#define AC_LLVM_BUILD_H
#include <stdbool.h>
-#include <llvm-c/TargetMachine.h>
+#include <llvm-c/Core.h>
#include "compiler/nir/nir.h"
#include "amd_family.h"
extern "C" {
#endif
-#define HAVE_32BIT_POINTERS (HAVE_LLVM >= 0x0700)
-
enum {
- /* CONST is the only address space that selects SMEM loads */
- AC_CONST_ADDR_SPACE = HAVE_LLVM >= 0x700 ? 4 : 2,
- AC_LOCAL_ADDR_SPACE = 3,
- AC_CONST_32BIT_ADDR_SPACE = 6, /* same as CONST, but the pointer type has 32 bits */
+ AC_ADDR_SPACE_FLAT = 0, /* Slower than global. */
+ AC_ADDR_SPACE_GLOBAL = 1,
+ AC_ADDR_SPACE_GDS = 2,
+ AC_ADDR_SPACE_LDS = 3,
+ AC_ADDR_SPACE_CONST = 4, /* Global allowing SMEM. */
+ AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
};
/* Combine these with & instead of |. */
LLVMTypeRef v3i32;
LLVMTypeRef v4i32;
LLVMTypeRef v2f32;
+ LLVMTypeRef v3f32;
LLVMTypeRef v4f32;
LLVMTypeRef v8i32;
+ LLVMValueRef i8_0;
+ LLVMValueRef i8_1;
+ LLVMValueRef i16_0;
+ LLVMValueRef i16_1;
LLVMValueRef i32_0;
LLVMValueRef i32_1;
LLVMValueRef i64_0;
LLVMValueRef i64_1;
+ LLVMValueRef f16_0;
+ LLVMValueRef f16_1;
LLVMValueRef f32_0;
LLVMValueRef f32_1;
LLVMValueRef f64_0;
LLVMTypeRef ac_to_integer_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
LLVMValueRef ac_to_integer(struct ac_llvm_context *ctx, LLVMValueRef v);
+LLVMValueRef ac_to_integer_or_pointer(struct ac_llvm_context *ctx, LLVMValueRef v);
LLVMTypeRef ac_to_float_type(struct ac_llvm_context *ctx, LLVMTypeRef t);
LLVMValueRef ac_to_float(struct ac_llvm_context *ctx, LLVMValueRef v);
LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
+LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
+ LLVMValueRef value);
LLVMValueRef ac_build_vote_all(struct ac_llvm_context *ctx, LLVMValueRef value);
LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
LLVMValueRef value,
unsigned num_channels);
+LLVMValueRef ac_build_round(struct ac_llvm_context *ctx, LLVMValueRef value);
LLVMValueRef
ac_build_fdiv(struct ac_llvm_context *ctx,
LLVMValueRef num,
LLVMValueRef den);
+LLVMValueRef ac_build_fast_udiv(struct ac_llvm_context *ctx,
+ LLVMValueRef num,
+ LLVMValueRef multiplier,
+ LLVMValueRef pre_shift,
+ LLVMValueRef post_shift,
+ LLVMValueRef increment);
+LLVMValueRef ac_build_fast_udiv_nuw(struct ac_llvm_context *ctx,
+ LLVMValueRef num,
+ LLVMValueRef multiplier,
+ LLVMValueRef pre_shift,
+ LLVMValueRef post_shift,
+ LLVMValueRef increment);
+LLVMValueRef ac_build_fast_udiv_u31_d_not_one(struct ac_llvm_context *ctx,
+ LLVMValueRef num,
+ LLVMValueRef multiplier,
+ LLVMValueRef post_shift);
+
void
ac_prepare_cube_coords(struct ac_llvm_context *ctx,
bool is_deriv, bool is_array, bool is_lod,
LLVMValueRef i,
LLVMValueRef j);
+LLVMValueRef
+ac_build_fs_interp_f16(struct ac_llvm_context *ctx,
+ LLVMValueRef llvm_chan,
+ LLVMValueRef attr_number,
+ LLVMValueRef params,
+ LLVMValueRef i,
+ LLVMValueRef j);
+
LLVMValueRef
ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
LLVMValueRef parameter,
LLVMValueRef attr_number,
LLVMValueRef params);
+LLVMValueRef
+ac_build_gep_ptr(struct ac_llvm_context *ctx,
+ LLVMValueRef base_ptr,
+ LLVMValueRef index);
+
LLVMValueRef
ac_build_gep0(struct ac_llvm_context *ctx,
LLVMValueRef base_ptr,
bool slc,
bool writeonly_memory,
bool swizzle_enable_hint);
+
+void
