#include "helpers.h"
#include "compiler.h"
#include "midgard_quirks.h"
+#include "panfrost-quirks.h"
+#include "panfrost/util/pan_lower_framebuffer.h"
#include "disassemble.h"
.dest = ~0, \
.src = { ~0, ~0, ~0, ~0 }, \
.swizzle = SWIZZLE_IDENTITY_4, \
+ .op = midgard_op_##name, \
.load_store = { \
- .op = midgard_op_##name, \
.address = address \
} \
}; \
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);
M_LOAD(ld_int4, nir_type_uint32);
M_STORE(st_int4, nir_type_uint32);
M_LOAD(ld_color_buffer_32u, nir_type_uint32);
+M_LOAD(ld_color_buffer_as_fp16, nir_type_float16);
+M_LOAD(ld_color_buffer_as_fp32, nir_type_float32);
M_STORE(st_vary_32, nir_type_uint32);
M_LOAD(ld_cubemap_coords, nir_type_uint32);
M_LOAD(ld_compute_id, nir_type_uint32);
return ins;
}
-static midgard_branch_extended
-midgard_create_branch_extended( midgard_condition cond,
- midgard_jmp_writeout_op op,
- unsigned dest_tag,
- signed quadword_offset)
-{
- /* The condition code is actually a LUT describing a function to
- * combine multiple condition codes. However, we only support a single
- * condition code at the moment, so we just duplicate over a bunch of
- * times. */
-
- uint16_t duplicated_cond =
- (cond << 14) |
- (cond << 12) |
- (cond << 10) |
- (cond << 8) |
- (cond << 6) |
- (cond << 4) |
- (cond << 2) |
- (cond << 0);
-
- midgard_branch_extended branch = {
- .op = op,
- .dest_tag = dest_tag,
- .offset = quadword_offset,
- .cond = duplicated_cond
- };
-
- return branch;
-}
-
static void
attach_constants(compiler_context *ctx, midgard_instruction *ins, void *constants, int name)
{
return progress;
}
-/* Midgard can't write depth and stencil separately. It has to happen in a
- * single store operation containing both. Let's add a panfrost specific
- * intrinsic and turn all depth/stencil stores into a packed depth+stencil
- * one.
+static const nir_variable *
+search_var(nir_shader *nir, nir_variable_mode mode, unsigned driver_loc)
+{
+ nir_foreach_variable_with_modes(var, nir, mode) {
+ if (var->data.driver_location == driver_loc)
+ return var;
+ }
+
+ return NULL;
+}
+
+/* Midgard can write all of color, depth and stencil in a single writeout
+ * operation, so we merge depth/stencil stores with color stores.
+ * If there are no color stores, we add a write to the "depth RT".
*/
static bool
midgard_nir_lower_zs_store(nir_shader *nir)
nir_variable *z_var = NULL, *s_var = NULL;
- nir_foreach_variable(var, &nir->outputs) {
+ nir_foreach_shader_out_variable(var, nir) {
if (var->data.location == FRAG_RESULT_DEPTH)
z_var = var;
else if (var->data.location == FRAG_RESULT_STENCIL)
nir_foreach_function(function, nir) {
if (!function->impl) continue;
- nir_intrinsic_instr *z_store = NULL, *s_store = NULL, *last_store = NULL;
+ nir_intrinsic_instr *z_store = NULL, *s_store = NULL;
nir_foreach_block(block, function->impl) {
nir_foreach_instr_safe(instr, block) {
if (z_var && nir_intrinsic_base(intr) == z_var->data.driver_location) {
assert(!z_store);
z_store = intr;
- last_store = intr;
}
if (s_var && nir_intrinsic_base(intr) == s_var->data.driver_location) {
assert(!s_store);
s_store = intr;
- last_store = intr;
}
}
}
if (!z_store && !s_store) continue;
- nir_builder b;
- nir_builder_init(&b, function->impl);
+ bool replaced = false;
- b.cursor = nir_before_instr(&last_store->instr);
+ nir_foreach_block(block, function->impl) {
+ nir_foreach_instr_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
- nir_ssa_def *zs_store_src;
+ nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+ if (intr->intrinsic != nir_intrinsic_store_output)
+ continue;
- if (z_store && s_store) {
- nir_ssa_def *srcs[2] = {
- nir_ssa_for_src(&b, z_store->src[0], 1),
- nir_ssa_for_src(&b, s_store->src[0], 1),
- };
+ const nir_variable *var = search_var(nir, nir_var_shader_out, nir_intrinsic_base(intr));
+ assert(var);
- zs_store_src = nir_vec(&b, srcs, 2);
- } else {
- zs_store_src = nir_ssa_for_src(&b, last_store->src[0], 1);
+ if (var->data.location != FRAG_RESULT_COLOR &&
+ var->data.location < FRAG_RESULT_DATA0)
+ continue;
+
+ if (var->data.index)
+ continue;
+
+ assert(nir_src_is_const(intr->src[1]) && "no indirect outputs");
+
+ nir_builder b;
+ nir_builder_init(&b, function->impl);
+
+ assert(!z_store || z_store->instr.block == instr->block);
+ assert(!s_store || s_store->instr.block == instr->block);
+ b.cursor = nir_after_block_before_jump(instr->block);
+
+ nir_intrinsic_instr *combined_store;
+ combined_store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_combined_output_pan);
+
+ combined_store->num_components = intr->src[0].ssa->num_components;
+
+ nir_intrinsic_set_base(combined_store, nir_intrinsic_base(intr));
+
+ unsigned writeout = PAN_WRITEOUT_C;
+ if (z_store)
+ writeout |= PAN_WRITEOUT_Z;
+ if (s_store)
+ writeout |= PAN_WRITEOUT_S;
+
+ nir_intrinsic_set_component(combined_store, writeout);
+
+ struct nir_ssa_def *zero = nir_imm_int(&b, 0);
+
+ struct nir_ssa_def *src[4] = {
+ intr->src[0].ssa,
+ intr->src[1].ssa,
+ z_store ? z_store->src[0].ssa : zero,
+ s_store ? s_store->src[0].ssa : zero,
+ };
+
+ for (int i = 0; i < 4; ++i)
+ combined_store->src[i] = nir_src_for_ssa(src[i]);
+
+ nir_builder_instr_insert(&b, &combined_store->instr);
+
+ nir_instr_remove(instr);
+
+ replaced = true;
+ }
}
- nir_intrinsic_instr *zs_store;
+ /* Insert a store to the depth RT (0xff) if needed */
+ if (!replaced) {
+ nir_builder b;
+ nir_builder_init(&b, function->impl);
- zs_store = nir_intrinsic_instr_create(b.shader,
- nir_intrinsic_store_zs_output_pan);
- zs_store->src[0] = nir_src_for_ssa(zs_store_src);
- zs_store->num_components = z_store && s_store ? 2 : 1;
- nir_intrinsic_set_component(zs_store, z_store ? 0 : 1);
+ nir_block *block = NULL;
+ if (z_store && s_store)
+ assert(z_store->instr.block == s_store->instr.block);
