if (store) { \
i.src[0] = ssa; \
i.src_types[0] = T; \
+ i.dest_type = T; \
} else { \
i.dest = ssa; \
i.dest_type = T; \
#define M_LOAD(name, T) M_LOAD_STORE(name, false, T)
#define M_STORE(name, T) M_LOAD_STORE(name, true, T)
-/* Inputs a NIR ALU source, with modifiers attached if necessary, and outputs
- * the corresponding Midgard source */
-
-static midgard_vector_alu_src
-vector_alu_modifiers(bool abs, bool neg, bool is_int,
- bool half, bool sext)
-{
- /* Figure out how many components there are so we can adjust.
- * Specifically we want to broadcast the last channel so things like
- * ball2/3 work.
- */
-
- midgard_vector_alu_src alu_src = {
- .rep_low = 0,
- .rep_high = 0,
- .half = half
- };
-
- if (is_int) {
- alu_src.mod = midgard_int_normal;
-
- /* Sign/zero-extend if needed */
-
- if (half) {
- alu_src.mod = sext ?
- midgard_int_sign_extend
- : midgard_int_zero_extend;
- }
-
- /* These should have been lowered away */
- assert(!(abs || neg));
- } else {
- alu_src.mod = (abs << 0) | (neg << 1);
- }
-
- return alu_src;
-}
-
M_LOAD(ld_attr_32, nir_type_uint32);
M_LOAD(ld_vary_32, nir_type_uint32);
M_LOAD(ld_ubo_int4, nir_type_uint32);
assert(src_bitsize == dst_bitsize); \
break;
+#define ALU_CHECK_CMP(sext) \
+ assert(src_bitsize == 16 || src_bitsize == 32); \
+ assert(dst_bitsize == 16 || dst_bitsize == 32); \
+
#define ALU_CASE_BCAST(nir, _op, count) \
case nir_op_##nir: \
op = midgard_alu_op_##_op; \
broadcast_swizzle = count; \
- assert(src_bitsize == dst_bitsize); \
+ ALU_CHECK_CMP(true); \
break;
-#define ALU_CHECK_CMP(sext) \
- if (src_bitsize == 16 && dst_bitsize == 32) { \
- half_1 = true; \
- half_2 = true; \
- sext_1 = sext; \
- sext_2 = sext; \
- } else { \
- assert(src_bitsize == dst_bitsize); \
- } \
-
#define ALU_CASE_CMP(nir, _op, sext) \
case nir_op_##nir: \
op = midgard_alu_op_##_op; \
unsigned dst_bitsize = nir_dest_bit_size(instr->dest.dest);
unsigned max_bitsize = MAX2(src_bitsize, dst_bitsize);
+ /* We don't have fp16 LUTs, so we'll want to emit code like:
+ *
+ * vlut.fsinr hr0, hr0
+ *
+ * where both input and output are 16-bit but the operation is carried
+ * out in 32-bit
+ */
+
+ switch (instr->op) {
+ case nir_op_fsqrt:
+ case nir_op_frcp:
+ case nir_op_frsq:
+ case nir_op_fsin:
+ case nir_op_fcos:
+ case nir_op_fexp2:
+ case nir_op_flog2:
+ max_bitsize = MAX2(max_bitsize, 32);
+ default:
+ break;
+ }
+
switch (max_bitsize) {
case 8:
return midgard_reg_mode_8;
{
switch (op) {
case nir_op_ior:
- case nir_op_iand:
+ case nir_op_iand: /* TODO: b2f16 */
case nir_op_ixor:
return true;
case nir_op_b32csel:
midgard_reg_mode reg_mode =
reg_mode_for_nir(instr);
- /* Do we need a destination override? Used for inline
- * type conversion */
-
- midgard_dest_override dest_override =
- midgard_dest_override_none;
-
- /* Should we use a smaller respective source and sign-extend? */
-
- bool half_1 = false, sext_1 = false;
- bool half_2 = false, sext_2 = false;
-
/* Should we swap arguments? */
bool flip_src12 = false;
*/
ALU_CASE_CMP(b2f32, iand, true);
+ ALU_CASE_CMP(b2f16, iand, true);
