From 9915bb2c40b0cad628536d179eac47ccf3325860 Mon Sep 17 00:00:00 2001 From: Alyssa Rosenzweig Date: Thu, 7 May 2020 10:12:38 -0400 Subject: [PATCH] pan/mdg: Remove mir_*size routines We'd rather use the actual type information than inferring modes all over the place. Signed-off-by: Alyssa Rosenzweig Part-of: --- src/panfrost/midgard/compiler.h | 8 +- src/panfrost/midgard/midgard_compile.c | 1 + src/panfrost/midgard/midgard_emit.c | 18 +-- src/panfrost/midgard/midgard_opt_dce.c | 6 +- src/panfrost/midgard/midgard_print.c | 4 +- src/panfrost/midgard/midgard_ra.c | 64 +++++---- src/panfrost/midgard/midgard_schedule.c | 9 +- src/panfrost/midgard/mir.c | 139 +++----------------- src/panfrost/midgard/mir_promote_uniforms.c | 3 +- src/panfrost/util/pan_ir.c | 4 + 10 files changed, 78 insertions(+), 178 deletions(-) diff --git a/src/panfrost/midgard/compiler.h b/src/panfrost/midgard/compiler.h index 8132a9e91a9..fdcd4c7a8bb 100644 --- a/src/panfrost/midgard/compiler.h +++ b/src/panfrost/midgard/compiler.h @@ -484,13 +484,9 @@ bool mir_single_use(compiler_context *ctx, unsigned value); unsigned mir_use_count(compiler_context *ctx, unsigned value); uint16_t mir_bytemask_of_read_components(midgard_instruction *ins, unsigned node); uint16_t mir_bytemask_of_read_components_index(midgard_instruction *ins, unsigned i); -midgard_reg_mode mir_typesize(midgard_instruction *ins); -midgard_reg_mode mir_srcsize(midgard_instruction *ins, unsigned i); -unsigned mir_bytes_for_mode(midgard_reg_mode mode); -midgard_reg_mode mir_mode_for_destsize(unsigned size); -uint16_t mir_from_bytemask(uint16_t bytemask, midgard_reg_mode mode); +uint16_t mir_from_bytemask(uint16_t bytemask, unsigned bits); uint16_t mir_bytemask(midgard_instruction *ins); -uint16_t mir_round_bytemask_up(uint16_t mask, midgard_reg_mode mode); +uint16_t mir_round_bytemask_up(uint16_t mask, unsigned bits); void mir_set_bytemask(midgard_instruction *ins, uint16_t bytemask); unsigned mir_upper_override(midgard_instruction *ins); diff --git a/src/panfrost/midgard/midgard_compile.c b/src/panfrost/midgard/midgard_compile.c index 9fbcd491883..4a387010981 100644 --- a/src/panfrost/midgard/midgard_compile.c +++ b/src/panfrost/midgard/midgard_compile.c @@ -117,6 +117,7 @@ schedule_barrier(compiler_context *ctx) if (store) { \ i.src[0] = ssa; \ i.src_types[0] = T; \ + i.dest_type = T; \ } else { \ i.dest = ssa; \ i.dest_type = T; \ diff --git a/src/panfrost/midgard/midgard_emit.c b/src/panfrost/midgard/midgard_emit.c index 8bfb76e2a4c..99134ddcfb9 100644 --- a/src/panfrost/midgard/midgard_emit.c +++ b/src/panfrost/midgard/midgard_emit.c @@ -188,12 +188,12 @@ mir_pack_swizzle_alu(midgard_instruction *ins) unsigned packed = 0; if (ins->alu.reg_mode == midgard_reg_mode_64) { - midgard_reg_mode mode = mir_srcsize(ins, i); - unsigned components = 16 / mir_bytes_for_mode(mode); + unsigned sz = nir_alu_type_get_type_size(ins->src_types[i]); + unsigned components = 64 / sz; packed = mir_pack_swizzle_64(ins->swizzle[i], components); - if (mode == midgard_reg_mode_32) { + if (sz == 32) { bool