if (cbuf >= 0 && cbuf < 8) {
struct r600_context *rctx = si_shader_ctx->rctx;
compressed = (si_shader_ctx->key.export_16bpc >> cbuf) & 0x1;
+
+ if (compressed)
+ si_shader_ctx->shader->spi_shader_col_format |=
+ V_028714_SPI_SHADER_FP16_ABGR << (4 * cbuf);
+ else
+ si_shader_ctx->shader->spi_shader_col_format |=
+ V_028714_SPI_SHADER_32_ABGR << (4 * cbuf);
}
}
LLVMInt32TypeInContext(base->gallivm->context),
args, 2,
LLVMReadNoneAttribute);
- args[chan + 7] = args[chan + 5];
+ args[chan + 7] = args[chan + 5] =
+ LLVMBuildBitCast(base->gallivm->builder,
+ args[chan + 5],
+ LLVMFloatTypeInContext(base->gallivm->context),
+ "");
}
/* Set COMPR flag */
struct lp_build_context * uint =
&si_shader_ctx->radeon_bld.soa.bld_base.uint_bld;
struct tgsi_parse_context *parse = &si_shader_ctx->parse;
+ LLVMValueRef args[9];
LLVMValueRef last_args[9] = { 0 };
unsigned color_count = 0;
unsigned param_count = 0;
+ int depth_index = -1, stencil_index = -1;
while (!tgsi_parse_end_of_tokens(parse)) {
struct tgsi_full_declaration *d =
&parse->FullToken.FullDeclaration;
- LLVMValueRef args[9];
unsigned target;
unsigned index;
int i;
/* Select the correct target */
switch(d->Semantic.Name) {
case TGSI_SEMANTIC_PSIZE:
- case TGSI_SEMANTIC_POSITION:
target = V_008DFC_SQ_EXP_POS;
break;
+ case TGSI_SEMANTIC_POSITION:
+ if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
+ target = V_008DFC_SQ_EXP_POS;
+ break;
+ } else {
+ depth_index = index;
+ continue;
+ }
+ case TGSI_SEMANTIC_STENCIL:
+ stencil_index = index;
+ continue;
case TGSI_SEMANTIC_COLOR:
if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX) {
case TGSI_SEMANTIC_BCOLOR:
}
}
+ if (depth_index >= 0 || stencil_index >= 0) {
+ LLVMValueRef out_ptr;
+ unsigned mask = 0;
+
+ /* Specify the target we are exporting */
+ args[3] = lp_build_const_int32(base->gallivm, V_008DFC_SQ_EXP_MRTZ);
+
+ if (depth_index >= 0) {
+ out_ptr = si_shader_ctx->radeon_bld.soa.outputs[depth_index][2];
+ args[5] = LLVMBuildLoad(base->gallivm->builder, out_ptr, "");
+ mask |= 0x1;
+
+ if (stencil_index < 0) {
+ args[6] =
+ args[7] =
+ args[8] = args[5];
+ }
+ }
+
+ if (stencil_index >= 0) {
+ out_ptr = si_shader_ctx->radeon_bld.soa.outputs[stencil_index][1];
+ args[7] =
+ args[8] =
+ args[6] = LLVMBuildLoad(base->gallivm->builder, out_ptr, "");
+ mask |= 0x2;
+
+ if (depth_index < 0)
+ args[5] = args[6];
+ }
+
+ /* Specify which components to enable */
+ args[0] = lp_build_const_int32(base->gallivm, mask);
+
+ args[1] =
+ args[2] =
+ args[4] = uint->zero;
+
+ if (last_args[0])
+ lp_build_intrinsic(base->gallivm->builder,
+ "llvm.SI.export",
+ LLVMVoidTypeInContext(base->gallivm->context),
+ args, 9);
+ else
+ memcpy(last_args, args, sizeof(args));
+ }
+
if (!last_args[0]) {
assert(si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT);
last_args[6]= uint->zero;
last_args[7]= uint->zero;
last_args[8]= uint->zero;
+
+ si_shader_ctx->shader->spi_shader_col_format |=
+ V_028714_SPI_SHADER_32_ABGR;
}
/* Specify whether the EXEC mask represents the valid mask */
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
const struct tgsi_full_instruction * inst = emit_data->inst;
LLVMValueRef ptr;
LLVMValueRef offset;
+ LLVMValueRef coords[5];
+ unsigned chan;
/* WriteMask */
/* XXX: should be optimized using emit_data->inst->Dst[0].Register.WriteMask*/
/* Coordinates */
/* XXX: Not all sample instructions need 4 address arguments. */
- if (inst->Instruction.Opcode == TGSI_OPCODE_TXP) {
- LLVMValueRef src_w;
