/* Copyright © 2011 Intel Corporation * * Permission is hereby granted, free of charge, to any person obtaining a * copy of this software and associated documentation files (the "Software"), * to deal in the Software without restriction, including without limitation * the rights to use, copy, modify, merge, publish, distribute, sublicense, * and/or sell copies of the Software, and to permit persons to whom the * Software is furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice (including the next * paragraph) shall be included in all copies or substantial portions of the * Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS * IN THE SOFTWARE. */ #include "elk_vec4.h" #include "elk_cfg.h" #include "elk_eu.h" #include "elk_disasm_info.h" #include "dev/intel_debug.h" #include "util/mesa-sha1.h" using namespace elk; static void generate_math1_gfx4(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg src) { elk_gfx4_math(p, dst, elk_math_function(inst->opcode), inst->base_mrf, src, ELK_MATH_PRECISION_FULL); } static void check_gfx6_math_src_arg(struct elk_reg src) { /* Source swizzles are ignored. */ assert(!src.abs); assert(!src.negate); assert(src.swizzle == ELK_SWIZZLE_XYZW); } static void generate_math_gfx6(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg src0, struct elk_reg src1) { /* Can't do writemask because math can't be align16. */ assert(dst.writemask == WRITEMASK_XYZW); /* Source swizzles are ignored. */ check_gfx6_math_src_arg(src0); if (src1.file == ELK_GENERAL_REGISTER_FILE) check_gfx6_math_src_arg(src1); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_gfx6_math(p, dst, elk_math_function(inst->opcode), src0, src1); elk_set_default_access_mode(p, ELK_ALIGN_16); } static void generate_math2_gfx4(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg src0, struct elk_reg src1) { /* From the Ironlake PRM, Volume 4, Part 1, Section 6.1.13 * "Message Payload": * * "Operand0[7]. For the INT DIV functions, this operand is the * denominator." * ... * "Operand1[7]. For the INT DIV functions, this operand is the * numerator." */ bool is_int_div = inst->opcode != ELK_SHADER_OPCODE_POW; struct elk_reg &op0 = is_int_div ? src1 : src0; struct elk_reg &op1 = is_int_div ? src0 : src1; elk_push_insn_state(p); elk_set_default_saturate(p, false); elk_set_default_predicate_control(p, ELK_PREDICATE_NONE); elk_set_default_flag_reg(p, 0, 0); elk_MOV(p, retype(elk_message_reg(inst->base_mrf + 1), op1.type), op1); elk_pop_insn_state(p); elk_gfx4_math(p, dst, elk_math_function(inst->opcode), inst->base_mrf, op0, ELK_MATH_PRECISION_FULL); } static void generate_tex(struct elk_codegen *p, struct elk_vue_prog_data *prog_data, gl_shader_stage stage, vec4_instruction *inst, struct elk_reg dst, struct elk_reg src, struct elk_reg surface_index, struct elk_reg sampler_index) { const struct intel_device_info *devinfo = p->devinfo; int msg_type = -1; if (devinfo->ver >= 5) { switch (inst->opcode) { case ELK_SHADER_OPCODE_TEX: case ELK_SHADER_OPCODE_TXL: if (inst->shadow_compare) { msg_type = GFX5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE; } else { msg_type = GFX5_SAMPLER_MESSAGE_SAMPLE_LOD; } break; case ELK_SHADER_OPCODE_TXD: if (inst->shadow_compare) { /* Gfx7.5+. Otherwise, lowered by elk_lower_texture_gradients(). */ assert(devinfo->verx10 == 75); msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE; } else { msg_type = GFX5_SAMPLER_MESSAGE_SAMPLE_DERIVS; } break; case ELK_SHADER_OPCODE_TXF: msg_type = GFX5_SAMPLER_MESSAGE_SAMPLE_LD; break; case ELK_SHADER_OPCODE_TXF_CMS: if (devinfo->ver >= 7) msg_type = GFX7_SAMPLER_MESSAGE_SAMPLE_LD2DMS; else msg_type = GFX5_SAMPLER_MESSAGE_SAMPLE_LD; break; case ELK_SHADER_OPCODE_TXF_MCS: assert(devinfo->ver >= 7); msg_type = GFX7_SAMPLER_MESSAGE_SAMPLE_LD_MCS; break; case ELK_SHADER_OPCODE_TXS: msg_type = GFX5_SAMPLER_MESSAGE_SAMPLE_RESINFO; break; case ELK_SHADER_OPCODE_TG4: if (inst->shadow_compare) { msg_type = GFX7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C; } else { msg_type = GFX7_SAMPLER_MESSAGE_SAMPLE_GATHER4; } break; case ELK_SHADER_OPCODE_TG4_OFFSET: if (inst->shadow_compare) { msg_type = GFX7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C; } else { msg_type = GFX7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO; } break; case ELK_SHADER_OPCODE_SAMPLEINFO: msg_type = GFX6_SAMPLER_MESSAGE_SAMPLE_SAMPLEINFO; break; default: unreachable("should not get here: invalid vec4 texture opcode"); } } else { switch (inst->opcode) { case ELK_SHADER_OPCODE_TEX: case ELK_SHADER_OPCODE_TXL: if (inst->shadow_compare) { msg_type = ELK_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD_COMPARE; assert(inst->mlen == 3); } else { msg_type = ELK_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD; assert(inst->mlen == 2); } break; case ELK_SHADER_OPCODE_TXD: /* There is no sample_d_c message; comparisons are done manually. */ msg_type = ELK_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_GRADIENTS; assert(inst->mlen == 4); break; case ELK_SHADER_OPCODE_TXF: msg_type = ELK_SAMPLER_MESSAGE_SIMD4X2_LD; assert(inst->mlen == 2); break; case ELK_SHADER_OPCODE_TXS: msg_type = ELK_SAMPLER_MESSAGE_SIMD4X2_RESINFO; assert(inst->mlen == 2); break; default: unreachable("should not get here: invalid vec4 texture opcode"); } } assert(msg_type != -1); assert(sampler_index.type == ELK_REGISTER_TYPE_UD); /* Load the message header if present. If there's a texture offset, we need * to set it up explicitly and load the offset bitfield. Otherwise, we can * use an implied move from g0 to the first message register. */ if (inst->header_size != 0) { if (devinfo->ver < 6 && !inst->offset) { /* Set up an implied move from g0 to the MRF. */ src = elk_vec8_grf(0, 0); } else { struct elk_reg header = retype(elk_message_reg(inst->base_mrf), ELK_REGISTER_TYPE_UD); uint32_t dw2 = 0; /* Explicitly set up the message header by copying g0 to the MRF. */ elk_push_insn_state(p); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, header, retype(elk_vec8_grf(0, 0), ELK_REGISTER_TYPE_UD)); elk_set_default_access_mode(p, ELK_ALIGN_1); if (inst->offset) /* Set the texel offset bits in DWord 2. */ dw2 = inst->offset; /* The VS, DS, and FS stages have the g0.2 payload delivered as 0, * so header0.2 is 0 when g0 is copied. The HS and GS stages do * not, so we must set to to 0 to avoid setting undesirable bits * in the message header. */ if (dw2 || stage == MESA_SHADER_TESS_CTRL || stage == MESA_SHADER_GEOMETRY) { elk_MOV(p, get_element_ud(header, 2), elk_imm_ud(dw2)); } elk_adjust_sampler_state_pointer(p, header, sampler_index); elk_pop_insn_state(p); } } uint32_t return_format; switch (dst.type) { case ELK_REGISTER_TYPE_D: return_format = ELK_SAMPLER_RETURN_FORMAT_SINT32; break; case ELK_REGISTER_TYPE_UD: return_format = ELK_SAMPLER_RETURN_FORMAT_UINT32; break; default: return_format = ELK_SAMPLER_RETURN_FORMAT_FLOAT32; break; } /* Stomp the resinfo output type to UINT32. On gens 4-5, the output type * is set as part of the message descriptor. On gfx4, the PRM seems to * allow UINT32 and FLOAT32 (i965 PRM, Vol. 4 Section 4.8.1.1), but on * later gens UINT32 is required. Once you hit Sandy Bridge, the bit is * gone from the message descriptor entirely and you just get UINT32 all * the time regasrdless. Since we can really only do non-UINT32 on gfx4, * just stomp it to UINT32 all the time. */ if (inst->opcode == ELK_SHADER_OPCODE_TXS) return_format = ELK_SAMPLER_RETURN_FORMAT_UINT32; if (surface_index.file == ELK_IMMEDIATE_VALUE && sampler_index.file == ELK_IMMEDIATE_VALUE) { uint32_t surface = surface_index.ud; uint32_t sampler = sampler_index.ud; elk_SAMPLE(p, dst, inst->base_mrf, src, surface, sampler % 16, msg_type, 1, /* response length */ inst->mlen, inst->header_size != 0, ELK_SAMPLER_SIMD_MODE_SIMD4X2, return_format); } else { /* Non-constant sampler index. */ struct elk_reg addr = vec1(retype(elk_address_reg(0), ELK_REGISTER_TYPE_UD)); struct elk_reg surface_reg = vec1(retype(surface_index, ELK_REGISTER_TYPE_UD)); struct