#include "gallivm/lp_bld_intr.h"
#include "gallivm/lp_bld_logic.h"
#include "gallivm/lp_bld_arit.h"
+#include "gallivm/lp_bld_bitarit.h"
#include "gallivm/lp_bld_flow.h"
#include "radeon/r600_cs.h"
#include "radeon/radeon_llvm.h"
int param_streamout_write_index;
int param_streamout_offset[4];
int param_vertex_id;
+ int param_rel_auto_id;
+ int param_vs_prim_id;
int param_instance_id;
+ int param_tes_u;
+ int param_tes_v;
+ int param_tes_rel_patch_id;
+ int param_tes_patch_id;
int param_es2gs_offset;
LLVMTargetMachineRef tm;
LLVMValueRef const_md;
- LLVMValueRef const_resource[SI_NUM_CONST_BUFFERS];
- LLVMValueRef ddxy_lds;
+ LLVMValueRef const_buffers[SI_NUM_CONST_BUFFERS];
+ LLVMValueRef lds;
LLVMValueRef *constants[SI_NUM_CONST_BUFFERS];
- LLVMValueRef resources[SI_NUM_SAMPLER_VIEWS];
- LLVMValueRef samplers[SI_NUM_SAMPLER_STATES];
+ LLVMValueRef sampler_views[SI_NUM_SAMPLER_VIEWS];
+ LLVMValueRef sampler_states[SI_NUM_SAMPLER_STATES];
LLVMValueRef so_buffers[4];
LLVMValueRef esgs_ring;
- LLVMValueRef gsvs_ring;
- LLVMValueRef gs_next_vertex;
+ LLVMValueRef gsvs_ring[4];
+ LLVMValueRef gs_next_vertex[4];
};
static struct si_shader_context * si_shader_context(
assert(index <= 1);
return 2 + index;
case TGSI_SEMANTIC_GENERIC:
- assert(index <= 63-4);
- return 4 + index;
+ if (index <= 63-4)
+ return 4 + index;
+ else
+ /* same explanation as in the default statement,
+ * the only user hitting this is st/nine.
+ */
+ return 0;
+
+ /* patch indices are completely separate and thus start from 0 */
+ case TGSI_SEMANTIC_TESSOUTER:
+ return 0;
+ case TGSI_SEMANTIC_TESSINNER:
+ return 1;
+ case TGSI_SEMANTIC_PATCH:
+ return 2 + index;
default:
/* Don't fail here. The result of this function is only used
}
}
-/**
- * Given a semantic name and index of a parameter and a mask of used parameters
- * (inputs or outputs), return the index of the parameter in the list of all
- * used parameters.
- *
- * For example, assume this list of parameters:
- * POSITION, PSIZE, GENERIC0, GENERIC2
- * which has the mask:
- * 11000000000101
- * Then:
- * querying POSITION returns 0,
- * querying PSIZE returns 1,
- * querying GENERIC0 returns 2,
- * querying GENERIC2 returns 3.
- *
- * Which can be used as an offset to a parameter buffer in units of vec4s.
- */
-static int get_param_index(unsigned semantic_name, unsigned index,
- uint64_t mask)
-{
- unsigned unique_index = si_shader_io_get_unique_index(semantic_name, index);
- int i, param_index = 0;
-
- /* If not present... */
- if (!((1llu << unique_index) & mask))
- return -1;
-
- for (i = 0; mask; i++) {
- uint64_t bit = 1llu << i;
-
- if (bit & mask) {
- if (i == unique_index)
- return param_index;
-
- mask &= ~bit;
- param_index++;
- }
- }
-
- assert(!"unreachable");
- return -1;
-}
-
/**
* Get the value of a shader input parameter and extract a bitfield.
*/
return value;
}
+static LLVMValueRef get_rel_patch_id(struct si_shader_context *si_shader_ctx)
+{
+ switch (si_shader_ctx->type) {
+ case TGSI_PROCESSOR_TESS_CTRL:
+ return unpack_param(si_shader_ctx, SI_PARAM_REL_IDS, 0, 8);
+
+ case TGSI_PROCESSOR_TESS_EVAL:
+ return LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ si_shader_ctx->param_tes_rel_patch_id);
+
+ default:
+ assert(0);
+ return NULL;
+ }
+}
+
+/* Tessellation shaders pass outputs to the next shader using LDS.
+ *
+ * LS outputs = TCS inputs
+ * TCS outputs = TES inputs
+ *
+ * The LDS layout is:
+ * - TCS inputs for patch 0
+ * - TCS inputs for patch 1
+ * - TCS inputs for patch 2 = get_tcs_in_current_patch_offset (if RelPatchID==2)
+ * - ...
+ * - TCS outputs for patch 0 = get_tcs_out_patch0_offset
+ * - Per-patch TCS outputs for patch 0 = get_tcs_out_patch0_patch_data_offset
+ * - TCS outputs for patch 1
+ * - Per-patch TCS outputs for patch 1
+ * - TCS outputs for patch 2 = get_tcs_out_current_patch_offset (if RelPatchID==2)
+ * - Per-patch TCS outputs for patch 2 = get_tcs_out_current_patch_data_offset (if RelPatchID==2)
+ * - ...
+ *
+ * All three shaders VS(LS), TCS, TES share the same LDS space.
+ */
+
+static LLVMValueRef
+get_tcs_in_patch_stride(struct si_shader_context *si_shader_ctx)
+{
+ if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX)
+ return unpack_param(si_shader_ctx, SI_PARAM_LS_OUT_LAYOUT, 0, 13);
+ else if (si_shader_ctx->type == TGSI_PROCESSOR_TESS_CTRL)
+ return unpack_param(si_shader_ctx, SI_PARAM_TCS_IN_LAYOUT, 0, 13);
+ else {
+ assert(0);
+ return NULL;
+ }
+}
+
+static LLVMValueRef
+get_tcs_out_patch_stride(struct si_shader_context *si_shader_ctx)
+{
+ return unpack_param(si_shader_ctx, SI_PARAM_TCS_OUT_LAYOUT, 0, 13);
+}
+
+static LLVMValueRef
+get_tcs_out_patch0_offset(struct si_shader_context *si_shader_ctx)
+{
+ return lp_build_mul_imm(&si_shader_ctx->radeon_bld.soa.bld_base.uint_bld,
+ unpack_param(si_shader_ctx,
+ SI_PARAM_TCS_OUT_OFFSETS,
+ 0, 16),
+ 4);
+}
+
+static LLVMValueRef
+get_tcs_out_patch0_patch_data_offset(struct si_shader_context *si_shader_ctx)
+{
+ return lp_build_mul_imm(&si_shader_ctx->radeon_bld.soa.bld_base.uint_bld,
+ unpack_param(si_shader_ctx,
+ SI_PARAM_TCS_OUT_OFFSETS,
+ 16, 16),
+ 4);
+}
+
+static LLVMValueRef
+get_tcs_in_current_patch_offset(struct si_shader_context *si_shader_ctx)
+{
+ struct gallivm_state *gallivm = &si_shader_ctx->radeon_bld.gallivm;
+ LLVMValueRef patch_stride = get_tcs_in_patch_stride(si_shader_ctx);
+ LLVMValueRef rel_patch_id = get_rel_patch_id(si_shader_ctx);
+
+ return LLVMBuildMul(gallivm->builder, patch_stride, rel_patch_id, "");
+}
+
+static LLVMValueRef
+get_tcs_out_current_patch_offset(struct si_shader_context *si_shader_ctx)
+{
+ struct gallivm_state *gallivm = &si_shader_ctx->radeon_bld.gallivm;
+ LLVMValueRef patch0_offset = get_tcs_out_patch0_offset(si_shader_ctx);
+ LLVMValueRef patch_stride = get_tcs_out_patch_stride(si_shader_ctx);
+ LLVMValueRef rel_patch_id = get_rel_patch_id(si_shader_ctx);
+
+ return LLVMBuildAdd(gallivm->builder, patch0_offset,
+ LLVMBuildMul(gallivm->builder, patch_stride,
+ rel_patch_id, ""),
+ "");
+}
+
+static LLVMValueRef
+get_tcs_out_current_patch_data_offset(struct si_shader_context *si_shader_ctx)
+{
+ struct gallivm_state *gallivm = &si_shader_ctx->radeon_bld.gallivm;
+ LLVMValueRef patch0_patch_data_offset =
+ get_tcs_out_patch0_patch_data_offset(si_shader_ctx);
+ LLVMValueRef patch_stride = get_tcs_out_patch_stride(si_shader_ctx);
+ LLVMValueRef rel_patch_id = get_rel_patch_id(si_shader_ctx);
+
+ return LLVMBuildAdd(gallivm->builder, patch0_patch_data_offset,
+ LLVMBuildMul(gallivm->builder, patch_stride,
+ rel_patch_id, ""),
+ "");
+}
+
+static void build_indexed_store(struct si_shader_context *si_shader_ctx,
+ LLVMValueRef base_ptr, LLVMValueRef index,
+ LLVMValueRef value)
+{
+ struct lp_build_tgsi_context *bld_base = &si_shader_ctx->radeon_bld.soa.bld_base;
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMValueRef indices[2], pointer;
+
+ indices[0] = bld_base->uint_bld.zero;
+ indices[1] = index;
+
+ pointer = LLVMBuildGEP(gallivm->builder, base_ptr, indices, 2, "");
+ LLVMBuildStore(gallivm->builder, value, pointer);
+}
+
/**
* Build an LLVM bytecode indexed load using LLVMBuildGEP + LLVMBuildLoad.
* It's equivalent to doing a load from &base_ptr[index].
LLVMValueRef input;
/* Load the T list */
- t_list_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_VERTEX_BUFFER);
+ t_list_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_VERTEX_BUFFERS);
t_offset = lp_build_const_int32(gallivm, input_index);
args[0] = t_list;
args[1] = attribute_offset;
args[2] = buffer_index;
- input = build_intrinsic(gallivm->builder,
+ input = lp_build_intrinsic(gallivm->builder,
"llvm.SI.vs.load.input", vec4_type, args, 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
}
}
+static LLVMValueRef get_primitive_id(struct lp_build_tgsi_context *bld_base,
+ unsigned swizzle)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+
+ if (swizzle > 0)
+ return bld_base->uint_bld.zero;
+
+ switch (si_shader_ctx->type) {
+ case TGSI_PROCESSOR_VERTEX:
+ return LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ si_shader_ctx->param_vs_prim_id);
+ case TGSI_PROCESSOR_TESS_CTRL:
+ return LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ SI_PARAM_PATCH_ID);
+ case TGSI_PROCESSOR_TESS_EVAL:
+ return LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ si_shader_ctx->param_tes_patch_id);
+ case TGSI_PROCESSOR_GEOMETRY:
+ return LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ SI_PARAM_PRIMITIVE_ID);
+ default:
+ assert(0);
+ return bld_base->uint_bld.zero;
+ }
+}
+
+/**
+ * Return the value of tgsi_ind_register for indexing.
+ * This is the indirect index with the constant offset added to it.
+ */
+static LLVMValueRef get_indirect_index(struct si_shader_context *si_shader_ctx,
+ const struct tgsi_ind_register *ind,
+ int rel_index)
+{
+ struct gallivm_state *gallivm = si_shader_ctx->radeon_bld.soa.bld_base.base.gallivm;
+ LLVMValueRef result;
+
+ result = si_shader_ctx->radeon_bld.soa.addr[ind->Index][ind->Swizzle];
+ result = LLVMBuildLoad(gallivm->builder, result, "");
+ result = LLVMBuildAdd(gallivm->builder, result,
+ lp_build_const_int32(gallivm, rel_index), "");
+ return result;
+}
+
+/**
+ * Calculate a dword address given an input or output register and a stride.
+ */
+static LLVMValueRef get_dw_address(struct si_shader_context *si_shader_ctx,
+ const struct tgsi_full_dst_register *dst,
+ const struct tgsi_full_src_register *src,
+ LLVMValueRef vertex_dw_stride,
+ LLVMValueRef base_addr)
+{
+ struct gallivm_state *gallivm = si_shader_ctx->radeon_bld.soa.bld_base.base.gallivm;
+ struct tgsi_shader_info *info = &si_shader_ctx->shader->selector->info;
+ ubyte *name, *index, *array_first;
+ int first, param;
+ struct tgsi_full_dst_register reg;
+
+ /* Set the register description. The address computation is the same
+ * for sources and destinations. */
+ if (src) {
+ reg.Register.File = src->Register.File;
+ reg.Register.Index = src->Register.Index;
+ reg.Register.Indirect = src->Register.Indirect;
+ reg.Register.Dimension = src->Register.Dimension;
+ reg.Indirect = src->Indirect;
+ reg.Dimension = src->Dimension;
+ reg.DimIndirect = src->DimIndirect;
+ } else
+ reg = *dst;
+
+ /* If the register is 2-dimensional (e.g. an array of vertices
+ * in a primitive), calculate the base address of the vertex. */
+ if (reg.Register.Dimension) {
+ LLVMValueRef index;
+
+ if (reg.Dimension.Indirect)
+ index = get_indirect_index(si_shader_ctx, ®.DimIndirect,
+ reg.Dimension.Index);
+ else
+ index = lp_build_const_int32(gallivm, reg.Dimension.Index);
+
+ base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
+ LLVMBuildMul(gallivm->builder, index,
+ vertex_dw_stride, ""), "");
+ }
+
+ /* Get information about the register. */
+ if (reg.Register.File == TGSI_FILE_INPUT) {
+ name = info->input_semantic_name;
+ index = info->input_semantic_index;
+ array_first = info->input_array_first;
+ } else if (reg.Register.File == TGSI_FILE_OUTPUT) {
+ name = info->output_semantic_name;
+ index = info->output_semantic_index;
+ array_first = info->output_array_first;
+ } else {
+ assert(0);
+ return NULL;
+ }
+
+ if (reg.Register.Indirect) {
+ /* Add the relative address of the element. */
+ LLVMValueRef ind_index;
+
+ if (reg.Indirect.ArrayID)
+ first = array_first[reg.Indirect.ArrayID];
+ else
+ first = reg.Register.Index;
+
+ ind_index = get_indirect_index(si_shader_ctx, ®.Indirect,
+ reg.Register.Index - first);
+
+ base_addr = LLVMBuildAdd(gallivm->builder, base_addr,
+ LLVMBuildMul(gallivm->builder, ind_index,
+ lp_build_const_int32(gallivm, 4), ""), "");
+
+ param = si_shader_io_get_unique_index(name[first], index[first]);
+ } else {
+ param = si_shader_io_get_unique_index(name[reg.Register.Index],
+ index[reg.Register.Index]);
+ }
+
+ /* Add the base address of the element. */
+ return LLVMBuildAdd(gallivm->builder, base_addr,
+ lp_build_const_int32(gallivm, param * 4), "");
+}
+
+/**
+ * Load from LDS.
+ *
+ * \param type output value type
+ * \param swizzle offset (typically 0..3); it can be ~0, which loads a vec4
+ * \param dw_addr address in dwords
+ */
+static LLVMValueRef lds_load(struct lp_build_tgsi_context *bld_base,
+ enum tgsi_opcode_type type, unsigned swizzle,
+ LLVMValueRef dw_addr)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMValueRef value;
+
+ if (swizzle == ~0) {
+ LLVMValueRef values[TGSI_NUM_CHANNELS];
+
+ for (unsigned chan = 0; chan < TGSI_NUM_CHANNELS; chan++)
+ values[chan] = lds_load(bld_base, type, chan, dw_addr);
+
+ return lp_build_gather_values(bld_base->base.gallivm, values,
+ TGSI_NUM_CHANNELS);
+ }
+
+ dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
+ lp_build_const_int32(gallivm, swizzle));
+
+ value = build_indexed_load(si_shader_ctx, si_shader_ctx->lds, dw_addr);
+ if (type == TGSI_TYPE_DOUBLE) {
+ LLVMValueRef value2;
+ dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
+ lp_build_const_int32(gallivm, swizzle + 1));
+ value2 = build_indexed_load(si_shader_ctx, si_shader_ctx->lds, dw_addr);
+ return radeon_llvm_emit_fetch_double(bld_base, value, value2);
+ }
+
+ return LLVMBuildBitCast(gallivm->builder, value,
+ tgsi2llvmtype(bld_base, type), "");
+}
+
+/**
+ * Store to LDS.
