summaryrefslogtreecommitdiffstats
path: root/src/gallium
diff options
context:
space:
mode:
authorRoland Scheidegger <[email protected]>2015-03-27 02:39:59 +0100
committerRoland Scheidegger <[email protected]>2015-03-27 19:25:53 +0100
commit56076be2ac60f005770516a9ac305e60e6285497 (patch)
tree1b52258911bece7c5ebd234a8f65a2796a4ba95d /src/gallium
parent8dad9455ff748c543635b24908566c3b94cb93a9 (diff)
gallivm: use llvm function calls for texturing instead of inlining
There are issues with inlining everything, most notably llvm will use much more memory (and be slower) when compiling. Ideally we'd probably use functions for shader functions too but texture sampling usually is responsible for quite some IR (it can easily reach 80% of total IR instructions) so this seems like a good start. This still generates a different function for all different combinations just like before, however it is possible llvm is missing some optimization opportunities - it is believed though such opportunities should be somewhat rare, but at least for now it can still be switched off (at compile time only). It should probably make compiled code also smaller because the same function should be used for different variants in the same module (so for the opaque/partial or linear/elts variants). No piglit change (though it does indeed speed up unrealistic tests like fp-indirections2 by a factor of 30 or so). Has a small negative performance impact in openarena - I suspect this could be fixed by running some IPO passes (despite the private linkage, llvm right now does NO optimization at all wrt anything going past the call, even if there's just one caller - so things like values stored before the call and then always written by the function etc. will not be optimized away, nor will dead arguments (which we mostly shouldn't have) be eliminated, always constant arguments promoted etc.). v2: use proper return values instead of pointer function arguments. llvm supports aggregate return values, which do wonders here eliminating unnecessary stack variables - everything in fact will be returned in registers even without any IPO optimizations. It makes the code simpler too. With this I could not measure a peformance impact in openarena any longer (though since there's still no constant value propagation etc. into the tex functions this does not mean it couldn't have a negative impact elsewhere). v3: fix some minor issues suggested by Jose, and do disassembly (and the profiling) without hacks. Reviewed-by: Jose Fonseca <[email protected]>
Diffstat (limited to 'src/gallium')
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_init.c44
-rw-r--r--src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c420
2 files changed, 438 insertions, 26 deletions
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_init.c b/src/gallium/auxiliary/gallivm/lp_bld_init.c
index 6133883e149..7b906c27ed5 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_init.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_init.c
@@ -113,6 +113,11 @@ create_pass_manager(struct gallivm_state *gallivm)
gallivm->passmgr = LLVMCreateFunctionPassManagerForModule(gallivm->module);
if (!gallivm->passmgr)
return FALSE;
+ /*
+ * TODO: some per module pass manager with IPO passes might be helpful -
+ * the generated texture functions may benefit from inlining if they are
+ * simple, or constant propagation into them, etc.
+ */
// Old versions of LLVM get the DataLayout from the pass manager.
LLVMAddTargetData(gallivm->target, gallivm->passmgr);
@@ -556,6 +561,37 @@ gallivm_compile_module(struct gallivm_state *gallivm)
assert(gallivm->engine);
++gallivm->compiled;
+
+ if (gallivm_debug & GALLIVM_DEBUG_ASM) {
+ LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
+
+ while (llvm_func) {
+ /*
+ * Need to filter out functions which don't have an implementation,
+ * such as the intrinsics. May not be sufficient in case of IPO?
+ * LLVMGetPointerToGlobal() will abort otherwise.
