.builder,
ac_build_gep_ptr(&ctx->ac, address, offset), "");
+
+ if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
+ LLVMSetOrdering(values[i], LLVMAtomicOrderingMonotonic);
}
return ac_build_gather_values(&ctx->ac, values, instr->dest.ssa.num_components);
} else {
LLVMGetPointerAddressSpace(LLVMTypeOf(address)));
address = LLVMBuildBitCast(ctx->ac.builder, address, ptr_type , "");
LLVMValueRef val = LLVMBuildLoad(ctx->ac.builder, address, "");
+
+ if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
+ LLVMSetOrdering(val, LLVMAtomicOrderingMonotonic);
return val;
}
}
val = LLVMBuildBitCast(ctx->ac.builder, val,
LLVMGetElementType(LLVMTypeOf(address)), "");
- LLVMBuildStore(ctx->ac.builder, val, address);
+ LLVMValueRef store = LLVMBuildStore(ctx->ac.builder, val, address);
+
+ if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
+ LLVMSetOrdering(store, LLVMAtomicOrderingMonotonic);
} else {
LLVMTypeRef val_type = LLVMTypeOf(val);
- if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMFixedVectorTypeKind)
+ if (LLVMGetTypeKind(LLVMTypeOf(val)) == LLVMVectorTypeKind)
val_type = LLVMGetElementType(val_type);
LLVMTypeRef ptr_type = LLVMPointerType(val_type,
chan);
src = LLVMBuildBitCast(ctx->ac.builder, src,
LLVMGetElementType(LLVMTypeOf(ptr)), "");
- LLVMBuildStore(ctx->ac.builder, src, ptr);
+ LLVMValueRef store = LLVMBuildStore(ctx->ac.builder, src, ptr);
+
+ if (nir_intrinsic_access(instr) & (ACCESS_COHERENT | ACCESS_VOLATILE))
+ LLVMSetOrdering(store, LLVMAtomicOrderingMonotonic);
}
}
break;
LLVMValueRef res;
enum glsl_sampler_dim dim;
- enum gl_access_qualifier access;
+ enum gl_access_qualifier access = nir_intrinsic_access(instr);
bool is_array;
if (bindless) {
dim = nir_intrinsic_image_dim(instr);
- access = nir_intrinsic_access(instr);
is_array = nir_intrinsic_image_array(instr);
} else {
const nir_deref_instr *image_deref = get_image_deref(instr);
const struct glsl_type *type = image_deref->type;
const nir_variable *var = nir_deref_instr_get_variable(image_deref);
dim = glsl_get_sampler_dim(type);
- access = var->data.access;
+ access |= var->data.access;
is_array = glsl_sampler_type_is_array(type);
}
vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
ctx->ac.i32_0, "");
+ assert(instr->dest.is_ssa);
bool can_speculate = access & ACCESS_CAN_REORDER;
res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
ctx->ac.i32_0, num_channels,
args.cache_policy,
- can_speculate);
+ can_speculate,
+ instr->dest.ssa.bit_size == 16);
res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
args.dmask = 15;
args.attributes = AC_FUNC_ATTR_READONLY;
+ assert(instr->dest.is_ssa);
+ args.d16 = instr->dest.ssa.bit_size == 16;
+
res = ac_build_image_opcode(&ctx->ac, &args);
}
return exit_waterfall(ctx, &wctx, res);
}
enum glsl_sampler_dim dim;
- enum gl_access_qualifier access;
+ enum gl_access_qualifier access = nir_intrinsic_access(instr);
bool is_array;
if (bindless) {
dim = nir_intrinsic_image_dim(instr);
- access = nir_intrinsic_access(instr);
is_array = nir_intrinsic_image_array(instr);
} else {
const nir_deref_instr *image_deref = get_image_deref(instr);
const struct glsl_type *type = image_deref->type;
const nir_variable *var = nir_deref_instr_get_variable(image_deref);
dim = glsl_get_sampler_dim(type);
- access = var->data.access;
+ access |= var->data.access;
is_array = glsl_sampler_type_is_array(type);
}
ctx->ac.i32_0, "");
ac_build_buffer_store_format(&ctx->ac, rsrc, src, vindex,
- ctx->ac.i32_0, src_channels,
- args.cache_policy);
+ ctx->ac.i32_0, args.cache_policy);
} else {
bool level_zero = nir_src_is_const(instr->src[4]) && nir_src_as_uint(instr->src[4]) == 0;
if (!level_zero)
args.lod = get_src(ctx, instr->src[4]);
args.dmask = 15;
+ args.d16 = ac_get_elem_bits(&ctx->ac, LLVMTypeOf(args.data[0])) == 16;
ac_build_image_opcode(&ctx->ac, &args);
}
case nir_intrinsic_image_deref_atomic_inc_wrap: {
atomic_name = "inc";
atomic_subop = ac_atomic_inc_wrap;
- /* ATOMIC_INC instruction does:
- * value = (value + 1) % (data + 1)
- * but we want:
- * value = (value + 1) % data
- * So replace 'data' by 'data - 1'.
