*/
bool unminify_coords;
- /* for looking up which system value is which */
- unsigned sysval_semantics[8];
+ /* on a4xx, for array textures we need to add 0.5 to the array
+ * index coordinate:
+ */
+ bool array_index_add_half;
+
+ /* on a4xx, bitmask of samplers which need astc+srgb workaround: */
+ unsigned astc_srgb;
+
+ unsigned max_texture_index;
/* set if we encounter something we can't handle yet, so we
* can bail cleanly and fallback to TGSI compiler f/e
bool error;
};
+/* gpu pointer size in units of 32bit registers/slots */
+static unsigned pointer_size(struct ir3_compile *ctx)
+{
+ return (ctx->compiler->gpu_id >= 500) ? 2 : 1;
+}
static struct ir3_instruction * create_immed(struct ir3_block *block, uint32_t val);
static struct ir3_block * get_block(struct ir3_compile *ctx, nir_block *nblock);
ctx->flat_bypass = true;
ctx->levels_add_one = false;
ctx->unminify_coords = false;
+ ctx->array_index_add_half = true;
+
+ if (so->type == SHADER_VERTEX)
+ ctx->astc_srgb = so->key.vastc_srgb;
+ else if (so->type == SHADER_FRAGMENT)
+ ctx->astc_srgb = so->key.fastc_srgb;
+
} else {
/* no special handling for "flat" */
ctx->flat_bypass = false;
ctx->levels_add_one = true;
ctx->unminify_coords = true;
+ ctx->array_index_add_half = false;
}
ctx->compiler = compiler;
nir_print_shader(ctx->s, stdout);
}
- so->first_driver_param = so->first_immediate = ctx->s->num_uniforms;
+ so->num_uniforms = ctx->s->num_uniforms;
+ so->num_ubos = ctx->s->info->num_ubos;
- /* Layout of constant registers:
+ /* Layout of constant registers, each section aligned to vec4. Note
+ * that pointer size (ubo, etc) changes depending on generation.
*
- * num_uniform * vec4 - user consts
- * 4 * vec4 - UBO addresses
+ * user consts
+ * UBO addresses
* if (vertex shader) {
- * N * vec4 - driver params (IR3_DP_*)
- * 1 * vec4 - stream-out addresses
+ * driver params (IR3_DP_*)
+ * if (stream_output.num_outputs > 0)
+ * stream-out addresses
* }
+ * immediates
*
- * TODO this could be made more dynamic, to at least skip sections
- * that we don't need..
+ * Immediates go last mostly because they are inserted in the CP pass
+ * after the nir -> ir3 frontend.
*/
+ unsigned constoff = align(ctx->s->num_uniforms, 4);
+ unsigned ptrsz = pointer_size(ctx);
+
+ memset(&so->constbase, ~0, sizeof(so->constbase));
- /* reserve 4 (vec4) slots for ubo base addresses: */
- so->first_immediate += 4;
+ if (so->num_ubos > 0) {
+ so->constbase.ubo = constoff;
+ constoff += align(ctx->s->info->num_ubos * ptrsz, 4) / 4;
+ }
if (so->type == SHADER_VERTEX) {
- /* driver params (see ir3_driver_param): */
- so->first_immediate += IR3_DP_COUNT/4; /* convert to vec4 */
- /* one (vec4) slot for stream-output base addresses: */
- so->first_immediate++;
+ so->constbase.driver_param = constoff;
+ constoff += align(IR3_DP_COUNT, 4) / 4;
+
+ if ((compiler->gpu_id < 500) &&
+ so->shader->stream_output.num_outputs > 0) {
+ so->constbase.tfbo = constoff;
+ constoff += align(PIPE_MAX_SO_BUFFERS * ptrsz, 4) / 4;
+ }
}
+ so->constbase.immediate = constoff;
+
return ctx;
}
declare_var(struct ir3_compile *ctx, nir_variable *var)
{
unsigned length = glsl_get_length(var->type) * 4; /* always vec4, at least with ttn */
- struct ir3_array *arr = ralloc(ctx, struct ir3_array);
+ struct ir3_array *arr = rzalloc(ctx, struct ir3_array);
arr->id = ++ctx->num_arrays;
arr->length = length;
arr->var = var;
return __get_dst(ctx, dst, n);
}
-static struct ir3_instruction **
+static struct ir3_instruction * const *
