#include "gallivm/lp_bld_arit.h"
#include "gallivm/lp_bld_arit_overflow.h"
+#include "gallivm/lp_bld_bitarit.h"
+#include "gallivm/lp_bld_gather.h"
#include "gallivm/lp_bld_logic.h"
#include "gallivm/lp_bld_const.h"
#include "gallivm/lp_bld_swizzle.h"
#include "util/u_math.h"
#include "util/u_pointer.h"
#include "util/u_string.h"
-#include "util/u_simple_list.h"
+#include "util/simple_list.h"
#define DEBUG_STORE 0
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 {
LLVMValueRef input;
};
-static INLINE const struct draw_gs_llvm_iface *
+static inline const struct draw_gs_llvm_iface *
draw_gs_llvm_iface(const struct lp_build_tgsi_gs_iface *iface)
{
return (const struct draw_gs_llvm_iface *)iface;
elem_types[DRAW_JIT_DVBUFFER_SIZE] = int32_type;
dvbuffer_type = LLVMStructTypeInContext(gallivm->context, elem_types,
- Elements(elem_types), 0);
+ ARRAY_SIZE(elem_types), 0);
+ (void) target; /* silence unused var warning for non-debug build */
LP_CHECK_MEMBER_OFFSET(struct draw_vertex_buffer, map,
target, dvbuffer_type,
DRAW_JIT_DVBUFFER_MAP);
LLVMArrayType(int32_type, PIPE_MAX_TEXTURE_LEVELS);
texture_type = LLVMStructTypeInContext(gallivm->context, elem_types,
- Elements(elem_types), 0);
+ ARRAY_SIZE(elem_types), 0);
+ (void) target; /* silence unused var warning for non-debug build */
LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, width,
target, texture_type,
DRAW_JIT_TEXTURE_WIDTH);
LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
sampler_type = LLVMStructTypeInContext(gallivm->context, elem_types,
- Elements(elem_types), 0);
+ ARRAY_SIZE(elem_types), 0);
+ (void) target; /* silence unused var warning for non-debug build */
LP_CHECK_MEMBER_OFFSET(struct draw_jit_sampler, min_lod,
target, sampler_type,
DRAW_JIT_SAMPLER_MIN_LOD);
LP_MAX_TGSI_CONST_BUFFERS);
elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
DRAW_TOTAL_CLIP_PLANES), 0);
- elem_types[3] = LLVMPointerType(float_type, 0); /* viewport */
+ elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
elem_types[4] = LLVMArrayType(texture_type,
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,
- Elements(elem_types), 0);
+ ARRAY_SIZE(elem_types), 0);
+
+ (void) target; /* silence unused var warning for non-debug build */
LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, vs_constants,
target, context_type, DRAW_JIT_CTX_CONSTANTS);
LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, num_vs_constants,
target, context_type, DRAW_JIT_CTX_NUM_CONSTANTS);
LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, planes,
target, context_type, DRAW_JIT_CTX_PLANES);
- LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewport,
+ LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, viewports,
target, context_type, DRAW_JIT_CTX_VIEWPORT);
LP_CHECK_MEMBER_OFFSET(struct draw_jit_context, textures,
target, context_type,
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);
LP_MAX_TGSI_CONST_BUFFERS);
elem_types[2] = LLVMPointerType(LLVMArrayType(LLVMArrayType(float_type, 4),
DRAW_TOTAL_CLIP_PLANES), 0);
- elem_types[3] = LLVMPointerType(float_type, 0); /* viewport */
+ elem_types[3] = LLVMPointerType(float_type, 0); /* viewports */
elem_types[4] = LLVMArrayType(texture_type,
PIPE_MAX_SHADER_SAMPLER_VIEWS); /* textures */
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,
- Elements(elem_types), 0);
+ ARRAY_SIZE(elem_types), 0);
+ (void) target; /* silence unused var warning for non-debug build */
LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, constants,
target, context_type, DRAW_GS_JIT_CTX_CONSTANTS);
LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, num_constants,
target, context_type, DRAW_GS_JIT_CTX_NUM_CONSTANTS);
LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, planes,
target, context_type, DRAW_GS_JIT_CTX_PLANES);
- LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewport,
+ LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, viewports,
target, context_type, DRAW_GS_JIT_CTX_VIEWPORT);
LP_CHECK_MEMBER_OFFSET(struct draw_gs_jit_context, textures,
target, context_type,
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,
- Elements(elem_types), 0);
+ ARRAY_SIZE(elem_types), 0);
+ (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);
create_jit_vertex_header(struct gallivm_state *gallivm, int data_elems)
{
LLVMTargetDataRef target = gallivm->target;
- LLVMTypeRef elem_types[4];
+ LLVMTypeRef elem_types[3];
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] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
- elem_types[DRAW_JIT_VERTEX_PRE_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
+ elem_types[DRAW_JIT_VERTEX_CLIP_POS] = LLVMArrayType(LLVMFloatTypeInContext(gallivm->context), 4);
elem_types[DRAW_JIT_VERTEX_DATA] = LLVMArrayType(elem_types[1], data_elems);
vertex_header = LLVMStructTypeInContext(gallivm->context, elem_types,
- Elements(elem_types), 0);
+ ARRAY_SIZE(elem_types), 0);
/* these are bit-fields and we can't take address of them
LP_CHECK_MEMBER_OFFSET(struct vertex_header, clipmask,
