static void
-draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var,
- boolean elts);
+draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *var);
struct draw_gs_llvm_iface {
PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
elem_types[5] = LLVMArrayType(sampler_type,
PIPE_MAX_SAMPLERS); /* samplers */
+ elem_types[6] = LLVMArrayType(LLVMPointerType(int_type, 0), /* vs_ssbo */
+ LP_MAX_TGSI_SHADER_BUFFERS);
+ elem_types[7] = LLVMArrayType(int_type, /* num_vs_ssbos */
+ LP_MAX_TGSI_SHADER_BUFFERS);
context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
ARRAY_SIZE(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, samplers,
target, context_type,
DRAW_JIT_CTX_SAMPLERS);
+ LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_ssbos,
+ target, context_type, DRAW_JIT_CTX_SSBOS);
+ LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_ssbos,
+ target, context_type, DRAW_JIT_CTX_NUM_SSBOS);
LP_CHECK_STRUCT_SIZE(struct draw_jit_context,
target, context_type);
elem_types[8] = LLVMPointerType(LLVMVectorType(int_type,
vector_length), 0);
+ elem_types[9] = LLVMArrayType(LLVMPointerType(int_type, 0), /* ssbos */
+ LP_MAX_TGSI_SHADER_BUFFERS);
+ elem_types[10] = LLVMArrayType(int_type, /* num_ssbos */
+ LP_MAX_TGSI_SHADER_BUFFERS);
context_type = LLVMStructTypeInContext(gallivm->context, elem_types,
ARRAY_SIZE(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, emitted_prims,
target, context_type,
DRAW_GS_JIT_CTX_EMITTED_PRIMS);
+ LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, ssbos,
+ target, context_type, DRAW_GS_JIT_CTX_SSBOS);
+ LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_ssbos,
+ target, context_type, DRAW_GS_JIT_CTX_NUM_SSBOS);
LP_CHECK_STRUCT_SIZE(struct draw_gs_jit_context,
target, context_type);
LLVMTypeRef elem_types[4];
LLVMTypeRef vb_type;
- elem_types[0] =
- elem_types[1] = LLVMInt32TypeInContext(gallivm->context);
- elem_types[2] =
+ elem_types[0] = LLVMInt16TypeInContext(gallivm->context);
+ elem_types[1] = LLVMInt8TypeInContext(gallivm->context);
+ elem_types[2] = LLVMInt32TypeInContext(gallivm->context);
elem_types[3] = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
vb_type = LLVMStructTypeInContext(gallivm->context, elem_types,
(void) target; /* silence unused var warning for non-debug build */
LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, stride,
target, vb_type, 0);
- LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
+ LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, is_user_buffer,
target, vb_type, 1);
+ LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer_offset,
+ target, vb_type, 2);
+ LP_CHECK_MEMBER_OFFSET(struct pipe_vertex_buffer, buffer.resource,
+ target, vb_type, 3);
LP_CHECK_STRUCT_SIZE(struct pipe_vertex_buffer, target, vb_type);
LLVMTypeRef vertex_header;
char struct_name[24];
- util_snprintf(struct_name, 23, "vertex_header%d", data_elems);
+ snprintf(struct_name, 23, "vertex_header%d", data_elems);
elem_types[DRAW_JIT_VERTEX_VERTEX_ID] = LLVMIntTypeInContext(gallivm->context, 32);
elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
variant->llvm = llvm;
variant->shader = shader;
- util_snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
- variant->shader->variants_cached);
+ snprintf(module_name, sizeof(module_name), "draw_llvm_vs_variant%u",
+ variant->shader->variants_cached);
variant->gallivm = gallivm_create(module_name, llvm->context);
variant->vertex_header_ptr_type = LLVMPointerType(vertex_header, 0);
- draw_llvm_generate(llvm, variant, FALSE); /* linear */
- draw_llvm_generate(llvm, variant, TRUE); /* elts */
+ draw_llvm_generate(llvm, variant);
gallivm_compile_module(variant->gallivm);
variant->jit_func = (draw_jit_vert_func)
gallivm_jit_function(variant->gallivm, variant->function);
- variant->jit_func_elts = (draw_jit_vert_func_elts)
- gallivm_jit_function(variant->gallivm, variant->function_elts);
-
gallivm_free_ir(variant->gallivm);
variant->list_item_global.base = variant;
const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS],
const struct lp_bld_tgsi_system_values *system_values,
LLVMValueRef context_ptr,
- struct lp_build_sampler_soa *draw_sampler,
- boolean clamp_vertex_color)
+ const struct lp_build_sampler_soa *draw_sampler,
