struct si_shader_output_values
{
LLVMValueRef values[4];
- unsigned name;
- unsigned sid;
+ unsigned semantic_name;
+ unsigned semantic_index;
+ ubyte vertex_stream[4];
};
static void si_init_shader_ctx(struct si_shader_context *ctx,
param);
if (LLVMGetTypeKind(LLVMTypeOf(value)) == LLVMFloatTypeKind)
- value = bitcast(&ctx->soa.bld_base,
+ value = bitcast(&ctx->bld_base,
TGSI_TYPE_UNSIGNED, value);
if (rshift)
static LLVMValueRef
get_tcs_out_patch0_offset(struct si_shader_context *ctx)
{
- return lp_build_mul_imm(&ctx->soa.bld_base.uint_bld,
+ return lp_build_mul_imm(&ctx->bld_base.uint_bld,
unpack_param(ctx,
SI_PARAM_TCS_OUT_OFFSETS,
0, 16),
static LLVMValueRef
get_tcs_out_patch0_patch_data_offset(struct si_shader_context *ctx)
{
- return lp_build_mul_imm(&ctx->soa.bld_base.uint_bld,
+ return lp_build_mul_imm(&ctx->bld_base.uint_bld,
unpack_param(ctx,
SI_PARAM_TCS_OUT_OFFSETS,
16, 16),
LLVMValueRef base_ptr, LLVMValueRef index,
LLVMValueRef value)
{
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMBuildStore(gallivm->builder, value,
LLVMValueRef base_ptr, LLVMValueRef index,
bool uniform)
{
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMValueRef pointer;
unsigned param_start_instance, unsigned divisor)
{
struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->soa.bld_base);
- struct gallivm_state *gallivm = radeon_bld->soa.bld_base.base.gallivm;
+ si_shader_context(&radeon_bld->bld_base);
+ struct gallivm_state *gallivm = radeon_bld->bld_base.base.gallivm;
LLVMValueRef result = LLVMGetParam(radeon_bld->main_fn,
ctx->param_instance_id);
const struct tgsi_full_declaration *decl,
LLVMValueRef out[4])
{
- struct lp_build_context *base = &ctx->soa.bld_base.base;
+ struct lp_build_context *base = &ctx->bld_base.base;
struct gallivm_state *gallivm = base->gallivm;
unsigned chan;
const struct tgsi_ind_register *ind,
int rel_index)
{
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
LLVMValueRef result;
- result = ctx->soa.addr[ind->Index][ind->Swizzle];
+ result = ctx->addrs[ind->Index][ind->Swizzle];
result = LLVMBuildLoad(gallivm->builder, result, "");
result = LLVMBuildAdd(gallivm->builder, result,
lp_build_const_int32(gallivm, rel_index), "");
LLVMValueRef vertex_dw_stride,
LLVMValueRef base_addr)
{
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
struct tgsi_shader_info *info = &ctx->shader->selector->info;
ubyte *name, *index, *array_first;
int first, param;
LLVMValueRef vertex_index,
LLVMValueRef param_index)
{
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
LLVMValueRef base_addr, vertices_per_patch, num_patches, total_vertices;
LLVMValueRef param_stride, constant16;
const struct tgsi_full_dst_register *dst,
const struct tgsi_full_src_register *src)
{
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
struct tgsi_shader_info *info = &ctx->shader->selector->info;
ubyte *name, *index, *array_first;
struct tgsi_full_src_register reg;
struct lp_build_context *base = &bld_base->base;
struct si_shader_context *ctx = si_shader_context(bld_base);
struct si_shader *shader = ctx->shader;
- struct lp_build_context *uint = &ctx->soa.bld_base.uint_bld;
+ struct lp_build_context *uint = &ctx->bld_base.uint_bld;
struct gallivm_state *gallivm = base->gallivm;
LLVMValueRef vtx_offset;
LLVMValueRef args[9];
LLVMValueRef face,
LLVMValueRef result[4])
{
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
struct lp_build_context *base = &bld_base->base;
struct lp_build_context *uint = &bld_base->uint_bld;
struct gallivm_state *gallivm = base->gallivm;
const struct tgsi_full_declaration *decl,
LLVMValueRef out[4])
{
- struct lp_build_context *base = &radeon_bld->soa.bld_base.base;
+ struct lp_build_context *base = &radeon_bld->bld_base.base;
struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->soa.bld_base);
+ si_shader_context(&radeon_bld->bld_base);
struct si_shader *shader = ctx->shader;
LLVMValueRef main_fn = radeon_bld->main_fn;
LLVMValueRef interp_param = NULL;
static LLVMValueRef get_sample_id(struct si_shader_context *radeon_bld)
{
- return unpack_param(si_shader_context(&radeon_bld->soa.bld_base),
+ return unpack_param(si_shader_context(&radeon_bld->bld_base),
SI_PARAM_ANCILLARY, 8, 4);
}
static LLVMValueRef load_sample_position(struct si_shader_context *radeon_bld, LLVMValueRef sample_id)
{
struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->soa.bld_base);
- struct lp_build_context *uint_bld = &radeon_bld->soa.bld_base.uint_bld;
+ si_shader_context(&radeon_bld->bld_base);
+ struct lp_build_context *uint_bld = &radeon_bld->bld_base.uint_bld;
struct gallivm_state *gallivm = &radeon_bld->gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef desc = LLVMGetParam(ctx->main_fn, SI_PARAM_RW_BUFFERS);
const struct tgsi_full_declaration *decl)
{
struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->soa.bld_base);
- struct lp_build_context *bld = &radeon_bld->soa.bld_base.base;
+ si_shader_context(&radeon_bld->bld_base);
+ struct lp_build_context *bld = &radeon_bld->bld_base.base;
struct gallivm_state *gallivm = &radeon_bld->gallivm;
LLVMValueRef value = 0;
LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_X_FLOAT),
LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_Y_FLOAT),
LLVMGetParam(radeon_bld->main_fn, SI_PARAM_POS_Z_FLOAT),