ALU_CASE_CMP(b2i32, iand, true);
/* Likewise, we don't have a dedicated f2b32 instruction, but
ALU_CASE_BCAST(b32all_fequal2, fball_eq, 2);
ALU_CASE_BCAST(b32all_fequal3, fball_eq, 3);
- ALU_CASE(b32all_fequal4, fball_eq);
+ ALU_CASE_CMP(b32all_fequal4, fball_eq, true);
ALU_CASE_BCAST(b32any_fnequal2, fbany_neq, 2);
ALU_CASE_BCAST(b32any_fnequal3, fbany_neq, 3);
- ALU_CASE(b32any_fnequal4, fbany_neq);
+ ALU_CASE_CMP(b32any_fnequal4, fbany_neq, true);
ALU_CASE_BCAST(b32all_iequal2, iball_eq, 2);
ALU_CASE_BCAST(b32all_iequal3, iball_eq, 3);
- ALU_CASE(b32all_iequal4, iball_eq);
+ ALU_CASE_CMP(b32all_iequal4, iball_eq, true);
ALU_CASE_BCAST(b32any_inequal2, ibany_neq, 2);
ALU_CASE_BCAST(b32any_inequal3, ibany_neq, 3);
- ALU_CASE(b32any_inequal4, ibany_neq);
+ ALU_CASE_CMP(b32any_inequal4, ibany_neq, true);
/* Source mods will be shoved in later */
ALU_CASE(fabs, fmov);
case nir_op_i2i16:
case nir_op_i2i32:
case nir_op_i2i64:
- /* If we end up upscale, we'll need a sign-extend on the
- * operand (the second argument) */
-
- sext_2 = true;
- /* fallthrough */
case nir_op_u2u8:
case nir_op_u2u16:
case nir_op_u2u32:
else
op = midgard_alu_op_imov;
- if (dst_bitsize == (src_bitsize * 2)) {
- /* Converting up */
- half_2 = true;
- } else if (src_bitsize == (dst_bitsize * 2)) {
- /* Converting down */
- dest_override = midgard_dest_override_lower;
- }
-
break;
}
return;
}
+ /* Promote imov to fmov if it might help inline a constant */
+ if (op == midgard_alu_op_imov && nir_src_is_const(instr->src[0].src)
+ && nir_src_bit_size(instr->src[0].src) == 32
+ && nir_is_same_comp_swizzle(instr->src[0].swizzle,
+ nir_src_num_components(instr->src[0].src))) {
+ op = midgard_alu_op_fmov;
+ }
+
/* Midgard can perform certain modifiers on output of an ALU op */
unsigned outmod = 0;
-
- bool abs[4] = { false };
- bool neg[4] = { false };
bool is_int = midgard_is_integer_op(op);
if (midgard_is_integer_out_op(op)) {
if (quirk_flipped_r24) {
ins.src[0] = ~0;
- mir_copy_src(&ins, instr, 0, 1, &abs[1], &neg[1], &ins.src_invert[1], is_int, broadcast_swizzle);
+ mir_copy_src(&ins, instr, 0, 1, &ins.src_abs[1], &ins.src_neg[1], &ins.src_invert[1], is_int, broadcast_swizzle);
} else {
for (unsigned i = 0; i < nr_inputs; ++i) {
unsigned to = i;
to = 1 - to;
}
- mir_copy_src(&ins, instr, i, to, &abs[to], &neg[to], &ins.src_invert[to], is_int, broadcast_swizzle);
+ mir_copy_src(&ins, instr, i, to, &ins.src_abs[to], &ins.src_neg[to], &ins.src_invert[to], is_int, broadcast_swizzle);
/* (!c) ? a : b = c ? b : a */
if (instr->op == nir_op_b32csel && ins.src_invert[2]) {
if (instr->op == nir_op_fneg || instr->op == nir_op_fabs) {
/* Lowered to move */
if (instr->op == nir_op_fneg)
- neg[1] = !neg[1];
+ ins.src_neg[1] ^= true;
if (instr->op == nir_op_fabs)
- abs[1] = true;
+ ins.src_abs[1] = true;
}
ins.mask = mask_of(nr_components);
midgard_vector_alu alu = {
.op = op,
.reg_mode = reg_mode,
- .dest_override = dest_override,
.outmod = outmod,
-
- .src1 = vector_alu_srco_unsigned(vector_alu_modifiers(abs[0], neg[0], is_int, half_1, sext_1)),
