void
vec4_generator::generate_gs_set_write_offset(struct brw_reg dst,
struct brw_reg src0,
struct brw_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 { Align1 WE_all }
*/
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_MUL(p, suboffset(stride(dst, 2, 2, 1), 3), stride(src0, 8, 2, 4),
src1);
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_pop_insn_state(p);
}
void
vec4_generator::generate_gs_set_vertex_count(struct brw_reg dst,
struct brw_reg src)
{
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_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 }
*/
brw_MOV(p, suboffset(stride(retype(dst, BRW_REGISTER_TYPE_UW), 2, 2, 1), 4),
stride(retype(src, BRW_REGISTER_TYPE_UW), 8, 1, 0));
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_pop_insn_state(p);
}
void
vec4_generator::generate_gs_set_dword_2_immed(struct brw_reg dst,
struct brw_reg src)
{
assert(src.file == BRW_IMMEDIATE_VALUE);
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_MOV(p, suboffset(vec1(dst), 2), src);
brw_set_default_access_mode(p, BRW_ALIGN_16);
brw_pop_insn_state(p);
}
void
vec4_generator::generate_gs_prepare_channel_masks(struct brw_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);
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_SHL(p, dst, dst, brw_imm_ud(4));
brw_pop_insn_state(p);
}
void
vec4_generator::generate_unpack_flags(vec4_instruction *inst,
struct brw_reg dst)
{
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_set_default_access_mode(p, BRW_ALIGN_1);
struct brw_reg flags = brw_flag_reg(0, 0);
struct brw_reg dst_0 = suboffset(vec1(dst), 0);
struct brw_reg dst_4 = suboffset(vec1(dst), 4);
brw_AND(p, dst_0, flags, brw_imm_ud(0x0f));
brw_AND(p, dst_4, flags, brw_imm_ud(0xf0));
brw_SHR(p, dst_4, dst_4, brw_imm_ud(4));
brw_pop_insn_state(p);
}
void
brw_init_codegen(const struct brw_device_info *devinfo,
struct brw_codegen *p, void *mem_ctx)
{
memset(p, 0, sizeof(*p));
p->devinfo = devinfo;
/*
* Set the initial instruction store array size to 1024, if found that
* isn't enough, then it will double the store size at brw_next_insn()
* until out of memory.
*/
p->store_size = 1024;
p->store = rzalloc_array(mem_ctx, brw_inst, p->store_size);
p->nr_insn = 0;
p->current = p->stack;
p->compressed = false;
memset(p->current, 0, sizeof(p->current[0]));
p->mem_ctx = mem_ctx;
/* Some defaults?
*/
brw_set_default_exec_size(p, BRW_EXECUTE_8);
brw_set_default_mask_control(p, BRW_MASK_ENABLE); /* what does this do? */
brw_set_default_saturate(p, 0);
brw_set_default_compression_control(p, BRW_COMPRESSION_NONE);
/* Set up control flow stack */
p->if_stack_depth = 0;
p->if_stack_array_size = 16;
p->if_stack = rzalloc_array(mem_ctx, int, p->if_stack_array_size);
p->loop_stack_depth = 0;
p->loop_stack_array_size = 16;
p->loop_stack = rzalloc_array(mem_ctx, int, p->loop_stack_array_size);
p->if_depth_in_loop = rzalloc_array(mem_ctx, int, p->loop_stack_array_size);
brw_init_compaction_tables(devinfo);
}
void
vec4_generator::generate_oword_dual_block_offsets(struct brw_reg m1,
struct brw_reg index)
{
int second_vertex_offset;
if (brw->gen >= 6)
second_vertex_offset = 1;
else
second_vertex_offset = 16;
m1 = retype(m1, BRW_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 brw_reg m1_0 = suboffset(vec1(m1), 0);
struct brw_reg m1_4 = suboffset(vec1(m1), 4);
struct brw_reg index_0 = suboffset(vec1(index), 0);
struct brw_reg index_4 = suboffset(vec1(index), 4);
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_MOV(p, m1_0, index_0);
if (index.file == BRW_IMMEDIATE_VALUE) {
index_4.dw1.ud += second_vertex_offset;
brw_MOV(p, m1_4, index_4);
} else {
brw_ADD(p, m1_4, index_4, brw_imm_d(second_vertex_offset));
}
brw_pop_insn_state(p);
}
void
brw_codegen_ff_gs_prog(struct brw_context *brw,
struct brw_ff_gs_prog_key *key)
{
struct brw_ff_gs_compile c;
const GLuint *program;
void *mem_ctx;
GLuint program_size;
memset(&c, 0, sizeof(c));
c.key = *key;
c.vue_map = brw->vs.prog_data->base.vue_map;
c.nr_regs = (c.vue_map.num_slots + 1)/2;
mem_ctx = ralloc_context(NULL);
/* Begin the compilation:
*/
brw_init_codegen(brw->intelScreen->devinfo, &c.func, mem_ctx);
c.func.single_program_flow = 1;
/* For some reason the thread is spawned with only 4 channels
* unmasked.
