struct ac_shader_abi *abi;
gl_shader_stage stage;
+ shader_info *info;
LLVMValueRef *ssa_defs;
enum ac_image_dim dim = get_ac_sampler_dim(ctx, sdim, is_array);
if (dim == ac_image_cube ||
- (ctx->chip_class <= VI && dim == ac_image_3d))
+ (ctx->chip_class <= GFX8 && dim == ac_image_3d))
dim = ac_image_2darray;
return dim;
src0 = ac_to_float(ctx, src0);
result = LLVMBuildFPTrunc(ctx->builder, src0, ctx->f16, "");
- if (ctx->chip_class >= VI) {
+ if (ctx->chip_class >= GFX8) {
LLVMValueRef args[2];
/* Check if the result is a denormal - and flush to 0 if so. */
args[0] = result;
/* need to convert back up to f32 */
result = LLVMBuildFPExt(ctx->builder, result, ctx->f32, "");
- if (ctx->chip_class >= VI)
+ if (ctx->chip_class >= GFX8)
result = LLVMBuildSelect(ctx->builder, cond, ctx->f32_0, result, "");
else {
- /* for SI/CIK */
+ /* for GFX6-GFX7 */
/* 0x38800000 is smallest half float value (2^-14) in 32-bit float,
* so compare the result and flush to 0 if it's smaller.
*/
src[i] = get_alu_src(ctx, instr->src[i], src_components);
switch (instr->op) {
- case nir_op_fmov:
- case nir_op_imov:
+ case nir_op_mov:
result = src[0];
break;
case nir_op_fneg:
case nir_op_i2f16:
case nir_op_i2f32:
case nir_op_i2f64:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
result = LLVMBuildSIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
break;
case nir_op_u2f16:
case nir_op_u2f32:
case nir_op_u2f64:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
result = LLVMBuildUIToFP(ctx->ac.builder, src[0], ac_to_float_type(&ctx->ac, def_type), "");
break;
case nir_op_f2f16_rtz:
case nir_op_u2u16:
case nir_op_u2u32:
case nir_op_u2u64:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
result = LLVMBuildZExt(ctx->ac.builder, src[0], def_type, "");
else
case nir_op_i2i16:
case nir_op_i2i32:
case nir_op_i2i64:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
if (ac_get_elem_bits(&ctx->ac, LLVMTypeOf(src[0])) < ac_get_elem_bits(&ctx->ac, def_type))
result = LLVMBuildSExt(ctx->ac.builder, src[0], def_type, "");
else
result = emit_bcsel(&ctx->ac, src[0], src[1], src[2]);
break;
case nir_op_find_lsb:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
result = ac_find_lsb(&ctx->ac, ctx->ac.i32, src[0]);
break;
case nir_op_ufind_msb:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
result = ac_build_umsb(&ctx->ac, src[0], ctx->ac.i32);
break;
case nir_op_ifind_msb:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
result = ac_build_imsb(&ctx->ac, src[0], ctx->ac.i32);
break;
case nir_op_uadd_carry:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
- src[1] = ac_to_integer(&ctx->ac, src[1]);
result = emit_uint_carry(&ctx->ac, "llvm.uadd.with.overflow.i32", src[0], src[1]);
break;
case nir_op_usub_borrow:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
- src[1] = ac_to_integer(&ctx->ac, src[1]);
result = emit_uint_carry(&ctx->ac, "llvm.usub.with.overflow.i32", src[0], src[1]);
break;
case nir_op_b2f16:
result = emit_b2i(&ctx->ac, src[0], instr->dest.dest.ssa.bit_size);
break;
case nir_op_i2b32:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
result = emit_i2b(&ctx->ac, src[0]);
break;
case nir_op_fquantize2f16:
result = emit_f2f16(&ctx->ac, src[0]);
break;
case nir_op_umul_high:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
- src[1] = ac_to_integer(&ctx->ac, src[1]);
result = emit_umul_high(&ctx->ac, src[0], src[1]);
break;
case nir_op_imul_high:
- src[0] = ac_to_integer(&ctx->ac, src[0]);
- src[1] = ac_to_integer(&ctx->ac, src[1]);
result = emit_imul_high(&ctx->ac, src[0], src[1]);
break;
case nir_op_pack_half_2x16:
}
case nir_op_pack_64_2x32_split: {
- LLVMValueRef tmp = LLVMGetUndef(ctx->ac.v2i32);
- tmp = ac_build_gather_values(&ctx->ac, src, 2);
+ LLVMValueRef tmp = ac_build_gather_values(&ctx->ac, src, 2);
result = LLVMBuildBitCast(ctx->ac.builder, tmp, ctx->ac.i64, "");
