X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;f=src%2Fpanfrost%2Fbifrost%2Fbifrost_compile.c;h=413c8d3f2c84764db524f3b37aca493f789fdb1d;hb=HEAD;hp=0f1922c9513f0a080f1f13463d542c6d38a312f9;hpb=330e9a66968056330ac030e4ecb738eb3c8489e9;p=mesa.git diff --git a/src/panfrost/bifrost/bifrost_compile.c b/src/panfrost/bifrost/bifrost_compile.c index 0f1922c9513..413c8d3f2c8 100644 --- a/src/panfrost/bifrost/bifrost_compile.c +++ b/src/panfrost/bifrost/bifrost_compile.c @@ -27,19 +27,33 @@ #include "main/mtypes.h" #include "compiler/glsl/glsl_to_nir.h" #include "compiler/nir_types.h" -#include "main/imports.h" #include "compiler/nir/nir_builder.h" +#include "util/u_debug.h" #include "disassemble.h" #include "bifrost_compile.h" +#include "bifrost_nir.h" #include "compiler.h" #include "bi_quirks.h" #include "bi_print.h" +static const struct debug_named_value debug_options[] = { + {"msgs", BIFROST_DBG_MSGS, "Print debug messages"}, + {"shaders", BIFROST_DBG_SHADERS, "Dump shaders in NIR and MIR"}, + DEBUG_NAMED_VALUE_END +}; + +DEBUG_GET_ONCE_FLAGS_OPTION(bifrost_debug, "BIFROST_MESA_DEBUG", debug_options, 0) + +int bifrost_debug = 0; + +#define DBG(fmt, ...) \ + do { if (bifrost_debug & BIFROST_DBG_MSGS) \ + fprintf(stderr, "%s:%d: "fmt, \ + __FUNCTION__, __LINE__, ##__VA_ARGS__); } while (0) + static bi_block *emit_cf_list(bi_context *ctx, struct exec_list *list); static bi_instruction *bi_emit_branch(bi_context *ctx); -static void bi_block_add_successor(bi_block *block, bi_block *successor); -static void bi_schedule_barrier(bi_context *ctx); static void emit_jump(bi_context *ctx, nir_jump_instr *instr) @@ -48,23 +62,16 @@ emit_jump(bi_context *ctx, nir_jump_instr *instr) switch (instr->type) { case nir_jump_break: - branch->branch.target = ctx->break_block; + branch->branch_target = ctx->break_block; break; case nir_jump_continue: - branch->branch.target = ctx->continue_block; + branch->branch_target = ctx->continue_block; break; default: unreachable("Unhandled jump type"); } - bi_block_add_successor(ctx->current_block, branch->branch.target); -} - -/* Gets a bytemask for a complete vecN write */ -static unsigned -bi_mask_for_channels_32(unsigned i) -{ - return (1 << (4 * i)) - 1; + pan_block_add_successor(&ctx->current_block->base, &branch->branch_target->base); } static bi_instruction @@ -72,17 +79,18 @@ bi_load(enum bi_class T, nir_intrinsic_instr *instr) { bi_instruction load = { .type = T, - .writemask = bi_mask_for_channels_32(instr->num_components), + .vector_channels = instr->num_components, .src = { BIR_INDEX_CONSTANT }, + .src_types = { nir_type_uint32 }, .constant = { .u64 = nir_intrinsic_base(instr) }, }; const nir_intrinsic_info *info = &nir_intrinsic_infos[instr->intrinsic]; if (info->has_dest) - load.dest = bir_dest_index(&instr->dest); + load.dest = pan_dest_index(&instr->dest); - if (info->has_dest && info->index_map[NIR_INTRINSIC_TYPE] > 0) + if (info->has_dest && nir_intrinsic_has_type(instr)) load.dest_type = nir_intrinsic_type(instr); nir_src *offset = nir_get_io_offset_src(instr); @@ -90,7 +98,7 @@ bi_load(enum bi_class T, nir_intrinsic_instr *instr) if (nir_src_is_const(*offset)) load.constant.u64 += nir_src_as_uint(*offset); else - load.src[0] = bir_src_index(offset); + load.src[0] = pan_src_index(offset); return load; } @@ -102,7 +110,16 @@ bi_emit_ld_vary(bi_context *ctx, nir_intrinsic_instr *instr) ins.load_vary.interp_mode = BIFROST_INTERP_DEFAULT; /* TODO */ ins.load_vary.reuse = false; /* TODO */ ins.load_vary.flat = instr->intrinsic != nir_intrinsic_load_interpolated_input; - ins.dest_type = nir_type_float | nir_dest_bit_size(instr->dest), + ins.dest_type = nir_type_float | nir_dest_bit_size(instr->dest); + + if (nir_src_is_const(*nir_get_io_offset_src(instr))) { + /* Zero it out for direct */ + ins.src[1] = BIR_INDEX_ZERO; + } else { + /* R61 contains sample mask stuff, TODO RA XXX */ + ins.src[1] = BIR_INDEX_REGISTER | 61; + } + bi_emit(ctx, ins); } @@ -111,11 +128,24 @@ bi_emit_frag_out(bi_context *ctx, nir_intrinsic_instr *instr) { if (!ctx->emitted_atest) { bi_instruction ins = { - .type = BI_ATEST + .type = BI_ATEST, + .src = { + BIR_INDEX_REGISTER | 60 /* TODO: RA */, + pan_src_index(&instr->src[0]) + }, + .src_types = { + nir_type_uint32, + nir_intrinsic_type(instr) + }, + .swizzle = { + { 0 }, + { 3, 0 } /* swizzle out the alpha */ + }, + .dest = BIR_INDEX_REGISTER | 60 /* TODO: RA */, + .dest_type = nir_type_uint32, }; bi_emit(ctx, ins); - bi_schedule_barrier(ctx); ctx->emitted_atest = true; } @@ -123,36 +153,74 @@ bi_emit_frag_out(bi_context *ctx, nir_intrinsic_instr *instr) .type = BI_BLEND, .blend_location = nir_intrinsic_base(instr), .src = { - bir_src_index(&instr->src[0]) + pan_src_index(&instr->src[0]), + BIR_INDEX_REGISTER | 60 /* Can this be arbitrary? */, + }, + .src_types = { + nir_intrinsic_type(instr), + nir_type_uint32 }, .swizzle = { - { 0, 1, 2, 3 } - } + { 0, 1, 2, 3 }, + { 0 } + }, + .dest = BIR_INDEX_REGISTER | 48 /* Looks like magic */, + .dest_type = nir_type_uint32, + .vector_channels = 4 }; + assert(blend.blend_location < 8); + assert(ctx->blend_types); + assert(blend.src_types[0]); + ctx->blend_types[blend.blend_location] = blend.src_types[0]; + bi_emit(ctx, blend); - bi_schedule_barrier(ctx); +} + +static bi_instruction +bi_load_with_r61(enum bi_class T, nir_intrinsic_instr *instr) +{ + bi_instruction ld = bi_load(T, instr); + ld.src[1] = BIR_INDEX_REGISTER | 61; /* TODO: RA */ + ld.src[2] = BIR_INDEX_REGISTER | 62; + ld.src[3] = 0; + ld.src_types[1] = nir_type_uint32; + ld.src_types[2] = nir_type_uint32; + ld.src_types[3] = nir_intrinsic_type(instr); + return ld; } static void bi_emit_st_vary(bi_context *ctx, nir_intrinsic_instr *instr) { - bi_instruction address = bi_load(BI_LOAD_VAR_ADDRESS, instr); + bi_instruction address = bi_load_with_r61(BI_LOAD_VAR_ADDRESS, instr); address.dest = bi_make_temp(ctx); - address.dest_type = nir_type_uint64; - address.writemask = (1 << 8) - 1; + address.dest_type = nir_type_uint32; + address.vector_channels = 3; + + unsigned nr = nir_intrinsic_src_components(instr, 0); + assert(nir_intrinsic_write_mask(instr) == ((1 << nr) - 1)); bi_instruction st = { .type = BI_STORE_VAR, .src = { - address.dest, - bir_src_index(&instr->src[0]) + pan_src_index(&instr->src[0]), + address.dest, address.dest, address.dest, + }, + .src_types = { + nir_type_uint32, + nir_type_uint32, nir_type_uint32, nir_type_uint32, }, .swizzle = { - { 0, 1, 2, 3 } - } + { 0 }, + { 0 }, { 1 }, { 2} + }, + .vector_channels = nr, }; + for (unsigned i = 0; i < nr; ++i) + st.swizzle[0][i] = i; + bi_emit(ctx, address); bi_emit(ctx, st); } @@ -162,9 +230,173 @@ bi_emit_ld_uniform(bi_context *ctx, nir_intrinsic_instr *instr) { bi_instruction ld = bi_load(BI_LOAD_UNIFORM, instr); ld.src[1] = BIR_INDEX_ZERO; /* TODO: UBO index */ + + /* TODO: Indirect access, since we need to multiply by the element + * size. I believe we can get this lowering automatically via + * nir_lower_io (as mul instructions) with the proper options, but this + * is TODO */ + assert(ld.src[0] & BIR_INDEX_CONSTANT); + ld.constant.u64 += ctx->sysvals.sysval_count; + ld.constant.u64 *= 16; + bi_emit(ctx, ld); } +static void +bi_emit_sysval(bi_context *ctx, nir_instr *instr, + unsigned nr_components, unsigned offset) +{ + nir_dest nir_dest; + + /* Figure out which uniform this is */ + int sysval = panfrost_sysval_for_instr(instr, &nir_dest); + void *val = _mesa_hash_table_u64_search(ctx->sysvals.sysval_to_id, sysval); + + /* Sysvals are prefix uniforms */ + unsigned uniform = ((uintptr_t) val) - 1; + + /* Emit the read itself -- this is never indirect */ + + bi_instruction load = { + .type = BI_LOAD_UNIFORM, + .vector_channels = nr_components, + .src = { BIR_INDEX_CONSTANT, BIR_INDEX_ZERO }, + .src_types = { nir_type_uint32, nir_type_uint32 }, + .constant = { (uniform * 16) + offset }, + .dest = pan_dest_index(&nir_dest), + .dest_type = nir_type_uint32, /* TODO */ + }; + + bi_emit(ctx, load); +} + +/* gl_FragCoord.xy = u16_to_f32(R59.xy) + 0.5 + * gl_FragCoord.z = ld_vary(fragz) + * gl_FragCoord.w = ld_vary(fragw) + */ + +static void +bi_emit_ld_frag_coord(bi_context *ctx, nir_intrinsic_instr *instr) +{ + /* Future proofing for mediump fragcoord at some point.. */ + nir_alu_type T = nir_type_float32; + + /* First, sketch a combine */ + bi_instruction