target, vertex_header,
DRAW_JIT_VERTEX_VERTEX_ID);
*/
- LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip,
- target, vertex_header,
- DRAW_JIT_VERTEX_CLIP);
- LP_CHECK_MEMBER_OFFSET(struct vertex_header, pre_clip_pos,
+ (void) target; /* silence unused var warning for non-debug build */
+ LP_CHECK_MEMBER_OFFSET(struct vertex_header, clip_pos,
target, vertex_header,
- DRAW_JIT_VERTEX_PRE_CLIP_POS);
+ DRAW_JIT_VERTEX_CLIP_POS);
LP_CHECK_MEMBER_OFFSET(struct vertex_header, data,
target, vertex_header,
DRAW_JIT_VERTEX_DATA);
variant = MALLOC(sizeof *variant +
shader->variant_key_size -
sizeof variant->key);
- if (variant == NULL)
+ if (!variant)
return NULL;
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);
memcpy(&variant->key, key, shader->variant_key_size);
+ if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
+ tgsi_dump(llvm->draw->vs.vertex_shader->state.tokens, 0);
+ draw_llvm_dump_variant_key(&variant->key);
+ }
+
vertex_header = create_jit_vertex_header(variant->gallivm, num_inputs);
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);
-
- if (gallivm_debug & (GALLIVM_DEBUG_TGSI | GALLIVM_DEBUG_IR)) {
- tgsi_dump(tokens, 0);
- draw_llvm_dump_variant_key(&variant->key);
- }
+ 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,
- draw_sampler,
- &llvm->draw->vs.vertex_shader->info,
- NULL);
+ ¶ms,
+ outputs);
{
LLVMValueRef out;
}
}
+
static void
-generate_fetch(struct gallivm_state *gallivm,
- struct draw_context *draw,
- LLVMValueRef vbuffers_ptr,
- LLVMValueRef *res,
- struct pipe_vertex_element *velem,
- LLVMValueRef vbuf,
- LLVMValueRef index,
- LLVMValueRef instance_id,
- LLVMValueRef start_instance)
+fetch_instanced(struct gallivm_state *gallivm,
+ const struct util_format_description *format_desc,
+ struct lp_type vs_type,
+ LLVMValueRef vb_stride,
+ LLVMValueRef map_ptr,
+ LLVMValueRef buffer_size_adj,
+ LLVMValueRef *inputs,
+ LLVMValueRef index)
{
- const struct util_format_description *format_desc =
- util_format_description(velem->src_format);
- 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 indices =
- LLVMConstInt(LLVMInt64TypeInContext(gallivm->context),
- velem->vertex_buffer_index, 0);
- LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
- &indices, 1, "");
- LLVMValueRef vb_stride = draw_jit_vbuffer_stride(gallivm, vbuf);
- LLVMValueRef vb_buffer_offset = draw_jit_vbuffer_offset(gallivm, vbuf);
- LLVMValueRef map_ptr = draw_jit_dvbuffer_map(gallivm, vbuffer_ptr);
- LLVMValueRef buffer_size = draw_jit_dvbuffer_size(gallivm, vbuffer_ptr);
- LLVMValueRef stride;
- LLVMValueRef buffer_overflowed;
- LLVMValueRef needed_buffer_size;
- LLVMValueRef temp_ptr =
- lp_build_alloca(gallivm,
- lp_build_vec_type(gallivm, lp_float32_vec4_type()), "");
- LLVMValueRef ofbit = NULL;
- struct lp_build_if_state if_ctx;
+ LLVMValueRef stride, buffer_overflowed, aos, index_valid;
+ unsigned i;
- if (velem->instance_divisor) {
- /* 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)
- */
- LLVMValueRef current_instance;
- current_instance = LLVMBuildUDiv(builder, instance_id,
- lp_build_const_int32(gallivm, velem->instance_divisor),
- "instance_divisor");
- index = lp_build_uadd_overflow(gallivm, start_instance,
- current_instance, &ofbit);
- }
+ aosf_t = lp_build_vec_type(gallivm, lp_float32_vec4_type());
+ aosi_t = lp_build_vec_type(gallivm, lp_int32_vec4_type());
- stride = lp_build_umul_overflow(gallivm, vb_stride, index, &ofbit);
- stride = lp_build_uadd_overflow(gallivm, stride, vb_buffer_offset, &ofbit);
- stride = lp_build_uadd_overflow(
- gallivm, stride,
- lp_build_const_int32(gallivm, velem->src_offset), &ofbit);
- needed_buffer_size = lp_build_uadd_overflow(
- gallivm, stride,
- lp_build_const_int32(gallivm,
- util_format_get_blocksize(velem->src_format)),
- &ofbit);
-
- buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
- needed_buffer_size, buffer_size,
+ /* This mul can overflow. Wraparound is ok. */
+ stride = LLVMBuildMul(builder, vb_stride, index, "");
+
+ buffer_overflowed = LLVMBuildICmp(builder, LLVMIntUGE,
+ stride, buffer_size_adj,
"buffer_overflowed");
- buffer_overflowed = LLVMBuildOr(builder, buffer_overflowed, ofbit, "");
-#if 0
- lp_build_printf(gallivm, "vbuf index = %u, vb_stride is %u\n",
- index, vb_stride);
- lp_build_printf(gallivm, " vb_buffer_offset = %u, src_offset is %u\n",
- vb_buffer_offset,
- lp_build_const_int32(gallivm, velem->src_offset));
- lp_build_print_value(gallivm, " blocksize = ",
- lp_build_const_int32(
- gallivm,
- util_format_get_blocksize(velem->src_format)));
- lp_build_printf(gallivm, " instance_id = %u\n", instance_id);
- lp_build_printf(gallivm, " stride = %u\n", stride);
- lp_build_printf(gallivm, " buffer size = %u\n", buffer_size);
- lp_build_printf(gallivm, " needed_buffer_size = %u\n", needed_buffer_size);
- lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
-#endif
- 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;
- map_ptr = LLVMBuildGEP(builder, map_ptr, &stride, 1, "");
-
- val = lp_build_fetch_rgba_aos(gallivm,
- format_desc,
