nir/vtn: SPIR-V bit count opcodes (core and extension) dest size mismatches nir

[mesa.git] / src / compiler / spirv / vtn_opencl.c

diff --git a/src/compiler/spirv/vtn_opencl.c b/src/compiler/spirv/vtn_opencl.c
index 0c524b61605c93c0cbab6bbb624d51380c2996a6..b5e92dc3b85e1d53ccb321747bf8e74ec2201c73 100644 (file)
--- a/src/compiler/spirv/vtn_opencl.c
+++ b/src/compiler/spirv/vtn_opencl.c
@@ -39,21 +39,18 @@ static void
 handle_instr(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
              const uint32_t *w, unsigned count, nir_handler handler)
 {
-   const struct glsl_type *dest_type =
-      vtn_value(b, w[1], vtn_value_type_type)->type->type;
+   const struct glsl_type *dest_type = vtn_get_type(b, w[1])->type;
 
    unsigned num_srcs = count - 5;
    nir_ssa_def *srcs[3] = { NULL };
    vtn_assert(num_srcs <= ARRAY_SIZE(srcs));
    for (unsigned i = 0; i < num_srcs; i++) {
-      srcs[i] = vtn_ssa_value(b, w[i + 5])->def;
+      srcs[i] = vtn_get_nir_ssa(b, w[i + 5]);
    }
 
    nir_ssa_def *result = handler(b, opcode, num_srcs, srcs, dest_type);
    if (result) {
-      struct vtn_value *val = vtn_push_value(b, w[2], vtn_value_type_ssa);
-      val->ssa = vtn_create_ssa_value(b, dest_type);
-      val->ssa->def = result;
+      vtn_push_nir_ssa(b, w[2], result);
    } else {
       vtn_assert(dest_type == glsl_void_type());
    }
@@ -92,6 +89,7 @@ nir_alu_op_for_opencl_opcode(struct vtn_builder *b,
    case OpenCLstd_Native_recip: return nir_op_frcp;
    case OpenCLstd_Native_rsqrt: return nir_op_frsq;
    case OpenCLstd_Native_sin: return nir_op_fsin;
+   case OpenCLstd_Native_sqrt: return nir_op_fsqrt;
    case OpenCLstd_SMul_hi: return nir_op_imul_high;
    case OpenCLstd_UMul_hi: return nir_op_umul_high;
    case OpenCLstd_Popcount: return nir_op_bit_count;
@@ -119,8 +117,11 @@ handle_alu(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
            unsigned num_srcs, nir_ssa_def **srcs,
            const struct glsl_type *dest_type)
 {
-   return nir_build_alu(&b->nb, nir_alu_op_for_opencl_opcode(b, opcode),
-                        srcs[0], srcs[1], srcs[2], NULL);
+   nir_ssa_def *ret = nir_build_alu(&b->nb, nir_alu_op_for_opencl_opcode(b, opcode),
+                                    srcs[0], srcs[1], srcs[2], NULL);
+   if (opcode == OpenCLstd_Popcount)
+      ret = nir_u2u(&b->nb, ret, glsl_get_bit_size(dest_type));
+   return ret;
 }
 
 static nir_ssa_def *
@@ -224,15 +225,15 @@ _handle_v_load_store(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
 {
    struct vtn_type *type;
    if (load)
-      type = vtn_value(b, w[1], vtn_value_type_type)->type;
+      type = vtn_get_type(b, w[1]);
    else
-      type = vtn_untyped_value(b, w[5])->type;
+      type = vtn_get_value_type(b, w[5]);
    unsigned a = load ? 0 : 1;
 
    const struct glsl_type *dest_type = type->type;
    unsigned components = glsl_get_vector_elements(dest_type);
 
-   nir_ssa_def *offset = vtn_ssa_value(b, w[5 + a])->def;
+   nir_ssa_def *offset = vtn_get_nir_ssa(b, w[5 + a]);
    struct vtn_value *p = vtn_value(b, w[6 + a], vtn_value_type_pointer);
 
    struct vtn_ssa_value *comps[NIR_MAX_VEC_COMPONENTS];
@@ -256,9 +257,7 @@ _handle_v_load_store(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
       }
    }
    if (load) {
-      struct vtn_ssa_value *ssa = vtn_create_ssa_value(b, dest_type);
-      ssa->def = nir_vec(&b->nb, ncomps, components);
-      vtn_push_ssa(b, w[2], type, ssa);
+      vtn_push_nir_ssa(b, w[2], nir_vec(&b->nb, ncomps, components));
    }
 }
 
@@ -285,6 +284,21 @@ handle_printf(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
    return nir_imm_int(&b->nb, -1);
 }
 
+static nir_ssa_def *
+handle_round(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode,
+             unsigned num_srcs, nir_ssa_def **srcs,
+             const struct glsl_type *dest_type)
+{
+   nir_ssa_def *src = srcs[0];
+   nir_builder *nb = &b->nb;
+   nir_ssa_def *half = nir_imm_floatN_t(nb, 0.5, src->bit_size);
+   nir_ssa_def *truncated = nir_ftrunc(nb, src);
+   nir_ssa_def *remainder = nir_fsub(nb, src, truncated);
+
+   return nir_bcsel(nb, nir_fge(nb, nir_fabs(nb, remainder), half),
+                    nir_fadd(nb, truncated, nir_fsign(nb, src)), truncated);
+}
+
 static nir_ssa_def *
 handle_shuffle(struct vtn_builder *b, enum OpenCLstd_Entrypoints opcode, unsigned num_srcs,
                nir_ssa_def **srcs, const struct glsl_type *dest_type)
@@ -366,6 +380,7 @@ vtn_handle_opencl_instruction(struct vtn_builder *b, SpvOp ext_opcode,
    case OpenCLstd_Native_recip:
    case OpenCLstd_Native_rsqrt:
    case OpenCLstd_Native_sin:
+   case OpenCLstd_Native_sqrt:
    case OpenCLstd_Fmod:
    case OpenCLstd_SMul_hi:
    case OpenCLstd_UMul_hi:
@@ -438,6 +453,9 @@ vtn_handle_opencl_instruction(struct vtn_builder *b, SpvOp ext_opcode,
    case OpenCLstd_Shuffle2:
       handle_instr(b, cl_opcode, w, count, handle_shuffle2);
       return true;
+   case OpenCLstd_Round:
+      handle_instr(b, cl_opcode, w, count, handle_round);
+      return true;
    case OpenCLstd_Printf:
       handle_instr(b, cl_opcode, w, count, handle_printf);
       return true;