#include "ir3.h"
#include "ir3_context.h"
+void
+ir3_handle_bindless_cat6(struct ir3_instruction *instr, nir_src rsrc)
+{
+ nir_intrinsic_instr *intrin = ir3_bindless_resource(rsrc);
+ if (!intrin)
+ return;
+
+ instr->flags |= IR3_INSTR_B;
+ instr->cat6.base = nir_intrinsic_desc_set(intrin);
+}
static struct ir3_instruction *
create_indirect_load(struct ir3_context *ctx, unsigned arrsz, int n,
instr->cat6.type = TYPE_U32;
instr->cat6.iim_val = 1;
} else {
- instr = ir3_BARY_F(block, inloc, 0, ctx->ij_pixel, 0);
+ instr = ir3_BARY_F(block, inloc, 0, ctx->ij[IJ_PERSP_PIXEL], 0);
instr->regs[2]->wrmask = 0x3;
}
{
/* first four vec4 sysval's reserved for UBOs: */
/* NOTE: dp is in scalar, but there can be >4 dp components: */
- struct ir3_const_state *const_state = &ctx->so->shader->const_state;
+ struct ir3_const_state *const_state = ir3_const_state(ctx->so);
unsigned n = const_state->offsets.driver_param;
unsigned r = regid(n + dp / 4, dp % 4);
return create_uniform(ctx->block, r);
* src instruction and create a mov. This is easier for cp
* to eliminate.
*
+ * NOTE: a3xx definitely seen not working with flat bary.f. Same test
+ * uses ldlv on a4xx+, so not definitive. Seems rare enough to apply
+ * everywhere.
+ *
* TODO probably opc_cat==4 is ok too
*/
if (alu->src[0].src.is_ssa &&
+ src[0]->opc != OPC_BARY_F &&
(list_length(&alu->src[0].src.ssa->uses) == 1) &&
((opc_cat(src[0]->opc) == 2) || (opc_cat(src[0]->opc) == 3))) {
src[0]->flags |= IR3_INSTR_SAT;
dst[0]->cat5.type = TYPE_F32;
break;
case nir_op_fddx_fine:
- dst[0] = ir3_DSXPP_1(b, src[0], 0);
+ dst[0] = ir3_DSXPP_MACRO(b, src[0], 0);
dst[0]->cat5.type = TYPE_F32;
break;
case nir_op_fddy:
break;
break;
case nir_op_fddy_fine:
- dst[0] = ir3_DSYPP_1(b, src[0], 0);
+ dst[0] = ir3_DSYPP_MACRO(b, src[0], 0);
dst[0]->cat5.type = TYPE_F32;
break;
case nir_op_flt:
dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
dst[0]->cat2.condition = IR3_COND_EQ;
break;
- case nir_op_fne:
+ case nir_op_fneu:
dst[0] = ir3_CMPS_F(b, src[0], 0, src[1], 0);
dst[0]->cat2.condition = IR3_COND_NE;
break;
ldc->cat6.d = nir_intrinsic_base(intr);
ldc->cat6.type = TYPE_U32;
- nir_intrinsic_instr *bindless = ir3_bindless_resource(intr->src[0]);
- if (bindless) {
- ldc->flags |= IR3_INSTR_B;
- ldc->cat6.base = nir_intrinsic_desc_set(bindless);
+ ir3_handle_bindless_cat6(ldc, intr->src[0]);
+ if (ldc->flags & IR3_INSTR_B)
ctx->so->bindless_ubo = true;
- }
ir3_split_dest(b, dst, ldc, 0, ncomp);
}
{
struct ir3_block *b = ctx->block;
struct ir3_instruction *base_lo, *base_hi, *addr, *src0, *src1;
- struct ir3_const_state *const_state = &ctx->so->shader->const_state;
+ const struct ir3_const_state *const_state = ir3_const_state(ctx->so);
unsigned ubo = regid(const_state->offsets.ubo, 0);
const unsigned ptrsz = ir3_pointer_size(ctx->compiler);
base_lo = create_uniform(b, ubo + (src0->regs[1]->iim_val * ptrsz));
base_hi = create_uniform(b, ubo + (src0->regs[1]->iim_val * ptrsz) + 1);
} else {
- base_lo = create_uniform_indirect(b, ubo, ir3_get_addr0(ctx, src0, ptrsz));
- base_hi = create_uniform_indirect(b, ubo + 1, ir3_get_addr0(ctx, src0, ptrsz));
+ base_lo = create_uniform_indirect(b, ubo, TYPE_U32, ir3_get_addr0(ctx, src0, ptrsz));
+ base_hi = create_uniform_indirect(b, ubo + 1, TYPE_U32, ir3_get_addr0(ctx, src0, ptrsz));
/* NOTE: since relative addressing is used, make sure constlen is
* at least big enough to cover all the UBO addresses, since the
emit_intrinsic_ssbo_size(struct ir3_context *ctx, nir_intrinsic_instr *intr,
struct ir3_instruction **dst)
{
+ if (ir3_bindless_resource(intr->src[0])) {
+ struct ir3_block *b = ctx->block;
+ struct ir3_instruction *ibo = ir3_ssbo_to_ibo(ctx, intr->src[0]);
+ struct ir3_instruction *resinfo = ir3_RESINFO(b, ibo, 0);
+ resinfo->cat6.iim_val = 1;
+ resinfo->cat6.d = 1;
+ resinfo->cat6.type = TYPE_U32;
+ resinfo->cat6.typed = false;
+ /* resinfo has no writemask and always writes out 3 components */
+ resinfo->regs[0]->wrmask = MASK(3);
+ ir3_handle_bindless_cat6(resinfo, intr->src[0]);
+ struct ir3_instruction *resinfo_dst;
+ ir3_split_dest(b, &resinfo_dst, resinfo, 0, 1);
+ /* Unfortunately resinfo returns the array length, i.e. in dwords,
+ * while NIR expects us to return the size in bytes.