+ *
+ * \param swizzle offset (typically 0..3)
+ * \param dw_addr address in dwords
+ * \param value value to store
+ */
+static void lds_store(struct lp_build_tgsi_context * bld_base,
+ unsigned swizzle, LLVMValueRef dw_addr,
+ LLVMValueRef value)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+
+ dw_addr = lp_build_add(&bld_base->uint_bld, dw_addr,
+ lp_build_const_int32(gallivm, swizzle));
+
+ value = LLVMBuildBitCast(gallivm->builder, value,
+ LLVMInt32TypeInContext(gallivm->context), "");
+ build_indexed_store(si_shader_ctx, si_shader_ctx->lds,
+ dw_addr, value);
+}
+
+static LLVMValueRef fetch_input_tcs(
+ struct lp_build_tgsi_context *bld_base,
+ const struct tgsi_full_src_register *reg,
+ enum tgsi_opcode_type type, unsigned swizzle)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ LLVMValueRef dw_addr, stride;
+
+ stride = unpack_param(si_shader_ctx, SI_PARAM_TCS_IN_LAYOUT, 13, 8);
+ dw_addr = get_tcs_in_current_patch_offset(si_shader_ctx);
+ dw_addr = get_dw_address(si_shader_ctx, NULL, reg, stride, dw_addr);
+
+ return lds_load(bld_base, type, swizzle, dw_addr);
+}
+
+static LLVMValueRef fetch_output_tcs(
+ struct lp_build_tgsi_context *bld_base,
+ const struct tgsi_full_src_register *reg,
+ enum tgsi_opcode_type type, unsigned swizzle)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ LLVMValueRef dw_addr, stride;
+
+ if (reg->Register.Dimension) {
+ stride = unpack_param(si_shader_ctx, SI_PARAM_TCS_OUT_LAYOUT, 13, 8);
+ dw_addr = get_tcs_out_current_patch_offset(si_shader_ctx);
+ dw_addr = get_dw_address(si_shader_ctx, NULL, reg, stride, dw_addr);
+ } else {
+ dw_addr = get_tcs_out_current_patch_data_offset(si_shader_ctx);
+ dw_addr = get_dw_address(si_shader_ctx, NULL, reg, NULL, dw_addr);
+ }
+
+ return lds_load(bld_base, type, swizzle, dw_addr);
+}
+
+static LLVMValueRef fetch_input_tes(
+ struct lp_build_tgsi_context *bld_base,
+ const struct tgsi_full_src_register *reg,
+ enum tgsi_opcode_type type, unsigned swizzle)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ LLVMValueRef dw_addr, stride;
+
+ if (reg->Register.Dimension) {
+ stride = unpack_param(si_shader_ctx, SI_PARAM_TCS_OUT_LAYOUT, 13, 8);
+ dw_addr = get_tcs_out_current_patch_offset(si_shader_ctx);
+ dw_addr = get_dw_address(si_shader_ctx, NULL, reg, stride, dw_addr);
+ } else {
+ dw_addr = get_tcs_out_current_patch_data_offset(si_shader_ctx);
+ dw_addr = get_dw_address(si_shader_ctx, NULL, reg, NULL, dw_addr);
+ }
+
+ return lds_load(bld_base, type, swizzle, dw_addr);
+}
+
+static void store_output_tcs(struct lp_build_tgsi_context * bld_base,
+ const struct tgsi_full_instruction * inst,
+ const struct tgsi_opcode_info * info,
+ LLVMValueRef dst[4])
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ const struct tgsi_full_dst_register *reg = &inst->Dst[0];
+ unsigned chan_index;
+ LLVMValueRef dw_addr, stride;
+
+ /* Only handle per-patch and per-vertex outputs here.
+ * Vectors will be lowered to scalars and this function will be called again.
+ */
+ if (reg->Register.File != TGSI_FILE_OUTPUT ||
+ (dst[0] && LLVMGetTypeKind(LLVMTypeOf(dst[0])) == LLVMVectorTypeKind)) {
+ radeon_llvm_emit_store(bld_base, inst, info, dst);
+ return;
+ }
+
+ if (reg->Register.Dimension) {
+ stride = unpack_param(si_shader_ctx, SI_PARAM_TCS_OUT_LAYOUT, 13, 8);
+ dw_addr = get_tcs_out_current_patch_offset(si_shader_ctx);
+ dw_addr = get_dw_address(si_shader_ctx, reg, NULL, stride, dw_addr);
+ } else {
+ dw_addr = get_tcs_out_current_patch_data_offset(si_shader_ctx);
+ dw_addr = get_dw_address(si_shader_ctx, reg, NULL, NULL, dw_addr);
+ }
+
+ TGSI_FOR_EACH_DST0_ENABLED_CHANNEL(inst, chan_index) {
+ LLVMValueRef value = dst[chan_index];
+
+ if (inst->Instruction.Saturate)
+ value = radeon_llvm_saturate(bld_base, value);
+
+ lds_store(bld_base, chan_index, dw_addr, value);
+ }
+}
+
static LLVMValueRef fetch_input_gs(
struct lp_build_tgsi_context *bld_base,
const struct tgsi_full_src_register *reg,
struct tgsi_shader_info *info = &shader->selector->info;
unsigned semantic_name = info->input_semantic_name[reg->Register.Index];
unsigned semantic_index = info->input_semantic_index[reg->Register.Index];
+ unsigned param;
+ LLVMValueRef value;
- if (swizzle != ~0 && semantic_name == TGSI_SEMANTIC_PRIMID) {
- if (swizzle == 0)
- return LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
- SI_PARAM_PRIMITIVE_ID);
- else
- return uint->zero;
- }
+ if (swizzle != ~0 && semantic_name == TGSI_SEMANTIC_PRIMID)
+ return get_primitive_id(bld_base, swizzle);
if (!reg->Register.Dimension)
return NULL;
vtx_offset_param),
4);
+ param = si_shader_io_get_unique_index(semantic_name, semantic_index);
args[0] = si_shader_ctx->esgs_ring;
args[1] = vtx_offset;
- args[2] = lp_build_const_int32(gallivm,
- (get_param_index(semantic_name, semantic_index,
- shader->selector->gs_used_inputs) * 4 +
- swizzle) * 256);
+ args[2] = lp_build_const_int32(gallivm, (param * 4 + swizzle) * 256);
args[3] = uint->zero;
args[4] = uint->one; /* OFFEN */
args[5] = uint->zero; /* IDXEN */
args[7] = uint->zero; /* SLC */
args[8] = uint->zero; /* TFE */
+ value = lp_build_intrinsic(gallivm->builder,
+ "llvm.SI.buffer.load.dword.i32.i32",
+ i32, args, 9,
+ LLVMReadOnlyAttribute | LLVMNoUnwindAttribute);
+ if (type == TGSI_TYPE_DOUBLE) {
+ LLVMValueRef value2;
+ args[2] = lp_build_const_int32(gallivm, (param * 4 + swizzle + 1) * 256);
+ value2 = lp_build_intrinsic(gallivm->builder,
+ "llvm.SI.buffer.load.dword.i32.i32",
+ i32, args, 9,
+ LLVMReadOnlyAttribute | LLVMNoUnwindAttribute);
+ return radeon_llvm_emit_fetch_double(bld_base,
+ value, value2);
+ }
return LLVMBuildBitCast(gallivm->builder,
- build_intrinsic(gallivm->builder,
- "llvm.SI.buffer.load.dword.i32.i32",
- i32, args, 9,
- LLVMReadOnlyAttribute | LLVMNoUnwindAttribute),
+ value,
tgsi2llvmtype(bld_base, type), "");
}
+static int lookup_interp_param_index(unsigned interpolate, unsigned location)
+{
+ switch (interpolate) {
+ case TGSI_INTERPOLATE_CONSTANT:
+ return 0;
+
+ case TGSI_INTERPOLATE_LINEAR:
+ if (location == TGSI_INTERPOLATE_LOC_SAMPLE)
+ return SI_PARAM_LINEAR_SAMPLE;
+ else if (location == TGSI_INTERPOLATE_LOC_CENTROID)
+ return SI_PARAM_LINEAR_CENTROID;
+ else
+ return SI_PARAM_LINEAR_CENTER;
+ break;
+ case TGSI_INTERPOLATE_COLOR:
+ case TGSI_INTERPOLATE_PERSPECTIVE:
+ if (location == TGSI_INTERPOLATE_LOC_SAMPLE)
+ return SI_PARAM_PERSP_SAMPLE;
+ else if (location == TGSI_INTERPOLATE_LOC_CENTROID)
+ return SI_PARAM_PERSP_CENTROID;
+ else
+ return SI_PARAM_PERSP_CENTER;
+ break;
+ default:
+ fprintf(stderr, "Warning: Unhandled interpolation mode.\n");
+ return -1;
+ }
+}
+
+/* This shouldn't be used by explicit INTERP opcodes. */
+static LLVMValueRef get_interp_param(struct si_shader_context *si_shader_ctx,
+ unsigned param)
+{
+ struct gallivm_state *gallivm = &si_shader_ctx->radeon_bld.gallivm;
+ unsigned sample_param = 0;
+ LLVMValueRef default_ij, sample_ij, force_sample;
+
+ default_ij = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, param);
+
+ /* If the shader doesn't use center/centroid, just return the parameter.
+ *
+ * If the shader only uses one set of (i,j), "si_emit_spi_ps_input" can
+ * switch between center/centroid and sample without shader changes.
+ */
+ switch (param) {
+ case SI_PARAM_PERSP_CENTROID:
+ case SI_PARAM_PERSP_CENTER:
+ if (!si_shader_ctx->shader->selector->forces_persample_interp_for_persp)
+ return default_ij;
+
+ sample_param = SI_PARAM_PERSP_SAMPLE;
+ break;
+
+ case SI_PARAM_LINEAR_CENTROID:
+ case SI_PARAM_LINEAR_CENTER:
+ if (!si_shader_ctx->shader->selector->forces_persample_interp_for_linear)
+ return default_ij;
+
+ sample_param = SI_PARAM_LINEAR_SAMPLE;
+ break;
+
+ default:
+ return default_ij;
+ }
+
+ /* Otherwise, we have to select (i,j) based on a user data SGPR. */
+ sample_ij = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, sample_param);
+
+ /* TODO: this can be done more efficiently by switching between
+ * 2 prologs.
+ */
+ force_sample = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ SI_PARAM_PS_STATE_BITS);
+ force_sample = LLVMBuildTrunc(gallivm->builder, force_sample,
+ LLVMInt1TypeInContext(gallivm->context), "");
+ return LLVMBuildSelect(gallivm->builder, force_sample,
+ sample_ij, default_ij, "");
+}
+
static void declare_input_fs(
struct radeon_llvm_context *radeon_bld,
unsigned input_index,
LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context);
LLVMValueRef main_fn = radeon_bld->main_fn;
- LLVMValueRef interp_param;
+ LLVMValueRef interp_param = NULL;
+ int interp_param_idx;
const char * intr_name;
/* This value is:
attr_number = lp_build_const_int32(gallivm,
shader->ps_input_param_offset[input_index]);
- switch (decl->Interp.Interpolate) {
- case TGSI_INTERPOLATE_CONSTANT:
- interp_param = 0;
- break;
- case TGSI_INTERPOLATE_LINEAR:
- if (decl->Interp.Location == TGSI_INTERPOLATE_LOC_SAMPLE)
- interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_SAMPLE);
- else if (decl->Interp.Location == TGSI_INTERPOLATE_LOC_CENTROID)
- interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_CENTROID);
- else
- interp_param = LLVMGetParam(main_fn, SI_PARAM_LINEAR_CENTER);
- break;
- case TGSI_INTERPOLATE_COLOR:
- case TGSI_INTERPOLATE_PERSPECTIVE:
- if (decl->Interp.Location == TGSI_INTERPOLATE_LOC_SAMPLE)
- interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_SAMPLE);
- else if (decl->Interp.Location == TGSI_INTERPOLATE_LOC_CENTROID)
- interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTROID);
- else
- interp_param = LLVMGetParam(main_fn, SI_PARAM_PERSP_CENTER);
- break;
- default:
- fprintf(stderr, "Warning: Unhandled interpolation mode.\n");
+ shader->ps_input_interpolate[input_index] = decl->Interp.Interpolate;
+ interp_param_idx = lookup_interp_param_index(decl->Interp.Interpolate,
+ decl->Interp.Location);
+ if (interp_param_idx == -1)
return;
- }
+ else if (interp_param_idx)
+ interp_param = get_interp_param(si_shader_ctx, interp_param_idx);
/* fs.constant returns the param from the middle vertex, so it's not
* really useful for flat shading. It's meant to be used for custom
args[0] = llvm_chan;
args[1] = attr_number;
- front = build_intrinsic(gallivm->builder, intr_name,
+ front = lp_build_intrinsic(gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
args[1] = back_attr_number;
- back = build_intrinsic(gallivm->builder, intr_name,
+ back = lp_build_intrinsic(gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
args[2] = params;
args[3] = interp_param;
radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 0)] =
- build_intrinsic(gallivm->builder, intr_name,
+ lp_build_intrinsic(gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
radeon_bld->inputs[radeon_llvm_reg_index_soa(input_index, 1)] =
args[2] = params;
args[3] = interp_param;
radeon_bld->inputs[soa_index] =
- build_intrinsic(gallivm->builder, intr_name,
+ lp_build_intrinsic(gallivm->builder, intr_name,
input_type, args, args[3] ? 4 : 3,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
}
{
LLVMValueRef args[2] = {resource, offset};
- return build_intrinsic(builder, "llvm.SI.load.const", return_type, args, 2,
+ return lp_build_intrinsic(builder, "llvm.SI.load.const", return_type, args, 2,
LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
}
+static LLVMValueRef load_sample_position(struct radeon_llvm_context *radeon_bld, LLVMValueRef sample_id)
+{
+ struct si_shader_context *si_shader_ctx =
+ si_shader_context(&radeon_bld->soa.bld_base);
+ struct lp_build_context *uint_bld = &radeon_bld->soa.bld_base.uint_bld;
+ struct gallivm_state *gallivm = &radeon_bld->gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMValueRef desc = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST_BUFFERS);
+ LLVMValueRef buf_index = lp_build_const_int32(gallivm, SI_DRIVER_STATE_CONST_BUF);
+ LLVMValueRef resource = build_indexed_load_const(si_shader_ctx, desc, buf_index);
+
+ /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
+ LLVMValueRef offset0 = lp_build_mul_imm(uint_bld, sample_id, 8);
+ LLVMValueRef offset1 = LLVMBuildAdd(builder, offset0, lp_build_const_int32(gallivm, 4), "");
+
+ LLVMValueRef pos[4] = {
+ buffer_load_const(builder, resource, offset0, radeon_bld->soa.bld_base.base.elem_type),
+ buffer_load_const(builder, resource, offset1, radeon_bld->soa.bld_base.base.elem_type),
+ lp_build_const_float(gallivm, 0),
+ lp_build_const_float(gallivm, 0)
+ };
+
+ return lp_build_gather_values(gallivm, pos, 4);
+}
+
static void declare_system_value(
struct radeon_llvm_context * radeon_bld,
unsigned index,
{
struct si_shader_context *si_shader_ctx =
si_shader_context(&radeon_bld->soa.bld_base);
+ struct lp_build_context *bld = &radeon_bld->soa.bld_base.base;
struct lp_build_context *uint_bld = &radeon_bld->soa.bld_base.uint_bld;
struct gallivm_state *gallivm = &radeon_bld->gallivm;
LLVMValueRef value = 0;
break;
case TGSI_SEMANTIC_INVOCATIONID:
- value = LLVMGetParam(radeon_bld->main_fn,
- SI_PARAM_GS_INSTANCE_ID);
+ if (si_shader_ctx->type == TGSI_PROCESSOR_TESS_CTRL)
+ value = unpack_param(si_shader_ctx, SI_PARAM_REL_IDS, 8, 5);
+ else if (si_shader_ctx->type == TGSI_PROCESSOR_GEOMETRY)
+ value = LLVMGetParam(radeon_bld->main_fn,
+ SI_PARAM_GS_INSTANCE_ID);
+ else
+ assert(!"INVOCATIONID not implemented");
+ break;
+
+ case TGSI_SEMANTIC_SAMPLEID:
+ value = get_sample_id(radeon_bld);
+ break;
+
+ case TGSI_SEMANTIC_SAMPLEPOS:
+ value = load_sample_position(radeon_bld, get_sample_id(radeon_bld));
+ break;
+
+ case TGSI_SEMANTIC_SAMPLEMASK:
+ /* Smoothing isn't MSAA in GL, but it's MSAA in hardware.