- unsigned chan;
- LLVMValueRef coords[4];
-
- emit_data->dst_type = LLVMVectorType(bld_base->base.elem_type, 4);
- src_w = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W);
-
- for (chan = 0; chan < 3; chan++ ) {
- LLVMValueRef arg = lp_build_emit_fetch(bld_base,
- emit_data->inst, 0, chan);
+ if (inst->Instruction.Opcode == TGSI_OPCODE_TXP)
+ coords[3] = lp_build_emit_fetch(bld_base, emit_data->inst, 0, TGSI_CHAN_W)
+;
+
+ for (chan = 0; chan < 3; chan++ ) {
+ coords[chan] = lp_build_emit_fetch(bld_base,
+ emit_data->inst, 0,
+ chan);
+ if (inst->Instruction.Opcode == TGSI_OPCODE_TXP)
coords[chan] = lp_build_emit_llvm_binary(bld_base,
TGSI_OPCODE_DIV,
- arg, src_w);
- }
- coords[3] = bld_base->base.one;
- emit_data->args[1] = lp_build_gather_values(bld_base->base.gallivm,
- coords, 4);
- } else
- emit_data->args[1] = lp_build_emit_fetch(bld_base, emit_data->inst,
- 0, LP_CHAN_ALL);
+ coords[chan],
+ coords[3]);
+ }
+
+ coords[3] = bld_base->base.one;
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_TEX2 ||
+ inst->Instruction.Opcode == TGSI_OPCODE_TXB2 ||
+ inst->Instruction.Opcode == TGSI_OPCODE_TXL2) {
+ /* These instructions have additional operand that should be packed
+ * into the cube coord vector by radeon_llvm_emit_prepare_cube_coords.
+ * That operand should be passed as a float value in the args array
+ * right after the coord vector. After packing it's not used anymore,
+ * that's why arg_count is not increased */
+ coords[4] = lp_build_emit_fetch(bld_base, inst, 1, 0);
+ }
+
+ if ((inst->Texture.Texture == TGSI_TEXTURE_CUBE ||
+ inst->Texture.Texture == TGSI_TEXTURE_SHADOWCUBE) &&
+ inst->Instruction.Opcode != TGSI_OPCODE_TXQ) {
+ radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords);
+ }
+
+ for (chan = 0; chan < 4; chan++ ) {
+ coords[chan] = LLVMBuildBitCast(gallivm->builder,
+ coords[chan],
+ LLVMInt32TypeInContext(gallivm->context),
+ "");
+ }
+
+ emit_data->args[1] = lp_build_gather_values(gallivm, coords, 4);
/* Resource */
ptr = use_sgpr(bld_base->base.gallivm, SGPR_CONST_PTR_V8I32, SI_SGPR_RESOURCE);
ptr, offset);
/* Dimensions */
- /* XXX: We might want to pass this information to the shader at some. */
-/* emit_data->args[4] = lp_build_const_int32(bld_base->base.gallivm,
+ emit_data->args[4] = lp_build_const_int32(bld_base->base.gallivm,
emit_data->inst->Texture.Texture);
-*/
- emit_data->arg_count = 4;
+ emit_data->arg_count = 5;
/* XXX: To optimize, we could use a float or v2f32, if the last bits of
* the writemask are clear */
emit_data->dst_type = LLVMVectorType(
static const struct lp_build_tgsi_action tex_action = {
.fetch_args = tex_fetch_args,
.emit = lp_build_tgsi_intrinsic,
- .intr_name = "llvm.SI.sample"
+ .intr_name = "llvm.SI.sample."
};
static const struct lp_build_tgsi_action txb_action = {
.fetch_args = tex_fetch_args,
.emit = lp_build_tgsi_intrinsic,
- .intr_name = "llvm.SI.sample.bias"
+ .intr_name = "llvm.SI.sampleb."
};
static const struct lp_build_tgsi_action txl_action = {
.fetch_args = tex_fetch_args,
.emit = lp_build_tgsi_intrinsic,
- .intr_name = "llvm.SI.sample.lod"
+ .intr_name = "llvm.SI.samplel."
};
bld_base = &si_shader_ctx.radeon_bld.soa.bld_base;
tgsi_scan_shader(sel->tokens, &shader_info);
+ if (shader_info.indirect_files != 0) {
+ fprintf(stderr, "Indirect addressing not fully handled yet\n");
+ return -ENOSYS;
+ }
+
shader->shader.uses_kill = shader_info.uses_kill;
bld_base->info = &shader_info;
bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;