+ *
+ * TODO: fix this in NIR.
+ */
+ *dst = ir3_SHL_B(b, resinfo_dst, 0, create_immed(b, 2), 0);
+ return;
+ }
+
/* SSBO size stored as a const starting at ssbo_sizes: */
- struct ir3_const_state *const_state = &ctx->so->shader->const_state;
+ const struct ir3_const_state *const_state = ir3_const_state(ctx->so);
unsigned blk_idx = nir_src_as_uint(intr->src[0]);
unsigned idx = regid(const_state->offsets.ssbo_sizes, 0) +
const_state->ssbo_size.off[blk_idx];
struct ir3_block *b = ctx->block;
struct ir3_instruction *stl, *offset;
struct ir3_instruction * const *value;
- unsigned base, wrmask;
+ unsigned base, wrmask, ncomp;
value = ir3_get_src(ctx, &intr->src[0]);
offset = ir3_get_src(ctx, &intr->src[1])[0];
base = nir_intrinsic_base(intr);
wrmask = nir_intrinsic_write_mask(intr);
+ ncomp = ffs(~wrmask) - 1;
- /* Combine groups of consecutive enabled channels in one write
- * message. We use ffs to find the first enabled channel and then ffs on
- * the bit-inverse, down-shifted writemask to determine the length of
- * the block of enabled bits.
- *
- * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
- */
- while (wrmask) {
- unsigned first_component = ffs(wrmask) - 1;
- unsigned length = ffs(~(wrmask >> first_component)) - 1;
-
- stl = ir3_STL(b, offset, 0,
- ir3_create_collect(ctx, &value[first_component], length), 0,
- create_immed(b, length), 0);
- stl->cat6.dst_offset = first_component + base;
- stl->cat6.type = utype_src(intr->src[0]);
- stl->barrier_class = IR3_BARRIER_SHARED_W;
- stl->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
-
- array_insert(b, b->keeps, stl);
-
- /* Clear the bits in the writemask that we just wrote, then try
- * again to see if more channels are left.
- */
- wrmask &= (15 << (first_component + length));
- }
+ assert(wrmask == BITFIELD_MASK(intr->num_components));
+
+ stl = ir3_STL(b, offset, 0,
+ ir3_create_collect(ctx, value, ncomp), 0,
+ create_immed(b, ncomp), 0);
+ stl->cat6.dst_offset = base;
+ stl->cat6.type = utype_src(intr->src[0]);
+ stl->barrier_class = IR3_BARRIER_SHARED_W;
+ stl->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
+
+ array_insert(b, b->keeps, stl);
}
/* src[] = { offset }. const_index[] = { base } */
create_immed(b, intr->num_components), 0,
create_immed(b, base), 0);
+ /* for a650, use LDL for tess ctrl inputs: */
+ if (ctx->so->type == MESA_SHADER_TESS_CTRL && ctx->compiler->tess_use_shared)
+ load->opc = OPC_LDL;
+
load->cat6.type = utype_dst(intr->dest);
load->regs[0]->wrmask = MASK(intr->num_components);
ir3_split_dest(b, dst, load, 0, intr->num_components);
}
-/* src[] = { value, offset }. const_index[] = { base, write_mask } */
+/* src[] = { value, offset }. const_index[] = { base } */
static void
emit_intrinsic_store_shared_ir3(struct ir3_context *ctx, nir_intrinsic_instr *intr)
{
struct ir3_block *b = ctx->block;
struct ir3_instruction *store, *offset;
struct ir3_instruction * const *value;
- unsigned base, wrmask;
value = ir3_get_src(ctx, &intr->src[0]);
offset = ir3_get_src(ctx, &intr->src[1])[0];
- base = nir_intrinsic_base(intr);
- wrmask = nir_intrinsic_write_mask(intr);
-
- /* Combine groups of consecutive enabled channels in one write
- * message. We use ffs to find the first enabled channel and then ffs on
- * the bit-inverse, down-shifted writemask to determine the length of
- * the block of enabled bits.