break;
}
LLVMValueRef in[3];
for (unsigned chan = 0; chan < 3; chan++)
in[chan] = ac_llvm_extract_elem(&ctx->ac, src[0], chan);
- results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
+ results[0] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
- results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubesc",
+ results[1] = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubetc",
ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
+ LLVMValueRef ma = ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.cubema",
+ ctx->ac.f32, in, 3, AC_FUNC_ATTR_READNONE);
+ results[0] = ac_build_fdiv(&ctx->ac, results[0], ma);
+ results[1] = ac_build_fdiv(&ctx->ac, results[1], ma);
+ LLVMValueRef offset = LLVMConstReal(ctx->ac.f32, 0.5);
+ results[0] = LLVMBuildFAdd(ctx->ac.builder, results[0], offset, "");
+ results[1] = LLVMBuildFAdd(ctx->ac.builder, results[1], offset, "");
result = ac_build_gather_values(&ctx->ac, results, 2);
break;
}
switch (instr->def.bit_size) {
case 8:
values[i] = LLVMConstInt(element_type,
- instr->value.u8[i], false);
+ instr->value[i].u8, false);
break;
case 16:
values[i] = LLVMConstInt(element_type,
- instr->value.u16[i], false);
+ instr->value[i].u16, false);
break;
case 32:
values[i] = LLVMConstInt(element_type,
- instr->value.u32[i], false);
+ instr->value[i].u32, false);
break;
case 64:
values[i] = LLVMConstInt(element_type,
- instr->value.u64[i], false);
+ instr->value[i].u64, false);
break;
default:
fprintf(stderr,
LLVMBuildExtractElement(ctx->ac.builder, descriptor,
LLVMConstInt(ctx->ac.i32, 2, false), "");
- /* VI only */
- if (ctx->ac.chip_class == VI && in_elements) {
- /* On VI, the descriptor contains the size in bytes,
+ /* GFX8 only */
+ if (ctx->ac.chip_class == GFX8 && in_elements) {
+ /* On GFX8, the descriptor contains the size in bytes,
* but TXQ must return the size in elements.
* The stride is always non-zero for resources using TXQ.
*/
LLVMConstInt(ctx->i32, 0x14000000, false), "");
/* replace the NUM FORMAT in the descriptor */
- tmp2 = LLVMBuildAnd(ctx->builder, tmp2, LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT_GFX6, false), "");
+ tmp2 = LLVMBuildAnd(ctx->builder, tmp2, LLVMConstInt(ctx->i32, C_008F14_NUM_FORMAT, false), "");
tmp2 = LLVMBuildOr(ctx->builder, tmp2, tmp, "");
args->resource = LLVMBuildInsertElement(ctx->builder, args->resource, tmp2, ctx->i32_1, "");
break;
}
- if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= VI) {
+ if (instr->op == nir_texop_tg4 && ctx->ac.chip_class <= GFX8) {
nir_deref_instr *texture_deref_instr = get_tex_texture_deref(instr);
nir_variable *var = nir_deref_instr_get_variable(texture_deref_instr);
const struct glsl_type *type = glsl_without_array(var->type);
}
args->attributes = AC_FUNC_ATTR_READNONE;
+ bool cs_derivs = ctx->stage == MESA_SHADER_COMPUTE &&
+ ctx->info->cs.derivative_group != DERIVATIVE_GROUP_NONE;
+ if (ctx->stage == MESA_SHADER_FRAGMENT || cs_derivs) {
+ /* Prevent texture instructions with implicit derivatives from being
+ * sinked into branches. */
+ switch (instr->op) {
+ case nir_texop_tex:
+ case nir_texop_txb:
+ case nir_texop_lod:
+ args->attributes |= AC_FUNC_ATTR_CONVERGENT;
+ break;
+ default:
+ break;
+ }
+ }
+
return ac_build_image_opcode(&ctx->ac, args);
}
}
}
- ptr = ac_build_gep0(&ctx->ac, ctx->abi->push_constants, addr);
+ ptr = LLVMBuildGEP(ctx->ac.builder, ctx->abi->push_constants, &addr, 1, "");
if (instr->dest.ssa.bit_size == 8) {
unsigned load_dwords = instr->dest.ssa.num_components > 1 ? 2 : 1;
{
unsigned cache_policy = 0;
- /* SI has a TC L1 bug causing corruption of 8bit/16bit stores. All
+ /* GFX6 has a TC L1 bug causing corruption of 8bit/16bit stores. All
* store opcodes not aligned to a dword are affected. The only way to
* get unaligned stores is through shader images.