+ */
+ if (!LLVMIsDeclaration(llvm_func)) {
+ void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
+ lp_disassemble(llvm_func, func_code);
+ }
+ llvm_func = LLVMGetNextFunction(llvm_func);
+ }
+ }
+
+#if defined(PROFILE)
+ {
+ LLVMValueRef llvm_func = LLVMGetFirstFunction(gallivm->module);
+
+ while (llvm_func) {
+ if (!LLVMIsDeclaration(llvm_func)) {
+ void *func_code = LLVMGetPointerToGlobal(gallivm->engine, llvm_func);
+ lp_profile(llvm_func, func_code);
+ }
+ llvm_func = LLVMGetNextFunction(llvm_func);
+ }
+ }
+#endif
}
@@ -574,13 +610,5 @@ gallivm_jit_function(struct gallivm_state *gallivm,
assert(code);
jit_func = pointer_to_func(code);
- if (gallivm_debug & GALLIVM_DEBUG_ASM) {
- lp_disassemble(func, code);
- }
-
-#if defined(PROFILE)
- lp_profile(func, code);
-#endif
-
return jit_func;
}
diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
index a90278ea889..9e74bd6f5d9 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_soa.c
@@ -2357,30 +2357,30 @@ lp_build_sample_nop(struct gallivm_state *gallivm,
/**
- * Build texture sampling code.
+ * Build the actual texture sampling code.
* 'texel' will return a vector of four LLVMValueRefs corresponding to
* R, G, B, A.
* \param type vector float type to use for coords, etc.
* \param is_fetch if this is a texel fetch instruction.
* \param derivs partial derivatives of (s,t,r,q) with respect to x and y
*/
-void
-lp_build_sample_soa(struct gallivm_state *gallivm,
- const struct lp_static_texture_state *static_texture_state,
- const struct lp_static_sampler_state *static_sampler_state,
- struct lp_sampler_dynamic_state *dynamic_state,
- struct lp_type type,
- boolean is_fetch,
- unsigned texture_index,
- unsigned sampler_index,
- LLVMValueRef context_ptr,
- const LLVMValueRef *coords,
- const LLVMValueRef *offsets,
- const struct lp_derivatives *derivs, /* optional */
- LLVMValueRef lod_bias, /* optional */
- LLVMValueRef explicit_lod, /* optional */
- enum lp_sampler_lod_property lod_property,
- LLVMValueRef texel_out[4])
+static void
+lp_build_sample_soa_code(struct gallivm_state *gallivm,
+ const struct lp_static_texture_state *static_texture_state,
+ const struct lp_static_sampler_state *static_sampler_state,
+ struct lp_sampler_dynamic_state *dynamic_state,
+ struct lp_type type,
+ boolean is_fetch,
+ unsigned texture_index,
+ unsigned sampler_index,
+ LLVMValueRef context_ptr,
+ const LLVMValueRef *coords,
+ const LLVMValueRef *offsets,
+ const struct lp_derivatives *derivs, /* optional */
+ LLVMValueRef lod_bias, /* optional */
+ LLVMValueRef explicit_lod, /* optional */
+ enum lp_sampler_lod_property lod_property,
+ LLVMValueRef texel_out[4])
{
unsigned target = static_texture_state->target;
unsigned dims = texture_dims(target);
@@ -2891,6 +2891,390 @@ lp_build_sample_soa(struct gallivm_state *gallivm,
}
+#define USE_TEX_FUNC_CALL 1
+
+#define LP_MAX_TEX_FUNC_ARGS 32
+
+#define LP_SAMPLER_FUNC_LOD_BIAS (1 << 0)
+#define LP_SAMPLER_FUNC_LOD_EXPLICIT (1 << 1)
+#define LP_SAMPLER_FUNC_EXPLICITDERIVS (1 << 2)
+#define LP_SAMPLER_FUNC_SHADOW (1 << 3)
+#define LP_SAMPLER_FUNC_OFFSETS (1 << 4)
+#define LP_SAMPLER_FUNC_LOD_PROPERTY_SHIFT 5
+
+
+static inline void
+get_target_info(enum pipe_texture_target target,
+ unsigned *num_coords, unsigned *num_derivs,
+ unsigned *num_offsets, unsigned *layer)
+{
+ unsigned dims = texture_dims(target);
+ *num_coords = dims;
+ *num_offsets = dims;
+ *num_derivs = (target == PIPE_TEXTURE_CUBE ||
+ target == PIPE_TEXTURE_CUBE_ARRAY) ? 3 : dims;
+ *layer = has_layer_coord(target) ? 2: 0;
+ if (target == PIPE_TEXTURE_CUBE_ARRAY) {
+ /*
+ * dims doesn't include r coord for cubes - this is handled
+ * by layer instead, but need to fix up for cube arrays...