+ boolean clamp_vertex_color,
+ struct lp_build_mask_context *bld_mask)
{
struct draw_llvm *llvm = variant->llvm;
const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
draw_jit_context_vs_constants(variant->gallivm, context_ptr);
LLVMValueRef num_consts_ptr =
draw_jit_context_num_vs_constants(variant->gallivm, context_ptr);
+ LLVMValueRef ssbos_ptr =
+ draw_jit_context_vs_ssbos(variant->gallivm, context_ptr);
+ LLVMValueRef num_ssbos_ptr =
+ draw_jit_context_num_vs_ssbos(variant->gallivm, context_ptr);
+
+ struct lp_build_tgsi_params params;
+ memset(¶ms, 0, sizeof(params));
+
+ params.type = vs_type;
+ params.mask = bld_mask;
+ params.consts_ptr = consts_ptr;
+ params.const_sizes_ptr = num_consts_ptr;
+ params.system_values = system_values;
+ params.inputs = inputs;
+ params.context_ptr = context_ptr;
+ params.sampler = draw_sampler;
+ params.info = &llvm->draw->vs.vertex_shader->info;
+ params.ssbo_ptr = ssbos_ptr;
+ params.ssbo_sizes_ptr = num_ssbos_ptr;
lp_build_tgsi_soa(variant->gallivm,
tokens,
- vs_type,
- NULL /*struct lp_build_mask_context *mask*/,
- consts_ptr,
- num_consts_ptr,
- system_values,
- inputs,
- outputs,
- context_ptr,
- NULL,
- draw_sampler,
- &llvm->draw->vs.vertex_shader->info,
- NULL);
+ ¶ms,
+ outputs);
{
LLVMValueRef out;
LLVMValueRef vb_stride,
LLVMValueRef map_ptr,
LLVMValueRef buffer_size_adj,
- LLVMValueRef ofbit,
LLVMValueRef *inputs,
LLVMValueRef index)
{
- LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
+ LLVMTypeRef i32_t = LLVMInt32TypeInContext(gallivm->context);
+ LLVMTypeRef aosf_t, aosi_t;
+ LLVMValueRef zero = LLVMConstNull(i32_t);
LLVMBuilderRef builder = gallivm->builder;
- LLVMValueRef stride, buffer_overflowed, aos;
- LLVMValueRef temp_ptr =
- lp_build_alloca(gallivm,
- lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
- struct lp_build_if_state if_ctx;
+ LLVMValueRef stride, buffer_overflowed, aos, index_valid;
unsigned i;
- stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
+ aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
+ aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
+
+ /* This mul can overflow. Wraparound is ok. */
+ stride = LLVMBuildMul(builder, vb_stride, index, "");
- buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
+ buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
stride, buffer_size_adj,
"buffer_overflowed");
- buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
if (0) {
lp_build_print_value(gallivm, " instance index = ", index);
lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
}
- lp_build_if(&if_ctx, gallivm, buffer_overflowed);
- {
- LLVMValueRef val =
- lp_build_const_vec(gallivm, lp_float32_vec4_type(), 0);
- LLVMBuildStore(builder, val, temp_ptr);
- }
- lp_build_else(&if_ctx);
- {
- LLVMValueRef val;
-
- val = lp_build_fetch_rgba_aos(gallivm,
- format_desc,
- lp_float32_vec4_type(),
- FALSE,
- map_ptr,
- stride, zero, zero,
- NULL);
- LLVMBuildStore(builder, val, temp_ptr);
- }
- lp_build_endif(&if_ctx);
+ index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
+ index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
+ stride = LLVMBuildAnd(builder, stride, index_valid, "");
- aos = LLVMBuildLoad(builder, temp_ptr, "aos");
+ aos = lp_build_fetch_rgba_aos(gallivm,
+ format_desc,
+ lp_float32_vec4_type(),
+ FALSE,
+ map_ptr,
+ stride, zero, zero,
+ NULL);
+
+ index_valid = lp_build_broadcast(gallivm, aosi_t, index_valid);
+ aos = LLVMBuildBitCast(builder, aos, aosi_t, "");
+ aos = LLVMBuildAnd(builder, aos, index_valid, "");
+ aos = LLVMBuildBitCast(builder, aos, aosf_t, "");
for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
LLVMValueRef index = lp_build_const_int32(gallivm, i);
}
-static void
-convert_to_soa(struct gallivm_state *gallivm,
- LLVMValueRef src_aos[LP_MAX_VECTOR_WIDTH / 32],
- LLVMValueRef dst_soa[TGSI_NUM_CHANNELS],
- const struct lp_type soa_type)
-{
- unsigned j, k;
- struct lp_type aos_channel_type = soa_type;
-
- LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
- unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
-
- debug_assert(TGSI_NUM_CHANNELS == 4);
- debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
-
- aos_channel_type.length >>= 1;
-