get_src(struct ir3_compile *ctx, nir_src *src)
{
struct hash_entry *entry;
{
struct ir3_instruction *mov;
- mov = ir3_instr_create(block, 1, 0);
+ mov = ir3_instr_create(block, OPC_MOV);
mov->cat1.src_type = TYPE_U32;
mov->cat1.dst_type = TYPE_U32;
ir3_reg_create(mov, 0, 0);
{
struct ir3_instruction *mov;
- mov = ir3_instr_create(ctx->block, 1, 0);
+ mov = ir3_instr_create(ctx->block, OPC_MOV);
/* TODO get types right? */
mov->cat1.src_type = TYPE_F32;
mov->cat1.dst_type = TYPE_F32;
{
struct ir3_instruction *mov;
- mov = ir3_instr_create(ctx->block, 1, 0);
+ mov = ir3_instr_create(ctx->block, OPC_MOV);
mov->cat1.src_type = TYPE_U32;
mov->cat1.dst_type = TYPE_U32;
ir3_reg_create(mov, 0, 0);
if (arrsz == 0)
return NULL;
- collect = ir3_instr_create2(block, -1, OPC_META_FI, 1 + arrsz);
+ collect = ir3_instr_create2(block, OPC_META_FI, 1 + arrsz);
ir3_reg_create(collect, 0, 0); /* dst */
for (unsigned i = 0; i < arrsz; i++)
ir3_reg_create(collect, 0, IR3_REG_SSA)->instr = arr[i];
struct ir3_instruction *mov;
struct ir3_register *src;
- mov = ir3_instr_create(block, 1, 0);
+ mov = ir3_instr_create(block, OPC_MOV);
mov->cat1.src_type = TYPE_U32;
mov->cat1.dst_type = TYPE_U32;
ir3_reg_create(mov, 0, 0);
struct ir3_instruction *mov;
struct ir3_register *src;
- mov = ir3_instr_create(block, 1, 0);
+ mov = ir3_instr_create(block, OPC_MOV);
mov->cat1.src_type = TYPE_U32;
mov->cat1.dst_type = TYPE_U32;
ir3_reg_create(mov, 0, 0);
struct ir3_instruction *mov;
struct ir3_register *dst;
- mov = ir3_instr_create(block, 1, 0);
+ mov = ir3_instr_create(block, OPC_MOV);
mov->cat1.src_type = TYPE_U32;
mov->cat1.dst_type = TYPE_U32;
dst = ir3_reg_create(mov, 0, IR3_REG_ARRAY |
{
struct ir3_instruction *in;
- in = ir3_instr_create(block, -1, OPC_META_INPUT);
+ in = ir3_instr_create(block, OPC_META_INPUT);
in->inout.block = block;
ir3_reg_create(in, n, 0);
}
}
-/* NOTE: this creates the "TGSI" style fragface (ie. input slot
- * VARYING_SLOT_FACE). For NIR style nir_intrinsic_load_front_face
- * we can just use the value from hw directly (since it is boolean)
- */
-static struct ir3_instruction *
-create_frag_face(struct ir3_compile *ctx, unsigned comp)
-{
- struct ir3_block *block = ctx->block;
- struct ir3_instruction *instr;
-
- switch (comp) {
- case 0: /* .x */
- compile_assert(ctx, !ctx->frag_face);
-
- ctx->frag_face = create_input(block, 0);
- ctx->frag_face->regs[0]->flags |= IR3_REG_HALF;
-
- /* for faceness, we always get -1 or 0 (int).. but TGSI expects
- * positive vs negative float.. and piglit further seems to
- * expect -1.0 or 1.0:
- *
- * mul.s tmp, hr0.x, 2
- * add.s tmp, tmp, 1
- * mov.s32f32, dst, tmp
- *
- */
- instr = ir3_MUL_S(block, ctx->frag_face, 0,
- create_immed(block, 2), 0);
- instr = ir3_ADD_S(block, instr, 0,
- create_immed(block, 1), 0);
- instr = ir3_COV(block, instr, TYPE_S32, TYPE_F32);
-
- return instr;
- case 1: /* .y */
- case 2: /* .z */
- return create_immed(block, fui(0.0));
- default:
- case 3: /* .w */
- return create_immed(block, fui(1.0));
- }
-}
-
static struct ir3_instruction *
create_driver_param(struct ir3_compile *ctx, enum ir3_driver_param dp)
{
/* first four vec4 sysval's reserved for UBOs: */
/* NOTE: dp is in scalar, but there can be >4 dp components: */
- unsigned n = ctx->so->first_driver_param + IR3_DRIVER_PARAM_OFF;
+ unsigned n = ctx->so->constbase.driver_param;
unsigned r = regid(n + dp / 4, dp % 4);
return create_uniform(ctx, r);
}
*/
static void