lo = ins->swizzle[i][0] >= COMPONENT_Z; bool hi = ins->swizzle[i][1] >= COMPONENT_Z; unsigned mask = mir_bytemask(ins); @@ -207,7 +207,7 @@ mir_pack_swizzle_alu(midgard_instruction *ins) } else { src[i].rep_low |= hi; } - } else if (mode < midgard_reg_mode_32) { + } else if (sz < 32) { unreachable("Cannot encode 8/16 swizzle in 64-bit"); } } else { @@ -221,7 +221,7 @@ mir_pack_swizzle_alu(midgard_instruction *ins) bool upper = ins->swizzle[i][first] > 3; if (upper && ins->mask) - assert(mir_srcsize(ins, i) <= midgard_reg_mode_16); + assert(nir_alu_type_get_type_size(ins->src_types[i]) <= 16); for (unsigned c = 0; c < 4; ++c) { unsigned v = ins->swizzle[i][c]; @@ -306,13 +306,13 @@ mir_pack_swizzle_tex(midgard_instruction *ins) static void mir_pack_ldst_mask(midgard_instruction *ins) { - midgard_reg_mode mode = mir_typesize(ins); + unsigned sz = nir_alu_type_get_type_size(ins->dest_type); unsigned packed = ins->mask; - if (mode == midgard_reg_mode_64) { + if (sz == 64) { packed = ((ins->mask & 0x2) ? (0x8 | 0x4) : 0) | ((ins->mask & 0x1) ? (0x2 | 0x1) : 0); - } else if (mode == midgard_reg_mode_16) { + } else if (sz == 16) { packed = 0; for (unsigned i = 0; i < 4; ++i) { @@ -323,6 +323,8 @@ mir_pack_ldst_mask(midgard_instruction *ins) packed |= (u << i); } + } else { + assert(sz == 32); } ins->load_store.mask = packed; diff --git a/src/panfrost/midgard/midgard_opt_dce.c b/src/panfrost/midgard/midgard_opt_dce.c index e9e51a6451d..be9307e712a 100644 --- a/src/panfrost/midgard/midgard_opt_dce.c +++ b/src/panfrost/midgard/midgard_opt_dce.c @@ -72,11 +72,11 @@ midgard_opt_dead_code_eliminate_block(compiler_context *ctx, midgard_block *bloc mir_foreach_instr_in_block_rev(block, ins) { if (can_cull_mask(ctx, ins)) { - midgard_reg_mode mode = mir_typesize(ins); + unsigned type_size = nir_alu_type_get_type_size(ins->dest_type); unsigned oldmask = ins->mask; - unsigned rounded = mir_round_bytemask_up(live[ins->dest], mode); - unsigned cmask = mir_from_bytemask(rounded, mode); + unsigned rounded = mir_round_bytemask_up(live[ins->dest], type_size); + unsigned cmask = mir_from_bytemask(rounded, type_size); ins->mask &= cmask; progress |= (ins->mask != oldmask); diff --git a/src/panfrost/midgard/midgard_print.c b/src/panfrost/midgard/midgard_print.c index fa7792e33e1..c74b75bcd2f 100644 --- a/src/panfrost/midgard/midgard_print.c +++ b/src/panfrost/midgard/midgard_print.c @@ -230,7 +230,7 @@ mir_print_constant_component(FILE *fp, const midgard_constants *consts, unsigned static void mir_print_embedded_constant(midgard_instruction *ins, unsigned src_idx) { - unsigned type_size = mir_bytes_for_mode(ins->alu.reg_mode); + unsigned type_size = nir_alu_type_get_type_size(ins->dest_type); midgard_vector_alu_src src; assert(src_idx <= 1); @@ -242,7 +242,7 @@ mir_print_embedded_constant(midgard_instruction *ins, unsigned src_idx) unsigned *swizzle = ins->swizzle[src_idx]; unsigned comp_mask = effective_writemask(&ins->alu, ins->mask); unsigned num_comp = util_bitcount(comp_mask); - unsigned max_comp = 16 / type_size; + unsigned max_comp = 64 / type_size; bool first = true; printf("#"); diff --git a/src/panfrost/midgard/midgard_ra.c b/src/panfrost/midgard/midgard_ra.c index 6049efc3f8f..729d540af96 100644 --- a/src/panfrost/midgard/midgard_ra.c +++ b/src/panfrost/midgard/midgard_ra.c @@ -42,17 +42,17 @@ struct phys_reg { /* Shift up by reg_offset and horizontally by dst_offset. */ static void -offset_swizzle(unsigned *swizzle, unsigned reg_offset, unsigned srcsize, unsigned dst_offset) +offset_swizzle(unsigned *swizzle, unsigned reg_offset, unsigned srcsize, unsigned dstsize, unsigned dst_offset) { unsigned out[MIR_VEC_COMPONENTS]; signed reg_comp = reg_offset / srcsize; - signed dst_comp = dst_offset / srcsize; + signed dst_comp = dst_offset / dstsize; unsigned max_component = (16 / srcsize) - 1; assert(reg_comp * srcsize == reg_offset); - assert(dst_comp * srcsize == dst_offset); + assert(dst_comp * dstsize == dst_offset); for (signed c = 0; c < MIR_VEC_COMPONENTS; ++c) { signed comp = MAX2(c - dst_comp, 0); @@ -65,12 +65,12 @@ offset_swizzle(unsigned *swizzle, unsigned reg_offset, unsigned srcsize, unsigne /* Helper to return the default phys_reg for a given register */ static struct phys_reg -default_phys_reg(int reg, midgard_reg_mode size) +default_phys_reg(int reg, unsigned size) { struct phys_reg r = { .reg = reg, .offset = 0, - .size = mir_bytes_for_mode(size) + .size = size }; return r; @@ -80,7 +80,7 @@ default_phys_reg(int reg, midgard_reg_mode size) * register corresponds to */ static struct phys_reg -index_to_reg(compiler_context *ctx, struct lcra_state *l, unsigned reg, midgard_reg_mode size) +index_to_reg(compiler_context *ctx, struct lcra_state *l, unsigned reg, unsigned size) { /* Check for special cases */ if (reg == ~0) @@ -93,7 +93,7 @@ index_to_reg(compiler_context *ctx, struct lcra_state *l, unsigned reg, midgard_ struct phys_reg r = { .reg = l->solutions[reg] / 16, .offset = l->solutions[reg] & 0xF, - .size = mir_bytes_for_mode(size) + .size = size }; /* Report that we actually use this register, and return it */ @@ -292,7 +292,7 @@ mir_lower_special_reads(compiler_context *ctx) } else { idx = spill_idx++; m = v_mov(i, idx); - m.mask = mir_from_bytemask(mir_bytemask_of_read_components(pre_use, i), midgard_reg_mode_32); + m.mask = mir_from_bytemask(mir_bytemask_of_read_components(pre_use, i), 32); mir_insert_instruction_before(ctx, pre_use, m); mir_rewrite_index_src_single(pre_use, i, idx); } @@ -496,10 +496,10 @@ allocate_registers(compiler_context *ctx, bool *spilled) if (ins->type == TAG_LOAD_STORE_4 && OP_HAS_ADDRESS(ins->load_store.op)) { mir_foreach_src(ins, v) { unsigned s = ins->src[v]; - unsigned size = mir_srcsize(ins, v); + unsigned size = nir_alu_type_get_type_size(ins->src_types[v]); if (s < ctx->temp_count) - min_alignment[s] = (size == midgard_reg_mode_64) ? 3 : 2; + min_alignment[s] = (size == 64) ? 