- /* Replace the Z and S store by a ZS store */
- nir_builder_instr_insert(&b, &zs_store->instr);
+ if (z_store)
+ block = z_store->instr.block;
+ else
+ block = s_store->instr.block;
+
+ b.cursor = nir_after_block_before_jump(block);
+
+ nir_intrinsic_instr *combined_store;
+ combined_store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_combined_output_pan);
+
+ combined_store->num_components = 4;
+
+ unsigned base;
+ if (z_store)
+ base = nir_intrinsic_base(z_store);
+ else
+ base = nir_intrinsic_base(s_store);
+ nir_intrinsic_set_base(combined_store, base);
+
+ unsigned writeout = 0;
+ if (z_store)
+ writeout |= PAN_WRITEOUT_Z;
+ if (s_store)
+ writeout |= PAN_WRITEOUT_S;
+
+ nir_intrinsic_set_component(combined_store, writeout);
+
+ struct nir_ssa_def *zero = nir_imm_int(&b, 0);
+
+ struct nir_ssa_def *src[4] = {
+ nir_imm_vec4(&b, 0, 0, 0, 0),
+ zero,
+ z_store ? z_store->src[0].ssa : zero,
+ s_store ? s_store->src[0].ssa : zero,
+ };
+
+ for (int i = 0; i < 4; ++i)
+ combined_store->src[i] = nir_src_for_ssa(src[i]);
+
+ nir_builder_instr_insert(&b, &combined_store->instr);
+ }
if (z_store)
nir_instr_remove(&z_store->instr);
return progress;
}
+/* Real writeout stores, which break execution, need to be moved to after
+ * dual-source stores, which are just standard register writes. */
+static bool
+midgard_nir_reorder_writeout(nir_shader *nir)
+{
+ bool progress = false;
+
+ nir_foreach_function(function, nir) {
+ if (!function->impl) continue;
+
+ nir_foreach_block(block, function->impl) {
+ nir_instr *last_writeout = NULL;
+
+ nir_foreach_instr_reverse_safe(instr, block) {
+ if (instr->type != nir_instr_type_intrinsic)
+ continue;
+
+ nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
+ if (intr->intrinsic != nir_intrinsic_store_output)
+ continue;
+
+ const nir_variable *var = search_var(nir, nir_var_shader_out, nir_intrinsic_base(intr));
+
+ if (var->data.index) {
+ if (!last_writeout)
+ last_writeout = instr;
+ continue;
+ }
+
+ if (!last_writeout)
+ continue;
+
+ /* This is a real store, so move it to after dual-source stores */
+ exec_node_remove(&instr->node);
+ exec_node_insert_after(&last_writeout->node, &instr->node);
+
+ progress = true;
+ }
+ }
+ }
+
+ return progress;
+}
+
/* Flushes undefined values to zero */
static void
-optimise_nir(nir_shader *nir, unsigned quirks)
+optimise_nir(nir_shader *nir, unsigned quirks, bool is_blend)
{
bool progress;
unsigned lower_flrp =
if (quirks & MIDGARD_BROKEN_LOD)
NIR_PASS_V(nir, midgard_nir_lod_errata);
+ NIR_PASS(progress, nir, midgard_nir_lower_algebraic_early);
+
do {
progress = false;
NIR_PASS(progress, nir, nir_opt_vectorize);
} while (progress);
+ /* Run after opts so it can hit more */
+ if (!is_blend)
+ NIR_PASS(progress, nir, nir_fuse_io_16);
+
/* Must be run at the end to prevent creation of fsin/fcos ops */
NIR_PASS(progress, nir, midgard_nir_scale_trig);
assert(src_bitsize == dst_bitsize); \
break;
+#define ALU_CASE_RTZ(nir, _op) \
+ case nir_op_##nir: \
+ op = midgard_alu_op_##_op; \
+ roundmode = MIDGARD_RTZ; \
+ 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;
-/* Analyze the sizes of the inputs to determine which reg mode. Ops needed
- * special treatment override this anyway. */
-static midgard_reg_mode
-reg_mode_for_nir(nir_alu_instr *instr)
-{
- unsigned src_bitsize = nir_src_bit_size(instr->src[0].src);
-
- switch (src_bitsize) {
- case 8:
- return midgard_reg_mode_8;
- case 16:
- return midgard_reg_mode_16;
- case 32:
- return midgard_reg_mode_32;
- case 64:
- return midgard_reg_mode_64;
- default:
- unreachable("Invalid bit size");
- }
-}
+#define ALU_CASE_CMP(nir, _op, sext) \
+ case nir_op_##nir: \
+ op = midgard_alu_op_##_op; \
+ ALU_CHECK_CMP(sext); \
+ break;
/* Compare mir_lower_invert */
static bool
{
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:
return false;
}
+/* Look for floating point mods. We have the mods fsat, fsat_signed,
+ * and fpos. We also have the relations (note 3 * 2 = 6 cases):
+ *
+ * fsat_signed(fpos(x)) = fsat(x)
+ * fsat_signed(fsat(x)) = fsat(x)
+ * fpos(fsat_signed(x)) = fsat(x)
+ * fpos(fsat(x)) = fsat(x)
+ * fsat(fsat_signed(x)) = fsat(x)
+ * fsat(fpos(x)) = fsat(x)
+ *
+ * So by cases any composition of output modifiers is equivalent to
+ * fsat alone.
+ */
+static unsigned
+mir_determine_float_outmod(compiler_context *ctx, nir_dest **dest, unsigned prior_outmod)
+{
+ bool fpos = mir_accept_dest_mod(ctx, dest, nir_op_fclamp_pos);
+ bool fsat = mir_accept_dest_mod(ctx, dest, nir_op_fsat);
+ bool ssat = mir_accept_dest_mod(ctx, dest, nir_op_fsat_signed);
+ bool prior = (prior_outmod != midgard_outmod_none);
+ int count = (int) prior + (int) fpos + (int) ssat + (int) fsat;
+
+ return ((count > 1) || fsat) ? midgard_outmod_sat :
+ fpos ? midgard_outmod_pos :
+ ssat ? midgard_outmod_sat_signed :
+ prior_outmod;
+}
+
static void
-mir_copy_src(midgard_instruction *ins, nir_alu_instr *instr, unsigned i, unsigned to, bool *abs, bool *neg, bool *not, bool is_int, unsigned bcast_count)
+mir_copy_src(midgard_instruction *ins, nir_alu_instr *instr, unsigned i, unsigned to, bool *abs, bool *neg, bool *not, enum midgard_roundmode *roundmode, bool is_int, unsigned bcast_count)
{
nir_alu_src src = instr->src[i];
if (nir_accepts_inot(instr->op, i) && pan_has_source_mod(&src, nir_op_inot))
*not = true;
+ if (roundmode) {
+ if (pan_has_source_mod(&src, nir_op_fround_even))
+ *roundmode = MIDGARD_RTE;
+
+ if (pan_has_source_mod(&src, nir_op_ftrunc))
+ *roundmode = MIDGARD_RTZ;
+
+ if (pan_has_source_mod(&src, nir_op_ffloor))
+ *roundmode = MIDGARD_RTN;
+
+ if (pan_has_source_mod(&src, nir_op_fceil))
+ *roundmode = MIDGARD_RTP;
+ }
+
unsigned bits = nir_src_bit_size(src.src);
ins->src[to] = nir_src_index(NULL, &src.src);
}
}
+/* Midgard features both fcsel and icsel, depending on whether you want int or