- for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
- LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
-
- assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
-
- for (k = 0; k < pixels_per_channel; ++k) {
- channel[k] = src_aos[j + TGSI_NUM_CHANNELS * k];
- }
-
- aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
- }
-
- lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa);
-}
-
-
static void
fetch_vector(struct gallivm_state *gallivm,
const struct util_format_description *format_desc,
LLVMValueRef vb_stride,
LLVMValueRef map_ptr,
LLVMValueRef buffer_size_adj,
- LLVMValueRef ofmask,
LLVMValueRef *inputs,
LLVMValueRef indices)
{
- LLVMValueRef zero = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
LLVMBuilderRef builder = gallivm->builder;
struct lp_build_context blduivec;
- LLVMValueRef offset, tmp, valid_mask;
- LLVMValueRef aos_fetch[LP_MAX_VECTOR_WIDTH / 32];
+ struct lp_type fetch_type = vs_type;
+ LLVMValueRef offset, valid_mask;
unsigned i;
lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
- /*
- * Sort of interestingly, with interleaved attribs, llvm 3.7+ will
- * recognize these calculations to be constant with different attribs
- * (the different offset has been added to map_ptr).
- * llvm 3.3, however, will not (I can't get llvm 3.4-3.6 to link...)
- *
- * XXX: could actually avoid this altogether (replacing by simple
- * non-widening mul) by precalculating the max index instead outside
- * the loop (at the cost of one scalar udiv per vertex element).
- */
- offset = lp_build_mul_32_lohi_cpu(&blduivec, vb_stride, indices, &tmp);
+ /* This mul can overflow. Wraparound is ok. */
+ offset = lp_build_mul(&blduivec, vb_stride, indices);
- tmp = lp_build_compare(gallivm, blduivec.type,
- PIPE_FUNC_EQUAL, tmp, blduivec.zero);
- valid_mask = lp_build_andnot(&blduivec, tmp, ofmask);
-
- tmp = lp_build_compare(gallivm, blduivec.type,
- PIPE_FUNC_LEQUAL, offset, buffer_size_adj);
- valid_mask = LLVMBuildAnd(builder, tmp, valid_mask, "");
+ valid_mask = lp_build_compare(gallivm, blduivec.type,
+ PIPE_FUNC_LESS, offset, buffer_size_adj);
/* not valid elements use offset 0 */
offset = LLVMBuildAnd(builder, offset, valid_mask, "");
}
/*
- * Note: we probably really want to use SoA fetch, not AoS one (albeit
- * for most formats it will amount to the same as this isn't very
- * optimized). But looks dangerous since it assumes alignment.
+ * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
+ * This should always produce better code.
*/
- for (i = 0; i < vs_type.length; i++) {
- LLVMValueRef offset1, elem;
- elem = lp_build_const_int32(gallivm, i);
- offset1 = LLVMBuildExtractElement(builder, offset, elem, "");
- aos_fetch[i] = lp_build_fetch_rgba_aos(gallivm, format_desc,
- lp_float32_vec4_type(),
- FALSE, map_ptr, offset1,
- zero, zero, NULL);
+ /* The type handling is annoying here... */
+ if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
+ format_desc->channel[0].pure_integer) {
+ if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_SIGNED) {
+ fetch_type = lp_type_int_vec(vs_type.width, vs_type.width * vs_type.length);
+ }
+ else if (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED) {
+ fetch_type = lp_type_uint_vec(vs_type.width, vs_type.width * vs_type.length);
+ }
}
- convert_to_soa(gallivm, aos_fetch, inputs, vs_type);
+ lp_build_fetch_rgba_soa(gallivm, format_desc,
+ fetch_type, FALSE, map_ptr, offset,
+ blduivec.zero, blduivec.zero,
+ NULL, inputs);
+
+ for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
+ inputs[i] = LLVMBuildBitCast(builder, inputs[i],
+ lp_build_vec_type(gallivm, vs_type), "");
+ }
+
+ /* out-of-bound fetches return all zeros */
for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
inputs[i] = LLVMBuildBitCast(builder, inputs[i], blduivec.vec_type, "");
inputs[i] = LLVMBuildAnd(builder, inputs[i], valid_mask, "");
inputs[i] = LLVMBuildBitCast(builder, inputs[i],
lp_build_vec_type(gallivm, vs_type), "");
-
}
}
/**
* Returns boolean if any clipping has occurred
- * Used zero/non-zero i32 value to represent boolean
+ * Used zero/one i8 value to represent boolean
*/
static LLVMValueRef
-clipmask_booli32(struct gallivm_state *gallivm,
- const struct lp_type vs_type,