split_dest(struct ir3_block *block, struct ir3_instruction **dst,
- struct ir3_instruction *src, unsigned n)
+ struct ir3_instruction *src, unsigned base, unsigned n)
{
struct ir3_instruction *prev = NULL;
for (int i = 0, j = 0; i < n; i++) {
- struct ir3_instruction *split =
- ir3_instr_create(block, -1, OPC_META_FO);
+ struct ir3_instruction *split = ir3_instr_create(block, OPC_META_FO);
ir3_reg_create(split, 0, IR3_REG_SSA);
ir3_reg_create(split, 0, IR3_REG_SSA)->instr = src;
- split->fo.off = i;
+ split->fo.off = i + base;
if (prev) {
split->cp.left = prev;
}
prev = split;
- if (src->regs[0]->wrmask & (1 << i))
+ if (src->regs[0]->wrmask & (1 << (i + base)))
dst[j++] = split;
}
}
}
switch (alu->op) {
- case nir_op_f2i:
+ case nir_op_f2i32:
dst[0] = ir3_COV(b, src[0], TYPE_F32, TYPE_S32);
break;
- case nir_op_f2u:
+ case nir_op_f2u32:
dst[0] = ir3_COV(b, src[0], TYPE_F32, TYPE_U32);
break;
- case nir_op_i2f:
+ case nir_op_i2f32:
dst[0] = ir3_COV(b, src[0], TYPE_S32, TYPE_F32);
break;
- case nir_op_u2f:
+ case nir_op_u2f32:
dst[0] = ir3_COV(b, src[0], TYPE_U32, TYPE_F32);
break;
case nir_op_imov:
struct ir3_instruction **dst)
{
struct ir3_block *b = ctx->block;
- struct ir3_instruction *addr, *src0, *src1;
+ struct ir3_instruction *base_lo, *base_hi, *addr, *src0, *src1;
nir_const_value *const_offset;
/* UBO addresses are the first driver params: */
- unsigned ubo = regid(ctx->so->first_driver_param + IR3_UBOS_OFF, 0);
+ unsigned ubo = regid(ctx->so->constbase.ubo, 0);
+ const unsigned ptrsz = pointer_size(ctx);
+
int off = 0;
/* First src is ubo index, which could either be an immed or not: */
src0 = get_src(ctx, &intr->src[0])[0];
if (is_same_type_mov(src0) &&
(src0->regs[1]->flags & IR3_REG_IMMED)) {
- addr = create_uniform(ctx, ubo + src0->regs[1]->iim_val);
+ base_lo = create_uniform(ctx, ubo + (src0->regs[1]->iim_val * ptrsz));
+ base_hi = create_uniform(ctx, ubo + (src0->regs[1]->iim_val * ptrsz) + 1);
} else {
- addr = create_uniform_indirect(ctx, ubo, get_addr(ctx, src0));
+ base_lo = create_uniform_indirect(ctx, ubo, get_addr(ctx, src0));
+ base_hi = create_uniform_indirect(ctx, ubo + 1, get_addr(ctx, src0));
}
+ /* note: on 32bit gpu's base_hi is ignored and DCE'd */
+ addr = base_lo;
+
const_offset = nir_src_as_const_value(intr->src[1]);
if (const_offset) {
- off += const_offset->u[0];
+ off += const_offset->u32[0];
} else {
/* For load_ubo_indirect, second src is indirect offset: */
src1 = get_src(ctx, &intr->src[1])[0];
off -= off2;
}
+ if (ptrsz == 2) {
+ struct ir3_instruction *carry;
+
+ /* handle 32b rollover, ie:
+ * if (addr < base_lo)
+ * base_hi++
+ */
+ carry = ir3_CMPS_U(b, addr, 0, base_lo, 0);
+ carry->cat2.condition = IR3_COND_LT;
+ base_hi = ir3_ADD_S(b, base_hi, 0, carry, 0);
+
+ addr = create_collect(b, (struct ir3_instruction*[]){ addr, base_hi }, 2);
+ }
+
for (int i = 0; i < intr->num_components; i++) {
struct ir3_instruction *load =
ir3_LDG(b, addr, 0, create_immed(b, 1), 0);
nir_deref_var *dvar = intr->variables[0];
nir_deref_array *darr = nir_deref_as_array(dvar->deref.child);
struct ir3_array *arr = get_var(ctx, dvar->var);
- struct ir3_instruction *addr, **src;
+ struct ir3_instruction *addr;
+ struct ir3_instruction * const *src;
unsigned wrmask = nir_intrinsic_write_mask(intr);
compile_assert(ctx, dvar->deref.child &&
default:
compile_error(ctx, "Unhandled store deref type: %u\n",
darr->deref_array_type);
- break;
+ return;
}
for (int i = 0; i < intr->num_components; i++) {
so->inputs[n].slot = slot;