+ * float modifiers. NIR's csel is typeless, so we want a heuristic to guess if
+ * we should emit an int or float csel depending on what modifiers could be
+ * placed. In the absense of modifiers, this is probably arbitrary. */
+
+static bool
+mir_is_bcsel_float(nir_alu_instr *instr)
+{
+ nir_op intmods[] = {
+ nir_op_i2i8, nir_op_i2i16,
+ nir_op_i2i32, nir_op_i2i64
+ };
+
+ nir_op floatmods[] = {
+ nir_op_fabs, nir_op_fneg,
+ nir_op_f2f16, nir_op_f2f32,
+ nir_op_f2f64
+ };
+
+ nir_op floatdestmods[] = {
+ nir_op_fsat, nir_op_fsat_signed, nir_op_fclamp_pos,
+ nir_op_f2f16, nir_op_f2f32
+ };
+
+ signed score = 0;
+
+ for (unsigned i = 1; i < 3; ++i) {
+ nir_alu_src s = instr->src[i];
+ for (unsigned q = 0; q < ARRAY_SIZE(intmods); ++q) {
+ if (pan_has_source_mod(&s, intmods[q]))
+ score--;
+ }
+ }
+
+ for (unsigned i = 1; i < 3; ++i) {
+ nir_alu_src s = instr->src[i];
+ for (unsigned q = 0; q < ARRAY_SIZE(floatmods); ++q) {
+ if (pan_has_source_mod(&s, floatmods[q]))
+ score++;
+ }
+ }
+
+ for (unsigned q = 0; q < ARRAY_SIZE(floatdestmods); ++q) {
+ nir_dest *dest = &instr->dest.dest;
+ if (pan_has_dest_mod(&dest, floatdestmods[q]))
+ score++;
+ }
+
+ return (score > 0);
+}
+
static void
emit_alu(compiler_context *ctx, nir_alu_instr *instr)
{
unsigned broadcast_swizzle = 0;
- /* What register mode should we operate in? */
- 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;
- unsigned src_bitsize = nir_src_bit_size(instr->src[0].src);
- unsigned dst_bitsize = nir_dest_bit_size(*dest);
+ ASSERTED unsigned src_bitsize = nir_src_bit_size(instr->src[0].src);
+ ASSERTED unsigned dst_bitsize = nir_dest_bit_size(*dest);
+
+ enum midgard_roundmode roundmode = MIDGARD_RTE;
switch (instr->op) {
ALU_CASE(fadd, fadd);
ALU_CASE(iadd, iadd);
ALU_CASE(isub, isub);
ALU_CASE(imul, imul);
+ ALU_CASE(imul_high, imul);
+ ALU_CASE(umul_high, imul);
/* Zero shoved as second-arg */
ALU_CASE(iabs, iabsdiff);
ALU_CASE(mov, imov);
- ALU_CASE(feq32, feq);
- ALU_CASE(fne32, fne);
- ALU_CASE(flt32, flt);
- ALU_CASE(ieq32, ieq);
- ALU_CASE(ine32, ine);
- ALU_CASE(ilt32, ilt);
- ALU_CASE(ult32, ult);
+ ALU_CASE_CMP(feq32, feq, false);
+ ALU_CASE_CMP(fneu32, fne, false);
+ ALU_CASE_CMP(flt32, flt, false);
+ ALU_CASE_CMP(ieq32, ieq, true);
+ ALU_CASE_CMP(ine32, ine, true);
+ ALU_CASE_CMP(ilt32, ilt, true);
+ ALU_CASE_CMP(ult32, ult, false);
/* We don't have a native b2f32 instruction. Instead, like many
* GPUs, we exploit booleans as 0/~0 for false/true, and
* At the end of emit_alu (as MIR), we'll fix-up the constant
*/
- ALU_CASE(b2f32, iand);
- ALU_CASE(b2i32, iand);
+ 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
* we can do a "not equal to 0.0" test. */
- ALU_CASE(f2b32, fne);
- ALU_CASE(i2b32, ine);
+ ALU_CASE_CMP(f2b32, fne, false);
+ ALU_CASE_CMP(i2b32, ine, true);
ALU_CASE(frcp, frcp);
ALU_CASE(frsq, frsqrt);
ALU_CASE(fexp2, fexp2);
ALU_CASE(flog2, flog2);
- ALU_CASE(f2i64, f2i_rtz);
- ALU_CASE(f2u64, f2u_rtz);
- ALU_CASE(i2f64, i2f_rtz);
- ALU_CASE(u2f64, u2f_rtz);
+ ALU_CASE_RTZ(f2i64, f2i_rte);
+ ALU_CASE_RTZ(f2u64, f2u_rte);
+ ALU_CASE_RTZ(i2f64, i2f_rte);
+ ALU_CASE_RTZ(u2f64, u2f_rte);
+
+ ALU_CASE_RTZ(f2i32, f2i_rte);
+ ALU_CASE_RTZ(f2u32, f2u_rte);
+ ALU_CASE_RTZ(i2f32, i2f_rte);
+ ALU_CASE_RTZ(u2f32, u2f_rte);
- ALU_CASE(f2i32, f2i_rtz);
- ALU_CASE(f2u32, f2u_rtz);
- ALU_CASE(i2f32, i2f_rtz);
- ALU_CASE(u2f32, u2f_rtz);
+ ALU_CASE_RTZ(f2i8, f2i_rte);
+ ALU_CASE_RTZ(f2u8, f2u_rte);
- ALU_CASE(f2i16, f2i_rtz);
- ALU_CASE(f2u16, f2u_rtz);
- ALU_CASE(i2f16, i2f_rtz);
- ALU_CASE(u2f16, u2f_rtz);
+ ALU_CASE_RTZ(f2i16, f2i_rte);
+ ALU_CASE_RTZ(f2u16, f2u_rte);
+ ALU_CASE_RTZ(i2f16, i2f_rte);
+ ALU_CASE_RTZ(u2f16, u2f_rte);
ALU_CASE(fsin, fsin);
ALU_CASE(fcos, fcos);
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;
-
- /* Use a greater register mode */
- reg_mode++;
- } else if (src_bitsize == (dst_bitsize * 2)) {
- /* Converting down */
- dest_override = midgard_dest_override_lower;
- }
-
break;
}
0;
flip_src12 = true;
+ ALU_CHECK_CMP(false);
break;
}
case nir_op_b32csel: {
- /* Midgard features both fcsel and icsel, depending on
- * the type of the arguments/output. However, as long
- * as we're careful we can _always_ use icsel and
- * _never_ need fcsel, since the latter does additional
- * floating-point-specific processing whereas the
- * former just moves bits on the wire. It's not obvious
- * why these are separate opcodes, save for the ability
- * to do things like sat/pos/abs/neg for free */
-
bool mixed = nir_is_non_scalar_swizzle(&instr->src[0], nr_components);
- op = mixed ? midgard_alu_op_icsel_v : midgard_alu_op_icsel;
+ bool is_float = mir_is_bcsel_float(instr);
+ op = is_float ?
+ (mixed ? midgard_alu_op_fcsel_v : midgard_alu_op_fcsel) :
+ (mixed ? midgard_alu_op_icsel_v : midgard_alu_op_icsel);
+
+ break;
+ }
+ case nir_op_unpack_32_2x16:
+ case nir_op_unpack_32_4x8:
+ case nir_op_pack_32_2x16:
+ case nir_op_pack_32_4x8: {
+ op = midgard_alu_op_imov;
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 (instr->op == nir_op_umul_high || instr->op == nir_op_imul_high) {
+ outmod = midgard_outmod_int_high;
+ } else if (midgard_is_integer_out_op(op)) {
outmod = midgard_outmod_int_wrap;
} else if (instr->op == nir_op_fsat) {
outmod = midgard_outmod_sat;
unsigned opcode_props = alu_opcode_props[op].props;
bool quirk_flipped_r24 = opcode_props & QUIRK_FLIPPED_R24;
- /* Look for floating point mods. We have the mods fsat, fsat_signed,
- * and fpos. We also have the relations (note 3 * 2 = 6 cases):
- *
- * fsat_signed(fpos(x)) = fsat(x)
- * fsat_signed(fsat(x)) = fsat(x)
- * fpos(fsat_signed(x)) = fsat(x)
- * fpos(fsat(x)) = fsat(x)
- * fsat(fsat_signed(x)) = fsat(x)
- * fsat(fpos(x)) = fsat(x)
- *
- * So by cases any composition of output modifiers is equivalent to
- * fsat alone.