3 : 2; } } @@ -613,6 +613,13 @@ install_registers_instr( struct lcra_state *l, midgard_instruction *ins) { + unsigned src_size[MIR_SRC_COUNT]; + + for (unsigned i = 0; i < MIR_SRC_COUNT; ++i) + src_size[i] = MAX2(nir_alu_type_get_type_size(ins->src_types[i]) / 8, 1); + + unsigned dest_size = MAX2(nir_alu_type_get_type_size(ins->dest_type) / 8, 1); + switch (ins->type) { case TAG_ALU_4: case TAG_ALU_8: @@ -621,9 +628,9 @@ install_registers_instr( if (ins->compact_branch) return; - struct phys_reg src1 = index_to_reg(ctx, l, ins->src[0], mir_srcsize(ins, 0)); - struct phys_reg src2 = index_to_reg(ctx, l, ins->src[1], mir_srcsize(ins, 1)); - struct phys_reg dest = index_to_reg(ctx, l, ins->dest, mir_typesize(ins)); + struct phys_reg src1 = index_to_reg(ctx, l, ins->src[0], src_size[0]); + struct phys_reg src2 = index_to_reg(ctx, l, ins->src[1], src_size[1]); + struct phys_reg dest = index_to_reg(ctx, l, ins->dest, dest_size); mir_set_bytemask(ins, mir_bytemask(ins) << dest.offset); @@ -631,7 +638,7 @@ install_registers_instr( GET_CHANNEL_COUNT(alu_opcode_props[ins->alu.op].props) ? 0 : dest.offset; - offset_swizzle(ins->swizzle[0], src1.offset, src1.size, dest_offset); + offset_swizzle(ins->swizzle[0], src1.offset, src1.size, dest.size, dest_offset); ins->registers.src1_reg = src1.reg; @@ -651,7 +658,7 @@ install_registers_instr( } else { midgard_vector_alu_src mod2 = vector_alu_from_unsigned(ins->alu.src2); - offset_swizzle(ins->swizzle[1], src2.offset, src2.size, dest_offset); + offset_swizzle(ins->swizzle[1], src2.offset, src2.size, dest.size, dest_offset); ins->alu.src2 = vector_alu_srco_unsigned(mod2); ins->registers.src2_reg = src2.reg; @@ -669,16 +676,16 @@ install_registers_instr( bool encodes_src = OP_IS_STORE(ins->load_store.op); if (encodes_src) { - struct phys_reg src = index_to_reg(ctx, l, ins->src[0], mir_srcsize(ins, 0)); + struct phys_reg src = index_to_reg(ctx, l, ins->src[0], src_size[0]); assert(src.reg == 26 || src.reg == 27); ins->load_store.reg = src.reg - 26; - offset_swizzle(ins->swizzle[0], src.offset, src.size, 0); + offset_swizzle(ins->swizzle[0], src.offset, src.size, 1, 0); } else { - struct phys_reg dst = index_to_reg(ctx, l, ins->dest, mir_typesize(ins)); + struct phys_reg dst = index_to_reg(ctx, l, ins->dest, dest_size); ins->load_store.reg = dst.reg; - offset_swizzle(ins->swizzle[0], 0, 4, dst.offset); + offset_swizzle(ins->swizzle[0], 0, 4, dst.size, dst.offset); mir_set_bytemask(ins, mir_bytemask(ins) << dst.offset); } @@ -686,17 +693,16 @@ install_registers_instr( unsigned src2 = ins->src[1]; unsigned src3 = ins->src[2]; - midgard_reg_mode m32 = midgard_reg_mode_32; if (src2 != ~0) { - struct phys_reg src = index_to_reg(ctx, l, src2, m32); + struct phys_reg src = index_to_reg(ctx, l, src2, 4); unsigned component = src.offset / src.size; assert(component * src.size == src.offset); ins->load_store.arg_1 |= midgard_ldst_reg(src.reg, component); } if (src3 != ~0) { - struct phys_reg src = index_to_reg(ctx, l, src3, m32); + struct phys_reg src = index_to_reg(ctx, l, src3, 4); unsigned component = src.offset / src.size; assert(component * src.size == src.offset); ins->load_store.arg_2 |= midgard_ldst_reg(src.reg, component); @@ -710,22 +716,22 @@ install_registers_instr( break; /* Grab RA results */ - struct phys_reg dest = index_to_reg(ctx, l, ins->dest, mir_typesize(ins)); - struct phys_reg coord = index_to_reg(ctx, l, ins->src[1], mir_srcsize(ins, 1)); - struct phys_reg lod = index_to_reg(ctx, l, ins->src[2], mir_srcsize(ins, 2)); - struct phys_reg offset = index_to_reg(ctx, l, ins->src[3], mir_srcsize(ins, 2)); + struct phys_reg dest = index_to_reg(ctx, l, ins->dest, dest_size); + struct phys_reg coord = index_to_reg(ctx, l, ins->src[1], src_size[1]); + struct phys_reg lod = index_to_reg(ctx, l, ins->src[2], src_size[2]); + struct phys_reg offset = index_to_reg(ctx, l, ins->src[3], src_size[3]); /* First, install the texture coordinate */ ins->texture.in_reg_full = 1; ins->texture.in_reg_upper = 0; ins->texture.in_reg_select = coord.reg & 1; - offset_swizzle(ins->swizzle[1], coord.offset, coord.size, 0); + offset_swizzle(ins->swizzle[1], coord.offset, coord.size, dest.size, 0); /* Next, install the destination */ ins->texture.out_full = 1; ins->texture.out_upper = 0; ins->texture.out_reg_select = dest.reg & 1; - offset_swizzle(ins->swizzle[0], 0, 4, dest.offset); + offset_swizzle(ins->swizzle[0], 0, 4, dest.size, dest.offset); mir_set_bytemask(ins, mir_bytemask(ins) << dest.offset); /* If there is a register LOD/bias, use it */ @@ -901,7 +907,7 @@ mir_spill_register( /* Mask the load based on the component count * actually needed to prevent RA loops */ - st.mask = mir_from_bytemask(read_bytemask, midgard_reg_mode_32); + st.mask = mir_from_bytemask(read_bytemask, 32); mir_insert_instruction_before_scheduled(ctx, block, before, st); } else { diff --git a/src/panfrost/midgard/midgard_schedule.c b/src/panfrost/midgard/midgard_schedule.c index c77eee16fce..1277608e373 100644 --- a/src/panfrost/midgard/midgard_schedule.c +++ b/src/panfrost/midgard/midgard_schedule.c @@ -381,8 +381,6 @@ mir_adjust_constants(midgard_instruction *ins, return true; unsigned r_constant = SSA_FIXED_REGISTER(REGISTER_CONSTANT); - midgard_reg_mode dst_mode = mir_typesize(ins); - unsigned bundle_constant_mask = pred->constant_mask; unsigned comp_mapping[2][16] = { }; uint8_t bundle_constants[16]; @@ -396,13 +394,12 @@ mir_adjust_constants(midgard_instruction *ins, if (ins->src[src] != SSA_FIXED_REGISTER(REGISTER_CONSTANT)) continue; - midgard_reg_mode src_mode = mir_srcsize(ins, src); - unsigned type_size = mir_bytes_for_mode(src_mode); + unsigned type_size = nir_alu_type_get_type_size(ins->src_types[src]) / 8; unsigned max_comp = 16 / type_size; unsigned comp_mask = mir_from_bytemask(mir_round_bytemask_up( mir_bytemask_of_read_components_index(ins, src), - dst_mode), - dst_mode); + type_size * 8), + type_size * 8); unsigned type_mask = (1 << type_size) - 1; for (unsigned comp = 0; comp < max_comp; comp++) { diff --git a/src/panfrost/midgard/mir.c b/src/panfrost/midgard/mir.c index 0289e5e3496..63665ddfe91 100644 --- a/src/panfrost/midgard/mir.c +++ b/src/panfrost/midgard/mir.c @@ -176,120 +176,11 @@ mir_nontrivial_outmod(midgard_instruction *ins) return mod != midgard_outmod_none; } -/* Grabs the type size. */ - -midgard_reg_mode -mir_typesize(midgard_instruction *ins) -{ - if (ins->compact_branch) - return midgard_reg_mode_32; - - /* TODO: Type sizes for texture */ - if (ins->type == TAG_TEXTURE_4) - return midgard_reg_mode_32; - - if (ins->type == TAG_LOAD_STORE_4) - return GET_LDST_SIZE(load_store_opcode_props[ins->load_store.op].props); - - if (ins->type == TAG_ALU_4) { - midgard_reg_mode mode = ins->alu.reg_mode; - - /* If we have an override, step down by half */ - if (ins->alu.dest_override != midgard_dest_override_none) { - assert(mode > midgard_reg_mode_8); - mode--; - } - - return mode; - } - - unreachable("Invalid instruction type"); -} - -/* Grabs the size of a source */ - -midgard_reg_mode -mir_srcsize(midgard_instruction *ins, unsigned i) -{ - if (ins->type == TAG_LOAD_STORE_4) { - if (OP_HAS_ADDRESS(ins->load_store.op)) { - if (i == 1) - return midgard_reg_mode_64; - else if (i == 2) { - bool zext = ins->load_store.arg_1 & 0x80; - return zext ? midgard_reg_mode_32 : midgard_reg_mode_64; - } - } - } - - /* TODO: 16-bit textures/ldst */ - if (ins->type == TAG_TEXTURE_4 || ins->type == TAG_LOAD_STORE_4) - return midgard_reg_mode_32; - - /* TODO: 16-bit branches */ - if (ins->compact_branch) - return midgard_reg_mode_32; - - if (i >= 2) { - /* TODO: 16-bit conditions, ffma */ - return midgard_reg_mode_32; - } - - /* Default to type of the instruction */ - - midgard_reg_mode mode = ins->alu.reg_mode; - - /* If we have a half modifier, step down by half */ - - if ((mir_get_alu_src(ins, i)).half) { - assert(mode > midgard_reg_mode_8); - mode--; - } - - return mode; -} - -midgard_reg_mode -mir_mode_for_destsize(unsigned size) -{ - switch (size) { - 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("Unknown destination size"); - } -} - -/* ...and the inverse */ - -unsigned -mir_bytes_for_mode(midgard_reg_mode mode) -{ - switch (mode) { - case midgard_reg_mode_8: - return 1; - case midgard_reg_mode_16: - return 2; - case midgard_reg_mode_32: - return 4; - case midgard_reg_mode_64: - return 8; - default: - unreachable("Invalid register mode"); - } -} - uint16_t -mir_from_bytemask(uint16_t bytemask, midgard_reg_mode mode) +mir_from_bytemask(uint16_t bytemask, unsigned bits) { unsigned value = 0; - unsigned count = mir_bytes_for_mode(mode); + unsigned count = bits / 8; for (unsigned c = 0, d = 0; c < 16; c += count, ++d) { bool a = (bytemask & (1 << c)) != 0; @@ -307,9 +198,9 @@ mir_from_bytemask(uint16_t bytemask, midgard_reg_mode mode) * component, and check if any bytes in the component are masked on */ uint16_t -mir_round_bytemask_up(uint16_t mask, midgard_reg_mode mode) +mir_round_bytemask_up(uint16_t mask, unsigned bits) { - unsigned bytes = mir_bytes_for_mode(mode); + unsigned bytes = bits / 8; unsigned maxmask = mask_of(bytes); unsigned channels = 16 / bytes; @@ -328,13 +219,15 @@ mir_round_bytemask_up(uint16_t mask, midgard_reg_mode mode) uint16_t mir_bytemask(midgard_instruction *ins) { - return pan_to_bytemask(mir_bytes_for_mode(mir_typesize(ins)) * 8, ins->mask); + unsigned type_size = nir_alu_type_get_type_size(ins->dest_type); + return