gallivm: optimize lp_build_minify for sse
authorRoland Scheidegger <sroland@vmware.com>
Tue, 5 Nov 2013 18:21:25 +0000 (19:21 +0100)
committerRoland Scheidegger <sroland@vmware.com>
Tue, 5 Nov 2013 22:32:24 +0000 (23:32 +0100)
SSE can't handle true vector shifts (with variable shift count),
so llvm is turning them into a mess of extracts, scalar shifts and inserts.
It is however possible to emulate them in lp_build_minify with float muls,
which should be way faster (saves over 20 instructions per 8-wide
lp_build_minify). This wouldn't work for "generic" 32bit shifts though
since we've got only 24bits of mantissa (actually for left shifts it would
work by using sse41 int mul instead of float mul but not for right shifts).
Note that this has very limited scope for now, since this is only used with
per-pixel lod (otherwise we're avoiding the non-constant shift count by doing
per-quad shifts manually), and only 1d textures even then (though the latter
should change).

Reviewed-by: Brian Paul <brianp@vmware.com>
Reviewed-by: Jose Fonseca <jfonseca@vmware.com>
src/gallium/auxiliary/gallivm/lp_bld_sample.c
src/gallium/auxiliary/gallivm/lp_bld_sample.h
src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c

index a032d9d68954f8b9cd55ca4558d340c246157715..e60a035a8829a11ac45cdb14633172578f6c2574 100644 (file)
@@ -36,6 +36,7 @@
 #include "pipe/p_state.h"
 #include "util/u_format.h"
 #include "util/u_math.h"
+#include "util/u_cpu_detect.h"
 #include "lp_bld_arit.h"
 #include "lp_bld_const.h"
 #include "lp_bld_debug.h"
@@ -248,7 +249,7 @@ lp_build_rho(struct lp_build_sample_context *bld,
    first_level = bld->dynamic_state->first_level(bld->dynamic_state,
                                                  bld->gallivm, texture_unit);
    first_level_vec = lp_build_broadcast_scalar(int_size_bld, first_level);
-   int_size = lp_build_minify(int_size_bld, bld->int_size, first_level_vec);
+   int_size = lp_build_minify(int_size_bld, bld->int_size, first_level_vec, TRUE);
    float_size = lp_build_int_to_float(float_size_bld, int_size);
 
    if (cube_rho) {
@@ -1089,12 +1090,14 @@ lp_build_get_mip_offsets(struct lp_build_sample_context *bld,
 