- LLVMValueRef clipmask_bool_ptr,
- boolean edgeflag_in_clipmask)
+clipmask_booli8(struct gallivm_state *gallivm,
+ const struct lp_type vs_type,
+ LLVMValueRef clipmask_bool_ptr,
+ boolean edgeflag_in_clipmask)
{
LLVMBuilderRef builder = gallivm->builder;
- LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
+ LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
LLVMValueRef clipmask_bool = LLVMBuildLoad(builder, clipmask_bool_ptr, "");
- LLVMValueRef ret = LLVMConstNull(int32_type);
- LLVMValueRef temp;
- int i;
+ LLVMValueRef ret;
+ struct lp_build_context bldivec;
+
+ lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
/*
* We need to invert the edgeflag bit from the clipmask here
* and we (may) need it if edgeflag was 0).
*/
if (edgeflag_in_clipmask) {
- struct lp_type i32_type = lp_int_type(vs_type);
- LLVMValueRef edge = lp_build_const_int_vec(gallivm, i32_type,
+ LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
1LL << DRAW_TOTAL_CLIP_PLANES);
clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
}
+
/*
- * Could do much better with just cmp/movmskps.
+ * XXX: probably should mask off bits from the mask which come from
+ * vertices which were beyond the count (i.e. indices_valid for
+ * linear fetches, for elts ones we don't have the correct mask
+ * right now). Otherwise might run the pipeline for nothing,
+ * though everything should still work.
*/
- for (i=0; i < vs_type.length; i++) {
- temp = LLVMBuildExtractElement(builder, clipmask_bool,
- lp_build_const_int32(gallivm, i) , "");
- ret = LLVMBuildOr(builder, ret, temp, "");
- }
+ ret = lp_build_any_true_range(&bldivec, vs_type.length, clipmask_bool);
+ ret = LLVMBuildZExt(builder, ret, int8_type, "");
return ret;
}
}
static void
-draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant,
- boolean elts)
+draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
{
struct gallivm_state *gallivm = variant->gallivm;
LLVMContextRef context = gallivm->context;
LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
LLVMTypeRef arg_types[11];
- unsigned num_arg_types =
- elts ? ARRAY_SIZE(arg_types) : ARRAY_SIZE(arg_types) - 1;
+ unsigned num_arg_types = ARRAY_SIZE(arg_types);
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
char func_name[64];
struct lp_type vs_type;
- LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
- LLVMValueRef vertex_id_offset, start_instance, start;
+ LLVMValueRef count, fetch_elts, start_or_maxelt;
+ LLVMValueRef vertex_id_offset, start_instance;
LLVMValueRef stride, step, io_itr;
- LLVMValueRef ind_vec;
+ LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
LLVMValueRef buffer_size_adj[PIPE_MAX_ATTRIBS];
- LLVMValueRef ofmask[PIPE_MAX_ATTRIBS];
LLVMValueRef instance_index[PIPE_MAX_ATTRIBS];
LLVMValueRef fake_buf_ptr, fake_buf;
struct draw_context *draw = llvm->draw;
const struct tgsi_shader_info *vs_info = &draw->vs.vertex_shader->info;
unsigned i, j;
- struct lp_build_context bld, bldivec, blduivec;
+ struct lp_build_context bld, blduivec;
struct lp_build_loop_state lp_loop;
+ struct lp_build_if_state if_ctx;
const int vector_length = lp_native_vector_width / 32;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
- LLVMValueRef fetch_max;
struct lp_build_sampler_soa *sampler = 0;
LLVMValueRef ret, clipmask_bool_ptr;
struct draw_llvm_variant_key *key = &variant->key;
memset(&system_values, 0, sizeof(system_values));
- util_snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u_%s",
- variant->shader->variants_cached, elts ? "elts" : "linear");
+ snprintf(func_name, sizeof(func_name), "draw_llvm_vs_variant%u",
+ variant->shader->variants_cached);
i = 0;
arg_types[i++] = get_context_ptr_type(variant); /* context */
arg_types[i++] = get_vertex_header_ptr_type(variant); /* vertex_header */
arg_types[i++] = get_buffer_ptr_type(variant); /* vbuffers */
- if (elts) {
- arg_types[i++] = LLVMPointerType(int32_type, 0);/* fetch_elts */
- arg_types[i++] = int32_type; /* fetch_elt_max */
- } else
- arg_types[i++] = int32_type; /* start */
- arg_types[i++] = int32_type; /* fetch_count / count */
- arg_types[i++] = int32_type; /* stride */
- arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
- arg_types[i++] = int32_type; /* instance_id */