- lp_float32_vec4_type(),
- map_ptr,
- zero, zero, zero);
- LLVMBuildStore(builder, val, temp_ptr);
+ if (0) {
+ lp_build_print_value(gallivm, " instance index = ", index);
+ lp_build_print_value(gallivm, " buffer overflowed = ", buffer_overflowed);
}
- lp_build_endif(&if_ctx);
- *res = LLVMBuildLoad(builder, temp_ptr, "aos");
+ index_valid = LLVMBuildNot(builder, buffer_overflowed, "");
+ index_valid = LLVMBuildSExt(builder, index_valid, i32_t, "");
+ stride = LLVMBuildAnd(builder, stride, index_valid, "");
+
+ 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);
+ inputs[i] = lp_build_extract_broadcast(gallivm,
+ lp_float32_vec4_type(),
+ vs_type, aos, index);
+ }
}
+
static void
-convert_to_soa(struct gallivm_state *gallivm,
- LLVMValueRef (*src_aos)[LP_MAX_VECTOR_WIDTH / 32],
- LLVMValueRef (*dst_soa)[TGSI_NUM_CHANNELS],
- unsigned num_attribs, const struct lp_type soa_type)
+fetch_vector(struct gallivm_state *gallivm,
+ const struct util_format_description *format_desc,
+ struct lp_type vs_type,
+ LLVMValueRef vb_stride,
+ LLVMValueRef map_ptr,
+ LLVMValueRef buffer_size_adj,
+ LLVMValueRef *inputs,
+ LLVMValueRef indices)
{
- unsigned i, j, k;
- struct lp_type aos_channel_type = soa_type;
+ LLVMBuilderRef builder = gallivm->builder;
+ struct lp_build_context blduivec;
+ struct lp_type fetch_type = vs_type;
+ LLVMValueRef offset, valid_mask;
+ unsigned i;
- debug_assert(TGSI_NUM_CHANNELS == 4);
- debug_assert((soa_type.length % TGSI_NUM_CHANNELS) == 0);
+ lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
- aos_channel_type.length >>= 1;
+ vb_stride = lp_build_broadcast_scalar(&blduivec, vb_stride);
+ buffer_size_adj = lp_build_broadcast_scalar(&blduivec, buffer_size_adj);
- for (i = 0; i < num_attribs; ++i) {
- LLVMValueRef aos_channels[TGSI_NUM_CHANNELS];
- unsigned pixels_per_channel = soa_type.length / TGSI_NUM_CHANNELS;
+ /* This mul can overflow. Wraparound is ok. */
+ offset = lp_build_mul(&blduivec, vb_stride, indices);
- for (j = 0; j < TGSI_NUM_CHANNELS; ++j) {
- LLVMValueRef channel[LP_MAX_VECTOR_LENGTH] = { 0 };
+ valid_mask = lp_build_compare(gallivm, blduivec.type,
+ PIPE_FUNC_LESS, offset, buffer_size_adj);
- assert(pixels_per_channel <= LP_MAX_VECTOR_LENGTH);
+ /* not valid elements use offset 0 */
+ offset = LLVMBuildAnd(builder, offset, valid_mask, "");
- for (k = 0; k < pixels_per_channel; ++k) {
- channel[k] = src_aos[i][j + TGSI_NUM_CHANNELS * k];
- }
+ if (0) {
+ lp_build_print_value(gallivm, " indices = ", indices);
+ lp_build_print_value(gallivm, " offsets = ", offset);
+ lp_build_print_value(gallivm, " valid_mask = ", valid_mask);
+ }
+
+ /*
+ * Unlike fetch_instanced, use SoA fetch instead of multiple AoS fetches.
+ * This should always produce better code.
+ */
- aos_channels[j] = lp_build_concat(gallivm, channel, aos_channel_type, pixels_per_channel);
+ /* 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);
+ }
+ }
- lp_build_transpose_aos(gallivm, soa_type, aos_channels, dst_soa[i]);
+ 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), "");
}
}
#endif
/* Unaligned store due to the vertex header */
- lp_set_store_alignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
+ LLVMSetAlignment(LLVMBuildStore(builder, value, data_ptr), sizeof(float));
}
/**
* struct vertex_header {
* unsigned clipmask:DRAW_TOTAL_CLIP_PLANES;
* unsigned edgeflag:1;
- * unsigned have_clipdist:1;
+ * unsigned pad:1;
* unsigned vertex_id:16;
* [...]
* }
* {
* return (x >> 16) | // vertex_id
* ((x & 0x3fff) << 18) | // clipmask
- * ((x & 0x4000) << 3) | // have_clipdist
+ * ((x & 0x4000) << 3) | // pad
* ((x & 0x8000) << 1); // edgeflag
* }
*/
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef vertex_id;
LLVMValueRef clipmask;
- LLVMValueRef have_clipdist;
+ LLVMValueRef pad;
LLVMValueRef edgeflag;
vertex_id = LLVMBuildLShr(builder, mask, lp_build_const_int32(gallivm, 16), "");
clipmask = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x3fff), "");
clipmask = LLVMBuildShl(builder, clipmask, lp_build_const_int32(gallivm, 18), "");
- have_clipdist = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
- have_clipdist = LLVMBuildShl(builder, have_clipdist, lp_build_const_int32(gallivm, 3), "");
+ if (0) {
+ pad = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x4000), "");
+ pad = LLVMBuildShl(builder, pad, lp_build_const_int32(gallivm, 3), "");
+ }
edgeflag = LLVMBuildAnd(builder, mask, lp_build_const_int32(gallivm, 0x8000), "");
edgeflag = LLVMBuildShl(builder, edgeflag, lp_build_const_int32(gallivm, 1), "");
mask = LLVMBuildOr(builder, vertex_id, clipmask, "");
- mask = LLVMBuildOr(builder, mask, have_clipdist, "");
+ if (0) {
+ mask = LLVMBuildOr(builder, mask, pad, "");
+ }
mask = LLVMBuildOr(builder, mask, edgeflag, "");
#endif
return mask;
int attrib,
int num_outputs,
LLVMValueRef clipmask,
- boolean have_clipdist)
+ boolean need_edgeflag)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef attr_index = lp_build_const_int32(gallivm, attrib);
* code here. See struct vertex_header in draw_private.h.