*/
- if (((may_store_unaligned && ctx->ac.chip_class == SI) ||
+ if (((may_store_unaligned && ctx->ac.chip_class == GFX6) ||
/* If this is write-only, don't keep data in L1 to prevent
* evicting L1 cache lines that may be needed by other
* instructions.
u_bit_scan_consecutive_range(&writemask, &start, &count);
- /* Due to an LLVM limitation, split 3-element writes
- * into a 2-element and a 1-element write. */
- if (count == 3) {
+ /* Due to an LLVM limitation with LLVM < 9, split 3-element
+ * writes into a 2-element and a 1-element write. */
+ if (count == 3 &&
+ (elem_size_bytes != 4 || !ac_has_vec3_support(ctx->ac.chip_class, false))) {
writemask |= 1 << (start + 2);
count = 2;
}
case 16: /* v4f32 */
data_type = ctx->ac.v4f32;
break;
+ case 12: /* v3f32 */
+ data_type = ctx->ac.v3f32;
+ break;
case 8: /* v2f32 */
data_type = ctx->ac.v2f32;
break;
static LLVMValueRef visit_atomic_ssbo(struct ac_nir_context *ctx,
const nir_intrinsic_instr *instr)
{
+ LLVMTypeRef return_type = LLVMTypeOf(get_src(ctx, instr->src[2]));
const char *op;
- char name[64];
+ char name[64], type[8];
LLVMValueRef params[6];
int arg_count = 0;
- if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
- params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
- }
- params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
- params[arg_count++] = ctx->abi->load_ssbo(ctx->abi,
- get_src(ctx, instr->src[0]),
- true);
- params[arg_count++] = ctx->ac.i32_0; /* vindex */
- params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
- params[arg_count++] = ctx->ac.i1false; /* slc */
-
switch (instr->intrinsic) {
case nir_intrinsic_ssbo_atomic_add:
op = "add";
abort();
}
- if (HAVE_LLVM >= 0x900 &&
- instr->intrinsic != nir_intrinsic_ssbo_atomic_comp_swap) {
+ if (instr->intrinsic == nir_intrinsic_ssbo_atomic_comp_swap) {
+ params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[3]), 0);
+ }
+ params[arg_count++] = ac_llvm_extract_elem(&ctx->ac, get_src(ctx, instr->src[2]), 0);
+ params[arg_count++] = ctx->abi->load_ssbo(ctx->abi,
+ get_src(ctx, instr->src[0]),
+ true);
+
+ if (HAVE_LLVM >= 0x900) {
+ /* XXX: The new raw/struct atomic intrinsics are buggy with
+ * LLVM 8, see r358579.
+ */
+ params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
+ params[arg_count++] = ctx->ac.i32_0; /* soffset */
+ params[arg_count++] = ctx->ac.i32_0; /* slc */
+
+ ac_build_type_name_for_intr(return_type, type, sizeof(type));
snprintf(name, sizeof(name),
- "llvm.amdgcn.buffer.atomic.%s.i32", op);
+ "llvm.amdgcn.raw.buffer.atomic.%s.%s", op, type);
} else {
+ params[arg_count++] = ctx->ac.i32_0; /* vindex */
+ params[arg_count++] = get_src(ctx, instr->src[1]); /* voffset */
+ params[arg_count++] = ctx->ac.i1false; /* slc */
+
+ assert(return_type == ctx->ac.i32);
snprintf(name, sizeof(name),
"llvm.amdgcn.buffer.atomic.%s", op);
}
- return ac_build_intrinsic(&ctx->ac, name, ctx->ac.i32, params, arg_count, 0);
+ return ac_build_intrinsic(&ctx->ac, name, return_type, params,
+ arg_count, 0);
}
static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
if (vertex_index_out)
*vertex_index_out = 0;
} else {
- nir_const_value *v = nir_src_as_const_value(path.path[idx_lvl]->arr.index);
- assert(v);
- *vertex_index_out = v->u32[0];
+ *vertex_index_out = nir_src_as_uint(path.path[idx_lvl]->arr.index);
}
++idx_lvl;
}
if (var->data.compact) {
assert(instr->deref_type == nir_deref_type_array);
- nir_const_value *v = nir_src_as_const_value(instr->arr.index);
- assert(v);
- const_offset = v->u32[0];
+ const_offset = nir_src_as_uint(instr->arr.index);
goto out;
}
return 0;
}
-
-/* Adjust the sample index according to FMASK.
