panvk: add support for {s,u}dot_4x8_{sat}

Generate IDPADD instruction to support integer dot product
Support is added for both signed/unsigned dot product as well as
saturated dot product.
Support is only for v9+.

Reviewed-by: Boris Brezillon <boris.brezillon@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34812>

src/panfrost/compiler/bifrost_compile.c

@@ -3515,6 +3515,20 @@ bi_emit_alu(bi_builder *b, nir_alu_instr *instr)
       bi_isub_u32_to(b, dst, bi_imm_u32(src_sz - 1), clz, false);
       break;
    }
+   case nir_op_udot_4x8_uadd_sat:
+   case nir_op_udot_4x8_uadd: {
+      assert(b->shader->arch >= 9);
+      bi_idpadd_v4u8_to(b, dst, s0, s1, s2,
+                        instr->op == nir_op_udot_4x8_uadd_sat);
+      break;
+   }
+   case nir_op_sdot_4x8_iadd_sat:
+   case nir_op_sdot_4x8_iadd: {
+      assert(b->shader->arch >= 9);
+      bi_idpadd_v4s8_to(b, dst, s0, s1, s2,
+                        instr->op == nir_op_sdot_4x8_iadd_sat);
+      break;
+   }
 
    default:
       fprintf(stderr, "Unhandled ALU op %s\n", nir_op_infos[instr->op].name);

src/panfrost/compiler/bifrost_compile.h

@@ -148,6 +148,10 @@ void bifrost_compile_shader_nir(nir_shader *nir,
       .support_indirect_inputs = (uint8_t)BITFIELD_MASK(PIPE_SHADER_TYPES),    \
       .lower_hadd = arch >= 11,                                                \
       .discard_is_demote = true,                                               \
+      .has_udot_4x8 = arch >= 9,                                               \
+      .has_udot_4x8_sat = arch >= 9,                                           \
+      .has_sdot_4x8 = arch >= 9,                                               \
+      .has_sdot_4x8_sat = arch >= 9,                                           \
    };
 
 DEFINE_OPTIONS(6);

src/panfrost/compiler/valhall/ISA.xml

@@ -2235,21 +2235,20 @@
     <src absneg="true">Z coordinate as 32-bit floating point</src>
   </ins>
 
-  <group name="IDP" title="8-bit dot product" dests="1" opcode="0xC2" unused="true" unit="FMA">
+  <group name="IDPADD" title="8-bit dot product and accumulate" dests="1" opcode="0xC2" unit="FMA">
     <desc>
       8-bit integer dot product between 4 channel vectors, intended for machine
       learning. Available in both unsigned and signed variants, controlling
       sign-extension/zero-extension behaviour to the final 32-bit destination.
       Saturation is available. Corresponds to the `cl_arm_integer_dot_product_*`
-      family of OpenCL extensions. Not for actual use, just for completeness.
-      Instead, use your platform's neural accelerator.
+      family of OpenCL extensions.
 
       For $A, B \in \{ 0, \ldots, 255 \}^4$ and $\text{Accumulator} \in
       \mathbb{Z}$, calculates $(A \cdot B) + \text{Accumulator}$ and optionally
       saturates.
     </desc>
-    <ins name="IDP.v4s8" opcode2="0"/>
-    <ins name="IDP.v4u8" opcode2="1"/>
+    <ins name="IDPADD.v4s8" opcode2="0"/>
+    <ins name="IDPADD.v4u8" opcode2="1"/>
     <src>A</src>
     <src>B</src>
     <src>Accumulator</src>

src/panfrost/compiler/valhall/valhall.c.py

@@ -88,8 +88,6 @@ SKIP = set([
         "NOT_OLD.i64",
 
         # TODO
-        "IDP.v4s8",
-        "IDP.v4u8",
         "FATAN_ASSIST.f32",
         "SEG_ADD.u64",
         "TEX_DUAL",