*/
assert(DRAW_TOTAL_CLIP_PLANES==14);
- /* initialize vertex id:16 = 0xffff, have_clipdist:1 = 0, edgeflag:1 = 1 */
- vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
- if (have_clipdist)
- vertex_id_pad_edgeflag |= 1 << (DRAW_TOTAL_CLIP_PLANES+1);
- val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type), vertex_id_pad_edgeflag);
+ /* initialize vertex id:16 = 0xffff, pad:1 = 0, edgeflag:1 = 1 */
+ if (!need_edgeflag) {
+ vertex_id_pad_edgeflag = (0xffff << 16) | (1 << DRAW_TOTAL_CLIP_PLANES);
+ }
+ else {
+ vertex_id_pad_edgeflag = (0xffff << 16);
+ }
+ val = lp_build_const_int_vec(gallivm, lp_int_type(soa_type),
+ vertex_id_pad_edgeflag);
/* OR with the clipmask */
cliptmp = LLVMBuildOr(builder, val, clipmask, "");
for (i = 0; i < vector_length; i++) {
LLVMValueRef clipmask,
int num_outputs,
struct lp_type soa_type,
- boolean have_clipdist)
+ boolean need_edgeflag)
{
LLVMBuilderRef builder = gallivm->builder;
unsigned chan, attrib, i;
aos,
attrib,
num_outputs,
- clipmask, have_clipdist);
+ clipmask,
+ need_edgeflag);
}
#if DEBUG_STORE
lp_build_printf(gallivm, " # storing end\n");
const struct lp_type vs_type,
LLVMValueRef io_ptr,
LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
- boolean pre_clip_pos, int idx)
+ int idx)
{
LLVMBuilderRef builder = gallivm->builder;
LLVMValueRef soa[4];
soa[2] = LLVMBuildLoad(builder, outputs[idx][2], ""); /*z0 z1 .. zn*/
soa[3] = LLVMBuildLoad(builder, outputs[idx][3], ""); /*w0 w1 .. wn*/
- if (!pre_clip_pos) {
- for (i = 0; i < vs_type.length; i++) {
- clip_ptrs[i] = draw_jit_header_clip(gallivm, io_ptrs[i]);
- }
- } else {
- for (i = 0; i < vs_type.length; i++) {
- clip_ptrs[i] = draw_jit_header_pre_clip_pos(gallivm, io_ptrs[i]);
- }
+ for (i = 0; i < vs_type.length; i++) {
+ clip_ptrs[i] = draw_jit_header_clip_pos(gallivm, io_ptrs[i]);
}
lp_build_transpose_aos(gallivm, vs_type, soa, soa);
clip_ptr = LLVMBuildPointerCast(builder, clip_ptr, clip_ptr_type, "");
/* Unaligned store */
- lp_set_store_alignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
+ LLVMSetAlignment(LLVMBuildStore(builder, aos[j], clip_ptr), sizeof(float));
}
}
LLVMTypeRef vs_type_llvm = lp_build_vec_type(gallivm, vs_type);
LLVMValueRef out3 = LLVMBuildLoad(builder, outputs[pos][3], ""); /*w0 w1 .. wn*/
LLVMValueRef const1 = lp_build_const_vec(gallivm, f32_type, 1.0); /*1.0 1.0 1.0 1.0*/
- LLVMValueRef vp_ptr = draw_jit_context_viewport(gallivm, context_ptr);
+ LLVMValueRef vp_ptr = draw_jit_context_viewports(gallivm, context_ptr);
+
+ /* We treat pipe_viewport_state as a float array */
+ const int scale_index_offset = offsetof(struct pipe_viewport_state, scale) / sizeof(float);
+ const int trans_index_offset = offsetof(struct pipe_viewport_state, translate) / sizeof(float);
/* for 1/w convention*/
out3 = LLVMBuildFDiv(builder, const1, out3, "");
LLVMValueRef trans_i;
LLVMValueRef index;
- index = lp_build_const_int32(gallivm, i);
+ index = lp_build_const_int32(gallivm, i + scale_index_offset);
scale_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
- index = lp_build_const_int32(gallivm, i+4);
+ index = lp_build_const_int32(gallivm, i + trans_index_offset);
trans_i = LLVMBuildGEP(builder, vp_ptr, &index, 1, "");
scale = lp_build_broadcast(gallivm, vs_type_llvm,
/* divide by w */
out = LLVMBuildFMul(builder, out, out3, "");
- /* mult by scale */
- out = LLVMBuildFMul(builder, out, scale, "");
- /* add translation */
- out = LLVMBuildFAdd(builder, out, trans, "");
+ /* mult by scale, add translation */
+ out = lp_build_fmuladd(builder, out, scale, trans);
/* store transformed outputs */
LLVMBuildStore(builder, out, outputs[pos][i]);
struct gallivm_state *gallivm,
struct lp_type vs_type,
LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS],
- boolean clip_xy,
- boolean clip_z,
- boolean clip_user,
- boolean clip_halfz,
- unsigned ucp_enable,
+ struct draw_llvm_variant_key *key,
LLVMValueRef context_ptr,
boolean *have_clipdist)
{
const unsigned pos = llvm->draw->vs.position_output;
const unsigned cv = llvm->draw->vs.clipvertex_output;
int num_written_clipdistance = llvm->draw->vs.vertex_shader->info.num_written_clipdistance;
- bool have_cd = false;
+ boolean have_cd = false;
+ boolean clip_user = key->clip_user;
+ unsigned ucp_enable = key->ucp_enable;
unsigned cd[2];
- cd[0] = llvm->draw->vs.clipdistance_output[0];
- cd[1] = llvm->draw->vs.clipdistance_output[1];
+ cd[0] = llvm->draw->vs.ccdistance_output[0];
+ cd[1] = llvm->draw->vs.ccdistance_output[1];
if (cd[0] != pos || cd[1] != pos)
have_cd = true;
cv_w = pos_w;
}
+ /*
+ * Be careful with the comparisons and NaNs (using llvm's unordered
+ * comparisons here).
+ */
/* Cliptest, for hardwired planes */
- if (clip_xy) {
+ /*
+ * XXX should take guardband into account (currently not in key).
+ * Otherwise might run the draw pipeline stages for nothing.
+ */
+ if (key->clip_xy) {
/* plane 1 */
test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, pos_x , pos_w);
temp = shift;
mask = LLVMBuildOr(builder, mask, test, "");
}
- if (clip_z) {
+ if (key->clip_z) {
temp = lp_build_const_int_vec(gallivm, i32_type, 16);
- if (clip_halfz) {
+ if (key->clip_halfz) {
/* plane 5 */
test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, pos_z);
test = LLVMBuildAnd(builder, test, temp, "");
test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, clipdist);
is_nan_or_inf = lp_build_is_inf_or_nan(gallivm, vs_type, clipdist);
test = LLVMBuildOr(builder, test, is_nan_or_inf, "");
- temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
+ temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
test = LLVMBuildAnd(builder, test, temp, "");
mask = LLVMBuildOr(builder, mask, test, "");
} else {
plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_y");
planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
- test = LLVMBuildFMul(builder, planes, cv_y, "");
- sum = LLVMBuildFAdd(builder, sum, test, "");
+ sum = lp_build_fmuladd(builder, planes, cv_y, sum);
indices[2] = lp_build_const_int32(gallivm, 2);
plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_z");
planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
- test = LLVMBuildFMul(builder, planes, cv_z, "");
- sum = LLVMBuildFAdd(builder, sum, test, "");
+ sum = lp_build_fmuladd(builder, planes, cv_z, sum);
indices[2] = lp_build_const_int32(gallivm, 3);
plane_ptr = LLVMBuildGEP(builder, planes_ptr, indices, 3, "");
plane1 = LLVMBuildLoad(builder, plane_ptr, "plane_w");
planes = lp_build_broadcast(gallivm, vs_type_llvm, plane1);
- test = LLVMBuildFMul(builder, planes, cv_w, "");
- sum = LLVMBuildFAdd(builder, sum, test, "");
+ sum = lp_build_fmuladd(builder, planes, cv_w, sum);
test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_GREATER, zero, sum);
- temp = lp_build_const_int_vec(gallivm, i32_type, 1 << plane_idx);
+ temp = lp_build_const_int_vec(gallivm, i32_type, 1LL << plane_idx);
test = LLVMBuildAnd(builder, test, temp, "");
mask = LLVMBuildOr(builder, mask, test, "");
}
}
}
+ if (key->need_edgeflags) {
+ /*
+ * This isn't really part of clipmask but stored the same in vertex
+ * header later, so do it here.