*/
brw_set_default_mask_control(&c.func, BRW_MASK_DISABLE);
if (brw->gen >= 6) {
unsigned num_verts;
bool check_edge_flag;
/* On Sandybridge, we use the GS for implementing transform feedback
* (called "Stream Out" in the PRM).
*/
switch (key->primitive) {
case _3DPRIM_POINTLIST:
num_verts = 1;
check_edge_flag = false;
break;
case _3DPRIM_LINELIST:
case _3DPRIM_LINESTRIP:
case _3DPRIM_LINELOOP:
num_verts = 2;
check_edge_flag = false;
break;
case _3DPRIM_TRILIST:
case _3DPRIM_TRIFAN:
case _3DPRIM_TRISTRIP:
case _3DPRIM_RECTLIST:
num_verts = 3;
check_edge_flag = false;
break;
case _3DPRIM_QUADLIST:
case _3DPRIM_QUADSTRIP:
case _3DPRIM_POLYGON:
num_verts = 3;
check_edge_flag = true;
break;
default:
unreachable("Unexpected primitive type in Gen6 SOL program.");
}
gen6_sol_program(&c, key, num_verts, check_edge_flag);
} else {
/* On Gen4-5, we use the GS to decompose certain types of primitives.
* Note that primitives which don't require a GS program have already
* been weeded out by now.
*/
switch (key->primitive) {
case _3DPRIM_QUADLIST:
brw_ff_gs_quads( &c, key );
break;
case _3DPRIM_QUADSTRIP:
brw_ff_gs_quad_strip( &c, key );
break;
case _3DPRIM_LINELOOP:
brw_ff_gs_lines( &c );
break;
default:
ralloc_free(mem_ctx);
return;
}
}
brw_compact_instructions(&c.func, 0, 0, NULL);
/* get the program
*/
program = brw_get_program(&c.func, &program_size);
if (unlikely(INTEL_DEBUG & DEBUG_GS)) {
fprintf(stderr, "gs:\n");
brw_disassemble(brw->intelScreen->devinfo, c.func.store,
0, program_size, stderr);
fprintf(stderr, "\n");
}
brw_upload_cache(&brw->cache, BRW_CACHE_FF_GS_PROG,
&c.key, sizeof(c.key),
program, program_size,
&c.prog_data, sizeof(c.prog_data),
&brw->ff_gs.prog_offset, &brw->ff_gs.prog_data);
ralloc_free(mem_ctx);
}
static void compile_clip_prog( struct brw_context *brw,
struct brw_clip_prog_key *key )
{
struct brw_clip_compile c;
const GLuint *program;
void *mem_ctx;
GLuint program_size;
memset(&c, 0, sizeof(c));
mem_ctx = ralloc_context(NULL);
/* Begin the compilation:
*/
brw_init_codegen(&brw->screen->devinfo, &c.func, mem_ctx);
c.func.single_program_flow = 1;
c.key = *key;
c.vue_map = brw->vue_map_geom_out;
/* nr_regs is the number of registers filled by reading data from the VUE.
* This program accesses the entire VUE, so nr_regs needs to be the size of
* the VUE (measured in pairs, since two slots are stored in each
* register).
*/
c.nr_regs = (c.vue_map.num_slots + 1)/2;
c.prog_data.clip_mode = c.key.clip_mode; /* XXX */
/* For some reason the thread is spawned with only 4 channels
* unmasked.
