#include "gallivm/lp_bld_debug.h"
#include "gallivm/lp_bld_tgsi.h"
#include "gallivm/lp_bld_printf.h"
+#include "gallivm/lp_bld_intr.h"
+#include "gallivm/lp_bld_init.h"
#include "tgsi/tgsi_exec.h"
#include "tgsi/tgsi_dump.h"
#include "util/u_cpu_detect.h"
+#include "util/u_math.h"
#include "util/u_pointer.h"
#include "util/u_string.h"
elem_types[DRAW_JIT_TEXTURE_DEPTH] = LLVMInt32Type();
elem_types[DRAW_JIT_TEXTURE_LAST_LEVEL] = LLVMInt32Type();
elem_types[DRAW_JIT_TEXTURE_ROW_STRIDE] =
- LLVMArrayType(LLVMInt32Type(), DRAW_MAX_TEXTURE_LEVELS);
+ LLVMArrayType(LLVMInt32Type(), PIPE_MAX_TEXTURE_LEVELS);
elem_types[DRAW_JIT_TEXTURE_IMG_STRIDE] =
- LLVMArrayType(LLVMInt32Type(), DRAW_MAX_TEXTURE_LEVELS);
+ LLVMArrayType(LLVMInt32Type(), PIPE_MAX_TEXTURE_LEVELS);
elem_types[DRAW_JIT_TEXTURE_DATA] =
LLVMArrayType(LLVMPointerType(LLVMInt8Type(), 0),
- DRAW_MAX_TEXTURE_LEVELS);
+ PIPE_MAX_TEXTURE_LEVELS);
+ elem_types[DRAW_JIT_TEXTURE_MIN_LOD] = LLVMFloatType();
+ elem_types[DRAW_JIT_TEXTURE_MAX_LOD] = LLVMFloatType();
+ elem_types[DRAW_JIT_TEXTURE_LOD_BIAS] = LLVMFloatType();
+ elem_types[DRAW_JIT_TEXTURE_BORDER_COLOR] =
+ LLVMArrayType(LLVMFloatType(), 4);
texture_type = LLVMStructType(elem_types, Elements(elem_types), 0);
LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, data,
llvm->target, texture_type,
DRAW_JIT_TEXTURE_DATA);
+ LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, min_lod,
+ llvm->target, texture_type,
+ DRAW_JIT_TEXTURE_MIN_LOD);
+ LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, max_lod,
+ llvm->target, texture_type,
+ DRAW_JIT_TEXTURE_MAX_LOD);
+ LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, lod_bias,
+ llvm->target, texture_type,
+ DRAW_JIT_TEXTURE_LOD_BIAS);
+ LP_CHECK_MEMBER_OFFSET(struct draw_jit_texture, border_color,
+ llvm->target, texture_type,
+ DRAW_JIT_TEXTURE_BORDER_COLOR);
LP_CHECK_STRUCT_SIZE(struct draw_jit_texture,
llvm->target, texture_type);
{
struct draw_llvm *llvm;
-#ifdef PIPE_ARCH_X86
- util_cpu_detect();
- /* require SSE2 due to LLVM PR6960. */
- if (!util_cpu_caps.has_sse2)
- return NULL;
-#endif
-
llvm = CALLOC_STRUCT( draw_llvm );
if (!llvm)
return NULL;
/* These are the passes currently listed in llvm-c/Transforms/Scalar.h,
* but there are more on SVN. */
/* TODO: Add more passes */
+
LLVMAddCFGSimplificationPass(llvm->pass);
- LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
- LLVMAddConstantPropagationPass(llvm->pass);
+
+ if (HAVE_LLVM >= 0x207 && sizeof(void*) == 4) {
+ /* For LLVM >= 2.7 and 32-bit build, use this order of passes to
+ * avoid generating bad code.
+ * Test with piglit glsl-vs-sqrt-zero test.
