ac/llvm: add better code for fsign

[mesa.git] / src / amd / llvm / ac_llvm_build.h

diff --git a/src/amd/llvm/ac_llvm_build.h b/src/amd/llvm/ac_llvm_build.h
index 013bf00041ac5876db24c1624b09f57abbfd1010..6491843cd78241cbe1fffaa7e0dabb0fb0c3d0b9 100644 (file)
--- a/src/amd/llvm/ac_llvm_build.h
+++ b/src/amd/llvm/ac_llvm_build.h
@@ -30,6 +30,8 @@
 #include "compiler/nir/nir.h"
 #include "amd_family.h"
 #include "ac_shader_util.h"
+#include "ac_shader_args.h"
+#include "ac_shader_abi.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -63,17 +65,23 @@ struct ac_llvm_context {
        LLVMModuleRef module;
        LLVMBuilderRef builder;
 
+       LLVMValueRef main_function;
+
        LLVMTypeRef voidt;
        LLVMTypeRef i1;
        LLVMTypeRef i8;
        LLVMTypeRef i16;
        LLVMTypeRef i32;
        LLVMTypeRef i64;
+       LLVMTypeRef i128;
        LLVMTypeRef intptr;
        LLVMTypeRef f16;
        LLVMTypeRef f32;
        LLVMTypeRef f64;
        LLVMTypeRef v2i16;
+       LLVMTypeRef v4i16;
+       LLVMTypeRef v2f16;
+       LLVMTypeRef v4f16;
        LLVMTypeRef v2i32;
        LLVMTypeRef v3i32;
        LLVMTypeRef v4i32;
@@ -92,6 +100,8 @@ struct ac_llvm_context {
        LLVMValueRef i32_1;
        LLVMValueRef i64_0;
        LLVMValueRef i64_1;
+       LLVMValueRef i128_0;
+       LLVMValueRef i128_1;
        LLVMValueRef f16_0;
        LLVMValueRef f16_1;
        LLVMValueRef f32_0;
@@ -101,6 +111,12 @@ struct ac_llvm_context {
        LLVMValueRef i1true;
        LLVMValueRef i1false;
 
+       /* Temporary helper to implement demote_to_helper:
+         * True = live lanes
+         * False = demoted lanes
+         */
+       LLVMValueRef postponed_kill;
+
        /* Since ac_nir_translate makes a local copy of ac_llvm_context, there
         * are two ac_llvm_contexts. Declare a pointer here, so that the control
         * flow stack is shared by both ac_llvm_contexts.
@@ -110,8 +126,6 @@ struct ac_llvm_context {
        unsigned range_md_kind;
        unsigned invariant_load_md_kind;
        unsigned uniform_md_kind;
-       unsigned fpmath_md_kind;
-       LLVMValueRef fpmath_md_2p5_ulp;
        LLVMValueRef empty_md;
 
        enum chip_class chip_class;
@@ -120,6 +134,8 @@ struct ac_llvm_context {
        unsigned wave_size;
        unsigned ballot_mask_bits;
 
+       unsigned float_mode;
+
        LLVMValueRef lds;
 };
 
@@ -168,7 +184,8 @@ void ac_build_s_barrier(struct ac_llvm_context *ctx);
 void ac_build_optimization_barrier(struct ac_llvm_context *ctx,
                                   LLVMValueRef *pvgpr);
 
-LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx);
+LLVMValueRef ac_build_shader_clock(struct ac_llvm_context *ctx,
+                                  nir_scope scope);
 
 LLVMValueRef ac_build_ballot(struct ac_llvm_context *ctx, LLVMValueRef value);
 LLVMValueRef ac_get_i1_sgpr_mask(struct ac_llvm_context *ctx,
@@ -293,8 +310,7 @@ ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
                            LLVMValueRef voffset,
                            LLVMValueRef soffset,
                            unsigned inst_offset,
-                           unsigned cache_policy,
-                           bool swizzle_enable_hint);
+                           unsigned cache_policy);
 
 void
 ac_build_buffer_store_format(struct ac_llvm_context *ctx,
@@ -302,7 +318,6 @@ ac_build_buffer_store_format(struct ac_llvm_context *ctx,
                             LLVMValueRef data,
                             LLVMValueRef vindex,
                             LLVMValueRef voffset,
-                            unsigned num_channels,
                             unsigned cache_policy);
 
 LLVMValueRef
@@ -323,7 +338,8 @@ LLVMValueRef ac_build_buffer_load_format(struct ac_llvm_context *ctx,
                                         LLVMValueRef voffset,
                                         unsigned num_channels,
                                         unsigned cache_policy,
-                                        bool can_speculate);
+                                        bool can_speculate,
+                                        bool d16);
 
 LLVMValueRef
 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
@@ -527,6 +543,7 @@ enum ac_image_cache_policy {
        ac_glc = 1 << 0, /* per-CU cache control */
        ac_slc = 1 << 1, /* global L2 cache control */
        ac_dlc = 1 << 2, /* per-shader-array cache control */
+       ac_swizzled = 1 << 3, /* the access is swizzled, disabling load/store merging */
 };
 
 struct ac_image_args {
@@ -537,6 +554,7 @@ struct ac_image_args {
        unsigned cache_policy : 3;
        bool unorm : 1;
        bool level_zero : 1;
+       bool d16 : 1; /* data and return values are 16-bit, requires GFX8+ */
        unsigned attributes; /* additional call-site specific AC_FUNC_ATTRs */
 
