* Rob Clark <robclark@freedesktop.org>
*/
-#include "util/u_math.h"
-
#include "ir3_compiler.h"
#include "ir3_context.h"
+#include "ir3_image.h"
#include "ir3_shader.h"
#include "ir3_nir.h"
}
}
+ if (compiler->gpu_id >= 600) {
+ ctx->funcs = &ir3_a6xx_funcs;
+ } else if (compiler->gpu_id >= 400) {
+ ctx->funcs = &ir3_a4xx_funcs;
+ }
+
ctx->compiler = compiler;
ctx->so = so;
ctx->def_ht = _mesa_hash_table_create(ctx,
*/
NIR_PASS_V(ctx->s, nir_lower_bool_to_int32);
NIR_PASS_V(ctx->s, nir_lower_locals_to_regs);
+
+ /* We want to lower nir_op_imul as late as possible, to catch also
+ * those generated by earlier passes (e.g, nir_lower_locals_to_regs).
+ * However, we want a final swing of a few passes to have a chance
+ * at optimizing the result.
+ */
+ bool progress = false;
+ NIR_PASS(progress, ctx->s, ir3_nir_lower_imul);
+ if (progress) {
+ NIR_PASS_V(ctx->s, nir_opt_algebraic);
+ NIR_PASS_V(ctx->s, nir_opt_copy_prop_vars);
+ NIR_PASS_V(ctx->s, nir_opt_dead_write_vars);
+ NIR_PASS_V(ctx->s, nir_opt_dce);
+ NIR_PASS_V(ctx->s, nir_opt_constant_folding);
+ }
+
NIR_PASS_V(ctx->s, nir_convert_from_ssa, true);
if (ir3_shader_debug & IR3_DBG_DISASM) {
nir_print_shader(ctx->s, stderr);
}
- ir3_nir_scan_driver_consts(ctx->s, &so->const_layout);
-
- so->num_uniforms = ctx->s->num_uniforms;
- so->num_ubos = ctx->s->info.num_ubos;
-
- /* Layout of constant registers, each section aligned to vec4. Note
- * that pointer size (ubo, etc) changes depending on generation.
- *
- * user consts
- * UBO addresses
- * SSBO sizes
- * if (vertex shader) {
- * driver params (IR3_DP_*)
- * if (stream_output.num_outputs > 0)
- * stream-out addresses
- * }
- * immediates
- *
- * 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 = ir3_pointer_size(ctx);
-
- memset(&so->constbase, ~0, sizeof(so->constbase));
-
- if (so->num_ubos > 0) {
- so->constbase.ubo = constoff;
- constoff += align(ctx->s->info.num_ubos * ptrsz, 4) / 4;
- }
-
- if (so->const_layout.ssbo_size.count > 0) {
- unsigned cnt = so->const_layout.ssbo_size.count;
- so->constbase.ssbo_sizes = constoff;
- constoff += align(cnt, 4) / 4;
- }
-
- if (so->const_layout.image_dims.count > 0) {
- unsigned cnt = so->const_layout.image_dims.count;
- so->constbase.image_dims = constoff;
- constoff += align(cnt, 4) / 4;
- }
-
- unsigned num_driver_params = 0;
- if (so->type == MESA_SHADER_VERTEX) {
- num_driver_params = IR3_DP_VS_COUNT;
- } else if (so->type == MESA_SHADER_COMPUTE) {
- num_driver_params = IR3_DP_CS_COUNT;
- }
-
- so->constbase.driver_param = constoff;
- constoff += align(num_driver_params, 4) / 4;
-
- if ((so->type == MESA_SHADER_VERTEX) &&
- (compiler->gpu_id < 500) &&
- so->shader->stream_output.num_outputs > 0) {
- so->constbase.tfbo = constoff;
- constoff += align(IR3_MAX_SO_BUFFERS * ptrsz, 4) / 4;
- }
-
- so->constbase.immediate = constoff;