so->inputs[n].compmask = 1;
so->inputs[n].regid = r;
- so->inputs[n].interpolate = INTERP_QUALIFIER_FLAT;
+ so->inputs[n].interpolate = INTERP_MODE_FLAT;
so->total_in++;
ctx->ir->ninputs = MAX2(ctx->ir->ninputs, r + 1);
emit_intrinsic(struct ir3_compile *ctx, nir_intrinsic_instr *intr)
{
const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic];
- struct ir3_instruction **dst, **src;
+ struct ir3_instruction **dst;
+ struct ir3_instruction * const *src;
struct ir3_block *b = ctx->block;
nir_const_value *const_offset;
int idx;
idx = nir_intrinsic_base(intr);
const_offset = nir_src_as_const_value(intr->src[0]);
if (const_offset) {
- idx += const_offset->u[0];
+ idx += const_offset->u32[0];
for (int i = 0; i < intr->num_components; i++) {
unsigned n = idx * 4 + i;
dst[i] = create_uniform(ctx, n);
idx = nir_intrinsic_base(intr);
const_offset = nir_src_as_const_value(intr->src[0]);
if (const_offset) {
- idx += const_offset->u[0];
+ idx += const_offset->u32[0];
for (int i = 0; i < intr->num_components; i++) {
unsigned n = idx * 4 + i;
dst[i] = ctx->ir->inputs[n];
idx = nir_intrinsic_base(intr);
const_offset = nir_src_as_const_value(intr->src[1]);
compile_assert(ctx, const_offset != NULL);
- idx += const_offset->u[0];
+ idx += const_offset->u32[0];
src = get_src(ctx, &intr->src[0]);
for (int i = 0; i < intr->num_components; i++) {
dst[0] = ctx->basevertex;
break;
case nir_intrinsic_load_vertex_id_zero_base:
+ case nir_intrinsic_load_vertex_id:
if (!ctx->vertex_id) {
+ gl_system_value sv = (intr->intrinsic == nir_intrinsic_load_vertex_id) ?
+ SYSTEM_VALUE_VERTEX_ID : SYSTEM_VALUE_VERTEX_ID_ZERO_BASE;
ctx->vertex_id = create_input(b, 0);
- add_sysval_input(ctx, SYSTEM_VALUE_VERTEX_ID_ZERO_BASE,
- ctx->vertex_id);
+ add_sysval_input(ctx, sv, ctx->vertex_id);
}
dst[0] = ctx->vertex_id;
break;
ctx->frag_face = create_input(b, 0);
ctx->frag_face->regs[0]->flags |= IR3_REG_HALF;
}
- dst[0] = ir3_ADD_S(b, ctx->frag_face, 0, create_immed(b, 1), 0);
+ /* for fragface, we always get -1 or 0, but that is inverse
+ * of what nir expects (where ~0 is true). Unfortunately
+ * trying to widen from half to full in add.s seems to do a
+ * non-sign-extending widen (resulting in something that
+ * gets interpreted as float Inf??)
+ */
+ dst[0] = ir3_COV(b, ctx->frag_face, TYPE_S16, TYPE_S32);
+ dst[0] = ir3_ADD_S(b, dst[0], 0, create_immed(b, 1), 0);
break;
case nir_intrinsic_discard_if:
case nir_intrinsic_discard: {
struct ir3_instruction **dst = get_dst_ssa(ctx, &instr->def,
instr->def.num_components);
for (int i = 0; i < instr->def.num_components; i++)
- dst[i] = create_immed(ctx->block, instr->value.u[i]);
+ dst[i] = create_immed(ctx->block, instr->value.u32[i]);
}
static void
{
struct ir3_block *b = ctx->block;
struct ir3_instruction **dst, *sam, *src0[12], *src1[4];
- struct ir3_instruction **coord, *lod, *compare, *proj, **off, **ddx, **ddy;
- struct ir3_instruction *const_off[4];
+ struct ir3_instruction * const *coord, * const *off, * const *ddx, * const *ddy;
+ struct ir3_instruction *lod, *compare, *proj;
bool has_bias = false, has_lod = false, has_proj = false, has_off = false;
unsigned i, coords, flags;
unsigned nsrc0 = 0, nsrc1 = 0;
lod = get_src(ctx, &tex->src[i].src)[0];
has_lod = true;
break;
- case nir_tex_src_comparitor: /* shadow comparator */
+ case nir_tex_src_comparator: /* shadow comparator */
compare = get_src(ctx, &tex->src[i].src)[0];
break;
case nir_tex_src_projector:
case nir_texop_query_levels:
case nir_texop_texture_samples:
case nir_texop_samples_identical:
+ case nir_texop_txf_ms_mcs:
compile_error(ctx, "Unhandled NIR tex type: %d\n", tex->op);
return;
}
tex_info(tex, &flags, &coords);
- if (!has_off) {
- /* could still have a constant offset: */
- if (tex->const_offset[0] || tex->const_offset[1] ||
- tex->const_offset[2] || tex->const_offset[3]) {
- off = const_off;
-
- off[0] = create_immed(b, tex->const_offset[0]);
- off[1] = create_immed(b, tex->const_offset[1]);
- off[2] = create_immed(b, tex->const_offset[2]);
- off[3] = create_immed(b, tex->const_offset[3]);
-
- has_off = true;
- }
- }
-
- /* scale up integer coords for TXF based on the LOD */
- if (ctx->unminify_coords && (opc == OPC_ISAML)) {
- assert(has_lod);
- for (i = 0; i < coords; i++)
- coord[i] = ir3_SHL_B(b, coord[i], 0, lod, 0);
- }
-
- /* the array coord for cube arrays needs 0.5 added to it */
- if (tex->sampler_dim == GLSL_SAMPLER_DIM_CUBE && tex->is_array &&
- opc != OPC_ISAML)
- coord[3] = ir3_ADD_F(b, coord[3], 0, create_immed(b, fui(0.5)), 0);
-
/*
* lay out the first argument in the proper order:
* - actual coordinates first
/* insert tex coords: */
for (i = 0; i < coords; i++)
- src0[nsrc0++] = coord[i];
+ src0[i] = coord[i];
+
+ nsrc0 = i;
+
+ /* scale up integer coords for TXF based on the LOD */
+ if (ctx->unminify_coords && (opc == OPC_ISAML)) {
+ assert(has_lod);
+ for (i = 0; i < coords; i++)
+ src0[i] = ir3_SHL_B(b, src0[i], 0, lod, 0);
+ }
if (coords == 1) {
/* hw doesn't do 1d, so we treat it as 2d with
if (tex->is_shadow && tex->op != nir_texop_lod)
src0[nsrc0++] = compare;
- if (tex->is_array && tex->op != nir_texop_lod)
- src0[nsrc0++] = coord[coords];
+ if (tex->is_array && tex->op != nir_texop_lod) {
+ struct ir3_instruction *idx = coord[coords];
+
+ /* the array coord for cube arrays needs 0.5 added to it */
+ if (ctx->array_index_add_half && (opc != OPC_ISAML))
+ idx = ir3_ADD_F(b, idx, 0, create_immed(b, fui(0.5)), 0);
+
+ src0[nsrc0++] = idx;
+ }
if (has_proj) {
src0[nsrc0++] = proj;
if (opc == OPC_GETLOD)
type = TYPE_U32;
- sam = ir3_SAM(b, opc, type, TGSI_WRITEMASK_XYZW,
- flags, tex->texture_index, tex->texture_index,
- create_collect(b, src0, nsrc0),
- create_collect(b, src1, nsrc1));
+ unsigned tex_idx = tex->texture_index;
+
+ ctx->max_texture_index = MAX2(ctx->max_texture_index, tex_idx);
+
+ struct ir3_instruction *col0 = create_collect(b, src0, nsrc0);
+ struct ir3_instruction *col1 = create_collect(b, src1, nsrc1);
+
+ sam = ir3_SAM(b, opc, type, TGSI_WRITEMASK_XYZW, flags,
+ tex_idx, tex_idx, col0, col1);
- split_dest(b, dst, sam, 4);
+ if ((ctx->astc_srgb & (1 << tex_idx)) && !nir_tex_instr_is_query(tex)) {
+ /* only need first 3 components: */
+ sam->regs[0]->wrmask = 0x7;
+ split_dest(b, dst, sam, 0, 3);
+
+ /* we need to sample the alpha separately with a non-ASTC
+ * texture state:
+ */
+ sam = ir3_SAM(b, opc, type, TGSI_WRITEMASK_W, flags,
+ tex_idx, tex_idx, col0, col1);
+
+ array_insert(ctx->ir->astc_srgb, sam);
+
+ /* fixup .w component: */
+ split_dest(b, &dst[3], sam, 3, 1);
+ } else {
+ /* normal (non-workaround) case: */
+ split_dest(b, dst, sam, 0, 4);
+ }
/* GETLOD returns results in 4.8 fixed point */
if (opc == OPC_GETLOD) {
/* even though there is only one component, since it ends
* up in .z rather than .x, we need a split_dest()
*/
- split_dest(b, dst, sam, 3);
+ split_dest(b, dst, sam, 0, 3);
/* The # of levels comes from getinfo.z. We need to add 1 to it, since
* the value in TEX_CONST_0 is zero-based.