- arg_types[i++] = int32_type; /* vertex_id_offset */
- arg_types[i++] = int32_type; /* start_instance */
-
- func_type = LLVMFunctionType(int32_type, arg_types, num_arg_types, 0);
+ arg_types[i++] = int32_type; /* count */
+ arg_types[i++] = int32_type; /* start/fetch_elt_max */
+ arg_types[i++] = int32_type; /* stride */
+ arg_types[i++] = get_vb_ptr_type(variant); /* pipe_vertex_buffer's */
+ arg_types[i++] = int32_type; /* instance_id */
+ arg_types[i++] = int32_type; /* vertex_id_offset */
+ arg_types[i++] = int32_type; /* start_instance */
+ arg_types[i++] = LLVMPointerType(int32_type, 0); /* fetch_elts */
+
+ func_type = LLVMFunctionType(LLVMInt8TypeInContext(context),
+ arg_types, num_arg_types, 0);
variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
-
- if (elts)
- variant->function_elts = variant_func;
- else
- variant->function = variant_func;
+ variant->function = variant_func;
LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
for (i = 0; i < num_arg_types; ++i)
context_ptr = LLVMGetParam(variant_func, 0);
io_ptr = LLVMGetParam(variant_func, 1);
vbuffers_ptr = LLVMGetParam(variant_func, 2);
+ count = LLVMGetParam(variant_func, 3);
+ /*
+ * XXX: the maxelt part is unused. Not really useful, since we cannot
+ * get index buffer overflows due to vsplit (which provides its own
+ * elts buffer, with a different size than what's passed in here).
+ */
+ start_or_maxelt = LLVMGetParam(variant_func, 4);
/*
* XXX: stride is actually unused. The stride we use is strictly calculated
* from the number of outputs (including the draw_extra outputs).
* Should probably fix some day (we need a new vs just because of extra
* outputs which the generated vs won't touch).
*/
- stride = LLVMGetParam(variant_func, 5 + (elts ? 1 : 0));
- vb_ptr = LLVMGetParam(variant_func, 6 + (elts ? 1 : 0));
- system_values.instance_id = LLVMGetParam(variant_func, 7 + (elts ? 1 : 0));
- vertex_id_offset = LLVMGetParam(variant_func, 8 + (elts ? 1 : 0));
- start_instance = LLVMGetParam(variant_func, 9 + (elts ? 1 : 0));
+ stride = LLVMGetParam(variant_func, 5);
+ vb_ptr = LLVMGetParam(variant_func, 6);
+ system_values.instance_id = LLVMGetParam(variant_func, 7);
+ vertex_id_offset = LLVMGetParam(variant_func, 8);
+ start_instance = LLVMGetParam(variant_func, 9);
+ fetch_elts = LLVMGetParam(variant_func, 10);
lp_build_name(context_ptr, "context");
lp_build_name(io_ptr, "io");
lp_build_name(vbuffers_ptr, "vbuffers");
+ lp_build_name(count, "count");
+ lp_build_name(start_or_maxelt, "start_or_maxelt");
lp_build_name(stride, "stride");
lp_build_name(vb_ptr, "vb");
lp_build_name(system_values.instance_id, "instance_id");
lp_build_name(vertex_id_offset, "vertex_id_offset");
lp_build_name(start_instance, "start_instance");
-
- if (elts) {
- fetch_elts = LLVMGetParam(variant_func, 3);
- fetch_elt_max = LLVMGetParam(variant_func, 4);
- fetch_count = LLVMGetParam(variant_func, 5);
- lp_build_name(fetch_elts, "fetch_elts");
- lp_build_name(fetch_elt_max, "fetch_elt_max");
- lp_build_name(fetch_count, "fetch_count");
- start = count = NULL;
- }
- else {
- start = LLVMGetParam(variant_func, 3);
- count = LLVMGetParam(variant_func, 4);
- lp_build_name(start, "start");
- lp_build_name(count, "count");
- fetch_elts = fetch_count = NULL;
- }
+ lp_build_name(fetch_elts, "fetch_elts");
/*
* Function body
vs_type.length = vector_length;
lp_build_context_init(&bld, gallivm, lp_type_uint(32));
- lp_build_context_init(&bldivec, gallivm, lp_int_type(vs_type));
lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
/* hold temporary "bool" clipmask */
ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
}
+ have_elts = LLVMBuildICmp(builder, LLVMIntNE,
+ LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
- if (elts) {
- fetch_max = fetch_count;
- count = fetch_count;
- start = blduivec.zero;
- }
- else {
- fetch_max = lp_build_add(&bld, start, count);
- start = lp_build_broadcast_scalar(&blduivec, start);
- ind_vec = lp_build_add(&blduivec, start, ind_vec);
- }
-
- fetch_max = LLVMBuildSub(builder, fetch_max, bld.one, "fetch_max");
+ fetch_max = LLVMBuildSub(builder, count, bld.one, "fetch_max");
fetch_max = lp_build_broadcast_scalar(&blduivec, fetch_max);
+ /*
+ * Only needed for non-indexed path.