+ */
+ unsigned edge_attr = llvm->draw->vs.edgeflag_output;
+ LLVMValueRef one = lp_build_const_vec(gallivm, f32_type, 1.0);
+ LLVMValueRef edgeflag = LLVMBuildLoad(builder, outputs[edge_attr][0], "");
+ test = lp_build_compare(gallivm, f32_type, PIPE_FUNC_EQUAL, one, edgeflag);
+ temp = lp_build_const_int_vec(gallivm, i32_type,
+ 1LL << DRAW_TOTAL_CLIP_PLANES);
+ test = LLVMBuildAnd(builder, test, temp, "");
+ mask = LLVMBuildOr(builder, mask, test, "");
+ }
return mask;
}
/**
* 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)
+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));
/*
- * Can do this with log2(vector length) pack instructions and one extract
- * (as we don't actually need a or) with sse2 which would be way better.
+ * We need to invert the edgeflag bit from the clipmask here
+ * (because the result is really if we want to run the pipeline or not
+ * and we (may) need it if edgeflag was 0).
*/
- for (i=0; i < vs_type.length; i++) {
- temp = LLVMBuildExtractElement(builder, clipmask_bool,
- lp_build_const_int32(gallivm, i) , "");
- ret = LLVMBuildOr(builder, ret, temp, "");
+ if (edgeflag_in_clipmask) {
+ LLVMValueRef edge = lp_build_const_int_vec(gallivm, bldivec.type,
+ 1LL << DRAW_TOTAL_CLIP_PLANES);
+ clipmask_bool = LLVMBuildXor(builder, clipmask_bool, edge, "");
}
+
+ /*
+ * 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.
+ */
+ 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 ? Elements(arg_types) : Elements(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 end, start;
- LLVMValueRef count, fetch_elts, fetch_elt_max, fetch_count;
+ LLVMValueRef count, fetch_elts, start_or_maxelt;
LLVMValueRef vertex_id_offset, start_instance;
LLVMValueRef stride, step, io_itr;
+ LLVMValueRef ind_vec, start_vec, have_elts, fetch_max, tmp;
LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
- LLVMValueRef zero = lp_build_const_int32(gallivm, 0);
- LLVMValueRef one = lp_build_const_int32(gallivm, 1);
+ LLVMValueRef vb_stride[PIPE_MAX_ATTRIBS];
+ LLVMValueRef map_ptr[PIPE_MAX_ATTRIBS];
+ LLVMValueRef buffer_size_adj[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;
+ 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;
/* If geometry shader is present we need to skip both the viewport
* transformation and clipping otherwise the inputs to the geometry
* shader will be incorrect.
+ * The code can't handle vp transform when vs writes vp index neither
+ * (though this would be fixable here, but couldn't just broadcast
+ * the values).
*/
- const boolean bypass_viewport = key->has_gs || key->bypass_viewport;
+ const boolean bypass_viewport = key->has_gs || key->bypass_viewport ||
+ vs_info->writes_viewport_index;
const boolean enable_cliptest = !key->has_gs && (key->clip_xy ||
- key->clip_z ||
- key->clip_user);
+ key->clip_z ||
+ key->clip_user ||
+ key->need_edgeflags);
LLVMValueRef variant_func;
- const unsigned pos = llvm->draw->vs.position_output;
- const unsigned cv = llvm->draw->vs.clipvertex_output;
+ const unsigned pos = draw->vs.position_output;
+ const unsigned cv = draw->vs.clipvertex_output;
boolean have_clipdist = FALSE;
struct lp_bld_tgsi_system_values system_values;
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)
if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
- LLVMAddAttribute(LLVMGetParam(variant_func, i),
- LLVMNoAliasAttribute);
+ lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
context_ptr = LLVMGetParam(variant_func, 0);
io_ptr = LLVMGetParam(variant_func, 1);
vbuffers_ptr = LLVMGetParam(variant_func, 2);
- 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));
+ 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);
+ 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
builder = gallivm->builder;
LLVMPositionBuilderAtEnd(builder, block);
- lp_build_context_init(&bld, gallivm, lp_type_int(32));
-
memset(&vs_type, 0, sizeof vs_type);
vs_type.floating = TRUE; /* floating point values */
vs_type.sign = TRUE; /* values are signed */
vs_type.width = 32; /* 32-bit float */
vs_type.length = vector_length;
+ lp_build_context_init(&bld, gallivm, lp_type_uint(32));
+ lp_build_context_init(&blduivec, gallivm, lp_uint_type(vs_type));
+
/* hold temporary "bool" clipmask */
- clipmask_bool_ptr = lp_build_alloca(gallivm, lp_build_int_vec_type(gallivm, vs_type), "");
- LLVMBuildStore(builder, lp_build_zero(gallivm, lp_int_type(vs_type)), clipmask_bool_ptr);
+ clipmask_bool_ptr = lp_build_alloca(gallivm, blduivec.vec_type, "");
+
+ fake_buf = lp_build_alloca_undef(gallivm,
+ LLVMVectorType(LLVMInt64TypeInContext(context), 4), "");
+ fake_buf = LLVMBuildBitCast(builder, fake_buf,
+ LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
+ fake_buf_ptr = LLVMBuildGEP(builder, fake_buf, &bld.zero, 1, "");
/* code generated texture sampling */
- sampler = draw_llvm_sampler_soa_create(
- draw_llvm_variant_key_samplers(key),
- context_ptr);
-
- if (elts) {
- start = zero;
- end = fetch_count;
- count = fetch_count;
- }
- else {
- end = lp_build_add(&bld, start, count);
- }
+ sampler = draw_llvm_sampler_soa_create(draw_llvm_variant_key_samplers(key));
step = lp_build_const_int32(gallivm, vector_length);
- fetch_max = LLVMBuildSub(builder, end, one, "fetch_max");
+ ind_vec = blduivec.undef;
+ for (i = 0; i < vs_type.length; i++) {
+ LLVMValueRef index = lp_build_const_int32(gallivm, i);
+ ind_vec = LLVMBuildInsertElement(builder, ind_vec, index, index, "");
+ }
+
+ have_elts = LLVMBuildICmp(builder, LLVMIntNE,
+ LLVMConstPointerNull(arg_types[10]), fetch_elts, "");
+
+ 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.