- *
- * For uncompressed MSAA surfaces, FMASK should return 0x76543210,
- * which is the identity mapping. Each nibble says which physical sample
- * should be fetched to get that sample.
- *
- * For example, 0x11111100 means there are only 2 samples stored and
- * the second sample covers 3/4 of the pixel. When reading samples 0
- * and 1, return physical sample 0 (determined by the first two 0s
- * in FMASK), otherwise return physical sample 1.
- *
- * The sample index should be adjusted as follows:
- * sample_index = (fmask >> (sample_index * 4)) & 0xF;
- */
static LLVMValueRef adjust_sample_index_using_fmask(struct ac_llvm_context *ctx,
LLVMValueRef coord_x, LLVMValueRef coord_y,
LLVMValueRef coord_z,
LLVMValueRef sample_index,
LLVMValueRef fmask_desc_ptr)
{
- struct ac_image_args args = {0};
- LLVMValueRef res;
+ unsigned sample_chan = coord_z ? 3 : 2;
+ LLVMValueRef addr[4] = {coord_x, coord_y, coord_z};
+ addr[sample_chan] = sample_index;
- args.coords[0] = coord_x;
- args.coords[1] = coord_y;
- if (coord_z)
- args.coords[2] = coord_z;
-
- args.opcode = ac_image_load;
- args.dim = coord_z ? ac_image_2darray : ac_image_2d;
- args.resource = fmask_desc_ptr;
- args.dmask = 0xf;
- args.attributes = AC_FUNC_ATTR_READNONE;
-
- res = ac_build_image_opcode(ctx, &args);
-
- res = ac_to_integer(ctx, res);
- LLVMValueRef four = LLVMConstInt(ctx->i32, 4, false);
- LLVMValueRef F = LLVMConstInt(ctx->i32, 0xf, false);
-
- LLVMValueRef fmask = LLVMBuildExtractElement(ctx->builder,
- res,
- ctx->i32_0, "");
-
- LLVMValueRef sample_index4 =
- LLVMBuildMul(ctx->builder, sample_index, four, "");
- LLVMValueRef shifted_fmask =
- LLVMBuildLShr(ctx->builder, fmask, sample_index4, "");
- LLVMValueRef final_sample =
- LLVMBuildAnd(ctx->builder, shifted_fmask, F, "");
-
- /* Don't rewrite the sample index if WORD1.DATA_FORMAT of the FMASK
- * resource descriptor is 0 (invalid),
- */
- LLVMValueRef fmask_desc =
- LLVMBuildBitCast(ctx->builder, fmask_desc_ptr,
- ctx->v8i32, "");
-
- LLVMValueRef fmask_word1 =
- LLVMBuildExtractElement(ctx->builder, fmask_desc,
- ctx->i32_1, "");
-
- LLVMValueRef word1_is_nonzero =
- LLVMBuildICmp(ctx->builder, LLVMIntNE,
- fmask_word1, ctx->i32_0, "");
-
- /* Replace the MSAA sample index. */
- sample_index =
- LLVMBuildSelect(ctx->builder, word1_is_nonzero,
- final_sample, sample_index, "");
- return sample_index;
+ ac_apply_fmask_to_sample(ctx, fmask_desc_ptr, addr, coord_z != NULL);
+ return addr[sample_chan];
}
static nir_deref_instr *get_image_deref(const nir_intrinsic_instr *instr)
}
static LLVMValueRef get_image_buffer_descriptor(struct ac_nir_context *ctx,
- const nir_intrinsic_instr *instr, bool write)
+ const nir_intrinsic_instr *instr,
+ bool write, bool atomic)
{
LLVMValueRef rsrc = get_image_descriptor(ctx, instr, AC_DESC_BUFFER, write);
- if (ctx->abi->gfx9_stride_size_workaround) {
