X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fpanfrost%2Fmidgard%2Fmidgard_compile.c;h=1b9ea1bd7516c7ab8adf020251cc18997a9e72fe;hb=3e2cb21e53cf9d5a3ef984ef5abbfd9683e516a8;hp=71191083794e58ca1383a0beb64404682f017cff;hpb=5f7e0185cd9e93e19e3bbf9236458234f6448044;p=mesa.git diff --git a/src/panfrost/midgard/midgard_compile.c b/src/panfrost/midgard/midgard_compile.c index 71191083794..1b9ea1bd751 100644 --- a/src/panfrost/midgard/midgard_compile.c +++ b/src/panfrost/midgard/midgard_compile.c @@ -110,8 +110,8 @@ schedule_barrier(compiler_context *ctx) .dest = ~0, \ .src = { ~0, ~0, ~0, ~0 }, \ .swizzle = SWIZZLE_IDENTITY_4, \ + .op = midgard_op_##name, \ .load_store = { \ - .op = midgard_op_##name, \ .address = address \ } \ }; \ @@ -160,37 +160,6 @@ v_branch(bool conditional, bool invert) 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) { @@ -736,6 +705,34 @@ mir_accept_dest_mod(compiler_context *ctx, nir_dest **dest, nir_op op) 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, enum midgard_roundmode *roundmode, bool is_int, unsigned bcast_count) { @@ -861,8 +858,8 @@ emit_alu(compiler_context *ctx, nir_alu_instr *instr) /* 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; @@ -884,6 +881,8 @@ emit_alu(compiler_context *ctx, nir_alu_instr *instr) 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); @@ -891,7 +890,7 @@ emit_alu(compiler_context *ctx, nir_alu_instr *instr) ALU_CASE(mov, imov); ALU_CASE_CMP(feq32, feq, false); - ALU_CASE_CMP(fne32, fne, 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); @@ -1062,7 +1061,9 @@ emit_alu(compiler_context *ctx, nir_alu_instr *instr) unsigned outmod = 0; 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; @@ -1076,31 +1077,8 @@ emit_alu(compiler_context *ctx, nir_alu_instr *instr) 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 (!midgard_is_integer_out_op(op)) { - 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; + outmod = mir_determine_float_outmod(ctx, &dest, outmod); } midgard_instruction ins = { @@ -1161,10 +1139,6 @@ emit_alu(compiler_context *ctx, nir_alu_instr *instr) ins.mask = mask_of(nr_components); - midgard_vector_alu alu = { - .outmod = outmod, - }; - /* 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.*/ @@ -1172,9 +1146,8 @@ emit_alu(compiler_context *ctx, nir_alu_instr *instr) if (!is_ssa) ins.mask &= instr->dest.write_mask; - ins.alu = alu; - ins.op = op; + ins.outmod = outmod; /* Late fixup for emulated instructions */ @@ -1412,16 +1385,16 @@ emit_varying_read( switch (type) { case nir_type_uint32: case nir_type_bool32: - ins.load_store.op = midgard_op_ld_vary_32u; + ins.op = midgard_op_ld_vary_32u; break; case nir_type_int32: - ins.load_store.op = midgard_op_ld_vary_32i; + ins.op = midgard_op_ld_vary_32i; break; case nir_type_float32: - ins.load_store.op = midgard_op_ld_vary_32; + ins.op = midgard_op_ld_vary_32; break; case nir_type_float16: - ins.load_store.op = midgard_op_ld_vary_16; + ins.op = midgard_op_ld_vary_16; break; default: unreachable("Attempted to load unknown type"); @@ -1446,13 +1419,13 @@ emit_attr_read( 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"); @@ -1583,7 +1556,7 @@ 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.load_store.op = midgard_op_ld_color_buffer_32u_old; + ld.op = midgard_op_ld_color_buffer_32u_old; ld.load_store.address = 97; ld.load_store.arg_2 = 0x1E; @@ -1600,13 +1573,7 @@ emit_control_barrier(compiler_context *ctx) .type = TAG_TEXTURE_4, .dest = ~0, .src = { ~0, ~0, ~0, ~0 }, - .texture = { - .op = TEXTURE_OP_BARRIER, - - /* TODO: optimize */ - .out_of_order = MIDGARD_BARRIER_BUFFER | - MIDGARD_BARRIER_SHARED , - } + .op = TEXTURE_OP_BARRIER, }; emit_mir_instruction(ctx, ins); @@ -1776,7 +1743,7 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr) } if (ctx->quirks & MIDGARD_OLD_BLEND) { - ld.load_store.op = midgard_op_ld_color_buffer_32u_old; + ld.op = midgard_op_ld_color_buffer_32u_old; ld.load_store.address = 16; ld.load_store.arg_2 = 0x1E; } @@ -1803,9 +1770,9 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr) if (ctx->quirks & MIDGARD_OLD_BLEND) { if (bits == 16) - ld.load_store.op = midgard_op_ld_color_buffer_as_fp16_old; + ld.op = midgard_op_ld_color_buffer_as_fp16_old; else - ld.load_store.op = midgard_op_ld_color_buffer_as_fp32_old; + ld.op = midgard_op_ld_color_buffer_as_fp32_old; ld.load_store.address = 1; ld.load_store.arg_2 = 0x1E; } @@ -1873,7 +1840,7 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr) else if (combined) rt = MIDGARD_ZS_RT; else - assert(0); + unreachable("bad rt"); unsigned reg_z = ~0, reg_s = ~0; if (combined) { @@ -1908,13 +1875,13 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr) 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"); @@ -2005,25 +1972,26 @@ emit_intrinsic(compiler_context *ctx, nir_intrinsic_instr *instr) } } +/* 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"); @@ -2072,21 +2040,30 @@ emit_texop_native(compiler_context *ctx, nir_tex_instr *instr, /* 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, @@ -2419,6 +2396,13 @@ max_bitsize_for_alu(midgard_instruction *ins) 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; } @@ -2803,7 +2787,7 @@ emit_cf_list(struct compiler_context *ctx, struct exec_list *list) * 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); @@ -2869,7 +2853,6 @@ midgard_compile_shader_nir(nir_shader *nir, panfrost_program *program, bool is_b 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; @@ -2883,7 +2866,6 @@ midgard_compile_shader_nir(nir_shader *nir, panfrost_program *program, bool is_b /* 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 @@ -2926,7 +2908,7 @@ midgard_compile_shader_nir(nir_shader *nir, panfrost_program *program, bool is_b /* 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); @@ -2999,124 +2981,6 @@ midgard_compile_shader_nir(nir_shader *nir, panfrost_program *program, bool is_b 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_tilebuf_wait = ins->branch.target_type == TARGET_TILEBUF_WAIT; - bool is_special = is_discard || is_tilebuf_wait; - 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 : - is_tilebuf_wait ? bundle->tag : - 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 (is_tilebuf_wait) { - quadword_offset = -1; - } 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_tilebuf_wait ? midgard_jmp_writeout_op_tilebuffer_pending : - 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_special) { - 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 */