 /**
  * Codegen equivalent for u_minify().
+ * @param lod_scalar  if lod is a (broadcasted) scalar
  * Return max(1, base_size >> level);
  */
 LLVMValueRef
 lp_build_minify(struct lp_build_context *bld,
                 LLVMValueRef base_size,
-                LLVMValueRef level)
+                LLVMValueRef level,
+                boolean lod_scalar)
 {
    LLVMBuilderRef builder = bld->gallivm->builder;
    assert(lp_check_value(bld->type, base_size));
@@ -1105,10 +1108,49 @@ lp_build_minify(struct lp_build_context *bld,
       return base_size;
    }
    else {
-      LLVMValueRef size =
-         LLVMBuildLShr(builder, base_size, level, "minify");
+      LLVMValueRef size;
       assert(bld->type.sign);
-      size = lp_build_max(bld, size, bld->one);
+      if (lod_scalar ||
+         (util_cpu_caps.has_avx2 || !util_cpu_caps.has_sse)) {
+         size = LLVMBuildLShr(builder, base_size, level, "minify");
+         size = lp_build_max(bld, size, bld->one);
+      }
+      else {
+         /*
+          * emulate shift with float mul, since intel "forgot" shifts with
+          * per-element shift count until avx2, which results in terrible
+          * scalar extraction (both count and value), scalar shift,
+          * vector reinsertion. Should not be an issue on any non-x86 cpu
+          * with a vector instruction set.
+          * On cpus with AMD's XOP this should also be unnecessary but I'm
+          * not sure if llvm would emit this with current flags.
+          */
+         LLVMValueRef const127, const23, lf;
+         struct lp_type ftype;
+         struct lp_build_context fbld;
+         ftype = lp_type_float_vec(32, bld->type.length * bld->type.width);
+         lp_build_context_init(&fbld, bld->gallivm, ftype);
+         const127 = lp_build_const_int_vec(bld->gallivm, bld->type, 127);
+         const23 = lp_build_const_int_vec(bld->gallivm, bld->type, 23);
+
+         /* calculate 2^(-level) float */
+         lf = lp_build_sub(bld, const127, level);
+         lf = lp_build_shl(bld, lf, const23);
+         lf = LLVMBuildBitCast(builder, lf, fbld.vec_type, "");
+
+         /* finish shift operation by doing float mul */
+         base_size = lp_build_int_to_float(&fbld, base_size);
+         size = lp_build_mul(&fbld, base_size, lf);
+         /*
+          * do the max also with floats because
+          * a) non-emulated int max requires sse41
+          *    (this is actually a lie as we could cast to 16bit values
+          *    as 16bit is sufficient and 16bit int max is sse2)
+          * b) with avx we can do int max 4-wide but float max 8-wide
+          */
+         size = lp_build_max(&fbld, size, fbld.one);
+         size = lp_build_itrunc(&fbld, size);
+      }
       return size;
    }
 }
@@ -1185,7 +1227,7 @@ lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
     */
    if (bld->num_mips == 1) {
       ilevel_vec = lp_build_broadcast_scalar(&bld->int_size_bld, ilevel);
-      *out_size = lp_build_minify(&bld->int_size_bld, bld->int_size, ilevel_vec);
+      *out_size = lp_build_minify(&bld->int_size_bld, bld->int_size, ilevel_vec, TRUE);
    }
    else {
       LLVMValueRef int_size_vec;
@@ -1229,7 +1271,7 @@ lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
                                                  bld4.type,
                                                  ilevel,
                                                  indexi);
-            tmp[i] = lp_build_minify(&bld4, int_size_vec, ileveli);
+            tmp[i] = lp_build_minify(&bld4, int_size_vec, ileveli, TRUE);
          }
          /*
           * out_size is [w0, h0, d0, _, w1, h1, d1, _, ...] vector for dims > 1,
@@ -1248,7 +1290,6 @@ lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
          * with 4-wide vector pack all elements into a 8xi16 vector
          * (on which we can still do useful math) instead of using a 16xi32
          * vector.
-         * FIXME: some callers can't handle this yet.
          * For dims == 1 this will create [w0, w1, w2, w3, ...] vector.
          * For dims > 1 this will create [w0, h0, d0, _, w1, h1, d1, _, ...] vector.
          */
@@ -1257,8 +1298,7 @@ lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
             assert(bld->int_size_in_bld.type.length == 1);
             int_size_vec = lp_build_broadcast_scalar(&bld->int_coord_bld,
                                                      bld->int_size);
-            /* vector shift with variable shift count alert... */
-            *out_size = lp_build_minify(&bld->int_coord_bld, int_size_vec, ilevel);
+            *out_size = lp_build_minify(&bld->int_coord_bld, int_size_vec, ilevel, FALSE);
          }
          else {
             LLVMValueRef ilevel1;
@@ -1267,7 +1307,7 @@ lp_build_mipmap_level_sizes(struct lp_build_sample_context *bld,
                ilevel1 = lp_build_extract_broadcast(bld->gallivm, bld->int_coord_type,
                                                     bld->int_size_in_bld.type, ilevel, indexi);
                tmp[i] = bld->int_size;
-               tmp[i] = lp_build_minify(&bld->int_size_in_bld, tmp[i], ilevel1);
+               tmp[i] = lp_build_minify(&bld->int_size_in_bld, tmp[i], ilevel1, TRUE);
             }
             *out_size = lp_build_concat(bld->gallivm, tmp,
                                         bld->int_size_in_bld.type,
index 5039128a203e5af03148732ce25ba856bab65959..fd4e0532607dbfef0ab82d73b1a35a8a303e8d9d 100644 (file)
@@ -547,7 +547,8 @@ lp_build_sample_nop(struct gallivm_state *gallivm,
 LLVMValueRef
 lp_build_minify(struct lp_build_context *bld,
                 LLVMValueRef base_size,
-                LLVMValueRef level);
+                LLVMValueRef level,
+                boolean lod_scalar);
 
 
 #endif /* LP_BLD_SAMPLE_H */
index 2d833318aeee4f9b2f361cd19921b9eca275c16f..e8c04d1e6c53343a9a9503202b1dcd49b378dcf4 100644 (file)
@@ -2940,7 +2940,7 @@ lp_build_size_query_soa(struct gallivm_state *gallivm,
                                     lp_build_const_int32(gallivm, 2), "");
    }
 
-   size = lp_build_minify(&bld_int_vec4, size, lod);
+   size = lp_build_minify(&bld_int_vec4, size, lod, TRUE);
 
    if (has_array)
       size = LLVMBuildInsertElement(gallivm->builder, size,