+ if (ctx->abi->gfx9_stride_size_workaround ||
+ (ctx->abi->gfx9_stride_size_workaround_for_atomic && atomic)) {
LLVMValueRef elem_count = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 2, 0), "");
LLVMValueRef stride = LLVMBuildExtractElement(ctx->ac.builder, rsrc, LLVMConstInt(ctx->ac.i32, 1, 0), "");
stride = LLVMBuildLShr(ctx->ac.builder, stride, LLVMConstInt(ctx->ac.i32, 16, 0), "");
unsigned num_channels = util_last_bit(mask);
LLVMValueRef rsrc, vindex;
- rsrc = get_image_buffer_descriptor(ctx, instr, false);
+ rsrc = get_image_buffer_descriptor(ctx, instr, false, false);
vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
ctx->ac.i32_0, "");
res = ac_build_image_opcode(&ctx->ac, &args);
}
- return ac_to_integer(&ctx->ac, res);
+ return res;
}
static void visit_image_store(struct ac_nir_context *ctx,
args.cache_policy = get_cache_policy(ctx, access, true, writeonly_memory);
if (dim == GLSL_SAMPLER_DIM_BUF) {
- LLVMValueRef rsrc = get_image_buffer_descriptor(ctx, instr, true);
+ LLVMValueRef rsrc = get_image_buffer_descriptor(ctx, instr, true, false);
LLVMValueRef src = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
unsigned src_channels = ac_get_llvm_num_components(src);
LLVMValueRef vindex;
ac_build_buffer_store_format(&ctx->ac, rsrc, src, vindex,
ctx->ac.i32_0, src_channels,
- args.cache_policy & ac_glc,
+ args.cache_policy & ac_glc, false,
writeonly_memory);
} else {
args.opcode = ac_image_store;
MAYBE_UNUSED int length;
enum glsl_sampler_dim dim;
- bool is_unsigned;
+ bool is_unsigned = false;
bool is_array;
if (bindless) {
- if (instr->intrinsic == nir_intrinsic_image_atomic_min ||
- instr->intrinsic == nir_intrinsic_image_atomic_max) {
+ if (instr->intrinsic == nir_intrinsic_bindless_image_atomic_min ||
+ instr->intrinsic == nir_intrinsic_bindless_image_atomic_max) {
const GLenum format = nir_intrinsic_format(instr);
assert(format == GL_R32UI || format == GL_R32I);
is_unsigned = format == GL_R32UI;
params[param_count++] = get_src(ctx, instr->src[3]);
if (dim == GLSL_SAMPLER_DIM_BUF) {
- params[param_count++] = get_image_buffer_descriptor(ctx, instr, true);
+ params[param_count++] = get_image_buffer_descriptor(ctx, instr, true, true);
params[param_count++] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
ctx->ac.i32_0, ""); /* vindex */
params[param_count++] = ctx->ac.i32_0; /* voffset */
- if (HAVE_LLVM >= 0x800) {
+ if (HAVE_LLVM >= 0x900) {
+ /* XXX: The new raw/struct atomic intrinsics are buggy
+ * with LLVM 8, see r358579.
+ */
params[param_count++] = ctx->ac.i32_0; /* soffset */
params[param_count++] = ctx->ac.i32_0; /* slc */
void ac_emit_barrier(struct ac_llvm_context *ac, gl_shader_stage stage)
{
- /* SI only (thanks to a hw bug workaround):
+ /* GFX6 only (thanks to a hw bug workaround):
* The real barrier instruction isn’t needed, because an entire patch
* always fits into a single wave.