*/
brw_set_default_mask_control(&c.func, BRW_MASK_DISABLE);
/* Would ideally have the option of producing a program which could
* do all three:
*/
switch (key->primitive) {
case GL_TRIANGLES:
if (key->do_unfilled)
brw_emit_unfilled_clip( &c );
else
brw_emit_tri_clip( &c );
break;
case GL_LINES:
brw_emit_line_clip( &c );
break;
case GL_POINTS:
brw_emit_point_clip( &c );
break;
default:
unreachable("not reached");
}
brw_compact_instructions(&c.func, 0, 0, NULL);
/* get the program
*/
program = brw_get_program(&c.func, &program_size);
if (unlikely(INTEL_DEBUG & DEBUG_CLIP)) {
fprintf(stderr, "clip:\n");
brw_disassemble(&brw->screen->devinfo, c.func.store,
0, program_size, stderr);
fprintf(stderr, "\n");
}
brw_upload_cache(&brw->cache,
BRW_CACHE_CLIP_PROG,
&c.key, sizeof(c.key),
program, program_size,
&c.prog_data, sizeof(c.prog_data),
&brw->clip.prog_offset, &brw->clip.prog_data);
ralloc_free(mem_ctx);
}
void
vec4_generator::generate_pull_constant_load_gen7(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg surf_index,
struct brw_reg offset)
{
assert(surf_index.type == BRW_REGISTER_TYPE_UD);
if (surf_index.file == BRW_IMMEDIATE_VALUE) {
brw_inst *insn = brw_next_insn(p, BRW_OPCODE_SEND);
brw_set_dest(p, insn, dst);
brw_set_src0(p, insn, offset);
brw_set_sampler_message(p, insn,
surf_index.dw1.ud,
0, /* LD message ignores sampler unit */
GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1, /* rlen */
1, /* mlen */
false, /* no header */
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
0);
brw_mark_surface_used(&prog_data->base, surf_index.dw1.ud);
} else {
struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_set_default_access_mode(p, BRW_ALIGN_1);
/* a0.0 = surf_index & 0xff */
brw_inst *insn_and = brw_next_insn(p, BRW_OPCODE_AND);
brw_inst_set_exec_size(p->brw, insn_and, BRW_EXECUTE_1);
brw_set_dest(p, insn_and, addr);
brw_set_src0(p, insn_and, vec1(retype(surf_index, BRW_REGISTER_TYPE_UD)));
brw_set_src1(p, insn_and, brw_imm_ud(0x0ff));
/* a0.0 |= <descriptor> */
brw_inst *insn_or = brw_next_insn(p, BRW_OPCODE_OR);
brw_set_sampler_message(p, insn_or,
0 /* surface */,
0 /* sampler */,
GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
1 /* rlen */,
1 /* mlen */,
false /* header */,
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
0);
brw_inst_set_exec_size(p->brw, insn_or, BRW_EXECUTE_1);
brw_inst_set_src1_reg_type(p->brw, insn_or, BRW_REGISTER_TYPE_UD);
brw_set_src0(p, insn_or, addr);
brw_set_dest(p, insn_or, addr);
/* dst = send(offset, a0.0) */
brw_inst *insn_send = brw_next_insn(p, BRW_OPCODE_SEND);
brw_set_dest(p, insn_send, dst);
brw_set_src0(p, insn_send, offset);
brw_set_indirect_send_descriptor(p, insn_send, BRW_SFID_SAMPLER, addr);
brw_pop_insn_state(p);
/* visitor knows more than we do about the surface limit required,
* so has already done marking.
