gallivm: Add no_rho_approx debug option
authorRoland Scheidegger <sroland@vmware.com>
Thu, 18 Apr 2013 15:04:01 +0000 (17:04 +0200)
committerRoland Scheidegger <sroland@vmware.com>
Thu, 18 Apr 2013 15:04:01 +0000 (17:04 +0200)
This will calculate rho correctly as
sqrt(max((ds/dx)^2 + (dt/dx)^2 + (dr/dx)^2), (ds/dx)^2 + (dt/dx)^2 + (dr/dx)^2))
instead of max(|ds/dx|,|dt/dx|,|dr/dx|,|ds/dy|,|dt/dy,|dr/dy|)
(for 3 coords - 2 coords work analogous, for 1 coord there's no point doing
the exact version), for both implicit and explicit derivatives.
While such approximation seems to be allowed in OpenGL some APIs may be less
forgiving, and the error can be quite large (sqrt(2) for 2 coords, sqrt(3) for
3 coords so wrong by nearly one mip level in the latter case).
This also helps to single out "real" bugs from "expected" ones, so it is debug
only (though at least combined with no_brilinear I didn't really see much of a
performance difference but only tested with a debug build - at least with
implicit mipmaps the instruction count is almost exactly the same though the
instructions are more complex (1 sqrt and mul/adds instead of and/max mostly).
The code when the option isn't set stays exactly the same.

v2: rename no_rho_opt to no_rho_approx.

Reviewed-by: Brian Paul <brianp@vmware.com>
src/gallium/auxiliary/gallivm/lp_bld_debug.h
src/gallium/auxiliary/gallivm/lp_bld_init.c
src/gallium/auxiliary/gallivm/lp_bld_sample.c

index ab83d98feedfdbd2a9261e542851c4973c45a6c5..4f38edf11f857d21fe1215504709debe17c97948 100644 (file)
@@ -42,7 +42,8 @@
 #define GALLIVM_DEBUG_NO_OPT        (1 << 3)
 #define GALLIVM_DEBUG_PERF          (1 << 4)
 #define GALLIVM_DEBUG_NO_BRILINEAR  (1 << 5)
-#define GALLIVM_DEBUG_GC            (1 << 6)
+#define GALLIVM_DEBUG_NO_RHO_APPROX (1 << 6)
+#define GALLIVM_DEBUG_GC            (1 << 7)
 