- lp_build_emit_llvm_unary(&radeon_bld->soa.bld_base, TGSI_OPCODE_RCP,
+ lp_build_emit_llvm_unary(&radeon_bld->bld_base, TGSI_OPCODE_RCP,
LLVMGetParam(radeon_bld->main_fn,
SI_PARAM_POS_W_FLOAT)),
};
lp_build_const_float(gallivm, 0),
lp_build_const_float(gallivm, 0)
};
- pos[0] = lp_build_emit_llvm_unary(&radeon_bld->soa.bld_base,
+ pos[0] = lp_build_emit_llvm_unary(&radeon_bld->bld_base,
TGSI_OPCODE_FRC, pos[0]);
- pos[1] = lp_build_emit_llvm_unary(&radeon_bld->soa.bld_base,
+ pos[1] = lp_build_emit_llvm_unary(&radeon_bld->bld_base,
TGSI_OPCODE_FRC, pos[1]);
value = lp_build_gather_values(gallivm, pos, 4);
break;
addr = get_tcs_tes_buffer_address(ctx, NULL,
lp_build_const_int32(gallivm, param));
- value = buffer_load(&radeon_bld->soa.bld_base, TGSI_TYPE_FLOAT,
+ value = buffer_load(&radeon_bld->bld_base, TGSI_TYPE_FLOAT,
~0, buffer, base, addr);
break;
}
case TGSI_SEMANTIC_PRIMID:
- value = get_primitive_id(&radeon_bld->soa.bld_base, 0);
+ value = get_primitive_id(&radeon_bld->bld_base, 0);
break;
case TGSI_SEMANTIC_GRID_SIZE:
value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_THREAD_ID);
break;
-#if HAVE_LLVM >= 0x0309
case TGSI_SEMANTIC_HELPER_INVOCATION:
- value = lp_build_intrinsic(gallivm->builder,
- "llvm.amdgcn.ps.live",
- ctx->i1, NULL, 0,
- LP_FUNC_ATTR_READNONE);
- value = LLVMBuildNot(gallivm->builder, value, "");
- value = LLVMBuildSExt(gallivm->builder, value, ctx->i32, "");
+ if (HAVE_LLVM >= 0x0309) {
+ value = lp_build_intrinsic(gallivm->builder,
+ "llvm.amdgcn.ps.live",
+ ctx->i1, NULL, 0,
+ LP_FUNC_ATTR_READNONE);
+ value = LLVMBuildNot(gallivm->builder, value, "");
+ value = LLVMBuildSExt(gallivm->builder, value, ctx->i32, "");
+ } else {
+ assert(!"TGSI_SEMANTIC_HELPER_INVOCATION unsupported");
+ return;
+ }
break;
-#endif
default:
assert(!"unknown system value");
const struct tgsi_full_declaration *decl)
{
struct si_shader_context *ctx =
- si_shader_context(&radeon_bld->soa.bld_base);
+ si_shader_context(&radeon_bld->bld_base);
struct si_shader_selector *sel = ctx->shader->selector;
struct gallivm_state *gallivm = &radeon_bld->gallivm;
bufp = load_const_buffer_desc(ctx, buf);
if (reg->Register.Indirect) {
- addr = ctx->soa.addr[ireg->Index][ireg->Swizzle];
+ addr = ctx->addrs[ireg->Index][ireg->Swizzle];
addr = LLVMBuildLoad(base->gallivm->builder, addr, "load addr reg");
addr = lp_build_mul_imm(&bld_base->uint_bld, addr, 16);
addr = lp_build_add(&bld_base->uint_bld, addr,
LLVMValueRef *args)
{
struct si_shader_context *ctx = si_shader_context(bld_base);
- struct lp_build_context *uint =
- &ctx->soa.bld_base.uint_bld;
+ struct lp_build_context *uint = &ctx->bld_base.uint_bld;
struct lp_build_context *base = &bld_base->base;
struct gallivm_state *gallivm = base->gallivm;
LLVMBuilderRef builder = base->gallivm->builder;
{
struct si_shader_context *ctx = si_shader_context(bld_base);
struct lp_build_context *base = &bld_base->base;
- struct lp_build_context *uint = &ctx->soa.bld_base.uint_bld;
+ struct lp_build_context *uint = &ctx->bld_base.uint_bld;
unsigned reg_index;
unsigned chan;
unsigned const_chan;
}
}
-/* On SI, the vertex shader is responsible for writing streamout data
- * to buffers. */
-static void si_llvm_emit_streamout(struct si_shader_context *ctx,
- struct si_shader_output_values *outputs,
- unsigned noutput)
+static void emit_streamout_output(struct si_shader_context *ctx,
+ LLVMValueRef const *so_buffers,
+ LLVMValueRef const *so_write_offsets,
+ struct pipe_stream_output *stream_out,
+ struct si_shader_output_values *shader_out)
{
- struct pipe_stream_output_info *so = &ctx->shader->selector->so;
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
- int i, j;
- struct lp_build_if_state if_ctx;
- LLVMValueRef so_buffers[4];
- LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
- SI_PARAM_RW_BUFFERS);
+ unsigned buf_idx = stream_out->output_buffer;
+ unsigned start = stream_out->start_component;
+ unsigned num_comps = stream_out->num_components;
+ LLVMValueRef out[4];
- /* Load the descriptors. */
- for (i = 0; i < 4; ++i) {
- if (ctx->shader->selector->so.stride[i]) {
- LLVMValueRef offset = lp_build_const_int32(gallivm,
- SI_VS_STREAMOUT_BUF0 + i);
+ assert(num_comps && num_comps <= 4);
+ if (!num_comps || num_comps > 4)
+ return;
- so_buffers[i] = build_indexed_load_const(ctx, buf_ptr, offset);
+ /* Load the output as int. */
+ for (int j = 0; j < num_comps; j++) {
+ assert(stream_out->stream == shader_out->vertex_stream[start + j]);
+
+ out[j] = LLVMBuildBitCast(builder,
+ shader_out->values[start + j],
+ ctx->i32, "");
+ }
+
+ /* Pack the output. */
+ LLVMValueRef vdata = NULL;
+
+ switch (num_comps) {
+ case 1: /* as i32 */
+ vdata = out[0];
+ break;
+ case 2: /* as v2i32 */
+ case 3: /* as v4i32 (aligned to 4) */
+ case 4: /* as v4i32 */
+ vdata = LLVMGetUndef(LLVMVectorType(ctx->i32, util_next_power_of_two(num_comps)));
+ for (int j = 0; j < num_comps; j++) {
+ vdata = LLVMBuildInsertElement(builder, vdata, out[j],
+ LLVMConstInt(ctx->i32, j, 0), "");
}
+ break;
}
+ build_tbuffer_store_dwords(ctx, so_buffers[buf_idx],
+ vdata, num_comps,
+ so_write_offsets[buf_idx],
+ LLVMConstInt(ctx->i32, 0, 0),
+ stream_out->dst_offset * 4);
+}
+
+/**
+ * Write streamout data to buffers for vertex stream @p stream (different
+ * vertex streams can occur for GS copy shaders).