+ */
+ for (j = 0; j < key->nr_vertex_elements; ++j) {
+ LLVMValueRef vb_buffer_offset, buffer_size, temp_ptr;
+ LLVMValueRef vb_info, vbuffer_ptr, buf_offset, ofbit;
+ struct pipe_vertex_element *velem = &key->vertex_element[j];
+ LLVMValueRef vb_index =
+ lp_build_const_int32(gallivm, velem->vertex_buffer_index);
+ LLVMValueRef bsize = lp_build_const_int32(gallivm,
+ util_format_get_blocksize(velem->src_format));
+ LLVMValueRef src_offset = lp_build_const_int32(gallivm,
+ velem->src_offset);
+ struct lp_build_if_state if_ctx;
+
+ if (velem->src_format != PIPE_FORMAT_NONE) {
+ 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;
+ /*
+ * 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);
+
+ /*
+ * We can't easily set fake vertex buffers outside the generated code.
+ * Hence, set fake vertex buffers here instead basically, so fetch
+ * code can always fetch using offset 0, eliminating all control flow
+ * inside the main loop.
+ * (Alternatively, could have control flow per vector skipping fetch
+ * if ofbit is true.)
+ */
+ if (velem->instance_divisor) {
+ /*
+ * 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).
+ * 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,
+ lp_build_const_int32(gallivm,
+ velem->instance_divisor),
+ "instance_divisor");
+ instance_index[j] = lp_build_uadd_overflow(gallivm, start_instance,
+ current_instance, &ofbit);
+ }
+
+ buffer_size_adj[j] = LLVMBuildSelect(builder, ofbit, bld.zero,
+ buffer_size_adj[j], "");
+
+ temp_ptr = lp_build_alloca_undef(gallivm,
+ LLVMPointerType(LLVMInt8TypeInContext(context), 0), "");
- lp_build_loop_begin(&lp_loop, gallivm, zero);
+ 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",
+ lp_build_const_int32(gallivm, j),
+ vb_index, vb_stride[j]);
+ lp_build_printf(gallivm,
+ " vb_buffer_offset = %u, src_offset = %u, buf_offset = %u\n",
+ vb_buffer_offset, src_offset, buf_offset);
+ lp_build_printf(gallivm, " buffer size = %u, blocksize = %u\n",
+ buffer_size, bsize);
+ lp_build_printf(gallivm, " instance_id = %u\n", system_values.instance_id);
+ }
+ }
+ }
+
+ lp_build_loop_begin(&lp_loop, gallivm, bld.zero);
{
LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS];
- LLVMValueRef aos_attribs[PIPE_MAX_SHADER_INPUTS][LP_MAX_VECTOR_WIDTH / 32] = { { 0 } };
LLVMValueRef io;
LLVMValueRef clipmask; /* holds the clipmask value */
+ LLVMValueRef true_index_array, index_store;
const LLVMValueRef (*ptr_aos)[TGSI_NUM_CHANNELS];
io_itr = lp_loop.counter;
lp_build_printf(gallivm, " --- io %d = %p, loop counter %d\n",
io_itr, io, lp_loop.counter);
#endif
- system_values.vertex_id = lp_build_zero(gallivm, lp_type_uint_vec(32, 32*vector_length));
- for (i = 0; i < vector_length; ++i) {
- LLVMValueRef vert_index =
- LLVMBuildAdd(builder,
- lp_loop.counter,
- lp_build_const_int32(gallivm, i), "");
- LLVMValueRef true_index =
- LLVMBuildAdd(builder, start, vert_index, "");
- LLVMValueRef vertex_id;
-
- /* make sure we're not out of bounds which can happen
- * if fetch_count % 4 != 0, because on the last iteration
- * a few of the 4 vertex fetches will be out of bounds */
- true_index = lp_build_min(&bld, true_index, fetch_max);
-
- if (elts) {
- LLVMValueRef fetch_ptr;
- LLVMValueRef index_overflowed;
- LLVMValueRef index_ptr =
- lp_build_alloca(
- gallivm,
- lp_build_vec_type(gallivm, lp_type_int(32)), "");
- struct lp_build_if_state if_ctx;
- index_overflowed = LLVMBuildICmp(builder, LLVMIntUGT,
- true_index, fetch_elt_max,
- "index_overflowed");
-
- lp_build_if(&if_ctx, gallivm, index_overflowed);
- {
- /* Generate maximum possible index so that
- * generate_fetch can treat it just like
- * any other overflow and return zeros.
- * We don't have to worry about the restart
- * primitive index because it has already been
- * handled
- */
- LLVMValueRef val =
- lp_build_const_int32(gallivm, 0xffffffff);
- LLVMBuildStore(builder, val, index_ptr);
- }
- lp_build_else(&if_ctx);
- {
- LLVMValueRef val;
- fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
- &true_index, 1, "");
- val = LLVMBuildLoad(builder, fetch_ptr, "");
- LLVMBuildStore(builder, val, index_ptr);
+
+ 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);
+ /*
+ * 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(&blduivec, true_index_array, fetch_max);
+
+ 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.
+ * There used to be one here but it was incorrect: overflow was
+ * detected if index > fetch_elt_max - but the correct condition
+ * would be index >= fetch_elt_max (since this is just size of elts
+ * buffer / element size).
+ * Using the correct condition however will cause failures - due to
+ * vsplit/vcache code which rebases indices. So, as an example, if
+ * fetch_elt_max is just 1 and fetch_count 2, vsplit cache will
+ * replace all invalid indices with 0 - which in case of elt_bias
+ * 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, 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).