+ ir3_ibo_mapping_init(&so->image_mapping, ctx->s->info.num_textures);
return ctx;
}
for (unsigned i = 0; i < num_components; i++) {
unsigned n = src->reg.base_offset * reg->num_components + i;
compile_assert(ctx, n < arr->length);
- value[i] = ir3_create_array_load(ctx, arr, n, addr);
+ value[i] = ir3_create_array_load(ctx, arr, n, addr, reg->bit_size);
}
return value;
}
void
-put_dst(struct ir3_context *ctx, nir_dest *dst)
+ir3_put_dst(struct ir3_context *ctx, nir_dest *dst)
{
unsigned bit_size = nir_dest_bit_size(*dst);
+ /* add extra mov if dst value is HIGH reg.. in some cases not all
+ * instructions can read from HIGH regs, in cases where they can
+ * ir3_cp will clean up the extra mov:
+ */
+ for (unsigned i = 0; i < ctx->last_dst_n; i++) {
+ if (!ctx->last_dst[i])
+ continue;
+ if (ctx->last_dst[i]->regs[0]->flags & IR3_REG_HIGH) {
+ ctx->last_dst[i] = ir3_MOV(ctx->block, ctx->last_dst[i], TYPE_U32);
+ }
+ }
+
if (bit_size < 32) {
for (unsigned i = 0; i < ctx->last_dst_n; i++) {
struct ir3_instruction *dst = ctx->last_dst[i];
ralloc_free(ctx->last_dst);
}
+
ctx->last_dst = NULL;
ctx->last_dst_n = 0;
}
ir3_reg_create(collect, 0, IR3_REG_SSA | flags)->instr = elem;
}
+ collect->regs[0]->wrmask = MASK(arrsz);
+
return collect;
}
return;
}
+ unsigned flags = src->regs[0]->flags & (IR3_REG_HALF | IR3_REG_HIGH);
+
for (int i = 0, j = 0; i < n; i++) {
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;
+ ir3_reg_create(split, 0, IR3_REG_SSA | flags);
+ ir3_reg_create(split, 0, IR3_REG_SSA | flags)->instr = src;
split->fo.off = i + base;
if (prev) {
}
}
-void
+NORETURN void
ir3_context_error(struct ir3_context *ctx, const char *format, ...)
{
struct hash_table *errors = NULL;
nir_print_shader_annotated(ctx->s, stdout, errors);
ralloc_free(errors);
ctx->error = true;
- debug_assert(0);
+ unreachable("");
}
static struct ir3_instruction *
/* relative (indirect) if address!=NULL */
struct ir3_instruction *
ir3_create_array_load(struct ir3_context *ctx, struct ir3_array *arr, int n,
- struct ir3_instruction *address)
+ struct ir3_instruction *address, unsigned bitsize)
{
struct ir3_block *block = ctx->block;
struct ir3_instruction *mov;
struct ir3_register *src;
+ unsigned flags = 0;
mov = ir3_instr_create(block, OPC_MOV);
- mov->cat1.src_type = TYPE_U32;
- mov->cat1.dst_type = TYPE_U32;
+ if (bitsize < 32) {
+ mov->cat1.src_type = TYPE_U16;
+ mov->cat1.dst_type = TYPE_U16;
+ flags |= IR3_REG_HALF;
+ } else {
+ mov->cat1.src_type = TYPE_U32;
+ mov->cat1.dst_type = TYPE_U32;
+ }
+
mov->barrier_class = IR3_BARRIER_ARRAY_R;
mov->barrier_conflict = IR3_BARRIER_ARRAY_W;
- ir3_reg_create(mov, 0, 0);
+ ir3_reg_create(mov, 0, flags);
src = ir3_reg_create(mov, 0, IR3_REG_ARRAY |
- COND(address, IR3_REG_RELATIV));
+ COND(address, IR3_REG_RELATIV) | flags);
src->instr = arr->last_write;
src->size = arr->length;
src->array.id = arr->id;