+ac_build_buffer_store_format(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef data,
+ LLVMValueRef vindex,
+ LLVMValueRef voffset,
+ unsigned num_channels,
+ bool glc,
+ bool slc,
+ bool writeonly_memory);
+
LLVMValueRef
ac_build_buffer_load(struct ac_llvm_context *ctx,
LLVMValueRef rsrc,
LLVMValueRef
ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
LLVMValueRef rsrc,
- LLVMValueRef vindex,
LLVMValueRef voffset,
- LLVMValueRef soffset,
- LLVMValueRef immoffset);
+ LLVMValueRef soffset,
+ LLVMValueRef immoffset,
+ bool glc);
+
+LLVMValueRef
+ac_build_tbuffer_load_byte(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ LLVMValueRef immoffset,
+ bool glc);
+
+LLVMValueRef
+ac_build_struct_tbuffer_load(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef vindex,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ LLVMValueRef immoffset,
+ unsigned num_channels,
+ unsigned dfmt,
+ unsigned nfmt,
+ bool glc,
+ bool slc,
+ bool can_speculate);
+
+LLVMValueRef
+ac_build_raw_tbuffer_load(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ LLVMValueRef immoffset,
+ unsigned num_channels,
+ unsigned dfmt,
+ unsigned nfmt,
+ bool glc,
+ bool slc,
+ bool can_speculate);
+
+/* For ac_build_fetch_format.
+ *
+ * Note: FLOAT must be 0 (used for convenience of encoding in radeonsi).
+ */
+enum {
+ AC_FETCH_FORMAT_FLOAT = 0,
+ AC_FETCH_FORMAT_FIXED,
+ AC_FETCH_FORMAT_UNORM,
+ AC_FETCH_FORMAT_SNORM,
+ AC_FETCH_FORMAT_USCALED,
+ AC_FETCH_FORMAT_SSCALED,
+ AC_FETCH_FORMAT_UINT,
+ AC_FETCH_FORMAT_SINT,
+};
+
+LLVMValueRef
+ac_build_opencoded_load_format(struct ac_llvm_context *ctx,
+ unsigned log_size,
+ unsigned num_channels,
+ unsigned format,
+ bool reverse,
+ bool known_aligned,
+ LLVMValueRef rsrc,
+ LLVMValueRef vindex,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ bool glc,
+ bool slc,
+ bool can_speculate);
+
+void
+ac_build_tbuffer_store_short(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef vdata,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ bool glc,
+ bool writeonly_memory);
+
+void
+ac_build_tbuffer_store_byte(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef vdata,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ bool glc,
+ bool writeonly_memory);
+
+void
+ac_build_struct_tbuffer_store(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef vdata,
+ LLVMValueRef vindex,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ LLVMValueRef immoffset,
+ unsigned num_channels,
+ unsigned dfmt,
+ unsigned nfmt,
+ bool glc,
+ bool slc,
+ bool writeonly_memory);
+
+void
+ac_build_raw_tbuffer_store(struct ac_llvm_context *ctx,
+ LLVMValueRef rsrc,
+ LLVMValueRef vdata,
+ LLVMValueRef voffset,
+ LLVMValueRef soffset,
+ LLVMValueRef immoffset,
+ unsigned num_channels,
+ unsigned dfmt,
+ unsigned nfmt,
+ bool glc,
+ bool slc,
+ bool writeonly_memory);
LLVMValueRef
ac_get_thread_id(struct ac_llvm_context *ctx);
LLVMValueRef ac_build_imax(struct ac_llvm_context *ctx, LLVMValueRef a,
LLVMValueRef b);
LLVMValueRef ac_build_umin(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
+LLVMValueRef ac_build_umax(struct ac_llvm_context *ctx, LLVMValueRef a, LLVMValueRef b);
LLVMValueRef ac_build_clamp(struct ac_llvm_context *ctx, LLVMValueRef value);
struct ac_export_args {