combine = { + .type = BI_COMBINE, + .dest_type = nir_type_uint32, + .dest = pan_dest_index(&instr->dest), + .src_types = { T, T, T, T }, + }; + + /* Second, handle xy */ + for (unsigned i = 0; i < 2; ++i) { + bi_instruction conv = { + .type = BI_CONVERT, + .dest_type = T, + .dest = bi_make_temp(ctx), + .src = { + /* TODO: RA XXX */ + BIR_INDEX_REGISTER | 59 + }, + .src_types = { nir_type_uint16 }, + .swizzle = { { i } } + }; + + bi_instruction add = { + .type = BI_ADD, + .dest_type = T, + .dest = bi_make_temp(ctx), + .src = { conv.dest, BIR_INDEX_CONSTANT }, + .src_types = { T, T }, + }; + + float half = 0.5; + memcpy(&add.constant.u32, &half, sizeof(float)); + + bi_emit(ctx, conv); + bi_emit(ctx, add); + + combine.src[i] = add.dest; + } + + /* Third, zw */ + for (unsigned i = 0; i < 2; ++i) { + bi_instruction load = { + .type = BI_LOAD_VAR, + .load_vary = { + .interp_mode = BIFROST_INTERP_DEFAULT, + .reuse = false, + .flat = true + }, + .vector_channels = 1, + .dest_type = nir_type_float32, + .dest = bi_make_temp(ctx), + .src = { BIR_INDEX_CONSTANT, BIR_INDEX_ZERO }, + .src_types = { nir_type_uint32, nir_type_uint32 }, + .constant = { + .u32 = (i == 0) ? BIFROST_FRAGZ : BIFROST_FRAGW + } + }; + + bi_emit(ctx, load); + + combine.src[i + 2] = load.dest; + } + + /* Finally, emit the combine */ + bi_emit(ctx, combine); +} + +static void +bi_emit_discard(bi_context *ctx, nir_intrinsic_instr *instr) +{ + /* Goofy lowering */ + bi_instruction discard = { + .type = BI_DISCARD, + .cond = BI_COND_EQ, + .src_types = { nir_type_uint32, nir_type_uint32 }, + .src = { BIR_INDEX_ZERO, BIR_INDEX_ZERO }, + }; + + bi_emit(ctx, discard); +} + +static void +bi_fuse_cond(bi_instruction *csel, nir_alu_src cond, + unsigned *constants_left, unsigned *constant_shift, + unsigned comps, bool float_only); + +static void +bi_emit_discard_if(bi_context *ctx, nir_intrinsic_instr *instr) +{ + nir_src cond = instr->src[0]; + nir_alu_type T = nir_type_uint | nir_src_bit_size(cond); + + bi_instruction discard = { + .type = BI_DISCARD, + .cond = BI_COND_NE, + .src_types = { T, T }, + .src = { + pan_src_index(&cond), + BIR_INDEX_ZERO + }, + }; + + /* Try to fuse in the condition */ + unsigned constants_left = 1, constant_shift = 0; + + /* Scalar so no swizzle */ + nir_alu_src wrap = { + .src = instr->src[0] + }; + + /* May or may not succeed but we're optimistic */ + bi_fuse_cond(&discard, wrap, &constants_left, &constant_shift, 1, true); + + bi_emit(ctx, discard); +} + static void emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr) { @@ -178,7 +410,7 @@ emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr) if (ctx->stage == MESA_SHADER_FRAGMENT) bi_emit_ld_vary(ctx, instr); else if (ctx->stage == MESA_SHADER_VERTEX) - bi_emit(ctx, bi_load(BI_LOAD_ATTR, instr)); + bi_emit(ctx, bi_load_with_r61(BI_LOAD_ATTR, instr)); else { unreachable("Unsupported shader stage"); } @@ -197,8 +429,35 @@ emit_intrinsic(bi_context *ctx, nir_intrinsic_instr *instr) bi_emit_ld_uniform(ctx, instr); break; + case nir_intrinsic_load_frag_coord: + bi_emit_ld_frag_coord(ctx, instr); + break; + + case nir_intrinsic_discard: + bi_emit_discard(ctx, instr); + break; + + case nir_intrinsic_discard_if: + bi_emit_discard_if(ctx, instr); + break; + + case nir_intrinsic_load_ssbo_address: + bi_emit_sysval(ctx, &instr->instr, 1, 0); + break; + + case nir_intrinsic_get_buffer_size: + bi_emit_sysval(ctx, &instr->instr, 1, 8); + break; + + case nir_intrinsic_load_viewport_scale: + case nir_intrinsic_load_viewport_offset: + case nir_intrinsic_load_num_work_groups: + case nir_intrinsic_load_sampler_lod_parameters_pan: + bi_emit_sysval(ctx, &instr->instr, 3, 0); + break; + default: - /* todo */ + unreachable("Unknown intrinsic"); break; } } @@ -207,24 +466,558 @@ static void emit_load_const(bi_context *ctx, nir_load_const_instr *instr) { /* Make sure we've been lowered */ - assert(instr->def.num_components == 1); + assert(instr->def.num_components <= (32 / instr->def.bit_size)); + + /* Accumulate all the channels of the constant, as if we did an + * implicit SEL over them */ + uint32_t acc = 0; + + for (unsigned i = 0; i < instr->def.num_components; ++i) { + unsigned