+ */
+ start_vec = lp_build_broadcast_scalar(&blduivec, start_or_maxelt);
/*
* Pre-calculate everything which is constant per shader invocation.
vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr, &vb_index, 1, "");
vb_info = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
vb_stride[j] = draw_jit_vbuffer_stride(gallivm, vb_info);
+ vb_stride[j] = LLVMBuildZExt(gallivm->builder, vb_stride[j],
+ LLVMInt32TypeInContext(context), "");
vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vb_info);
map_ptr[j] = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
ofbit = NULL;
- buf_offset = lp_build_uadd_overflow(gallivm, vb_buffer_offset,
- src_offset, &ofbit);
- buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, bsize,
+ /*
+ * We'll set buffer_size_adj to zero if we have of, so it will
+ * always overflow later automatically without having to keep ofbit.
+ * Overflows (with normal wraparound) doing the actual offset
+ * calculation should be ok, just not for the buffer size calc.
+ * It would also be possible to detect such overflows and return
+ * zeros if that happens, but this would be more complex.
+ */
+ buf_offset = lp_build_add(&bld, vb_buffer_offset, src_offset);
+ tmp = lp_build_sub(&bld, bsize, bld.one);
+ buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size, tmp,
&ofbit);
buffer_size_adj[j] = lp_build_usub_overflow(gallivm, buffer_size_adj[j],
buf_offset, &ofbit);
* inside the main loop.
* (Alternatively, could have control flow per vector skipping fetch
* if ofbit is true.)
- * For instanced elements, we keep the control flow for now as it's a
- * scalar fetch, making things easier.
*/
if (velem->instance_divisor) {
- /* Index is equal to the start instance plus the number of current
+ /*
+ * Index is equal to the start instance plus the number of current
* instance divided by the divisor. In this case we compute it as:
- * index = start_instance + (instance_id / divisor)
+ * index = start_instance + (instance_id / divisor).
+ * Note we could actually do the fetch here, outside the loop -
+ * it's all constant, hopefully llvm recognizes this.