+ */
+ *layer = 3;
+ *num_coords = 3;
+ }
+}
+
+
+/**
+ * Generate the function body for a texture sampling function.
+ */
+static void
+lp_build_sample_gen_func(struct gallivm_state *gallivm,
+ const struct lp_static_texture_state *static_texture_state,
+ const struct lp_static_sampler_state *static_sampler_state,
+ struct lp_sampler_dynamic_state *dynamic_state,
+ struct lp_type type,
+ boolean is_fetch,
+ unsigned texture_index,
+ unsigned sampler_index,
+ LLVMValueRef function,
+ unsigned num_args,
+ unsigned sampler_bits,
+ enum lp_sampler_lod_property lod_property)
+{
+
+ LLVMBuilderRef old_builder;
+ LLVMBasicBlockRef block;
+ LLVMValueRef coords[5];
+ LLVMValueRef offsets[3] = { NULL };
+ LLVMValueRef lod_bias = NULL;
+ LLVMValueRef explicit_lod = NULL;
+ LLVMValueRef context_ptr;
+ LLVMValueRef texel_out[4];
+ struct lp_derivatives derivs;
+ struct lp_derivatives *deriv_ptr = NULL;
+ unsigned num_param = 0;
+ unsigned i, num_coords, num_derivs, num_offsets, layer;
+
+ get_target_info(static_texture_state->target,
+ &num_coords, &num_derivs, &num_offsets, &layer);
+
+ /* "unpack" arguments */
+ context_ptr = LLVMGetParam(function, num_param++);
+ for (i = 0; i < num_coords; i++) {
+ coords[i] = LLVMGetParam(function, num_param++);
+ }
+ for (i = num_coords; i < 5; i++) {
+ /* This is rather unfortunate... */
+ coords[i] = lp_build_undef(gallivm, type);
+ }
+ if (layer) {
+ coords[layer] = LLVMGetParam(function, num_param++);
+ }
+ if (sampler_bits & LP_SAMPLER_FUNC_SHADOW) {
+ coords[4] = LLVMGetParam(function, num_param++);
+ }
+ if (sampler_bits & LP_SAMPLER_FUNC_OFFSETS) {
+ for (i = 0; i < num_offsets; i++) {
+ offsets[i] = LLVMGetParam(function, num_param++);
+ }
+ }
+ if (sampler_bits & LP_SAMPLER_FUNC_LOD_BIAS) {
+ lod_bias = LLVMGetParam(function, num_param++);
+ }
+ else if (sampler_bits & LP_SAMPLER_FUNC_LOD_EXPLICIT) {
+ explicit_lod = LLVMGetParam(function, num_param++);
+ }
+ else if (sampler_bits & LP_SAMPLER_FUNC_EXPLICITDERIVS) {
+ for (i = 0; i < num_derivs; i++) {
+ derivs.ddx[i] = LLVMGetParam(function, num_param++);
+ derivs.ddy[i] = LLVMGetParam(function, num_param++);
+ }
+ deriv_ptr = &derivs;
+ }
+
+ assert(num_args == num_param);
+
+ /*
+ * Function body
+ */
+
+ old_builder = gallivm->builder;
+ block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
+ gallivm->builder = LLVMCreateBuilderInContext(gallivm->context);
+ LLVMPositionBuilderAtEnd(gallivm->builder, block);
+
+ lp_build_sample_soa_code(gallivm,
+ static_texture_state,
+ static_sampler_state,
+ dynamic_state,
+ type,
+ is_fetch,
+ texture_index,
+ sampler_index,
+ context_ptr,
+ coords,
+ offsets,
+ deriv_ptr,
+ lod_bias,
+ explicit_lod,
+ lod_property,
+ texel_out);
+
+ LLVMBuildAggregateRet(gallivm->builder, texel_out, 4);
+
+ LLVMDisposeBuilder(gallivm->builder);
+ gallivm->builder = old_builder;
+
+ gallivm_verify_function(gallivm, function);
+}
+
+
+/**
+ * Call the matching function for texture sampling.