- */
-
- if (!is_int && !(opcode_props & OP_TYPE_CONVERT)) {
- bool fpos = mir_accept_dest_mod(ctx, &dest, nir_op_fclamp_pos);
- bool fsat = mir_accept_dest_mod(ctx, &dest, nir_op_fsat);
- bool ssat = mir_accept_dest_mod(ctx, &dest, nir_op_fsat_signed);
- bool prior = (outmod != midgard_outmod_none);
- int count = (int) prior + (int) fpos + (int) ssat + (int) fsat;
-
- outmod = ((count > 1) || fsat) ? midgard_outmod_sat :
- fpos ? midgard_outmod_pos :
- ssat ? midgard_outmod_sat_signed :
- outmod;
+ if (!midgard_is_integer_out_op(op)) {
+ outmod = mir_determine_float_outmod(ctx, &dest, outmod);
}
midgard_instruction ins = {
.dest = nir_dest_index(dest),
.dest_type = nir_op_infos[instr->op].output_type
| nir_dest_bit_size(*dest),
+ .roundmode = roundmode,
};
+ enum midgard_roundmode *roundptr = (opcode_props & MIDGARD_ROUNDS) ?
+ &ins.roundmode : NULL;
+
for (unsigned i = nr_inputs; i < ARRAY_SIZE(ins.src); ++i)
ins.src[i] = ~0;
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], roundptr, 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], roundptr, 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
* components that don't exist. Note modifier => SSA => !reg => no
* writemask, so we don't have to worry about writemasks here.*/
if (!is_ssa)
ins.mask &= instr->dest.write_mask;
- 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;
- }
+ ins.op = op;
+ ins.outmod = outmod;
/* Late fixup for emulated instructions */
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) {
/* Lots of instructions need a 0 plonked in */
ins.has_inline_constant = false;
ins.src[1] = SSA_FIXED_REGISTER(REGISTER_CONSTANT);
- ins.src_types[1] = nir_type_uint32;
+ ins.src_types[1] = ins.src_types[0];
ins.has_constants = true;
ins.constants.u32[0] = 0;
for (unsigned c = 0; c < 16; ++c)
ins.swizzle[1][c] = 0;
+ } else if (instr->op == nir_op_pack_32_2x16) {
+ ins.dest_type = nir_type_uint16;
+ ins.mask = mask_of(nr_components * 2);
+ ins.is_pack = true;
+ } else if (instr->op == nir_op_pack_32_4x8) {
+ ins.dest_type = nir_type_uint8;
+ ins.mask = mask_of(nr_components * 4);
+ ins.is_pack = true;
+ } else if (instr->op == nir_op_unpack_32_2x16) {
+ ins.dest_type = nir_type_uint32;
+ ins.mask = mask_of(nr_components >> 1);
+ ins.is_pack = true;
+ } else if (instr->op == nir_op_unpack_32_4x8) {
+ ins.dest_type = nir_type_uint32;
+ ins.mask = mask_of(nr_components >> 2);
+ ins.is_pack = true;
}
if ((opcode_props & UNITS_ALL) == UNIT_VLUT) {
unsigned swizzle_back[MIR_VEC_COMPONENTS];
memcpy(&swizzle_back, ins.swizzle[0], sizeof(swizzle_back));
+ midgard_instruction ins_split[MIR_VEC_COMPONENTS];
+ unsigned ins_count = 0;
+
for (int i = 0; i < nr_components; ++i) {
/* Mask the associated component, dropping the
* instruction if needed */
ins.mask = 1 << i;
ins.mask &= orig_mask;
+ for (unsigned j = 0; j < ins_count; ++j) {
+ if (swizzle_back[i] == ins_split[j].swizzle[0][0]) {
+ ins_split[j].mask |= ins.mask;
+ ins.mask = 0;
+ break;
+ }
+ }
+
if (!ins.mask)
continue;
for (unsigned j = 0; j < MIR_VEC_COMPONENTS; ++j)
ins.swizzle[0][j] = swizzle_back[i]; /* Pull from the correct component */
- emit_mir_instruction(ctx, ins);
+ ins_split[ins_count] = ins;
+
+ ++ins_count;
+ }
+
+ for (unsigned i = 0; i < ins_count; ++i) {
+ emit_mir_instruction(ctx, ins_split[i]);
}
} else {
emit_mir_instruction(ctx, ins);
/* 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
ins.src[2] = nir_src_index(ctx, indirect_offset);
ins.src_types[2] = nir_type_uint32;
ins.load_store.arg_2 = (indirect_shift << 5);
+
+ /* X component for the whole swizzle to prevent register
+ * pressure from ballooning from the extra components */
+ for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[2]); ++i)
+ ins.swizzle[2][i] = 0;
} else {
ins.load_store.arg_2 = 0x1E;
}
midgard_instruction ins = m_ld_vary_32(dest, offset);
ins.mask = mask_of(nr_comp);
+ ins.dest_type = type;
+
+ if (type == nir_type_float16) {
+ /* Ensure we are aligned so we can pack it later */
+ ins.mask = mask_of(ALIGN_POT(nr_comp, 2));
+ }
for (unsigned i = 0; i < ARRAY_SIZE(ins.swizzle[0]); ++i)
ins.swizzle[0][i] = MIN2(i + component, COMPONENT_W);
/* Use the type appropriate load */
switch (type) {
- case nir_type_uint:
- case nir_type_bool:
- ins.load_store.op = midgard_op_ld_vary_32u;
+ case nir_type_uint32:
+ case nir_type_bool32:
+ ins.op = midgard_op_ld_vary_32u;
break;
- case nir_type_int:
- ins.load_store.op = midgard_op_ld_vary_32i;
+ case nir_type_int32:
+ ins.op = midgard_op_ld_vary_32i;
break;
- case nir_type_float:
- ins.load_store.op = midgard_op_ld_vary_32;
+ case nir_type_float32:
+ ins.op = midgard_op_ld_vary_32;
+ break;
+ case nir_type_float16:
+ ins.op = midgard_op_ld_vary_16;
break;
default:
unreachable("Attempted to load unknown type");
switch (t) {
case nir_type_uint:
case nir_type_bool:
- ins.load_store.op = midgard_op_ld_attr_32u;
+ ins.op = midgard_op_ld_attr_32u;
break;
case nir_type_int:
- ins.load_store.op = midgard_op_ld_attr_32i;
+ ins.op = midgard_op_ld_attr_32i;
break;
case nir_type_float:
- ins.load_store.op = midgard_op_ld_attr_32;
+ ins.op = midgard_op_ld_attr_32;
break;
default:
unreachable("Attempted to load unknown type");
}
static void
-emit_fragment_store(compiler_context *ctx, unsigned src, enum midgard_rt_id rt)
+emit_fragment_store(compiler_context *ctx, unsigned src, unsigned src_z, unsigned src_s, enum midgard_rt_id rt)
{
assert(rt < ARRAY_SIZE(ctx->writeout_branch));
struct midgard_instruction ins =
v_branch(false, false);
- ins.writeout = true;
+ bool depth_only = (rt == MIDGARD_ZS_RT);
+
+ ins.writeout = depth_only ? 0 : PAN_WRITEOUT_C;
/* Add dependencies */
ins.src[0] = src;
ins.src_types[0] = nir_type_uint32;
- ins.constants.u32[0] = rt == MIDGARD_ZS_RT ?