- *
- * (trick stolen from i965's fs_visitor::nir_emit_cs_intrinsic())
- */
- while (wrmask) {
- unsigned first_component = ffs(wrmask) - 1;
- unsigned length = ffs(~(wrmask >> first_component)) - 1;
+ store = ir3_STLW(b, offset, 0,
+ ir3_create_collect(ctx, value, intr->num_components), 0,
+ create_immed(b, intr->num_components), 0);
- store = ir3_STLW(b, offset, 0,
- ir3_create_collect(ctx, &value[first_component], length), 0,
- create_immed(b, length), 0);
+ /* for a650, use STL for vertex outputs used by tess ctrl shader: */
+ if (ctx->so->type == MESA_SHADER_VERTEX && ctx->so->key.tessellation &&
+ ctx->compiler->tess_use_shared)
+ store->opc = OPC_STL;
- store->cat6.dst_offset = first_component + base;
- store->cat6.type = utype_src(intr->src[0]);
- store->barrier_class = IR3_BARRIER_SHARED_W;
- store->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
+ store->cat6.dst_offset = nir_intrinsic_base(intr);
+ store->cat6.type = utype_src(intr->src[0]);
+ store->barrier_class = IR3_BARRIER_SHARED_W;
+ store->barrier_conflict = IR3_BARRIER_SHARED_R | IR3_BARRIER_SHARED_W;
- array_insert(b, b->keeps, store);
-
- /* Clear the bits in the writemask that we just wrote, then try
- * again to see if more channels are left.
- */
- wrmask &= (15 << (first_component + length));
- }
+ array_insert(b, b->keeps, store);
}
/*
ir3_split_dest(b, dst, sam, 0, 4);
}
-static void
-emit_intrinsic_image_size(struct ir3_context *ctx, nir_intrinsic_instr *intr,
+/* A4xx version of image_size, see ir3_a6xx.c for newer resinfo version. */
+void
+emit_intrinsic_image_size_tex(struct ir3_context *ctx, nir_intrinsic_instr *intr,
struct ir3_instruction **dst)
{
struct ir3_block *b = ctx->block;
TYPE_U16 : TYPE_U32;
info.flags |= flags;
+ assert(nir_src_as_uint(intr->src[1]) == 0);
lod = create_immed(b, 0);
sam = emit_sam(ctx, OPC_GETSIZE, info, dst_type, 0b1111, lod, NULL);
* bytes-per-pixel should have been emitted in 2nd slot of
* image_dims. See ir3_shader::emit_image_dims().
*/
- struct ir3_const_state *const_state = &ctx->so->shader->const_state;
+ const struct ir3_const_state *const_state =
+ ir3_const_state(ctx->so);
unsigned cb = regid(const_state->offsets.image_dims, 0) +
const_state->image_dims.off[nir_src_as_uint(intr->src[0])];
struct ir3_instruction *aux = create_uniform(b, cb + 1);
}
static struct ir3_instruction *
-get_barycentric_centroid(struct ir3_context *ctx)
+get_barycentric(struct ir3_context *ctx, enum ir3_bary bary)
{
- if (!ctx->ij_centroid) {
+ static const gl_system_value sysval_base = SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL;
+
+ STATIC_ASSERT(sysval_base + IJ_PERSP_PIXEL == SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL);
+ STATIC_ASSERT(sysval_base + IJ_PERSP_SAMPLE == SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE);
+ STATIC_ASSERT(sysval_base + IJ_PERSP_CENTROID == SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID);
+ STATIC_ASSERT(sysval_base + IJ_PERSP_SIZE == SYSTEM_VALUE_BARYCENTRIC_PERSP_SIZE);
+ STATIC_ASSERT(sysval_base + IJ_LINEAR_PIXEL == SYSTEM_VALUE_BARYCENTRIC_LINEAR_PIXEL);
+ STATIC_ASSERT(sysval_base + IJ_LINEAR_CENTROID == SYSTEM_VALUE_BARYCENTRIC_LINEAR_CENTROID);
+ STATIC_ASSERT(sysval_base + IJ_LINEAR_SAMPLE == SYSTEM_VALUE_BARYCENTRIC_LINEAR_SAMPLE);
+
+ if (!ctx->ij[bary]) {
struct ir3_instruction *xy[2];
struct ir3_instruction *ij;
- ij = create_sysval_input(ctx, SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID, 0x3);
+ ij = create_sysval_input(ctx, sysval_base + bary, 0x3);
ir3_split_dest(ctx->block, xy, ij, 0, 2);
- ctx->ij_centroid = ir3_create_collect(ctx, xy, 2);
+ ctx->ij[bary] = ir3_create_collect(ctx, xy, 2);
}
- return ctx->ij_centroid;
+ return ctx->ij[bary];
}
-static struct ir3_instruction *
-get_barycentric_sample(struct ir3_context *ctx)
+/* TODO: make this a common NIR helper?