emit_tex_txs(struct ir3_compile *ctx, nir_tex_instr *tex)
{
struct ir3_block *b = ctx->block;
- struct ir3_instruction **dst, *sam, *lod;
+ struct ir3_instruction **dst, *sam;
+ struct ir3_instruction *lod;
unsigned flags, coords;
tex_info(tex, &flags, &coords);
sam = ir3_SAM(b, OPC_GETSIZE, TYPE_U32, TGSI_WRITEMASK_XYZW, flags,
tex->texture_index, tex->texture_index, lod, NULL);
- split_dest(b, dst, sam, 4);
+ split_dest(b, dst, sam, 0, 4);
/* Array size actually ends up in .w rather than .z. This doesn't
* matter for miplevel 0, but for higher mips the value in z is
dst = get_dst(ctx, &nphi->dest, 1);
- phi = ir3_instr_create2(ctx->block, -1, OPC_META_PHI,
+ phi = ir3_instr_create2(ctx->block, OPC_META_PHI,
1 + exec_list_length(&nphi->srcs));
ir3_reg_create(phi, 0, 0); /* dst */
phi->phi.nphi = nphi;
nir_phi_instr *nphi;
/* phi's only come at start of block: */
- if (!(is_meta(instr) && (instr->opc == OPC_META_PHI)))
+ if (instr->opc != OPC_META_PHI)
break;
if (!instr->phi.nphi)
foreach_list_typed(nir_phi_src, nsrc, node, &nphi->srcs) {
struct ir3_instruction *src = get_src(ctx, &nsrc->src)[0];
+
+ /* NOTE: src might not be in the same block as it comes from
+ * according to the phi.. but in the end the backend assumes
+ * it will be able to assign the same register to each (which
+ * only works if it is assigned in the src block), so insert
+ * an extra mov to make sure the phi src is assigned in the
+ * block it comes from:
+ */
+ src = ir3_MOV(get_block(ctx, nsrc->pred), src, TYPE_U32);
+
ir3_reg_create(instr, 0, IR3_REG_SSA)->instr = src;
}
}
_mesa_hash_table_destroy(ctx->addr_ht, NULL);
ctx->addr_ht = NULL;
- nir_foreach_instr(nblock, instr) {
+ nir_foreach_instr(instr, nblock) {
emit_instr(ctx, instr);
if (ctx->error)
return;
unsigned stride = strmout->stride[i];
struct ir3_instruction *base, *off;
- base = create_uniform(ctx, regid(v->first_driver_param + IR3_TFBOS_OFF, i));
+ base = create_uniform(ctx, regid(v->constbase.tfbo, i));
/* 24-bit should be enough: */
off = ir3_MUL_U(ctx->block, vtxcnt, 0,
* out, we guarantee that all exit paths flow into the stream-
* out instructions.
*/
- if ((ctx->so->shader->stream_output.num_outputs > 0) &&
+ if ((ctx->compiler->gpu_id < 500) &&
+ (ctx->so->shader->stream_output.num_outputs > 0) &&
!ctx->so->key.binning_pass) {
debug_assert(ctx->so->type == SHADER_VERTEX);
emit_stream_out(ctx);
DBG("; in: slot=%u, len=%ux%u, drvloc=%u",
slot, array_len, ncomp, n);
+ /* let's pretend things other than vec4 don't exist: */
+ ncomp = MAX2(ncomp, 4);
+ compile_assert(ctx, ncomp == 4);
+
so->inputs[n].slot = slot;
so->inputs[n].compmask = (1 << ncomp) - 1;
so->inputs_count = MAX2(so->inputs_count, n + 1);
so->inputs[n].bary = false;
so->frag_coord = true;
instr = create_frag_coord(ctx, i);
- } else if (slot == VARYING_SLOT_FACE) {
- so->inputs[n].bary = false;
- so->frag_face = true;
- instr = create_frag_face(ctx, i);
+ } else if (slot == VARYING_SLOT_PNTC) {
+ /* see for example st_get_generic_varying_index().. this is
+ * maybe a bit mesa/st specific. But we need things to line
+ * up for this in fdN_program:
+ * unsigned texmask = 1 << (slot - VARYING_SLOT_VAR0);
+ * if (emit->sprite_coord_enable & texmask) {
+ * ...