+ */
+static void si_llvm_emit_streamout(struct si_shader_context *ctx,
+ struct si_shader_output_values *outputs,
+ unsigned noutput, unsigned stream)
+{
+ struct si_shader_selector *sel = ctx->shader->selector;
+ struct pipe_stream_output_info *so = &sel->so;
+ struct gallivm_state *gallivm = &ctx->gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ int i;
+ struct lp_build_if_state if_ctx;
+
/* Get bits [22:16], i.e. (so_param >> 16) & 127; */
LLVMValueRef so_vtx_count =
unpack_param(ctx, ctx->param_streamout_config, 16, 7);
LLVMValueRef can_emit =
LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");
- LLVMValueRef stream_id =
- unpack_param(ctx, ctx->param_streamout_config, 24, 2);
-
/* Emit the streamout code conditionally. This actually avoids
* out-of-bounds buffer access. The hw tells us via the SGPR
* (so_vtx_count) which threads are allowed to emit streamout data. */
/* Compute (streamout_write_index + thread_id). */
so_write_index = LLVMBuildAdd(builder, so_write_index, tid, "");
- /* Compute the write offset for each enabled buffer. */
+ /* Load the descriptor and compute the write offset for each
+ * enabled buffer. */
LLVMValueRef so_write_offset[4] = {};
+ LLVMValueRef so_buffers[4];
+ LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
+ SI_PARAM_RW_BUFFERS);
+
for (i = 0; i < 4; i++) {
if (!so->stride[i])
continue;
+ LLVMValueRef offset = lp_build_const_int32(gallivm,
+ SI_VS_STREAMOUT_BUF0 + i);
+
+ so_buffers[i] = build_indexed_load_const(ctx, buf_ptr, offset);
+
LLVMValueRef so_offset = LLVMGetParam(ctx->main_fn,
ctx->param_streamout_offset[i]);
so_offset = LLVMBuildMul(builder, so_offset, LLVMConstInt(ctx->i32, 4, 0), "");
/* Write streamout data. */
for (i = 0; i < so->num_outputs; i++) {
- unsigned buf_idx = so->output[i].output_buffer;
unsigned reg = so->output[i].register_index;
- unsigned start = so->output[i].start_component;
- unsigned num_comps = so->output[i].num_components;
- unsigned stream = so->output[i].stream;
- LLVMValueRef out[4];
- struct lp_build_if_state if_ctx_stream;
-
- assert(num_comps && num_comps <= 4);
- if (!num_comps || num_comps > 4)
- continue;
if (reg >= noutput)
continue;
- /* Load the output as int. */
- for (j = 0; j < num_comps; j++) {
- out[j] = LLVMBuildBitCast(builder,
- outputs[reg].values[start+j],
- ctx->i32, "");
- }
-
- /* Pack the output. */
- LLVMValueRef vdata = NULL;
-
- switch (num_comps) {
- case 1: /* as i32 */
- vdata = out[0];
- break;
- case 2: /* as v2i32 */
- case 3: /* as v4i32 (aligned to 4) */
- case 4: /* as v4i32 */
- vdata = LLVMGetUndef(LLVMVectorType(ctx->i32, util_next_power_of_two(num_comps)));
- for (j = 0; j < num_comps; j++) {
- vdata = LLVMBuildInsertElement(builder, vdata, out[j],
- LLVMConstInt(ctx->i32, j, 0), "");
- }
- break;
- }
+ if (stream != so->output[i].stream)
+ continue;
- LLVMValueRef can_emit_stream =
- LLVMBuildICmp(builder, LLVMIntEQ,
- stream_id,
- lp_build_const_int32(gallivm, stream), "");
-
- lp_build_if(&if_ctx_stream, gallivm, can_emit_stream);
- build_tbuffer_store_dwords(ctx, so_buffers[buf_idx],
- vdata, num_comps,
- so_write_offset[buf_idx],
- LLVMConstInt(ctx->i32, 0, 0),
- so->output[i].dst_offset*4);
- lp_build_endif(&if_ctx_stream);
+ emit_streamout_output(ctx, so_buffers, so_write_offset,
+ &so->output[i], &outputs[reg]);
}
}
lp_build_endif(&if_ctx);
struct si_shader_context *ctx = si_shader_context(bld_base);
struct si_shader *shader = ctx->shader;
struct lp_build_context *base = &bld_base->base;
- struct lp_build_context *uint =
- &ctx->soa.bld_base.uint_bld;
+ struct lp_build_context *uint = &ctx->bld_base.uint_bld;
LLVMValueRef args[9];
LLVMValueRef pos_args[4][9] = { { 0 } };
LLVMValueRef psize_value = NULL, edgeflag_value = NULL, layer_value = NULL, viewport_index_value = NULL;
unsigned pos_idx;
int i;
- if (outputs && ctx->shader->selector->so.num_outputs) {
- si_llvm_emit_streamout(ctx, outputs, noutput);
- }
-
for (i = 0; i < noutput; i++) {
- semantic_name = outputs[i].name;
- semantic_index = outputs[i].sid;
