struct ir3_instruction **dst)
{
struct ir3_block *b = ctx->block;
- struct ir3_instruction *ldgb, *src0, *src1, *offset;
- nir_const_value *const_offset;
+ struct ir3_instruction *ldgb, *src0, *src1, *byte_offset, *offset;
- /* can this be non-const buffer_index? how do we handle that? */
- const_offset = nir_src_as_const_value(intr->src[0]);
- compile_assert(ctx, const_offset);
+ struct ir3_instruction *ssbo = ir3_ssbo_to_ibo(ctx, intr->src[0]);
- offset = ir3_get_src(ctx, &intr->src[1])[0];
+ byte_offset = ir3_get_src(ctx, &intr->src[1])[0];
+ offset = ir3_get_src(ctx, &intr->src[2])[0];
/* src0 is uvec2(offset*4, 0), src1 is offset.. nir already *= 4: */
src0 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
- offset,
+ byte_offset,
create_immed(b, 0),
}, 2);
- src1 = ir3_SHR_B(b, offset, 0, create_immed(b, 2), 0);
+ src1 = offset;
- ldgb = ir3_LDGB(b, create_immed(b, const_offset->u32[0]), 0,
+ ldgb = ir3_LDGB(b, ssbo, 0,
src0, 0, src1, 0);
ldgb->regs[0]->wrmask = MASK(intr->num_components);
ldgb->cat6.iim_val = intr->num_components;
emit_intrinsic_store_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
struct ir3_block *b = ctx->block;
- struct ir3_instruction *stgb, *src0, *src1, *src2, *offset;
- nir_const_value *const_offset;
- /* TODO handle wrmask properly, see _store_shared().. but I think
- * it is more a PITA than that, since blob ends up loading the
- * masked components and writing them back out.
- */
- unsigned wrmask = intr->const_index[0];
+ struct ir3_instruction *stgb, *src0, *src1, *src2, *byte_offset, *offset;
+ unsigned wrmask = nir_intrinsic_write_mask(intr);
unsigned ncomp = ffs(~wrmask) - 1;
- /* can this be non-const buffer_index? how do we handle that? */
- const_offset = nir_src_as_const_value(intr->src[1]);
- compile_assert(ctx, const_offset);
+ assert(wrmask == BITFIELD_MASK(intr->num_components));
- offset = ir3_get_src(ctx, &intr->src[2])[0];
+ struct ir3_instruction *ssbo = ir3_ssbo_to_ibo(ctx, intr->src[1]);
+
+ byte_offset = ir3_get_src(ctx, &intr->src[2])[0];
+ offset = ir3_get_src(ctx, &intr->src[3])[0];
/* src0 is value, src1 is offset, src2 is uvec2(offset*4, 0)..
* nir already *= 4:
*/
src0 = ir3_create_collect(ctx, ir3_get_src(ctx, &intr->src[0]), ncomp);
- src1 = ir3_SHR_B(b, offset, 0, create_immed(b, 2), 0);
+ src1 = offset;
src2 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
- offset,
+ byte_offset,
create_immed(b, 0),
}, 2);
- stgb = ir3_STGB(b, create_immed(b, const_offset->u32[0]), 0,
- src0, 0, src1, 0, src2, 0);
+ stgb = ir3_STGB(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
stgb->cat6.iim_val = ncomp;
stgb->cat6.d = 4;
stgb->cat6.type = TYPE_U32;
emit_intrinsic_atomic_ssbo(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
struct ir3_block *b = ctx->block;
- struct ir3_instruction *atomic, *ssbo, *src0, *src1, *src2, *offset;
- nir_const_value *const_offset;
+ struct ir3_instruction *atomic, *ssbo, *src0, *src1, *src2, *byte_offset,
+ *offset;
type_t type = TYPE_U32;
- /* can this be non-const buffer_index? how do we handle that? */
- const_offset = nir_src_as_const_value(intr->src[0]);
- compile_assert(ctx, const_offset);
- ssbo = create_immed(b, const_offset->u32[0]);
+ ssbo = ir3_ssbo_to_ibo(ctx, intr->src[0]);
- offset = ir3_get_src(ctx, &intr->src[1])[0];
+ byte_offset = ir3_get_src(ctx, &intr->src[1])[0];
+ offset = ir3_get_src(ctx, &intr->src[3])[0];
/* src0 is data (or uvec2(data, compare))
* src1 is offset
* Note that nir already multiplies the offset by four
*/
src0 = ir3_get_src(ctx, &intr->src[2])[0];
- src1 = ir3_SHR_B(b, offset, 0, create_immed(b, 2), 0);
+ src1 = offset;
src2 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
- offset,
+ byte_offset,
create_immed(b, 0),
}, 2);
switch (intr->intrinsic) {
- case nir_intrinsic_ssbo_atomic_add:
+ case nir_intrinsic_ssbo_atomic_add_ir3:
atomic = ir3_ATOMIC_ADD_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_ssbo_atomic_imin:
+ case nir_intrinsic_ssbo_atomic_imin_ir3:
atomic = ir3_ATOMIC_MIN_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