*/
}
}
void
vec4_generator::generate_gs_set_channel_masks(struct brw_reg dst,
struct brw_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 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 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, BRW_REGISTER_TYPE_UB);
src = retype(src, BRW_REGISTER_TYPE_UB);
brw_push_insn_state(p);
brw_set_default_access_mode(p, BRW_ALIGN_1);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_OR(p, suboffset(vec1(dst), 21), vec1(src), suboffset(vec1(src), 16));
brw_pop_insn_state(p);
}
void
vec4_generator::generate_tex(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg src,
struct brw_reg sampler_index)
{
int msg_type = -1;
if (brw->gen >= 5) {
switch (inst->opcode) {
case SHADER_OPCODE_TEX:
case SHADER_OPCODE_TXL:
if (inst->shadow_compare) {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
} else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
}
break;
case SHADER_OPCODE_TXD:
if (inst->shadow_compare) {
/* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
assert(brw->gen >= 8 || brw->is_haswell);
msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE;
} else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
}
break;
case SHADER_OPCODE_TXF:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
break;
case SHADER_OPCODE_TXF_CMS:
if (brw->gen >= 7)
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
else
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
break;
case SHADER_OPCODE_TXF_MCS:
assert(brw->gen >= 7);
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS;
break;
case SHADER_OPCODE_TXS:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
break;
case SHADER_OPCODE_TG4:
if (inst->shadow_compare) {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C;
} else {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4;
}
break;
case SHADER_OPCODE_TG4_OFFSET:
if (inst->shadow_compare) {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C;
} else {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO;
}
break;
default:
unreachable("should not get here: invalid vec4 texture opcode");
}
} else {
switch (inst->opcode) {
case SHADER_OPCODE_TEX:
case SHADER_OPCODE_TXL:
if (inst->shadow_compare) {
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD_COMPARE;
assert(inst->mlen == 3);
} else {
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD;
assert(inst->mlen == 2);
}
break;
case SHADER_OPCODE_TXD:
/* There is no sample_d_c message; comparisons are done manually. */
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_GRADIENTS;
assert(inst->mlen == 4);
break;
case SHADER_OPCODE_TXF:
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_LD;
assert(inst->mlen == 2);
break;
case SHADER_OPCODE_TXS:
msg_type = BRW_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 == BRW_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_present) {
if (brw->gen < 6 && !inst->texture_offset) {
/* Set up an implied move from g0 to the MRF. */
src = brw_vec8_grf(0, 0);
} else {
struct brw_reg header =
retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD);
/* Explicitly set up the message header by copying g0 to the MRF. */
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_MOV(p, header, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
brw_set_default_access_mode(p, BRW_ALIGN_1);
if (inst->texture_offset) {
/* Set the texel offset bits in DWord 2. */
brw_MOV(p, get_element_ud(header, 2),
brw_imm_ud(inst->texture_offset));
}
brw_adjust_sampler_state_pointer(p, header, sampler_index, dst);
brw_pop_insn_state(p);
}
}
uint32_t return_format;
switch (dst.type) {
case BRW_REGISTER_TYPE_D:
return_format = BRW_SAMPLER_RETURN_FORMAT_SINT32;
break;
case BRW_REGISTER_TYPE_UD:
return_format = BRW_SAMPLER_RETURN_FORMAT_UINT32;
break;
default:
return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
break;
}
uint32_t base_binding_table_index = (inst->opcode == SHADER_OPCODE_TG4 ||
inst->opcode == SHADER_OPCODE_TG4_OFFSET)
? prog_data->base.binding_table.gather_texture_start
: prog_data->base.binding_table.texture_start;
if (sampler_index.file == BRW_IMMEDIATE_VALUE) {
uint32_t sampler = sampler_index.dw1.ud;
brw_SAMPLE(p,
dst,
inst->base_mrf,
src,
sampler + base_binding_table_index,
sampler % 16,
msg_type,
1, /* response length */
inst->mlen,
inst->header_present,
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
return_format);
brw_mark_surface_used(&prog_data->base, sampler + base_binding_table_index);
} else {
/* Non-constant sampler index. */
/* Note: this clobbers `dst` as a temporary before emitting the send */
struct brw_reg addr = vec1(retype(brw_address_reg(0), BRW_REGISTER_TYPE_UD));
struct brw_reg temp = vec1(retype(dst, BRW_REGISTER_TYPE_UD));
struct brw_reg sampler_reg = vec1(retype(sampler_index, BRW_REGISTER_TYPE_UD));
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_set_default_access_mode(p, BRW_ALIGN_1);
/* Some care required: `sampler` and `temp` may alias:
* addr = sampler & 0xff
* temp = (sampler << 8) & 0xf00
* addr = addr | temp
*/
brw_ADD(p, addr, sampler_reg, brw_imm_ud(base_binding_table_index));
brw_SHL(p, temp, sampler_reg, brw_imm_ud(8u));
brw_AND(p, temp, temp, brw_imm_ud(0x0f00));
brw_AND(p, addr, addr, brw_imm_ud(0x0ff));
brw_OR(p, addr, addr, temp);
/* a0.0 |= <descriptor> */
brw_inst *insn_or = brw_next_insn(p, BRW_OPCODE_OR);
brw_set_sampler_message(p, insn_or,
0 /* surface */,
0 /* sampler */,
msg_type,
1 /* rlen */,
inst->mlen /* mlen */,
inst->header_present /* header */,
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
return_format);
brw_inst_set_exec_size(p->brw, insn_or, BRW_EXECUTE_1);
brw_inst_set_src1_reg_type(p->brw, insn_or, BRW_REGISTER_TYPE_UD);
brw_set_src0(p, insn_or, addr);
brw_set_dest(p, insn_or, addr);
/* dst = send(offset, a0.0) */
brw_inst *insn_send = brw_next_insn(p, BRW_OPCODE_SEND);
brw_set_dest(p, insn_send, dst);
brw_set_src0(p, insn_send, src);
brw_set_indirect_send_descriptor(p, insn_send, BRW_SFID_SAMPLER, addr);
brw_pop_insn_state(p);
/* visitor knows more than we do about the surface limit required,
* so has already done marking.