+ */
+ LLVMAddConstantPropagationPass(llvm->pass);
+ LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
+ }
+ else {
+ LLVMAddPromoteMemoryToRegisterPass(llvm->pass);
+ LLVMAddConstantPropagationPass(llvm->pass);
+ }
+
if(util_cpu_caps.has_sse4_1) {
/* FIXME: There is a bug in this pass, whereby the combination of fptosi
* and sitofp (necessary for trunc/floor/ceil/round implementation)
}
struct draw_llvm_variant *
-draw_llvm_create_variant(struct draw_llvm *llvm, int num_inputs)
+draw_llvm_create_variant(struct draw_llvm *llvm,
+ unsigned num_inputs,
+ const struct draw_llvm_variant_key *key)
{
- struct draw_llvm_variant *variant = MALLOC(sizeof(struct draw_llvm_variant));
+ struct draw_llvm_variant *variant;
struct llvm_vertex_shader *shader =
llvm_vertex_shader(llvm->draw->vs.vertex_shader);
+ variant = MALLOC(sizeof *variant +
+ shader->variant_key_size -
+ sizeof variant->key);
+ if (variant == NULL)
+ return NULL;
+
variant->llvm = llvm;
- draw_llvm_make_variant_key(llvm, &variant->key);
+ memcpy(&variant->key, key, shader->variant_key_size);
llvm->vertex_header_ptr_type = create_vertex_header(llvm, num_inputs);
LLVMValueRef (*outputs)[NUM_CHANNELS],
const LLVMValueRef (*inputs)[NUM_CHANNELS],
LLVMValueRef context_ptr,
- struct lp_build_sampler_soa *sampler)
+ struct lp_build_sampler_soa *draw_sampler)
{
const struct tgsi_token *tokens = llvm->draw->vs.vertex_shader->state.tokens;
struct lp_type vs_type;
LLVMValueRef consts_ptr = draw_jit_context_vs_constants(builder, context_ptr);
+ struct lp_build_sampler_soa *sampler = 0;
memset(&vs_type, 0, sizeof vs_type);
vs_type.floating = TRUE; /* floating point values */
tgsi_dump(tokens, 0);
}
+ if (llvm->draw->num_sampler_views &&
+ llvm->draw->num_samplers)
+ sampler = draw_sampler;
+
lp_build_tgsi_soa(builder,
tokens,
vs_type,
LLVMValueRef *res,
struct pipe_vertex_element *velem,
LLVMValueRef vbuf,
- LLVMValueRef index)
+ LLVMValueRef index,
+ LLVMValueRef instance_id)
{
LLVMValueRef indices = LLVMConstInt(LLVMInt64Type(), velem->vertex_buffer_index, 0);
LLVMValueRef vbuffer_ptr = LLVMBuildGEP(builder, vbuffers_ptr,
LLVMValueRef cond;
LLVMValueRef stride;
- cond = LLVMBuildICmp(builder, LLVMIntULE, index, vb_max_index, "");
+ if (velem->instance_divisor) {
+ /* array index = instance_id / instance_divisor */
+ index = LLVMBuildUDiv(builder, instance_id,
+ LLVMConstInt(LLVMInt32Type(), velem->instance_divisor, 0),
+ "instance_divisor");
+ }
+ /* limit index to min(inex, vb_max_index) */
+ cond = LLVMBuildICmp(builder, LLVMIntULE, index, vb_max_index, "");
index = LLVMBuildSelect(builder, cond, index, vb_max_index, "");
stride = LLVMBuildMul(builder, vb_stride, index, "");
static void
draw_llvm_generate(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
{
- LLVMTypeRef arg_types[7];
+ LLVMTypeRef arg_types[8];
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef start, end, count, stride, step, io_itr;
LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
+ LLVMValueRef instance_id;
struct draw_context *draw = llvm->draw;
unsigned i, j;
struct lp_build_context bld;
struct lp_build_loop_state lp_loop;
- struct lp_type vs_type = lp_type_float_vec(32);
const int max_vertices = 4;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
void *code;
arg_types[4] = LLVMInt32Type(); /* count */
arg_types[5] = LLVMInt32Type(); /* stride */
arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
+ arg_types[7] = LLVMInt32Type(); /* instance_id */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
count = LLVMGetParam(variant->function, 4);
stride = LLVMGetParam(variant->function, 5);
vb_ptr = LLVMGetParam(variant->function, 6);
+ instance_id = LLVMGetParam(variant->function, 7);
lp_build_name(context_ptr, "context");