+ */
+ tmp = lp_build_shl_imm(&blduivec, true_index_array, 2);
+ fetch_elts = LLVMBuildBitCast(builder, fetch_elts,
+ LLVMPointerType(LLVMInt8TypeInContext(context),
+ 0), "");
+ 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];
+ const struct util_format_description *format_desc =
+ util_format_description(velem->src_format);
+
+ if (format_desc->format == PIPE_FORMAT_NONE) {
+ for (i = 0; i < TGSI_NUM_CHANNELS; i++) {
+ inputs[j][i] = lp_build_zero(gallivm, vs_type);
}
- lp_build_endif(&if_ctx);
- true_index = LLVMBuildLoad(builder, index_ptr, "true_index");
}
- /* 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
- * get back to the original index.
- * in the linear paths vertex id has to be unaffected by the
- * original start index and because we abuse the 'start' variable
- * to either represent the actual start index or the index at which
- * the primitive was split (we split rendering into chunks of at
- * most 4095-vertices) we need to back out the original start
- * index out of our vertex id here.
- */
- vertex_id = LLVMBuildSub(builder, true_index, vertex_id_offset, "");
-
- system_values.vertex_id = LLVMBuildInsertElement(
- gallivm->builder,
- system_values.vertex_id, vertex_id,
- lp_build_const_int32(gallivm, i), "");
-
- for (j = 0; j < draw->pt.nr_vertex_elements; ++j) {
- struct pipe_vertex_element *velem = &draw->pt.vertex_element[j];
- LLVMValueRef vb_index =
- lp_build_const_int32(gallivm, velem->vertex_buffer_index);
- LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr, &vb_index, 1, "");
- generate_fetch(gallivm, draw, vbuffers_ptr,
- &aos_attribs[j][i], velem, vb, true_index,
- system_values.instance_id, start_instance);
+ else if (velem->instance_divisor) {
+ fetch_instanced(gallivm, format_desc, vs_type,
+ vb_stride[j], map_ptr[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],
+ inputs[j], true_index_array);
}
}
- convert_to_soa(gallivm, aos_attribs, inputs,
- draw->pt.nr_vertex_elements, vs_type);
+
+ 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
+ * get back to the original index.
+ * in the linear paths vertex id has to be unaffected by the
+ * original start index and because we abuse the 'start' variable
+ * to either represent the actual start index or the index at which
+ * the primitive was split (we split rendering into chunks of at
+ * most 4095-vertices) we need to back out the original start
+ * index out of our vertex id here.
+ */
+ system_values.basevertex = lp_build_broadcast_scalar(&blduivec,
+ vertex_id_offset);
+ system_values.vertex_id = true_index_array;
+ system_values.vertex_id_nobase = LLVMBuildSub(builder, true_index_array,
+ system_values.basevertex, "");
ptr_aos = (const LLVMValueRef (*)[TGSI_NUM_CHANNELS]) inputs;
generate_vs(variant,
&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, FALSE, key->clip_user ? cv : pos);
- store_clip(gallivm, vs_type, io, outputs, TRUE, pos);
+ store_clip(gallivm, vs_type, io, outputs, pos);
/* do cliptest */
if (enable_cliptest) {
gallivm,
vs_type,
outputs,
- key->clip_xy,
- key->clip_z,
- key->clip_user,
- key->clip_halfz,
- key->ucp_enable,
+ key,
context_ptr, &have_clipdist);
temp = LLVMBuildOr(builder, clipmask, temp, "");
/* store temporary clipping boolean value */
LLVMBuildStore(builder, temp, clipmask_bool_ptr);
}
else {
- clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
+ clipmask = blduivec.zero;
}
/* do viewport mapping */
}
}
else {
- clipmask = lp_build_const_int_vec(gallivm, lp_int_type(vs_type), 0);
+ clipmask = blduivec.zero;
}
/* store clipmask in vertex header,
*/
convert_to_aos(gallivm, io, NULL, outputs, clipmask,
vs_info->num_outputs, vs_type,
- have_clipdist);
+ enable_cliptest && key->need_edgeflags);
}
lp_build_loop_end_cond(&lp_loop, count, step, LLVMIntUGE);
sampler->destroy(sampler);
/* return clipping boolean value for function */
- ret = clipmask_booli32(gallivm, vs_type, clipmask_bool_ptr);
+ ret = clipmask_booli8(gallivm, vs_type, clipmask_bool_ptr,
+ enable_cliptest && key->need_edgeflags);
LLVMBuildRet(builder, ret);
key = (struct draw_llvm_variant_key *)store;
- key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
+ memset(key, 0, offsetof(struct draw_llvm_variant_key, vertex_element[0]));
- /* Presumably all variants of the shader should have the same
- * number of vertex elements - ie the number of shader inputs.
- * NOTE: we NEED to store the needed number of needed inputs
- * here, not the number of provided elements to match keysize
- * (and the offset of sampler state in the key).
- */
- key->nr_vertex_elements = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
- assert(key->nr_vertex_elements <= llvm->draw->pt.nr_vertex_elements);
+ key->clamp_vertex_color = llvm->draw->rasterizer->clamp_vertex_color; /**/
/* will have to rig this up properly later */
key->clip_xy = llvm->draw->clip_xy;
key->clip_z = llvm->draw->clip_z;
key->clip_user = llvm->draw->clip_user;
- key->bypass_viewport = llvm->draw->identity_viewport;
+ key->bypass_viewport = llvm->draw->bypass_viewport;
key->clip_halfz = llvm->draw->rasterizer->clip_halfz;
+ /* XXX assumes edgeflag output not at 0 */
key->need_edgeflags = (llvm->draw->vs.edgeflag_output ? TRUE : FALSE);
key->ucp_enable = llvm->draw->rasterizer->clip_plane_enable;
key->has_gs = llvm->draw->gs.geometry_shader != NULL;
key->num_outputs = draw_total_vs_outputs(llvm->draw);
- key->pad1 = 0;
/* All variants of this shader will have the same value for
* nr_samplers. Not yet trying to compact away holes in the
key->nr_sampler_views = key->nr_samplers;
}
- draw_sampler = draw_llvm_variant_key_samplers(key);
-
+ /* Presumably all variants of the shader should have the same
+ * number of vertex elements - ie the number of shader inputs.
+ * NOTE: we NEED to store the needed number of needed inputs
+ * here, not the number of provided elements to match keysize
+ * (and the offset of sampler state in the key).
+ * If we have excess number of vertex elements, this is valid,
+ * but the excess ones don't matter.