elk_reg sampler_reg = vec1(retype(sampler_index, ELK_REGISTER_TYPE_UD)); elk_push_insn_state(p); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_set_default_access_mode(p, ELK_ALIGN_1); if (elk_regs_equal(&surface_reg, &sampler_reg)) { elk_MUL(p, addr, sampler_reg, elk_imm_uw(0x101)); } else { if (sampler_reg.file == ELK_IMMEDIATE_VALUE) { elk_OR(p, addr, surface_reg, elk_imm_ud(sampler_reg.ud << 8)); } else { elk_SHL(p, addr, sampler_reg, elk_imm_ud(8)); elk_OR(p, addr, addr, surface_reg); } } elk_AND(p, addr, addr, elk_imm_ud(0xfff)); elk_pop_insn_state(p); if (inst->base_mrf != -1) elk_gfx6_resolve_implied_move(p, &src, inst->base_mrf); /* dst = send(offset, a0.0 | ) */ elk_send_indirect_message( p, ELK_SFID_SAMPLER, dst, src, addr, elk_message_desc(devinfo, inst->mlen, 1, inst->header_size) | elk_sampler_desc(devinfo, 0 /* surface */, 0 /* sampler */, msg_type, ELK_SAMPLER_SIMD_MODE_SIMD4X2, return_format), false /* EOT */); /* visitor knows more than we do about the surface limit required, * so has already done marking. */ } } static void generate_vs_urb_write(struct elk_codegen *p, vec4_instruction *inst) { elk_urb_WRITE(p, elk_null_reg(), /* dest */ inst->base_mrf, /* starting mrf reg nr */ elk_vec8_grf(0, 0), /* src */ inst->urb_write_flags, inst->mlen, 0, /* response len */ inst->offset, /* urb destination offset */ ELK_URB_SWIZZLE_INTERLEAVE); } static void generate_gs_urb_write(struct elk_codegen *p, vec4_instruction *inst) { struct elk_reg src = elk_message_reg(inst->base_mrf); elk_urb_WRITE(p, elk_null_reg(), /* dest */ inst->base_mrf, /* starting mrf reg nr */ src, inst->urb_write_flags, inst->mlen, 0, /* response len */ inst->offset, /* urb destination offset */ ELK_URB_SWIZZLE_INTERLEAVE); } static void generate_gs_urb_write_allocate(struct elk_codegen *p, vec4_instruction *inst) { struct elk_reg src = elk_message_reg(inst->base_mrf); /* We pass the temporary passed in src0 as the writeback register */ elk_urb_WRITE(p, inst->src[0].as_elk_reg(), /* dest */ inst->base_mrf, /* starting mrf reg nr */ src, ELK_URB_WRITE_ALLOCATE_COMPLETE, inst->mlen, 1, /* response len */ inst->offset, /* urb destination offset */ ELK_URB_SWIZZLE_INTERLEAVE); /* Now put allocated urb handle in dst.0 */ elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, get_element_ud(inst->dst.as_elk_reg(), 0), get_element_ud(inst->src[0].as_elk_reg(), 0)); elk_pop_insn_state(p); } static void generate_gs_thread_end(struct elk_codegen *p, vec4_instruction *inst) { struct elk_reg src = elk_message_reg(inst->base_mrf); elk_urb_WRITE(p, elk_null_reg(), /* dest */ inst->base_mrf, /* starting mrf reg nr */ src, ELK_URB_WRITE_EOT | inst->urb_write_flags, inst->mlen, 0, /* response len */ 0, /* urb destination offset */ ELK_URB_SWIZZLE_INTERLEAVE); } static void generate_gs_set_write_offset(struct elk_codegen *p, struct elk_reg dst, struct elk_reg src0, struct elk_reg src1) { /* From p22 of volume 4 part 2 of the Ivy Bridge PRM (2.4.3.1 Message * Header: M0.3): * * Slot 0 Offset. This field, after adding to the Global Offset field * in the message descriptor, specifies the offset (in 256-bit units) * from the start of the URB entry, as referenced by URB Handle 0, at * which the data will be accessed. * * Similar text describes DWORD M0.4, which is slot 1 offset. * * Therefore, we want to multiply DWORDs 0 and 4 of src0 (the x components * of the register for geometry shader invocations 0 and 1) by the * immediate value in src1, and store the result in DWORDs 3 and 4 of dst. * * We can do this with the following EU instruction: * * mul(2) dst.3<1>UD src0<8;2,4>UD src1<...>UW { Align1 WE_all } */ elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); assert(p->devinfo->ver >= 7 && src1.file == ELK_IMMEDIATE_VALUE && src1.type == ELK_REGISTER_TYPE_UD && src1.ud <= USHRT_MAX); if (src0.file == ELK_IMMEDIATE_VALUE) { elk_MOV(p, suboffset(stride(dst, 2, 2, 1), 3), elk_imm_ud(src0.ud * src1.ud)); } else { if (src1.file == ELK_IMMEDIATE_VALUE) { src1 = elk_imm_uw(src1.ud); } elk_MUL(p, suboffset(stride(dst, 2, 2, 1), 3), stride(src0, 8, 2, 4), retype(src1, ELK_REGISTER_TYPE_UW)); } elk_pop_insn_state(p); } static void generate_gs_set_vertex_count(struct elk_codegen *p, struct elk_reg dst, struct elk_reg src) { elk_push_insn_state(p); elk_set_default_mask_control(p, ELK_MASK_DISABLE); /* If we think of the src and dst registers as composed of 8 DWORDs each, * we want to pick up the contents of DWORDs 0 and 4 from src, truncate * them to WORDs, and then pack them into DWORD 2 of dst. * * It's easier to get the EU to do this if we think of the src and dst * registers as composed of 16 WORDS each; then, we want to pick up the * contents of WORDs 0 and 8 from src, and pack them into WORDs 4 and 5 * of dst. * * We can do that by the following EU instruction: * * mov (2) dst.4<1>:uw src<8;1,0>:uw { Align1, Q1, NoMask } */ elk_set_default_access_mode(p, ELK_ALIGN_1); elk_MOV(p, suboffset(stride(retype(dst, ELK_REGISTER_TYPE_UW), 2, 2, 1), 4), stride(retype(src, ELK_REGISTER_TYPE_UW), 8, 1, 0)); elk_pop_insn_state(p); } static void generate_gs_svb_write(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg src0, struct elk_reg src1) { int binding = inst->sol_binding; bool final_write = inst->sol_final_write; elk_push_insn_state(p); elk_set_default_exec_size(p, ELK_EXECUTE_4); /* Copy Vertex data into M0.x */ elk_MOV(p, stride(dst, 4, 4, 1), stride(retype(src0, ELK_REGISTER_TYPE_UD), 4, 4, 1)); elk_pop_insn_state(p); elk_push_insn_state(p); /* Send SVB Write */ elk_svb_write(p, final_write ? src1 : elk_null_reg(), /* dest == src1 */ 1, /* msg_reg_nr */ dst, /* src0 == previous dst */ ELK_GFX6_SOL_BINDING_START + binding, /* binding_table_index */ final_write); /* send_commit_msg */ /* Finally, wait for the write commit to occur so that we can proceed to * other things safely. * * From the Sandybridge PRM, Volume 4, Part 1, Section 3.3: * * The write commit does not modify the destination register, but * merely clears the dependency associated with the destination * register. Thus, a simple “mov” instruction using the register as a * source is sufficient to wait for the write commit to occur. */ if (final_write) { elk_MOV(p, src1, src1); } elk_pop_insn_state(p); } static void generate_gs_svb_set_destination_index(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg src) { int vertex = inst->sol_vertex; elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, get_element_ud(dst, 5), get_element_ud(src, vertex)); elk_pop_insn_state(p); } static void generate_gs_set_dword_2(struct elk_codegen *p, struct elk_reg dst, struct elk_reg src) { elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, suboffset(vec1(dst), 2), suboffset(vec1(src), 0)); elk_pop_insn_state(p); } static void generate_gs_prepare_channel_masks(struct elk_codegen *p, struct elk_reg dst) { /* We want to left shift just DWORD 4 (the x component belonging to the * second geometry shader invocation) by 4 bits. So generate the * instruction: * * shl(1) dst.4<1>UD dst.4<0,1,0>UD 4UD { align1 WE_all } */ dst = suboffset(vec1(dst), 4); elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_SHL(p, dst, dst, elk_imm_ud(4)); elk_pop_insn_state(p); } static void generate_gs_set_channel_masks(struct elk_codegen *p, struct elk_reg dst, struct elk_reg src) { /* From p21 of volume 4 part 2 of the Ivy Bridge PRM (2.4.3.1 Message * Header: M0.5): * * 15 Vertex 1 DATA [3] / Vertex 0 DATA[7] Channel Mask * * When Swizzle Control = URB_INTERLEAVED this bit controls Vertex 1 * DATA[3], when Swizzle Control = URB_NOSWIZZLE this bit controls * Vertex 0 DATA[7]. This bit is ANDed with the corresponding * channel enable to determine the final channel enable. For the * URB_READ_OWORD & URB_READ_HWORD messages, when final channel * enable is 1 it indicates that Vertex 1 DATA [3] will be included * in the writeback message. For the URB_WRITE_OWORD & * URB_WRITE_HWORD messages, when final channel enable is 1 it * indicates