intel/fs: Add support for subgroup quad operations

NIR has code to lower these away for us but we can do significantly
better in many cases with register regioning and SIMD4x2.

Acked-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>

5 files changed, 124 insertions, 0 deletions

diff --git a/src/intel/compiler/brw_eu_defines.h b/src/intel/compiler/brw_eu_defines.h
index 3c4c538ac17..332d627bc37 100644
--- a/src/intel/compiler/brw_eu_defines.h
+++ b/src/intel/compiler/brw_eu_defines.h
@@ -467,6 +467,11 @@ enum opcode {
     */
    SHADER_OPCODE_SEL_EXEC,
 
+   /* This turns into an align16 mov from src0 to dst with a swizzle
+    * provided as an immediate in src1.
+    */
+   SHADER_OPCODE_QUAD_SWIZZLE,
+
    /* Take every Nth element in src0 and broadcast it to the group of N
     * channels in which it lives in the destination.  The offset within the
     * cluster is given by src1 and the cluster size is given by src2.
diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index 9f1b8d0b184..53ba94ccccb 100644
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -5233,6 +5233,9 @@ get_lowered_simd_width(const struct gen_device_info *devinfo,
    case SHADER_OPCODE_URB_WRITE_SIMD8_MASKED_PER_SLOT:
       return MIN2(8, inst->exec_size);
 
+   case SHADER_OPCODE_QUAD_SWIZZLE:
+      return 8;
+
    case SHADER_OPCODE_MOV_INDIRECT: {
       /* From IVB and HSW PRMs:
        *
diff --git a/src/intel/compiler/brw_fs_generator.cpp b/src/intel/compiler/brw_fs_generator.cpp
index 382548f5c36..5371246fd24 100644
--- a/src/intel/compiler/brw_fs_generator.cpp
+++ b/src/intel/compiler/brw_fs_generator.cpp
@@ -2301,6 +2301,26 @@ fs_generator::generate_code(const cfg_t *cfg, int dispatch_width)
          brw_MOV(p, dst, src[0]);
          break;
 
+      case SHADER_OPCODE_QUAD_SWIZZLE:
+         /* This only works on 8-wide 32-bit values */
+         assert(inst->exec_size == 8);
+         assert(type_sz(src[0].type) == 4);
+         assert(inst->force_writemask_all);
+         assert(src[1].file == BRW_IMMEDIATE_VALUE);
+         assert(src[1].type == BRW_REGISTER_TYPE_UD);
+
+         if (src[0].file == BRW_IMMEDIATE_VALUE ||
+             (src[0].vstride == 0 && src[0].hstride == 0)) {
+            /* The value is uniform across all channels */
+            brw_MOV(p, dst, src[0]);
+         } else {
+            brw_set_default_access_mode(p, BRW_ALIGN_16);
+            struct brw_reg swiz_src = stride(src[0], 4, 4, 1);
+            swiz_src.swizzle = inst->src[1].ud;
+            brw_MOV(p, dst, swiz_src);
+         }
+         break;
+
       case SHADER_OPCODE_CLUSTER_BROADCAST: {
          assert(src[0].type == dst.type);
          assert(!src[0].negate && !src[0].abs);
diff --git a/src/intel/compiler/brw_fs_nir.cpp b/src/intel/compiler/brw_fs_nir.cpp
index e25fd38af91..dbd2105f7e9 100644
--- a/src/intel/compiler/brw_fs_nir.cpp
+++ b/src/intel/compiler/brw_fs_nir.cpp
@@ -4588,6 +4588,100 @@ fs_visitor::nir_emit_intrinsic(const fs_builder &bld, nir_intrinsic_instr *instr
       break;
    }
 