+ * If there's no match, generate a new one.
+ */
+static void
+lp_build_sample_soa_func(struct gallivm_state *gallivm,
+ const struct lp_static_texture_state *static_texture_state,
+ const struct lp_static_sampler_state *static_sampler_state,
+ struct lp_sampler_dynamic_state *dynamic_state,
+ struct lp_type type,
+ boolean is_fetch,
+ unsigned texture_index,
+ unsigned sampler_index,
+ LLVMValueRef context_ptr,
+ const LLVMValueRef *coords,
+ const LLVMValueRef *offsets,
+ const struct lp_derivatives *derivs, /* optional */
+ LLVMValueRef lod_bias, /* optional */
+ LLVMValueRef explicit_lod, /* optional */
+ enum lp_sampler_lod_property lod_property,
+ LLVMValueRef texel_out[4])
+{
+ LLVMBuilderRef builder = gallivm->builder;
+ LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(
+ LLVMGetInsertBlock(builder)));
+ LLVMValueRef function, inst;
+ LLVMValueRef args[LP_MAX_TEX_FUNC_ARGS];
+ LLVMBasicBlockRef bb;
+ LLVMValueRef tex_ret;
+ unsigned num_args = 0;
+ unsigned sampler_bits = 0;
+ char func_name[64];
+ unsigned i, num_coords, num_derivs, num_offsets, layer;
+
+ get_target_info(static_texture_state->target,
+ &num_coords, &num_derivs, &num_offsets, &layer);
+
+ /*
+ * texture function matches are found by name.
+ * Thus the name has to include both the texture and sampler unit
+ * (which covers all static state) plus the actual texture functions
+ * (which is determined here somewhat awkwardly by presence of the
+ * corresponding LLVMValueRefs). Additionally, lod_property also
+ * has to be included (it could change if the lod for instance comes
+ * from a shader uniform or a temp reg).
+ */
+ if (static_sampler_state->compare_mode != PIPE_TEX_COMPARE_NONE) {
+ sampler_bits |= LP_SAMPLER_FUNC_SHADOW;
+ }
+ if (offsets[0]) {
+ sampler_bits |= LP_SAMPLER_FUNC_OFFSETS;
+ }
+ if (lod_bias) {
+ sampler_bits |= LP_SAMPLER_FUNC_LOD_BIAS;
+ }
+ else if (explicit_lod) {
+ sampler_bits |= LP_SAMPLER_FUNC_LOD_EXPLICIT;
+ }
+ else if (derivs) {
+ sampler_bits |= LP_SAMPLER_FUNC_EXPLICITDERIVS;
+ }
+ sampler_bits |= lod_property << LP_SAMPLER_FUNC_LOD_PROPERTY_SHIFT;
+
+ util_snprintf(func_name, sizeof(func_name), "texfunc_res_%d_sam_%d_%x",
+ texture_index, sampler_index, sampler_bits);
+
+ function = LLVMGetNamedFunction(module, func_name);
+
+ if(!function) {
+ LLVMTypeRef arg_types[LP_MAX_TEX_FUNC_ARGS];
+ LLVMTypeRef ret_type;
+ LLVMTypeRef function_type;
+ LLVMTypeRef val_type[4];
+ unsigned num_param = 0;
+
+ /*
+ * Generate the function prototype.