v = nir_const_value_as_uint(instr->value[i], instr->def.bit_size); + acc |= (v << (i * instr->def.bit_size)); + } bi_instruction move = { .type = BI_MOV, - .dest = bir_ssa_index(&instr->def), - .dest_type = instr->def.bit_size | nir_type_uint, - .writemask = (1 << (instr->def.bit_size / 8)) - 1, + .dest = pan_ssa_index(&instr->def), + .dest_type = nir_type_uint32, .src = { BIR_INDEX_CONSTANT }, + .src_types = { + nir_type_uint32, + }, .constant = { - .u64 = nir_const_value_as_uint(instr->value[0], instr->def.bit_size) + .u32 = acc } }; bi_emit(ctx, move); } +#define BI_CASE_CMP(op) \ + case op##8: \ + case op##16: \ + case op##32: \ + +static enum bi_class +bi_class_for_nir_alu(nir_op op) +{ + switch (op) { + case nir_op_fadd: + case nir_op_fsub: + return BI_ADD; + + case nir_op_iadd: + case nir_op_isub: + return BI_IMATH; + + case nir_op_imul: + return BI_IMUL; + + case nir_op_iand: + case nir_op_ior: + case nir_op_ixor: + case nir_op_inot: + case nir_op_ishl: + return BI_BITWISE; + + BI_CASE_CMP(nir_op_flt) + BI_CASE_CMP(nir_op_fge) + BI_CASE_CMP(nir_op_feq) + BI_CASE_CMP(nir_op_fneu) + BI_CASE_CMP(nir_op_ilt) + BI_CASE_CMP(nir_op_ige) + BI_CASE_CMP(nir_op_ieq) + BI_CASE_CMP(nir_op_ine) + BI_CASE_CMP(nir_op_uge) + return BI_CMP; + + case nir_op_b8csel: + case nir_op_b16csel: + case nir_op_b32csel: + return BI_CSEL; + + case nir_op_i2i8: + case nir_op_i2i16: + case nir_op_i2i32: + case nir_op_i2i64: + case nir_op_u2u8: + case nir_op_u2u16: + case nir_op_u2u32: + case nir_op_u2u64: + case nir_op_f2i16: + case nir_op_f2i32: + case nir_op_f2i64: + case nir_op_f2u16: + case nir_op_f2u32: + case nir_op_f2u64: + case nir_op_i2f16: + case nir_op_i2f32: + case nir_op_i2f64: + case nir_op_u2f16: + case nir_op_u2f32: + case nir_op_u2f64: + case nir_op_f2f16: + case nir_op_f2f32: + case nir_op_f2f64: + case nir_op_f2fmp: + return BI_CONVERT; + + case nir_op_vec2: + case nir_op_vec3: + case nir_op_vec4: + return BI_COMBINE; + + case nir_op_vec8: + case nir_op_vec16: + unreachable("should've been lowered"); + + case nir_op_ffma: + case nir_op_fmul: + return BI_FMA; + + case nir_op_imin: + case nir_op_imax: + case nir_op_umin: + case nir_op_umax: + case nir_op_fmin: + case nir_op_fmax: + return BI_MINMAX; + + case nir_op_fsat: + case nir_op_fneg: + case nir_op_fabs: + return BI_FMOV; + case nir_op_mov: + return BI_MOV; + + case nir_op_fround_even: + case nir_op_fceil: + case nir_op_ffloor: + case nir_op_ftrunc: + return BI_ROUND; + + case nir_op_frcp: + case nir_op_frsq: + case nir_op_iabs: + return BI_SPECIAL; + + default: + unreachable("Unknown ALU op"); + } +} + +/* Gets a bi_cond for a given NIR comparison opcode. In soft mode, it will + * return BI_COND_ALWAYS as a sentinel if it fails to do so (when used for + * optimizations). Otherwise it will bail (when used for primary code + * generation). */ + +static enum bi_cond +bi_cond_for_nir(nir_op op, bool soft) +{ + switch (op) { + BI_CASE_CMP(nir_op_flt) + BI_CASE_CMP(nir_op_ilt) + return BI_COND_LT; + + BI_CASE_CMP(nir_op_fge) + BI_CASE_CMP(nir_op_ige) + BI_CASE_CMP(nir_op_uge) + return BI_COND_GE; + + BI_CASE_CMP(nir_op_feq) + BI_CASE_CMP(nir_op_ieq) + return BI_COND_EQ; + + BI_CASE_CMP(nir_op_fneu) + BI_CASE_CMP(nir_op_ine) + return BI_COND_NE; + default: + if (soft) + return BI_COND_ALWAYS; + else + unreachable("Invalid compare"); + } +} + +static void +bi_copy_src(bi_instruction *alu, nir_alu_instr *instr, unsigned i, unsigned to, + unsigned *constants_left, unsigned *constant_shift, unsigned comps) +{ + unsigned bits = nir_src_bit_size(instr->src[i].src); + unsigned dest_bits = nir_dest_bit_size(instr->dest.dest); + + alu->src_types[to] = nir_op_infos[instr->op].input_types[i] + | bits; + + /* Try to inline a constant */ + if (nir_src_is_const(instr->src[i].src) && *constants_left && (dest_bits == bits)) { + uint64_t mask = (1ull << dest_bits) - 1; + uint64_t cons = nir_src_as_uint(instr->src[i].src); + + /* Try to reuse a constant */ + for (unsigned i = 0; i < (*constant_shift); i += dest_bits) { + if (((alu->constant.u64 >> i) & mask) == cons) { + alu->src[to] = BIR_INDEX_CONSTANT | i; + return; + } + } + + alu->constant.u64 |= cons << *constant_shift; + alu->src[to] = BIR_INDEX_CONSTANT | (*constant_shift); + --(*constants_left); + (*constant_shift) += MAX2(dest_bits, 32); /* lo/hi */ + return; + } + + alu->src[to] = pan_src_index(&instr->src[i].src); + + /* Copy swizzle for all vectored components, replicating last component + * to fill undersized */ + + unsigned vec = alu->type == BI_COMBINE ? 