 
 #ifdef __cplusplus
index 1153411dd5255caf65efc4a42c85e4bb4ffcef2b..673c16e67623fc51ebd774efecc32744442e9567 100644 (file)
@@ -79,6 +79,7 @@ static const struct debug_named_value lp_bld_debug_flags[] = {
    { "nopt",   GALLIVM_DEBUG_NO_OPT, NULL },
    { "perf",   GALLIVM_DEBUG_PERF, NULL },
    { "no_brilinear", GALLIVM_DEBUG_NO_BRILINEAR, NULL },
+   { "no_rho_approx", GALLIVM_DEBUG_NO_RHO_APPROX, NULL },
    { "gc",     GALLIVM_DEBUG_GC, NULL },
    DEBUG_NAMED_VALUE_END
 };
index f8853631c8bc8f7a8796169fe92f0f92ce5988dd..666e5692100fabb18947fd4f5b8aa67ac41a8377 100644 (file)
@@ -257,31 +257,59 @@ lp_build_rho(struct lp_build_sample_context *bld,
                                       perquadf_bld->type, rho_vec, 0);
    }
    else if (derivs && !(bld->static_texture_state->target == PIPE_TEXTURE_CUBE)) {
-      LLVMValueRef ddmax[3];
+      LLVMValueRef ddmax[3], ddx[3], ddy[3];
       for (i = 0; i < dims; i++) {
-         LLVMValueRef ddx, ddy;
          LLVMValueRef floatdim;
          LLVMValueRef indexi = lp_build_const_int32(gallivm, i);
-         ddx = lp_build_abs(coord_bld, derivs->ddx[i]);
-         ddy = lp_build_abs(coord_bld, derivs->ddy[i]);
-         ddmax[i] = lp_build_max(coord_bld, ddx, ddy);
+
          floatdim = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
                                                coord_bld->type, float_size, indexi);
-         ddmax[i] = lp_build_mul(coord_bld, floatdim, ddmax[i]);
+
+         if ((gallivm_debug & GALLIVM_DEBUG_NO_RHO_APPROX) && (dims > 1)) {
+            ddx[i] = lp_build_mul(coord_bld, floatdim, derivs->ddx[i]);
+            ddy[i] = lp_build_mul(coord_bld, floatdim, derivs->ddy[i]);
+            ddx[i] = lp_build_mul(coord_bld, ddx[i], ddx[i]);
+            ddy[i] = lp_build_mul(coord_bld, ddy[i], ddy[i]);
+         }
+         else {
+            LLVMValueRef tmpx, tmpy;
+            tmpx = lp_build_abs(coord_bld, derivs->ddx[i]);
+            tmpy = lp_build_abs(coord_bld, derivs->ddy[i]);
+            ddmax[i] = lp_build_max(coord_bld, tmpx, tmpy);
+            ddmax[i] = lp_build_mul(coord_bld, floatdim, ddmax[i]);
+         }
       }
-      rho_vec = ddmax[0];
-      if (dims > 1) {
-         rho_vec = lp_build_max(coord_bld, rho_vec, ddmax[1]);
+      if ((gallivm_debug & GALLIVM_DEBUG_NO_RHO_APPROX) && (dims > 1)) {
+         rho_xvec = lp_build_add(coord_bld, ddx[0], ddx[1]);
+         rho_yvec = lp_build_add(coord_bld, ddy[0], ddy[1]);
          if (dims > 2) {
-            rho_vec = lp_build_max(coord_bld, rho_vec, ddmax[2]);
+            rho_xvec = lp_build_add(coord_bld, rho_xvec, ddx[2]);
+            rho_yvec = lp_build_add(coord_bld, rho_yvec, ddy[2]);
          }
+         rho_vec = lp_build_max(coord_bld, rho_xvec, rho_yvec);
+         rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
+                                         perquadf_bld->type, rho_vec, 0);
+         /*
+          * note that as long as we don't care about per-pixel lod could reduce math
+          * more (at some shuffle cost), but for now only do sqrt after packing.
+          */
+         rho = lp_build_sqrt(perquadf_bld, rho);
+      }
+      else {
+         rho_vec = ddmax[0];
+         if (dims > 1) {
+            rho_vec = lp_build_max(coord_bld, rho_vec, ddmax[1]);
+            if (dims > 2) {
+               rho_vec = lp_build_max(coord_bld, rho_vec, ddmax[2]);
+            }
+         }
+         /*
+          * rho_vec now still contains per-pixel rho, convert to scalar per quad
+          * since we can't handle per-pixel rho/lod from now on (TODO).
+          */
+         rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
+                                         perquadf_bld->type, rho_vec, 0);
       }
-      /*
-       * rho_vec now still contains per-pixel rho, convert to scalar per quad
-       * since we can't handle per-pixel rho/lod from now on (TODO).
-       */
-      rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
-                                      perquadf_bld->type, rho_vec, 0);
    }
    else {
       /*
@@ -289,6 +317,19 @@ lp_build_rho(struct lp_build_sample_context *bld,
        * (the shuffle code makes it look worse than it is).
        * Still, might not be ideal for all cases.
        */
+      static const unsigned char swizzle0[] = { /* no-op swizzle */
+         0, LP_BLD_SWIZZLE_DONTCARE,
+         LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+      };
+      static const unsigned char swizzle1[] = {
+         1, LP_BLD_SWIZZLE_DONTCARE,
+         LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+      };
+      static const unsigned char swizzle2[] = {
+         2, LP_BLD_SWIZZLE_DONTCARE,
+         LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+      };
+
       if (dims < 2) {
          ddx_ddy[0] = lp_build_packed_ddx_ddy_onecoord(coord_bld, s);
       }
@@ -299,127 +340,151 @@ lp_build_rho(struct lp_build_sample_context *bld,
          }
       }
 