+ semantic_name = outputs[i].semantic_name;
+ semantic_index = outputs[i].semantic_index;
bool export_param = true;
switch (semantic_name) {
break;
}
+ if (outputs[i].vertex_stream[0] != 0 &&
+ outputs[i].vertex_stream[1] != 0 &&
+ outputs[i].vertex_stream[2] != 0 &&
+ outputs[i].vertex_stream[3] != 0)
+ export_param = false;
+
handle_semantic:
/* Select the correct target */
switch(semantic_name) {
case TGSI_SEMANTIC_POSITION:
target = V_008DFC_SQ_EXP_POS;
break;
- case TGSI_SEMANTIC_COLOR:
- case TGSI_SEMANTIC_BCOLOR:
- if (!export_param)
- continue;
- target = V_008DFC_SQ_EXP_PARAM + param_count;
- assert(i < ARRAY_SIZE(shader->info.vs_output_param_offset));
- shader->info.vs_output_param_offset[i] = param_count;
- param_count++;
- break;
case TGSI_SEMANTIC_CLIPDIST:
if (shader->key.opt.hw_vs.clip_disable) {
semantic_name = TGSI_SEMANTIC_GENERIC;
continue;
si_llvm_emit_clipvertex(bld_base, pos_args, outputs[i].values);
continue;
+ case TGSI_SEMANTIC_COLOR:
+ case TGSI_SEMANTIC_BCOLOR:
case TGSI_SEMANTIC_PRIMID:
case TGSI_SEMANTIC_FOG:
case TGSI_SEMANTIC_TEXCOORD:
/* Write outputs to LDS. The next shader (TCS aka HS) will read
* its inputs from it. */
for (i = 0; i < info->num_outputs; i++) {
- LLVMValueRef *out_ptr = ctx->soa.outputs[i];
+ LLVMValueRef *out_ptr = ctx->outputs[i];
unsigned name = info->output_semantic_name[i];
unsigned index = info->output_semantic_index[i];
int param = si_shader_io_get_unique_index(name, index);
int i;
for (i = 0; i < info->num_outputs; i++) {
- LLVMValueRef *out_ptr =
- ctx->soa.outputs[i];
+ LLVMValueRef *out_ptr = ctx->outputs[i];
int param_index;
if (info->output_semantic_name[i] == TGSI_SEMANTIC_VIEWPORT_INDEX ||
}
for (j = 0; j < 4; j++) {
- addr = ctx->soa.outputs[i][j];
+ addr = ctx->outputs[i][j];
val = LLVMBuildLoad(gallivm->builder, addr, "");
val = si_llvm_saturate(bld_base, val);
LLVMBuildStore(gallivm->builder, val, addr);
}
for (i = 0; i < info->num_outputs; i++) {
- outputs[i].name = info->output_semantic_name[i];
- outputs[i].sid = info->output_semantic_index[i];
+ outputs[i].semantic_name = info->output_semantic_name[i];
+ outputs[i].semantic_index = info->output_semantic_index[i];
- for (j = 0; j < 4; j++)
+ for (j = 0; j < 4; j++) {
outputs[i].values[j] =
LLVMBuildLoad(gallivm->builder,
- ctx->soa.outputs[i][j],
+ ctx->outputs[i][j],
"");
+ outputs[i].vertex_stream[j] =
+ (info->output_streams[i] >> (2 * j)) & 3;
+ }
+
}
/* Return the primitive ID from the LLVM function. */
get_primitive_id(bld_base, 0)),
VS_EPILOG_PRIMID_LOC, "");
+ if (ctx->shader->selector->so.num_outputs)
+ si_llvm_emit_streamout(ctx, outputs, i, 0);
si_llvm_export_vs(bld_base, outputs, i);
FREE(outputs);
}
case TGSI_SEMANTIC_COLOR:
assert(semantic_index < 8);
for (j = 0; j < 4; j++) {
- LLVMValueRef ptr = ctx->soa.outputs[i][j];
+ LLVMValueRef ptr = ctx->outputs[i][j];
LLVMValueRef result = LLVMBuildLoad(builder, ptr, "");
color[semantic_index][j] = result;
}
break;
case TGSI_SEMANTIC_POSITION:
depth = LLVMBuildLoad(builder,
- ctx->soa.outputs[i][2], "");
+ ctx->outputs[i][2], "");
break;
case TGSI_SEMANTIC_STENCIL:
stencil = LLVMBuildLoad(builder,
- ctx->soa.outputs[i][1], "");
+ ctx->outputs[i][1], "");
break;
case TGSI_SEMANTIC_SAMPLEMASK:
samplemask = LLVMBuildLoad(builder,
- ctx->soa.outputs[i][0], "");
+ ctx->outputs[i][0], "");
break;
default:
fprintf(stderr, "Warning: SI unhandled fs output type:%d\n",
LLVMValueRef offset, ptr;
int addr_space;
- offset = lp_build_emit_fetch(&ctx->soa.bld_base, inst, arg, 0);
+ offset = lp_build_emit_fetch(&ctx->bld_base, inst, arg, 0);
offset = LLVMBuildBitCast(builder, offset, ctx->i32, "");
ptr = ctx->shared_memory;
struct lp_build_emit_data *emit_data)
{
const struct tgsi_full_instruction *inst = emit_data->inst;
- struct lp_build_context *base = &ctx->soa.bld_base.base;
+ struct lp_build_context *base = &ctx->bld_base.base;
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
unsigned writemask = inst->Dst[0].Register.WriteMask;
const struct tgsi_full_instruction *inst = emit_data->inst;