+ * }
+ */
+ so->inputs[n].slot = VARYING_SLOT_VAR8;
+ so->inputs[n].bary = true;
+ instr = create_frag_input(ctx, false);
} else {
bool use_ldlv = false;
* we need to do flat vs smooth shading depending on
* rast state:
*/
- if (in->data.interpolation == INTERP_QUALIFIER_NONE) {
+ if (in->data.interpolation == INTERP_MODE_NONE) {
switch (slot) {
case VARYING_SLOT_COL0:
case VARYING_SLOT_COL1:
}
if (ctx->flat_bypass) {
- if ((so->inputs[n].interpolate == INTERP_QUALIFIER_FLAT) ||
+ if ((so->inputs[n].interpolate == INTERP_MODE_FLAT) ||
(so->inputs[n].rasterflat && ctx->so->key.rasterflat))
use_ldlv = true;
}
instr = create_frag_input(ctx, use_ldlv);
}
+ compile_assert(ctx, idx < ctx->ir->ninputs);
+
ctx->ir->inputs[idx] = instr;
}
} else if (ctx->so->type == SHADER_VERTEX) {
for (int i = 0; i < ncomp; i++) {
unsigned idx = (n * 4) + i;
+ compile_assert(ctx, idx < ctx->ir->ninputs);
ctx->ir->inputs[idx] = create_input(ctx->block, idx);
}
} else {
DBG("; out: slot=%u, len=%ux%u, drvloc=%u",
slot, array_len, ncomp, n);
+ /* let's pretend things other than vec4 don't exist: */
+ ncomp = MAX2(ncomp, 4);
+ compile_assert(ctx, ncomp == 4);
+
if (ctx->so->type == SHADER_FRAGMENT) {
switch (slot) {
case FRAG_RESULT_DEPTH:
case VARYING_SLOT_FOGC:
case VARYING_SLOT_CLIP_DIST0:
case VARYING_SLOT_CLIP_DIST1:
- break;
case VARYING_SLOT_CLIP_VERTEX:
- /* handled entirely in nir_lower_clip: */
- return;
+ break;
default:
if (slot >= VARYING_SLOT_VAR0)
break;
for (int i = 0; i < ncomp; i++) {
unsigned idx = (n * 4) + i;
-
+ compile_assert(ctx, idx < ctx->ir->noutputs);
ctx->ir->outputs[idx] = create_immed(ctx->block, fui(0.0));
}
}
+static int
+max_drvloc(struct exec_list *vars)
+{
+ int drvloc = -1;
+ nir_foreach_variable(var, vars) {
+ drvloc = MAX2(drvloc, (int)var->data.driver_location);
+ }
+ return drvloc;
+}
+
static void
emit_instructions(struct ir3_compile *ctx)
{
unsigned ninputs, noutputs;
- nir_function_impl *fxn = NULL;
-
- /* Find the main function: */
- nir_foreach_function(ctx->s, function) {
- compile_assert(ctx, strcmp(function->name, "main") == 0);
- compile_assert(ctx, function->impl);
- fxn = function->impl;
- break;
- }
+ nir_function_impl *fxn = nir_shader_get_entrypoint(ctx->s);
- ninputs = exec_list_length(&ctx->s->inputs) * 4;
- noutputs = exec_list_length(&ctx->s->outputs) * 4;
+ ninputs = (max_drvloc(&ctx->s->inputs) + 1) * 4;
+ noutputs = (max_drvloc(&ctx->s->outputs) + 1) * 4;
/* or vtx shaders, we need to leave room for sysvals:
*/
if (ctx->so->type == SHADER_VERTEX) {
- ninputs += 8;
+ ninputs += 16;
}
ctx->ir = ir3_create(ctx->compiler, ninputs, noutputs);
list_addtail(&ctx->block->node, &ctx->ir->block_list);
if (ctx->so->type == SHADER_VERTEX) {
- ctx->ir->ninputs -= 8;
+ ctx->ir->ninputs -= 16;
}
/* for fragment shader, we have a single input register (usually
if (ctx->so->type == SHADER_FRAGMENT) {
// TODO maybe a helper for fi since we need it a few places..
struct ir3_instruction *instr;
- instr = ir3_instr_create(ctx->block, -1, OPC_META_FI);
+ instr = ir3_instr_create(ctx->block, OPC_META_FI);
ir3_reg_create(instr, 0, 0);
ir3_reg_create(instr, 0, IR3_REG_SSA); /* r0.x */
ir3_reg_create(instr, 0, IR3_REG_SSA); /* r0.y */
ir->inputs = inputs;
}
+/* Fixup tex sampler state for astc/srgb workaround instructions. We
+ * need to assign the tex state indexes for these after we know the
+ * max tex index.