+ * Therefore, force gl_SampleMaskIn to 1 for GL. */
+ if (si_shader_ctx->shader->key.ps.poly_line_smoothing)
+ value = uint_bld->one;
+ else
+ value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_SAMPLE_COVERAGE);
+ break;
+
+ case TGSI_SEMANTIC_TESSCOORD:
+ {
+ LLVMValueRef coord[4] = {
+ LLVMGetParam(radeon_bld->main_fn, si_shader_ctx->param_tes_u),
+ LLVMGetParam(radeon_bld->main_fn, si_shader_ctx->param_tes_v),
+ bld->zero,
+ bld->zero
+ };
+
+ /* For triangles, the vector should be (u, v, 1-u-v). */
+ if (si_shader_ctx->shader->selector->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] ==
+ PIPE_PRIM_TRIANGLES)
+ coord[2] = lp_build_sub(bld, bld->one,
+ lp_build_add(bld, coord[0], coord[1]));
+
+ value = lp_build_gather_values(gallivm, coord, 4);
break;
+ }
- case TGSI_SEMANTIC_SAMPLEID:
- value = get_sample_id(radeon_bld);
+ case TGSI_SEMANTIC_VERTICESIN:
+ value = unpack_param(si_shader_ctx, SI_PARAM_TCS_OUT_LAYOUT, 26, 6);
break;
- case TGSI_SEMANTIC_SAMPLEPOS:
+ case TGSI_SEMANTIC_TESSINNER:
+ case TGSI_SEMANTIC_TESSOUTER:
{
- LLVMBuilderRef builder = gallivm->builder;
- LLVMValueRef desc = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST);
- LLVMValueRef buf_index = lp_build_const_int32(gallivm, SI_DRIVER_STATE_CONST_BUF);
- LLVMValueRef resource = build_indexed_load_const(si_shader_ctx, desc, buf_index);
-
- /* offset = sample_id * 8 (8 = 2 floats containing samplepos.xy) */
- LLVMValueRef offset0 = lp_build_mul_imm(uint_bld, get_sample_id(radeon_bld), 8);
- LLVMValueRef offset1 = LLVMBuildAdd(builder, offset0, lp_build_const_int32(gallivm, 4), "");
-
- LLVMValueRef pos[4] = {
- buffer_load_const(builder, resource, offset0, radeon_bld->soa.bld_base.base.elem_type),
- buffer_load_const(builder, resource, offset1, radeon_bld->soa.bld_base.base.elem_type),
- lp_build_const_float(gallivm, 0),
- lp_build_const_float(gallivm, 0)
- };
- value = lp_build_gather_values(gallivm, pos, 4);
+ LLVMValueRef dw_addr;
+ int param = si_shader_io_get_unique_index(decl->Semantic.Name, 0);
+
+ dw_addr = get_tcs_out_current_patch_data_offset(si_shader_ctx);
+ dw_addr = LLVMBuildAdd(gallivm->builder, dw_addr,
+ lp_build_const_int32(gallivm, param * 4), "");
+
+ value = lds_load(&radeon_bld->soa.bld_base, TGSI_TYPE_FLOAT,
+ ~0, dw_addr);
break;
}
- case TGSI_SEMANTIC_SAMPLEMASK:
- /* Smoothing isn't MSAA in GL, but it's MSAA in hardware.
- * Therefore, force gl_SampleMaskIn to 1 for GL. */
- if (si_shader_ctx->shader->key.ps.poly_line_smoothing)
- value = uint_bld->one;
- else
- value = LLVMGetParam(radeon_bld->main_fn, SI_PARAM_SAMPLE_COVERAGE);
+ case TGSI_SEMANTIC_PRIMID:
+ value = get_primitive_id(&radeon_bld->soa.bld_base, 0);
break;
default:
const struct tgsi_ind_register *ireg = ®->Indirect;
unsigned buf, idx;
- LLVMValueRef addr;
+ LLVMValueRef addr, bufp;
LLVMValueRef result;
if (swizzle == LP_CHAN_ALL) {
buf = reg->Register.Dimension ? reg->Dimension.Index : 0;
idx = reg->Register.Index * 4 + swizzle;
- if (!reg->Register.Indirect) {
+ if (!reg->Register.Indirect && !reg->Dimension.Indirect) {
if (type != TGSI_TYPE_DOUBLE)
return bitcast(bld_base, type, si_shader_ctx->constants[buf][idx]);
else {
}
}
+ if (reg->Register.Dimension && reg->Dimension.Indirect) {
+ LLVMValueRef ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST_BUFFERS);
+ LLVMValueRef index;
+ index = get_indirect_index(si_shader_ctx, ®->DimIndirect,
+ reg->Dimension.Index);
+ bufp = build_indexed_load_const(si_shader_ctx, ptr, index);
+ } else
+ bufp = si_shader_ctx->const_buffers[buf];
+
addr = si_shader_ctx->radeon_bld.soa.addr[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,
lp_build_const_int32(base->gallivm, idx * 4));
- result = buffer_load_const(base->gallivm->builder, si_shader_ctx->const_resource[buf],
+ result = buffer_load_const(base->gallivm->builder, bufp,
addr, bld_base->base.elem_type);
if (type != TGSI_TYPE_DOUBLE)
addr2 = lp_build_add(&bld_base->uint_bld, addr2,
lp_build_const_int32(base->gallivm, idx * 4));
- result2 = buffer_load_const(base->gallivm->builder, si_shader_ctx->const_resource[buf],
+ result2 = buffer_load_const(base->gallivm->builder, si_shader_ctx->const_buffers[buf],
addr2, bld_base->base.elem_type);
result = radeon_llvm_emit_fetch_double(bld_base,
unsigned compressed = 0;
unsigned chan;
+ /* XXX: This controls which components of the output
+ * registers actually get exported. (e.g bit 0 means export
+ * X component, bit 1 means export Y component, etc.) I'm
+ * hard coding this to 0xf for now. In the future, we might
+ * want to do something else.
+ */
+ args[0] = lp_build_const_int32(base->gallivm, 0xf);
+
+ /* Specify whether the EXEC mask represents the valid mask */
+ args[1] = uint->zero;
+
+ /* Specify whether this is the last export */
+ args[2] = uint->zero;
+
+ /* Specify the target we are exporting */
+ args[3] = lp_build_const_int32(base->gallivm, target);
+
if (si_shader_ctx->type == TGSI_PROCESSOR_FRAGMENT) {
int cbuf = target - V_008DFC_SQ_EXP_MRT;
}
}
+ /* Set COMPR flag */
+ args[4] = compressed ? uint->one : uint->zero;
+
if (compressed) {
/* Pixel shader needs to pack output values before export */
- for (chan = 0; chan < 2; chan++ ) {
- args[0] = values[2 * chan];
- args[1] = values[2 * chan + 1];
- args[chan + 5] =
- build_intrinsic(base->gallivm->builder,
- "llvm.SI.packf16",
- LLVMInt32TypeInContext(base->gallivm->context),
- args, 2,
- LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
+ for (chan = 0; chan < 2; chan++) {
+ LLVMValueRef pack_args[2] = {
+ values[2 * chan],
+ values[2 * chan + 1]
+ };
+ LLVMValueRef packed;
+
+ packed = lp_build_intrinsic(base->gallivm->builder,
+ "llvm.SI.packf16",
+ LLVMInt32TypeInContext(base->gallivm->context),
+ pack_args, 2,
+ LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
args[chan + 7] = args[chan + 5] =
LLVMBuildBitCast(base->gallivm->builder,
- args[chan + 5],
+ packed,
LLVMFloatTypeInContext(base->gallivm->context),
"");
}
-
- /* Set COMPR flag */
- args[4] = uint->one;
- } else {
- for (chan = 0; chan < 4; chan++ )
- /* +5 because the first output value will be
- * the 6th argument to the intrinsic. */
- args[chan + 5] = values[chan];
-
- /* Clear COMPR flag */
- args[4] = uint->zero;
- }
-
- /* XXX: This controls which components of the output
- * registers actually get exported. (e.g bit 0 means export
- * X component, bit 1 means export Y component, etc.) I'm
- * hard coding this to 0xf for now. In the future, we might
- * want to do something else. */
- args[0] = lp_build_const_int32(base->gallivm, 0xf);
-
- /* Specify whether the EXEC mask represents the valid mask */
- args[1] = uint->zero;
-
- /* Specify whether this is the last export */
- args[2] = uint->zero;
-
- /* Specify the target we are exporting */
- args[3] = lp_build_const_int32(base->gallivm, target);
-
- /* XXX: We probably need to keep track of the output
- * values, so we know what we are passing to the next
- * stage. */
+ } else
+ memcpy(&args[5], values, sizeof(values[0]) * 4);
}
/* Load from output pointers and initialize arguments for the shader export intrinsic */
lp_build_const_float(gallivm, 1.0f),
lp_build_const_float(gallivm, -1.0f));
- build_intrinsic(gallivm->builder,
+ lp_build_intrinsic(gallivm->builder,
"llvm.AMDGPU.kill",
LLVMVoidTypeInContext(gallivm->context),
&arg, 1, 0);
} else {
- build_intrinsic(gallivm->builder,
+ lp_build_intrinsic(gallivm->builder,
"llvm.AMDGPU.kilp",
LLVMVoidTypeInContext(gallivm->context),
NULL, 0, 0);
SI_PARAM_SAMPLE_COVERAGE);
coverage = bitcast(bld_base, TGSI_TYPE_SIGNED, coverage);
- coverage = build_intrinsic(gallivm->builder, "llvm.ctpop.i32",
+ coverage = lp_build_intrinsic(gallivm->builder, "llvm.ctpop.i32",
bld_base->int_bld.elem_type,
&coverage, 1, LLVMReadNoneAttribute);
unsigned chan;
unsigned const_chan;
LLVMValueRef base_elt;
- LLVMValueRef ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST);
+ LLVMValueRef ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST_BUFFERS);
LLVMValueRef constbuf_index = lp_build_const_int32(base->gallivm, SI_DRIVER_STATE_CONST_BUF);
LLVMValueRef const_resource = build_indexed_load_const(si_shader_ctx, ptr, constbuf_index);
lp_build_intrinsic(gallivm->builder, name,
LLVMVoidTypeInContext(gallivm->context),
- args, Elements(args));
+ args, Elements(args), 0);
}
-static void build_streamout_store(struct si_shader_context *shader,
- LLVMValueRef rsrc,
- LLVMValueRef vdata,
- unsigned num_channels,
- LLVMValueRef vaddr,
- LLVMValueRef soffset,
- unsigned inst_offset)
+static void build_tbuffer_store_dwords(struct si_shader_context *shader,
+ LLVMValueRef rsrc,
+ LLVMValueRef vdata,
+ unsigned num_channels,
+ LLVMValueRef vaddr,
+ LLVMValueRef soffset,
+ unsigned inst_offset)
{
static unsigned dfmt[] = {
V_008F0C_BUF_DATA_FORMAT_32,
LLVMValueRef so_vtx_count =
unpack_param(shader, shader->param_streamout_config, 16, 7);
- LLVMValueRef tid = build_intrinsic(builder, "llvm.SI.tid", i32,
+ LLVMValueRef tid = lp_build_intrinsic(builder, "llvm.SI.tid", i32,
NULL, 0, LLVMReadNoneAttribute);
/* can_emit = tid < so_vtx_count; */
LLVMValueRef can_emit =
LLVMBuildICmp(builder, LLVMIntULT, tid, so_vtx_count, "");
+ LLVMValueRef stream_id =
+ unpack_param(shader, shader->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. */
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)
break;
}
- build_streamout_store(shader, shader->so_buffers[buf_idx],
- vdata, num_comps,
- so_write_offset[buf_idx],
- LLVMConstInt(i32, 0, 0),
- so->output[i].dst_offset*4);
+ 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(shader, shader->so_buffers[buf_idx],
+ vdata, num_comps,
+ so_write_offset[buf_idx],
+ LLVMConstInt(i32, 0, 0),
+ so->output[i].dst_offset*4);
+ lp_build_endif(&if_ctx_stream);
}
}
lp_build_endif(&if_ctx);
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- args, 9);
+ args, 9, 0);
}
if (semantic_name == TGSI_SEMANTIC_CLIPDIST) {
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- pos_args[i], 9);
+ pos_args[i], 9, 0);
+ }
+}
+
+/* This only writes the tessellation factor levels. */
+static void si_llvm_emit_tcs_epilogue(struct lp_build_tgsi_context *bld_base)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct si_shader *shader = si_shader_ctx->shader;
+ unsigned tess_inner_index, tess_outer_index;
+ LLVMValueRef lds_base, lds_inner, lds_outer;
+ LLVMValueRef tf_base, rel_patch_id, byteoffset, buffer, rw_buffers;
+ LLVMValueRef out[6], vec0, vec1, invocation_id;
+ unsigned stride, outer_comps, inner_comps, i;
+ struct lp_build_if_state if_ctx;
+
+ invocation_id = unpack_param(si_shader_ctx, SI_PARAM_REL_IDS, 8, 5);
+
+ /* Do this only for invocation 0, because the tess levels are per-patch,
+ * not per-vertex.
+ *
+ * This can't jump, because invocation 0 executes this. It should
+ * at least mask out the loads and stores for other invocations.
+ */
+ lp_build_if(&if_ctx, gallivm,
+ LLVMBuildICmp(gallivm->builder, LLVMIntEQ,
+ invocation_id, bld_base->uint_bld.zero, ""));
+
+ /* Determine the layout of one tess factor element in the buffer. */
+ switch (shader->key.tcs.prim_mode) {
+ case PIPE_PRIM_LINES:
+ stride = 2; /* 2 dwords, 1 vec2 store */
+ outer_comps = 2;
+ inner_comps = 0;
+ break;
+ case PIPE_PRIM_TRIANGLES:
+ stride = 4; /* 4 dwords, 1 vec4 store */
+ outer_comps = 3;
+ inner_comps = 1;
+ break;
+ case PIPE_PRIM_QUADS:
+ stride = 6; /* 6 dwords, 2 stores (vec4 + vec2) */
+ outer_comps = 4;
+ inner_comps = 2;
+ break;
+ default:
+ assert(0);
+ return;
+ }
+
+ /* Load tess_inner and tess_outer from LDS.
+ * Any invocation can write them, so we can't get them from a temporary.
+ */
+ tess_inner_index = si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSINNER, 0);
+ tess_outer_index = si_shader_io_get_unique_index(TGSI_SEMANTIC_TESSOUTER, 0);
+
+ lds_base = get_tcs_out_current_patch_data_offset(si_shader_ctx);
+ lds_inner = LLVMBuildAdd(gallivm->builder, lds_base,
+ lp_build_const_int32(gallivm,
+ tess_inner_index * 4), "");
+ lds_outer = LLVMBuildAdd(gallivm->builder, lds_base,
+ lp_build_const_int32(gallivm,
+ tess_outer_index * 4), "");
+
+ for (i = 0; i < outer_comps; i++)
+ out[i] = lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_outer);
+ for (i = 0; i < inner_comps; i++)
+ out[outer_comps+i] = lds_load(bld_base, TGSI_TYPE_SIGNED, i, lds_inner);
+
+ /* Convert the outputs to vectors for stores. */
+ vec0 = lp_build_gather_values(gallivm, out, MIN2(stride, 4));
+ vec1 = NULL;
+
+ if (stride > 4)
+ vec1 = lp_build_gather_values(gallivm, out+4, stride - 4);
+
+ /* Get the buffer. */
+ rw_buffers = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ SI_PARAM_RW_BUFFERS);
+ buffer = build_indexed_load_const(si_shader_ctx, rw_buffers,
+ lp_build_const_int32(gallivm, SI_RING_TESS_FACTOR));
+
+ /* Get the offset. */
+ tf_base = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ SI_PARAM_TESS_FACTOR_OFFSET);
+ rel_patch_id = get_rel_patch_id(si_shader_ctx);
+ byteoffset = LLVMBuildMul(gallivm->builder, rel_patch_id,
+ lp_build_const_int32(gallivm, 4 * stride), "");
+
+ /* Store the outputs. */
+ build_tbuffer_store_dwords(si_shader_ctx, buffer, vec0,
+ MIN2(stride, 4), byteoffset, tf_base, 0);
+ if (vec1)
+ build_tbuffer_store_dwords(si_shader_ctx, buffer, vec1,
+ stride - 4, byteoffset, tf_base, 16);
+ lp_build_endif(&if_ctx);
+}
+
+static void si_llvm_emit_ls_epilogue(struct lp_build_tgsi_context * bld_base)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct si_shader *shader = si_shader_ctx->shader;
+ struct tgsi_shader_info *info = &shader->selector->info;
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ unsigned i, chan;
+ LLVMValueRef vertex_id = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ si_shader_ctx->param_rel_auto_id);
+ LLVMValueRef vertex_dw_stride =
+ unpack_param(si_shader_ctx, SI_PARAM_LS_OUT_LAYOUT, 13, 8);
+ LLVMValueRef base_dw_addr = LLVMBuildMul(gallivm->builder, vertex_id,
+ vertex_dw_stride, "");
+
+ /* 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 = si_shader_ctx->radeon_bld.soa.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);
+ LLVMValueRef dw_addr = LLVMBuildAdd(gallivm->builder, base_dw_addr,
+ lp_build_const_int32(gallivm, param * 4), "");
+
+ for (chan = 0; chan < 4; chan++) {
+ lds_store(bld_base, chan, dw_addr,
+ LLVMBuildLoad(gallivm->builder, out_ptr[chan], ""));
+ }
}
}
info->output_semantic_name[i] == TGSI_SEMANTIC_LAYER)
continue;
- param_index = get_param_index(info->output_semantic_name[i],
- info->output_semantic_index[i],
- es->key.vs.gs_used_inputs);
- if (param_index < 0)
- continue;
+ param_index = si_shader_io_get_unique_index(info->output_semantic_name[i],
+ info->output_semantic_index[i]);
for (chan = 0; chan < 4; chan++) {
LLVMValueRef out_val = LLVMBuildLoad(gallivm->builder, out_ptr[chan], "");
args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_NOP | SENDMSG_GS_DONE);
args[1] = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_GS_WAVE_ID);
- build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
+ lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
LLVMVoidTypeInContext(gallivm->context), args, 2,
LLVMNoUnwindAttribute);
}
struct si_shader_output_values *outputs = NULL;
int i,j;
- outputs = MALLOC(info->num_outputs * sizeof(outputs[0]));
+ outputs = MALLOC((info->num_outputs + 1) * sizeof(outputs[0]));
+
+ /* Vertex color clamping.