struct gallivm_state *gallivm = &ctx->gallivm;
LLVMBuilderRef builder = gallivm->builder;
- struct lp_build_context *uint_bld = &ctx->soa.bld_base.uint_bld;
+ struct lp_build_context *uint_bld = &ctx->bld_base.uint_bld;
LLVMValueRef base_data = emit_data->args[0];
LLVMValueRef base_offset = emit_data->args[3];
unsigned writemask = inst->Dst[0].Register.WriteMask;
{
const struct tgsi_full_instruction *inst = emit_data->inst;
struct gallivm_state *gallivm = &ctx->gallivm;
- struct lp_build_context *base = &ctx->soa.bld_base.base;
+ struct lp_build_context *base = &ctx->bld_base.base;
LLVMBuilderRef builder = gallivm->builder;
unsigned writemask = inst->Dst[0].Register.WriteMask;
LLVMValueRef ptr, derived_ptr, data, index;
if (!(writemask & (1 << chan))) {
continue;
}
- data = lp_build_emit_fetch(&ctx->soa.bld_base, inst, 1, chan);
+ data = lp_build_emit_fetch(&ctx->bld_base, inst, 1, chan);
index = lp_build_const_int32(gallivm, chan);
derived_ptr = LLVMBuildGEP(builder, ptr, &index, 1, "");
LLVMBuildStore(builder, data, derived_ptr);
ptr = get_memory_ptr(ctx, inst, ctx->i32, 1);
- arg = lp_build_emit_fetch(&ctx->soa.bld_base, inst, 2, 0);
+ arg = lp_build_emit_fetch(&ctx->bld_base, inst, 2, 0);
arg = LLVMBuildBitCast(builder, arg, ctx->i32, "");
if (inst->Instruction.Opcode == TGSI_OPCODE_ATOMCAS) {
LLVMValueRef new_data;
- new_data = lp_build_emit_fetch(&ctx->soa.bld_base,
+ new_data = lp_build_emit_fetch(&ctx->bld_base,
inst, 3, 0);
new_data = LLVMBuildBitCast(builder, new_data, ctx->i32, "");
target == TGSI_TEXTURE_CUBE_ARRAY ||
target == TGSI_TEXTURE_SHADOWCUBE ||
target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
- si_prepare_cube_coords(bld_base, emit_data, coords, derivs);
+ ac_prepare_cube_coords(&ctx->ac,
+ opcode == TGSI_OPCODE_TXD,
+ target == TGSI_TEXTURE_CUBE_ARRAY ||
+ target == TGSI_TEXTURE_SHADOWCUBE_ARRAY,
+ coords, derivs);
if (opcode == TGSI_OPCODE_TXD)
for (int i = 0; i < num_deriv_channels * 2; i++)
/* add tex offsets */
if (inst->Texture.NumOffsets) {
struct lp_build_context *uint_bld = &bld_base->uint_bld;
- struct lp_build_tgsi_soa_context *bld = lp_soa_context(bld_base);
const struct tgsi_texture_offset *off = inst->TexOffsets;
assert(inst->Texture.NumOffsets == 1);
switch (target) {
case TGSI_TEXTURE_3D:
address[2] = lp_build_add(uint_bld, address[2],
- bld->immediates[off->Index][off->SwizzleZ]);
+ ctx->imms[off->Index * TGSI_NUM_CHANNELS + off->SwizzleZ]);
/* fall through */
case TGSI_TEXTURE_2D:
case TGSI_TEXTURE_SHADOW2D:
case TGSI_TEXTURE_SHADOW2D_ARRAY:
address[1] =
lp_build_add(uint_bld, address[1],
- bld->immediates[off->Index][off->SwizzleY]);
+ ctx->imms[off->Index * TGSI_NUM_CHANNELS + off->SwizzleY]);
/* fall through */
case TGSI_TEXTURE_1D:
case TGSI_TEXTURE_SHADOW1D:
case TGSI_TEXTURE_SHADOW1D_ARRAY:
address[0] =
lp_build_add(uint_bld, address[0],
- bld->immediates[off->Index][off->SwizzleX]);
+ ctx->imms[off->Index * TGSI_NUM_CHANNELS + off->SwizzleX]);
break;
/* texture offsets do not apply to other texture targets */
}
/* Get the component index from src1.x for Gather4. */
if (!tgsi_is_shadow_target(target)) {
- LLVMValueRef (*imms)[4] = lp_soa_context(bld_base)->immediates;
LLVMValueRef comp_imm;
struct tgsi_src_register src1 = inst->Src[1].Register;
assert(src1.File == TGSI_FILE_IMMEDIATE);
- comp_imm = imms[src1.Index][src1.SwizzleX];
+ comp_imm = ctx->imms[src1.Index * TGSI_NUM_CHANNELS + src1.SwizzleX];
gather_comp = LLVMConstIntGetZExtValue(comp_imm);
gather_comp = CLAMP(gather_comp, 0, 3);
}
set_tex_fetch_args(ctx, &txq_emit_data, TGSI_OPCODE_TXQ,
txq_inst.Texture.Texture,
emit_data->args[1], NULL,
- &ctx->soa.bld_base.uint_bld.zero,
+ &ctx->bld_base.uint_bld.zero,
1, 0xf);
- txq_emit(NULL, &ctx->soa.bld_base, &txq_emit_data);
+ txq_emit(NULL, &ctx->bld_base, &txq_emit_data);
/* Compute -0.5 / size. */
for (c = 0; c < 2; c++) {
LLVMConstInt(ctx->i32, c, 0), "");
half_texel[c] = LLVMBuildUIToFP(builder, half_texel[c], ctx->f32, "");
half_texel[c] =