that Vertex 1 DATA [3] will be written to the surface. * * 0: Vertex 1 DATA [3] / Vertex 0 DATA[7] channel not included * 1: Vertex DATA [3] / Vertex 0 DATA[7] channel included * * 14 Vertex 1 DATA [2] Channel Mask * 13 Vertex 1 DATA [1] Channel Mask * 12 Vertex 1 DATA [0] Channel Mask * 11 Vertex 0 DATA [3] Channel Mask * 10 Vertex 0 DATA [2] Channel Mask * 9 Vertex 0 DATA [1] Channel Mask * 8 Vertex 0 DATA [0] Channel Mask * * (This is from a section of the PRM that is agnostic to the particular * type of shader being executed, so "Vertex 0" and "Vertex 1" refer to * geometry shader invocations 0 and 1, respectively). Since we have the * enable flags for geometry shader invocation 0 in bits 3:0 of DWORD 0, * and the enable flags for geometry shader invocation 1 in bits 7:0 of * DWORD 4, we just need to OR them together and store the result in bits * 15:8 of DWORD 5. * * It's easier to get the EU to do this if we think of the src and dst * registers as composed of 32 bytes each; then, we want to pick up the * contents of bytes 0 and 16 from src, OR them together, and store them in * byte 21. * * We can do that by the following EU instruction: * * or(1) dst.21<1>UB src<0,1,0>UB src.16<0,1,0>UB { align1 WE_all } * * Note: this relies on the source register having zeros in (a) bits 7:4 of * DWORD 0 and (b) bits 3:0 of DWORD 4. We can rely on (b) because the * source register was prepared by ELK_GS_OPCODE_PREPARE_CHANNEL_MASKS (which * shifts DWORD 4 left by 4 bits), and we can rely on (a) because prior to * the execution of ELK_GS_OPCODE_PREPARE_CHANNEL_MASKS, DWORDs 0 and 4 need to * contain valid channel mask values (which are in the range 0x0-0xf). */ dst = retype(dst, ELK_REGISTER_TYPE_UB); src = retype(src, ELK_REGISTER_TYPE_UB); elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_OR(p, suboffset(vec1(dst), 21), vec1(src), suboffset(vec1(src), 16)); elk_pop_insn_state(p); } static void generate_gs_get_instance_id(struct elk_codegen *p, struct elk_reg dst) { /* We want to right shift R0.0 & R0.1 by GFX7_GS_PAYLOAD_INSTANCE_ID_SHIFT * and store into dst.0 & dst.4. So generate the instruction: * * shr(8) dst<1> R0<1,4,0> GFX7_GS_PAYLOAD_INSTANCE_ID_SHIFT { align1 WE_normal 1Q } */ elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); dst = retype(dst, ELK_REGISTER_TYPE_UD); struct elk_reg r0(retype(elk_vec8_grf(0, 0), ELK_REGISTER_TYPE_UD)); elk_SHR(p, dst, stride(r0, 1, 4, 0), elk_imm_ud(GFX7_GS_PAYLOAD_INSTANCE_ID_SHIFT)); elk_pop_insn_state(p); } static void generate_gs_ff_sync_set_primitives(struct elk_codegen *p, struct elk_reg dst, struct elk_reg src0, struct elk_reg src1, struct elk_reg src2) { elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); /* Save src0 data in 16:31 bits of dst.0 */ elk_AND(p, suboffset(vec1(dst), 0), suboffset(vec1(src0), 0), elk_imm_ud(0xffffu)); elk_SHL(p, suboffset(vec1(dst), 0), suboffset(vec1(dst), 0), elk_imm_ud(16)); /* Save src1 data in 0:15 bits of dst.0 */ elk_AND(p, suboffset(vec1(src2), 0), suboffset(vec1(src1), 0), elk_imm_ud(0xffffu)); elk_OR(p, suboffset(vec1(dst), 0), suboffset(vec1(dst), 0), suboffset(vec1(src2), 0)); elk_pop_insn_state(p); } static void generate_gs_ff_sync(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg src0, struct elk_reg src1) { /* This opcode uses an implied MRF register for: * - the header of the ff_sync message. And as such it is expected to be * initialized to r0 before calling here. * - the destination where we will write the allocated URB handle. */ struct elk_reg header = retype(elk_message_reg(inst->base_mrf), ELK_REGISTER_TYPE_UD); /* Overwrite dword 0 of the header (SO vertices to write) and * dword 1 (number of primitives written). */ elk_push_insn_state(p); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_MOV(p, get_element_ud(header, 0), get_element_ud(src1, 0)); elk_MOV(p, get_element_ud(header, 1), get_element_ud(src0, 0)); elk_pop_insn_state(p); /* Allocate URB handle in dst */ elk_ff_sync(p, dst, 0, header, 1, /* allocate */ 1, /* response length */ 0 /* eot */); /* Now put allocated urb handle in header.0 */ elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, get_element_ud(header, 0), get_element_ud(dst, 0)); /* src1 is not an immediate when we use transform feedback */ if (src1.file != ELK_IMMEDIATE_VALUE) { elk_set_default_exec_size(p, ELK_EXECUTE_4); elk_MOV(p, elk_vec4_grf(src1.nr, 0), elk_vec4_grf(dst.nr, 1)); } elk_pop_insn_state(p); } static void generate_gs_set_primitive_id(struct elk_codegen *p, struct elk_reg dst) { /* In gfx6, PrimitiveID is delivered in R0.1 of the payload */ struct elk_reg src = elk_vec8_grf(0, 0); elk_push_insn_state(p); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_MOV(p, get_element_ud(dst, 0), get_element_ud(src, 1)); elk_pop_insn_state(p); } static void generate_tcs_get_instance_id(struct elk_codegen *p, struct elk_reg dst) { const struct intel_device_info *devinfo = p->devinfo; const bool ivb = devinfo->platform == INTEL_PLATFORM_IVB || devinfo->platform == INTEL_PLATFORM_BYT; /* "Instance Count" comes as part of the payload in r0.2 bits 23:17. * * Since we operate in SIMD4x2 mode, we need run half as many threads * as necessary. So we assign (2i + 1, 2i) as the thread counts. We * shift right by one less to accomplish the multiplication by two. */ dst = retype(dst, ELK_REGISTER_TYPE_UD); struct elk_reg r0(retype(elk_vec8_grf(0, 0), ELK_REGISTER_TYPE_UD)); elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); const int mask = ivb ? INTEL_MASK(22, 16) : INTEL_MASK(23, 17); const int shift = ivb ? 16 : 17; elk_AND(p, get_element_ud(dst, 0), get_element_ud(r0, 2), elk_imm_ud(mask)); elk_SHR(p, get_element_ud(dst, 0), get_element_ud(dst, 0), elk_imm_ud(shift - 1)); elk_ADD(p, get_element_ud(dst, 4), get_element_ud(dst, 0), elk_imm_ud(1)); elk_pop_insn_state(p); } static void generate_tcs_urb_write(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg urb_header) { const struct intel_device_info *devinfo = p->devinfo; elk_inst *send = elk_next_insn(p, ELK_OPCODE_SEND); elk_set_dest(p, send, elk_null_reg()); elk_set_src0(p, send, urb_header); elk_set_desc(p, send, elk_message_desc(devinfo, inst->mlen, 0, true)); elk_inst_set_sfid(devinfo, send, ELK_SFID_URB); elk_inst_set_urb_opcode(devinfo, send, ELK_URB_OPCODE_WRITE_OWORD); elk_inst_set_urb_global_offset(devinfo, send, inst->offset); if (inst->urb_write_flags & ELK_URB_WRITE_EOT) { elk_inst_set_eot(devinfo, send, 1); } else { elk_inst_set_urb_per_slot_offset(devinfo, send, 1); elk_inst_set_urb_swizzle_control(devinfo, send, ELK_URB_SWIZZLE_INTERLEAVE); } /* what happens to swizzles? */ } static void generate_tcs_input_urb_offsets(struct elk_codegen *p, struct elk_reg dst, struct elk_reg vertex, struct elk_reg offset) { /* Generates an URB read/write message header for HS/DS operation. * Inputs are a vertex index, and a byte offset from the beginning of * the vertex. */ /* If `vertex` is not an immediate, we clobber a0.0 */ assert(vertex.file == ELK_IMMEDIATE_VALUE || vertex.file == ELK_GENERAL_REGISTER_FILE); assert(vertex.type == ELK_REGISTER_TYPE_UD || vertex.type == ELK_REGISTER_TYPE_D); assert(dst.file == ELK_GENERAL_REGISTER_FILE); elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, dst, elk_imm_ud(0)); /* m0.5 bits 8-15 are channel enables */ elk_MOV(p, get_element_ud(dst, 5), elk_imm_ud(0xff00)); /* m0.0-0.1: URB handles */ if (vertex.file == ELK_IMMEDIATE_VALUE) { uint32_t vertex_index = vertex.ud; struct elk_reg index_reg = elk_vec1_grf( 1 + (vertex_index >> 3), vertex_index & 7); elk_MOV(p, vec2(get_element_ud(dst, 0)), retype(index_reg, ELK_REGISTER_TYPE_UD)); } else { /* Use indirect addressing. ICP Handles are DWords (single channels * of a register) and start at g1.0. * * In order to start our region at g1.0, we add 8 to the vertex index, * effectively skipping over the 8 channels in g0.0. This gives us a * DWord offset to the ICP Handle. * * Indirect addressing works in terms of bytes, so we then multiply * the DWord offset