- 0xFF : (rt - MIDGARD_COLOR_RT0) * 0x100;
+ ins.constants.u32[0] = depth_only ? 0xFF : (rt - MIDGARD_COLOR_RT0) * 0x100;
+ for (int i = 0; i < 4; ++i)
+ ins.swizzle[0][i] = i;
+
+ if (~src_z) {
+ emit_explicit_constant(ctx, src_z, src_z);
+ ins.src[2] = src_z;
+ ins.src_types[2] = nir_type_uint32;
+ ins.writeout |= PAN_WRITEOUT_Z;
+ }
+ if (~src_s) {
+ emit_explicit_constant(ctx, src_s, src_s);
+ ins.src[3] = src_s;
+ ins.src_types[3] = nir_type_uint32;
+ ins.writeout |= PAN_WRITEOUT_S;
+ }
/* Emit the branch */
br = emit_mir_instruction(ctx, ins);
emit_attr_read(ctx, reg, vertex_builtin_arg(instr->intrinsic), 1, nir_type_int);
}
+static void
+emit_msaa_builtin(compiler_context *ctx, nir_intrinsic_instr *instr)
+{
+ unsigned reg = nir_dest_index(&instr->dest);
+
+ midgard_instruction ld = m_ld_color_buffer_32u(reg, 0);
+ ld.op = midgard_op_ld_color_buffer_32u_old;
+ ld.load_store.address = 97;
+ ld.load_store.arg_2 = 0x1E;
+
+ for (int i = 0; i < 4; ++i)
+ ld.swizzle[0][i] = COMPONENT_X;
+
+ emit_mir_instruction(ctx, ld);
+}
+
static void
emit_control_barrier(compiler_context *ctx)
{
midgard_instruction ins = {
.type = TAG_TEXTURE_4,
+ .dest = ~0,
.src = { ~0, ~0, ~0, ~0 },
- .texture = {
- .op = TEXTURE_OP_BARRIER,
-
- /* TODO: optimize */
- .barrier_buffer = 1,
- .barrier_shared = 1
- }
+ .op = TEXTURE_OP_BARRIER,
};
emit_mir_instruction(ctx, ins);
}
-static const nir_variable *
-search_var(struct exec_list *vars, unsigned driver_loc)
-{
- nir_foreach_variable(var, vars) {
- if (var->data.driver_location == driver_loc)
- return var;
- }
-
- return NULL;
-}
-
static unsigned
mir_get_branch_cond(nir_src *src, bool *invert)
{
return nir_src_index(NULL, &alu.src);
}
+static uint8_t
+output_load_rt_addr(compiler_context *ctx, nir_intrinsic_instr *instr)
+{
+ if (ctx->is_blend)
+ return ctx->blend_rt;
+
+ const nir_variable *var;
+ var = search_var(ctx->nir, nir_var_shader_out, nir_intrinsic_base(instr));
+ assert(var);
+
+ unsigned loc = var->data.location;
+
+ if (loc == FRAG_RESULT_COLOR)
+ loc = FRAG_RESULT_DATA0;
+
+ if (loc >= FRAG_RESULT_DATA0)
+ return loc - FRAG_RESULT_DATA0;
+
+ if (loc == FRAG_RESULT_DEPTH)
+ return 0x1F;
+ if (loc == FRAG_RESULT_STENCIL)
+ return 0x1E;
+
+ unreachable("Invalid RT to load from");
+}
+
static void
emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr)
{
} else if (is_global || is_shared) {
emit_global(ctx, &instr->instr, true, reg, src_offset, is_shared);
} else if (ctx->stage == MESA_SHADER_FRAGMENT && !ctx->is_blend) {
- emit_varying_read(ctx, reg, offset, nr_comp, component, indirect_offset, t, is_flat);
+ emit_varying_read(ctx, reg, offset, nr_comp, component, indirect_offset, t | nir_dest_bit_size(instr->dest), is_flat);
} else if (ctx->is_blend) {
- /* For blend shaders, load the input color, which is
- * preloaded to r0 */
+ /* ctx->blend_input will be precoloured to r0/r2, where
+ * the input is preloaded */
- midgard_instruction move = v_mov(SSA_FIXED_REGISTER(0), reg);
- emit_mir_instruction(ctx, move);
- schedule_barrier(ctx);
+ unsigned *input = offset ? &ctx->blend_src1 : &ctx->blend_input;
+
+ if (*input == ~0)
+ *input = reg;
+ else
+ emit_mir_instruction(ctx, v_mov(*input, reg));
} else if (ctx->stage == MESA_SHADER_VERTEX) {
emit_attr_read(ctx, reg, offset, nr_comp, t);
} else {
/* Reads 128-bit value raw off the tilebuffer during blending, tasty */
- case nir_intrinsic_load_raw_output_pan:
- case nir_intrinsic_load_output_u8_as_fp16_pan:
+ case nir_intrinsic_load_raw_output_pan: {
reg = nir_dest_index(&instr->dest);
- assert(ctx->is_blend);
/* T720 and below use different blend opcodes with slightly
* different semantics than T760 and up */
midgard_instruction ld = m_ld_color_buffer_32u(reg, 0);
- bool old_blend = ctx->quirks & MIDGARD_OLD_BLEND;
- if (instr->intrinsic == nir_intrinsic_load_output_u8_as_fp16_pan) {
- ld.load_store.op = old_blend ?
- midgard_op_ld_color_buffer_u8_as_fp16_old :
- midgard_op_ld_color_buffer_u8_as_fp16;
+ ld.load_store.arg_2 = output_load_rt_addr(ctx, instr);
- if (old_blend) {
- ld.load_store.address = 1;
- ld.load_store.arg_2 = 0x1E;
- }
+ if (nir_src_is_const(instr->src[0])) {
+ ld.load_store.arg_1 = nir_src_as_uint(instr->src[0]);
+ } else {
+ ld.load_store.varying_parameters = 2;
+ ld.src[1] = nir_src_index(ctx, &instr->src[0]);
+ ld.src_types[1] = nir_type_int32;
+ }
- for (unsigned c = 2; c < 16; ++c)
- ld.swizzle[0][c] = 0;
+ if (ctx->quirks & MIDGARD_OLD_BLEND) {
+ ld.op = midgard_op_ld_color_buffer_32u_old;
+ ld.load_store.address = 16;
+ ld.load_store.arg_2 = 0x1E;
}
emit_mir_instruction(ctx, ld);
break;
+ }
+
+ case nir_intrinsic_load_output: {
+ reg = nir_dest_index(&instr->dest);
+
+ unsigned bits = nir_dest_bit_size(instr->dest);
+
+ midgard_instruction ld;
+ if (bits == 16)
+ ld = m_ld_color_buffer_as_fp16(reg, 0);
+ else
+ ld = m_ld_color_buffer_as_fp32(reg, 0);
+
+ ld.load_store.arg_2 = output_load_rt_addr(ctx, instr);
+
+ for (unsigned c = 4; c < 16; ++c)
+ ld.swizzle[0][c] = 0;
+
+ if (ctx->quirks & MIDGARD_OLD_BLEND) {
+ if (bits == 16)
+ ld.op = midgard_op_ld_color_buffer_as_fp16_old;
+ else
+ ld.op = midgard_op_ld_color_buffer_as_fp32_old;
+ ld.load_store.address = 1;
+ ld.load_store.arg_2 = 0x1E;
+ }
+
+ emit_mir_instruction(ctx, ld);
+ break;
+ }
case nir_intrinsic_load_blend_const_color_rgba: {
assert(ctx->is_blend);
break;
}
- case nir_intrinsic_store_zs_output_pan: {
- assert(ctx->stage == MESA_SHADER_FRAGMENT);
- emit_fragment_store(ctx, nir_src_index(ctx, &instr->src[0]),
- MIDGARD_ZS_RT);
-
- midgard_instruction *br = ctx->writeout_branch[MIDGARD_ZS_RT];
-
- if (!nir_intrinsic_component(instr))
- br->writeout_depth = true;
- if (nir_intrinsic_component(instr) ||
- instr->num_components)
- br->writeout_stencil = true;
- assert(br->writeout_depth | br->writeout_stencil);
- break;
- }
-
case nir_intrinsic_store_output:
+ case nir_intrinsic_store_combined_output_pan:
assert(nir_src_is_const(instr->src[1]) && "no indirect outputs");
offset = nir_intrinsic_base(instr) + nir_src_as_uint(instr->src[1]);
reg = nir_src_index(ctx, &instr->src[0]);
if (ctx->stage == MESA_SHADER_FRAGMENT) {
- const nir_variable *var;
- enum midgard_rt_id rt;
+ bool combined = instr->intrinsic ==
+ nir_intrinsic_store_combined_output_pan;
- var = search_var(&ctx->nir->outputs,
+ const nir_variable *var;
+ var = search_var(ctx->nir, nir_var_shader_out,
nir_intrinsic_base(instr));
assert(var);
+
+ /* Dual-source blend writeout is done by leaving the
+ * value in r2 for the blend shader to use. */
+ if (var->data.index) {
+ if (instr->src[0].is_ssa) {
+ emit_explicit_constant(ctx, reg, reg);
+
+ unsigned out = make_compiler_temp(ctx);
+
+ midgard_instruction ins = v_mov(reg, out);
+ emit_mir_instruction(ctx, ins);
+
+ ctx->blend_src1 = out;
+ } else {
+ ctx->blend_src1 = reg;
+ }
+
+ break;
+ }
+
+ enum midgard_rt_id rt;
if (var->data.location == FRAG_RESULT_COLOR)
rt = MIDGARD_COLOR_RT0;
else if (var->data.location >= FRAG_RESULT_DATA0)
rt = MIDGARD_COLOR_RT0 + var->data.location -
FRAG_RESULT_DATA0;
+ else if (combined)
+ rt = MIDGARD_ZS_RT;
else
- assert(0);
+ unreachable("bad rt");
+
+ unsigned reg_z = ~0, reg_s = ~0;
+ if (combined) {
+ unsigned writeout = nir_intrinsic_component(instr);
+ if (writeout & PAN_WRITEOUT_Z)
+ reg_z = nir_src_index(ctx, &instr->src[2]);
+ if (writeout & PAN_WRITEOUT_S)
+ reg_s = nir_src_index(ctx, &instr->src[3]);
+ }
- emit_fragment_store(ctx, reg, rt);
+ emit_fragment_store(ctx, reg, reg_z, reg_s, rt);
} else if (ctx->stage == MESA_SHADER_VERTEX) {
+ assert(instr->intrinsic == nir_intrinsic_store_output);
+
/* We should have been vectorized, though we don't
* currently check that st_vary is emitted only once
* per slot (this is relevant, since there's not a mask
switch (nir_alu_type_get_base_type(nir_intrinsic_type(instr))) {
case nir_type_uint:
case nir_type_bool:
- st.load_store.op = midgard_op_st_vary_32u;
+ st.op = midgard_op_st_vary_32u;
break;
case nir_type_int:
- st.load_store.op = midgard_op_st_vary_32i;
+ st.op = midgard_op_st_vary_32i;
break;
case nir_type_float:
- st.load_store.op = midgard_op_st_vary_32;
+ st.op = midgard_op_st_vary_32;
break;
default:
unreachable("Attempted to store unknown type");
case nir_intrinsic_store_raw_output_pan:
assert (ctx->stage == MESA_SHADER_FRAGMENT);
reg = nir_src_index(ctx, &instr->src[0]);
-
- if (ctx->quirks & MIDGARD_OLD_BLEND) {
- /* Suppose reg = qr0.xyzw. That means 4 8-bit ---> 1 32-bit. So
- * reg = r0.x. We want to splatter. So we can do a 32-bit move
- * of:
- *
- * imov r0.xyzw, r0.xxxx
- */
-
- unsigned expanded = make_compiler_temp(ctx);
-
- midgard_instruction splatter = v_mov(reg, expanded);
-
- for (unsigned c = 0; c < 16; ++c)
- splatter.swizzle[1][c] = 0;
-
- emit_mir_instruction(ctx, splatter);
- emit_fragment_store(ctx, expanded, ctx->blend_rt);
- } else
- emit_fragment_store(ctx, reg, ctx->blend_rt);
-
+ emit_fragment_store(ctx, reg, ~0, ~0, ctx->blend_rt);
break;
case nir_intrinsic_store_global:
emit_vertex_builtin(ctx, instr);
break;
+ case nir_intrinsic_load_sample_id:
+ emit_msaa_builtin(ctx, instr);
+ break;
+
case nir_intrinsic_memory_barrier_buffer:
case nir_intrinsic_memory_barrier_shared:
break;
}
}
+/* Returns dimension with 0 special casing cubemaps */
static unsigned
midgard_tex_format(enum glsl_sampler_dim dim)
{
switch (dim) {
case GLSL_SAMPLER_DIM_1D:
case GLSL_SAMPLER_DIM_BUF:
- return MALI_TEX_1D;
+ return 1;
case GLSL_SAMPLER_DIM_2D:
+ case GLSL_SAMPLER_DIM_MS:
case GLSL_SAMPLER_DIM_EXTERNAL:
case GLSL_SAMPLER_DIM_RECT:
- return MALI_TEX_2D;
+ return 2;
case GLSL_SAMPLER_DIM_3D:
- return MALI_TEX_3D;
+ return 3;
case GLSL_SAMPLER_DIM_CUBE:
- return MALI_TEX_CUBE;
+ return 0;
default:
DBG("Unknown sampler dim type\n");
}
}
-/* 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
/* TODO */
//assert (!instr->sampler);
+ nir_dest *dest = &instr->dest;
+
int texture_index = instr->texture_index;
int sampler_index = texture_index;
nir_alu_type dest_base = nir_alu_type_get_base_type(instr->dest_type);
- nir_alu_type dest_type = dest_base | nir_dest_bit_size(instr->dest);
+ nir_alu_type dest_type = dest_base | nir_dest_bit_size(*dest);
+
+ /* texture instructions support float outmods */
+ unsigned outmod = midgard_outmod_none;
+ if (dest_base == nir_type_float) {
+ outmod = mir_determine_float_outmod(ctx, &dest, 0);
+ }
midgard_instruction ins = {
.type = TAG_TEXTURE_4,
.mask = 0xF,
- .dest = nir_dest_index(&instr->dest),
+ .dest = nir_dest_index(dest),
.src = { ~0, ~0, ~0, ~0 },
.dest_type = dest_type,
.swizzle = SWIZZLE_IDENTITY_4,
+ .outmod = outmod,
+ .op = midgard_texop,
.texture = {
- .op = midgard_texop,
.format = midgard_tex_format(instr->sampler_dim),
.texture_handle = texture_index,
.sampler_handle = sampler_index,
break;
};
- case nir_tex_src_comparator: {
+ case nir_tex_src_comparator:
+ case nir_tex_src_ms_index: {
unsigned comp = COMPONENT_Z;
/* mov coord_temp.foo, coords */
emit_texop_native(ctx, instr, TEXTURE_OP_LOD);
break;
case nir_texop_txf:
+ case nir_texop_txf_ms:
emit_texop_native(ctx, instr, TEXTURE_OP_TEXEL_FETCH);
break;
case nir_texop_txs:
}
}
+unsigned
+max_bitsize_for_alu(midgard_instruction *ins)
+{
+ unsigned max_bitsize = 0;
+ for (int i = 0; i < MIR_SRC_COUNT; i++) {
+ if (ins->src[i] == ~0) continue;
+ unsigned src_bitsize = nir_alu_type_get_type_size(ins->src_types[i]);
+ max_bitsize = MAX2(src_bitsize, max_bitsize);
+ }
+ unsigned dst_bitsize = nir_alu_type_get_type_size(ins->dest_type);
+ max_bitsize = MAX2(dst_bitsize, max_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 (ins->op) {
+ case midgard_alu_op_fsqrt:
+ case midgard_alu_op_frcp:
+ case midgard_alu_op_frsqrt:
+ case midgard_alu_op_fsin:
+ case midgard_alu_op_fcos:
+ case midgard_alu_op_fexp2:
+ case midgard_alu_op_flog2:
+ max_bitsize = MAX2(max_bitsize, 32);
+ break;
+
+ default:
+ break;
+ }
+