- */
- ctx->ssa_defs[instr->src[3].ssa->index] =
- LLVMBuildSub(ctx->ac.builder,
- ctx->ssa_defs[instr->src[3].ssa->index],
- ctx->ac.i32_1, "");
break;
}
case nir_intrinsic_bindless_image_atomic_dec_wrap:
args.dmask = 0xf;
args.resource = get_image_descriptor(ctx, instr, dynamic_index, AC_DESC_IMAGE, false);
args.opcode = ac_image_get_resinfo;
+ assert(nir_src_as_uint(instr->src[1]) == 0);
args.lod = ctx->ac.i32_0;
args.attributes = AC_FUNC_ATTR_READNONE;
case nir_intrinsic_deref_atomic_exchange:
op = LLVMAtomicRMWBinOpXchg;
break;
+#if LLVM_VERSION_MAJOR >= 10
+ case nir_intrinsic_shared_atomic_fadd:
+ case nir_intrinsic_deref_atomic_fadd:
+ op = LLVMAtomicRMWBinOpFAdd;
+ break;
+#endif
default:
return NULL;
}
- result = ac_build_atomic_rmw(&ctx->ac, op, ptr, ac_to_integer(&ctx->ac, src), sync_scope);
+ LLVMValueRef val;
+
+ if (instr->intrinsic == nir_intrinsic_shared_atomic_fadd ||
+ instr->intrinsic == nir_intrinsic_deref_atomic_fadd) {
+ val = ac_to_float(&ctx->ac, src);
+ } else {
+ val = ac_to_integer(&ctx->ac, src);
+ }
+
+ result = ac_build_atomic_rmw(&ctx->ac, op, ptr, val, sync_scope);
}
if (ctx->ac.postponed_kill)
result = visit_image_size(ctx, instr, false);
break;
case nir_intrinsic_shader_clock:
- result = ac_build_shader_clock(&ctx->ac);
+ result = ac_build_shader_clock(&ctx->ac,
+ nir_intrinsic_memory_scope(instr));
break;
case nir_intrinsic_discard:
case nir_intrinsic_discard_if:
case nir_intrinsic_memory_barrier_shared:
emit_membar(&ctx->ac, instr);
break;
+ case nir_intrinsic_scoped_barrier: {
+ assert(!(nir_intrinsic_memory_semantics(instr) &
+ (NIR_MEMORY_MAKE_AVAILABLE | NIR_MEMORY_MAKE_VISIBLE)));
+
+ nir_variable_mode modes = nir_intrinsic_memory_modes(instr);
+
+ unsigned wait_flags = 0;
+ if (modes & (nir_var_mem_global | nir_var_mem_ssbo))
+ wait_flags |= AC_WAIT_VLOAD | AC_WAIT_VSTORE;
+ if (modes & nir_var_mem_shared)
+ wait_flags |= AC_WAIT_LGKM;
+
+ if (wait_flags)
+ ac_build_waitcnt(&ctx->ac, wait_flags);
+
+ if (nir_intrinsic_execution_scope(instr) == NIR_SCOPE_WORKGROUP)
+ ac_emit_barrier(&ctx->ac, ctx->stage);
+ break;
+ }
case nir_intrinsic_memory_barrier_tcs_patch:
break;
case nir_intrinsic_control_barrier:
case nir_intrinsic_shared_atomic_or:
case nir_intrinsic_shared_atomic_xor:
case nir_intrinsic_shared_atomic_exchange:
- case nir_intrinsic_shared_atomic_comp_swap: {
+ case nir_intrinsic_shared_atomic_comp_swap:
+ case nir_intrinsic_shared_atomic_fadd: {
LLVMValueRef ptr = get_memory_ptr(ctx, instr->src[0],
instr->src[1].ssa->bit_size);
result = visit_var_atomic(ctx, instr, ptr, 1);
case nir_intrinsic_deref_atomic_or:
case nir_intrinsic_deref_atomic_xor:
case nir_intrinsic_deref_atomic_exchange:
- case nir_intrinsic_deref_atomic_comp_swap: {
+ case nir_intrinsic_deref_atomic_comp_swap:
+ case nir_intrinsic_deref_atomic_fadd: {
LLVMValueRef ptr = get_src(ctx, instr->src[0]);
result = visit_var_atomic(ctx, instr, ptr, 1);
break;
case nir_intrinsic_shuffle:
if (ctx->ac.chip_class == GFX8 ||
ctx->ac.chip_class == GFX9 ||
- (ctx->ac.chip_class == GFX10 && ctx->ac.wave_size == 32)) {
+ (ctx->ac.chip_class >= GFX10 && ctx->ac.wave_size == 32)) {
result = ac_build_shuffle(&ctx->ac, get_src(ctx, instr->src[0]),
get_src(ctx, instr->src[1]));
} else {
offset_src = i;
break;
case nir_tex_src_bias:
- if (instr->op == nir_texop_txb)
- args.bias = get_src(ctx, instr->src[i].src);
+ args.bias = get_src(ctx, instr->src[i].src);
break;
case nir_tex_src_lod: {
if (nir_src_is_const(instr->src[i].src) && nir_src_as_uint(instr->src[i].src) == 0)
}
}
+ assert(instr->dest.is_ssa);
+ args.d16 = instr->dest.ssa.bit_size == 16;
+
result = build_tex_intrinsic(ctx, instr, &args);
if (instr->op == nir_texop_query_levels)
LLVMTypeRef type = LLVMPointerType(pointee_type, address_space);
if (LLVMTypeOf(result) != type) {
- if (LLVMGetTypeKind(LLVMTypeOf(result)) == LLVMFixedVectorTypeKind) {
+ if (LLVMGetTypeKind(LLVMTypeOf(result)) == LLVMVectorTypeKind) {
result = LLVMBuildBitCast(ctx->ac.builder, result,
type, "");
} else {
{
int i, j;
ctx->num_locals = 0;
- nir_foreach_variable(variable, &func->impl->locals) {
+ nir_foreach_function_temp_variable(variable, func->impl) {
unsigned attrib_count = glsl_count_attribute_slots(variable->type, false);
variable->data.driver_location = ctx->num_locals * 4;
variable->data.location_frac = 0;
ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
- nir_foreach_variable(variable, &nir->outputs)
+ nir_foreach_shader_out_variable(variable, nir)
ac_handle_shader_output_decl(&ctx.ac, ctx.abi, nir, variable,
ctx.stage);
projects / mesa.git / blobdiff