by 4 (by shifting left by 2). */ struct elk_reg addr = elk_address_reg(0); /* bottom half: m0.0 = g[1.0 + vertex.0]UD */ elk_ADD(p, addr, retype(get_element_ud(vertex, 0), ELK_REGISTER_TYPE_UW), elk_imm_uw(0x8)); elk_SHL(p, addr, addr, elk_imm_uw(2)); elk_MOV(p, get_element_ud(dst, 0), deref_1ud(elk_indirect(0, 0), 0)); /* top half: m0.1 = g[1.0 + vertex.4]UD */ elk_ADD(p, addr, retype(get_element_ud(vertex, 4), ELK_REGISTER_TYPE_UW), elk_imm_uw(0x8)); elk_SHL(p, addr, addr, elk_imm_uw(2)); elk_MOV(p, get_element_ud(dst, 1), deref_1ud(elk_indirect(0, 0), 0)); } /* m0.3-0.4: 128bit-granular offsets into the URB from the handles */ if (offset.file != ARF) elk_MOV(p, vec2(get_element_ud(dst, 3)), stride(offset, 4, 1, 0)); elk_pop_insn_state(p); } static void generate_tcs_output_urb_offsets(struct elk_codegen *p, struct elk_reg dst, struct elk_reg write_mask, struct elk_reg offset) { /* Generates an URB read/write message header for HS/DS operation, for the patch URB entry. */ assert(dst.file == ELK_GENERAL_REGISTER_FILE || dst.file == ELK_MESSAGE_REGISTER_FILE); assert(write_mask.file == ELK_IMMEDIATE_VALUE); assert(write_mask.type == ELK_REGISTER_TYPE_UD); elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, dst, elk_imm_ud(0)); unsigned mask = write_mask.ud; /* m0.5 bits 15:12 and 11:8 are channel enables */ elk_MOV(p, get_element_ud(dst, 5), elk_imm_ud((mask << 8) | (mask << 12))); /* HS patch URB handle is delivered in r0.0 */ struct elk_reg urb_handle = elk_vec1_grf(0, 0); /* m0.0-0.1: URB handles */ elk_MOV(p, vec2(get_element_ud(dst, 0)), retype(urb_handle, ELK_REGISTER_TYPE_UD)); /* m0.3-0.4: 128bit-granular offsets into the URB from the handles */ if (offset.file != ARF) elk_MOV(p, vec2(get_element_ud(dst, 3)), stride(offset, 4, 1, 0)); elk_pop_insn_state(p); } static void generate_tes_create_input_read_header(struct elk_codegen *p, struct elk_reg dst) { elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); /* Initialize the register to 0 */ elk_MOV(p, dst, elk_imm_ud(0)); /* Enable all the channels in m0.5 bits 15:8 */ elk_MOV(p, get_element_ud(dst, 5), elk_imm_ud(0xff00)); /* Copy g1.3 (the patch URB handle) to m0.0 and m0.1. For safety, * mask out irrelevant "Reserved" bits, as they're not marked MBZ. */ elk_AND(p, vec2(get_element_ud(dst, 0)), retype(elk_vec1_grf(1, 3), ELK_REGISTER_TYPE_UD), elk_imm_ud(0x1fff)); elk_pop_insn_state(p); } static void generate_tes_add_indirect_urb_offset(struct elk_codegen *p, struct elk_reg dst, struct elk_reg header, struct elk_reg offset) { elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, dst, header); /* Uniforms will have a stride <0;4,1>, and we need to convert to <0;1,0>. * Other values get <4;1,0>. */ struct elk_reg restrided_offset; if (offset.vstride == ELK_VERTICAL_STRIDE_0 && offset.width == ELK_WIDTH_4 && offset.hstride == ELK_HORIZONTAL_STRIDE_1) { restrided_offset = stride(offset, 0, 1, 0); } else { restrided_offset = stride(offset, 4, 1, 0); } /* m0.3-0.4: 128-bit-granular offsets into the URB from the handles */ elk_MOV(p, vec2(get_element_ud(dst, 3)), restrided_offset); elk_pop_insn_state(p); } static void generate_vec4_urb_read(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg header) { const struct intel_device_info *devinfo = p->devinfo; assert(header.file == ELK_GENERAL_REGISTER_FILE); assert(header.type == ELK_REGISTER_TYPE_UD); elk_inst *send = elk_next_insn(p, ELK_OPCODE_SEND); elk_set_dest(p, send, dst); elk_set_src0(p, send, header); elk_set_desc(p, send, elk_message_desc(devinfo, 1, 1, true)); elk_inst_set_sfid(devinfo, send, ELK_SFID_URB); elk_inst_set_urb_opcode(devinfo, send, ELK_URB_OPCODE_READ_OWORD); elk_inst_set_urb_swizzle_control(devinfo, send, ELK_URB_SWIZZLE_INTERLEAVE); elk_inst_set_urb_per_slot_offset(devinfo, send, 1); elk_inst_set_urb_global_offset(devinfo, send, inst->offset); } static void generate_tcs_release_input(struct elk_codegen *p, struct elk_reg header, struct elk_reg vertex, struct elk_reg is_unpaired) { const struct intel_device_info *devinfo = p->devinfo; assert(vertex.file == ELK_IMMEDIATE_VALUE); assert(vertex.type == ELK_REGISTER_TYPE_UD); /* m0.0-0.1: URB handles */ struct elk_reg urb_handles = retype(elk_vec2_grf(1 + (vertex.ud >> 3), vertex.ud & 7), ELK_REGISTER_TYPE_UD); elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, header, elk_imm_ud(0)); elk_MOV(p, vec2(get_element_ud(header, 0)), urb_handles); elk_pop_insn_state(p); elk_inst *send = elk_next_insn(p, ELK_OPCODE_SEND); elk_set_dest(p, send, elk_null_reg()); elk_set_src0(p, send, header); elk_set_desc(p, send, elk_message_desc(devinfo, 1, 0, true)); elk_inst_set_sfid(devinfo, send, ELK_SFID_URB); elk_inst_set_urb_opcode(devinfo, send, ELK_URB_OPCODE_READ_OWORD); elk_inst_set_urb_complete(devinfo, send, 1); elk_inst_set_urb_swizzle_control(devinfo, send, is_unpaired.ud ? ELK_URB_SWIZZLE_NONE : ELK_URB_SWIZZLE_INTERLEAVE); } static void generate_tcs_thread_end(struct elk_codegen *p, vec4_instruction *inst) { struct elk_reg header = elk_message_reg(inst->base_mrf); elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_MOV(p, header, elk_imm_ud(0)); elk_MOV(p, get_element_ud(header, 5), elk_imm_ud(WRITEMASK_X << 8)); elk_MOV(p, get_element_ud(header, 0), retype(elk_vec1_grf(0, 0), ELK_REGISTER_TYPE_UD)); elk_MOV(p, elk_message_reg(inst->base_mrf + 1), elk_imm_ud(0u)); elk_pop_insn_state(p); elk_urb_WRITE(p, elk_null_reg(), /* dest */ inst->base_mrf, /* starting mrf reg nr */ header, ELK_URB_WRITE_EOT | ELK_URB_WRITE_OWORD | ELK_URB_WRITE_USE_CHANNEL_MASKS, inst->mlen, 0, /* response len */ 0, /* urb destination offset */ 0); } static void generate_tes_get_primitive_id(struct elk_codegen *p, struct elk_reg dst) { elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_MOV(p, dst, retype(elk_vec1_grf(1, 7), ELK_REGISTER_TYPE_D)); elk_pop_insn_state(p); } static void generate_tcs_get_primitive_id(struct elk_codegen *p, struct elk_reg dst) { elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_MOV(p, dst, retype(elk_vec1_grf(0, 1), ELK_REGISTER_TYPE_UD)); elk_pop_insn_state(p); } static void generate_tcs_create_barrier_header(struct elk_codegen *p, struct elk_vue_prog_data *prog_data, struct elk_reg dst) { const struct intel_device_info *devinfo = p->devinfo; const bool ivb = devinfo->platform == INTEL_PLATFORM_IVB || devinfo->platform == INTEL_PLATFORM_BYT; struct elk_reg m0_2 = get_element_ud(dst, 2); unsigned instances = ((struct elk_tcs_prog_data *) prog_data)->instances; elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); /* Zero the message header */ elk_MOV(p, retype(dst, ELK_REGISTER_TYPE_UD), elk_imm_ud(0u)); /* Copy "Barrier ID" from r0.2, bits 16:13 (Gfx7.5+) or 15:12 (Gfx7) */ elk_AND(p, m0_2, retype(elk_vec1_grf(0, 2), ELK_REGISTER_TYPE_UD), elk_imm_ud(ivb ? INTEL_MASK(15, 12) : INTEL_MASK(16, 13))); /* Shift it up to bits 27:24. */ elk_SHL(p, m0_2, get_element_ud(dst, 2), elk_imm_ud(ivb ? 