+ * there is a nir_system_value_from_intrinsic but it takes nir_intrinsic_op so it
+ * can't be extended to work with this
+ */
+static gl_system_value
+nir_intrinsic_barycentric_sysval(nir_intrinsic_instr *intr)
{
- if (!ctx->ij_sample) {
- struct ir3_instruction *xy[2];
- struct ir3_instruction *ij;
-
- ij = create_sysval_input(ctx, SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE, 0x3);
- ir3_split_dest(ctx->block, xy, ij, 0, 2);
+ enum glsl_interp_mode interp_mode = nir_intrinsic_interp_mode(intr);
+ gl_system_value sysval;
- ctx->ij_sample = ir3_create_collect(ctx, xy, 2);
+ switch (intr->intrinsic) {
+ case nir_intrinsic_load_barycentric_pixel:
+ if (interp_mode == INTERP_MODE_NOPERSPECTIVE)
+ sysval = SYSTEM_VALUE_BARYCENTRIC_LINEAR_PIXEL;
+ else
+ sysval = SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL;
+ break;
+ case nir_intrinsic_load_barycentric_centroid:
+ if (interp_mode == INTERP_MODE_NOPERSPECTIVE)
+ sysval = SYSTEM_VALUE_BARYCENTRIC_LINEAR_CENTROID;
+ else
+ sysval = SYSTEM_VALUE_BARYCENTRIC_PERSP_CENTROID;
+ break;
+ case nir_intrinsic_load_barycentric_sample:
+ if (interp_mode == INTERP_MODE_NOPERSPECTIVE)
+ sysval = SYSTEM_VALUE_BARYCENTRIC_LINEAR_SAMPLE;
+ else
+ sysval = SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE;
+ break;
+ default:
+ unreachable("invalid barycentric intrinsic");
}
- return ctx->ij_sample;
+ return sysval;
}
-static struct ir3_instruction *
-get_barycentric_pixel(struct ir3_context *ctx)
+static void
+emit_intrinsic_barycentric(struct ir3_context *ctx, nir_intrinsic_instr *intr,
+ struct ir3_instruction **dst)
{
- /* TODO when tgsi_to_nir supports "new-style" FS inputs switch
- * this to create ij_pixel only on demand:
- */
- return ctx->ij_pixel;
+ gl_system_value sysval = nir_intrinsic_barycentric_sysval(intr);
+
+ if (!ctx->so->key.msaa) {
+ if (sysval == SYSTEM_VALUE_BARYCENTRIC_PERSP_SAMPLE)
+ sysval = SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL;
+ if (sysval == SYSTEM_VALUE_BARYCENTRIC_LINEAR_SAMPLE)
+ sysval = SYSTEM_VALUE_BARYCENTRIC_LINEAR_PIXEL;
+ }
+
+ enum ir3_bary bary = sysval - SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL;
+
+ struct ir3_instruction *ij = get_barycentric(ctx, bary);
+ ir3_split_dest(ctx->block, dst, ij, 0, 2);
}
static struct ir3_instruction *
struct ir3_instruction **dst;
struct ir3_instruction * const *src;
struct ir3_block *b = ctx->block;
+ unsigned dest_components = nir_intrinsic_dest_components(intr);
int idx, comp;
if (info->has_dest) {
- unsigned n = nir_intrinsic_dest_components(intr);
- dst = ir3_get_dst(ctx, &intr->dest, n);
+ dst = ir3_get_dst(ctx, &intr->dest, dest_components);
} else {
dst = NULL;
}
- const unsigned primitive_param = ctx->so->shader->const_state.offsets.primitive_param * 4;
- const unsigned primitive_map = ctx->so->shader->const_state.offsets.primitive_map * 4;
+ const struct ir3_const_state *const_state = ir3_const_state(ctx->so);
+ const unsigned primitive_param = const_state->offsets.primitive_param * 4;
+ const unsigned primitive_map = const_state->offsets.primitive_map * 4;
switch (intr->intrinsic) {
case nir_intrinsic_load_uniform:
idx = nir_intrinsic_base(intr);
if (nir_src_is_const(intr->src[0])) {
idx += nir_src_as_uint(intr->src[0]);
- for (int i = 0; i < intr->num_components; i++) {
+ for (int i = 0; i < dest_components; i++) {
dst[i] = create_uniform_typed(b, idx + i,
nir_dest_bit_size(intr->dest) == 16 ? TYPE_F16 : TYPE_F32);
}
} else {
src = ir3_get_src(ctx, &intr->src[0]);