        LLVMValueRef resource;
@@ -548,6 +566,7 @@ struct ac_image_args {
        LLVMValueRef derivs[6];
        LLVMValueRef coords[4];
        LLVMValueRef lod; // also used by ac_image_get_resinfo
+       LLVMValueRef min_lod;
 };
 
 LLVMValueRef ac_build_image_opcode(struct ac_llvm_context *ctx,
@@ -578,17 +597,8 @@ void ac_build_waitcnt(struct ac_llvm_context *ctx, unsigned wait_flags);
 
 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_isign(struct ac_llvm_context *ctx, LLVMValueRef src0);
+LLVMValueRef ac_build_fsign(struct ac_llvm_context *ctx, LLVMValueRef src);
 LLVMValueRef ac_build_bit_count(struct ac_llvm_context *ctx, LLVMValueRef src0);
 
 LLVMValueRef ac_build_bitfield_reverse(struct ac_llvm_context *ctx,
@@ -598,6 +608,7 @@ void ac_optimize_vs_outputs(struct ac_llvm_context *ac,
                            LLVMValueRef main_fn,
                            uint8_t *vs_output_param_offset,
                            uint32_t num_outputs,
+                           uint32_t skip_output_mask,
                            uint8_t *num_param_exports);
 void ac_init_exec_full_mask(struct ac_llvm_context *ctx);
 
@@ -646,6 +657,9 @@ void ac_apply_fmask_to_sample(struct ac_llvm_context *ac, LLVMValueRef fmask,
 LLVMValueRef
 ac_build_ds_swizzle(struct ac_llvm_context *ctx, LLVMValueRef src, unsigned mask);
 
+LLVMValueRef ac_build_readlane_no_opt_barrier(struct ac_llvm_context *ctx,
+                                             LLVMValueRef src, LLVMValueRef lane);
+
 LLVMValueRef
 ac_build_readlane(struct ac_llvm_context *ctx, LLVMValueRef src, LLVMValueRef lane);
 
@@ -715,12 +729,19 @@ LLVMValueRef
 ac_build_frexp_mant(struct ac_llvm_context *ctx, LLVMValueRef src0,
                    unsigned bitsize);
 
+LLVMValueRef
+ac_build_canonicalize(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_is_helper_invocation(struct ac_llvm_context *ctx);
+
 LLVMValueRef ac_build_call(struct ac_llvm_context *ctx, LLVMValueRef func,
                           LLVMValueRef *args, unsigned num_args);
 
@@ -737,6 +758,53 @@ ac_export_mrt_z(struct ac_llvm_context *ctx, LLVMValueRef depth,
                LLVMValueRef stencil, LLVMValueRef samplemask,
                struct ac_export_args *args);
 
+void ac_build_sendmsg_gs_alloc_req(struct ac_llvm_context *ctx, LLVMValueRef wave_id,
+                                  LLVMValueRef vtx_cnt, LLVMValueRef prim_cnt);
+
+struct ac_ngg_prim {
+       unsigned num_vertices;
+       LLVMValueRef isnull;
+       LLVMValueRef index[3];
+       LLVMValueRef edgeflag[3];
+       LLVMValueRef passthrough;
+};
+
+LLVMValueRef ac_pack_prim_export(struct ac_llvm_context *ctx,
+                                const struct ac_ngg_prim *prim);
+void ac_build_export_prim(struct ac_llvm_context *ctx,
+                         const struct ac_ngg_prim *prim);
+
+static inline LLVMValueRef
+ac_get_arg(struct ac_llvm_context *ctx, struct ac_arg arg)
+{
+       assert(arg.used);
+       return LLVMGetParam(ctx->main_function, arg.arg_index);
+}
+
+enum ac_llvm_calling_convention {
+       AC_LLVM_AMDGPU_VS = 87,
+       AC_LLVM_AMDGPU_GS = 88,
+       AC_LLVM_AMDGPU_PS = 89,
+       AC_LLVM_AMDGPU_CS = 90,
+       AC_LLVM_AMDGPU_HS = 93,
+};
+
+LLVMValueRef ac_build_main(const struct ac_shader_args *args,
+                          struct ac_llvm_context *ctx,
+                          enum ac_llvm_calling_convention convention,
+                          const char *name, LLVMTypeRef ret_type,
+                          LLVMModuleRef module);
+void ac_build_s_endpgm(struct ac_llvm_context *ctx);
+
+LLVMValueRef ac_prefix_bitcount(struct ac_llvm_context *ctx,
+                               LLVMValueRef mask, LLVMValueRef index);
+LLVMValueRef ac_prefix_bitcount_2x64(struct ac_llvm_context *ctx,
+                                    LLVMValueRef mask[2], LLVMValueRef index);
+void ac_build_triangle_strip_indices_to_triangle(struct ac_llvm_context *ctx,
+                                                LLVMValueRef is_odd,
+                                                LLVMValueRef flatshade_first,
+                                                LLVMValueRef index[3]);
+
 #ifdef __cplusplus
 }
 #endif