12 : 11)); /* Set the Barrier Count and the enable bit */ elk_OR(p, m0_2, m0_2, elk_imm_ud(instances << 9 | (1 << 15))); elk_pop_insn_state(p); } static void generate_oword_dual_block_offsets(struct elk_codegen *p, struct elk_reg m1, struct elk_reg index) { int second_vertex_offset; if (p->devinfo->ver >= 6) second_vertex_offset = 1; else second_vertex_offset = 16; m1 = retype(m1, ELK_REGISTER_TYPE_D); /* Set up M1 (message payload). Only the block offsets in M1.0 and * M1.4 are used, and the rest are ignored. */ struct elk_reg m1_0 = suboffset(vec1(m1), 0); struct elk_reg m1_4 = suboffset(vec1(m1), 4); struct elk_reg index_0 = suboffset(vec1(index), 0); struct elk_reg index_4 = suboffset(vec1(index), 4); elk_push_insn_state(p); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_MOV(p, m1_0, index_0); if (index.file == ELK_IMMEDIATE_VALUE) { index_4.ud += second_vertex_offset; elk_MOV(p, m1_4, index_4); } else { elk_ADD(p, m1_4, index_4, elk_imm_d(second_vertex_offset)); } elk_pop_insn_state(p); } static void generate_unpack_flags(struct elk_codegen *p, struct elk_reg dst) { elk_push_insn_state(p); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_set_default_access_mode(p, ELK_ALIGN_1); struct elk_reg flags = elk_flag_reg(0, 0); struct elk_reg dst_0 = suboffset(vec1(dst), 0); struct elk_reg dst_4 = suboffset(vec1(dst), 4); elk_AND(p, dst_0, flags, elk_imm_ud(0x0f)); elk_AND(p, dst_4, flags, elk_imm_ud(0xf0)); elk_SHR(p, dst_4, dst_4, elk_imm_ud(4)); elk_pop_insn_state(p); } static void generate_scratch_read(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg index) { const struct intel_device_info *devinfo = p->devinfo; struct elk_reg header = elk_vec8_grf(0, 0); elk_gfx6_resolve_implied_move(p, &header, inst->base_mrf); generate_oword_dual_block_offsets(p, elk_message_reg(inst->base_mrf + 1), index); uint32_t msg_type; if (devinfo->ver >= 6) msg_type = GFX6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ; else if (devinfo->verx10 >= 45) msg_type = G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ; else msg_type = ELK_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ; const unsigned target_cache = devinfo->ver >= 7 ? GFX7_SFID_DATAPORT_DATA_CACHE : devinfo->ver >= 6 ? GFX6_SFID_DATAPORT_RENDER_CACHE : ELK_SFID_DATAPORT_READ; /* Each of the 8 channel enables is considered for whether each * dword is written. */ elk_inst *send = elk_next_insn(p, ELK_OPCODE_SEND); elk_inst_set_sfid(devinfo, send, target_cache); elk_set_dest(p, send, dst); elk_set_src0(p, send, header); if (devinfo->ver < 6) elk_inst_set_cond_modifier(devinfo, send, inst->base_mrf); elk_set_desc(p, send, elk_message_desc(devinfo, 2, 1, true) | elk_dp_read_desc(devinfo, elk_scratch_surface_idx(p), ELK_DATAPORT_OWORD_DUAL_BLOCK_1OWORD, msg_type, ELK_DATAPORT_READ_TARGET_RENDER_CACHE)); } static void generate_scratch_write(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg src, struct elk_reg index) { const struct intel_device_info *devinfo = p->devinfo; const unsigned target_cache = (devinfo->ver >= 7 ? GFX7_SFID_DATAPORT_DATA_CACHE : devinfo->ver >= 6 ? GFX6_SFID_DATAPORT_RENDER_CACHE : ELK_SFID_DATAPORT_WRITE); struct elk_reg header = elk_vec8_grf(0, 0); bool write_commit; /* If the instruction is predicated, we'll predicate the send, not * the header setup. */ elk_push_insn_state(p); elk_set_default_predicate_control(p, ELK_PREDICATE_NONE); elk_set_default_flag_reg(p, 0, 0); elk_gfx6_resolve_implied_move(p, &header, inst->base_mrf); generate_oword_dual_block_offsets(p, elk_message_reg(inst->base_mrf + 1), index); elk_MOV(p, retype(elk_message_reg(inst->base_mrf + 2), ELK_REGISTER_TYPE_D), retype(src, ELK_REGISTER_TYPE_D)); elk_pop_insn_state(p); uint32_t msg_type; if (devinfo->ver >= 7) msg_type = GFX7_DATAPORT_DC_OWORD_DUAL_BLOCK_WRITE; else if (devinfo->ver == 6) msg_type = GFX6_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE; else msg_type = ELK_DATAPORT_WRITE_MESSAGE_OWORD_DUAL_BLOCK_WRITE; elk_set_default_predicate_control(p, inst->predicate); /* Pre-gfx6, we have to specify write commits to ensure ordering * between reads and writes within a thread. Afterwards, that's * guaranteed and write commits only matter for inter-thread * synchronization. */ if (devinfo->ver >= 6) { write_commit = false; } else { /* The visitor set up our destination register to be g0. This * means that when the next read comes along, we will end up * reading from g0 and causing a block on the write commit. For * write-after-read, we are relying on the value of the previous * read being used (and thus blocking on completion) before our * write is executed. This means we have to be careful in * instruction scheduling to not violate this assumption. */ write_commit = true; } /* Each of the 8 channel enables is considered for whether each * dword is written. */ elk_inst *send = elk_next_insn(p, ELK_OPCODE_SEND); elk_inst_set_sfid(p->devinfo, send, target_cache); elk_set_dest(p, send, dst); elk_set_src0(p, send, header); if (devinfo->ver < 6) elk_inst_set_cond_modifier(p->devinfo, send, inst->base_mrf); elk_set_desc(p, send, elk_message_desc(devinfo, 3, write_commit, true) | elk_dp_write_desc(devinfo, elk_scratch_surface_idx(p), ELK_DATAPORT_OWORD_DUAL_BLOCK_1OWORD, msg_type, write_commit)); } static void generate_pull_constant_load(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg index, struct elk_reg offset) { const struct intel_device_info *devinfo = p->devinfo; const unsigned target_cache = (devinfo->ver >= 6 ? GFX6_SFID_DATAPORT_SAMPLER_CACHE : ELK_SFID_DATAPORT_READ); assert(index.file == ELK_IMMEDIATE_VALUE && index.type == ELK_REGISTER_TYPE_UD); uint32_t surf_index = index.ud; struct elk_reg header = elk_vec8_grf(0, 0); elk_gfx6_resolve_implied_move(p, &header, inst->base_mrf); if (devinfo->ver >= 6) { if (offset.file == ELK_IMMEDIATE_VALUE) { elk_MOV(p, retype(elk_message_reg(inst->base_mrf + 1), ELK_REGISTER_TYPE_D), elk_imm_d(offset.ud >> 4)); } else { elk_SHR(p, retype(elk_message_reg(inst->base_mrf + 1), ELK_REGISTER_TYPE_D), offset, elk_imm_d(4)); } } else { elk_MOV(p, retype(elk_message_reg(inst->base_mrf + 1), ELK_REGISTER_TYPE_D), offset); } uint32_t msg_type; if (devinfo->ver >= 6) msg_type = GFX6_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ; else if (devinfo->verx10 >= 45) msg_type = G45_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ; else msg_type = ELK_DATAPORT_READ_MESSAGE_OWORD_DUAL_BLOCK_READ; /* Each of the 8 channel enables is considered for whether each * dword is written. */ elk_inst *send = elk_next_insn(p, ELK_OPCODE_SEND); elk_inst_set_sfid(devinfo, send, target_cache); elk_set_dest(p, send, dst); elk_set_src0(p, send, header); if (devinfo->ver < 6) elk_inst_set_cond_modifier(p->devinfo, send, inst->base_mrf); elk_set_desc(p, send, elk_message_desc(devinfo, 2, 1, true) | elk_dp_read_desc(devinfo, surf_index, ELK_DATAPORT_OWORD_DUAL_BLOCK_1OWORD, msg_type, ELK_DATAPORT_READ_TARGET_DATA_CACHE)); } static void generate_get_buffer_size(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg src, struct elk_reg surf_index) { assert(p->devinfo->ver >= 7); assert(surf_index.type == ELK_REGISTER_TYPE_UD && surf_index.file == ELK_IMMEDIATE_VALUE); elk_SAMPLE(p, dst, inst->base_mrf, src, surf_index.ud, 0, GFX5_SAMPLER_MESSAGE_SAMPLE_RESINFO, 1, /* response length */ inst->mlen, inst->header_size > 0, ELK_SAMPLER_SIMD_MODE_SIMD4X2, ELK_SAMPLER_RETURN_FORMAT_SINT32); } static void generate_pull_constant_load_gfx7(struct elk_codegen *p, vec4_instruction *inst, struct elk_reg dst, struct elk_reg surf_index, struct elk_reg offset) { const struct intel_device_info *devinfo = p->devinfo; assert(surf_index.type == ELK_REGISTER_TYPE_UD); if (surf_index.file == ELK_IMMEDIATE_VALUE) { elk_inst *insn = elk_next_insn(p, ELK_OPCODE_SEND); elk_inst_set_sfid(devinfo, insn, ELK_SFID_SAMPLER); elk_set_dest(p, insn, dst); elk_set_src0(p, insn, offset); elk_set_desc(p, insn, elk_message_desc(devinfo, inst->mlen, 1, inst->header_size) | elk_sampler_desc(devinfo, surf_index.ud, 0, /* LD message ignores sampler unit */ GFX5_SAMPLER_MESSAGE_SAMPLE_LD, ELK_SAMPLER_SIMD_MODE_SIMD4X2, 0)); } else { struct elk_reg addr = vec1(retype(elk_address_reg(0), ELK_REGISTER_TYPE_UD)); elk_push_insn_state(p); elk_set_default_mask_control(p, ELK_MASK_DISABLE); elk_set_default_access_mode(p, ELK_ALIGN_1); /* a0.0 = surf_index & 0xff */ elk_inst *insn_and = elk_next_insn(p, ELK_OPCODE_AND); elk_inst_set_exec_size(devinfo, insn_and, ELK_EXECUTE_1); elk_set_dest(p, insn_and, addr); elk_set_src0(p, insn_and, vec1(retype(surf_index, ELK_REGISTER_TYPE_UD))); elk_set_src1(p, insn_and, elk_imm_ud(0x0ff)); elk_pop_insn_state(p); /* dst = send(offset, a0.0 | ) */ elk_send_indirect_message( p, ELK_SFID_SAMPLER, dst, offset, addr, elk_message_desc(devinfo, inst->mlen, 1, inst->header_size) | elk_sampler_desc(devinfo, 0 /* surface */, 0 /* sampler */, GFX5_SAMPLER_MESSAGE_SAMPLE_LD, ELK_SAMPLER_SIMD_MODE_SIMD4X2, 0), false /* EOT */); } } static void generate_mov_indirect(struct elk_codegen *p, vec4_instruction *, struct elk_reg dst, struct elk_reg