- for (int i = 0; i < intr->num_components; i++) {
+ for (int i = 0; i < dest_components; i++) {
dst[i] = create_uniform_indirect(b, idx + i,
+ nir_dest_bit_size(intr->dest) == 16 ? TYPE_F16 : TYPE_F32,
ir3_get_addr0(ctx, src[0], 1));
}
/* NOTE: if relative addressing is used, we set
* addr reg value can be:
*/
ctx->so->constlen = MAX2(ctx->so->constlen,
- ctx->so->shader->ubo_state.size / 16);
+ const_state->ubo_state.size / 16);
}
break;
case nir_intrinsic_store_global_ir3: {
struct ir3_instruction *value, *addr, *offset;
+ unsigned ncomp = nir_intrinsic_src_components(intr, 0);
addr = ir3_create_collect(ctx, (struct ir3_instruction*[]){
ir3_get_src(ctx, &intr->src[1])[0],
offset = ir3_get_src(ctx, &intr->src[2])[0];
- value = ir3_create_collect(ctx, ir3_get_src(ctx, &intr->src[0]),
- intr->num_components);
+ value = ir3_create_collect(ctx, ir3_get_src(ctx, &intr->src[0]), ncomp);
struct ir3_instruction *stg =
ir3_STG_G(ctx->block, addr, 0, value, 0,
- create_immed(ctx->block, intr->num_components), 0, offset, 0);
+ create_immed(ctx->block, ncomp), 0, offset, 0);
stg->cat6.type = TYPE_U32;
stg->cat6.iim_val = 1;
offset = ir3_get_src(ctx, &intr->src[1])[0];
struct ir3_instruction *load =
- ir3_LDG(b, addr, 0, create_immed(ctx->block, intr->num_components),
+ ir3_LDG(b, addr, 0, create_immed(ctx->block, dest_components),
0, offset, 0);
load->cat6.type = TYPE_U32;
- load->regs[0]->wrmask = MASK(intr->num_components);
+ load->regs[0]->wrmask = MASK(dest_components);
load->barrier_class = IR3_BARRIER_BUFFER_R;
load->barrier_conflict = IR3_BARRIER_BUFFER_W;
- ir3_split_dest(b, dst, load, 0, intr->num_components);
+ ir3_split_dest(b, dst, load, 0, dest_components);
break;
}
break;
}
case nir_intrinsic_load_size_ir3:
- if (!ctx->ij_size) {
- ctx->ij_size =
+ if (!ctx->ij[IJ_PERSP_SIZE]) {
+ ctx->ij[IJ_PERSP_SIZE] =
create_sysval_input(ctx, SYSTEM_VALUE_BARYCENTRIC_PERSP_SIZE, 0x1);
}
- dst[0] = ctx->ij_size;
+ dst[0] = ctx->ij[IJ_PERSP_SIZE];
break;
case nir_intrinsic_load_barycentric_centroid:
- ir3_split_dest(b, dst, get_barycentric_centroid(ctx), 0, 2);
- break;
case nir_intrinsic_load_barycentric_sample:
- if (ctx->so->key.msaa) {
- ir3_split_dest(b, dst, get_barycentric_sample(ctx), 0, 2);
- } else {
- ir3_split_dest(b, dst, get_barycentric_pixel(ctx), 0, 2);
- }
- break;
case nir_intrinsic_load_barycentric_pixel:
- ir3_split_dest(b, dst, get_barycentric_pixel(ctx), 0, 2);
+ emit_intrinsic_barycentric(ctx, intr, dst);
break;
case nir_intrinsic_load_interpolated_input:
idx = nir_intrinsic_base(intr);
if (nir_src_is_const(intr->src[1])) {
struct ir3_instruction *coord = ir3_create_collect(ctx, src, 2);
idx += nir_src_as_uint(intr->src[1]);
- for (int i = 0; i < intr->num_components; i++) {
+ for (int i = 0; i < dest_components; i++) {
unsigned inloc = idx * 4 + i + comp;
if (ctx->so->inputs[idx].bary &&
!ctx->so->inputs[idx].use_ldlv) {
comp = nir_intrinsic_component(intr);
if (nir_src_is_const(intr->src[0])) {
idx += nir_src_as_uint(intr->src[0]);
- for (int i = 0; i < intr->num_components; i++) {
+ for (int i = 0; i < dest_components; i++) {
unsigned n = idx * 4 + i + comp;
dst[i] = ctx->inputs[n];
compile_assert(ctx, ctx->inputs[n]);
struct ir3_instruction *collect =
ir3_create_collect(ctx, ctx->ir->inputs, ctx->ninputs);
struct ir3_instruction *addr = ir3_get_addr0(ctx, src[0], 4);
- for (int i = 0; i < intr->num_components; i++) {
+ for (int i = 0; i < dest_components; i++) {
unsigned n = idx * 4 + i + comp;
dst[i] = create_indirect_load(ctx, ctx->ninputs,
