#include "util/u_math.h"
#include <stdio.h>
+#if UTIL_ARCH_BIG_ENDIAN
+#include <byteswap.h>
+#endif
void
vtn_log(struct vtn_builder *b, enum nir_spirv_debug_level level,
return si;
}
-static char *
+static const char *
vtn_string_literal(struct vtn_builder *b, const uint32_t *words,
unsigned word_count, unsigned *words_used)
{
- char *dup = ralloc_strndup(b, (char *)words, word_count * sizeof(*words));
- if (words_used) {
- /* Ammount of space taken by the string (including the null) */
- unsigned len = strlen(dup) + 1;
- *words_used = DIV_ROUND_UP(len, sizeof(*words));
+ /* From the SPIR-V spec:
+ *
+ * "A string is interpreted as a nul-terminated stream of characters.
+ * The character set is Unicode in the UTF-8 encoding scheme. The UTF-8
+ * octets (8-bit bytes) are packed four per word, following the
+ * little-endian convention (i.e., the first octet is in the
+ * lowest-order 8 bits of the word). The final word contains the
+ * string’s nul-termination character (0), and all contents past the
+ * end of the string in the final word are padded with 0."
+ *
+ * On big-endian, we need to byte-swap.
+ */
+#if UTIL_ARCH_BIG_ENDIAN
+ {
+ uint32_t *copy = ralloc_array(b, uint32_t, word_count);
+ for (unsigned i = 0; i < word_count; i++)
+ copy[i] = bswap_32(words[i]);
+ words = copy;
}
- return dup;
+#endif
+
+ const char *str = (char *)words;
+ const char *end = memchr(str, 0, word_count * 4);
+ vtn_fail_if(end == NULL, "String is not null-terminated");
+
+ if (words_used)
+ *words_used = DIV_ROUND_UP(end - str + 1, sizeof(*words));
+
+ return str;
}
const uint32_t *
vtn_handle_extension(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, unsigned count)
{
- const char *ext = (const char *)&w[2];
switch (opcode) {
case SpvOpExtInstImport: {
struct vtn_value *val = vtn_push_value(b, w[1], vtn_value_type_extension);
+ const char *ext = vtn_string_literal(b, &w[2], count - 2, NULL);
if (strcmp(ext, "GLSL.std.450") == 0) {
val->ext_handler = vtn_handle_glsl450_instruction;
} else if ((strcmp(ext, "SPV_AMD_gcn_shader") == 0)
val->constant = vtn_null_constant(b, val->type);
break;
- case SpvOpConstantSampler:
- vtn_fail("OpConstantSampler requires Kernel Capability");
- break;
-
default:
vtn_fail_with_opcode("Unhandled opcode", opcode);
}
vtn_foreach_decoration(b, val, handle_workgroup_size_decoration_cb, NULL);
}
-SpvMemorySemanticsMask
-vtn_storage_class_to_memory_semantics(SpvStorageClass sc)
-{
- switch (sc) {
- case SpvStorageClassStorageBuffer:
- case SpvStorageClassPhysicalStorageBuffer:
- return SpvMemorySemanticsUniformMemoryMask;
- case SpvStorageClassWorkgroup:
- return SpvMemorySemanticsWorkgroupMemoryMask;
- default:
- return SpvMemorySemanticsMaskNone;
- }
-}
-
static void
vtn_split_barrier_semantics(struct vtn_builder *b,
SpvMemorySemanticsMask semantics,
SpvMemorySemanticsOutputMemoryMask);
const SpvMemorySemanticsMask other_semantics =
- semantics & ~(order_semantics | av_vis_semantics | storage_semantics);
+ semantics & ~(order_semantics | av_vis_semantics | storage_semantics |
+ SpvMemorySemanticsVolatileMask);
if (other_semantics)
vtn_warn("Ignoring unhandled memory semantics: %u\n", other_semantics);
fill_common_atomic_sources(struct vtn_builder *b, SpvOp opcode,
const uint32_t *w, nir_src *src)
{
+ const struct glsl_type *type = vtn_get_type(b, w[1])->type;
+ unsigned bit_size = glsl_get_bit_size(type);
+
switch (opcode) {
case SpvOpAtomicIIncrement:
- src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, 1));
+ src[0] = nir_src_for_ssa(nir_imm_intN_t(&b->nb, 1, bit_size));
break;
case SpvOpAtomicIDecrement:
- src[0] = nir_src_for_ssa(nir_imm_int(&b->nb, -1));
+ src[0] = nir_src_for_ssa(nir_imm_intN_t(&b->nb, -1, bit_size));
break;