- lp_build_emit_llvm_unary(&ctx->soa.bld_base,
+ lp_build_emit_llvm_unary(&ctx->bld_base,
TGSI_OPCODE_RCP, half_texel[c]);
half_texel[c] = LLVMBuildFMul(builder, half_texel[c],
LLVMConstReal(ctx->f32, -0.5), "");
static unsigned si_llvm_get_stream(struct lp_build_tgsi_context *bld_base,
struct lp_build_emit_data *emit_data)
{
- LLVMValueRef (*imms)[4] = lp_soa_context(bld_base)->immediates;
+ struct si_shader_context *ctx = si_shader_context(bld_base);
struct tgsi_src_register src0 = emit_data->inst->Src[0].Register;
+ LLVMValueRef imm;
unsigned stream;
assert(src0.File == TGSI_FILE_IMMEDIATE);
- stream = LLVMConstIntGetZExtValue(imms[src0.Index][src0.SwizzleX]) & 0x3;
+ imm = ctx->imms[src0.Index * TGSI_NUM_CHANNELS + src0.SwizzleX];
+ stream = LLVMConstIntGetZExtValue(imm) & 0x3;
return stream;
}
LLVMValueRef gs_next_vertex;
LLVMValueRef can_emit, kill;
LLVMValueRef args[2];
- unsigned chan;
+ unsigned chan, offset;
int i;
unsigned stream;
* further memory loads and may allow LLVM to skip to the end
* altogether.
*/
- can_emit = LLVMBuildICmp(gallivm->builder, LLVMIntULE, gs_next_vertex,
+ can_emit = LLVMBuildICmp(gallivm->builder, LLVMIntULT, gs_next_vertex,
lp_build_const_int32(gallivm,
shader->selector->gs_max_out_vertices), "");
lp_build_if(&if_state, gallivm, can_emit);
}
+ offset = 0;
for (i = 0; i < info->num_outputs; i++) {
- LLVMValueRef *out_ptr =
- ctx->soa.outputs[i];
+ LLVMValueRef *out_ptr = ctx->outputs[i];
for (chan = 0; chan < 4; chan++) {
+ if (!(info->output_usagemask[i] & (1 << chan)) ||
+ ((info->output_streams[i] >> (2 * chan)) & 3) != stream)
+ continue;
+
LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], "");
LLVMValueRef voffset =
- lp_build_const_int32(gallivm, (i * 4 + chan) *
+ lp_build_const_int32(gallivm, offset *
shader->selector->gs_max_out_vertices);
+ offset++;
voffset = lp_build_add(uint, voffset, gs_next_vertex);
voffset = lp_build_mul_imm(uint, voffset, 4);
static void create_meta_data(struct si_shader_context *ctx)
{
- struct gallivm_state *gallivm = ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
ctx->invariant_load_md_kind = LLVMGetMDKindIDInContext(gallivm->context,
"invariant.load", 14);
static void declare_tess_lds(struct si_shader_context *ctx)
{
struct gallivm_state *gallivm = &ctx->gallivm;
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
struct lp_build_context *uint = &bld_base->uint_bld;
unsigned lds_size = ctx->screen->b.chip_class >= CIK ? 65536 : 32768;
static void create_function(struct si_shader_context *ctx)
{
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
struct gallivm_state *gallivm = bld_base->base.gallivm;
struct si_shader *shader = ctx->shader;
LLVMTypeRef params[SI_NUM_PARAMS + SI_NUM_VERTEX_BUFFERS], v3i32;
*/
static void preload_ring_buffers(struct si_shader_context *ctx)
{
- struct gallivm_state *gallivm =
- ctx->soa.bld_base.base.gallivm;
+ struct gallivm_state *gallivm = ctx->bld_base.base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef buf_ptr = LLVMGetParam(ctx->main_fn,
SI_PARAM_RW_BUFFERS);
}
if (ctx->shader->is_gs_copy_shader) {
- LLVMValueRef offset = lp_build_const_int32(gallivm, SI_VS_RING_GSVS);
+ LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS);
ctx->gsvs_ring[0] =
build_indexed_load_const(ctx, buf_ptr, offset);
- }
- if (ctx->type == PIPE_SHADER_GEOMETRY) {
- int i;
- for (i = 0; i < 4; i++) {
- LLVMValueRef offset = lp_build_const_int32(gallivm, SI_GS_RING_GSVS0 + i);
+ } else if (ctx->type == PIPE_SHADER_GEOMETRY) {
+ const struct si_shader_selector *sel = ctx->shader->selector;
+ struct lp_build_context *uint = &ctx->bld_base.uint_bld;
+ LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS);
+ LLVMValueRef base_ring;
+
+ base_ring = build_indexed_load_const(ctx, buf_ptr, offset);
+
+ /* The conceptual layout of the GSVS ring is
+ * v0c0 .. vLv0 v0c1 .. vLc1 ..
+ * but the real memory layout is swizzled across
+ * threads:
+ * t0v0c0 .. t15v0c0 t0v1c0 .. t15v1c0 ... t15vLcL
+ * t16v0c0 ..
+ * Override the buffer descriptor accordingly.