*/
LLVMValueRef current_instance;
current_instance = LLVMBuildUDiv(builder, system_values.instance_id,
"instance_divisor");
instance_index[j] = lp_build_uadd_overflow(gallivm, start_instance,
current_instance, &ofbit);
- map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
-
- /* This is a scalar fetch, just keep the of bit */
- ofmask[j] = ofbit;
}
- else {
- temp_ptr = lp_build_alloca_undef(gallivm,
- LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
- lp_build_if(&if_ctx, gallivm, ofbit);
- {
- LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
- }
- lp_build_else(&if_ctx);
- {
- map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
- LLVMBuildStore(builder, map_ptr[j], temp_ptr);
- }
- lp_build_endif(&if_ctx);
- map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
+ buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
+ buffer_size_adj[j], "");
+
+ temp_ptr = lp_build_alloca_undef(gallivm,
+ LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
- /* Expand to vector mask */
- ofmask[j] = LLVMBuildSExt(builder, ofbit, int32_type, "");
- ofmask[j] = lp_build_broadcast_scalar(&blduivec, ofmask[j]);
+ lp_build_if(&if_ctx, gallivm, ofbit);
+ {
+ LLVMBuildStore(builder, fake_buf_ptr, temp_ptr);
}
+ lp_build_else(&if_ctx);
+ {
+ map_ptr[j] = LLVMBuildGEP(builder, map_ptr[j], &buf_offset, 1, "");
+ LLVMBuildStore(builder, map_ptr[j], temp_ptr);
+ }
+ lp_build_endif(&if_ctx);
+ map_ptr[j] = LLVMBuildLoad(builder, temp_ptr, "map_ptr");
if (0) {
lp_build_printf(gallivm, "velem %d, vbuf index = %u, vb_stride = %u\n",
LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
LLVMValueRef io;
LLVMValueRef clipmask; /* holds the clipmask value */
- LLVMValueRef true_index_array;
+ LLVMValueRef true_index_array, index_store;
const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
io_itr = lp_loop.counter;
true_index_array = lp_build_broadcast_scalar(&blduivec, lp_loop.counter);
true_index_array = LLVMBuildAdd(builder, true_index_array, ind_vec, "");
+ LLVMValueRef exec_mask = lp_build_cmp(&blduivec, PIPE_FUNC_LEQUAL, true_index_array, fetch_max);
/*
- * XXX: This code is really fishy. We are required to use a int min
- * here, not uint. The reason is that for some non-indexed draws, we
- * might get something like MAX_UINT - 3 as start value (due to start
- * vertex). So, the first 3 elements in the vector are huge, and
- * limiting them to fetch_max is incorrect. By using int min, we'll
- * pick that huge value - we rely on this creating an overflow (which
- * is guaranteed) in the stride mul later (using (signed) cmp and
- * incorporating the result into ofmask would also work).
- * For the later elements, this just wraps around the indices, which
- * is apparently ok...
+ * Limit indices to fetch_max, otherwise might try to access indices
+ * beyond index buffer (or rather vsplit elt buffer) size.
+ * Could probably safely (?) skip this for non-indexed draws and
+ * simplify things minimally (by removing it could combine the ind_vec
+ * and start_vec adds). I think the only effect for non-indexed draws will
+ * be that for the invalid elements they will be all fetched from the
+ * same location as the last valid one, but noone should really care.
*/
- true_index_array = lp_build_min(&bldivec, true_index_array, fetch_max);
+ true_index_array = lp_build_min(&blduivec, true_index_array, fetch_max);
- if (elts) {
+ index_store = lp_build_alloca_undef(gallivm, blduivec.vec_type, "index_store");
+ lp_build_if(&if_ctx, gallivm, have_elts);
+ {
/*
* Note: you'd expect some comparison/clamp against fetch_elt_max
* here.
* not being zero will get a different fetch index than the valid
* index 0. So, just rely on vsplit code preventing out-of-bounds
* fetches. This is also why it's safe to do elts fetch even if there
- * was no index buffer bound - the real buffer is never seen here.
+ * was no index buffer bound - the real buffer is never seen here, at
+ * least not if there are index buffer overflows...
*/
/*
* XXX should not have to do this, as scale can be handled
* natively by loads (hits asserts though).
*/
- true_index_array = lp_build_shl_imm(&blduivec, true_index_array, 2);
+ tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
LLVMPointerType(LLVMInt8TypeInContext(context),
0), "");
- true_index_array = lp_build_gather(gallivm, vs_type.length,
- 32, 32, TRUE,
- fetch_elts, true_index_array,
- FALSE);
+ tmp = lp_build_gather(gallivm, vs_type.length,
+ 32, bld.type, TRUE,
+ fetch_elts, tmp, FALSE);
+ LLVMBuildStore(builder, tmp, index_store);
+ }
+ lp_build_else(&if_ctx);
+ {
+ tmp = LLVMBuildAdd(builder, true_index_array, start_vec, "");