1 : + MAX2(1, 32 / dest_bits); + + for (unsigned j = 0; j < vec; ++j) + alu->swizzle[to][j] = instr->src[i].swizzle[MIN2(j, comps - 1)]; +} + +static void +bi_fuse_cond(bi_instruction *csel, nir_alu_src cond, + unsigned *constants_left, unsigned *constant_shift, + unsigned comps, bool float_only) +{ + /* Bail for vector weirdness */ + if (cond.swizzle[0] != 0) + return; + + if (!cond.src.is_ssa) + return; + + nir_ssa_def *def = cond.src.ssa; + nir_instr *parent = def->parent_instr; + + if (parent->type != nir_instr_type_alu) + return; + + nir_alu_instr *alu = nir_instr_as_alu(parent); + + /* Try to match a condition */ + enum bi_cond bcond = bi_cond_for_nir(alu->op, true); + + if (bcond == BI_COND_ALWAYS) + return; + + /* Some instructions can't compare ints */ + if (float_only) { + nir_alu_type T = nir_op_infos[alu->op].input_types[0]; + T = nir_alu_type_get_base_type(T); + + if (T != nir_type_float) + return; + } + + /* We found one, let's fuse it in */ + csel->cond = bcond; + bi_copy_src(csel, alu, 0, 0, constants_left, constant_shift, comps); + bi_copy_src(csel, alu, 1, 1, constants_left, constant_shift, comps); +} + +static void +emit_alu(bi_context *ctx, nir_alu_instr *instr) +{ + /* Try some special functions */ + switch (instr->op) { + case nir_op_fexp2: + bi_emit_fexp2(ctx, instr); + return; + case nir_op_flog2: + bi_emit_flog2(ctx, instr); + return; + default: + break; + } + + /* Otherwise, assume it's something we can handle normally */ + bi_instruction alu = { + .type = bi_class_for_nir_alu(instr->op), + .dest = pan_dest_index(&instr->dest.dest), + .dest_type = nir_op_infos[instr->op].output_type + | nir_dest_bit_size(instr->dest.dest), + }; + + /* TODO: Implement lowering of special functions for older Bifrost */ + assert((alu.type != BI_SPECIAL) || !(ctx->quirks & BIFROST_NO_FAST_OP)); + + unsigned comps = nir_dest_num_components(instr->dest.dest); + + if (alu.type != BI_COMBINE) + assert(comps <= MAX2(1, 32 / comps)); + + if (!instr->dest.dest.is_ssa) { + for (unsigned i = 0; i < comps; ++i) + assert(instr->dest.write_mask); + } + + /* We inline constants as we go. This tracks how many constants have + * been inlined, since we're limited to 64-bits of constants per + * instruction */ + + unsigned dest_bits = nir_dest_bit_size(instr->dest.dest); + unsigned constants_left = (64 / dest_bits); + unsigned constant_shift = 0; + + if (alu.type == BI_COMBINE) + constants_left = 0; + + /* Copy sources */ + + unsigned num_inputs = nir_op_infos[instr->op].num_inputs; + assert(num_inputs <= ARRAY_SIZE(alu.src)); + + for (unsigned i = 0; i < num_inputs; ++i) { + unsigned f = 0; + + if (i && alu.type == BI_CSEL) + f++; + + bi_copy_src(&alu, instr, i, i + f, &constants_left, &constant_shift, comps); + } + + /* Op-specific fixup */ + switch (instr->op) { + case nir_op_fmul: + alu.src[2] = BIR_INDEX_ZERO; /* FMA */ + alu.src_types[2] = alu.src_types[1]; + break; + case nir_op_fsat: + alu.outmod = BIFROST_SAT; /* FMOV */ + break; + case nir_op_fneg: + alu.src_neg[0] = true; /* FMOV */ + break; + case nir_op_fabs: + alu.src_abs[0] = true; /* FMOV */ + break; + case nir_op_fsub: + alu.src_neg[1] = true; /* FADD */ + break; + case nir_op_iadd: + alu.op.imath = BI_IMATH_ADD; + break; + case nir_op_isub: + alu.op.imath = BI_IMATH_SUB; + break; + case nir_op_iabs: + alu.op.special = BI_SPECIAL_IABS; + break; + case nir_op_inot: + /* no dedicated bitwise not, but we can invert sources. convert to ~a | 0 */ + alu.op.bitwise = BI_BITWISE_OR; + alu.bitwise.src_invert[0] = true; + alu.src[1] = BIR_INDEX_ZERO; + /* zero shift */ + alu.src[2] = BIR_INDEX_ZERO; + alu.src_types[2] = alu.src_types[1]; + break; + case nir_op_ishl: + alu.op.bitwise = BI_BITWISE_OR; + /* move src1 to src2 and replace with zero. underlying op is (src0 << src2) | src1 */ + alu.src[2] = alu.src[1]; + alu.src_types[2] = alu.src_types[1]; + alu.src[1] = BIR_INDEX_ZERO; + break; + case nir_op_imul: + alu.op.imul = BI_IMUL_IMUL; + break; + case nir_op_fmax: + case nir_op_imax: + case nir_op_umax: + alu.op.minmax = BI_MINMAX_MAX; /* MINMAX */ + break; + case nir_op_frcp: + alu.op.special = BI_SPECIAL_FRCP; + break; + case nir_op_frsq: + alu.op.special = BI_SPECIAL_FRSQ; + break; + BI_CASE_CMP(nir_op_flt) + BI_CASE_CMP(nir_op_ilt) + BI_CASE_CMP(nir_op_fge) + BI_CASE_CMP(nir_op_ige) + BI_CASE_CMP(nir_op_feq) + BI_CASE_CMP(nir_op_ieq) + BI_CASE_CMP(nir_op_fneu) + BI_CASE_CMP(nir_op_ine) + BI_CASE_CMP(nir_op_uge) + alu.cond = bi_cond_for_nir(instr->op, false); + break; + case nir_op_fround_even: + alu.roundmode = BIFROST_RTE; + break; + case nir_op_fceil: + alu.roundmode = BIFROST_RTP; + break; + case nir_op_ffloor: + alu.roundmode = BIFROST_RTN; + break; + case nir_op_ftrunc: + alu.roundmode = BIFROST_RTZ; + break; + case nir_op_iand: + alu.op.bitwise = BI_BITWISE_AND; + /* zero shift */ + alu.src[2] = BIR_INDEX_ZERO; + alu.src_types[2] = alu.src_types[1]; + break; + case nir_op_ior: + alu.op.bitwise = BI_BITWISE_OR; + /* zero shift */ + alu.src[2] = BIR_INDEX_ZERO; + alu.src_types[2] = alu.src_types[1]; + break; + case nir_op_ixor: + alu.op.bitwise = BI_BITWISE_XOR; + /* zero shift */ + alu.src[2] = BIR_INDEX_ZERO; + alu.src_types[2] = alu.src_types[1]; + break; + case nir_op_f2i32: + alu.roundmode = BIFROST_RTZ; + break; + + case nir_op_f2f16: + case nir_op_i2i16: + case nir_op_u2u16: { + if (nir_src_bit_size(instr->src[0].src) != 32) + break; + + /* Should have been const folded */ + assert(!nir_src_is_const(instr->src[0].src)); + + alu.src_types[1] = alu.src_types[0]; + alu.src[1] = alu.src[0]; + + unsigned last = nir_dest_num_components(instr->dest.dest) - 1; + assert(last <= 1); + + alu.swizzle[1][0] = instr->src[0].swizzle[last]; + break; + } + + default: + break; + } + + if (alu.type == BI_CSEL) { + /* Default to csel3 */ + alu.cond = BI_COND_NE; + alu.src[1] = BIR_INDEX_ZERO; + alu.src_types[1] = alu.src_types[0]; + + /* TODO: Reenable cond fusing when we can split up registers + * when scheduling */ +#if 0 + bi_fuse_cond(&alu, instr->src[0], + &constants_left, &constant_shift, comps, false); +#endif + } + + bi_emit(ctx, alu); +} + +/* TEX_COMPACT instructions assume normal 2D f32 operation but are more + * space-efficient and with simpler RA/scheduling requirements*/ + +static void +emit_tex_compact(bi_context *ctx, nir_tex_instr *instr) +{ + bi_instruction tex = { + .type = BI_TEX, + .op = { .texture = BI_TEX_COMPACT }, + .texture = { + .texture_index = instr->texture_index, + .sampler_index = instr->sampler_index, + }, + .dest = pan_dest_index(&instr->dest), + .dest_type = instr->dest_type, + .src_types = { nir_type_float32, nir_type_float32 }, + .vector_channels = 4 + }; + + for (unsigned i = 0; i < instr->num_srcs; ++i) { + int index = pan_src_index(&instr->src[i].src); + + /* We were checked ahead-of-time */ + if (instr->src[i].src_type == nir_tex_src_lod) + continue; + + assert (instr->src[i].src_type == nir_tex_src_coord); + + tex.src[0] = index; + tex.src[1] = index; + tex.swizzle[0][0] = 0; + tex.swizzle[1][0] = 1; + } + + bi_emit(ctx, tex); +} + +static void +emit_tex_full(bi_context *ctx, nir_tex_instr *instr) +{ + unreachable("stub"); +} + +/* Normal textures ops are tex for frag shaders and txl for vertex shaders with + * lod a constant 0. Anything else needs a full texture op. */ + +static bool +bi_is_normal_tex(gl_shader_stage stage, nir_tex_instr *instr) +{ + if (stage == MESA_SHADER_FRAGMENT) + return instr->op == nir_texop_tex; + + if (instr->op != nir_texop_txl) + return false; + + for (unsigned i = 0; i < instr->num_srcs; ++i) { + if (instr->src[i].src_type != nir_tex_src_lod) + continue; + + nir_src src = instr->src[i].src; + + if (!nir_src_is_const(src)) + continue; + + if (nir_src_as_uint(src) != 0) + continue; + } + + return true; +} + +static void +emit_tex(bi_context *ctx, nir_tex_instr *instr) +{ + nir_alu_type base = nir_alu_type_get_base_type(instr->dest_type); + unsigned sz = nir_dest_bit_size(instr->dest); + instr->dest_type = base | sz; + + bool is_normal = bi_is_normal_tex(ctx->stage, instr); + bool is_2d = instr->sampler_dim == GLSL_SAMPLER_DIM_2D || + instr->sampler_dim == GLSL_SAMPLER_DIM_EXTERNAL; + bool is_f = base == nir_type_float && (sz == 16 || sz == 32); + + bool is_compact = is_normal && is_2d && is_f && !instr->is_shadow; + + if (is_compact) + emit_tex_compact(ctx, instr); + else + emit_tex_full(ctx, instr); +} + static void emit_instr(bi_context *ctx, struct nir_instr *instr) { @@ -237,7 +1030,6 @@ emit_instr(bi_context *ctx, struct nir_instr *instr) emit_intrinsic(ctx, nir_instr_as_intrinsic(instr)); break; -#if 0 case nir_instr_type_alu: emit_alu(ctx, nir_instr_as_alu(instr)); break; @@ -245,7 +1037,6 @@ emit_instr(bi_context *ctx, struct nir_instr *instr) case nir_instr_type_tex: emit_tex(ctx, nir_instr_as_tex(instr)); break; -#endif case nir_instr_type_jump: emit_jump(ctx, nir_instr_as_jump(instr)); @@ -256,7 +1047,7 @@ emit_instr(bi_context *ctx, struct nir_instr *instr) break; default: - //unreachable("Unhandled instruction type"); + unreachable("Unhandled instruction type"); break; } } @@ -268,49 +1059,13 @@ create_empty_block(bi_context *ctx) { bi_block *blk = rzalloc(ctx, bi_block); - blk->predecessors = _mesa_set_create(blk, + blk->base.predecessors = _mesa_set_create(blk, _mesa_hash_pointer, _mesa_key_pointer_equal); - blk->name = ctx->block_name_count++; - return blk; } -static void -bi_block_add_successor(bi_block *block, bi_block *successor) -{ - assert(block); - assert(successor); - - for (unsigned i = 0; i < ARRAY_SIZE(block->successors); ++i) { - if (block->successors[i]) { - if (block->successors[i] == successor) - return; - else - continue; - } - - block->successors[i] = successor; - _mesa_set_add(successor->predecessors, block); - return; - } - - unreachable("Too many successors"); -} - -static void -bi_schedule_barrier(bi_context *ctx) -{ - bi_block *temp = ctx->after_block; - ctx->after_block = create_empty_block(ctx); - list_addtail(&ctx->after_block->link, &ctx->blocks); - list_inithead(&ctx->after_block->instructions); - bi_block_add_successor(ctx->current_block, ctx->after_block); - ctx->current_block = ctx->after_block; - ctx->after_block = temp; -} - static bi_block * emit_block(bi_context *ctx, nir_block *block) { @@ -321,8 +1076,8 @@ emit_block(bi_context *ctx, nir_block *block) ctx->current_block = create_empty_block(ctx); } - list_addtail(&ctx->current_block->link, &ctx->blocks); - list_inithead(&ctx->current_block->instructions); + list_addtail(&ctx->current_block->base.link, &ctx->blocks); + list_inithead(&ctx->current_block->base.instructions); nir_foreach_instr(instr, block) { emit_instr(ctx, instr); @@ -340,9 +1095,7 @@ bi_emit_branch(bi_context *ctx) { bi_instruction branch = { .type = BI_BRANCH, - .branch = { - .cond = BI_COND_ALWAYS - } + .cond = BI_COND_ALWAYS }; return bi_emit(ctx, branch); @@ -359,10 +1112,11 @@ static void bi_set_branch_cond(bi_instruction *branch, nir_src *cond, bool invert) { /* TODO: Try to unwrap instead of always bailing */ - branch->src[0] = bir_src_index(cond); + branch->src[0] = pan_src_index(cond); branch->src[1] = BIR_INDEX_ZERO; - branch->src_types[0] = branch->src_types[1] = nir_type_uint16; - branch->branch.cond = invert ? BI_COND_EQ : BI_COND_NE; + branch->src_types[0] = branch->src_types[1] = nir_type_uint | + nir_src_bit_size(*cond); + branch->cond = invert ? BI_COND_EQ : BI_COND_NE; } static void @@ -396,19 +1150,17 @@ emit_if(bi_context *ctx, nir_if *nif) if (ctx->instruction_count == count_in) { /* The else block is empty, so don't emit an exit jump */ bi_remove_instruction(then_exit); - then_branch->branch.target = ctx->after_block; + then_branch->branch_target = ctx->after_block; + pan_block_add_successor(&end_then_block->base, &ctx->after_block->base); /* fallthrough */ } else { - then_branch->branch.target = else_block; - then_exit->branch.target = ctx->after_block; - bi_block_add_successor(end_then_block, then_exit->branch.target); + then_branch->branch_target = else_block; + then_exit->branch_target = ctx->after_block; + pan_block_add_successor(&end_then_block->base, &then_exit->branch_target->base); + pan_block_add_successor(&end_else_block->base, &ctx->after_block->base); /* fallthrough */ } - /* Wire up the successors */ - - bi_block_add_successor(before_block, then_branch->branch.target); /* then_branch */ - - bi_block_add_successor(before_block, then_block); /* fallthrough */ - bi_block_add_successor(end_else_block, ctx->after_block); /* fallthrough */ + pan_block_add_successor(&before_block->base, &then_branch->branch_target->base); /* then_branch */ + pan_block_add_successor(&before_block->base, &then_block->base); /* fallthrough */ } static void @@ -429,9 +1181,9 @@ emit_loop(bi_context *ctx, nir_loop *nloop) /* Branch back to loop back */ bi_instruction *br_back = bi_emit_branch(ctx); - br_back->branch.target = ctx->continue_block; - bi_block_add_successor(start_block, ctx->continue_block); - bi_block_add_successor(ctx->current_block, ctx->continue_block); + br_back->branch_target = ctx->continue_block; + pan_block_add_successor(&start_block->base, &ctx->continue_block->base); + pan_block_add_successor(&ctx->current_block->base, &ctx->continue_block->base); ctx->after_block = ctx->break_block; @@ -520,8 +1272,7 @@ bi_optimize_nir(nir_shader *nir) nir, nir_lower_flrp, lower_flrp, - false /* always_precise */, - nir->options->lower_ffma); + false /* always_precise */); if (lower_flrp_progress) { NIR_PASS(progress, nir, nir_opt_constant_folding); @@ -542,25 +1293,22 @@ bi_optimize_nir(nir_shader *nir) } while (progress); NIR_PASS(progress, nir, nir_opt_algebraic_late); + NIR_PASS(progress, nir, nir_lower_bool_to_int32); + NIR_PASS(progress, nir, bifrost_nir_lower_algebraic_late); NIR_PASS(progress, nir, nir_lower_alu_to_scalar, NULL, NULL); NIR_PASS(progress, nir, nir_lower_load_const_to_scalar); /* Take us out of SSA */ NIR_PASS(progress, nir, nir_lower_locals_to_regs); - NIR_PASS(progress, nir, nir_convert_from_ssa, true); - - /* We're a primary scalar architecture but there's enough vector that - * we use a vector IR so let's not also deal with scalar hacks on top - * of the vector hacks */ - NIR_PASS(progress, nir, nir_move_vec_src_uses_to_dest); - NIR_PASS(progress, nir, nir_lower_vec_to_movs); - NIR_PASS(progress, nir, nir_opt_dce); + NIR_PASS(progress, nir, nir_convert_from_ssa, true); } void -bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned product_id) +bifrost_compile_shader_nir(nir_shader *nir, panfrost_program *program, unsigned product_id) { + bifrost_debug = debug_get_option_bifrost_debug(); + bi_context *ctx = rzalloc(NULL, bi_context); ctx->nir = nir; ctx->stage = nir->info.stage; @@ -582,11 +1330,21 @@ bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned p NIR_PASS_V(nir, nir_lower_global_vars_to_local); 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); + 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, nir_lower_mediump_outputs); bi_optimize_nir(nir); - nir_print_shader(nir, stdout); + + if (bifrost_debug & BIFROST_DBG_SHADERS) { + nir_print_shader(nir, stdout); + } + + 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->blend_types = program->blend_types; nir_foreach_function(func, nir) { if (!func->impl) @@ -597,8 +1355,39 @@ bifrost_compile_shader_nir(nir_shader *nir, bifrost_program *program, unsigned p break; /* TODO: Multi-function shaders */ } - bi_print_shader(ctx, stdout); + unsigned block_source_count = 0; + + bi_foreach_block(ctx, _block) { + bi_block *block = (bi_block *) _block; + + /* Name blocks now that we're done emitting so the order is + * consistent */ + block->base.name = block_source_count++; + + bi_lower_combine(ctx, block); + } + + bool progress = false; + + do { + progress = false; + + bi_foreach_block(ctx, _block) { + bi_block *block = (bi_block *) _block; + progress |= bi_opt_dead_code_eliminate(ctx, block); + } + } while(progress); + + if (bifrost_debug & BIFROST_DBG_SHADERS) + bi_print_shader(ctx, stdout); bi_schedule(ctx); + bi_register_allocate(ctx); + if (bifrost_debug & BIFROST_DBG_SHADERS) + bi_print_shader(ctx, stdout); + bi_pack(ctx, &program->compiled); + + if (bifrost_debug & BIFROST_DBG_SHADERS) + disassemble_bifrost(stdout, program->compiled.data, program->compiled.size, true); ralloc_free(ctx); }