*/
- if (ac->chip_class == SI && stage == MESA_SHADER_TESS_CTRL) {
+ if (ac->chip_class == GFX6 && stage == MESA_SHADER_TESS_CTRL) {
ac_build_waitcnt(ac, LGKM_CNT & VM_CNT);
return;
}
LLVMValueRef result;
LLVMValueRef src = get_src(ctx, instr->src[src_idx]);
+ const char *sync_scope = HAVE_LLVM >= 0x0900 ? "workgroup-one-as" : "workgroup";
+
if (instr->intrinsic == nir_intrinsic_shared_atomic_comp_swap ||
instr->intrinsic == nir_intrinsic_deref_atomic_comp_swap) {
LLVMValueRef src1 = get_src(ctx, instr->src[src_idx + 1]);
- result = LLVMBuildAtomicCmpXchg(ctx->ac.builder,
- ptr, src, src1,
- LLVMAtomicOrderingSequentiallyConsistent,
- LLVMAtomicOrderingSequentiallyConsistent,
- false);
+ result = ac_build_atomic_cmp_xchg(&ctx->ac, ptr, src, src1, sync_scope);
result = LLVMBuildExtractValue(ctx->ac.builder, result, 0, "");
} else {
LLVMAtomicRMWBinOp op;
return NULL;
}
- result = LLVMBuildAtomicRMW(ctx->ac.builder, op, ptr, ac_to_integer(&ctx->ac, src),
- LLVMAtomicOrderingSequentiallyConsistent,
- false);
+ result = ac_build_atomic_rmw(&ctx->ac, op, ptr, ac_to_integer(&ctx->ac, src), sync_scope);
}
return result;
}
unsigned array_size = glsl_count_attribute_slots(deref_instr->type, false);
LLVMValueRef offset;
- nir_const_value *const_value = nir_src_as_const_value(deref_instr->arr.index);
- if (const_value) {
- offset = LLVMConstInt(ctx->ac.i32, array_size * const_value->u32[0], false);
+ if (nir_src_is_const(deref_instr->arr.index)) {
+ offset = LLVMConstInt(ctx->ac.i32, array_size * nir_src_as_uint(deref_instr->arr.index), false);
} else {
LLVMValueRef indirect = get_src(ctx, deref_instr->arr.index);
instr->const_index[0]);
break;
case nir_intrinsic_quad_broadcast: {
- unsigned lane = nir_src_as_const_value(instr->src[1])->u32[0];
+ unsigned lane = nir_src_as_uint(instr->src[1]);
result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
lane, lane, lane, lane);
break;
case nir_intrinsic_quad_swap_diagonal:
result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), 3, 2, 1 ,0);
break;
+ case nir_intrinsic_quad_swizzle_amd: {
+ uint32_t mask = nir_intrinsic_swizzle_mask(instr);
+ result = ac_build_quad_swizzle(&ctx->ac, get_src(ctx, instr->src[0]),
+ mask & 0x3, (mask >> 2) & 0x3,
+ (mask >> 4) & 0x3, (mask >> 6) & 0x3);
+ break;
+ }
+ case nir_intrinsic_masked_swizzle_amd: {
+ uint32_t mask = nir_intrinsic_swizzle_mask(instr);
+ result = ac_build_ds_swizzle(&ctx->ac, get_src(ctx, instr->src[0]), mask);
+ break;
+ }
+ case nir_intrinsic_write_invocation_amd:
+ result = ac_build_writelane(&ctx->ac, get_src(ctx, instr->src[0]),
+ get_src(ctx, instr->src[1]),
+ get_src(ctx, instr->src[2]));
+ break;
+ case nir_intrinsic_mbcnt_amd:
+ result = ac_build_mbcnt(&ctx->ac, get_src(ctx, instr->src[0]));
+ break;
default:
fprintf(stderr, "Unknown intrinsic: ");
nir_print_instr(&instr->instr, stderr);
if (!array_size)
array_size = 1;
- nir_const_value *const_value = nir_src_as_const_value(deref_instr->arr.index);
- if (const_value) {
- constant_index += array_size * const_value->u32[0];
+ if (nir_src_is_const(deref_instr->arr.index)) {
+ constant_index += array_size * nir_src_as_uint(deref_instr->arr.index);
} else {
LLVMValueRef indirect = get_src(ctx, deref_instr->arr.index);
/* Disable anisotropic filtering if BASE_LEVEL == LAST_LEVEL.
*
- * SI-CI:
+ * GFX6-GFX7:
* If BASE_LEVEL == LAST_LEVEL, the shader must disable anisotropic
* filtering manually. The driver sets img7 to a mask clearing
* MAX_ANISO_RATIO if BASE_LEVEL == LAST_LEVEL. The shader must do:
* s_and_b32 samp0, samp0, img7
*
- * VI:
+ * GFX8:
* The ANISO_OVERRIDE sampler field enables this fix in TA.