+ LLVMBuildStore(builder, tmp, index_store);
}
+ lp_build_endif(&if_ctx);
+
+ true_index_array = LLVMBuildLoad(builder, index_store, "");
for (j = 0; j < key->nr_vertex_elements; ++j) {
struct pipe_vertex_element *velem = &key->vertex_element[j];
else if (velem->instance_divisor) {
fetch_instanced(gallivm, format_desc, vs_type,
vb_stride[j], map_ptr[j],
- buffer_size_adj[j], ofmask[j],
+ buffer_size_adj[j],
inputs[j], instance_index[j]);
}
else {
fetch_vector(gallivm, format_desc, vs_type,
vb_stride[j], map_ptr[j],
- buffer_size_adj[j], ofmask[j],
+ buffer_size_adj[j],
inputs[j], true_index_array);
}
}
+ struct lp_build_mask_context mask;
+
+ lp_build_mask_begin(&mask, gallivm, vs_type, exec_mask);
/* In the paths with elts vertex id has to be unaffected by the
* index bias and because indices inside our elements array have
* already had index bias applied we need to subtract it here to
&system_values,
context_ptr,
sampler,
- key->clamp_vertex_color);
+ key->clamp_vertex_color,
+ &mask);
+ lp_build_mask_end(&mask);
if (pos != -1 && cv != -1) {
/* store original positions in clip before further manipulation */
store_clip(gallivm, vs_type, io, outputs, pos);
sampler->destroy(sampler);
/* return clipping boolean value for function */
- ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr,
- enable_cliptest && key->need_edgeflags);
+ ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
+ enable_cliptest && key->need_edgeflags);
LLVMBuildRet(builder, ret);
void
draw_llvm_set_mapped_texture(struct draw_context *draw,
- unsigned shader_stage,
+ enum pipe_shader_type shader_stage,
unsigned sview_idx,
uint32_t width, uint32_t height, uint32_t depth,
uint32_t first_level, uint32_t last_level,
void
draw_llvm_set_sampler_state(struct draw_context *draw,
- unsigned shader_type)
+ enum pipe_shader_type shader_type)
{
unsigned i;
{
struct draw_llvm *llvm = variant->llvm;
+ if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
+ debug_printf("Deleting VS variant: %u vs variants,\t%u total variants\n",
+ variant->shader->variants_cached, llvm->nr_variants);
+ }
+
gallivm_destroy(variant->gallivm);
remove_from_list(&variant->list_item_local);
struct draw_gs_llvm_iface gs_iface;
const struct tgsi_token *tokens = variant->shader->base.state.tokens;
LLVMValueRef consts_ptr, num_consts_ptr;
+ LLVMValueRef ssbos_ptr, num_ssbos_ptr;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
struct lp_build_mask_context mask;
const struct tgsi_shader_info *gs_info = &variant->shader->base.info;
memset(&system_values, 0, sizeof(system_values));
- util_snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
- variant->shader->variants_cached);
+ snprintf(func_name, sizeof(func_name), "draw_llvm_gs_variant%u",
+ variant->shader->variants_cached);
assert(variant->vertex_header_ptr_type);
num_consts_ptr =
draw_gs_jit_context_num_constants(variant->gallivm, context_ptr);
+ ssbos_ptr = draw_gs_jit_context_ssbos(variant->gallivm, context_ptr);
+ num_ssbos_ptr =
+ draw_gs_jit_context_num_ssbos(variant->gallivm, context_ptr);
+
/* code generated texture sampling */
sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
draw_gs_llvm_dump_variant_key(&variant->key);
}
+ struct lp_build_tgsi_params params;
+ memset(¶ms, 0, sizeof(params));
+
+ params.type = gs_type;
+ params.mask = &mask;
+ params.consts_ptr = consts_ptr;
+ params.const_sizes_ptr = num_consts_ptr;
+ params.system_values = &system_values;
+ params.context_ptr = context_ptr;
+ params.sampler = sampler;
+ params.info = &llvm->draw->gs.geometry_shader->info;
+ params.gs_iface = (const struct lp_build_tgsi_gs_iface *)&gs_iface;
+ params.ssbo_ptr = ssbos_ptr;
+ params.ssbo_sizes_ptr = num_ssbos_ptr;
+
lp_build_tgsi_soa(variant->gallivm,
tokens,
- gs_type,
- &mask,
- consts_ptr,
- num_consts_ptr,
- &system_values,
- NULL,
- outputs,
- context_ptr,
- NULL,
- sampler,
- &llvm->draw->gs.geometry_shader->info,
- (const struct lp_build_tgsi_gs_iface *)&gs_iface);
+ ¶ms,
+ outputs);
sampler->destroy(sampler);
variant->llvm = llvm;
variant->shader = shader;
- util_snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
- variant->shader->variants_cached);
+ snprintf(module_name, sizeof(module_name), "draw_llvm_gs_variant%u",
+ variant->shader->variants_cached);
variant->gallivm = gallivm_create(module_name, llvm->context);
{
struct draw_llvm *llvm = variant->llvm;
+ if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
+ debug_printf("Deleting GS variant: %u gs variants,\t%u total variants\n",
+ variant->shader->variants_cached, llvm->nr_gs_variants);
+ }
+
gallivm_destroy(variant->gallivm);
remove_from_list(&variant->list_item_local);
projects / mesa.git / blobdiff