pan_to_bytemask(type_size, ins->mask); } void mir_set_bytemask(midgard_instruction *ins, uint16_t bytemask) { - ins->mask = mir_from_bytemask(bytemask, mir_typesize(ins)); + unsigned type_size = nir_alu_type_get_type_size(ins->dest_type); + ins->mask = mir_from_bytemask(bytemask, type_size); } /* Checks if we should use an upper destination override, rather than the lower @@ -350,14 +243,13 @@ mir_upper_override(midgard_instruction *ins) /* Make sure we didn't already lower somehow */ assert(ins->alu.dest_override == midgard_dest_override_lower); - /* What is the mask in terms of currently? */ - midgard_reg_mode type = mir_typesize(ins); - /* There are 16 bytes per vector, so there are (16/bytes) * components per vector. So the magic half is half of - * (16/bytes), which simplifies to 8/bytes */ + * (16/bytes), which simplifies to 8/bytes = 8 / (bits / 8) = 64 / bits + * */ - unsigned threshold = 8 / mir_bytes_for_mode(type); + unsigned type_size = nir_alu_type_get_type_size(ins->dest_type); + unsigned threshold = 64 / type_size; /* How many components did we shift over? */ unsigned zeroes = __builtin_ctz(ins->mask); @@ -375,7 +267,7 @@ mir_upper_override(midgard_instruction *ins) */ static uint16_t -mir_bytemask_of_read_components_single(unsigned *swizzle, unsigned inmask, midgard_reg_mode mode) +mir_bytemask_of_read_components_single(unsigned *swizzle, unsigned inmask, unsigned bits) { unsigned cmask = 0; @@ -384,7 +276,7 @@ mir_bytemask_of_read_components_single(unsigned *swizzle, unsigned inmask, midga cmask |= (1 << swizzle[c]); } - return pan_to_bytemask(mir_bytes_for_mode(mode) * 8, cmask); + return pan_to_bytemask(bits, cmask); } uint16_t @@ -423,7 +315,8 @@ mir_bytemask_of_read_components_index(midgard_instruction *ins, unsigned i) qmask = mask_of(channel_override); } - return mir_bytemask_of_read_components_single(ins->swizzle[i], qmask, mir_srcsize(ins, i)); + return mir_bytemask_of_read_components_single(ins->swizzle[i], qmask, + nir_alu_type_get_type_size(ins->src_types[i])); } uint16_t diff --git a/src/panfrost/midgard/mir_promote_uniforms.c b/src/panfrost/midgard/mir_promote_uniforms.c index d78484c3568..e29de78808d 100644 --- a/src/panfrost/midgard/mir_promote_uniforms.c +++ b/src/panfrost/midgard/mir_promote_uniforms.c @@ -197,7 +197,8 @@ midgard_promote_uniforms(compiler_context *ctx) if (ins->load_64) mov.alu.reg_mode = midgard_reg_mode_64; - uint16_t rounded = mir_round_bytemask_up(mir_bytemask(ins), mov.alu.reg_mode); + unsigned type_size = nir_alu_type_get_type_size(ins->dest_type); + uint16_t rounded = mir_round_bytemask_up(mir_bytemask(ins), type_size); mir_set_bytemask(&mov, rounded); mir_insert_instruction_before(ctx, ins, mov); } else { diff --git a/src/panfrost/util/pan_ir.c b/src/panfrost/util/pan_ir.c index 895c9b67f03..4fdb9ebabff 100644 --- a/src/panfrost/util/pan_ir.c +++ b/src/panfrost/util/pan_ir.c @@ -33,6 +33,10 @@ uint16_t pan_to_bytemask(unsigned bytes, unsigned mask) { switch (bytes) { + case 0: + assert(mask == 0); + return 0; + case 8: return mask; -- 2.30.2