+ *
+ * This uses a state constant loaded in a user data SGPR and
+ * an IF statement is added that clamps all colors if the constant
+ * is true.
+ */
+ if (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX &&
+ !si_shader_ctx->shader->is_gs_copy_shader) {
+ struct lp_build_if_state if_ctx;
+ LLVMValueRef cond = NULL;
+ LLVMValueRef addr, val;
+
+ for (i = 0; i < info->num_outputs; i++) {
+ if (info->output_semantic_name[i] != TGSI_SEMANTIC_COLOR &&
+ info->output_semantic_name[i] != TGSI_SEMANTIC_BCOLOR)
+ continue;
+
+ /* We've found a color. */
+ if (!cond) {
+ /* The state is in the first bit of the user SGPR. */
+ cond = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
+ SI_PARAM_VS_STATE_BITS);
+ cond = LLVMBuildTrunc(gallivm->builder, cond,
+ LLVMInt1TypeInContext(gallivm->context), "");
+ lp_build_if(&if_ctx, gallivm, cond);
+ }
+
+ for (j = 0; j < 4; j++) {
+ addr = si_shader_ctx->radeon_bld.soa.outputs[i][j];
+ val = LLVMBuildLoad(gallivm->builder, addr, "");
+ val = radeon_llvm_saturate(bld_base, val);
+ LLVMBuildStore(gallivm->builder, val, addr);
+ }
+ }
+
+ if (cond)
+ lp_build_endif(&if_ctx);
+ }
for (i = 0; i < info->num_outputs; i++) {
outputs[i].name = info->output_semantic_name[i];
"");
}
- si_llvm_export_vs(bld_base, outputs, info->num_outputs);
+ /* Export PrimitiveID when PS needs it. */
+ if (si_vs_exports_prim_id(si_shader_ctx->shader)) {
+ outputs[i].name = TGSI_SEMANTIC_PRIMID;
+ outputs[i].sid = 0;
+ outputs[i].values[0] = bitcast(bld_base, TGSI_TYPE_FLOAT,
+ get_primitive_id(bld_base, 0));
+ outputs[i].values[1] = bld_base->base.undef;
+ outputs[i].values[2] = bld_base->base.undef;
+ outputs[i].values[3] = bld_base->base.undef;
+ i++;
+ }
+
+ si_llvm_export_vs(bld_base, outputs, i);
FREE(outputs);
}
struct lp_build_context * base = &bld_base->base;
struct lp_build_context * uint = &bld_base->uint_bld;
struct tgsi_shader_info *info = &shader->selector->info;
+ LLVMBuilderRef builder = base->gallivm->builder;
LLVMValueRef args[9];
LLVMValueRef last_args[9] = { 0 };
int depth_index = -1, stencil_index = -1, samplemask_index = -1;
target = V_008DFC_SQ_EXP_MRT + semantic_index;
alpha_ptr = si_shader_ctx->radeon_bld.soa.outputs[i][3];
+ if (si_shader_ctx->shader->key.ps.clamp_color) {
+ for (int j = 0; j < 4; j++) {
+ LLVMValueRef ptr = si_shader_ctx->radeon_bld.soa.outputs[i][j];
+ LLVMValueRef result = LLVMBuildLoad(builder, ptr, "");
+
+ result = radeon_llvm_saturate(bld_base, result);
+ LLVMBuildStore(builder, result, ptr);
+ }
+ }
+
if (si_shader_ctx->shader->key.ps.alpha_to_one)
LLVMBuildStore(base->gallivm->builder,
base->one, alpha_ptr);
if (si_shader_ctx->shader->key.ps.poly_line_smoothing)
si_scale_alpha_by_sample_mask(bld_base, alpha_ptr);
+
break;
default:
target = 0;
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- last_args, 9);
+ last_args, 9, 0);
}
/* This instruction will be emitted at the end of the shader. */
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- args, 9);
+ args, 9, 0);
}
}
} else {
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- args, 9);
+ args, 9, 0);
}
}
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- args, 9);
+ args, 9, 0);
else
memcpy(last_args, args, sizeof(args));
}
lp_build_intrinsic(base->gallivm->builder,
"llvm.SI.export",
LLVMVoidTypeInContext(base->gallivm->context),
- last_args, 9);
+ last_args, 9, 0);
}
static void build_tex_intrinsic(const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data);
-static bool tgsi_is_shadow_sampler(unsigned target)
+static bool tgsi_is_array_sampler(unsigned target)
{
- return target == TGSI_TEXTURE_SHADOW1D ||
+ return target == TGSI_TEXTURE_1D_ARRAY ||
target == TGSI_TEXTURE_SHADOW1D_ARRAY ||
- target == TGSI_TEXTURE_SHADOW2D ||
+ target == TGSI_TEXTURE_2D_ARRAY ||
target == TGSI_TEXTURE_SHADOW2D_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE ||
+ target == TGSI_TEXTURE_CUBE_ARRAY ||
target == TGSI_TEXTURE_SHADOWCUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWRECT;
+ target == TGSI_TEXTURE_2D_ARRAY_MSAA;
+}
+
+static void set_tex_fetch_args(struct gallivm_state *gallivm,
+ struct lp_build_emit_data *emit_data,
+ unsigned opcode, unsigned target,
+ LLVMValueRef res_ptr, LLVMValueRef samp_ptr,
+ LLVMValueRef *param, unsigned count,
+ unsigned dmask)
+{
+ unsigned num_args;
+ unsigned is_rect = target == TGSI_TEXTURE_RECT;
+ LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context);
+
+ /* Pad to power of two vector */
+ while (count < util_next_power_of_two(count))
+ param[count++] = LLVMGetUndef(i32);
+
+ /* Texture coordinates. */
+ if (count > 1)
+ emit_data->args[0] = lp_build_gather_values(gallivm, param, count);
+ else
+ emit_data->args[0] = param[0];
+
+ /* Resource. */
+ emit_data->args[1] = res_ptr;
+ num_args = 2;
+
+ if (opcode == TGSI_OPCODE_TXF || opcode == TGSI_OPCODE_TXQ)
+ emit_data->dst_type = LLVMVectorType(i32, 4);
+ else {
+ emit_data->dst_type = LLVMVectorType(
+ LLVMFloatTypeInContext(gallivm->context), 4);
+
+ emit_data->args[num_args++] = samp_ptr;
+ }
+
+ emit_data->args[num_args++] = lp_build_const_int32(gallivm, dmask);
+ emit_data->args[num_args++] = lp_build_const_int32(gallivm, is_rect); /* unorm */
+ emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* r128 */
+ emit_data->args[num_args++] = lp_build_const_int32(gallivm,
+ tgsi_is_array_sampler(target)); /* da */
+ emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* glc */
+ emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* slc */
+ emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* tfe */
+ emit_data->args[num_args++] = lp_build_const_int32(gallivm, 0); /* lwe */
+
+ emit_data->arg_count = num_args;
}
static const struct lp_build_tgsi_action tex_action;
+static void tex_fetch_ptrs(
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data,
+ LLVMValueRef *res_ptr, LLVMValueRef *samp_ptr, LLVMValueRef *fmask_ptr)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ const struct tgsi_full_instruction * inst = emit_data->inst;
+ unsigned target = inst->Texture.Texture;
+ unsigned sampler_src;
+ unsigned sampler_index;
+
+ sampler_src = emit_data->inst->Instruction.NumSrcRegs - 1;
+ sampler_index = emit_data->inst->Src[sampler_src].Register.Index;
+
+ if (emit_data->inst->Src[sampler_src].Register.Indirect) {
+ const struct tgsi_full_src_register *reg = &emit_data->inst->Src[sampler_src];
+ LLVMValueRef ind_index;
+
+ ind_index = get_indirect_index(si_shader_ctx, ®->Indirect, reg->Register.Index);
+
+ *res_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_SAMPLER_VIEWS);
+ *res_ptr = build_indexed_load_const(si_shader_ctx, *res_ptr, ind_index);
+
+ *samp_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_SAMPLER_STATES);
+ *samp_ptr = build_indexed_load_const(si_shader_ctx, *samp_ptr, ind_index);
+
+ if (target == TGSI_TEXTURE_2D_MSAA ||
+ target == TGSI_TEXTURE_2D_ARRAY_MSAA) {
+ ind_index = LLVMBuildAdd(gallivm->builder, ind_index,
+ lp_build_const_int32(gallivm,
+ SI_FMASK_TEX_OFFSET), "");
+ *fmask_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_SAMPLER_VIEWS);
+ *fmask_ptr = build_indexed_load_const(si_shader_ctx, *fmask_ptr, ind_index);
+ }
+ } else {
+ *res_ptr = si_shader_ctx->sampler_views[sampler_index];
+ *samp_ptr = si_shader_ctx->sampler_states[sampler_index];
+ *fmask_ptr = si_shader_ctx->sampler_views[SI_FMASK_TEX_OFFSET + sampler_index];
+ }
+}
+
static void tex_fetch_args(
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
const struct tgsi_full_instruction * inst = emit_data->inst;
unsigned opcode = inst->Instruction.Opcode;
unsigned target = inst->Texture.Texture;
- LLVMValueRef coords[5];
+ LLVMValueRef coords[5], derivs[6];
LLVMValueRef address[16];
int ref_pos;
unsigned num_coords = tgsi_util_get_texture_coord_dim(target, &ref_pos);
unsigned count = 0;
unsigned chan;
- unsigned sampler_src = emit_data->inst->Instruction.NumSrcRegs - 1;
- unsigned sampler_index = emit_data->inst->Src[sampler_src].Register.Index;
- bool has_offset = HAVE_LLVM >= 0x0305 ? inst->Texture.NumOffsets > 0 : false;
+ unsigned num_deriv_channels = 0;
+ bool has_offset = inst->Texture.NumOffsets > 0;
+ LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
+ LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context);
+ unsigned dmask = 0xf;
+
+ tex_fetch_ptrs(bld_base, emit_data, &res_ptr, &samp_ptr, &fmask_ptr);
+
+ if (opcode == TGSI_OPCODE_TXQ) {
+ if (target == TGSI_TEXTURE_BUFFER) {
+ LLVMTypeRef v8i32 = LLVMVectorType(i32, 8);
+
+ /* Read the size from the buffer descriptor directly. */
+ LLVMValueRef res = LLVMBuildBitCast(builder, res_ptr, v8i32, "");
+ LLVMValueRef size = LLVMBuildExtractElement(builder, res,
+ lp_build_const_int32(gallivm, 6), "");
+
+ if (si_shader_ctx->screen->b.chip_class >= VI) {
+ /* On VI, 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.
+ */
+ LLVMValueRef stride =
+ LLVMBuildExtractElement(builder, res,
+ lp_build_const_int32(gallivm, 5), "");
+ stride = LLVMBuildLShr(builder, stride,
+ lp_build_const_int32(gallivm, 16), "");
+ stride = LLVMBuildAnd(builder, stride,
+ lp_build_const_int32(gallivm, 0x3FFF), "");
+
+ size = LLVMBuildUDiv(builder, size, stride, "");
+ }
+
+ emit_data->args[0] = size;
+ return;
+ }
+
+ /* Textures - set the mip level. */
+ address[count++] = lp_build_emit_fetch(bld_base, inst, 0, TGSI_CHAN_X);
+
+ set_tex_fetch_args(gallivm, emit_data, opcode, target, res_ptr,
+ NULL, address, count, 0xf);
+ return;
+ }
if (target == TGSI_TEXTURE_BUFFER) {
LLVMTypeRef i128 = LLVMIntTypeInContext(gallivm->context, 128);
LLVMTypeRef v16i8 = LLVMVectorType(i8, 16);
/* Bitcast and truncate v8i32 to v16i8. */
- LLVMValueRef res = si_shader_ctx->resources[sampler_index];
+ LLVMValueRef res = res_ptr;
res = LLVMBuildBitCast(gallivm->builder, res, v2i128, "");
res = LLVMBuildExtractElement(gallivm->builder, res, bld_base->uint_bld.one, "");
res = LLVMBuildBitCast(gallivm->builder, res, v16i8, "");
address[count++] = lp_build_emit_fetch(bld_base, inst, 1, 0);
/* Pack depth comparison value */
- if (tgsi_is_shadow_sampler(target) && opcode != TGSI_OPCODE_LODQ) {
+ if (tgsi_is_shadow_target(target) && opcode != TGSI_OPCODE_LODQ) {
if (target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
address[count++] = lp_build_emit_fetch(bld_base, inst, 1, 0);
} else {
}
}
- if (target == TGSI_TEXTURE_CUBE ||
- target == TGSI_TEXTURE_CUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
- radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords);
-
/* Pack user derivatives */
if (opcode == TGSI_OPCODE_TXD) {
- int num_deriv_channels, param;
+ int param, num_src_deriv_channels;
switch (target) {
case TGSI_TEXTURE_3D:
+ num_src_deriv_channels = 3;
num_deriv_channels = 3;
break;
case TGSI_TEXTURE_2D:
case TGSI_TEXTURE_SHADOWRECT:
case TGSI_TEXTURE_2D_ARRAY:
case TGSI_TEXTURE_SHADOW2D_ARRAY:
+ num_src_deriv_channels = 2;
+ num_deriv_channels = 2;
+ break;
case TGSI_TEXTURE_CUBE:
case TGSI_TEXTURE_SHADOWCUBE:
case TGSI_TEXTURE_CUBE_ARRAY:
case TGSI_TEXTURE_SHADOWCUBE_ARRAY:
+ /* Cube derivatives will be converted to 2D. */
+ num_src_deriv_channels = 3;
num_deriv_channels = 2;
break;
case TGSI_TEXTURE_1D:
case TGSI_TEXTURE_SHADOW1D:
case TGSI_TEXTURE_1D_ARRAY:
case TGSI_TEXTURE_SHADOW1D_ARRAY:
+ num_src_deriv_channels = 1;
num_deriv_channels = 1;
break;
default:
- assert(0); /* no other targets are valid here */
+ unreachable("invalid target");
}
- for (param = 1; param <= 2; param++)
- for (chan = 0; chan < num_deriv_channels; chan++)
- address[count++] = lp_build_emit_fetch(bld_base, inst, param, chan);
+ for (param = 0; param < 2; param++)
+ for (chan = 0; chan < num_src_deriv_channels; chan++)
+ derivs[param * num_src_deriv_channels + chan] =
+ lp_build_emit_fetch(bld_base, inst, param+1, chan);
}
+ if (target == TGSI_TEXTURE_CUBE ||
+ target == TGSI_TEXTURE_CUBE_ARRAY ||
+ target == TGSI_TEXTURE_SHADOWCUBE ||
+ target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
+ radeon_llvm_emit_prepare_cube_coords(bld_base, emit_data, coords, derivs);
+
+ if (opcode == TGSI_OPCODE_TXD)
+ for (int i = 0; i < num_deriv_channels * 2; i++)
+ address[count++] = derivs[i];
+
/* Pack texture coordinates */
address[count++] = coords[0];
if (num_coords > 1)
for (chan = 0; chan < count; chan++ ) {
address[chan] = LLVMBuildBitCast(gallivm->builder,
- address[chan],
- LLVMInt32TypeInContext(gallivm->context),
- "");
+ address[chan], i32, "");
}
/* Adjust the sample index according to FMASK.