+ */
+ LLVMTypeRef v2i64 = LLVMVectorType(ctx->i64, 2);
+ uint64_t stream_offset = 0;
+
+ for (unsigned stream = 0; stream < 4; ++stream) {
+ unsigned num_components;
+ unsigned stride;
+ unsigned num_records;
+ LLVMValueRef ring, tmp;
- ctx->gsvs_ring[i] =
- build_indexed_load_const(ctx, buf_ptr, offset);
+ num_components = sel->info.num_stream_output_components[stream];
+ if (!num_components)
+ continue;
+
+ stride = 4 * num_components * sel->gs_max_out_vertices;
+
+ /* Limit on the stride field for <= CIK. */
+ assert(stride < (1 << 14));
+
+ num_records = 64;
+
+ ring = LLVMBuildBitCast(builder, base_ring, v2i64, "");
+ tmp = LLVMBuildExtractElement(builder, ring, uint->zero, "");
+ tmp = LLVMBuildAdd(builder, tmp,
+ LLVMConstInt(ctx->i64,
+ stream_offset, 0), "");
+ stream_offset += stride * 64;
+
+ ring = LLVMBuildInsertElement(builder, ring, tmp, uint->zero, "");
+ ring = LLVMBuildBitCast(builder, ring, ctx->v4i32, "");
+ tmp = LLVMBuildExtractElement(builder, ring, uint->one, "");
+ tmp = LLVMBuildOr(builder, tmp,
+ LLVMConstInt(ctx->i32,
+ S_008F04_STRIDE(stride) |
+ S_008F04_SWIZZLE_ENABLE(1), 0), "");
+ ring = LLVMBuildInsertElement(builder, ring, tmp, uint->one, "");
+ ring = LLVMBuildInsertElement(builder, ring,
+ LLVMConstInt(ctx->i32, num_records, 0),
+ LLVMConstInt(ctx->i32, 2, 0), "");
+ ring = LLVMBuildInsertElement(builder, ring,
+ LLVMConstInt(ctx->i32,
+ S_008F0C_DST_SEL_X(V_008F0C_SQ_SEL_X) |
+ S_008F0C_DST_SEL_Y(V_008F0C_SQ_SEL_Y) |
+ S_008F0C_DST_SEL_Z(V_008F0C_SQ_SEL_Z) |
+ S_008F0C_DST_SEL_W(V_008F0C_SQ_SEL_W) |
+ S_008F0C_NUM_FORMAT(V_008F0C_BUF_NUM_FORMAT_FLOAT) |
+ S_008F0C_DATA_FORMAT(V_008F0C_BUF_DATA_FORMAT_32) |
+ S_008F0C_ELEMENT_SIZE(1) | /* element_size = 4 (bytes) */
+ S_008F0C_INDEX_STRIDE(1) | /* index_stride = 16 (elements) */
+ S_008F0C_ADD_TID_ENABLE(1),
+ 0),
+ LLVMConstInt(ctx->i32, 3, 0), "");
+ ring = LLVMBuildBitCast(builder, ring, ctx->v16i8, "");
+
+ ctx->gsvs_ring[stream] = ring;
}
}
}
LLVMValueRef param_rw_buffers,
unsigned param_pos_fixed_pt)
{
- struct lp_build_tgsi_context *bld_base =
- &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef slot, desc, offset, row, bit, address[2];
struct si_shader *shader,
struct pipe_debug_callback *debug,
unsigned processor,
- FILE *file)
+ FILE *file,
+ bool check_debug_option)
{
struct si_shader_config *conf = &shader->config;
unsigned num_inputs = shader->selector ? shader->selector->info.num_inputs : 0;
if (lds_per_wave)
max_simd_waves = MIN2(max_simd_waves, 16384 / lds_per_wave);
- if (file != stderr ||
+ if (!check_debug_option ||
r600_can_dump_shader(&sscreen->b, processor)) {
if (processor == PIPE_SHADER_FRAGMENT) {
fprintf(file, "*** SHADER CONFIG ***\n"
void si_shader_dump(struct si_screen *sscreen, struct si_shader *shader,
struct pipe_debug_callback *debug, unsigned processor,
- FILE *file)
+ FILE *file, bool check_debug_option)
{
- if (file != stderr ||
+ if (!check_debug_option ||
r600_can_dump_shader(&sscreen->b, processor))
si_dump_shader_key(processor, &shader->key, file);
- if (file != stderr && shader->binary.llvm_ir_string) {
+ if (!check_debug_option && shader->binary.llvm_ir_string) {
fprintf(file, "\n%s - main shader part - LLVM IR:\n\n",
si_get_shader_name(shader, processor));
fprintf(file, "%s\n", shader->binary.llvm_ir_string);
}
- if (file != stderr ||
+ if (!check_debug_option ||
(r600_can_dump_shader(&sscreen->b, processor) &&
!(sscreen->b.debug_flags & DBG_NO_ASM))) {
fprintf(file, "\n%s:\n", si_get_shader_name(shader, processor));
fprintf(file, "\n");
}
- si_shader_dump_stats(sscreen, shader, debug, processor, file);
+ si_shader_dump_stats(sscreen, shader, debug, processor, file,
+ check_debug_option);
}
int si_compile_llvm(struct si_screen *sscreen,
struct si_shader_context ctx;
struct si_shader *shader;
struct gallivm_state *gallivm = &ctx.gallivm;
- struct lp_build_tgsi_context *bld_base = &ctx.soa.bld_base;
+ LLVMBuilderRef builder;
+ struct lp_build_tgsi_context *bld_base = &ctx.bld_base;
struct lp_build_context *uint = &bld_base->uint_bld;
struct si_shader_output_values *outputs;
struct tgsi_shader_info *gsinfo = &gs_selector->info;
si_init_shader_ctx(&ctx, sscreen, shader, tm);
ctx.type = PIPE_SHADER_VERTEX;
+ builder = gallivm->builder;
+
create_meta_data(&ctx);
create_function(&ctx);
preload_ring_buffers(&ctx);
args[7] = uint->one; /* SLC */
args[8] = uint->zero; /* TFE */
- /* Fetch vertex data from GSVS ring */