type = TYPE_S32;
break;
- case nir_intrinsic_ssbo_atomic_umin:
+ case nir_intrinsic_ssbo_atomic_umin_ir3:
atomic = ir3_ATOMIC_MIN_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_ssbo_atomic_imax:
+ case nir_intrinsic_ssbo_atomic_imax_ir3:
atomic = ir3_ATOMIC_MAX_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
type = TYPE_S32;
break;
- case nir_intrinsic_ssbo_atomic_umax:
+ case nir_intrinsic_ssbo_atomic_umax_ir3:
atomic = ir3_ATOMIC_MAX_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_ssbo_atomic_and:
+ case nir_intrinsic_ssbo_atomic_and_ir3:
atomic = ir3_ATOMIC_AND_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_ssbo_atomic_or:
+ case nir_intrinsic_ssbo_atomic_or_ir3:
atomic = ir3_ATOMIC_OR_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_ssbo_atomic_xor:
+ case nir_intrinsic_ssbo_atomic_xor_ir3:
atomic = ir3_ATOMIC_XOR_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_ssbo_atomic_exchange:
+ case nir_intrinsic_ssbo_atomic_exchange_ir3:
atomic = ir3_ATOMIC_XCHG_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_ssbo_atomic_comp_swap:
+ case nir_intrinsic_ssbo_atomic_comp_swap_ir3:
/* for cmpxchg, src0 is [ui]vec2(data, compare): */
src0 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
ir3_get_src(ctx, &intr->src[3])[0],
src0,
}, 2);
+ src1 = ir3_get_src(ctx, &intr->src[4])[0];
atomic = ir3_ATOMIC_CMPXCHG_G(b, ssbo, 0, src0, 0, src1, 0, src2, 0);
break;
default:
}
static struct ir3_instruction *
-get_image_offset(struct ir3_context *ctx, const nir_variable *var,
+get_image_offset(struct ir3_context *ctx, const nir_intrinsic_instr *instr,
struct ir3_instruction * const *coords, bool byteoff)
{
struct ir3_block *b = ctx->block;
struct ir3_instruction *offset;
- unsigned ncoords = ir3_get_image_coords(var, NULL);
+ unsigned index = nir_src_as_uint(instr->src[0]);
+ unsigned ncoords = ir3_get_image_coords(instr, NULL);
/* to calculate the byte offset (yes, uggg) we need (up to) three
* const values to know the bytes per pixel, and y and z stride:
*/
- unsigned cb = regid(ctx->so->constbase.image_dims, 0) +
- ctx->so->const_layout.image_dims.off[var->data.driver_location];
+ struct ir3_const_state *const_state = &ctx->so->shader->const_state;
+ unsigned cb = regid(const_state->offsets.image_dims, 0) +
+ const_state->image_dims.off[index];
- debug_assert(ctx->so->const_layout.image_dims.mask &
- (1 << var->data.driver_location));
+ debug_assert(const_state->image_dims.mask & (1 << index));
/* offset = coords.x * bytes_per_pixel: */
- offset = ir3_MUL_S(b, coords[0], 0, create_uniform(b, cb + 0), 0);
+ offset = ir3_MUL_S24(b, coords[0], 0, create_uniform(b, cb + 0), 0);
if (ncoords > 1) {
/* offset += coords.y * y_pitch: */
offset = ir3_MAD_S24(b, create_uniform(b, cb + 1), 0,
}, 2);
}
-/* src[] = { deref, coord, sample_index, value }. const_index[] = {} */
+/* src[] = { index, coord, sample_index, value }. const_index[] = {} */
static void
emit_intrinsic_store_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
struct ir3_block *b = ctx->block;
- const nir_variable *var = nir_intrinsic_get_var(intr, 0);
struct ir3_instruction *stib, *offset;
struct ir3_instruction * const *value = ir3_get_src(ctx, &intr->src[3]);
struct ir3_instruction * const *coords = ir3_get_src(ctx, &intr->src[1]);
- unsigned ncoords = ir3_get_image_coords(var, NULL);
- unsigned tex_idx = ir3_get_image_slot(ctx, nir_src_as_deref(intr->src[0]));
- unsigned ncomp = ir3_get_num_components_for_glformat(var->data.image.format);
+ struct ir3_instruction * ibo = ir3_image_to_ibo(ctx, intr->src[0]);
+ unsigned ncoords = ir3_get_image_coords(intr, NULL);
+ unsigned ncomp = ir3_get_num_components_for_image_format(nir_intrinsic_format(intr));
/* src0 is value
* src1 is coords
* src2 is 64b byte offset
*/
- offset = get_image_offset(ctx, var, coords, true);
+ offset = get_image_offset(ctx, intr, coords, true);
/* NOTE: stib seems to take byte offset, but stgb.typed can be used
* too and takes a dword offset.. not quite sure yet why blob uses
* one over the other in various cases.
