OpenHarmony/third_party_mesa3d
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

intel/fs: switch compute push constant loads to LSC

Browse Source 
We're now able to load up to 8 GRFs in one send.

v2: Switch to use transpose + vector of up to 64 (Thanks Curro!)

v3: Increase parallelism by not reusing the same register for push
    constant offset (Curro)

v4: Drop dead ADD() instruction (Curro)

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Reviewed-by: Rohan Garg <rohan.garg@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17555>
This commit is contained in:
Lionel Landwerlin
2022-07-15 13:08:23 +03:00
committed by Marge Bot
parent 1e7a131fd1
commit aa65f83203
1 changed files with 29 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
src/intel/compiler/brw_fs.cpp
View File

@@ -1586,67 +1586,56 @@ fs_visitor::assign_curb_setup()
assert(devinfo->verx10 >= 125); assert(devinfo->verx10 >= 125);
assert(uniform_push_length <= 1); assert(uniform_push_length <= 1);
} else if (is_compute && devinfo->verx10 >= 125) { } else if (is_compute && devinfo->verx10 >= 125) {
fs_builder ubld = bld.exec_all().group(8, 0).at( assert(devinfo->has_lsc);
fs_builder ubld = bld.exec_all().group(1, 0).at(
cfg->first_block(), cfg->first_block()->start()); cfg->first_block(), cfg->first_block()->start());
/* The base address for our push data is passed in as R0.0[31:6]. We /* The base offset for our push data is passed in as R0.0[31:6]. We have
* have to mask off the bottom 6 bits. * to mask off the bottom 6 bits.
*/ */
fs_reg base_addr = ubld.vgrf(BRW_REGISTER_TYPE_UD); fs_reg base_addr = ubld.vgrf(BRW_REGISTER_TYPE_UD);
ubld.group(1, 0).AND(base_addr, ubld.AND(base_addr,
retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD), retype(brw_vec1_grf(0, 0), BRW_REGISTER_TYPE_UD),
brw_imm_ud(INTEL_MASK(31, 6))); brw_imm_ud(INTEL_MASK(31, 6)));
fs_reg header0 = ubld.vgrf(BRW_REGISTER_TYPE_UD);
ubld.MOV(header0, brw_imm_ud(0));
ubld.group(1, 0).SHR(component(header0, 2), base_addr, brw_imm_ud(4));
/* On Gfx12-HP we load constants at the start of the program using A32 /* On Gfx12-HP we load constants at the start of the program using A32
* stateless messages. * stateless messages.
*/ */
for (unsigned i = 0; i < uniform_push_length;) { for (unsigned i = 0; i < uniform_push_length;) {
/* Limit ourselves to HW limit of 8 Owords (8 * 16bytes = 128 bytes /* Limit ourselves to LSC HW limit of 8 GRFs (256bytes D32V64). */
* or 4 registers). unsigned num_regs = MIN2(uniform_push_length - i, 8);
*/
unsigned num_regs = MIN2(uniform_push_length - i, 4);
assert(num_regs > 0); assert(num_regs > 0);
num_regs = 1 << util_logbase2(num_regs); num_regs = 1 << util_logbase2(num_regs);
fs_reg header; fs_reg addr = ubld.vgrf(BRW_REGISTER_TYPE_UD);
if (i == 0) { ubld.ADD(addr, base_addr, brw_imm_ud(i * REG_SIZE));
header = header0;
} else {
header = ubld.vgrf(BRW_REGISTER_TYPE_UD);
ubld.MOV(header, brw_imm_ud(0));
ubld.group(1, 0).ADD(component(header, 2),
component(header0, 2),
brw_imm_ud(i * 2));
}
fs_reg srcs[4] = { fs_reg srcs[4] = {
brw_imm_ud(0), /* desc */ brw_imm_ud(0), /* desc */
brw_imm_ud(0), /* ex_desc */ brw_imm_ud(0), /* ex_desc */
header, /* payload */ addr, /* payload */
fs_reg(), /* payload2 */ fs_reg(), /* payload2 */
}; };
fs_reg dest = retype(brw_vec8_grf(payload.num_regs + i, 0), fs_reg dest = retype(brw_vec8_grf(payload.num_regs + i, 0),
BRW_REGISTER_TYPE_UD); BRW_REGISTER_TYPE_UD);
fs_inst *send = ubld.emit(SHADER_OPCODE_SEND, dest, srcs, 4);
/* This instruction has to be run SIMD16 if we're filling more than a send->sfid = GFX12_SFID_UGM;
* single register. send->desc = lsc_msg_desc(devinfo, LSC_OP_LOAD,
*/ 1 /* exec_size */,
unsigned send_width = MIN2(16, num_regs * 8); LSC_ADDR_SURFTYPE_FLAT,
LSC_ADDR_SIZE_A32,
fs_inst *send = ubld.group(send_width, 0).emit(SHADER_OPCODE_SEND, 1 /* num_coordinates */,
dest, srcs, 4); LSC_DATA_SIZE_D32,
send->sfid = GFX7_SFID_DATAPORT_DATA_CACHE; num_regs * 8 /* num_channels */,
send->desc = brw_dp_desc(devinfo, GFX8_BTI_STATELESS_NON_COHERENT, true /* transpose */,
GFX7_DATAPORT_DC_OWORD_BLOCK_READ, LSC_CACHE_LOAD_L1STATE_L3MOCS,
BRW_DATAPORT_OWORD_BLOCK_OWORDS(num_regs * 2)); true /* has_dest */);
send->header_size = 1; send->header_size = 0;
send->mlen = 1; send->mlen = lsc_msg_desc_src0_len(devinfo, send->desc);
send->size_written = num_regs * REG_SIZE; send->size_written =
lsc_msg_desc_dest_len(devinfo, send->desc) * REG_SIZE;
send->send_is_volatile = true; send->send_is_volatile = true;
i += num_regs; i += num_regs;