*/
}
}
void
vec4_generator::generate_code(const cfg_t *cfg)
{
struct annotation_info annotation;
memset(&annotation, 0, sizeof(annotation));
foreach_block_and_inst (block, vec4_instruction, inst, cfg) {
struct brw_reg src[3], dst;
if (unlikely(debug_flag))
annotate(brw, &annotation, cfg, inst, p->next_insn_offset);
for (unsigned int i = 0; i < 3; i++) {
src[i] = inst->get_src(this->prog_data, i);
}
dst = inst->get_dst();
brw_set_default_predicate_control(p, inst->predicate);
brw_set_default_predicate_inverse(p, inst->predicate_inverse);
brw_set_default_saturate(p, inst->saturate);
brw_set_default_mask_control(p, inst->force_writemask_all);
brw_set_default_acc_write_control(p, inst->writes_accumulator);
unsigned pre_emit_nr_insn = p->nr_insn;
generate_vec4_instruction(inst, dst, src);
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");
brw_inst *last = &p->store[pre_emit_nr_insn];
brw_inst_set_cond_modifier(brw, last, inst->conditional_mod);
brw_inst_set_no_dd_clear(brw, last, inst->no_dd_clear);
brw_inst_set_no_dd_check(brw, last, inst->no_dd_check);
}
}
brw_set_uip_jip(p);
annotation_finalize(&annotation, p->next_insn_offset);
int before_size = p->next_insn_offset;
brw_compact_instructions(p, 0, annotation.ann_count, annotation.ann);
int after_size = p->next_insn_offset;
if (unlikely(debug_flag)) {
if (shader_prog) {
fprintf(stderr, "Native code for %s vertex shader %d:\n",
shader_prog->Label ? shader_prog->Label : "unnamed",
shader_prog->Name);
} else {
fprintf(stderr, "Native code for vertex program %d:\n", prog->Id);
}
fprintf(stderr, "vec4 shader: %d instructions. Compacted %d to %d"
" bytes (%.0f%%)\n",
before_size / 16, before_size, after_size,
100.0f * (before_size - after_size) / before_size);
dump_assembly(p->store, annotation.ann_count, annotation.ann, brw, prog);
ralloc_free(annotation.ann);
}
}
void
vec4_generator::generate_tex(vec4_instruction *inst,
struct brw_reg dst,
struct brw_reg src)
{
int msg_type = -1;
if (brw->gen >= 5) {
switch (inst->opcode) {
case SHADER_OPCODE_TEX:
case SHADER_OPCODE_TXL:
if (inst->shadow_compare) {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD_COMPARE;
} else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LOD;
}
break;
case SHADER_OPCODE_TXD:
if (inst->shadow_compare) {
/* Gen7.5+. Otherwise, lowered by brw_lower_texture_gradients(). */
assert(brw->gen >= 8 || brw->is_haswell);
msg_type = HSW_SAMPLER_MESSAGE_SAMPLE_DERIV_COMPARE;
} else {
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_DERIVS;
}
break;
case SHADER_OPCODE_TXF:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
break;
case SHADER_OPCODE_TXF_CMS:
if (brw->gen >= 7)
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD2DMS;
else
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_LD;
break;
case SHADER_OPCODE_TXF_MCS:
assert(brw->gen >= 7);
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_LD_MCS;
break;
case SHADER_OPCODE_TXS:
msg_type = GEN5_SAMPLER_MESSAGE_SAMPLE_RESINFO;
break;
case SHADER_OPCODE_TG4:
if (inst->shadow_compare) {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_C;
} else {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4;
}
break;