+   case nir_intrinsic_quad_broadcast: {
+      const fs_reg value = get_nir_src(instr->src[0]);
+      nir_const_value *index = nir_src_as_const_value(instr->src[1]);
+      assert(nir_src_bit_size(instr->src[1]) == 32);
+
+      bld.emit(SHADER_OPCODE_CLUSTER_BROADCAST, retype(dest, value.type),
+               value, brw_imm_ud(index->u32[0]), brw_imm_ud(4));
+      break;
+   }
+
+   case nir_intrinsic_quad_swap_horizontal: {
+      const fs_reg value = get_nir_src(instr->src[0]);
+      const fs_reg tmp = bld.vgrf(value.type);
+      const fs_builder ubld = bld.exec_all().group(dispatch_width / 2, 0);
+
+      const fs_reg src_left = horiz_stride(value, 2);
+      const fs_reg src_right = horiz_stride(horiz_offset(value, 1), 2);
+      const fs_reg tmp_left = horiz_stride(tmp, 2);
+      const fs_reg tmp_right = horiz_stride(horiz_offset(tmp, 1), 2);
+
+      /* From the Cherryview PRM Vol. 7, "Register Region Restrictiosn":
+       *
+       *    "When source or destination datatype is 64b or operation is
+       *    integer DWord multiply, regioning in Align1 must follow
+       *    these rules:
+       *
+       *    [...]
+       *
+       *    3. Source and Destination offset must be the same, except
+       *       the case of scalar source."
+       *
+       * In order to work around this, we have to emit two 32-bit MOVs instead
+       * of a single 64-bit MOV to do the shuffle.
+       */
+      if (type_sz(value.type) > 4 &&
+          (devinfo->is_cherryview || gen_device_info_is_9lp(devinfo))) {
+         ubld.MOV(subscript(tmp_left, BRW_REGISTER_TYPE_D, 0),
+                  subscript(src_right, BRW_REGISTER_TYPE_D, 0));
+         ubld.MOV(subscript(tmp_left, BRW_REGISTER_TYPE_D, 1),
+                  subscript(src_right, BRW_REGISTER_TYPE_D, 1));
+         ubld.MOV(subscript(tmp_right, BRW_REGISTER_TYPE_D, 0),
+                  subscript(src_left, BRW_REGISTER_TYPE_D, 0));
+         ubld.MOV(subscript(tmp_right, BRW_REGISTER_TYPE_D, 1),
+                  subscript(src_left, BRW_REGISTER_TYPE_D, 1));
+      } else {
+         ubld.MOV(tmp_left, src_right);
+         ubld.MOV(tmp_right, src_left);
+      }
+      bld.MOV(retype(dest, value.type), tmp);
+      break;
+   }
+
+   case nir_intrinsic_quad_swap_vertical: {
+      const fs_reg value = get_nir_src(instr->src[0]);
+      if (nir_src_bit_size(instr->src[0]) == 32) {
+         /* For 32-bit, we can use a SIMD4x2 instruction to do this easily */
+         const fs_reg tmp = bld.vgrf(value.type);
+         const fs_builder ubld = bld.exec_all();
+         ubld.emit(SHADER_OPCODE_QUAD_SWIZZLE, tmp, value,
+                   brw_imm_ud(BRW_SWIZZLE4(2,3,0,1)));
+         bld.MOV(retype(dest, value.type), tmp);
+      } else {
+         /* For larger data types, we have to either emit dispatch_width many
+          * MOVs or else fall back to doing indirects.
+          */
+         fs_reg idx = bld.vgrf(BRW_REGISTER_TYPE_W);
+         bld.XOR(idx, nir_system_values[SYSTEM_VALUE_SUBGROUP_INVOCATION],
+                      brw_imm_w(0x2));
+         bld.emit(SHADER_OPCODE_SHUFFLE, retype(dest, value.type), value, idx);
+      }
+      break;
+   }
+
+   case nir_intrinsic_quad_swap_diagonal: {
+      const fs_reg value = get_nir_src(instr->src[0]);
+      if (nir_src_bit_size(instr->src[0]) == 32) {
+         /* For 32-bit, we can use a SIMD4x2 instruction to do this easily */
+         const fs_reg tmp = bld.vgrf(value.type);
+         const fs_builder ubld = bld.exec_all();
+         ubld.emit(SHADER_OPCODE_QUAD_SWIZZLE, tmp, value,
+                   brw_imm_ud(BRW_SWIZZLE4(3,2,1,0)));
+         bld.MOV(retype(dest, value.type), tmp);
+      } else {
+         /* For larger data types, we have to either emit dispatch_width many
+          * MOVs or else fall back to doing indirects.
+          */
+         fs_reg idx = bld.vgrf(BRW_REGISTER_TYPE_W);
+         bld.XOR(idx, nir_system_values[SYSTEM_VALUE_SUBGROUP_INVOCATION],
+                      brw_imm_w(0x3));
+         bld.emit(SHADER_OPCODE_SHUFFLE, retype(dest, value.type), value, idx);
+      }
+      break;
+   }
+
    case nir_intrinsic_reduce: {
       fs_reg src = get_nir_src(instr->src[0]);
       nir_op redop = (nir_op)nir_intrinsic_reduction_op(instr);
diff --git a/src/intel/compiler/brw_shader.cpp b/src/intel/compiler/brw_shader.cpp
index e822c100e9f..ffe8a7403da 100644
--- a/src/intel/compiler/brw_shader.cpp
+++ b/src/intel/compiler/brw_shader.cpp
@@ -334,6 +334,8 @@ brw_instruction_name(const struct gen_device_info *devinfo, enum opcode op)
       return "shuffle";
    case SHADER_OPCODE_SEL_EXEC:
       return "sel_exec";
+   case SHADER_OPCODE_QUAD_SWIZZLE:
+      return "quad_swizzle";
    case SHADER_OPCODE_CLUSTER_BROADCAST:
       return "cluster_broadcast";