}
txf_address[3] = bld_base->uint_bld.zero;
- /* Pad to a power-of-two size. */
- while (txf_count < util_next_power_of_two(txf_count))
- txf_address[txf_count++] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
-
/* Read FMASK using TXF. */
inst.Instruction.Opcode = TGSI_OPCODE_TXF;
- inst.Texture.Texture = target == TGSI_TEXTURE_2D_MSAA ? TGSI_TEXTURE_2D : TGSI_TEXTURE_2D_ARRAY;
+ inst.Texture.Texture = target;
txf_emit_data.inst = &inst;
txf_emit_data.chan = 0;
- txf_emit_data.dst_type = LLVMVectorType(
- LLVMInt32TypeInContext(gallivm->context), 4);
- txf_emit_data.args[0] = lp_build_gather_values(gallivm, txf_address, txf_count);
- txf_emit_data.args[1] = si_shader_ctx->resources[SI_FMASK_TEX_OFFSET + sampler_index];
- txf_emit_data.args[2] = lp_build_const_int32(gallivm, inst.Texture.Texture);
- txf_emit_data.arg_count = 3;
-
+ set_tex_fetch_args(gallivm, &txf_emit_data, TGSI_OPCODE_TXF,
+ target, fmask_ptr, NULL,
+ txf_address, txf_count, 0xf);
build_tex_intrinsic(&tex_action, bld_base, &txf_emit_data);
/* Initialize some constants. */
* resource descriptor is 0 (invalid),
*/
LLVMValueRef fmask_desc =
- LLVMBuildBitCast(gallivm->builder,
- si_shader_ctx->resources[SI_FMASK_TEX_OFFSET + sampler_index],
+ LLVMBuildBitCast(gallivm->builder, fmask_ptr,
LLVMVectorType(uint_bld->elem_type, 8), "");
LLVMValueRef fmask_word1 =
final_sample, address[sample_chan], "");
}
- /* Resource */
- emit_data->args[1] = si_shader_ctx->resources[sampler_index];
-
if (opcode == TGSI_OPCODE_TXF) {
/* add tex offsets */
if (inst->Texture.NumOffsets) {
case TGSI_TEXTURE_SHADOW2D:
case TGSI_TEXTURE_RECT:
case TGSI_TEXTURE_SHADOWRECT:
- case TGSI_TEXTURE_2D_ARRAY:
- case TGSI_TEXTURE_SHADOW2D_ARRAY:
- address[1] =
- lp_build_add(uint_bld, address[1],
- bld->immediates[off->Index][off->SwizzleY]);
- /* fall through */
- case TGSI_TEXTURE_1D:
- case TGSI_TEXTURE_SHADOW1D:
- case TGSI_TEXTURE_1D_ARRAY:
- case TGSI_TEXTURE_SHADOW1D_ARRAY:
- address[0] =
- lp_build_add(uint_bld, address[0],
- bld->immediates[off->Index][off->SwizzleX]);
- break;
- /* texture offsets do not apply to other texture targets */
- }
- }
-
- emit_data->args[2] = lp_build_const_int32(gallivm, target);
- emit_data->arg_count = 3;
-
- emit_data->dst_type = LLVMVectorType(
- LLVMInt32TypeInContext(gallivm->context),
- 4);
- } else if (opcode == TGSI_OPCODE_TG4 ||
- opcode == TGSI_OPCODE_LODQ ||
- has_offset) {
- unsigned is_array = target == TGSI_TEXTURE_1D_ARRAY ||
- target == TGSI_TEXTURE_SHADOW1D_ARRAY ||
- target == TGSI_TEXTURE_2D_ARRAY ||
- target == TGSI_TEXTURE_SHADOW2D_ARRAY ||
- target == TGSI_TEXTURE_CUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY;
- unsigned is_rect = target == TGSI_TEXTURE_RECT;
- unsigned dmask = 0xf;
-
- if (opcode == TGSI_OPCODE_TG4) {
- unsigned gather_comp = 0;
-
- /* 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.
- */
-
- /* Get the component index from src1.x for Gather4. */
- if (!tgsi_is_shadow_sampler(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];
- gather_comp = LLVMConstIntGetZExtValue(comp_imm);
- gather_comp = CLAMP(gather_comp, 0, 3);
- }
-
- dmask = 1 << gather_comp;
- }
-
- emit_data->args[2] = si_shader_ctx->samplers[sampler_index];
- emit_data->args[3] = lp_build_const_int32(gallivm, dmask);
- emit_data->args[4] = lp_build_const_int32(gallivm, is_rect); /* unorm */
- emit_data->args[5] = lp_build_const_int32(gallivm, 0); /* r128 */
- emit_data->args[6] = lp_build_const_int32(gallivm, is_array); /* da */
- emit_data->args[7] = lp_build_const_int32(gallivm, 0); /* glc */
- emit_data->args[8] = lp_build_const_int32(gallivm, 0); /* slc */
- emit_data->args[9] = lp_build_const_int32(gallivm, 0); /* tfe */
- emit_data->args[10] = lp_build_const_int32(gallivm, 0); /* lwe */
-
- emit_data->arg_count = 11;
-
- emit_data->dst_type = LLVMVectorType(
- LLVMFloatTypeInContext(gallivm->context),
- 4);
- } else {
- emit_data->args[2] = si_shader_ctx->samplers[sampler_index];
- emit_data->args[3] = lp_build_const_int32(gallivm, target);
- emit_data->arg_count = 4;
-
- emit_data->dst_type = LLVMVectorType(
- LLVMFloatTypeInContext(gallivm->context),
- 4);
- }
-
- /* The fetch opcode has been converted to a 2D array fetch.
- * This simplifies the LLVM backend. */
- if (target == TGSI_TEXTURE_CUBE_ARRAY)
- target = TGSI_TEXTURE_2D_ARRAY;
- else if (target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
- target = TGSI_TEXTURE_SHADOW2D_ARRAY;
-
- /* Pad to power of two vector */
- while (count < util_next_power_of_two(count))
- address[count++] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
-
- emit_data->args[0] = lp_build_gather_values(gallivm, address, count);
-}
-
-static void build_tex_intrinsic(const struct lp_build_tgsi_action * action,
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
-{
- struct lp_build_context * base = &bld_base->base;
- unsigned opcode = emit_data->inst->Instruction.Opcode;
- unsigned target = emit_data->inst->Texture.Texture;
- char intr_name[127];
- bool has_offset = HAVE_LLVM >= 0x0305 ?
- emit_data->inst->Texture.NumOffsets > 0 : false;
-
- if (target == TGSI_TEXTURE_BUFFER) {
- emit_data->output[emit_data->chan] = build_intrinsic(
- base->gallivm->builder,
- "llvm.SI.vs.load.input", emit_data->dst_type,
- emit_data->args, emit_data->arg_count,
- LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
- return;
- }
-
- if (opcode == TGSI_OPCODE_TG4 ||
- opcode == TGSI_OPCODE_LODQ ||
- (opcode != TGSI_OPCODE_TXF && has_offset)) {
- bool is_shadow = tgsi_is_shadow_sampler(target);
- const char *name = "llvm.SI.image.sample";
- const char *infix = "";
-
- switch (opcode) {
- case TGSI_OPCODE_TEX:
- case TGSI_OPCODE_TEX2:
- case TGSI_OPCODE_TXP:
- break;
- case TGSI_OPCODE_TXB:
- case TGSI_OPCODE_TXB2:
- infix = ".b";
- break;
- case TGSI_OPCODE_TXL:
- case TGSI_OPCODE_TXL2:
- infix = ".l";
- break;
- case TGSI_OPCODE_TXD:
- infix = ".d";
- break;
- case TGSI_OPCODE_TG4:
- name = "llvm.SI.gather4";
- break;
- case TGSI_OPCODE_LODQ:
- name = "llvm.SI.getlod";
- is_shadow = false;
- has_offset = false;
- break;
- default:
- assert(0);
- return;
- }
-
- /* Add the type and suffixes .c, .o if needed. */
- sprintf(intr_name, "%s%s%s%s.v%ui32", name,
- is_shadow ? ".c" : "", infix, has_offset ? ".o" : "",
- LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0])));
-
- emit_data->output[emit_data->chan] = build_intrinsic(
- base->gallivm->builder, intr_name, emit_data->dst_type,
- emit_data->args, emit_data->arg_count,
- LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
- } else {
- LLVMTypeRef i8, v16i8, v32i8;
- const char *name;
-
- switch (opcode) {
- case TGSI_OPCODE_TEX:
- case TGSI_OPCODE_TEX2:
- case TGSI_OPCODE_TXP:
- name = "llvm.SI.sample";
- break;
- case TGSI_OPCODE_TXB:
- case TGSI_OPCODE_TXB2:
- name = "llvm.SI.sampleb";
- break;
- case TGSI_OPCODE_TXD:
- name = "llvm.SI.sampled";
- break;
- case TGSI_OPCODE_TXF:
- name = "llvm.SI.imageload";
- break;
- case TGSI_OPCODE_TXL:
- case TGSI_OPCODE_TXL2:
- name = "llvm.SI.samplel";
- break;
- default:
- assert(0);
- return;
- }
-
- i8 = LLVMInt8TypeInContext(base->gallivm->context);
- v16i8 = LLVMVectorType(i8, 16);
- v32i8 = LLVMVectorType(i8, 32);
-
- emit_data->args[1] = LLVMBuildBitCast(base->gallivm->builder,
- emit_data->args[1], v32i8, "");
- if (opcode != TGSI_OPCODE_TXF) {
- emit_data->args[2] = LLVMBuildBitCast(base->gallivm->builder,
- emit_data->args[2], v16i8, "");
+ case TGSI_TEXTURE_2D_ARRAY:
+ case TGSI_TEXTURE_SHADOW2D_ARRAY:
+ address[1] =
+ lp_build_add(uint_bld, address[1],
+ bld->immediates[off->Index][off->SwizzleY]);
+ /* fall through */
+ case TGSI_TEXTURE_1D:
+ case TGSI_TEXTURE_SHADOW1D:
+ case TGSI_TEXTURE_1D_ARRAY:
+ case TGSI_TEXTURE_SHADOW1D_ARRAY:
+ address[0] =
+ lp_build_add(uint_bld, address[0],
+ bld->immediates[off->Index][off->SwizzleX]);
+ break;
+ /* texture offsets do not apply to other texture targets */
+ }
}
-
- sprintf(intr_name, "%s.v%ui32", name,
- LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0])));
-
- emit_data->output[emit_data->chan] = build_intrinsic(
- base->gallivm->builder, intr_name, emit_data->dst_type,
- emit_data->args, emit_data->arg_count,
- LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
}
-}
-
-static void txq_fetch_args(
- struct lp_build_tgsi_context * bld_base,
- struct lp_build_emit_data * emit_data)
-{
- struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
- const struct tgsi_full_instruction *inst = emit_data->inst;
- struct gallivm_state *gallivm = bld_base->base.gallivm;
- unsigned target = inst->Texture.Texture;
- if (target == TGSI_TEXTURE_BUFFER) {
- LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context);
- LLVMTypeRef v8i32 = LLVMVectorType(i32, 8);
-
- /* Read the size from the buffer descriptor directly. */
- LLVMValueRef size = si_shader_ctx->resources[inst->Src[1].Register.Index];
- size = LLVMBuildBitCast(gallivm->builder, size, v8i32, "");
- size = LLVMBuildExtractElement(gallivm->builder, size,
- lp_build_const_int32(gallivm, 6), "");
- emit_data->args[0] = size;
- return;
- }
+ if (opcode == TGSI_OPCODE_TG4) {
+ unsigned gather_comp = 0;
- /* Mip level */
- emit_data->args[0] = lp_build_emit_fetch(bld_base, inst, 0, TGSI_CHAN_X);
+ /* 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.
+ */
- /* Resource */
- emit_data->args[1] = si_shader_ctx->resources[inst->Src[1].Register.Index];
+ /* 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;
- /* Texture target */
- if (target == TGSI_TEXTURE_CUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)
- target = TGSI_TEXTURE_2D_ARRAY;
+ assert(src1.File == TGSI_FILE_IMMEDIATE);
- emit_data->args[2] = lp_build_const_int32(bld_base->base.gallivm,
- target);
+ comp_imm = imms[src1.Index][src1.SwizzleX];
+ gather_comp = LLVMConstIntGetZExtValue(comp_imm);
+ gather_comp = CLAMP(gather_comp, 0, 3);
+ }
- emit_data->arg_count = 3;
+ dmask = 1 << gather_comp;
+ }
- emit_data->dst_type = LLVMVectorType(
- LLVMInt32TypeInContext(bld_base->base.gallivm->context),
- 4);
+ set_tex_fetch_args(gallivm, emit_data, opcode, target, res_ptr,
+ samp_ptr, address, count, dmask);
}
-static void build_txq_intrinsic(const struct lp_build_tgsi_action * action,
+static void build_tex_intrinsic(const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
struct lp_build_emit_data * emit_data)
{
+ struct lp_build_context * base = &bld_base->base;
+ unsigned opcode = emit_data->inst->Instruction.Opcode;
unsigned target = emit_data->inst->Texture.Texture;
+ char intr_name[127];
+ bool has_offset = emit_data->inst->Texture.NumOffsets > 0;
+ bool is_shadow = tgsi_is_shadow_target(target);
+ char type[64];
+ const char *name = "llvm.SI.image.sample";
+ const char *infix = "";
- if (target == TGSI_TEXTURE_BUFFER) {
+ if (opcode == TGSI_OPCODE_TXQ && target == TGSI_TEXTURE_BUFFER) {
/* Just return the buffer size. */
emit_data->output[emit_data->chan] = emit_data->args[0];
return;
}
- build_tgsi_intrinsic_nomem(action, bld_base, emit_data);
+ if (target == TGSI_TEXTURE_BUFFER) {
+ emit_data->output[emit_data->chan] = lp_build_intrinsic(
+ base->gallivm->builder,
+ "llvm.SI.vs.load.input", emit_data->dst_type,
+ emit_data->args, emit_data->arg_count,
+ LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
+ return;
+ }
+
+ switch (opcode) {
+ case TGSI_OPCODE_TXF:
+ name = target == TGSI_TEXTURE_2D_MSAA ||
+ target == TGSI_TEXTURE_2D_ARRAY_MSAA ?
+ "llvm.SI.image.load" :
+ "llvm.SI.image.load.mip";
+ is_shadow = false;
+ has_offset = false;
+ break;
+ case TGSI_OPCODE_TXQ:
+ name = "llvm.SI.getresinfo";
+ is_shadow = false;
+ has_offset = false;
+ break;
+ case TGSI_OPCODE_LODQ:
+ name = "llvm.SI.getlod";
+ is_shadow = false;
+ has_offset = false;
+ break;
+ case TGSI_OPCODE_TEX:
+ case TGSI_OPCODE_TEX2:
+ case TGSI_OPCODE_TXP:
+ break;
+ case TGSI_OPCODE_TXB:
+ case TGSI_OPCODE_TXB2:
+ infix = ".b";
+ break;
+ case TGSI_OPCODE_TXL:
+ case TGSI_OPCODE_TXL2:
+ infix = ".l";
+ break;
+ case TGSI_OPCODE_TXD:
+ infix = ".d";
+ break;
+ case TGSI_OPCODE_TG4:
+ name = "llvm.SI.gather4";
+ break;
+ default:
+ assert(0);
+ return;
+ }
+
+ if (LLVMGetTypeKind(LLVMTypeOf(emit_data->args[0])) == LLVMVectorTypeKind)
+ sprintf(type, ".v%ui32",
+ LLVMGetVectorSize(LLVMTypeOf(emit_data->args[0])));
+ else
+ strcpy(type, ".i32");
+
+ /* Add the type and suffixes .c, .o if needed. */
+ sprintf(intr_name, "%s%s%s%s%s",
+ name, is_shadow ? ".c" : "", infix,
+ has_offset ? ".o" : "", type);
+
+ emit_data->output[emit_data->chan] = lp_build_intrinsic(
+ base->gallivm->builder, intr_name, emit_data->dst_type,
+ emit_data->args, emit_data->arg_count,
+ LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
/* Divide the number of layers by 6 to get the number of cubes. */
- if (target == TGSI_TEXTURE_CUBE_ARRAY ||
- target == TGSI_TEXTURE_SHADOWCUBE_ARRAY) {
+ if (opcode == TGSI_OPCODE_TXQ &&
+ (target == TGSI_TEXTURE_CUBE_ARRAY ||
+ target == TGSI_TEXTURE_SHADOWCUBE_ARRAY)) {
LLVMBuilderRef builder = bld_base->base.gallivm->builder;
LLVMValueRef two = lp_build_const_int32(bld_base->base.gallivm, 2);
LLVMValueRef six = lp_build_const_int32(bld_base->base.gallivm, 6);
}
}
+static void si_llvm_emit_txqs(
+ const struct lp_build_tgsi_action * action,
+ struct lp_build_tgsi_context * bld_base,
+ struct lp_build_emit_data * emit_data)
+{
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMTypeRef i32 = LLVMInt32TypeInContext(gallivm->context);
+ LLVMTypeRef v8i32 = LLVMVectorType(i32, 8);
+ LLVMValueRef res, samples;
+ LLVMValueRef res_ptr, samp_ptr, fmask_ptr = NULL;
+
+ tex_fetch_ptrs(bld_base, emit_data, &res_ptr, &samp_ptr, &fmask_ptr);
+
+
+ /* Read the samples from the descriptor directly. */
+ res = LLVMBuildBitCast(builder, res_ptr, v8i32, "");
+ samples = LLVMBuildExtractElement(
+ builder, res,
+ lp_build_const_int32(gallivm, 3), "");
+ samples = LLVMBuildLShr(builder, samples,
+ lp_build_const_int32(gallivm, 16), "");
+ samples = LLVMBuildAnd(builder, samples,
+ lp_build_const_int32(gallivm, 0xf), "");
+ samples = LLVMBuildShl(builder, lp_build_const_int32(gallivm, 1),
+ samples, "");
+
+ emit_data->output[emit_data->chan] = samples;
+}
+
+/*
+ * SI implements derivatives using the local data store (LDS)
+ * All writes to the LDS happen in all executing threads at
+ * the same time. TID is the Thread ID for the current
+ * thread and is a value between 0 and 63, representing
+ * the thread's position in the wavefront.