-      ddx_ddy[0] = lp_build_abs(coord_bld, ddx_ddy[0]);
-      if (dims > 2) {
-         ddx_ddy[1] = lp_build_abs(coord_bld, ddx_ddy[1]);
-      }
-
-      if (dims < 2) {
-         static const unsigned char swizzle1[] = { /* no-op swizzle */
-            0, LP_BLD_SWIZZLE_DONTCARE,
+      if ((gallivm_debug & GALLIVM_DEBUG_NO_RHO_APPROX) && (dims > 1)) {
+         static const unsigned char swizzle01[] = { /* no-op swizzle */
+            0, 1,
             LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
          };
-         static const unsigned char swizzle2[] = {
-            2, LP_BLD_SWIZZLE_DONTCARE,
+         static const unsigned char swizzle23[] = {
+            2, 3,
             LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
          };
-         rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle1);
-         rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle2);
-      }
-      else if (dims == 2) {
-         static const unsigned char swizzle1[] = {
-            0, 2,
-            LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
-         };
-         static const unsigned char swizzle2[] = {
-            1, 3,
-            LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
-         };
-         rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle1);
-         rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle2);
-      }
-      else {
-         LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH];
-         LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH];
-         assert(dims == 3);
+         LLVMValueRef ddx_ddys, ddx_ddyt, floatdim, shuffles[LP_MAX_VECTOR_LENGTH / 4];
+
          for (i = 0; i < num_quads; i++) {
-            shuffles1[4*i + 0] = lp_build_const_int32(gallivm, 4*i);
-            shuffles1[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 2);
-            shuffles1[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i);
-            shuffles1[4*i + 3] = i32undef;
-            shuffles2[4*i + 0] = lp_build_const_int32(gallivm, 4*i + 1);
-            shuffles2[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 3);
-            shuffles2[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i + 2);
-            shuffles2[4*i + 3] = i32undef;
+            shuffles[i*4+0] = shuffles[i*4+1] = index0;
+            shuffles[i*4+2] = shuffles[i*4+3] = index1;
          }
-         rho_xvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
-                                           LLVMConstVector(shuffles1, length), "");
-         rho_yvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
-                                           LLVMConstVector(shuffles2, length), "");
-      }
-
-      rho_vec = lp_build_max(coord_bld, rho_xvec, rho_yvec);
+         floatdim = LLVMBuildShuffleVector(builder, float_size, float_size,
+                                           LLVMConstVector(shuffles, length), "");
+         ddx_ddy[0] = lp_build_mul(coord_bld, ddx_ddy[0], floatdim);
+         ddx_ddy[0] = lp_build_mul(coord_bld, ddx_ddy[0], ddx_ddy[0]);
+         ddx_ddys = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle01);
+         ddx_ddyt = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle23);
+         rho_vec = lp_build_add(coord_bld, ddx_ddys, ddx_ddyt);
 
-      if (bld->coord_type.length > 4) {
-         /* expand size to each quad */
-         if (dims > 1) {
-            /* could use some broadcast_vector helper for this? */
-            int num_quads = bld->coord_type.length / 4;
-            LLVMValueRef src[LP_MAX_VECTOR_LENGTH/4];
-            for (i = 0; i < num_quads; i++) {
-               src[i] = float_size;
-            }
-            float_size = lp_build_concat(bld->gallivm, src, float_size_bld->type, num_quads);
+         if (dims > 2) {
+            static const unsigned char swizzle02[] = {
+               0, 2,
+               LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+            };
+            floatdim = lp_build_extract_broadcast(gallivm, bld->float_size_in_type,
+                                                  coord_bld->type, float_size, index2);
+            ddx_ddy[1] = lp_build_mul(coord_bld, ddx_ddy[1], floatdim);
+            ddx_ddy[1] = lp_build_mul(coord_bld, ddx_ddy[1], ddx_ddy[1]);
+            ddx_ddy[1] = lp_build_swizzle_aos(coord_bld, ddx_ddy[1], swizzle02);
+            rho_vec = lp_build_add(coord_bld, rho_vec, ddx_ddy[1]);
          }
-         else {
-            float_size = lp_build_broadcast_scalar(coord_bld, float_size);
+         rho_xvec = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle0);
+         rho_yvec = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
+         rho_vec = lp_build_max(coord_bld, rho_xvec, rho_yvec);
+
+         rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
+                                         perquadf_bld->type, rho_vec, 0);
+         rho = lp_build_sqrt(perquadf_bld, rho);
+      }
+      else {
+         ddx_ddy[0] = lp_build_abs(coord_bld, ddx_ddy[0]);
+         if (dims > 2) {
+            ddx_ddy[1] = lp_build_abs(coord_bld, ddx_ddy[1]);
          }
-         rho_vec = lp_build_mul(coord_bld, rho_vec, float_size);
 