*/
- stib = ir3_STIB(b, create_immed(b, tex_idx), 0,
+ stib = ir3_STIB(b, ibo, 0,
ir3_create_collect(ctx, value, ncomp), 0,
ir3_create_collect(ctx, coords, ncoords), 0,
offset, 0);
stib->cat6.iim_val = ncomp;
stib->cat6.d = ncoords;
- stib->cat6.type = ir3_get_image_type(var);
+ stib->cat6.type = ir3_get_type_for_image_intrinsic(intr);
stib->cat6.typed = true;
stib->barrier_class = IR3_BARRIER_IMAGE_W;
stib->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;
emit_intrinsic_atomic_image(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
struct ir3_block *b = ctx->block;
- const nir_variable *var = nir_intrinsic_get_var(intr, 0);
- struct ir3_instruction *atomic, *image, *src0, *src1, *src2;
+ struct ir3_instruction *atomic, *src0, *src1, *src2;
struct ir3_instruction * const *coords = ir3_get_src(ctx, &intr->src[1]);
- unsigned ncoords = ir3_get_image_coords(var, NULL);
-
- image = create_immed(b, ir3_get_image_slot(ctx, nir_src_as_deref(intr->src[0])));
+ struct ir3_instruction * image = ir3_image_to_ibo(ctx, intr->src[0]);
+ unsigned ncoords = ir3_get_image_coords(intr, NULL);
/* src0 is value (or uvec2(value, compare))
* src1 is coords
*/
src0 = ir3_get_src(ctx, &intr->src[3])[0];
src1 = ir3_create_collect(ctx, coords, ncoords);
- src2 = get_image_offset(ctx, var, coords, false);
+ src2 = get_image_offset(ctx, intr, coords, false);
switch (intr->intrinsic) {
- case nir_intrinsic_image_deref_atomic_add:
+ case nir_intrinsic_image_atomic_add:
atomic = ir3_ATOMIC_ADD_G(b, image, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_image_deref_atomic_min:
+ case nir_intrinsic_image_atomic_imin:
+ case nir_intrinsic_image_atomic_umin:
atomic = ir3_ATOMIC_MIN_G(b, image, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_image_deref_atomic_max:
+ case nir_intrinsic_image_atomic_imax:
+ case nir_intrinsic_image_atomic_umax:
atomic = ir3_ATOMIC_MAX_G(b, image, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_image_deref_atomic_and:
+ case nir_intrinsic_image_atomic_and:
atomic = ir3_ATOMIC_AND_G(b, image, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_image_deref_atomic_or:
+ case nir_intrinsic_image_atomic_or:
atomic = ir3_ATOMIC_OR_G(b, image, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_image_deref_atomic_xor:
+ case nir_intrinsic_image_atomic_xor:
atomic = ir3_ATOMIC_XOR_G(b, image, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_image_deref_atomic_exchange:
+ case nir_intrinsic_image_atomic_exchange:
atomic = ir3_ATOMIC_XCHG_G(b, image, 0, src0, 0, src1, 0, src2, 0);
break;
- case nir_intrinsic_image_deref_atomic_comp_swap:
+ case nir_intrinsic_image_atomic_comp_swap:
/* for cmpxchg, src0 is [ui]vec2(data, compare): */
src0 = ir3_create_collect(ctx, (struct ir3_instruction*[]){
ir3_get_src(ctx, &intr->src[4])[0],
atomic->cat6.iim_val = 1;
atomic->cat6.d = ncoords;
- atomic->cat6.type = ir3_get_image_type(var);
+ atomic->cat6.type = ir3_get_type_for_image_intrinsic(intr);
atomic->cat6.typed = true;
atomic->barrier_class = IR3_BARRIER_IMAGE_W;
atomic->barrier_conflict = IR3_BARRIER_IMAGE_R | IR3_BARRIER_IMAGE_W;
.emit_intrinsic_atomic_ssbo = emit_intrinsic_atomic_ssbo,
.emit_intrinsic_store_image = emit_intrinsic_store_image,
.emit_intrinsic_atomic_image = emit_intrinsic_atomic_image,
+ .emit_intrinsic_image_size = emit_intrinsic_image_size_tex,
};
projects / mesa.git / blobdiff