nir_ssa_def *shift_ssa;
nir_ssa_def *new_offset = NULL;
- b->cursor = nir_after_instr(&alu->instr);
-
/* the first src could be something like ssa_18.x, but we only want
* the single component. Otherwise the ishl/ishr/ushr could turn
* into a vec4 operation:
unsigned ir3_ssbo_opcode, uint8_t offset_src_idx)
{
unsigned num_srcs = nir_intrinsic_infos[intrinsic->intrinsic].num_srcs;
+ int shift = 2;
bool has_dest = nir_intrinsic_infos[intrinsic->intrinsic].has_dest;
nir_ssa_def *new_dest = NULL;
+ /* for 16-bit ssbo access, offset is in 16-bit words instead of dwords */
+ if ((has_dest && intrinsic->dest.ssa.bit_size == 16) ||
+ (!has_dest && intrinsic->src[0].ssa->bit_size == 16))
+ shift = 1;
+
/* Here we create a new intrinsic and copy over all contents from the old one. */
nir_intrinsic_instr *new_intrinsic;
nir_src *target_src;
+ b->cursor = nir_before_instr(&intrinsic->instr);
+
/* 'offset_src_idx' holds the index of the source that represent the offset. */
new_intrinsic =
nir_intrinsic_instr_create(b->shader, ir3_ssbo_opcode);
* Here we use the convention that shifting right is negative while shifting
* left is positive. So 'x / 4' ~ 'x >> 2' or 'x << -2'.
*/
- nir_ssa_def *new_offset = ir3_nir_try_propagate_bit_shift(b, offset, -2);
+ nir_ssa_def *new_offset = ir3_nir_try_propagate_bit_shift(b, offset, -shift);
/* The new source that will hold the dword-offset is always the last
* one for every intrinsic.
for (unsigned i = 0; i < num_srcs; i++)
new_intrinsic->src[i] = nir_src_for_ssa(intrinsic->src[i].ssa);
- for (unsigned i = 0; i < NIR_INTRINSIC_MAX_CONST_INDEX; i++)
- new_intrinsic->const_index[i] = intrinsic->const_index[i];
+ nir_intrinsic_copy_const_indices(new_intrinsic, intrinsic);
new_intrinsic->num_components = intrinsic->num_components;
- b->cursor = nir_before_instr(&intrinsic->instr);
-
/* If we managed to propagate the division by 4, just use the new offset
* register and don't emit the SHR.
*/
if (new_offset)
offset = new_offset;
else
- offset = nir_ushr(b, offset, nir_imm_int(b, 2));
+ offset = nir_ushr(b, offset, nir_imm_int(b, shift));
/* Insert the new intrinsic right before the old one. */
nir_builder_instr_insert(b, &new_intrinsic->instr);
}
static bool
-lower_offset_for_ubo(nir_intrinsic_instr *intrinsic, nir_builder *b)
+lower_offset_for_ubo(nir_intrinsic_instr *intrinsic, nir_builder *b, int gpu_id)
{
- /* We only need to lower offset if using LDC. Currently, we only use LDC
- * in the bindless mode. Also, LDC is introduced on A6xx, but currently we
- * only use bindless in turnip which is A6xx only.
- *
- * TODO: We should be using LDC always on A6xx+.
+ /* We only need to lower offset if using LDC, which takes an offset in
+ * vec4 units and has the start component baked into the instruction.
*/
- if (!ir3_bindless_resource(intrinsic->src[0]))
+ if (gpu_id < 600)
return false;
/* TODO handle other bitsizes, including non-dword-aligned loads */
nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_ubo_ir3);
debug_assert(intrinsic->dest.is_ssa);
- new_intrinsic->src[0] = nir_src_for_ssa(intrinsic->src[0].ssa);
+ new_intrinsic->src[0] = intrinsic->src[0];
nir_ssa_def *offset = intrinsic->src[1].ssa;
nir_ssa_def *new_offset = ir3_nir_try_propagate_bit_shift(b, offset, -4);
}
static bool
-lower_io_offsets_block(nir_block *block, nir_builder *b, void *mem_ctx)
+lower_io_offsets_block(nir_block *block, nir_builder *b, void *mem_ctx, int gpu_id)
{
bool progress = false;
/* UBO */
if (intr->intrinsic == nir_intrinsic_load_ubo) {
- progress |= lower_offset_for_ubo(intr, b);
+ progress |= lower_offset_for_ubo(intr, b, gpu_id);
continue;
}
}
static bool
-lower_io_offsets_func(nir_function_impl *impl)
+lower_io_offsets_func(nir_function_impl *impl, int gpu_id)
{
void *mem_ctx = ralloc_parent(impl);
nir_builder b;
bool progress = false;
nir_foreach_block_safe (block, impl) {
- progress |= lower_io_offsets_block(block, &b, mem_ctx);
+ progress |= lower_io_offsets_block(block, &b, mem_ctx, gpu_id);
}
if (progress) {
}
bool
-ir3_nir_lower_io_offsets(nir_shader *shader)
+ir3_nir_lower_io_offsets(nir_shader *shader, int gpu_id)
{
bool progress = false;
nir_foreach_function (function, shader) {
if (function->impl)
- progress |= lower_io_offsets_func(function->impl);
+ progress |= lower_io_offsets_func(function->impl, gpu_id);
}
return progress;