projects / mesa.git / blobdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | inline | side by side
ac: use new LLVM 8 intrinsic when storing 16-bit values
[mesa.git] / src / amd / common / ac_llvm_build.h
diff --git a/src/amd/common/ac_llvm_build.h b/src/amd/common/ac_llvm_build.h
index e90c8c21ad4ebf2dd462cd08616542357c57d9ca..723aec53cbe1254f953a6042916d83db05c3c9b2 100644 (file)
--- a/src/amd/common/ac_llvm_build.h
+++ b/src/amd/common/ac_llvm_build.h
@@ -34,14 +34,12 @@
 extern "C" {
 #endif
 
-#define HAVE_32BIT_POINTERS (HAVE_LLVM >= 0x0700)
-
 enum {
-       AC_ADDR_SPACE_FLAT = HAVE_LLVM >= 0x0700 ? 0 : 4, /* Slower than global. */
+       AC_ADDR_SPACE_FLAT = 0, /* Slower than global. */
        AC_ADDR_SPACE_GLOBAL = 1,
-       AC_ADDR_SPACE_GDS = HAVE_LLVM >= 0x0700 ? 2 : 5,
+       AC_ADDR_SPACE_GDS = 2,
        AC_ADDR_SPACE_LDS = 3,
-       AC_ADDR_SPACE_CONST = HAVE_LLVM >= 0x0700 ? 4 : 2, /* Global allowing SMEM. */
+       AC_ADDR_SPACE_CONST = 4, /* Global allowing SMEM. */
        AC_ADDR_SPACE_CONST_32BIT = 6, /* same as CONST, but the pointer type has 32 bits */
 };
 
@@ -173,9 +171,6 @@ LLVMValueRef
 ac_build_gather_values(struct ac_llvm_context *ctx,
                       LLVMValueRef *values,
                       unsigned value_count);
-LLVMValueRef ac_build_expand(struct ac_llvm_context *ctx,
-                            LLVMValueRef value,
-                            unsigned src_channels, unsigned dst_channels);
 LLVMValueRef ac_build_expand_to_vec4(struct ac_llvm_context *ctx,
                                     LLVMValueRef value,
                                     unsigned num_channels);
@@ -218,6 +213,14 @@ ac_build_fs_interp(struct ac_llvm_context *ctx,
                   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,
@@ -225,6 +228,11 @@ ac_build_fs_interp_mov(struct ac_llvm_context *ctx,
                       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,
@@ -258,6 +266,17 @@ ac_build_buffer_store_dword(struct ac_llvm_context *ctx,
                            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 writeonly_memory);
+
 LLVMValueRef
 ac_build_buffer_load(struct ac_llvm_context *ctx,
                     LLVMValueRef rsrc,
@@ -292,11 +311,75 @@ LLVMValueRef ac_build_buffer_load_format_gfx9_safe(struct ac_llvm_context *ctx,
 LLVMValueRef
 ac_build_tbuffer_load_short(struct ac_llvm_context *ctx,
                            LLVMValueRef rsrc,
-                           LLVMValueRef vindex,
                            LLVMValueRef voffset,
-                               LLVMValueRef soffset,
-                               LLVMValueRef immoffset,
-                               LLVMValueRef glc);
+                           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);
+
+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_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);
@@ -482,6 +565,7 @@ void ac_build_continue(struct ac_llvm_context *ctx);
 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,
@@ -525,6 +609,42 @@ ac_build_exclusive_scan(struct ac_llvm_context *ctx, LLVMValueRef src, nir_op op
 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);