case SpvOpAtomicISub:
image.lod = nir_imm_int(&b->nb, 0);
}
- /* TODO: Volatile. */
+ if (operands & SpvImageOperandsVolatileTexelMask)
+ access |= ACCESS_VOLATILE;
break;
}
image.lod = nir_imm_int(&b->nb, 0);
}
- /* TODO: Volatile. */
+ if (operands & SpvImageOperandsVolatileTexelMask)
+ access |= ACCESS_VOLATILE;
break;
}
vtn_fail_with_opcode("Invalid image opcode", opcode);
}
+ if (semantics & SpvMemorySemanticsVolatileMask)
+ access |= ACCESS_VOLATILE;
+
nir_intrinsic_op op;
switch (opcode) {
#define OP(S, N) case SpvOp##S: op = nir_intrinsic_image_deref_##N; break;
SpvScope scope = SpvScopeInvocation;
SpvMemorySemanticsMask semantics = 0;
+ enum gl_access_qualifier access = 0;
switch (opcode) {
case SpvOpAtomicLoad:
vtn_fail_with_opcode("Invalid SPIR-V atomic", opcode);
}
+ if (semantics & SpvMemorySemanticsVolatileMask)
+ access |= ACCESS_VOLATILE;
+
/* uniform as "atomic counter uniform" */
if (ptr->mode == vtn_variable_mode_atomic_counter) {
nir_deref_instr *deref = vtn_pointer_to_deref(b, ptr);
nir_intrinsic_op op = get_ssbo_nir_atomic_op(b, opcode);
atomic = nir_intrinsic_instr_create(b->nb.shader, op);
- nir_intrinsic_set_access(atomic, ACCESS_COHERENT);
+ nir_intrinsic_set_access(atomic, access | ACCESS_COHERENT);
int src = 0;
switch (opcode) {
atomic->src[0] = nir_src_for_ssa(&deref->dest.ssa);
if (ptr->mode != vtn_variable_mode_workgroup)
- nir_intrinsic_set_access(atomic, ACCESS_COHERENT);
+ access |= ACCESS_COHERENT;
+
+ nir_intrinsic_set_access(atomic, access);
switch (opcode) {
case SpvOpAtomicLoad:
/* Atomic ordering operations will implicitly apply to the atomic operation
* storage class, so include that too.
*/
- semantics |= vtn_storage_class_to_memory_semantics(ptr->ptr_type->storage_class);
+ semantics |= vtn_mode_to_memory_semantics(ptr->mode);
SpvMemorySemanticsMask before_semantics;
SpvMemorySemanticsMask after_semantics;
spv_check_supported(kernel_image, cap);
break;
+ case SpvCapabilityLiteralSampler:
+ spv_check_supported(literal_sampler, cap);
+ break;
+
case SpvCapabilityImageReadWrite:
case SpvCapabilityImageMipmap:
case SpvCapabilityPipes:
case SpvCapabilityDeviceEnqueue:
- case SpvCapabilityLiteralSampler:
case SpvCapabilityGenericPointer:
vtn_warn("Unsupported OpenCL-style SPIR-V capability: %s",
spirv_capability_to_string(cap));
assert(nir_address_format_bit_size(b->options->shared_addr_format) == 32);
assert(nir_address_format_num_components(b->options->shared_addr_format) == 1);
if (!b->options->constant_as_global) {
- assert(nir_address_format_bit_size(b->options->ubo_addr_format) == 32);
- assert(nir_address_format_num_components(b->options->ubo_addr_format) == 1);
+ assert(nir_address_format_bit_size(b->options->constant_addr_format) == 32);
+ assert(nir_address_format_num_components(b->options->constant_addr_format) == 1);
}
break;
case SpvAddressingModelPhysical64:
assert(nir_address_format_bit_size(b->options->shared_addr_format) == 64);
assert(nir_address_format_num_components(b->options->shared_addr_format) == 1);
if (!b->options->constant_as_global) {
- assert(nir_address_format_bit_size(b->options->ubo_addr_format) == 64);
- assert(nir_address_format_num_components(b->options->ubo_addr_format) == 1);
+ assert(nir_address_format_bit_size(b->options->constant_addr_format) == 64);
+ assert(nir_address_format_num_components(b->options->constant_addr_format) == 1);
}
break;
case SpvAddressingModelLogical:
case SpvOpConstantFalse:
case SpvOpConstant:
case SpvOpConstantComposite:
- case SpvOpConstantSampler:
case SpvOpConstantNull:
case SpvOpSpecConstantTrue:
case SpvOpSpecConstantFalse:
case SpvOpUndef:
case SpvOpVariable:
+ case SpvOpConstantSampler:
vtn_handle_variables(b, opcode, w, count);
break;
projects / mesa.git / blobdiff