+ /* High implies computing at a higher bitsize, e.g umul_high of 32-bit
+ * requires computing at 64-bit */
+ if (midgard_is_integer_out_op(ins->op) && ins->outmod == midgard_outmod_int_high) {
+ max_bitsize *= 2;
+ assert(max_bitsize <= 64);
+ }
+
+ return max_bitsize;
+}
+
+midgard_reg_mode
+reg_mode_for_bitsize(unsigned bitsize)
+{
+ switch (bitsize) {
+ /* use 16 pipe for 8 since we don't support vec16 yet */
+ case 8:
+ case 16:
+ return midgard_reg_mode_16;
+ case 32:
+ return midgard_reg_mode_32;
+ case 64:
+ return midgard_reg_mode_64;
+ default:
+ unreachable("invalid bit size");
+ }
+}
+
/* Midgard supports two types of constants, embedded constants (128-bit) and
* inline constants (16-bit). Sometimes, especially with scalar ops, embedded
* constants can be demoted to inline constants, for space savings and
/* Blend constants must not be inlined by definition */
if (ins->has_blend_constant) continue;
+ unsigned max_bitsize = max_bitsize_for_alu(ins);
+
/* We can inline 32-bit (sometimes) or 16-bit (usually) */
- bool is_16 = ins->alu.reg_mode == midgard_reg_mode_16;
- bool is_32 = ins->alu.reg_mode == midgard_reg_mode_32;
+ bool is_16 = max_bitsize == 16;
+ bool is_32 = max_bitsize == 32;
if (!(is_16 || is_32))
continue;
* restrictions. So, if possible we try to flip the arguments
* in that case */
- int op = ins->alu.op;
+ int op = ins->op;
if (ins->src[0] == SSA_FIXED_REGISTER(REGISTER_CONSTANT) &&
alu_opcode_props[op].props & OP_COMMUTES) {
}
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;
unsigned component = ins->swizzle[1][first_comp];
/* Scale constant appropriately, if we can legally */
- uint16_t scaled_constant = 0;
+ int16_t scaled_constant = 0;
if (is_16) {
scaled_constant = ins->constants.u16[component];
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. */
uint32_t value = is_16 ? cons->u16[component] : cons->u32[component];
bool is_vector = false;
- unsigned mask = effective_writemask(&ins->alu, ins->mask);
+ unsigned mask = effective_writemask(ins->op, ins->mask);
for (unsigned c = 0; c < MIR_VEC_COMPONENTS; ++c) {
/* We only care if this component is actually used */
}
}
+/* We want to force the invert on AND/OR to the second slot to legalize into
+ * iandnot/iornot. The relevant patterns are for AND (and OR respectively)
+ *
+ * ~a & #b = ~a & ~(#~b)
+ * ~a & b = b & ~a
+ */
+
+static void
+midgard_legalize_invert(compiler_context *ctx, midgard_block *block)
+{
+ mir_foreach_instr_in_block(block, ins) {
+ if (ins->type != TAG_ALU_4) continue;
+
+ if (ins->op != midgard_alu_op_iand &&
+ ins->op != midgard_alu_op_ior) continue;
+
+ if (ins->src_invert[1] || !ins->src_invert[0]) continue;
+
+ if (ins->has_inline_constant) {
+ /* ~(#~a) = ~(~#a) = a, so valid, and forces both
+ * inverts on */
+ ins->inline_constant = ~ins->inline_constant;
+ ins->src_invert[1] = true;
+ } else {
+ /* Flip to the right invert order. Note
+ * has_inline_constant false by assumption on the
+ * branch, so flipping makes sense. */
+ mir_flip(ins);
+ }
+ }
+}
+
static unsigned
emit_fragment_epilogue(compiler_context *ctx, unsigned rt)
{
/* Loop to ourselves */
midgard_instruction *br = ctx->writeout_branch[rt];
struct midgard_instruction ins = v_branch(false, false);
- ins.writeout = true;
- ins.writeout_depth = br->writeout_depth;
- ins.writeout_stencil = br->writeout_stencil;
+ ins.writeout = br->writeout;
ins.branch.target_block = ctx->block_count - 1;
ins.constants.u32[0] = br->constants.u32[0];
+ memcpy(&ins.src_types, &br->src_types, sizeof(ins.src_types));
emit_mir_instruction(ctx, ins);
ctx->current_block->epilogue = true;
}
static midgard_block *
-emit_block(compiler_context *ctx, nir_block *block)
+emit_block_init(compiler_context *ctx)
{
midgard_block *this_block = ctx->after_block;
ctx->after_block = NULL;
list_inithead(&this_block->base.instructions);
ctx->current_block = this_block;
+ return this_block;
+}
+
+static midgard_block *
+emit_block(compiler_context *ctx, nir_block *block)
+{
+ midgard_block *this_block = emit_block_init(ctx);
+
nir_foreach_instr(instr, block) {
emit_instr(ctx, instr);
++ctx->instruction_count;
* stream and in branch targets. An initial block might be empty, so iterate
* until we find one that 'works' */
-static unsigned
+unsigned
midgard_get_first_tag_from_block(compiler_context *ctx, unsigned block_idx)
{
midgard_block *initial_block = mir_get_block(ctx, block_idx);
}
int
-midgard_compile_shader_nir(nir_shader *nir, panfrost_program *program, bool is_blend, unsigned blend_rt, unsigned gpu_id, bool shaderdb)
+midgard_compile_shader_nir(nir_shader *nir, panfrost_program *program, bool is_blend, unsigned blend_rt, unsigned gpu_id, bool shaderdb, bool silent)
{
struct util_dynarray *compiled = &program->compiled;
ctx->nir = nir;
ctx->stage = nir->info.stage;
ctx->is_blend = is_blend;
- ctx->alpha_ref = program->alpha_ref;
ctx->blend_rt = MIDGARD_COLOR_RT0 + blend_rt;
+ ctx->blend_input = ~0;
+ ctx->blend_src1 = ~0;
ctx->quirks = midgard_get_quirks(gpu_id);
/* Start off with a safe cutoff, allowing usage of all 16 work
/* Initialize at a global (not block) level hash tables */
ctx->ssa_constants = _mesa_hash_table_u64_create(NULL);
- ctx->hash_to_temp = _mesa_hash_table_u64_create(NULL);
/* Lower gl_Position pre-optimisation, but after lowering vars to ssa
* (so we don't accidentally duplicate the epilogue since mesa/st has
NIR_PASS_V(nir, nir_lower_var_copies);
NIR_PASS_V(nir, nir_lower_vars_to_ssa);
- NIR_PASS_V(nir, nir_lower_io, nir_var_all, glsl_type_size, 0);
+ unsigned pan_quirks = panfrost_get_quirks(gpu_id);
+ NIR_PASS_V(nir, pan_lower_framebuffer,
+ program->rt_formats, is_blend, pan_quirks);
+
+ NIR_PASS_V(nir, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
+ glsl_type_size, 0);
NIR_PASS_V(nir, nir_lower_ssbo);
NIR_PASS_V(nir, midgard_nir_lower_zs_store);
/* Optimisation passes */
- optimise_nir(nir, ctx->quirks);
+ optimise_nir(nir, ctx->quirks, is_blend);