*/
static LLVMValueRef sici_fix_sampler_aniso(struct ac_nir_context *ctx,
LLVMBuilderRef builder = ctx->ac.builder;
LLVMValueRef img7, samp0;
- if (ctx->ac.chip_class >= VI)
+ if (ctx->ac.chip_class >= GFX8)
return samp;
img7 = LLVMBuildExtractElement(builder, res,
{
nir_deref_instr *texture_deref_instr = NULL;
nir_deref_instr *sampler_deref_instr = NULL;
+ int plane = -1;
for (unsigned i = 0; i < instr->num_srcs; i++) {
switch (instr->src[i].src_type) {
case nir_tex_src_sampler_deref:
sampler_deref_instr = nir_src_as_deref(instr->src[i].src);
break;
+ case nir_tex_src_plane:
+ plane = nir_src_as_int(instr->src[i].src);
+ break;
default:
break;
}
if (!sampler_deref_instr)
sampler_deref_instr = texture_deref_instr;
- if (instr->sampler_dim == GLSL_SAMPLER_DIM_BUF)
- *res_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_BUFFER, &instr->instr, false, false);
- else
- *res_ptr = get_sampler_desc(ctx, texture_deref_instr, AC_DESC_IMAGE, &instr->instr, false, false);
+ enum ac_descriptor_type main_descriptor = instr->sampler_dim == GLSL_SAMPLER_DIM_BUF ? AC_DESC_BUFFER : AC_DESC_IMAGE;
+
+ if (plane >= 0) {
+ assert(instr->op != nir_texop_txf_ms &&
+ instr->op != nir_texop_samples_identical);
+ assert(instr->sampler_dim != GLSL_SAMPLER_DIM_BUF);
+
+ main_descriptor = AC_DESC_PLANE_0 + plane;
+ }
+
+ *res_ptr = get_sampler_desc(ctx, texture_deref_instr, main_descriptor, &instr->instr, false, false);
+
if (samp_ptr) {
*samp_ptr = get_sampler_desc(ctx, sampler_deref_instr, AC_DESC_SAMPLER, &instr->instr, false, false);
if (instr->sampler_dim < GLSL_SAMPLER_DIM_RECT)
args.bias = get_src(ctx, instr->src[i].src);
break;
case nir_tex_src_lod: {
- nir_const_value *val = nir_src_as_const_value(instr->src[i].src);
-
- if (val && val->i32[0] == 0)
+ if (nir_src_is_const(instr->src[i].src) && nir_src_as_uint(instr->src[i].src) == 0)
args.level_zero = true;
else
args.lod = get_src(ctx, instr->src[i].src);
* It's unnecessary if the original texture format was
* Z32_FLOAT, but we don't know that here.
*/
- if (args.compare && ctx->ac.chip_class >= VI && ctx->abi->clamp_shadow_reference)
+ if (args.compare && ctx->ac.chip_class >= GFX8 && ctx->abi->clamp_shadow_reference)
args.compare = ac_build_clamp(&ctx->ac, ac_to_float(&ctx->ac, args.compare));
/* pack derivatives */
}
if (args.offset && instr->op == nir_texop_txf) {
- nir_const_value *const_offset =
- nir_src_as_const_value(instr->src[offset_src].src);
int num_offsets = instr->src[offset_src].src.ssa->num_components;
- assert(const_offset);
num_offsets = MIN2(num_offsets, instr->coord_components);
for (unsigned i = 0; i < num_offsets; ++i) {
args.coords[i] = LLVMBuildAdd(
ctx->ac.builder, args.coords[i],
- LLVMConstInt(ctx->ac.i32, const_offset->i32[i], false), "");
+ LLVMConstInt(ctx->ac.i32, nir_src_comp_as_uint(instr->src[offset_src].src, i), false), "");
}
args.offset = NULL;
}
- /* TODO TG4 support */
+ /* DMASK was repurposed for GATHER4. 4 components are always
+ * returned and DMASK works like a swizzle - it selects
+ * the component to fetch. The only valid DMASK values are
+ * 1=red, 2=green, 4=blue, 8=alpha. (e.g. 1 returns
+ * (red,red,red,red) etc.) The ISA document doesn't mention
+ * this.
+ */
args.dmask = 0xf;
if (instr->op == nir_texop_tg4) {
if (instr->is_shadow)
ctx.abi = abi;
ctx.stage = nir->info.stage;
+ ctx.info = &nir->info;
ctx.main_function = LLVMGetBasicBlockParent(LLVMGetInsertBlock(ctx.ac.builder));
* by the reality that LLVM 5.0 doesn't have working VGPR indexing
* on GFX9.
*/
- bool llvm_has_working_vgpr_indexing = chip_class <= VI;
+ bool llvm_has_working_vgpr_indexing = chip_class <= GFX8;
/* TODO: Indirect indexing of GS inputs is unimplemented.
*