AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

aco: refactor visit_store_scratch() to use new helpers

Browse Source 
Should support 8/16-bit stores now

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4639>
This commit is contained in:
Rhys Perry
2020-04-16 19:27:13 +01:00
committed by Marge Bot
parent 18817041f7
commit 936b70c8cf
1 changed files with 15 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/amd/compiler/aco_instruction_selection.cpp
+15 -32
View File
@@ -6798,46 +6798,29 @@ void visit_load_scratch(isel_context *ctx, nir_intrinsic_instr *instr) {
}
void visit_store_scratch(isel_context *ctx, nir_intrinsic_instr *instr) {
assert(instr->src[0].ssa->bit_size == 32 || instr->src[0].ssa->bit_size == 64);
Builder bld(ctx->program, ctx->block);
Temp rsrc = get_scratch_resource(ctx);
Temp data = as_vgpr(ctx, get_ssa_temp(ctx, instr->src[0].ssa));
Temp offset = as_vgpr(ctx, get_ssa_temp(ctx, instr->src[1].ssa));
unsigned elem_size_bytes = instr->src[0].ssa->bit_size / 8;
unsigned writemask = nir_intrinsic_write_mask(instr);
unsigned writemask = widen_mask(nir_intrinsic_write_mask(instr), elem_size_bytes);
while (writemask) {
int start, count;
u_bit_scan_consecutive_range(&writemask, &start, &count);
int num_bytes = count * elem_size_bytes;
if (num_bytes > 16) {
assert(elem_size_bytes == 8);
writemask |= (((count - 2) << 1) - 1) << (start + 2);
count = 2;
num_bytes = 16;
}
// TODO: check alignment of sub-dword stores
// TODO: split 3 bytes. there is no store instruction for that
Temp write_data;
if (count != instr->num_components) {
aco_ptr<Pseudo_instruction> vec{create_instruction<Pseudo_instruction>(aco_opcode::p_create_vector, Format::PSEUDO, count, 1)};
for (int i = 0; i < count; i++) {
Temp elem = emit_extract_vector(ctx, data, start + i, RegClass(RegType::vgpr, elem_size_bytes / 4));
vec->operands[i] = Operand(elem);
}
write_data = bld.tmp(RegClass(RegType::vgpr, count * elem_size_bytes / 4));
vec->definitions[0] = Definition(write_data);
ctx->block->instructions.emplace_back(std::move(vec));
} else {
write_data = data;
}
unsigned write_count = 0;
Temp write_datas[32];
unsigned offsets[32];
split_buffer_store(ctx, instr, false, RegType::vgpr, data, writemask,
16, &write_count, write_datas, offsets);
for (unsigned i = 0; i < write_count; i++) {
aco_opcode op;
switch (num_bytes) {
switch (write_datas[i].bytes()) {
case 1:
op = aco_opcode::buffer_store_byte;
break;
case 2:
op = aco_opcode::buffer_store_short;
break;
case 4:
op = aco_opcode::buffer_store_dword;
break;
@@ -6854,7 +6837,7 @@ void visit_store_scratch(isel_context *ctx, nir_intrinsic_instr *instr) {
unreachable("Invalid data size for nir_intrinsic_store_scratch.");
}
bld.mubuf(op, rsrc, offset, ctx->program->scratch_offset, write_data, start * elem_size_bytes, true);
bld.mubuf(op, rsrc, offset, ctx->program->scratch_offset, write_datas[i], offsets[i], true);
}
}