static void
gs_write_so(struct gs_compile_context *gcc,
struct toy_dst dst,
struct toy_src index, struct toy_src out,
bool send_write_commit_message,
int binding_table_index)
{
struct toy_compiler *tc = &gcc->tc;
struct toy_dst mrf_header;
struct toy_src desc;
mrf_header = tdst_d(tdst(TOY_FILE_MRF, gcc->first_free_mrf, 0));
/* m0.5: destination index */
gs_COPY1(tc, mrf_header, 5, index, 0);
/* m0.0 - m0.3: RGBA */
gs_COPY4(tc, mrf_header, 0, tsrc_type(out, mrf_header.type), 0);
desc = tsrc_imm_mdesc_data_port(tc, false,
1, send_write_commit_message,
true, send_write_commit_message,
GEN6_MSG_DP_SVB_WRITE, 0,
binding_table_index);
tc_SEND(tc, dst, tsrc_from(mrf_header), desc,
GEN6_SFID_DP_RC);
}
static void
vs_lower_opcode_tgsi_const_gen7(struct vs_compile_context *vcc,
struct toy_dst dst, int dim,
struct toy_src idx)
{
struct toy_compiler *tc = &vcc->tc;
const struct toy_dst offset =
tdst_ud(tdst(TOY_FILE_MRF, vcc->first_free_mrf, 0));
struct toy_src desc;
if (vs_lower_opcode_tgsi_const_pcb(vcc, dst, dim, idx))
return;
/*
* In 259b65e2e7938de4aab323033cfe2b33369ddb07, pull constant load was
* changed from OWord Dual Block Read to ld to increase performance in the
* classic driver. Since we use the constant cache instead of the data
* cache, I wonder if we still want to follow the classic driver.
*/
/* set offset */
tc_MOV(tc, offset, idx);
desc = tsrc_imm_mdesc_sampler(tc, 1, 1, false,
GEN6_MSG_SAMPLER_SIMD4X2,
GEN6_MSG_SAMPLER_LD,
0,
vcc->shader->bt.const_base + dim);
tc_SEND(tc, dst, tsrc_from(offset), desc, GEN6_SFID_SAMPLER);
}
static void
vs_lower_opcode_tgsi_const_gen6(struct vs_compile_context *vcc,
struct toy_dst dst, int dim,
struct toy_src idx)
{
const struct toy_dst header =
tdst_ud(tdst(TOY_FILE_MRF, vcc->first_free_mrf, 0));
const struct toy_dst block_offsets =
tdst_ud(tdst(TOY_FILE_MRF, vcc->first_free_mrf + 1, 0));
const struct toy_src r0 = tsrc_ud(tsrc(TOY_FILE_GRF, 0, 0));
struct toy_compiler *tc = &vcc->tc;
unsigned msg_type, msg_ctrl, msg_len;
struct toy_inst *inst;
struct toy_src desc;
if (vs_lower_opcode_tgsi_const_pcb(vcc, dst, dim, idx))
return;
/* set message header */
inst = tc_MOV(tc, header, r0);
inst->mask_ctrl = GEN6_MASKCTRL_NOMASK;
/* set block offsets */
tc_MOV(tc, block_offsets, idx);
msg_type = GEN6_MSG_DP_OWORD_DUAL_BLOCK_READ;
msg_ctrl = GEN6_MSG_DP_OWORD_DUAL_BLOCK_SIZE_1;
msg_len = 2;
desc = tsrc_imm_mdesc_data_port(tc, false, msg_len, 1, true, false,
msg_type, msg_ctrl, vcc->shader->bt.const_base + dim);
tc_SEND(tc, dst, tsrc_from(header), desc, vcc->const_cache);
}
static void
gs_ff_sync(struct gs_compile_context *gcc, struct toy_dst dst,
struct toy_src num_prims)
{
struct toy_compiler *tc = &gcc->tc;
struct toy_dst mrf_header =
tdst_d(tdst(TOY_FILE_MRF, gcc->first_free_mrf, 0));
struct toy_src desc;
bool allocate;
gs_COPY8(tc, mrf_header, gcc->payload.header);
/* set NumSOVertsToWrite and NumSOPrimsNeeded */
if (gcc->write_so) {
if (num_prims.file == TOY_FILE_IMM) {
const uint32_t v =
(num_prims.val32 * gcc->in_vue_count) << 16 | num_prims.val32;
gs_COPY1(tc, mrf_header, 0, tsrc_imm_d(v), 0);
}
else {
struct toy_dst m0_0 = tdst_d(gcc->vars.tmp);
tc_MUL(tc, m0_0, num_prims, tsrc_imm_d(gcc->in_vue_count << 16));
tc_OR(tc, m0_0, tsrc_from(m0_0), num_prims);
gs_COPY1(tc, mrf_header, 0, tsrc_from(m0_0), 0);
}
}
/* set NumGSPrimsGenerated */
if (gcc->write_vue)
gs_COPY1(tc, mrf_header, 1, num_prims, 0);
/*
* From the Sandy Bridge PRM, volume 2 part 1, page 173:
*
* "Programming Note: If the GS stage is enabled, software must always
* allocate at least one GS URB Entry. This is true even if the GS
* thread never needs to output vertices to the pipeline, e.g., when
* only performing stream output. This is an artifact of the need to
* pass the GS thread an initial destination URB handle."