n, addr, collect);
break;
case nir_intrinsic_image_size:
case nir_intrinsic_bindless_image_size:
- emit_intrinsic_image_size(ctx, intr, dst);
+ ctx->funcs->emit_intrinsic_image_size(ctx, intr, dst);
break;
case nir_intrinsic_image_atomic_add:
case nir_intrinsic_bindless_image_atomic_add:
idx += nir_src_as_uint(intr->src[1]);
src = ir3_get_src(ctx, &intr->src[0]);
- for (int i = 0; i < intr->num_components; i++) {
+ for (int i = 0; i < nir_intrinsic_src_components(intr, 0); i++) {
unsigned n = idx * 4 + i + comp;
ctx->outputs[n] = src[i];
}
}
dst[0] = ctx->basevertex;
break;
+ case nir_intrinsic_load_draw_id:
+ if (!ctx->draw_id) {
+ ctx->draw_id = create_driver_param(ctx, IR3_DP_DRAWID);
+ }
+ dst[0] = ctx->draw_id;
+ break;
case nir_intrinsic_load_base_instance:
if (!ctx->base_instance) {
ctx->base_instance = create_driver_param(ctx, IR3_DP_INSTID_BASE);
}
dst[0] = ctx->base_instance;
break;
+ case nir_intrinsic_load_view_index:
+ if (!ctx->view_index) {
+ ctx->view_index = create_sysval_input(ctx, SYSTEM_VALUE_VIEW_INDEX, 0x1);
+ }
+ dst[0] = ctx->view_index;
+ break;
case nir_intrinsic_load_vertex_id_zero_base:
case nir_intrinsic_load_vertex_id:
if (!ctx->vertex_id) {
break;
case nir_intrinsic_load_user_clip_plane:
idx = nir_intrinsic_ucp_id(intr);
- for (int i = 0; i < intr->num_components; i++) {
+ for (int i = 0; i < dest_components; i++) {
unsigned n = idx * 4 + i;
dst[i] = create_driver_param(ctx, IR3_DP_UCP0_X + n);
}
ir3_split_dest(b, dst, ctx->work_group_id, 0, 3);
break;
case nir_intrinsic_load_num_work_groups:
- for (int i = 0; i < intr->num_components; i++) {
+ for (int i = 0; i < dest_components; i++) {
dst[i] = create_driver_param(ctx, IR3_DP_NUM_WORK_GROUPS_X + i);
}
break;
case nir_intrinsic_load_local_group_size:
- for (int i = 0; i < intr->num_components; i++) {
+ for (int i = 0; i < dest_components; i++) {
dst[i] = create_driver_param(ctx, IR3_DP_LOCAL_GROUP_SIZE_X + i);
}
break;
array_insert(ctx->ir, ctx->ir->predicates, kill);
array_insert(b, b->keeps, kill);
- ctx->so->no_earlyz = true;
+ ctx->so->has_kill = true;
break;
}
sam = ir3_META_TEX_PREFETCH(b);
__ssa_dst(sam)->wrmask = MASK(ncomp); /* dst */
- __ssa_src(sam, get_barycentric_pixel(ctx), 0);
+ __ssa_src(sam, get_barycentric(ctx, IJ_PERSP_PIXEL), 0);
sam->prefetch.input_offset =
ir3_nir_coord_offset(tex->src[idx].src.ssa);
/* make sure not to add irrelevant flags like S2EN */
* stripped out in the backend.
*/
for (unsigned i = 0; i < IR3_MAX_SO_BUFFERS; i++) {
- struct ir3_const_state *const_state = &ctx->so->shader->const_state;
+ const struct ir3_const_state *const_state =
+ ir3_const_state(ctx->so);
unsigned stride = strmout->stride[i];
struct ir3_instruction *base, *off;
ctx->inputs[idx] = instr;
}
} else if (ctx->so->type == MESA_SHADER_VERTEX) {
- struct ir3_instruction *input = NULL, *in;
+ struct ir3_instruction *input = NULL;
struct ir3_instruction *components[4];
+ /* input as setup as frac=0 with "ncomp + frac" components,
+ * this avoids getting a sparse writemask
+ */
unsigned mask = (1 << (ncomp + frac)) - 1;
foreach_input (in, ctx->ir) {
input = create_input(ctx, mask);
input->input.inidx = n;
} else {
+ /* For aliased inputs, just append to the wrmask.. ie. if we
+ * first see a vec2 index at slot N, and then later a vec4,
+ * the wrmask of the resulting overlapped vec2 and vec4 is 0xf
+ *
+ * If the new input that aliases a previously processed input
+ * sets no new bits, then just bail as there is nothing to see
+ * here.