reg, struct elk_reg indirect) { assert(indirect.type == ELK_REGISTER_TYPE_UD); assert(p->devinfo->ver >= 6); unsigned imm_byte_offset = reg.nr * REG_SIZE + reg.subnr * (REG_SIZE / 2); /* This instruction acts in align1 mode */ assert(dst.writemask == WRITEMASK_XYZW); if (indirect.file == ELK_IMMEDIATE_VALUE) { imm_byte_offset += indirect.ud; reg.nr = imm_byte_offset / REG_SIZE; reg.subnr = (imm_byte_offset / (REG_SIZE / 2)) % 2; unsigned shift = (imm_byte_offset / 4) % 4; reg.swizzle += ELK_SWIZZLE4(shift, shift, shift, shift); elk_MOV(p, dst, reg); } else { elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); struct elk_reg addr = vec8(elk_address_reg(0)); /* We need to move the indirect value into the address register. In * order to make things make some sense, we want to respect at least the * X component of the swizzle. In order to do that, we need to convert * the subnr (probably 0) to an align1 subnr and add in the swizzle. */ assert(elk_is_single_value_swizzle(indirect.swizzle)); indirect.subnr = (indirect.subnr * 4 + ELK_GET_SWZ(indirect.swizzle, 0)); /* We then use a region of <8,4,0>:uw to pick off the first 2 bytes of * the indirect and splat it out to all four channels of the given half * of a0. */ indirect.subnr *= 2; indirect = stride(retype(indirect, ELK_REGISTER_TYPE_UW), 8, 4, 0); elk_ADD(p, addr, indirect, elk_imm_uw(imm_byte_offset)); /* Now we need to incorporate the swizzle from the source register */ if (reg.swizzle != ELK_SWIZZLE_XXXX) { uint32_t uv_swiz = ELK_GET_SWZ(reg.swizzle, 0) << 2 | ELK_GET_SWZ(reg.swizzle, 1) << 6 | ELK_GET_SWZ(reg.swizzle, 2) << 10 | ELK_GET_SWZ(reg.swizzle, 3) << 14; uv_swiz |= uv_swiz << 16; elk_ADD(p, addr, addr, elk_imm_uv(uv_swiz)); } elk_MOV(p, dst, retype(elk_VxH_indirect(0, 0), reg.type)); elk_pop_insn_state(p); } } static void generate_zero_oob_push_regs(struct elk_codegen *p, struct elk_stage_prog_data *prog_data, struct elk_reg scratch, struct elk_reg bit_mask_in) { const uint64_t want_zero = prog_data->zero_push_reg; assert(want_zero); assert(bit_mask_in.file == ELK_GENERAL_REGISTER_FILE); assert(ELK_GET_SWZ(bit_mask_in.swizzle, 1) == ELK_GET_SWZ(bit_mask_in.swizzle, 0) + 1); bit_mask_in.subnr += ELK_GET_SWZ(bit_mask_in.swizzle, 0) * 4; bit_mask_in.type = ELK_REGISTER_TYPE_W; /* Scratch should be 3 registers in the GRF */ assert(scratch.file == ELK_GENERAL_REGISTER_FILE); scratch = vec8(scratch); struct elk_reg mask_w16 = retype(scratch, ELK_REGISTER_TYPE_W); struct elk_reg mask_d16 = retype(byte_offset(scratch, REG_SIZE), ELK_REGISTER_TYPE_D); elk_push_insn_state(p); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_set_default_mask_control(p, ELK_MASK_DISABLE); for (unsigned i = 0; i < 64; i++) { if (i % 16 == 0 && (want_zero & BITFIELD64_RANGE(i, 16))) { elk_set_default_exec_size(p, ELK_EXECUTE_8); elk_SHL(p, suboffset(mask_w16, 8), vec1(byte_offset(bit_mask_in, i / 8)), elk_imm_v(0x01234567)); elk_SHL(p, mask_w16, suboffset(mask_w16, 8), elk_imm_w(8)); elk_set_default_exec_size(p, ELK_EXECUTE_16); elk_ASR(p, mask_d16, mask_w16, elk_imm_w(15)); } if (want_zero & BITFIELD64_BIT(i)) { unsigned push_start = prog_data->dispatch_grf_start_reg; struct elk_reg push_reg = retype(elk_vec8_grf(push_start + i, 0), ELK_REGISTER_TYPE_D); elk_set_default_exec_size(p, ELK_EXECUTE_8); elk_AND(p, push_reg, push_reg, vec1(suboffset(mask_d16, i))); } } elk_pop_insn_state(p); } static void generate_code(struct elk_codegen *p, const struct elk_compiler *compiler, const struct elk_compile_params *params, const nir_shader *nir, struct elk_vue_prog_data *prog_data, const struct elk_cfg_t *cfg, const performance &perf, struct elk_compile_stats *stats, bool debug_enabled) { const struct intel_device_info *devinfo = p->devinfo; const char *stage_abbrev = _mesa_shader_stage_to_abbrev(nir->info.stage); struct elk_disasm_info *elk_disasm_info = elk_disasm_initialize(p->isa, cfg); /* `send_count` explicitly does not include spills or fills, as we'd * like to use it as a metric for intentional memory access or other * shared function use. Otherwise, subtle changes to scheduling or * register allocation could cause it to fluctuate wildly - and that * effect is already counted in spill/fill counts. */ int spill_count = 0, fill_count = 0; int loop_count = 0, send_count = 0; foreach_block_and_inst (block, vec4_instruction, inst, cfg) { struct elk_reg src[3], dst; if (unlikely(debug_enabled)) elk_disasm_annotate(elk_disasm_info, inst, p->next_insn_offset); for (unsigned int i = 0; i < 3; i++) { src[i] = inst->src[i].as_elk_reg(); } dst = inst->dst.as_elk_reg(); elk_set_default_predicate_control(p, inst->predicate); elk_set_default_predicate_inverse(p, inst->predicate_inverse); elk_set_default_flag_reg(p, inst->flag_subreg / 2, inst->flag_subreg % 2); elk_set_default_saturate(p, inst->saturate); elk_set_default_mask_control(p, inst->force_writemask_all); elk_set_default_acc_write_control(p, inst->writes_accumulator); assert(inst->group % inst->exec_size == 0); assert(inst->group % 4 == 0); /* There are some instructions where the destination is 64-bit * but we retype it to a smaller type. In that case, we cannot * double the exec_size. */ const bool is_df = (get_exec_type_size(inst) == 8 || inst->dst.type == ELK_REGISTER_TYPE_DF) && inst->opcode != ELK_VEC4_OPCODE_PICK_LOW_32BIT && inst->opcode != ELK_VEC4_OPCODE_PICK_HIGH_32BIT && inst->opcode != ELK_VEC4_OPCODE_SET_LOW_32BIT && inst->opcode != ELK_VEC4_OPCODE_SET_HIGH_32BIT; unsigned exec_size = inst->exec_size; if (devinfo->verx10 == 70 && is_df) exec_size *= 2; elk_set_default_exec_size(p, cvt(exec_size) - 1); if (!inst->force_writemask_all) elk_set_default_group(p, inst->group); assert(inst->base_mrf + inst->mlen <= ELK_MAX_MRF(devinfo->ver)); assert(inst->mlen <= ELK_MAX_MSG_LENGTH); unsigned pre_emit_nr_insn = p->nr_insn; switch (inst->opcode) { case ELK_VEC4_OPCODE_UNPACK_UNIFORM: case ELK_OPCODE_MOV: case ELK_VEC4_OPCODE_MOV_FOR_SCRATCH: elk_MOV(p, dst, src[0]); break; case ELK_OPCODE_ADD: elk_ADD(p, dst, src[0], src[1]); break; case ELK_OPCODE_MUL: elk_MUL(p, dst, src[0], src[1]); break; case ELK_OPCODE_MACH: elk_MACH(p, dst, src[0], src[1]); break; case ELK_OPCODE_MAD: assert(devinfo->ver >= 6); elk_MAD(p, dst, src[0], src[1], src[2]); break; case ELK_OPCODE_FRC: elk_FRC(p, dst, src[0]); break; case ELK_OPCODE_RNDD: elk_RNDD(p, dst, src[0]); break; case ELK_OPCODE_RNDE: elk_RNDE(p, dst, src[0]); break; case ELK_OPCODE_RNDZ: elk_RNDZ(p, dst, src[0]); break; case ELK_OPCODE_AND: elk_AND(p, dst, src[0], src[1]); break; case ELK_OPCODE_OR: elk_OR(p, dst, src[0], src[1]); break; case ELK_OPCODE_XOR: elk_XOR(p, dst, src[0], src[1]); break; case ELK_OPCODE_NOT: elk_NOT(p, dst, src[0]); break; case ELK_OPCODE_ASR: elk_ASR(p, dst, src[0], src[1]); break; case ELK_OPCODE_SHR: elk_SHR(p, dst, src[0], src[1]); break; case ELK_OPCODE_SHL: elk_SHL(p, dst, src[0], src[1]); break; case ELK_OPCODE_CMP: elk_CMP(p, dst, inst->conditional_mod, src[0], src[1]); break; case ELK_OPCODE_CMPN: elk_CMPN(p, dst, inst->conditional_mod, src[0], src[1]); break; case ELK_OPCODE_SEL: elk_SEL(p, dst, src[0], src[1]); break; case ELK_OPCODE_DPH: elk_DPH(p, dst, src[0], src[1]); break; case ELK_OPCODE_DP4: elk_DP4(p, dst, src[0], src[1]); break; case ELK_OPCODE_DP3: elk_DP3(p, dst, src[0], src[1]); break; case ELK_OPCODE_DP2: elk_DP2(p, dst, src[0], src[1]); break; case ELK_OPCODE_F32TO16: assert(devinfo->ver >= 7); elk_F32TO16(p, dst, src[0]); break; case ELK_OPCODE_F16TO32: assert(devinfo->ver >= 7); elk_F16TO32(p, dst, src[0]); break; case ELK_OPCODE_LRP: assert(devinfo->ver >= 6); elk_LRP(p, dst, src[0], src[1], src[2]); break; case ELK_OPCODE_BFREV: assert(devinfo->ver >= 7); elk_BFREV(p, retype(dst, ELK_REGISTER_TYPE_UD), retype(src[0], ELK_REGISTER_TYPE_UD)); break; case ELK_OPCODE_FBH: assert(devinfo->ver >= 7); elk_FBH(p, retype(dst, src[0].type), src[0]); break; case ELK_OPCODE_FBL: assert(devinfo->ver >= 7); elk_FBL(p, retype(dst, ELK_REGISTER_TYPE_UD), retype(src[0], ELK_REGISTER_TYPE_UD)); break; case