case SHADER_OPCODE_TG4_OFFSET:
if (inst->shadow_compare) {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO_C;
} else {
msg_type = GEN7_SAMPLER_MESSAGE_SAMPLE_GATHER4_PO;
}
break;
default:
unreachable("should not get here: invalid vec4 texture opcode");
}
} else {
switch (inst->opcode) {
case SHADER_OPCODE_TEX:
case SHADER_OPCODE_TXL:
if (inst->shadow_compare) {
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD_COMPARE;
assert(inst->mlen == 3);
} else {
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_LOD;
assert(inst->mlen == 2);
}
break;
case SHADER_OPCODE_TXD:
/* There is no sample_d_c message; comparisons are done manually. */
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_SAMPLE_GRADIENTS;
assert(inst->mlen == 4);
break;
case SHADER_OPCODE_TXF:
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_LD;
assert(inst->mlen == 2);
break;
case SHADER_OPCODE_TXS:
msg_type = BRW_SAMPLER_MESSAGE_SIMD4X2_RESINFO;
assert(inst->mlen == 2);
break;
default:
unreachable("should not get here: invalid vec4 texture opcode");
}
}
assert(msg_type != -1);
/* 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_present) {
if (brw->gen < 6 && !inst->texture_offset) {
/* Set up an implied move from g0 to the MRF. */
src = brw_vec8_grf(0, 0);
} else {
struct brw_reg header =
retype(brw_message_reg(inst->base_mrf), BRW_REGISTER_TYPE_UD);
/* Explicitly set up the message header by copying g0 to the MRF. */
brw_push_insn_state(p);
brw_set_default_mask_control(p, BRW_MASK_DISABLE);
brw_MOV(p, header, retype(brw_vec8_grf(0, 0), BRW_REGISTER_TYPE_UD));
brw_set_default_access_mode(p, BRW_ALIGN_1);
if (inst->texture_offset) {
/* Set the texel offset bits in DWord 2. */
brw_MOV(p, get_element_ud(header, 2),
brw_imm_ud(inst->texture_offset));
}
if (inst->sampler >= 16) {
/* The "Sampler Index" field can only store values between 0 and 15.
* However, we can add an offset to the "Sampler State Pointer"
* field, effectively selecting a different set of 16 samplers.
*
* The "Sampler State Pointer" needs to be aligned to a 32-byte
* offset, and each sampler state is only 16-bytes, so we can't
* exclusively use the offset - we have to use both.
*/
assert(brw->gen >= 8 || brw->is_haswell);
brw_ADD(p,
get_element_ud(header, 3),
get_element_ud(brw_vec8_grf(0, 0), 3),
brw_imm_ud(16 * (inst->sampler / 16) *
sizeof(gen7_sampler_state)));
}
brw_pop_insn_state(p);
}
}
uint32_t return_format;
switch (dst.type) {
case BRW_REGISTER_TYPE_D:
return_format = BRW_SAMPLER_RETURN_FORMAT_SINT32;
break;
case BRW_REGISTER_TYPE_UD:
return_format = BRW_SAMPLER_RETURN_FORMAT_UINT32;
break;
default:
return_format = BRW_SAMPLER_RETURN_FORMAT_FLOAT32;
break;
}
uint32_t surface_index = ((inst->opcode == SHADER_OPCODE_TG4 ||
inst->opcode == SHADER_OPCODE_TG4_OFFSET)
? prog_data->base.binding_table.gather_texture_start
: prog_data->base.binding_table.texture_start) + inst->sampler;
brw_SAMPLE(p,
dst,
inst->base_mrf,
src,
surface_index,
inst->sampler % 16,
msg_type,
1, /* response length */
inst->mlen,
inst->header_present,
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
return_format);
brw_mark_surface_used(&prog_data->base, surface_index);
}