+ */
+ if (!(mask & ~input->regs[0]->wrmask))
+ return;
input->regs[0]->wrmask |= mask;
}
- ir3_split_dest(ctx->block, components, input, frac, ncomp);
+ ir3_split_dest(ctx->block, components, input, 0, ncomp + frac);
- for (int i = 0; i < ncomp; i++) {
- unsigned idx = (n * 4) + i + frac;
+ for (int i = 0; i < ncomp + frac; i++) {
+ unsigned idx = (n * 4) + i;
compile_assert(ctx, idx < ctx->ninputs);
+
+ /* With aliased inputs, since we add to the wrmask above, we
+ * can end up with stale meta:split instructions in the inputs
+ * table. This is basically harmless, since eventually they
+ * will get swept away by DCE, but the mismatch wrmask (since
+ * they would be using the previous wrmask before we OR'd in
+ * more bits) angers ir3_validate. So just preemptively clean
+ * them up. See:
+ *
+ * dEQP-GLES2.functional.attribute_location.bind_aliasing.cond_vec2
+ *
+ * Note however that split_dest() will return the src if it is
+ * scalar, so the previous ctx->inputs[idx] could be the input
+ * itself (which we don't want to remove)
+ */
+ if (ctx->inputs[idx] && (ctx->inputs[idx] != input)) {
+ list_del(&ctx->inputs[idx]->node);
+ }
+
ctx->inputs[idx] = components[i];
}
} else {
ir3_context_error(ctx, "unknown shader type: %d\n", ctx->so->type);
}
+ /* note: this can be wrong for sparse vertex inputs, this happens with
+ * vulkan, only a3xx/a4xx use this value for VS, so it shouldn't matter
+ */
if (so->inputs[n].bary || (ctx->so->type == MESA_SHADER_VERTEX)) {
so->total_in += ncomp;
}
case FRAG_RESULT_SAMPLE_MASK:
so->writes_smask = true;
break;
+ case FRAG_RESULT_STENCIL:
+ so->writes_stencilref = true;
+ break;
default:
+ slot += out->data.index; /* For dual-src blend */
if (slot >= FRAG_RESULT_DATA0)
break;
ir3_context_error(ctx, "unknown FS output name: %s\n",
ctx->inputs = rzalloc_array(ctx, struct ir3_instruction *, ctx->ninputs);
ctx->outputs = rzalloc_array(ctx, struct ir3_instruction *, ctx->noutputs);
- ctx->ir = ir3_create(ctx->compiler, ctx->so->type);
+ ctx->ir = ir3_create(ctx->compiler, ctx->so);
/* Create inputs in first block: */
ctx->block = get_block(ctx, nir_start_block(fxn));
* tgsi_to_nir)
*/
if (ctx->so->type == MESA_SHADER_FRAGMENT) {
- ctx->ij_pixel = create_input(ctx, 0x3);
+ ctx->ij[IJ_PERSP_PIXEL] = create_input(ctx, 0x3);
}
/* Setup inputs: */
- nir_foreach_variable (var, &ctx->s->inputs) {
+ nir_foreach_shader_in_variable (var, ctx->s) {
setup_input(ctx, var);
}
/* Defer add_sysval_input() stuff until after setup_inputs(),
* because sysvals need to be appended after varyings:
*/
- if (ctx->ij_pixel) {
+ if (ctx->ij[IJ_PERSP_PIXEL]) {
add_sysval_input_compmask(ctx, SYSTEM_VALUE_BARYCENTRIC_PERSP_PIXEL,
- 0x3, ctx->ij_pixel);
+ 0x3, ctx->ij[IJ_PERSP_PIXEL]);
}
}
/* Setup outputs: */
- nir_foreach_variable (var, &ctx->s->outputs) {
+ nir_foreach_shader_out_variable (var, ctx->s) {
setup_output(ctx, var);
}
- /* Find # of samplers: */
- nir_foreach_variable (var, &ctx->s->uniforms) {
- ctx->so->num_samp += glsl_type_get_sampler_count(var->type);
- /* just assume that we'll be reading from images.. if it
- * is write-only we don't have to count it, but not sure
- * if there is a good way to know?
- */
- ctx->so->num_samp += glsl_type_get_image_count(var->type);
- }
+ /* Find # of samplers. Just assume that we'll be reading from images.. if
+ * it is write-only we don't have to count it, but after lowering derefs
+ * is too late to compact indices for that.
+ */
+ ctx->so->num_samp = util_last_bit(ctx->s->info.textures_used) + ctx->s->info.num_images;
/* NOTE: need to do something more clever when we support >1 fxn */
nir_foreach_register (reg, &fxn->registers) {
so->outputs[j] = so->outputs[i];
/* fixup outidx to point to new output table entry: */
- struct ir3_instruction *out;
foreach_output (out, ir) {
if (out->collect.outidx == i) {
out->collect.outidx = j;
fetch->dst = instr->regs[0]->num;
fetch->src = instr->prefetch.input_offset;
+ /* These are the limits on a5xx/a6xx, we might need to
+ * revisit if SP_FS_PREFETCH[n] changes on later gens:
+ */
+ assert(fetch->dst <= 0x3f);
+ assert(fetch->tex_id <= 0x1f);
+ assert(fetch->samp_id < 0xf);
+
ctx->so->total_in =
MAX2(ctx->so->total_in, instr->prefetch.input_offset + 2);
- /* Disable half precision until supported. */
fetch->half_precision = !!(instr->regs[0]->flags & IR3_REG_HALF);
/* Remove the prefetch placeholder instruction: */
struct ir3_context *ctx;
struct ir3 *ir;
int ret = 0, max_bary;
+ bool progress;
assert(!so->ir);
}
}
- /* at this point, for binning pass, throw away unneeded outputs: */
- if (so->binning_pass && (ctx->compiler->gpu_id < 600))
- fixup_binning_pass(ctx);
-
- ir3_debug_print(ir, "BEFORE CF");
-
- ir3_cf(ir);
-
- ir3_debug_print(ir, "BEFORE CP");
-
- ir3_cp(ir, so);
-
- /* at this point, for binning pass, throw away unneeded outputs:
- * Note that for a6xx and later, we do this after ir3_cp to ensure
- * that the uniform/constant layout for BS and VS matches, so that
- * we can re-use same VS_CONST state group.