+ *
+ * For the pixel shader threads are grouped into quads of four pixels.
+ * The TIDs of the pixels of a quad are:
+ *
+ * +------+------+
+ * |4n + 0|4n + 1|
+ * +------+------+
+ * |4n + 2|4n + 3|
+ * +------+------+
+ *
+ * So, masking the TID with 0xfffffffc yields the TID of the top left pixel
+ * of the quad, masking with 0xfffffffd yields the TID of the top pixel of
+ * the current pixel's column, and masking with 0xfffffffe yields the TID
+ * of the left pixel of the current pixel's row.
+ *
+ * Adding 1 yields the TID of the pixel to the right of the left pixel, and
+ * adding 2 yields the TID of the pixel below the top pixel.
+ */
+/* masks for thread ID. */
+#define TID_MASK_TOP_LEFT 0xfffffffc
+#define TID_MASK_TOP 0xfffffffd
+#define TID_MASK_LEFT 0xfffffffe
+
static void si_llvm_emit_ddxy(
const struct lp_build_tgsi_action * action,
struct lp_build_tgsi_context * bld_base,
LLVMTypeRef i32;
unsigned swizzle[4];
unsigned c;
+ int idx;
+ unsigned mask;
i32 = LLVMInt32TypeInContext(gallivm->context);
indices[0] = bld_base->uint_bld.zero;
- indices[1] = build_intrinsic(gallivm->builder, "llvm.SI.tid", i32,
+ indices[1] = lp_build_intrinsic(gallivm->builder, "llvm.SI.tid", i32,
NULL, 0, LLVMReadNoneAttribute);
- store_ptr = LLVMBuildGEP(gallivm->builder, si_shader_ctx->ddxy_lds,
+ store_ptr = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
indices, 2, "");
+ if (opcode == TGSI_OPCODE_DDX_FINE)
+ mask = TID_MASK_LEFT;
+ else if (opcode == TGSI_OPCODE_DDY_FINE)
+ mask = TID_MASK_TOP;
+ else
+ mask = TID_MASK_TOP_LEFT;
+
indices[1] = LLVMBuildAnd(gallivm->builder, indices[1],
- lp_build_const_int32(gallivm, 0xfffffffc), "");
- load_ptr0 = LLVMBuildGEP(gallivm->builder, si_shader_ctx->ddxy_lds,
+ lp_build_const_int32(gallivm, mask), "");
+ load_ptr0 = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
indices, 2, "");
+ /* for DDX we want to next X pixel, DDY next Y pixel. */
+ idx = (opcode == TGSI_OPCODE_DDX || opcode == TGSI_OPCODE_DDX_FINE) ? 1 : 2;
indices[1] = LLVMBuildAdd(gallivm->builder, indices[1],
- lp_build_const_int32(gallivm,
- opcode == TGSI_OPCODE_DDX ? 1 : 2),
- "");
- load_ptr1 = LLVMBuildGEP(gallivm->builder, si_shader_ctx->ddxy_lds,
+ lp_build_const_int32(gallivm, idx), "");
+ load_ptr1 = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
indices, 2, "");
for (c = 0; c < 4; ++c) {
emit_data->output[0] = lp_build_gather_values(gallivm, result, 4);
}
+/*
+ * this takes an I,J coordinate pair,
+ * and works out the X and Y derivatives.
+ * it returns DDX(I), DDX(J), DDY(I), DDY(J).
+ */
+static LLVMValueRef si_llvm_emit_ddxy_interp(
+ struct lp_build_tgsi_context *bld_base,
+ LLVMValueRef interp_ij)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ struct lp_build_context *base = &bld_base->base;
+ LLVMValueRef indices[2];
+ LLVMValueRef store_ptr, load_ptr_x, load_ptr_y, load_ptr_ddx, load_ptr_ddy, temp, temp2;
+ LLVMValueRef tl, tr, bl, result[4];
+ LLVMTypeRef i32;
+ unsigned c;
+
+ i32 = LLVMInt32TypeInContext(gallivm->context);
+
+ indices[0] = bld_base->uint_bld.zero;
+ indices[1] = lp_build_intrinsic(gallivm->builder, "llvm.SI.tid", i32,
+ NULL, 0, LLVMReadNoneAttribute);
+ store_ptr = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ temp = LLVMBuildAnd(gallivm->builder, indices[1],
+ lp_build_const_int32(gallivm, TID_MASK_LEFT), "");
+
+ temp2 = LLVMBuildAnd(gallivm->builder, indices[1],
+ lp_build_const_int32(gallivm, TID_MASK_TOP), "");
+
+ indices[1] = temp;
+ load_ptr_x = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ indices[1] = temp2;
+ load_ptr_y = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ indices[1] = LLVMBuildAdd(gallivm->builder, temp,
+ lp_build_const_int32(gallivm, 1), "");
+ load_ptr_ddx = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ indices[1] = LLVMBuildAdd(gallivm->builder, temp2,
+ lp_build_const_int32(gallivm, 2), "");
+ load_ptr_ddy = LLVMBuildGEP(gallivm->builder, si_shader_ctx->lds,
+ indices, 2, "");
+
+ for (c = 0; c < 2; ++c) {
+ LLVMValueRef store_val;
+ LLVMValueRef c_ll = lp_build_const_int32(gallivm, c);
+
+ store_val = LLVMBuildExtractElement(gallivm->builder,
+ interp_ij, c_ll, "");
+ LLVMBuildStore(gallivm->builder,
+ store_val,
+ store_ptr);
+
+ tl = LLVMBuildLoad(gallivm->builder, load_ptr_x, "");
+ tl = LLVMBuildBitCast(gallivm->builder, tl, base->elem_type, "");
+
+ tr = LLVMBuildLoad(gallivm->builder, load_ptr_ddx, "");
+ tr = LLVMBuildBitCast(gallivm->builder, tr, base->elem_type, "");
+
+ result[c] = LLVMBuildFSub(gallivm->builder, tr, tl, "");
+
+ tl = LLVMBuildLoad(gallivm->builder, load_ptr_y, "");
+ tl = LLVMBuildBitCast(gallivm->builder, tl, base->elem_type, "");
+
+ bl = LLVMBuildLoad(gallivm->builder, load_ptr_ddy, "");
+ bl = LLVMBuildBitCast(gallivm->builder, bl, base->elem_type, "");
+
+ result[c + 2] = LLVMBuildFSub(gallivm->builder, bl, tl, "");
+ }
+
+ return lp_build_gather_values(gallivm, result, 4);
+}
+
+static void interp_fetch_args(
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ const struct tgsi_full_instruction *inst = emit_data->inst;
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET) {
+ /* offset is in second src, first two channels */
+ emit_data->args[0] = lp_build_emit_fetch(bld_base,
+ emit_data->inst, 1,
+ 0);
+ emit_data->args[1] = lp_build_emit_fetch(bld_base,
+ emit_data->inst, 1,
+ 1);
+ emit_data->arg_count = 2;
+ } else if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
+ LLVMValueRef sample_position;
+ LLVMValueRef sample_id;
+ LLVMValueRef halfval = lp_build_const_float(gallivm, 0.5f);
+
+ /* fetch sample ID, then fetch its sample position,
+ * and place into first two channels.
+ */
+ sample_id = lp_build_emit_fetch(bld_base,
+ emit_data->inst, 1, 0);
+ sample_id = LLVMBuildBitCast(gallivm->builder, sample_id,
+ LLVMInt32TypeInContext(gallivm->context),
+ "");
+ sample_position = load_sample_position(&si_shader_ctx->radeon_bld, sample_id);
+
+ emit_data->args[0] = LLVMBuildExtractElement(gallivm->builder,
+ sample_position,
+ lp_build_const_int32(gallivm, 0), "");
+
+ emit_data->args[0] = LLVMBuildFSub(gallivm->builder, emit_data->args[0], halfval, "");
+ emit_data->args[1] = LLVMBuildExtractElement(gallivm->builder,
+ sample_position,
+ lp_build_const_int32(gallivm, 1), "");
+ emit_data->args[1] = LLVMBuildFSub(gallivm->builder, emit_data->args[1], halfval, "");
+ emit_data->arg_count = 2;
+ }
+}
+
+static void build_interp_intrinsic(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
+ struct si_shader *shader = si_shader_ctx->shader;
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+ LLVMValueRef interp_param;
+ const struct tgsi_full_instruction *inst = emit_data->inst;
+ const char *intr_name;
+ int input_index;
+ int chan;
+ int i;
+ LLVMValueRef attr_number;
+ LLVMTypeRef input_type = LLVMFloatTypeInContext(gallivm->context);
+ LLVMValueRef params = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_PRIM_MASK);
+ int interp_param_idx;
+ unsigned location;
+
+ assert(inst->Src[0].Register.File == TGSI_FILE_INPUT);
+ input_index = inst->Src[0].Register.Index;
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
+ inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE)
+ location = TGSI_INTERPOLATE_LOC_CENTER;
+ else
+ location = TGSI_INTERPOLATE_LOC_CENTROID;
+
+ interp_param_idx = lookup_interp_param_index(shader->ps_input_interpolate[input_index],
+ location);
+ if (interp_param_idx == -1)
+ return;
+ else if (interp_param_idx)
+ interp_param = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, interp_param_idx);
+ else
+ interp_param = NULL;
+
+ attr_number = lp_build_const_int32(gallivm,
+ shader->ps_input_param_offset[input_index]);
+
+ if (inst->Instruction.Opcode == TGSI_OPCODE_INTERP_OFFSET ||
+ inst->Instruction.Opcode == TGSI_OPCODE_INTERP_SAMPLE) {
+ LLVMValueRef ij_out[2];
+ LLVMValueRef ddxy_out = si_llvm_emit_ddxy_interp(bld_base, interp_param);
+
+ /*
+ * take the I then J parameters, and the DDX/Y for it, and
+ * calculate the IJ inputs for the interpolator.