-         if (dims <= 1) {
-            rho = rho_vec;
+         if (dims < 2) {
+            rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle0);
+            rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle1);
+         }
+         else if (dims == 2) {
+            static const unsigned char swizzle02[] = {
+               0, 2,
+               LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+            };
+            static const unsigned char swizzle13[] = {
+               1, 3,
+               LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
+            };
+            rho_xvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle02);
+            rho_yvec = lp_build_swizzle_aos(coord_bld, ddx_ddy[0], swizzle13);
          }
          else {
-            if (dims >= 2) {
-               static const unsigned char swizzle1[] = {
-                  0, LP_BLD_SWIZZLE_DONTCARE,
-                  LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
-               };
-               static const unsigned char swizzle2[] = {
-                  1, LP_BLD_SWIZZLE_DONTCARE,
-                  LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
-               };
-               LLVMValueRef rho_s, rho_t, rho_r;
-
-               rho_s = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
-               rho_t = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle2);
-
-               rho = lp_build_max(coord_bld, rho_s, rho_t);
-
-               if (dims >= 3) {
-                  static const unsigned char swizzle3[] = {
-                     2, LP_BLD_SWIZZLE_DONTCARE,
-                     LP_BLD_SWIZZLE_DONTCARE, LP_BLD_SWIZZLE_DONTCARE
-                  };
-                  rho_r = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle3);
-                  rho = lp_build_max(coord_bld, rho, rho_r);
-               }
+            LLVMValueRef shuffles1[LP_MAX_VECTOR_LENGTH];
+            LLVMValueRef shuffles2[LP_MAX_VECTOR_LENGTH];
+            assert(dims == 3);
+            for (i = 0; i < num_quads; i++) {
+               shuffles1[4*i + 0] = lp_build_const_int32(gallivm, 4*i);
+               shuffles1[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 2);
+               shuffles1[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i);
+               shuffles1[4*i + 3] = i32undef;
+               shuffles2[4*i + 0] = lp_build_const_int32(gallivm, 4*i + 1);
+               shuffles2[4*i + 1] = lp_build_const_int32(gallivm, 4*i + 3);
+               shuffles2[4*i + 2] = lp_build_const_int32(gallivm, length + 4*i + 2);
+               shuffles2[4*i + 3] = i32undef;
             }
+            rho_xvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
+                                              LLVMConstVector(shuffles1, length), "");
+            rho_yvec = LLVMBuildShuffleVector(builder, ddx_ddy[0], ddx_ddy[1],
+                                              LLVMConstVector(shuffles2, length), "");
          }
-         rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
-                                         perquadf_bld->type, rho, 0);
-      }
-      else {
-         if (dims <= 1) {
-            rho_vec = LLVMBuildExtractElement(builder, rho_vec, index0, "");
-         }
-         rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
 
-         if (dims <= 1) {
-            rho = rho_vec;
+         rho_vec = lp_build_max(coord_bld, rho_xvec, rho_yvec);
+
+         if (bld->coord_type.length > 4) {
+            /* expand size to each quad */
+            if (dims > 1) {
+               /* could use some broadcast_vector helper for this? */
+               LLVMValueRef src[LP_MAX_VECTOR_LENGTH/4];
+               for (i = 0; i < num_quads; i++) {
+                  src[i] = float_size;
+               }
+               float_size = lp_build_concat(bld->gallivm, src, float_size_bld->type, num_quads);
+            }
+            else {
+               float_size = lp_build_broadcast_scalar(coord_bld, float_size);
+            }
+            rho_vec = lp_build_mul(coord_bld, rho_vec, float_size);
+
+            if (dims <= 1) {
+               rho = rho_vec;
+            }
+            else {
+               if (dims >= 2) {
+                  LLVMValueRef rho_s, rho_t, rho_r;
+
+                  rho_s = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle0);
+                  rho_t = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle1);
+
+                  rho = lp_build_max(coord_bld, rho_s, rho_t);
+
+                  if (dims >= 3) {
+                     rho_r = lp_build_swizzle_aos(coord_bld, rho_vec, swizzle2);
+                     rho = lp_build_max(coord_bld, rho, rho_r);
+                  }
+               }
+            }
+            rho = lp_build_pack_aos_scalars(bld->gallivm, coord_bld->type,
+                                            perquadf_bld->type, rho, 0);
          }
          else {
-            if (dims >= 2) {
-               LLVMValueRef rho_s, rho_t, rho_r;
+            if (dims <= 1) {
+               rho_vec = LLVMBuildExtractElement(builder, rho_vec, index0, "");
+            }
+            rho_vec = lp_build_mul(float_size_bld, rho_vec, float_size);
+
+            if (dims <= 1) {
+               rho = rho_vec;
+            }
+            else {
+               if (dims >= 2) {
+                  LLVMValueRef rho_s, rho_t, rho_r;
 
-               rho_s = LLVMBuildExtractElement(builder, rho_vec, index0, "");
-               rho_t = LLVMBuildExtractElement(builder, rho_vec, index1, "");
+                  rho_s = LLVMBuildExtractElement(builder, rho_vec, index0, "");
+                  rho_t = LLVMBuildExtractElement(builder, rho_vec, index1, "");
 
-               rho = lp_build_max(float_bld, rho_s, rho_t);
+                  rho = lp_build_max(float_bld, rho_s, rho_t);
 
-               if (dims >= 3) {
-                  rho_r = LLVMBuildExtractElement(builder, rho_vec, index2, "");
-                  rho = lp_build_max(float_bld, rho, rho_r);
+                  if (dims >= 3) {
+                     rho_r = LLVMBuildExtractElement(builder, rho_vec, index2, "");
+                     rho = lp_build_max(float_bld, rho, rho_r);
+                  }
                }
             }
          }