- .src2 = vector_alu_srco_unsigned(vector_alu_modifiers(abs[1], neg[1], is_int, half_2, sext_2)),
};
/* Apply writemask if non-SSA, keeping in mind that we can't write to
ins.alu = alu;
- /* Arrange for creation of iandnot/iornot */
- if (ins.src_invert[0] && !ins.src_invert[1]) {
- mir_flip(&ins);
- ins.src_invert[0] = false;
- ins.src_invert[1] = true;
- }
-
/* Late fixup for emulated instructions */
if (instr->op == nir_op_b2f32 || instr->op == nir_op_b2i32) {
else
ins.constants.i32[0] = 1;
+ for (unsigned c = 0; c < 16; ++c)
+ ins.swizzle[1][c] = 0;
+ } else if (instr->op == nir_op_b2f16) {
+ ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
+ ins.src_types[1] = nir_type_float16;
+ ins.has_constants = true;
+ ins.constants.i16[0] = _mesa_float_to_half(1.0);
+
for (unsigned c = 0; c < 16; ++c)
ins.swizzle[1][c] = 0;
} else if (nr_inputs == 1 && !quirk_flipped_r24) {
ins.swizzle[1][c] = 0;
}
+ /* Arrange for creation of iandnot/iornot */
+ if (ins.src_invert[0] && !ins.src_invert[1]) {
+ mir_flip(&ins);
+ ins.src_invert[0] = false;
+ ins.src_invert[1] = true;
+ }
+
if ((opcode_props & UNITS_ALL) == UNIT_VLUT) {
/* To avoid duplicating the lookup tables (probably), true LUT
* instructions can only operate as if they were scalars. Lower
/* Once we have the NIR mask, we need to normalize to work in 32-bit space */
unsigned bytemask = pan_to_bytemask(dsize, nir_mask);
mir_set_bytemask(ins, bytemask);
-
- if (dsize == 64)
- ins->load_64 = true;
+ ins->dest_type = nir_type_uint | dsize;
}
/* Uniforms and UBOs use a shared code path, as uniforms are just (slightly
{
midgard_instruction ins = {
.type = TAG_TEXTURE_4,
+ .dest = ~0,
.src = { ~0, ~0, ~0, ~0 },
.texture = {
.op = TEXTURE_OP_BARRIER,
ld.load_store.arg_2 = 0x1E;
}
- for (unsigned c = 2; c < 16; ++c)
+ for (unsigned c = 4; c < 16; ++c)
ld.swizzle[0][c] = 0;
+
+ ld.dest_type = nir_type_float16;
}
emit_mir_instruction(ctx, ld);
}
}
-/* Tries to attach an explicit LOD / bias as a constant. Returns whether this
+/* Tries to attach an explicit LOD or bias as a constant. Returns whether this
* was successful */
static bool
}
if (ins->src[1] == SSA_FIXED_REGISTER(REGISTER_CONSTANT)) {
- /* Extract the source information */
-
- midgard_vector_alu_src *src;
- int q = ins->alu.src2;
- midgard_vector_alu_src *m = (midgard_vector_alu_src *) &q;
- src = m;
-
/* Component is from the swizzle. Take a nonzero component */
assert(ins->mask);
unsigned first_comp = ffs(ins->mask) - 1;
continue;
}
- /* We don't know how to handle these with a constant */
-
- if (mir_nontrivial_source2_mod_simple(ins) || src->rep_low || src->rep_high) {
- DBG("Bailing inline constant...\n");
+ /* Should've been const folded */
+ if (ins->src_abs[1] || ins->src_neg[1])
continue;
- }
/* Make sure that the constant is not itself a vector
* by checking if all accessed values are the same. */
mir_foreach_block(ctx, _block) {
midgard_block *block = (midgard_block *) _block;
inline_alu_constants(ctx, block);
- midgard_opt_promote_fmov(ctx, block);
embedded_to_inline_constant(ctx, block);
}
/* MIR-level optimizations */