*/
allocate = true;
desc = tsrc_imm_mdesc_urb(tc, false, 1, 1,
false, false, allocate,
false, 0, 1);
tc_SEND(tc, dst, tsrc_from(mrf_header), desc, GEN6_SFID_URB);
}
static void
fs_lower_opcode_tgsi_const_gen6(struct fs_compile_context *fcc,
struct toy_dst dst, int dim, struct toy_src idx)
{
const struct toy_dst header =
tdst_ud(tdst(TOY_FILE_MRF, fcc->first_free_mrf, 0));
const struct toy_dst global_offset =
tdst_ud(tdst(TOY_FILE_MRF, fcc->first_free_mrf, 2 * 4));
const struct toy_src r0 = tsrc_ud(tsrc(TOY_FILE_GRF, 0, 0));
struct toy_compiler *tc = &fcc->tc;
unsigned msg_type, msg_ctrl, msg_len;
struct toy_inst *inst;
struct toy_src desc;
struct toy_dst tmp, real_dst[4];
int i;
/* set message header */
inst = tc_MOV(tc, header, r0);
inst->mask_ctrl = BRW_MASK_DISABLE;
/* set global offset */
inst = tc_MOV(tc, global_offset, idx);
inst->mask_ctrl = BRW_MASK_DISABLE;
inst->exec_size = BRW_EXECUTE_1;
inst->src[0].rect = TOY_RECT_010;
msg_type = BRW_DATAPORT_READ_MESSAGE_OWORD_BLOCK_READ;
msg_ctrl = BRW_DATAPORT_OWORD_BLOCK_1_OWORDLOW << 8;
msg_len = 1;
desc = tsrc_imm_mdesc_data_port(tc, false, msg_len, 1, true, false,
msg_type, msg_ctrl, ILO_WM_CONST_SURFACE(dim));
tmp = tc_alloc_tmp(tc);
tc_SEND(tc, tmp, tsrc_from(header), desc, fcc->const_cache);
tdst_transpose(dst, real_dst);
for (i = 0; i < 4; i++) {
const struct toy_src src =
tsrc_offset(tsrc_rect(tsrc_from(tmp), TOY_RECT_010), 0, i);
/* cast to type D to make sure these are raw moves */
tc_MOV(tc, tdst_d(real_dst[i]), tsrc_d(src));
}
}
static void
fs_lower_opcode_tgsi_const_gen7(struct fs_compile_context *fcc,
struct toy_dst dst, int dim, struct toy_src idx)
{
struct toy_compiler *tc = &fcc->tc;
const struct toy_dst offset =
tdst_ud(tdst(TOY_FILE_MRF, fcc->first_free_mrf, 0));
struct toy_src desc;
struct toy_inst *inst;
struct toy_dst tmp, real_dst[4];
int i;
/*
* In 4c1fdae0a01b3f92ec03b61aac1d3df500d51fc6, pull constant load was
* changed from OWord Block Read to ld to increase performance in the
* classic driver. Since we use the constant cache instead of the data
* cache, I wonder if we still want to follow the classic driver.
*/
/* set offset */
inst = tc_MOV(tc, offset, tsrc_rect(idx, TOY_RECT_010));
inst->exec_size = BRW_EXECUTE_8;
inst->mask_ctrl = BRW_MASK_DISABLE;
desc = tsrc_imm_mdesc_sampler(tc, 1, 1, false,
BRW_SAMPLER_SIMD_MODE_SIMD4X2,
GEN5_SAMPLER_MESSAGE_SAMPLE_LD,
0,
ILO_WM_CONST_SURFACE(dim));
tmp = tc_alloc_tmp(tc);
inst = tc_SEND(tc, tmp, tsrc_from(offset), desc, BRW_SFID_SAMPLER);
inst->exec_size = BRW_EXECUTE_8;
inst->mask_ctrl = BRW_MASK_DISABLE;
tdst_transpose(dst, real_dst);
for (i = 0; i < 4; i++) {
const struct toy_src src =
tsrc_offset(tsrc_rect(tsrc_from(tmp), TOY_RECT_010), 0, i);
/* cast to type D to make sure these are raw moves */
tc_MOV(tc, tdst_d(real_dst[i]), tsrc_d(src));
}
}
static void
cs_dummy(struct cs_compile_context *ccc)
{
struct toy_compiler *tc = &ccc->tc;
struct toy_dst header;
struct toy_src r0, desc;
struct toy_inst *inst;
header = tdst_ud(tdst(TOY_FILE_MRF, ccc->first_free_mrf, 0));
r0 = tsrc_ud(tsrc(TOY_FILE_GRF, 0, 0));
inst = tc_MOV(tc, header, r0);
inst->exec_size = GEN6_EXECSIZE_8;
inst->mask_ctrl = GEN6_MASKCTRL_NOMASK;
desc = tsrc_imm_mdesc(tc, true, 1, 0, true,
GEN6_MSG_TS_RESOURCE_SELECT_NO_DEREF |
GEN6_MSG_TS_REQUESTER_TYPE_ROOT |
GEN6_MSG_TS_OPCODE_DEREF);
tc_SEND(tc, tdst_null(), tsrc_from(header), desc, GEN6_SFID_SPAWNER);
}