+ * temp1 = ddx * offset/sample.x + I;
+ * interp_param.I = ddy * offset/sample.y + temp1;
+ * temp1 = ddx * offset/sample.x + J;
+ * interp_param.J = ddy * offset/sample.y + temp1;
+ */
+ for (i = 0; i < 2; i++) {
+ LLVMValueRef ix_ll = lp_build_const_int32(gallivm, i);
+ LLVMValueRef iy_ll = lp_build_const_int32(gallivm, i + 2);
+ LLVMValueRef ddx_el = LLVMBuildExtractElement(gallivm->builder,
+ ddxy_out, ix_ll, "");
+ LLVMValueRef ddy_el = LLVMBuildExtractElement(gallivm->builder,
+ ddxy_out, iy_ll, "");
+ LLVMValueRef interp_el = LLVMBuildExtractElement(gallivm->builder,
+ interp_param, ix_ll, "");
+ LLVMValueRef temp1, temp2;
+
+ interp_el = LLVMBuildBitCast(gallivm->builder, interp_el,
+ LLVMFloatTypeInContext(gallivm->context), "");
+
+ temp1 = LLVMBuildFMul(gallivm->builder, ddx_el, emit_data->args[0], "");
+
+ temp1 = LLVMBuildFAdd(gallivm->builder, temp1, interp_el, "");
+
+ temp2 = LLVMBuildFMul(gallivm->builder, ddy_el, emit_data->args[1], "");
+
+ temp2 = LLVMBuildFAdd(gallivm->builder, temp2, temp1, "");
+
+ ij_out[i] = LLVMBuildBitCast(gallivm->builder,
+ temp2,
+ LLVMIntTypeInContext(gallivm->context, 32), "");
+ }
+ interp_param = lp_build_gather_values(bld_base->base.gallivm, ij_out, 2);
+ }
+
+ intr_name = interp_param ? "llvm.SI.fs.interp" : "llvm.SI.fs.constant";
+ for (chan = 0; chan < 2; chan++) {
+ LLVMValueRef args[4];
+ LLVMValueRef llvm_chan;
+ unsigned schan;
+
+ schan = tgsi_util_get_full_src_register_swizzle(&inst->Src[0], chan);
+ llvm_chan = lp_build_const_int32(gallivm, schan);
+
+ args[0] = llvm_chan;
+ args[1] = attr_number;
+ args[2] = params;
+ args[3] = interp_param;
+
+ emit_data->output[chan] =
+ lp_build_intrinsic(gallivm->builder, intr_name,
+ input_type, args, args[3] ? 4 : 3,
+ LLVMReadNoneAttribute | LLVMNoUnwindAttribute);
+ }
+}
+
+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 tgsi_src_register src0 = emit_data->inst->Src[0].Register;
+ unsigned stream;
+
+ assert(src0.File == TGSI_FILE_IMMEDIATE);
+
+ stream = LLVMConstIntGetZExtValue(imms[src0.Index][src0.SwizzleX]) & 0x3;
+ return stream;
+}
+
/* Emit one vertex from the geometry shader */
static void si_llvm_emit_vertex(
const struct lp_build_tgsi_action *action,
LLVMValueRef args[2];
unsigned chan;
int i;
+ unsigned stream;
+
+ stream = si_llvm_get_stream(bld_base, emit_data);
/* Write vertex attribute values to GSVS ring */
- gs_next_vertex = LLVMBuildLoad(gallivm->builder, si_shader_ctx->gs_next_vertex, "");
+ gs_next_vertex = LLVMBuildLoad(gallivm->builder,
+ si_shader_ctx->gs_next_vertex[stream],
+ "");
/* If this thread has already emitted the declared maximum number of
* vertices, kill it: excessive vertex emissions are not supposed to
kill = lp_build_select(&bld_base->base, can_emit,
lp_build_const_float(gallivm, 1.0f),
lp_build_const_float(gallivm, -1.0f));
- build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill",
- LLVMVoidTypeInContext(gallivm->context), &kill, 1, 0);
+
+ lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.kill",
+ LLVMVoidTypeInContext(gallivm->context), &kill, 1, 0);
for (i = 0; i < info->num_outputs; i++) {
LLVMValueRef *out_ptr =
out_val = LLVMBuildBitCast(gallivm->builder, out_val, i32, "");
build_tbuffer_store(si_shader_ctx,
- si_shader_ctx->gsvs_ring,
+ si_shader_ctx->gsvs_ring[stream],
out_val, 1,
voffset, soffset, 0,
V_008F0C_BUF_DATA_FORMAT_32,
}
gs_next_vertex = lp_build_add(uint, gs_next_vertex,
lp_build_const_int32(gallivm, 1));
- LLVMBuildStore(gallivm->builder, gs_next_vertex, si_shader_ctx->gs_next_vertex);
+
+ LLVMBuildStore(gallivm->builder, gs_next_vertex, si_shader_ctx->gs_next_vertex[stream]);
/* Signal vertex emission */
- args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_EMIT | SENDMSG_GS);
+ args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_EMIT | SENDMSG_GS | (stream << 8));
args[1] = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_GS_WAVE_ID);
- build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
+ lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
LLVMVoidTypeInContext(gallivm->context), args, 2,
LLVMNoUnwindAttribute);
}
struct si_shader_context *si_shader_ctx = si_shader_context(bld_base);
struct gallivm_state *gallivm = bld_base->base.gallivm;
LLVMValueRef args[2];
+ unsigned stream;
/* Signal primitive cut */
- args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_CUT | SENDMSG_GS);
+ stream = si_llvm_get_stream(bld_base, emit_data);
+ args[0] = lp_build_const_int32(gallivm, SENDMSG_GS_OP_CUT | SENDMSG_GS | (stream << 8));
args[1] = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_GS_WAVE_ID);
- build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
+ lp_build_intrinsic(gallivm->builder, "llvm.SI.sendmsg",
LLVMVoidTypeInContext(gallivm->context), args, 2,
LLVMNoUnwindAttribute);
}
+static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
+ struct lp_build_tgsi_context *bld_base,
+ struct lp_build_emit_data *emit_data)
+{
+ struct gallivm_state *gallivm = bld_base->base.gallivm;
+
+ lp_build_intrinsic(gallivm->builder, "llvm.AMDGPU.barrier.local",
+ LLVMVoidTypeInContext(gallivm->context), NULL, 0,
+ LLVMNoUnwindAttribute);
+}
+
static const struct lp_build_tgsi_action tex_action = {
.fetch_args = tex_fetch_args,
.emit = build_tex_intrinsic,
};
-static const struct lp_build_tgsi_action txq_action = {
- .fetch_args = txq_fetch_args,
- .emit = build_txq_intrinsic,
- .intr_name = "llvm.SI.resinfo"
+static const struct lp_build_tgsi_action interp_action = {
+ .fetch_args = interp_fetch_args,
+ .emit = build_interp_intrinsic,
};
static void create_meta_data(struct si_shader_context *si_shader_ctx)
CONST_ADDR_SPACE);
}
+static void declare_streamout_params(struct si_shader_context *si_shader_ctx,
+ struct pipe_stream_output_info *so,
+ LLVMTypeRef *params, LLVMTypeRef i32,
+ unsigned *num_params)
+{
+ int i;
+
+ /* Streamout SGPRs. */
+ if (so->num_outputs) {
+ params[si_shader_ctx->param_streamout_config = (*num_params)++] = i32;
+ params[si_shader_ctx->param_streamout_write_index = (*num_params)++] = i32;
+ }
+ /* A streamout buffer offset is loaded if the stride is non-zero. */
+ for (i = 0; i < 4; i++) {
+ if (!so->stride[i])
+ continue;
+
+ params[si_shader_ctx->param_streamout_offset[i] = (*num_params)++] = i32;
+ }
+}
+
static void create_function(struct si_shader_context *si_shader_ctx)
{
struct lp_build_tgsi_context *bld_base = &si_shader_ctx->radeon_bld.soa.bld_base;
v16i8 = LLVMVectorType(i8, 16);
params[SI_PARAM_RW_BUFFERS] = const_array(v16i8, SI_NUM_RW_BUFFERS);
- params[SI_PARAM_CONST] = const_array(v16i8, SI_NUM_CONST_BUFFERS);
- params[SI_PARAM_SAMPLER] = const_array(v4i32, SI_NUM_SAMPLER_STATES);
- params[SI_PARAM_RESOURCE] = const_array(v8i32, SI_NUM_SAMPLER_VIEWS);
- last_array_pointer = SI_PARAM_RESOURCE;
+ params[SI_PARAM_CONST_BUFFERS] = const_array(v16i8, SI_NUM_CONST_BUFFERS);
+ params[SI_PARAM_SAMPLER_STATES] = const_array(v4i32, SI_NUM_SAMPLER_STATES);
+ params[SI_PARAM_SAMPLER_VIEWS] = const_array(v8i32, SI_NUM_SAMPLER_VIEWS);
+ last_array_pointer = SI_PARAM_SAMPLER_VIEWS;
switch (si_shader_ctx->type) {
case TGSI_PROCESSOR_VERTEX:
- params[SI_PARAM_VERTEX_BUFFER] = const_array(v16i8, SI_NUM_VERTEX_BUFFERS);
- last_array_pointer = SI_PARAM_VERTEX_BUFFER;
+ params[SI_PARAM_VERTEX_BUFFERS] = const_array(v16i8, SI_NUM_VERTEX_BUFFERS);
+ last_array_pointer = SI_PARAM_VERTEX_BUFFERS;
params[SI_PARAM_BASE_VERTEX] = i32;
params[SI_PARAM_START_INSTANCE] = i32;
num_params = SI_PARAM_START_INSTANCE+1;
if (shader->key.vs.as_es) {
params[si_shader_ctx->param_es2gs_offset = num_params++] = i32;
+ } else if (shader->key.vs.as_ls) {
+ params[SI_PARAM_LS_OUT_LAYOUT] = i32;
+ num_params = SI_PARAM_LS_OUT_LAYOUT+1;
} else {
if (shader->is_gs_copy_shader) {
- last_array_pointer = SI_PARAM_CONST;
- num_params = SI_PARAM_CONST+1;
+ last_array_pointer = SI_PARAM_CONST_BUFFERS;
+ num_params = SI_PARAM_CONST_BUFFERS+1;
+ } else {
+ params[SI_PARAM_VS_STATE_BITS] = i32;
+ num_params = SI_PARAM_VS_STATE_BITS+1;
}
/* The locations of the other parameters are assigned dynamically. */
-
- /* Streamout SGPRs. */
- if (shader->selector->so.num_outputs) {
- params[si_shader_ctx->param_streamout_config = num_params++] = i32;
- params[si_shader_ctx->param_streamout_write_index = num_params++] = i32;
- }
- /* A streamout buffer offset is loaded if the stride is non-zero. */
- for (i = 0; i < 4; i++) {
- if (!shader->selector->so.stride[i])
- continue;
-
- params[si_shader_ctx->param_streamout_offset[i] = num_params++] = i32;
- }
+ declare_streamout_params(si_shader_ctx, &shader->selector->so,
+ params, i32, &num_params);
}
last_sgpr = num_params-1;
/* VGPRs */
params[si_shader_ctx->param_vertex_id = num_params++] = i32;
- params[num_params++] = i32; /* unused*/
- params[num_params++] = i32; /* unused */
+ params[si_shader_ctx->param_rel_auto_id = num_params++] = i32;
+ params[si_shader_ctx->param_vs_prim_id = num_params++] = i32;
params[si_shader_ctx->param_instance_id = num_params++] = i32;
break;
+ case TGSI_PROCESSOR_TESS_CTRL:
+ params[SI_PARAM_TCS_OUT_OFFSETS] = i32;
+ params[SI_PARAM_TCS_OUT_LAYOUT] = i32;
+ params[SI_PARAM_TCS_IN_LAYOUT] = i32;
+ params[SI_PARAM_TESS_FACTOR_OFFSET] = i32;
+ last_sgpr = SI_PARAM_TESS_FACTOR_OFFSET;
+
+ /* VGPRs */
+ params[SI_PARAM_PATCH_ID] = i32;
+ params[SI_PARAM_REL_IDS] = i32;
+ num_params = SI_PARAM_REL_IDS+1;
+ break;
+
+ case TGSI_PROCESSOR_TESS_EVAL:
+ params[SI_PARAM_TCS_OUT_OFFSETS] = i32;
+ params[SI_PARAM_TCS_OUT_LAYOUT] = i32;
+ num_params = SI_PARAM_TCS_OUT_LAYOUT+1;
+
+ if (shader->key.tes.as_es) {
+ params[si_shader_ctx->param_es2gs_offset = num_params++] = i32;
+ } else {
+ declare_streamout_params(si_shader_ctx, &shader->selector->so,
+ params, i32, &num_params);
+ }
+ last_sgpr = num_params - 1;
+
+ /* VGPRs */
+ params[si_shader_ctx->param_tes_u = num_params++] = f32;
+ params[si_shader_ctx->param_tes_v = num_params++] = f32;
+ params[si_shader_ctx->param_tes_rel_patch_id = num_params++] = i32;
+ params[si_shader_ctx->param_tes_patch_id = num_params++] = i32;
+ break;
+
case TGSI_PROCESSOR_GEOMETRY:
params[SI_PARAM_GS2VS_OFFSET] = i32;
params[SI_PARAM_GS_WAVE_ID] = i32;
case TGSI_PROCESSOR_FRAGMENT:
params[SI_PARAM_ALPHA_REF] = f32;
+ params[SI_PARAM_PS_STATE_BITS] = i32;
params[SI_PARAM_PRIM_MASK] = i32;
last_sgpr = SI_PARAM_PRIM_MASK;
params[SI_PARAM_PERSP_SAMPLE] = v2i32;
if (bld_base->info &&
(bld_base->info->opcode_count[TGSI_OPCODE_DDX] > 0 ||
- bld_base->info->opcode_count[TGSI_OPCODE_DDY] > 0))
- si_shader_ctx->ddxy_lds =
+ bld_base->info->opcode_count[TGSI_OPCODE_DDY] > 0 ||
+ bld_base->info->opcode_count[TGSI_OPCODE_DDX_FINE] > 0 ||
+ bld_base->info->opcode_count[TGSI_OPCODE_DDY_FINE] > 0 ||
+ bld_base->info->opcode_count[TGSI_OPCODE_INTERP_OFFSET] > 0 ||
+ bld_base->info->opcode_count[TGSI_OPCODE_INTERP_SAMPLE] > 0))
+ si_shader_ctx->lds =
LLVMAddGlobalInAddressSpace(gallivm->module,
LLVMArrayType(i32, 64),
"ddxy_lds",
LOCAL_ADDR_SPACE);
+
+ if ((si_shader_ctx->type == TGSI_PROCESSOR_VERTEX && shader->key.vs.as_ls) ||
+ si_shader_ctx->type == TGSI_PROCESSOR_TESS_CTRL ||
+ si_shader_ctx->type == TGSI_PROCESSOR_TESS_EVAL) {
+ /* This is the upper bound, maximum is 32 inputs times 32 vertices */
+ unsigned vertex_data_dw_size = 32*32*4;
+ unsigned patch_data_dw_size = 32*4;
+ /* The formula is: TCS inputs + TCS outputs + TCS patch outputs. */
+ unsigned patch_dw_size = vertex_data_dw_size*2 + patch_data_dw_size;
+ unsigned lds_dwords = patch_dw_size;
+
+ /* The actual size is computed outside of the shader to reduce
+ * the number of shader variants. */
+ si_shader_ctx->lds =
+ LLVMAddGlobalInAddressSpace(gallivm->module,
+ LLVMArrayType(i32, lds_dwords),
+ "tess_lds",
+ LOCAL_ADDR_SPACE);
+ }
}
static void preload_constants(struct si_shader_context *si_shader_ctx)
struct gallivm_state * gallivm = bld_base->base.gallivm;
const struct tgsi_shader_info * info = bld_base->info;
unsigned buf;
- LLVMValueRef ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST);
+ LLVMValueRef ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_CONST_BUFFERS);
for (buf = 0; buf < SI_NUM_CONST_BUFFERS; buf++) {
unsigned i, num_const = info->const_file_max[buf] + 1;
si_shader_ctx->constants[buf] = CALLOC(num_const * 4, sizeof(LLVMValueRef));
/* Load the resource descriptor */
- si_shader_ctx->const_resource[buf] =
+ si_shader_ctx->const_buffers[buf] =
build_indexed_load_const(si_shader_ctx, ptr, lp_build_const_int32(gallivm, buf));
/* Load the constants, we rely on the code sinking to do the rest */
for (i = 0; i < num_const * 4; ++i) {
si_shader_ctx->constants[buf][i] =
buffer_load_const(gallivm->builder,
- si_shader_ctx->const_resource[buf],
+ si_shader_ctx->const_buffers[buf],
lp_build_const_int32(gallivm, i * 4),
bld_base->base.elem_type);
}
if (num_samplers == 0)
return;
- res_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_RESOURCE);
- samp_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_SAMPLER);
+ res_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_SAMPLER_VIEWS);
+ samp_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn, SI_PARAM_SAMPLER_STATES);
/* Load the resources and samplers, we rely on the code sinking to do the rest */
for (i = 0; i < num_samplers; ++i) {
/* Resource */
offset = lp_build_const_int32(gallivm, i);
- si_shader_ctx->resources[i] = build_indexed_load_const(si_shader_ctx, res_ptr, offset);
+ si_shader_ctx->sampler_views[i] = build_indexed_load_const(si_shader_ctx, res_ptr, offset);
/* Sampler */
offset = lp_build_const_int32(gallivm, i);
- si_shader_ctx->samplers[i] = build_indexed_load_const(si_shader_ctx, samp_ptr, offset);
+ si_shader_ctx->sampler_states[i] = build_indexed_load_const(si_shader_ctx, samp_ptr, offset);
/* FMASK resource */
if (info->is_msaa_sampler[i]) {
offset = lp_build_const_int32(gallivm, SI_FMASK_TEX_OFFSET + i);
- si_shader_ctx->resources[SI_FMASK_TEX_OFFSET + i] =
+ si_shader_ctx->sampler_views[SI_FMASK_TEX_OFFSET + i] =
build_indexed_load_const(si_shader_ctx, res_ptr, offset);
}
}
struct gallivm_state * gallivm = bld_base->base.gallivm;
unsigned i;
- if (si_shader_ctx->type != TGSI_PROCESSOR_VERTEX ||
- si_shader_ctx->shader->key.vs.as_es ||
- !si_shader_ctx->shader->selector->so.num_outputs)
+ /* Streamout can only be used if the shader is compiled as VS. */
+ if (!si_shader_ctx->shader->selector->so.num_outputs ||
+ (si_shader_ctx->type == TGSI_PROCESSOR_VERTEX &&
+ (si_shader_ctx->shader->key.vs.as_es ||
+ si_shader_ctx->shader->key.vs.as_ls)) ||
+ (si_shader_ctx->type == TGSI_PROCESSOR_TESS_EVAL &&
+ si_shader_ctx->shader->key.tes.as_es))
return;
LLVMValueRef buf_ptr = LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
if ((si_shader_ctx->type == TGSI_PROCESSOR_VERTEX &&
si_shader_ctx->shader->key.vs.as_es) ||
+ (si_shader_ctx->type == TGSI_PROCESSOR_TESS_EVAL &&
+ si_shader_ctx->shader->key.tes.as_es) ||
si_shader_ctx->type == TGSI_PROCESSOR_GEOMETRY) {
LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_ESGS);
build_indexed_load_const(si_shader_ctx, buf_ptr, offset);
}
- if (si_shader_ctx->type == TGSI_PROCESSOR_GEOMETRY ||
- si_shader_ctx->shader->is_gs_copy_shader) {
+ if (si_shader_ctx->shader->is_gs_copy_shader) {
LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS);
- si_shader_ctx->gsvs_ring =
+ si_shader_ctx->gsvs_ring[0] =
build_indexed_load_const(si_shader_ctx, buf_ptr, offset);
}
+ if (si_shader_ctx->type == TGSI_PROCESSOR_GEOMETRY) {
+ int i;
+ for (i = 0; i < 4; i++) {
+ LLVMValueRef offset = lp_build_const_int32(gallivm, SI_RING_GSVS + i);
+
+ si_shader_ctx->gsvs_ring[i] =
+ build_indexed_load_const(si_shader_ctx, buf_ptr, offset);
+ }
+ }
}
void si_shader_binary_read_config(const struct si_screen *sscreen,
shader->num_sgprs = MAX2(shader->num_sgprs, (G_00B028_SGPRS(value) + 1) * 8);
shader->num_vgprs = MAX2(shader->num_vgprs, (G_00B028_VGPRS(value) + 1) * 4);
shader->float_mode = G_00B028_FLOAT_MODE(value);
+ shader->rsrc1 = value;
break;
case R_00B02C_SPI_SHADER_PGM_RSRC2_PS:
shader->lds_size = MAX2(shader->lds_size, G_00B02C_EXTRA_LDS_SIZE(value));
break;
case R_00B84C_COMPUTE_PGM_RSRC2:
shader->lds_size = MAX2(shader->lds_size, G_00B84C_LDS_SIZE(value));
+ shader->rsrc2 = value;
break;
case R_0286CC_SPI_PS_INPUT_ENA:
shader->spi_ps_input_ena = value;
uint64_t scratch_va)
{
unsigned i;
- uint32_t scratch_rsrc_dword0 = scratch_va & 0xffffffff;
+ uint32_t scratch_rsrc_dword0 = scratch_va;
uint32_t scratch_rsrc_dword1 =
S_008F04_BASE_ADDRESS_HI(scratch_va >> 32)
| S_008F04_STRIDE(shader->scratch_bytes_per_wave / 64);
if (!shader->bo)
return -ENOMEM;
- ptr = sscreen->b.ws->buffer_map(shader->bo->cs_buf, NULL,
+ ptr = sscreen->b.ws->buffer_map(shader->bo->buf, NULL,
PIPE_TRANSFER_READ_WRITE);
util_memcpy_cpu_to_le32(ptr, binary->code, binary->code_size);
if (binary->rodata_size > 0) {
binary->rodata_size);
}
- sscreen->b.ws->buffer_unmap(shader->bo->cs_buf);
+ sscreen->b.ws->buffer_unmap(shader->bo->buf);
return 0;
}
-int si_shader_binary_read(struct si_screen *sscreen, struct si_shader *shader)
+static void si_shader_dump_disassembly(const struct radeon_shader_binary *binary,
+ struct pipe_debug_callback *debug)
+{
+ char *line, *p;
+ unsigned i, count;
+
+ if (binary->disasm_string) {
+ fprintf(stderr, "\nShader Disassembly:\n\n");
+ fprintf(stderr, "%s\n", binary->disasm_string);
+
+ if (debug && debug->debug_message) {
+ /* Very long debug messages are cut off, so send the
+ * disassembly one line at a time. This causes more
+ * overhead, but on the plus side it simplifies
+ * parsing of resulting logs.