+ /* Fetch the vertex stream ID.*/
+ LLVMValueRef stream_id;
+
+ if (gs_selector->so.num_outputs)
+ stream_id = unpack_param(&ctx, ctx.param_streamout_config, 24, 2);
+ else
+ stream_id = uint->zero;
+
+ /* Fill in output information. */
for (i = 0; i < gsinfo->num_outputs; ++i) {
- unsigned chan;
+ outputs[i].semantic_name = gsinfo->output_semantic_name[i];
+ outputs[i].semantic_index = gsinfo->output_semantic_index[i];
- outputs[i].name = gsinfo->output_semantic_name[i];
- outputs[i].sid = gsinfo->output_semantic_index[i];
+ for (int chan = 0; chan < 4; chan++) {
+ outputs[i].vertex_stream[chan] =
+ (gsinfo->output_streams[i] >> (2 * chan)) & 3;
+ }
+ }
- for (chan = 0; chan < 4; chan++) {
- args[2] = lp_build_const_int32(gallivm,
- (i * 4 + chan) *
- gs_selector->gs_max_out_vertices * 16 * 4);
+ LLVMBasicBlockRef end_bb;
+ LLVMValueRef switch_inst;
- outputs[i].values[chan] =
- LLVMBuildBitCast(gallivm->builder,
+ end_bb = LLVMAppendBasicBlockInContext(gallivm->context, ctx.main_fn, "end");
+ switch_inst = LLVMBuildSwitch(builder, stream_id, end_bb, 4);
+
+ for (int stream = 0; stream < 4; stream++) {
+ LLVMBasicBlockRef bb;
+ unsigned offset;
+
+ if (!gsinfo->num_stream_output_components[stream])
+ continue;
+
+ if (stream > 0 && !gs_selector->so.num_outputs)
+ continue;
+
+ bb = LLVMInsertBasicBlockInContext(gallivm->context, end_bb, "out");
+ LLVMAddCase(switch_inst, lp_build_const_int32(gallivm, stream), bb);
+ LLVMPositionBuilderAtEnd(builder, bb);
+
+ /* Fetch vertex data from GSVS ring */
+ offset = 0;
+ for (i = 0; i < gsinfo->num_outputs; ++i) {
+ for (unsigned chan = 0; chan < 4; chan++) {
+ if (!(gsinfo->output_usagemask[i] & (1 << chan)) ||
+ outputs[i].vertex_stream[chan] != stream) {
+ outputs[i].values[chan] = ctx.bld_base.base.undef;
+ continue;
+ }
+
+ args[2] = lp_build_const_int32(
+ gallivm,
+ offset * gs_selector->gs_max_out_vertices * 16 * 4);
+ offset++;
+
+ outputs[i].values[chan] =
+ LLVMBuildBitCast(gallivm->builder,
lp_build_intrinsic(gallivm->builder,
"llvm.SI.buffer.load.dword.i32.i32",
ctx.i32, args, 9,
LP_FUNC_ATTR_READONLY),
ctx.f32, "");
+ }
+ }
+
+ /* Streamout and exports. */
+ if (gs_selector->so.num_outputs) {
+ si_llvm_emit_streamout(&ctx, outputs,
+ gsinfo->num_outputs,
+ stream);
}
+
+ if (stream == 0)
+ si_llvm_export_vs(bld_base, outputs, gsinfo->num_outputs);
+
+ LLVMBuildBr(builder, end_bb);
}
- si_llvm_export_vs(bld_base, outputs, gsinfo->num_outputs);
+ LLVMPositionBuilderAtEnd(builder, end_bb);
LLVMBuildRetVoid(gallivm->builder);
if (r600_can_dump_shader(&sscreen->b, PIPE_SHADER_GEOMETRY))
fprintf(stderr, "GS Copy Shader:\n");
si_shader_dump(sscreen, ctx.shader, debug,
- PIPE_SHADER_GEOMETRY, stderr);
+ PIPE_SHADER_GEOMETRY, stderr, true);
r = si_shader_binary_upload(sscreen, ctx.shader);
}
(shader && shader->selector) ? &shader->selector->info : NULL,
(shader && shader->selector) ? shader->selector->tokens : NULL);
- bld_base = &ctx->soa.bld_base;
+ bld_base = &ctx->bld_base;
bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
bld_base->op_actions[TGSI_OPCODE_INTERP_CENTROID] = interp_action;
struct si_shader *shader)
{
struct si_shader_selector *sel = shader->selector;
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
switch (ctx->type) {
case PIPE_SHADER_VERTEX:
shader->info.uses_instanceid = sel->info.uses_instanceid;
- bld_base = &ctx.soa.bld_base;
+ bld_base = &ctx.bld_base;
ctx.load_system_value = declare_system_value;
if (!si_compile_tgsi_main(&ctx, shader)) {
union si_shader_part_key *key)
{
struct gallivm_state *gallivm = &ctx->gallivm;
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
LLVMTypeRef params[5];
int num_params, i;
union si_shader_part_key *key)
{
struct gallivm_state *gallivm = &ctx->gallivm;
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
LLVMTypeRef params[16];
LLVMValueRef func;
int last_sgpr, num_params;
union si_shader_part_key *key)
{
struct gallivm_state *gallivm = &ctx->gallivm;
- struct lp_build_tgsi_context *bld_base = &ctx->soa.bld_base;
+ struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
LLVMTypeRef params[16+8*4+3];
LLVMValueRef depth = NULL, stencil = NULL, samplemask = NULL;
int last_sgpr, num_params, i;
si_fix_resource_usage(sscreen, shader);
si_shader_dump(sscreen, shader, debug, sel->info.processor,
- stderr);
+ stderr, true);
/* Upload. */
r = si_shader_binary_upload(sscreen, shader);
projects / mesa.git / blobdiff