lp_build_name(io_ptr, "io");
lp_build_name(count, "count");
lp_build_name(stride, "stride");
lp_build_name(vb_ptr, "vb");
+ lp_build_name(instance_id, "instance_id");
/*
* Function body
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
- lp_build_context_init(&bld, builder, vs_type);
+ lp_build_context_init(&bld, builder, lp_type_int(32));
end = lp_build_add(&bld, start, count);
step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
/* code generated texture sampling */
- sampler = draw_llvm_sampler_soa_create(variant->key.sampler,
- context_ptr);
+ sampler = draw_llvm_sampler_soa_create(
+ draw_llvm_variant_key_samplers(&variant->key),
+ context_ptr);
#if DEBUG_STORE
lp_build_printf(builder, "start = %d, end = %d, step = %d\n",
LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
&vb_index, 1, "");
generate_fetch(builder, vbuffers_ptr,
- &aos_attribs[j][i], velem, vb, true_index);
+ &aos_attribs[j][i], velem, vb, true_index,
+ instance_id);
}
}
convert_to_soa(builder, aos_attribs, inputs,
sampler->destroy(sampler);
+#ifdef PIPE_ARCH_X86
+ /* Avoid corrupting the FPU stack on 32bit OSes. */
+ lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
+#endif
+
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
if (gallivm_debug & GALLIVM_DEBUG_ASM) {
lp_disassemble(code);
}
+ lp_func_delete_body(variant->function);
}
static void
draw_llvm_generate_elts(struct draw_llvm *llvm, struct draw_llvm_variant *variant)
{
- LLVMTypeRef arg_types[7];
+ LLVMTypeRef arg_types[8];
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef fetch_elts, fetch_count, stride, step, io_itr;
LLVMValueRef io_ptr, vbuffers_ptr, vb_ptr;
+ LLVMValueRef instance_id;
struct draw_context *draw = llvm->draw;
unsigned i, j;
struct lp_build_context bld;
- struct lp_build_context bld_int;
struct lp_build_loop_state lp_loop;
- struct lp_type vs_type = lp_type_float_vec(32);
const int max_vertices = 4;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
LLVMValueRef fetch_max;
arg_types[4] = LLVMInt32Type(); /* fetch_count */
arg_types[5] = LLVMInt32Type(); /* stride */
arg_types[6] = llvm->vb_ptr_type; /* pipe_vertex_buffer's */
+ arg_types[7] = LLVMInt32Type(); /* instance_id */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
- variant->function_elts = LLVMAddFunction(llvm->module, "draw_llvm_shader_elts", func_type);
+ variant->function_elts = LLVMAddFunction(llvm->module, "draw_llvm_shader_elts",
+ func_type);
LLVMSetFunctionCallConv(variant->function_elts, LLVMCCallConv);
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
- LLVMAddAttribute(LLVMGetParam(variant->function_elts, i), LLVMNoAliasAttribute);
+ LLVMAddAttribute(LLVMGetParam(variant->function_elts, i),
+ LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(variant->function_elts, 0);
io_ptr = LLVMGetParam(variant->function_elts, 1);
fetch_count = LLVMGetParam(variant->function_elts, 4);
stride = LLVMGetParam(variant->function_elts, 5);
vb_ptr = LLVMGetParam(variant->function_elts, 6);
+ instance_id = LLVMGetParam(variant->function_elts, 7);
lp_build_name(context_ptr, "context");
lp_build_name(io_ptr, "io");
lp_build_name(fetch_count, "fetch_count");
lp_build_name(stride, "stride");
lp_build_name(vb_ptr, "vb");
+ lp_build_name(instance_id, "instance_id");
/*
* Function body
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
- lp_build_context_init(&bld, builder, vs_type);
- lp_build_context_init(&bld_int, builder, lp_type_int(32));
+ lp_build_context_init(&bld, builder, lp_type_int(32));
step = LLVMConstInt(LLVMInt32Type(), max_vertices, 0);
/* code generated texture sampling */
- sampler = draw_llvm_sampler_soa_create(variant->key.sampler,
- context_ptr);
+ sampler = draw_llvm_sampler_soa_create(