+ * If we don't have enough vertex elements (which looks not really
+ * valid but we'll handle it gracefully) fill out missing ones with
+ * zero (we'll recognize these later by PIPE_FORMAT_NONE).
+ */
+ key->nr_vertex_elements =
+ llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_INPUT] + 1;
+
+ if (llvm->draw->pt.nr_vertex_elements < key->nr_vertex_elements) {
+ debug_printf("draw: vs with %d inputs but only have %d vertex elements\n",
+ key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements);
+ memset(key->vertex_element, 0,
+ sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
+ }
memcpy(key->vertex_element,
llvm->draw->pt.vertex_element,
- sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
+ sizeof(struct pipe_vertex_element) *
+ MIN2(key->nr_vertex_elements, llvm->draw->pt.nr_vertex_elements));
- memset(draw_sampler, 0, MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
+ draw_sampler = draw_llvm_variant_key_samplers(key);
+ memset(draw_sampler, 0,
+ MAX2(key->nr_samplers, key->nr_sampler_views) * sizeof *draw_sampler);
for (i = 0 ; i < key->nr_samplers; i++) {
lp_sampler_static_sampler_state(&draw_sampler[i].sampler_state,
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,
shader_stage == PIPE_SHADER_GEOMETRY);
if (shader_stage == PIPE_SHADER_VERTEX) {
- assert(sview_idx < Elements(draw->llvm->jit_context.textures));
+ assert(sview_idx < ARRAY_SIZE(draw->llvm->jit_context.textures));
jit_tex = &draw->llvm->jit_context.textures[sview_idx];
} else if (shader_stage == PIPE_SHADER_GEOMETRY) {
- assert(sview_idx < Elements(draw->llvm->gs_jit_context.textures));
+ assert(sview_idx < ARRAY_SIZE(draw->llvm->gs_jit_context.textures));
jit_tex = &draw->llvm->gs_jit_context.textures[sview_idx];
} else {
void
draw_llvm_set_sampler_state(struct draw_context *draw,
- unsigned shader_type)
+ enum pipe_shader_type shader_type)
{
unsigned i;
for (i = 0; i < draw->num_samplers[PIPE_SHADER_VERTEX]; i++) {
struct draw_jit_sampler *jit_sam = &draw->llvm->jit_context.samplers[i];
- if (draw->samplers[i]) {
+ if (draw->samplers[PIPE_SHADER_VERTEX][i]) {
const struct pipe_sampler_state *s
= draw->samplers[PIPE_SHADER_VERTEX][i];
jit_sam->min_lod = s->min_lod;
for (i = 0; i < draw->num_samplers[PIPE_SHADER_GEOMETRY]; i++) {
struct draw_jit_sampler *jit_sam = &draw->llvm->gs_jit_context.samplers[i];
- if (draw->samplers[i]) {
+ if (draw->samplers[PIPE_SHADER_GEOMETRY][i]) {
const struct pipe_sampler_state *s
= draw->samplers[PIPE_SHADER_GEOMETRY][i];
jit_sam->min_lod = s->min_lod;
{
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);
num_prims = lp_build_broadcast(gallivm, lp_build_vec_type(gallivm, mask_type),
variant->num_prims);
- for (i = 0; i <= gs_type.length; i++) {
+ for (i = 0; i < gs_type.length; i++) {
LLVMValueRef idx = lp_build_const_int32(gallivm, i);
mask_val = LLVMBuildInsertElement(builder, mask_val, idx, idx, "");
}
struct gallivm_state *gallivm = variant->gallivm;
LLVMContextRef context = gallivm->context;
LLVMTypeRef int32_type = LLVMInt32TypeInContext(context);
- LLVMTypeRef arg_types[6];
+ LLVMTypeRef arg_types[7];
LLVMTypeRef func_type;
LLVMValueRef variant_func;
LLVMValueRef context_ptr;
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);
arg_types[4] = int32_type; /* instance_id */
arg_types[5] = LLVMPointerType(
LLVMVectorType(int32_type, vector_length), 0); /* prim_id_ptr */
+ arg_types[6] = int32_type;
- func_type = LLVMFunctionType(int32_type, arg_types, Elements(arg_types), 0);
+ func_type = LLVMFunctionType(int32_type, arg_types, ARRAY_SIZE(arg_types), 0);
variant_func = LLVMAddFunction(gallivm->module, func_name, func_type);
LLVMSetFunctionCallConv(variant_func, LLVMCCallConv);
- for (i = 0; i < Elements(arg_types); ++i)
+ for (i = 0; i < ARRAY_SIZE(arg_types); ++i)
if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
- LLVMAddAttribute(LLVMGetParam(variant_func, i),
- LLVMNoAliasAttribute);
+ lp_add_function_attr(variant_func, i + 1, LP_FUNC_ATTR_NOALIAS);
context_ptr = LLVMGetParam(variant_func, 0);
input_array = LLVMGetParam(variant_func, 1);
num_prims = LLVMGetParam(variant_func, 3);
system_values.instance_id = LLVMGetParam(variant_func, 4);
prim_id_ptr = LLVMGetParam(variant_func, 5);
+ system_values.invocation_id = LLVMGetParam(variant_func, 6);
lp_build_name(context_ptr, "context");
lp_build_name(input_array, "input");
lp_build_name(num_prims, "num_prims");
lp_build_name(system_values.instance_id, "instance_id");
lp_build_name(prim_id_ptr, "prim_id_ptr");
+ lp_build_name(system_values.invocation_id, "invocation_id");
variant->context_ptr = context_ptr;
variant->io_ptr = io_ptr;
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,
- context_ptr);
+ sampler = draw_llvm_sampler_soa_create(variant->key.samplers);
mask_val = generate_mask_value(variant, gs_type);
lp_build_mask_begin(&mask, gallivm, gs_type, mask_val);
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,
- sampler,
- &llvm->draw->gs.geometry_shader->info,
- (const struct lp_build_tgsi_gs_iface *)&gs_iface);
+ ¶ms,
+ outputs);
sampler->destroy(sampler);
variant = MALLOC(sizeof *variant +
shader->variant_key_size -
sizeof variant->key);
- if (variant == NULL)
+ if (!variant)
return NULL;
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);
key = (struct draw_gs_llvm_variant_key *)store;
+ memset(key, 0, offsetof(struct draw_gs_llvm_variant_key, samplers[0]));
+
key->num_outputs = draw_total_gs_outputs(llvm->draw);
/* All variants of this shader will have the same value for