+
+ NIR_PASS_V(nir, midgard_nir_reorder_writeout);
- if (midgard_debug & MIDGARD_DBG_SHADERS) {
+ if ((midgard_debug & MIDGARD_DBG_SHADERS) && !silent) {
nir_print_shader(nir, stdout);
}
/* Assign sysvals and counts, now that we're sure
* (post-optimisation) */
- panfrost_nir_assign_sysvals(&ctx->sysvals, nir);
+ panfrost_nir_assign_sysvals(&ctx->sysvals, ctx, nir);
program->sysval_count = ctx->sysvals.sysval_count;
memcpy(program->sysvals, ctx->sysvals.sysvals, sizeof(ctx->sysvals.sysvals[0]) * ctx->sysvals.sysval_count);
ctx->func = func;
ctx->already_emitted = calloc(BITSET_WORDS(func->impl->ssa_alloc), sizeof(BITSET_WORD));
+ if (nir->info.outputs_read && !is_blend) {
+ emit_block_init(ctx);
+
+ struct midgard_instruction wait = v_branch(false, false);
+ wait.branch.target_type = TARGET_TILEBUF_WAIT;
+
+ emit_mir_instruction(ctx, wait);
+
+ ++ctx->instruction_count;
+ }
+
emit_cf_list(ctx, &func->impl->body);
free(ctx->already_emitted);
break; /* TODO: Multi-function shaders */
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 */
do {
progress = false;
+ progress |= midgard_opt_dead_code_eliminate(ctx);
mir_foreach_block(ctx, _block) {
midgard_block *block = (midgard_block *) _block;
progress |= midgard_opt_copy_prop(ctx, block);
- progress |= midgard_opt_dead_code_eliminate(ctx, block);
progress |= midgard_opt_combine_projection(ctx, block);
progress |= midgard_opt_varying_projection(ctx, block);
}
mir_foreach_block(ctx, _block) {
midgard_block *block = (midgard_block *) _block;
midgard_lower_derivatives(ctx, block);
+ midgard_legalize_invert(ctx, block);
midgard_cull_dead_branch(ctx, block);
}
midgard_schedule_program(ctx);
mir_ra(ctx);
- /* Now that all the bundles are scheduled and we can calculate block
- * sizes, emit actual branch instructions rather than placeholders */
-
- int br_block_idx = 0;
-
- mir_foreach_block(ctx, _block) {
- midgard_block *block = (midgard_block *) _block;
- util_dynarray_foreach(&block->bundles, midgard_bundle, bundle) {
- for (int c = 0; c < bundle->instruction_count; ++c) {
- midgard_instruction *ins = bundle->instructions[c];
-
- if (!midgard_is_branch_unit(ins->unit)) continue;
-
- /* Parse some basic branch info */
- bool is_compact = ins->unit == ALU_ENAB_BR_COMPACT;
- bool is_conditional = ins->branch.conditional;
- bool is_inverted = ins->branch.invert_conditional;
- bool is_discard = ins->branch.target_type == TARGET_DISCARD;
- bool is_writeout = ins->writeout;
-
- /* Determine the block we're jumping to */
- int target_number = ins->branch.target_block;
-
- /* Report the destination tag */
- int dest_tag = is_discard ? 0 : midgard_get_first_tag_from_block(ctx, target_number);
-
- /* Count up the number of quadwords we're
- * jumping over = number of quadwords until
- * (br_block_idx, target_number) */
-
- int quadword_offset = 0;
-
- if (is_discard) {
- /* Ignored */
- } else if (target_number > br_block_idx) {
- /* Jump forward */
-
- for (int idx = br_block_idx + 1; idx < target_number; ++idx) {
- midgard_block *blk = mir_get_block(ctx, idx);
- assert(blk);
-
- quadword_offset += blk->quadword_count;
- }
- } else {
- /* Jump backwards */
-
- for (int idx = br_block_idx; idx >= target_number; --idx) {
- midgard_block *blk = mir_get_block(ctx, idx);
- assert(blk);
-
- quadword_offset -= blk->quadword_count;
- }
- }
-
- /* Unconditional extended branches (far jumps)
- * have issues, so we always use a conditional
- * branch, setting the condition to always for
- * unconditional. For compact unconditional
- * branches, cond isn't used so it doesn't
- * matter what we pick. */
-
- midgard_condition cond =
- !is_conditional ? midgard_condition_always :
- is_inverted ? midgard_condition_false :
- midgard_condition_true;
-
- midgard_jmp_writeout_op op =
- is_discard ? midgard_jmp_writeout_op_discard :
- is_writeout ? midgard_jmp_writeout_op_writeout :
- (is_compact && !is_conditional) ? midgard_jmp_writeout_op_branch_uncond :
- midgard_jmp_writeout_op_branch_cond;
-
- if (!is_compact) {
- midgard_branch_extended branch =
- midgard_create_branch_extended(
- cond, op,
- dest_tag,
- quadword_offset);
-
- memcpy(&ins->branch_extended, &branch, sizeof(branch));
- } else if (is_conditional || is_discard) {
- midgard_branch_cond branch = {
- .op = op,
- .dest_tag = dest_tag,
- .offset = quadword_offset,
- .cond = cond
- };
-
- assert(branch.offset == quadword_offset);
-
- memcpy(&ins->br_compact, &branch, sizeof(branch));
- } else {
- assert(op == midgard_jmp_writeout_op_branch_uncond);
-
- midgard_branch_uncond branch = {
- .op = op,
- .dest_tag = dest_tag,
- .offset = quadword_offset,
- .unknown = 1
- };
-
- assert(branch.offset == quadword_offset);
-
- memcpy(&ins->br_compact, &branch, sizeof(branch));
- }
- }
- }
-
- ++br_block_idx;
- }
-
/* Emit flat binary from the instruction arrays. Iterate each block in
* sequence. Save instruction boundaries such that lookahead tags can
* be assigned easily */
if (!bundle->last_writeout && (current_bundle + 1 < bundle_count))
lookahead = source_order_bundles[current_bundle + 1]->tag;
- emit_binary_bundle(ctx, bundle, compiled, lookahead);
+ emit_binary_bundle(ctx, block, bundle, compiled, lookahead);
++current_bundle;
}
program->blend_patch_offset = ctx->blend_constant_offset;
program->tls_size = ctx->tls_size;
- if (midgard_debug & MIDGARD_DBG_SHADERS)
+ if ((midgard_debug & MIDGARD_DBG_SHADERS) && !silent)
disassemble_midgard(stdout, program->compiled.data, program->compiled.size, gpu_id, ctx->stage);
- if (midgard_debug & MIDGARD_DBG_SHADERDB || shaderdb) {
+ if ((midgard_debug & MIDGARD_DBG_SHADERDB || shaderdb) && !silent) {
unsigned nr_bundles = 0, nr_ins = 0;
/* Count instructions and bundles */
"%u registers, %u threads, %u loops, "
"%u:%u spills:fills\n",
SHADER_DB_COUNT++,
+ ctx->is_blend ? "PAN_SHADER_BLEND" :
gl_shader_stage_name(ctx->stage),
nr_ins, nr_bundles, ctx->quadword_count,
nr_registers, nr_threads,