+ */
+static void
+fixup_astc_srgb(struct ir3_compile *ctx)
+{
+ struct ir3_shader_variant *so = ctx->so;
+ /* indexed by original tex idx, value is newly assigned alpha sampler
+ * state tex idx. Zero is invalid since there is at least one sampler
+ * if we get here.
+ */
+ unsigned alt_tex_state[16] = {0};
+ unsigned tex_idx = ctx->max_texture_index + 1;
+ unsigned idx = 0;
+
+ so->astc_srgb.base = tex_idx;
+
+ for (unsigned i = 0; i < ctx->ir->astc_srgb_count; i++) {
+ struct ir3_instruction *sam = ctx->ir->astc_srgb[i];
+
+ compile_assert(ctx, sam->cat5.tex < ARRAY_SIZE(alt_tex_state));
+
+ if (alt_tex_state[sam->cat5.tex] == 0) {
+ /* assign new alternate/alpha tex state slot: */
+ alt_tex_state[sam->cat5.tex] = tex_idx++;
+ so->astc_srgb.orig_idx[idx++] = sam->cat5.tex;
+ so->astc_srgb.count++;
+ }
+
+ sam->cat5.tex = alt_tex_state[sam->cat5.tex];
+ }
+}
+
int
ir3_compile_shader_nir(struct ir3_compiler *compiler,
struct ir3_shader_variant *so)
if (so->key.half_precision) {
for (i = 0; i < ir->noutputs; i++) {
struct ir3_instruction *out = ir->outputs[i];
+
if (!out)
continue;
+
+ /* if frag shader writes z, that needs to be full precision: */
+ if (so->outputs[i/4].slot == FRAG_RESULT_DEPTH)
+ continue;
+
out->regs[0]->flags |= IR3_REG_HALF;
/* output could be a fanout (ie. texture fetch output)
* in which case we need to propagate the half-reg flag
* up to the definer so that RA sees it:
*/
- if (is_meta(out) && (out->opc == OPC_META_FO)) {
+ if (out->opc == OPC_META_FO) {
out = out->regs[1]->instr;
out->regs[0]->flags |= IR3_REG_HALF;
}
- if (out->category == 1) {
+ if (out->opc == OPC_MOV) {
out->cat1.dst_type = half_type(out->cat1.dst_type);
}
}
ir3_print(ir);
}
- ir3_cp(ir);
+ ir3_cp(ir, so);
if (fd_mesa_debug & FD_DBG_OPTMSGS) {
printf("BEFORE GROUPING:\n");
/* fixup input/outputs: */
for (i = 0; i < so->outputs_count; i++) {
so->outputs[i].regid = ir->outputs[i*4]->regs[0]->num;
- /* preserve hack for depth output.. tgsi writes depth to .z,
- * but what we give the hw is the scalar register:
- */
- if ((so->type == SHADER_FRAGMENT) &&
- (so->outputs[i].slot == FRAG_RESULT_DEPTH))
- so->outputs[i].regid += 2;
}
/* Note that some or all channels of an input may be unused: */
actual_in = 0;
inloc = 0;
for (i = 0; i < so->inputs_count; i++) {
- unsigned j, regid = ~0, compmask = 0;
+ unsigned j, regid = ~0, compmask = 0, maxcomp = 0;
so->inputs[i].ncomp = 0;
- so->inputs[i].inloc = inloc + 8;
+ so->inputs[i].inloc = inloc;
for (j = 0; j < 4; j++) {
struct ir3_instruction *in = inputs[(i*4) + j];
if (in && !(in->flags & IR3_INSTR_UNUSED)) {
if ((so->type == SHADER_FRAGMENT) && so->inputs[i].bary) {
/* assign inloc: */
assert(in->regs[1]->flags & IR3_REG_IMMED);
- in->regs[1]->iim_val = inloc++;
+ in->regs[1]->iim_val = inloc + j;
+ maxcomp = j + 1;
}
}
}
- if ((so->type == SHADER_FRAGMENT) && compmask && so->inputs[i].bary)
+ if ((so->type == SHADER_FRAGMENT) && compmask && so->inputs[i].bary) {
so->varying_in++;
+ so->inputs[i].compmask = (1 << maxcomp) - 1;
+ inloc += maxcomp;
+ } else {
+ so->inputs[i].compmask = compmask;
+ }
so->inputs[i].regid = regid;
- so->inputs[i].compmask = compmask;
}
+ if (ctx->astc_srgb)
+ fixup_astc_srgb(ctx);
+
/* We need to do legalize after (for frag shader's) the "bary.f"
* offsets (inloc) have been assigned.
*/