LLVMValueRef ac_build_fract(struct ac_llvm_context *ctx, LLVMValueRef src0,
unsigned bitsize);
+LLVMValueRef ac_build_fmed3(struct ac_llvm_context *ctx, LLVMValueRef src0,
+ LLVMValueRef src1, LLVMValueRef src2,
+ unsigned bitsize);
+
LLVMValueRef ac_build_isign(struct ac_llvm_context *ctx, LLVMValueRef src0,
unsigned bitsize);
LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src0,
unsigned bitsize);
+LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
+
+LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
+ LLVMValueRef src0);
+
void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
LLVMValueRef main_fn,
uint8_t *vs_output_param_offset,
void ac_build_else(struct ac_llvm_context *ctx, int lable_id);
void ac_build_endif(struct ac_llvm_context *ctx, int lable_id);
void ac_build_endloop(struct ac_llvm_context *ctx, int lable_id);
+void ac_build_ifcc(struct ac_llvm_context *ctx, LLVMValueRef cond, int label_id);
void ac_build_if(struct ac_llvm_context *ctx, LLVMValueRef value,
int lable_id);
void ac_build_uif(struct ac_llvm_context *ctx, LLVMValueRef value,
LLVMValueRef
ac_build_reduce(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op, unsigned cluster_size);
+/**
+ * Common arguments for a scan/reduce operation that accumulates per-wave
+ * values across an entire workgroup, while respecting the order of waves.
+ */
+struct ac_wg_scan {
+ bool enable_reduce;
+ bool enable_exclusive;
+ bool enable_inclusive;
+ nir_op op;
+ LLVMValueRef src; /* clobbered! */
+ LLVMValueRef result_reduce;
+ LLVMValueRef result_exclusive;
+ LLVMValueRef result_inclusive;
+ LLVMValueRef extra;
+ LLVMValueRef waveidx;
+ LLVMValueRef numwaves; /* only needed for "reduce" operations */
+
+ /* T addrspace(LDS) pointer to the same type as value, at least maxwaves entries */
+ LLVMValueRef scratch;
+ unsigned maxwaves;
+};
+
+void
+ac_build_wg_wavescan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
+void
+ac_build_wg_wavescan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
+void
+ac_build_wg_wavescan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
+
+void
+ac_build_wg_scan_top(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
+void
+ac_build_wg_scan_bottom(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
+void
+ac_build_wg_scan(struct ac_llvm_context *ctx, struct ac_wg_scan *ws);
+
LLVMValueRef
ac_build_quad_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src,
unsigned lane0, unsigned lane1, unsigned lane2, unsigned lane3);
LLVMValueRef
ac_build_shuffle(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef index);
+LLVMValueRef
+ac_build_frexp_exp(struct ac_llvm_context *ctx, LLVMValueRef src0,
+ unsigned bitsize);
+
+LLVMValueRef
+ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
+ unsigned bitsize);
+
+LLVMValueRef
+ac_build_ddxy_interp(struct ac_llvm_context *ctx, LLVMValueRef interp_ij);
+
+LLVMValueRef
+ac_build_load_helper_invocation(struct ac_llvm_context *ctx);
+
+LLVMValueRef ac_build_atomic_rmw(struct ac_llvm_context *ctx, LLVMAtomicRMWBinOp op,
+ LLVMValueRef ptr, LLVMValueRef val,
+ const char *sync_scope);
+
+LLVMValueRef ac_build_atomic_cmp_xchg(struct ac_llvm_context *ctx, LLVMValueRef ptr,
+ LLVMValueRef cmp, LLVMValueRef val,
+ const char *sync_scope);
+
#ifdef __cplusplus
}
#endif