+ */
+ pipe_debug_message(debug, SHADER_INFO,
+ "Shader Disassembly Begin");
+
+ line = binary->disasm_string;
+ while (*line) {
+ p = strchrnul(line, '\n');
+ count = p - line;
+
+ if (count) {
+ pipe_debug_message(debug, SHADER_INFO,
+ "%.*s", count, line);
+ }
+
+ if (!*p)
+ break;
+ line = p + 1;
+ }
+
+ pipe_debug_message(debug, SHADER_INFO,
+ "Shader Disassembly End");
+ }
+ } else {
+ fprintf(stderr, "SI CODE:\n");
+ for (i = 0; i < binary->code_size; i += 4) {
+ fprintf(stderr, "@0x%x: %02x%02x%02x%02x\n", i,
+ binary->code[i + 3], binary->code[i + 2],
+ binary->code[i + 1], binary->code[i]);
+ }
+ }
+}
+
+int si_shader_binary_read(struct si_screen *sscreen, struct si_shader *shader,
+ struct pipe_debug_callback *debug)
{
const struct radeon_shader_binary *binary = &shader->binary;
- unsigned i;
+ int r;
bool dump = r600_can_dump_shader(&sscreen->b,
shader->selector ? shader->selector->tokens : NULL);
si_shader_binary_read_config(sscreen, shader, 0);
- si_shader_binary_upload(sscreen, shader);
+ r = si_shader_binary_upload(sscreen, shader);
+ if (r)
+ return r;
if (dump) {
- if (!binary->disassembled) {
- fprintf(stderr, "SI CODE:\n");
- for (i = 0; i < binary->code_size; i+=4 ) {
- fprintf(stderr, "@0x%x: %02x%02x%02x%02x\n", i, binary->code[i + 3],
- binary->code[i + 2], binary->code[i + 1],
- binary->code[i]);
- }
- }
+ if (!(sscreen->b.debug_flags & DBG_NO_ASM))
+ si_shader_dump_disassembly(binary, debug);
fprintf(stderr, "*** SHADER STATS ***\n"
"SGPRS: %d\nVGPRS: %d\nCode Size: %d bytes\nLDS: %d blocks\n"
shader->num_sgprs, shader->num_vgprs, binary->code_size,
shader->lds_size, shader->scratch_bytes_per_wave);
}
+
+ pipe_debug_message(debug, SHADER_INFO,
+ "Shader Stats: SGPRS: %d VGPRS: %d Code Size: %d LDS: %d Scratch: %d",
+ shader->num_sgprs, shader->num_vgprs, binary->code_size,
+ shader->lds_size, shader->scratch_bytes_per_wave);
+
return 0;
}
int si_compile_llvm(struct si_screen *sscreen, struct si_shader *shader,
- LLVMTargetMachineRef tm, LLVMModuleRef mod)
+ LLVMTargetMachineRef tm, LLVMModuleRef mod,
+ struct pipe_debug_callback *debug, unsigned processor)
{
int r = 0;
- bool dump = r600_can_dump_shader(&sscreen->b,
- shader->selector ? shader->selector->tokens : NULL);
- r = radeon_llvm_compile(mod, &shader->binary,
- r600_get_llvm_processor_name(sscreen->b.family), dump, tm);
-
- if (r) {
- return r;
+ bool dump_asm = r600_can_dump_shader(&sscreen->b,
+ shader->selector ? shader->selector->tokens : NULL);
+ bool dump_ir = dump_asm && !(sscreen->b.debug_flags & DBG_NO_IR);
+ unsigned count = p_atomic_inc_return(&sscreen->b.num_compilations);
+
+ if (dump_ir || dump_asm)
+ fprintf(stderr, "radeonsi: Compiling shader %d\n", count);
+
+ if (!si_replace_shader(count, &shader->binary)) {
+ r = radeon_llvm_compile(mod, &shader->binary,
+ r600_get_llvm_processor_name(sscreen->b.family), dump_ir, dump_asm, tm,
+ debug);
+ if (r)
+ return r;
}
- r = si_shader_binary_read(sscreen, shader);
+
+ r = si_shader_binary_read(sscreen, shader, debug);
FREE(shader->binary.config);
FREE(shader->binary.rodata);
if (shader->scratch_bytes_per_wave == 0) {
FREE(shader->binary.code);
FREE(shader->binary.relocs);
- memset(&shader->binary, 0, sizeof(shader->binary));
+ memset(&shader->binary, 0,
+ offsetof(struct radeon_shader_binary, disasm_string));
}
return r;
}
/* Generate code for the hardware VS shader stage to go with a geometry shader */
static int si_generate_gs_copy_shader(struct si_screen *sscreen,
struct si_shader_context *si_shader_ctx,
- struct si_shader *gs, bool dump)
+ struct si_shader *gs, bool dump,
+ struct pipe_debug_callback *debug)
{
struct gallivm_state *gallivm = &si_shader_ctx->radeon_bld.gallivm;
struct lp_build_tgsi_context *bld_base = &si_shader_ctx->radeon_bld.soa.bld_base;
preload_streamout_buffers(si_shader_ctx);
preload_ring_buffers(si_shader_ctx);
- args[0] = si_shader_ctx->gsvs_ring;
+ args[0] = si_shader_ctx->gsvs_ring[0];
args[1] = lp_build_mul_imm(uint,
LLVMGetParam(si_shader_ctx->radeon_bld.main_fn,
si_shader_ctx->param_vertex_id),
outputs[i].values[chan] =
LLVMBuildBitCast(gallivm->builder,
- build_intrinsic(gallivm->builder,
+ lp_build_intrinsic(gallivm->builder,
"llvm.SI.buffer.load.dword.i32.i32",
LLVMInt32TypeInContext(gallivm->context),
args, 9,
fprintf(stderr, "Copy Vertex Shader for Geometry Shader:\n\n");
r = si_compile_llvm(sscreen, si_shader_ctx->shader,
- si_shader_ctx->tm, bld_base->base.gallivm->module);
+ si_shader_ctx->tm, bld_base->base.gallivm->module,
+ debug, TGSI_PROCESSOR_GEOMETRY);
radeon_llvm_dispose(&si_shader_ctx->radeon_bld);
return r;
}
-static void si_dump_key(unsigned shader, union si_shader_key *key)
+void si_dump_shader_key(unsigned shader, union si_shader_key *key, FILE *f)
{
int i;
- fprintf(stderr, "SHADER KEY\n");
+ fprintf(f, "SHADER KEY\n");
switch (shader) {
case PIPE_SHADER_VERTEX:
- fprintf(stderr, " instance_divisors = {");
+ fprintf(f, " instance_divisors = {");
for (i = 0; i < Elements(key->vs.instance_divisors); i++)
- fprintf(stderr, !i ? "%u" : ", %u",
+ fprintf(f, !i ? "%u" : ", %u",
key->vs.instance_divisors[i]);
- fprintf(stderr, "}\n");
+ fprintf(f, "}\n");
+ fprintf(f, " as_es = %u\n", key->vs.as_es);
+ fprintf(f, " as_ls = %u\n", key->vs.as_ls);
+ fprintf(f, " export_prim_id = %u\n", key->vs.export_prim_id);
+ break;
- if (key->vs.as_es)
- fprintf(stderr, " gs_used_inputs = 0x%"PRIx64"\n",
- key->vs.gs_used_inputs);
- fprintf(stderr, " as_es = %u\n", key->vs.as_es);
+ case PIPE_SHADER_TESS_CTRL:
+ fprintf(f, " prim_mode = %u\n", key->tcs.prim_mode);
+ break;
+
+ case PIPE_SHADER_TESS_EVAL:
+ fprintf(f, " as_es = %u\n", key->tes.as_es);
+ fprintf(f, " export_prim_id = %u\n", key->tes.export_prim_id);
break;
case PIPE_SHADER_GEOMETRY:
break;
case PIPE_SHADER_FRAGMENT:
- fprintf(stderr, " export_16bpc = 0x%X\n", key->ps.export_16bpc);
- fprintf(stderr, " last_cbuf = %u\n", key->ps.last_cbuf);
- fprintf(stderr, " color_two_side = %u\n", key->ps.color_two_side);
- fprintf(stderr, " alpha_func = %u\n", key->ps.alpha_func);
- fprintf(stderr, " alpha_to_one = %u\n", key->ps.alpha_to_one);
- fprintf(stderr, " poly_stipple = %u\n", key->ps.poly_stipple);
+ fprintf(f, " export_16bpc = 0x%X\n", key->ps.export_16bpc);
+ fprintf(f, " last_cbuf = %u\n", key->ps.last_cbuf);
+ fprintf(f, " color_two_side = %u\n", key->ps.color_two_side);
+ fprintf(f, " alpha_func = %u\n", key->ps.alpha_func);
+ fprintf(f, " alpha_to_one = %u\n", key->ps.alpha_to_one);
+ fprintf(f, " poly_stipple = %u\n", key->ps.poly_stipple);
+ fprintf(f, " clamp_color = %u\n", key->ps.clamp_color);
break;
default:
}
int si_shader_create(struct si_screen *sscreen, LLVMTargetMachineRef tm,
- struct si_shader *shader)
+ struct si_shader *shader,
+ struct pipe_debug_callback *debug)
{
struct si_shader_selector *sel = shader->selector;
struct tgsi_token *tokens = sel->tokens;
/* Dump TGSI code before doing TGSI->LLVM conversion in case the
* conversion fails. */
- if (dump) {
- si_dump_key(sel->type, &shader->key);
+ if (dump && !(sscreen->b.debug_flags & DBG_NO_TGSI)) {
+ si_dump_shader_key(sel->type, &shader->key, stderr);
tgsi_dump(tokens, 0);
si_dump_streamout(&sel->so);
}
bld_base->info = poly_stipple ? &stipple_shader_info : &sel->info;
bld_base->emit_fetch_funcs[TGSI_FILE_CONSTANT] = fetch_constant;
+ bld_base->op_actions[TGSI_OPCODE_INTERP_CENTROID] = interp_action;
+ bld_base->op_actions[TGSI_OPCODE_INTERP_SAMPLE] = interp_action;
+ bld_base->op_actions[TGSI_OPCODE_INTERP_OFFSET] = interp_action;
+
bld_base->op_actions[TGSI_OPCODE_TEX] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TEX2] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXB] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXL] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXL2] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TXP] = tex_action;
- bld_base->op_actions[TGSI_OPCODE_TXQ] = txq_action;
+ bld_base->op_actions[TGSI_OPCODE_TXQ] = tex_action;
bld_base->op_actions[TGSI_OPCODE_TG4] = tex_action;
bld_base->op_actions[TGSI_OPCODE_LODQ] = tex_action;
+ bld_base->op_actions[TGSI_OPCODE_TXQS].emit = si_llvm_emit_txqs;
bld_base->op_actions[TGSI_OPCODE_DDX].emit = si_llvm_emit_ddxy;
bld_base->op_actions[TGSI_OPCODE_DDY].emit = si_llvm_emit_ddxy;
+ bld_base->op_actions[TGSI_OPCODE_DDX_FINE].emit = si_llvm_emit_ddxy;
+ bld_base->op_actions[TGSI_OPCODE_DDY_FINE].emit = si_llvm_emit_ddxy;
bld_base->op_actions[TGSI_OPCODE_EMIT].emit = si_llvm_emit_vertex;
bld_base->op_actions[TGSI_OPCODE_ENDPRIM].emit = si_llvm_emit_primitive;
+ bld_base->op_actions[TGSI_OPCODE_BARRIER].emit = si_llvm_emit_barrier;
if (HAVE_LLVM >= 0x0306) {
bld_base->op_actions[TGSI_OPCODE_MAX].emit = build_tgsi_intrinsic_nomem;
switch (si_shader_ctx.type) {
case TGSI_PROCESSOR_VERTEX:
si_shader_ctx.radeon_bld.load_input = declare_input_vs;
- if (shader->key.vs.as_es) {
+ if (shader->key.vs.as_ls)
+ bld_base->emit_epilogue = si_llvm_emit_ls_epilogue;
+ else if (shader->key.vs.as_es)
bld_base->emit_epilogue = si_llvm_emit_es_epilogue;
- } else {
+ else
+ bld_base->emit_epilogue = si_llvm_emit_vs_epilogue;
+ break;
+ case TGSI_PROCESSOR_TESS_CTRL:
+ bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tcs;
+ bld_base->emit_fetch_funcs[TGSI_FILE_OUTPUT] = fetch_output_tcs;
+ bld_base->emit_store = store_output_tcs;
+ bld_base->emit_epilogue = si_llvm_emit_tcs_epilogue;
+ break;
+ case TGSI_PROCESSOR_TESS_EVAL:
+ bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_tes;
+ if (shader->key.tes.as_es)
+ bld_base->emit_epilogue = si_llvm_emit_es_epilogue;
+ else
bld_base->emit_epilogue = si_llvm_emit_vs_epilogue;
- }
break;
case TGSI_PROCESSOR_GEOMETRY:
bld_base->emit_fetch_funcs[TGSI_FILE_INPUT] = fetch_input_gs;
preload_ring_buffers(&si_shader_ctx);
if (si_shader_ctx.type == TGSI_PROCESSOR_GEOMETRY) {
- si_shader_ctx.gs_next_vertex =
- lp_build_alloca(bld_base->base.gallivm,
- bld_base->uint_bld.elem_type, "");
+ int i;
+ for (i = 0; i < 4; i++) {
+ si_shader_ctx.gs_next_vertex[i] =
+ lp_build_alloca(bld_base->base.gallivm,
+ bld_base->uint_bld.elem_type, "");
+ }
}
if (!lp_build_tgsi_llvm(bld_base, tokens)) {
radeon_llvm_finalize_module(&si_shader_ctx.radeon_bld);
mod = bld_base->base.gallivm->module;
- r = si_compile_llvm(sscreen, shader, tm, mod);
+ r = si_compile_llvm(sscreen, shader, tm, mod, debug, si_shader_ctx.type);
if (r) {
fprintf(stderr, "LLVM failed to compile shader\n");
goto out;
shader->gs_copy_shader->key = shader->key;
si_shader_ctx.shader = shader->gs_copy_shader;
if ((r = si_generate_gs_copy_shader(sscreen, &si_shader_ctx,
- shader, dump))) {
+ shader, dump, debug))) {
free(shader->gs_copy_shader);
shader->gs_copy_shader = NULL;
goto out;
return r;
}
-void si_shader_destroy(struct pipe_context *ctx, struct si_shader *shader)
+void si_shader_destroy(struct si_shader *shader)
{
- if (shader->gs_copy_shader)
- si_shader_destroy(ctx, shader->gs_copy_shader);
+ if (shader->gs_copy_shader) {
+ si_shader_destroy(shader->gs_copy_shader);
+ FREE(shader->gs_copy_shader);
+ }
if (shader->scratch_bo)
r600_resource_reference(&shader->scratch_bo, NULL);
FREE(shader->binary.code);
FREE(shader->binary.relocs);
+ FREE(shader->binary.disasm_string);
}
projects / mesa.git / blobdiff