- */
- if (so->binning_pass && (ctx->compiler->gpu_id >= 600))
- fixup_binning_pass(ctx);
-
/* for a6xx+, binning and draw pass VS use same VBO state, so we
* need to make sure not to remove any inputs that are used by
* the nonbinning VS.
}
}
- ir3_debug_print(ir, "BEFORE GROUPING");
+ /* at this point, for binning pass, throw away unneeded outputs: */
+ if (so->binning_pass && (ctx->compiler->gpu_id < 600))
+ fixup_binning_pass(ctx);
- ir3_sched_add_deps(ir);
+ ir3_debug_print(ir, "AFTER: nir->ir3");
+ ir3_validate(ir);
- /* Group left/right neighbors, inserting mov's where needed to
- * solve conflicts:
- */
- ir3_group(ir);
+ do {
+ progress = false;
- ir3_debug_print(ir, "AFTER GROUPING");
+ progress |= IR3_PASS(ir, ir3_cf);
+ progress |= IR3_PASS(ir, ir3_cp, so);
+ progress |= IR3_PASS(ir, ir3_dce, so);
+ } while (progress);
- ir3_dce(ir, so);
+ /* at this point, for binning pass, throw away unneeded outputs:
+ * Note that for a6xx and later, we do this after ir3_cp to ensure
+ * that the uniform/constant layout for BS and VS matches, so that
+ * we can re-use same VS_CONST state group.
+ */
+ if (so->binning_pass && (ctx->compiler->gpu_id >= 600)) {
+ fixup_binning_pass(ctx);
+ /* cleanup the result of removing unneeded outputs: */
+ while (IR3_PASS(ir, ir3_dce, so)) {}
+ }
- ir3_debug_print(ir, "AFTER DCE");
+ IR3_PASS(ir, ir3_sched_add_deps);
+
+ /* Group left/right neighbors, inserting mov's where needed to
+ * solve conflicts:
+ */
+ IR3_PASS(ir, ir3_group);
- /* do Sethi–Ullman numbering before scheduling: */
- ir3_sun(ir);
+ /* At this point, all the dead code should be long gone: */
+ assert(!IR3_PASS(ir, ir3_dce, so));
ret = ir3_sched(ir);
if (ret) {
goto out;
}
- ir3_debug_print(ir, "AFTER SCHED");
+ ir3_debug_print(ir, "AFTER: ir3_sched");
+
+ if (IR3_PASS(ir, ir3_cp_postsched)) {
+ /* cleanup the result of removing unneeded mov's: */
+ while (IR3_PASS(ir, ir3_dce, so)) {}
+ }
/* Pre-assign VS inputs on a6xx+ binning pass shader, to align
* with draw pass VS, so binning and draw pass can both use the
ret = ir3_ra(so, precolor, ARRAY_SIZE(precolor));
} else if (so->num_sampler_prefetch) {
assert(so->type == MESA_SHADER_FRAGMENT);
- struct ir3_instruction *instr, *precolor[2];
+ struct ir3_instruction *precolor[2];
int idx = 0;
foreach_input (instr, ir) {
goto out;
}
- ir3_postsched(ctx);
- ir3_debug_print(ir, "AFTER POSTSCHED");
+ IR3_PASS(ir, ir3_postsched, so);
if (compiler->gpu_id >= 600) {
- if (ir3_a6xx_fixup_atomic_dests(ir, so)) {
- ir3_debug_print(ir, "AFTER ATOMIC FIXUP");
- }
+ IR3_PASS(ir, ir3_a6xx_fixup_atomic_dests, so);
}
if (so->type == MESA_SHADER_FRAGMENT)
for (unsigned i = 0; i < so->outputs_count; i++)
so->outputs[i].regid = INVALID_REG;
- struct ir3_instruction *out;
foreach_output (out, ir) {
assert(out->opc == OPC_META_COLLECT);
unsigned outidx = out->collect.outidx;
so->outputs[outidx].half = !!(out->regs[0]->flags & IR3_REG_HALF);
}
- struct ir3_instruction *in;
foreach_input (in, ir) {
assert(in->opc == OPC_META_INPUT);
unsigned inidx = in->input.inidx;
/* We need to do legalize after (for frag shader's) the "bary.f"
* offsets (inloc) have been assigned.
*/
- ir3_legalize(ir, so, &max_bary);
-
- ir3_debug_print(ir, "AFTER LEGALIZE");
+ IR3_PASS(ir, ir3_legalize, so, &max_bary);
/* Set (ss)(sy) on first TCS and GEOMETRY instructions, since we don't
* know what we might have to wait on when coming in from VS chsh.
if (so->type == MESA_SHADER_FRAGMENT)
so->total_in = max_bary + 1;
- so->max_sun = ir->max_sun;
-
/* Collect sampling instructions eligible for pre-dispatch. */
collect_tex_prefetches(ctx, ir);