ELK_OPCODE_LZD: elk_LZD(p, dst, src[0]); break; case ELK_OPCODE_CBIT: assert(devinfo->ver >= 7); elk_CBIT(p, retype(dst, ELK_REGISTER_TYPE_UD), retype(src[0], ELK_REGISTER_TYPE_UD)); break; case ELK_OPCODE_ADDC: assert(devinfo->ver >= 7); elk_ADDC(p, dst, src[0], src[1]); break; case ELK_OPCODE_SUBB: assert(devinfo->ver >= 7); elk_SUBB(p, dst, src[0], src[1]); break; case ELK_OPCODE_MAC: elk_MAC(p, dst, src[0], src[1]); break; case ELK_OPCODE_BFE: assert(devinfo->ver >= 7); elk_BFE(p, dst, src[0], src[1], src[2]); break; case ELK_OPCODE_BFI1: assert(devinfo->ver >= 7); elk_BFI1(p, dst, src[0], src[1]); break; case ELK_OPCODE_BFI2: assert(devinfo->ver >= 7); elk_BFI2(p, dst, src[0], src[1], src[2]); break; case ELK_OPCODE_IF: if (!inst->src[0].is_null()) { /* The instruction has an embedded compare (only allowed on gfx6) */ assert(devinfo->ver == 6); elk_gfx6_IF(p, inst->conditional_mod, src[0], src[1]); } else { elk_inst *if_inst = elk_IF(p, ELK_EXECUTE_8); elk_inst_set_pred_control(p->devinfo, if_inst, inst->predicate); } break; case ELK_OPCODE_ELSE: elk_ELSE(p); break; case ELK_OPCODE_ENDIF: elk_ENDIF(p); break; case ELK_OPCODE_DO: elk_DO(p, ELK_EXECUTE_8); break; case ELK_OPCODE_BREAK: elk_BREAK(p); elk_set_default_predicate_control(p, ELK_PREDICATE_NONE); break; case ELK_OPCODE_CONTINUE: elk_CONT(p); elk_set_default_predicate_control(p, ELK_PREDICATE_NONE); break; case ELK_OPCODE_WHILE: elk_WHILE(p); loop_count++; break; case ELK_SHADER_OPCODE_RCP: case ELK_SHADER_OPCODE_RSQ: case ELK_SHADER_OPCODE_SQRT: case ELK_SHADER_OPCODE_EXP2: case ELK_SHADER_OPCODE_LOG2: case ELK_SHADER_OPCODE_SIN: case ELK_SHADER_OPCODE_COS: assert(inst->conditional_mod == ELK_CONDITIONAL_NONE); if (devinfo->ver >= 7) { elk_gfx6_math(p, dst, elk_math_function(inst->opcode), src[0], elk_null_reg()); } else if (devinfo->ver == 6) { generate_math_gfx6(p, inst, dst, src[0], elk_null_reg()); } else { generate_math1_gfx4(p, inst, dst, src[0]); send_count++; } break; case ELK_SHADER_OPCODE_POW: case ELK_SHADER_OPCODE_INT_QUOTIENT: case ELK_SHADER_OPCODE_INT_REMAINDER: assert(inst->conditional_mod == ELK_CONDITIONAL_NONE); if (devinfo->ver >= 7) { elk_gfx6_math(p, dst, elk_math_function(inst->opcode), src[0], src[1]); } else if (devinfo->ver == 6) { generate_math_gfx6(p, inst, dst, src[0], src[1]); } else { generate_math2_gfx4(p, inst, dst, src[0], src[1]); send_count++; } break; case ELK_SHADER_OPCODE_TEX: case ELK_SHADER_OPCODE_TXD: case ELK_SHADER_OPCODE_TXF: case ELK_SHADER_OPCODE_TXF_CMS: case ELK_SHADER_OPCODE_TXF_CMS_W: case ELK_SHADER_OPCODE_TXF_MCS: case ELK_SHADER_OPCODE_TXL: case ELK_SHADER_OPCODE_TXS: case ELK_SHADER_OPCODE_TG4: case ELK_SHADER_OPCODE_TG4_OFFSET: case ELK_SHADER_OPCODE_SAMPLEINFO: generate_tex(p, prog_data, nir->info.stage, inst, dst, src[0], src[1], src[2]); send_count++; break; case ELK_SHADER_OPCODE_GET_BUFFER_SIZE: generate_get_buffer_size(p, inst, dst, src[0], src[1]); send_count++; break; case ELK_VEC4_VS_OPCODE_URB_WRITE: generate_vs_urb_write(p, inst); send_count++; break; case ELK_SHADER_OPCODE_GFX4_SCRATCH_READ: generate_scratch_read(p, inst, dst, src[0]); fill_count++; break; case ELK_SHADER_OPCODE_GFX4_SCRATCH_WRITE: generate_scratch_write(p, inst, dst, src[0], src[1]); spill_count++; break; case ELK_VS_OPCODE_PULL_CONSTANT_LOAD: generate_pull_constant_load(p, inst, dst, src[0], src[1]); send_count++; break; case ELK_VS_OPCODE_PULL_CONSTANT_LOAD_GFX7: generate_pull_constant_load_gfx7(p, inst, dst, src[0], src[1]); send_count++; break; case ELK_VEC4_GS_OPCODE_URB_WRITE: generate_gs_urb_write(p, inst); send_count++; break; case ELK_VEC4_GS_OPCODE_URB_WRITE_ALLOCATE: generate_gs_urb_write_allocate(p, inst); send_count++; break; case ELK_GS_OPCODE_SVB_WRITE: generate_gs_svb_write(p, inst, dst, src[0], src[1]); send_count++; break; case ELK_GS_OPCODE_SVB_SET_DST_INDEX: generate_gs_svb_set_destination_index(p, inst, dst, src[0]); break; case ELK_GS_OPCODE_THREAD_END: generate_gs_thread_end(p, inst); send_count++; break; case ELK_GS_OPCODE_SET_WRITE_OFFSET: generate_gs_set_write_offset(p, dst, src[0], src[1]); break; case ELK_GS_OPCODE_SET_VERTEX_COUNT: generate_gs_set_vertex_count(p, dst, src[0]); break; case ELK_GS_OPCODE_FF_SYNC: generate_gs_ff_sync(p, inst, dst, src[0], src[1]); send_count++; break; case ELK_GS_OPCODE_FF_SYNC_SET_PRIMITIVES: generate_gs_ff_sync_set_primitives(p, dst, src[0], src[1], src[2]); break; case ELK_GS_OPCODE_SET_PRIMITIVE_ID: generate_gs_set_primitive_id(p, dst); break; case ELK_GS_OPCODE_SET_DWORD_2: generate_gs_set_dword_2(p, dst, src[0]); break; case ELK_GS_OPCODE_PREPARE_CHANNEL_MASKS: generate_gs_prepare_channel_masks(p, dst); break; case ELK_GS_OPCODE_SET_CHANNEL_MASKS: generate_gs_set_channel_masks(p, dst, src[0]); break; case ELK_GS_OPCODE_GET_INSTANCE_ID: generate_gs_get_instance_id(p, dst); break; case ELK_VEC4_OPCODE_UNTYPED_ATOMIC: assert(src[2].file == ELK_IMMEDIATE_VALUE); elk_untyped_atomic(p, dst, src[0], src[1], src[2].ud, inst->mlen, !inst->dst.is_null(), inst->header_size); send_count++; break; case ELK_VEC4_OPCODE_UNTYPED_SURFACE_READ: assert(!inst->header_size); assert(src[2].file == ELK_IMMEDIATE_VALUE); elk_untyped_surface_read(p, dst, src[0], src[1], inst->mlen, src[2].ud); send_count++; break; case ELK_VEC4_OPCODE_UNTYPED_SURFACE_WRITE: assert(src[2].file == ELK_IMMEDIATE_VALUE); elk_untyped_surface_write(p, src[0], src[1], inst->mlen, src[2].ud, inst->header_size); send_count++; break; case ELK_SHADER_OPCODE_MEMORY_FENCE: elk_memory_fence(p, dst, src[0], ELK_OPCODE_SEND, elk_message_target(inst->sfid), inst->desc, /* commit_enable */ false, /* bti */ 0); send_count++; break; case ELK_SHADER_OPCODE_FIND_LIVE_CHANNEL: elk_find_live_channel(p, dst, false); break; case ELK_SHADER_OPCODE_BROADCAST: assert(inst->force_writemask_all); elk_broadcast(p, dst, src[0], src[1]); break; case ELK_VS_OPCODE_UNPACK_FLAGS_SIMD4X2: generate_unpack_flags(p, dst); break; case ELK_VEC4_OPCODE_MOV_BYTES: { /* Moves the low byte from each channel, using an Align1 access mode * and a <4,1,0> source region. */ assert(src[0].type == ELK_REGISTER_TYPE_UB || src[0].type == ELK_REGISTER_TYPE_B); elk_set_default_access_mode(p, ELK_ALIGN_1); src[0].vstride = ELK_VERTICAL_STRIDE_4; src[0].width = ELK_WIDTH_1; src[0].hstride = ELK_HORIZONTAL_STRIDE_0; elk_MOV(p, dst, src[0]); elk_set_default_access_mode(p, ELK_ALIGN_16); break; } case ELK_VEC4_OPCODE_DOUBLE_TO_F32: case ELK_VEC4_OPCODE_DOUBLE_TO_D32: case ELK_VEC4_OPCODE_DOUBLE_TO_U32: { assert(type_sz(src[0].type) == 8); assert(type_sz(dst.type) == 8); elk_reg_type dst_type; switch (inst->opcode) { case ELK_VEC4_OPCODE_DOUBLE_TO_F32: dst_type = ELK_REGISTER_TYPE_F; break; case ELK_VEC4_OPCODE_DOUBLE_TO_D32: dst_type = ELK_REGISTER_TYPE_D; break; case ELK_VEC4_OPCODE_DOUBLE_TO_U32: dst_type = ELK_REGISTER_TYPE_UD; break; default: unreachable("Not supported conversion"); } dst = retype(dst, dst_type); elk_set_default_access_mode(p, ELK_ALIGN_1); /* When converting from DF->F, we set destination's stride as 2 as an * alignment requirement. But in IVB/BYT, each DF implicitly writes * two floats, being the first one the converted value. So we don't * need to explicitly set stride 2, but 1. */ struct elk_reg spread_dst; if (devinfo->verx10 == 70) spread_dst = stride(dst, 8, 4, 1); else spread_dst = stride(dst, 8, 4, 2); elk_MOV(p, spread_dst, src[0]); elk_set_default_access_mode(p, ELK_ALIGN_16); break; } case ELK_VEC4_OPCODE_TO_DOUBLE: { assert(type_sz(src[0].type) == 4); assert(type_sz(dst.type) == 8); elk_set_default_access_mode(p, ELK_ALIGN_1); elk_MOV(p, dst, src[0]); elk_set_default_access_mode(p, ELK_ALIGN_16); break; } case ELK_VEC4_OPCODE_PICK_LOW_32BIT: case ELK_VEC4_OPCODE_PICK_HIGH_32BIT: { /* Stores the low/high 32-bit of each 64-bit element in src[0] into * dst using ALIGN1 mode and a <8,4,2>:UD region on the source. */ assert(type_sz(src[0].type) == 8); assert(type_sz(dst.type) == 4); elk_set_default_access_mode(p, ELK_ALIGN_1); dst = retype(dst, ELK_REGISTER_TYPE_UD); dst.hstride = ELK_HORIZONTAL_STRIDE_1; src[0] = retype(src[0], ELK_REGISTER_TYPE_UD); if (inst->opcode == ELK_VEC4_OPCODE_PICK_HIGH_32BIT) src[0] = suboffset(src[0], 1); src[0] = spread(src[0], 2); elk_MOV(p, dst, src[0]); elk_set_default_access_mode(p, ELK_ALIGN_16); break; } case ELK_VEC4_OPCODE_SET_LOW_32BIT: case ELK_VEC4_OPCODE_SET_HIGH_32BIT: { /* Reads consecutive 32-bit elements from src[0] and writes * them to the low/high 32-bit of each 64-bit element in dst. */ assert(type_sz(src[0].type) == 4); assert(type_sz(dst.type) == 8); elk_set_default_access_mode(p, ELK_ALIGN_1); dst = retype(dst, ELK_REGISTER_TYPE_UD); if (inst->opcode == ELK_VEC4_OPCODE_SET_HIGH_32BIT) dst = suboffset(dst, 1); dst.hstride = ELK_HORIZONTAL_STRIDE_2; src[0] = retype(src[0], ELK_REGISTER_TYPE_UD); elk_MOV(p, dst, src[0]); elk_set_default_access_mode(p, ELK_ALIGN_16); break; } case ELK_VEC4_OPCODE_PACK_BYTES: { /* Is effectively: * * mov(8) dst<16,4,1>:UB src<4,1,0>:UB * * but destinations' only regioning is horizontal stride, so instead we * have to use two instructions: * * mov(4) dst<1>:UB src<4,1,0>:UB * mov(4) dst.16<1>:UB src.16<4,1,0>:UB * * where they pack the four bytes from the low and high four DW. */ assert(util_is_power_of_two_nonzero(dst.writemask)); unsigned offset = __builtin_ctz(dst.writemask); dst.type = ELK_REGISTER_TYPE_UB; elk_set_default_access_mode(p, ELK_ALIGN_1); src[0].type = ELK_REGISTER_TYPE_UB; src[0].vstride = ELK_VERTICAL_STRIDE_4; src[0].width = ELK_WIDTH_1; src[0].hstride = ELK_HORIZONTAL_STRIDE_0; dst.subnr = offset * 4; struct elk_inst *insn = elk_MOV(p, dst, src[0]); elk_inst_set_exec_size(p->devinfo, insn, ELK_EXECUTE_4); elk_inst_set_no_dd_clear(p->devinfo, insn, true); elk_inst_set_no_dd_check(p->devinfo, insn, inst->no_dd_check); src[0].subnr = 16; dst.subnr = 16 + offset * 4; insn = elk_MOV(p, dst, src[0]); elk_inst_set_exec_size(p->devinfo, insn, ELK_EXECUTE_4); elk_inst_set_no_dd_clear(p->devinfo, insn, inst->no_dd_clear); elk_inst_set_no_dd_check(p->devinfo, insn, true); elk_set_default_access_mode(p, ELK_ALIGN_16); break; } case ELK_VEC4_OPCODE_ZERO_OOB_PUSH_REGS: generate_zero_oob_push_regs(p, &prog_data->base, dst, src[0]); break; case ELK_VEC4_TCS_OPCODE_URB_WRITE: generate_tcs_urb_write(p, inst, src[0]); send_count++; break; case ELK_VEC4_OPCODE_URB_READ: generate_vec4_urb_read(p, inst, dst, src[0]); send_count++; break; case ELK_VEC4_TCS_OPCODE_SET_INPUT_URB_OFFSETS: generate_tcs_input_urb_offsets(p, dst, src[0], src[1]); break; case ELK_VEC4_TCS_OPCODE_SET_OUTPUT_URB_OFFSETS: generate_tcs_output_urb_offsets(p, dst, src[0], src[1]); break; case ELK_TCS_OPCODE_GET_INSTANCE_ID: generate_tcs_get_instance_id(p, dst); break; case ELK_TCS_OPCODE_GET_PRIMITIVE_ID: generate_tcs_get_primitive_id(p, dst); break; case ELK_TCS_OPCODE_CREATE_BARRIER_HEADER: generate_tcs_create_barrier_header(p, prog_data, dst); break; case ELK_TES_OPCODE_CREATE_INPUT_READ_HEADER: generate_tes_create_input_read_header(p, dst); break; case ELK_TES_OPCODE_ADD_INDIRECT_URB_OFFSET: generate_tes_add_indirect_urb_offset(p, dst, src[0], src[1]); break; case ELK_TES_OPCODE_GET_PRIMITIVE_ID: generate_tes_get_primitive_id(p, dst); break; case ELK_TCS_OPCODE_SRC0_010_IS_ZERO: /* If src_reg had stride like elk_fs_reg, we wouldn't need this. */ elk_MOV(p, elk_null_reg(), stride(src[0], 0, 1, 0)); break; case ELK_TCS_OPCODE_RELEASE_INPUT: generate_tcs_release_input(p, dst, src[0], src[1]); send_count++; break; case ELK_TCS_OPCODE_THREAD_END: generate_tcs_thread_end(p, inst); send_count++; break; case ELK_SHADER_OPCODE_BARRIER: elk_barrier(p, src[0]); elk_WAIT(p); send_count++; break; case ELK_SHADER_OPCODE_MOV_INDIRECT: generate_mov_indirect(p, inst, dst, src[0], src[1]); break; case ELK_OPCODE_DIM: assert(devinfo->verx10 == 75); assert(src[0].type == ELK_REGISTER_TYPE_DF); assert(dst.type == ELK_REGISTER_TYPE_DF); elk_DIM(p, dst, retype(src[0], ELK_REGISTER_TYPE_F)); break; case ELK_SHADER_OPCODE_RND_MODE: { assert(src[0].file == ELK_IMMEDIATE_VALUE); /* * Changes the floating point rounding mode updating the control * register field defined at cr0.0[5-6] bits. */ enum elk_rnd_mode mode = (enum elk_rnd_mode) (src[0].d << ELK_CR0_RND_MODE_SHIFT); elk_float_controls_mode(p, mode, ELK_CR0_RND_MODE_MASK); } break; default: unreachable("Unsupported opcode"); } if (inst->opcode == ELK_VEC4_OPCODE_PACK_BYTES) { /* Handled dependency hints in the generator. */ assert(!inst->conditional_mod); } else if (inst->no_dd_clear || inst->no_dd_check || inst->conditional_mod) { assert(p->nr_insn == pre_emit_nr_insn + 1 || !"conditional_mod, no_dd_check, or no_dd_clear set for IR " "emitting more than 1 instruction"); elk_inst *last = &p->store[pre_emit_nr_insn]; if (inst->conditional_mod) elk_inst_set_cond_modifier(p->devinfo, last, inst->conditional_mod); elk_inst_set_no_dd_clear(p->devinfo, last, inst->no_dd_clear); elk_inst_set_no_dd_check(p->devinfo, last, inst->no_dd_check); } } elk_set_uip_jip(p, 0); /* end of program sentinel */ elk_disasm_new_inst_group(elk_disasm_info, p->next_insn_offset); #ifndef NDEBUG bool validated = #else if (unlikely(debug_enabled)) #endif elk_validate_instructions(&compiler->isa, p->store, 0, p->next_insn_offset, elk_disasm_info); int before_size = p->next_insn_offset; elk_compact_instructions(p, 0, elk_disasm_info); int after_size = p->next_insn_offset; bool dump_shader_bin = elk_should_dump_shader_bin(); unsigned char sha1[21]; char sha1buf[41]; if (unlikely(debug_enabled || dump_shader_bin)) { _mesa_sha1_compute(p->store, p->next_insn_offset, sha1); _mesa_sha1_format(sha1buf, sha1); } if (unlikely(dump_shader_bin)) elk_dump_shader_bin(p->store, 0, p->next_insn_offset, sha1buf); if (unlikely(debug_enabled)) { fprintf(stderr, "Native code for %s %s shader %s (src_hash 0x%08x) (sha1 %s):\n", nir->info.label ? nir->info.label : "unnamed", _mesa_shader_stage_to_string(nir->info.stage), nir->info.name, params->source_hash, sha1buf); fprintf(stderr, "%s vec4 shader: %d instructions. %d loops. %u cycles. %d:%d " "spills:fills, %u sends. Compacted %d to %d bytes (%.0f%%)\n", stage_abbrev, before_size / 16, loop_count, perf.latency, spill_count, fill_count, send_count, before_size, after_size, 100.0f * (before_size - after_size) / before_size); /* overriding the shader makes elk_disasm_info invalid */ if (!elk_try_override_assembly(p, 0, sha1buf)) { elk_dump_assembly(p->store, 0, p->next_insn_offset, elk_disasm_info, perf.block_latency); } else { fprintf(stderr, "Successfully overrode shader with sha1 %s\n\n", sha1buf); } } ralloc_free(elk_disasm_info); assert(validated); elk_shader_debug_log(compiler, params->log_data, "%s vec4 shader: %d inst, %d loops, %u cycles, " "%d:%d spills:fills, %u sends, " "compacted %d to %d bytes.\n", stage_abbrev, before_size / 16, loop_count, perf.latency, spill_count, fill_count, send_count, before_size, after_size); if (stats) { stats->dispatch_width = 0; stats->max_dispatch_width = 0; stats->instructions = before_size / 16; stats->sends = send_count; stats->loops = loop_count; stats->cycles = perf.latency; stats->spills = spill_count; stats->fills = fill_count; } } extern "C" const unsigned * elk_vec4_generate_assembly(const struct elk_compiler *compiler, const struct elk_compile_params *params, const nir_shader *nir, struct elk_vue_prog_data *prog_data, const struct elk_cfg_t *cfg, const performance &perf, bool debug_enabled) { struct elk_codegen *p = rzalloc(params->mem_ctx, struct elk_codegen); elk_init_codegen(&compiler->isa, p, params->mem_ctx); elk_set_default_access_mode(p, ELK_ALIGN_16); generate_code(p, compiler, params, nir, prog_data, cfg, perf, params->stats, debug_enabled); assert(prog_data->base.const_data_size == 0); if (nir->constant_data_size > 0) { prog_data->base.const_data_size = nir->constant_data_size; prog_data->base.const_data_offset = elk_append_data(p, nir->constant_data, nir->constant_data_size, 32); } return elk_get_program(p, &prog_data->base.program_size); }