+ draw_llvm_variant_key_samplers(&variant->key),
+ context_ptr);
fetch_max = LLVMBuildSub(builder, fetch_count,
LLVMConstInt(LLVMInt32Type(), 1, 0),
/* 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_int, true_index, fetch_max);
+ true_index = lp_build_min(&bld, true_index, fetch_max);
fetch_ptr = LLVMBuildGEP(builder, fetch_elts,
&true_index, 1, "");
LLVMValueRef vb = LLVMBuildGEP(builder, vb_ptr,
&vb_index, 1, "");
generate_fetch(builder, vbuffers_ptr,
- &aos_attribs[j][i], velem, vb, true_index);
+ &aos_attribs[j][i], velem, vb, true_index,
+ instance_id);
}
}
convert_to_soa(builder, aos_attribs, inputs,
sampler->destroy(sampler);
+#ifdef PIPE_ARCH_X86
+ /* Avoid corrupting the FPU stack on 32bit OSes. */
+ lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
+#endif
+
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
if (gallivm_debug & GALLIVM_DEBUG_ASM) {
lp_disassemble(code);
}
+ lp_func_delete_body(variant->function_elts);
}
-void
-draw_llvm_make_variant_key(struct draw_llvm *llvm,
- struct draw_llvm_variant_key *key)
+
+struct draw_llvm_variant_key *
+draw_llvm_make_variant_key(struct draw_llvm *llvm, char *store)
{
unsigned i;
+ struct draw_llvm_variant_key *key;
+ struct lp_sampler_static_state *sampler;
- memset(key, 0, sizeof(struct draw_llvm_variant_key));
+ key = (struct draw_llvm_variant_key *)store;
+ /* Presumably all variants of the shader should have the same
+ * number of vertex elements - ie the number of shader inputs.
+ */
key->nr_vertex_elements = llvm->draw->pt.nr_vertex_elements;
+ /* All variants of this shader will have the same value for
+ * nr_samplers. Not yet trying to compact away holes in the
+ * sampler array.
+ */
+ key->nr_samplers = llvm->draw->vs.vertex_shader->info.file_max[TGSI_FILE_SAMPLER] + 1;
+
+ sampler = draw_llvm_variant_key_samplers(key);
+
memcpy(key->vertex_element,
llvm->draw->pt.vertex_element,
sizeof(struct pipe_vertex_element) * key->nr_vertex_elements);
+
+ memset(sampler, 0, key->nr_samplers * sizeof *sampler);
- memcpy(&key->vs,
- &llvm->draw->vs.vertex_shader->state,
- sizeof(struct pipe_shader_state));
-
- for(i = 0; i < PIPE_MAX_VERTEX_SAMPLERS; ++i) {
- struct draw_vertex_shader *shader = llvm->draw->vs.vertex_shader;
- if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
- lp_sampler_static_state(&key->sampler[i],
- llvm->draw->sampler_views[i],
- llvm->draw->samplers[i]);
+ for (i = 0 ; i < key->nr_samplers; i++) {
+ lp_sampler_static_state(&sampler[i],
+ llvm->draw->sampler_views[i],
+ llvm->draw->samplers[i]);
}
+
+ return key;
}
void
unsigned sampler_idx,
uint32_t width, uint32_t height, uint32_t depth,
uint32_t last_level,
- uint32_t row_stride[DRAW_MAX_TEXTURE_LEVELS],
- uint32_t img_stride[DRAW_MAX_TEXTURE_LEVELS],
- const void *data[DRAW_MAX_TEXTURE_LEVELS])
+ uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS],
+ uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS],
+ const void *data[PIPE_MAX_TEXTURE_LEVELS])
{
unsigned j;
struct draw_jit_texture *jit_tex;
- assert(sampler_idx <= PIPE_MAX_VERTEX_SAMPLERS);
+ assert(sampler_idx < PIPE_MAX_VERTEX_SAMPLERS);
jit_tex = &draw->llvm->jit_context.textures[sampler_idx];
}
}
+
+void
+draw_llvm_set_sampler_state(struct draw_context *draw)
+{
+ unsigned i;
+
+ for (i = 0; i < draw->num_samplers; i++) {
+ struct draw_jit_texture *jit_tex = &draw->llvm->jit_context.textures[i];
+
+ if (draw->samplers[i]) {
+ jit_tex->min_lod = draw->samplers[i]->min_lod;
+ jit_tex->max_lod = draw->samplers[i]->max_lod;
+ jit_tex->lod_bias = draw->samplers[i]->lod_bias;
+ COPY_4V(jit_tex->border_color, draw->samplers[i]->border_color);
+ }
+ }
+}
+
+
void
draw_llvm_destroy_variant(struct draw_llvm_variant *variant)
{
projects / mesa.git / blobdiff