+ */
+
+ arg_types[num_param++] = LLVMTypeOf(context_ptr);
+ for (i = 0; i < num_coords; i++) {
+ arg_types[num_param++] = LLVMTypeOf(coords[0]);
+ assert(LLVMTypeOf(coords[0]) == LLVMTypeOf(coords[i]));
+ }
+ if (layer) {
+ arg_types[num_param++] = LLVMTypeOf(coords[layer]);
+ assert(LLVMTypeOf(coords[0]) == LLVMTypeOf(coords[layer]));
+ }
+ if (sampler_bits & LP_SAMPLER_FUNC_SHADOW) {
+ arg_types[num_param++] = LLVMTypeOf(coords[0]);
+ }
+ if (sampler_bits & LP_SAMPLER_FUNC_OFFSETS) {
+ for (i = 0; i < num_offsets; i++) {
+ arg_types[num_param++] = LLVMTypeOf(offsets[0]);
+ assert(LLVMTypeOf(offsets[0]) == LLVMTypeOf(offsets[i]));
+ }
+ }
+ if (sampler_bits & LP_SAMPLER_FUNC_LOD_BIAS) {
+ arg_types[num_param++] = LLVMTypeOf(lod_bias);
+ }
+ else if (sampler_bits & LP_SAMPLER_FUNC_LOD_EXPLICIT) {
+ arg_types[num_param++] = LLVMTypeOf(explicit_lod);
+ }
+ else if (sampler_bits & LP_SAMPLER_FUNC_EXPLICITDERIVS) {
+ for (i = 0; i < num_derivs; i++) {
+ arg_types[num_param++] = LLVMTypeOf(derivs->ddx[i]);
+ arg_types[num_param++] = LLVMTypeOf(derivs->ddy[i]);
+ assert(LLVMTypeOf(derivs->ddx[0]) == LLVMTypeOf(derivs->ddx[i]));
+ assert(LLVMTypeOf(derivs->ddy[0]) == LLVMTypeOf(derivs->ddy[i]));
+ }
+ }
+
+ ret_type = LLVMVoidTypeInContext(gallivm->context);
+ val_type[0] = val_type[1] = val_type[2] = val_type[3] =
+ lp_build_vec_type(gallivm, type);
+ ret_type = LLVMStructTypeInContext(gallivm->context, val_type, 4, 0);
+ function_type = LLVMFunctionType(ret_type, arg_types, num_param, 0);
+ function = LLVMAddFunction(module, func_name, function_type);
+
+ for (i = 0; i < num_param; ++i) {
+ if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) {
+ LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
+ }
+ }
+
+ LLVMSetFunctionCallConv(function, LLVMFastCallConv);
+ LLVMSetLinkage(function, LLVMPrivateLinkage);
+
+ lp_build_sample_gen_func(gallivm,
+ static_texture_state,
+ static_sampler_state,
+ dynamic_state,
+ type,
+ is_fetch,
+ texture_index,
+ sampler_index,
+ function,
+ num_param,
+ sampler_bits,
+ lod_property);
+ }
+
+ num_args = 0;
+ args[num_args++] = context_ptr;
+ for (i = 0; i < num_coords; i++) {
+ args[num_args++] = coords[i];
+ }
+ if (layer) {
+ args[num_args++] = coords[layer];
+ }
+ if (sampler_bits & LP_SAMPLER_FUNC_SHADOW) {
+ args[num_args++] = coords[4];
+ }
+ if (sampler_bits & LP_SAMPLER_FUNC_OFFSETS) {
+ for (i = 0; i < num_offsets; i++) {
+ args[num_args++] = offsets[i];
+ }
+ }
+ if (sampler_bits & LP_SAMPLER_FUNC_LOD_BIAS) {
+ args[num_args++] = lod_bias;
+ }
+ else if (sampler_bits & LP_SAMPLER_FUNC_LOD_EXPLICIT) {
+ args[num_args++] = explicit_lod;
+ }
+ else if (sampler_bits & LP_SAMPLER_FUNC_EXPLICITDERIVS) {
+ for (i = 0; i < num_derivs; i++) {
+ args[num_args++] = derivs->ddx[i];
+ args[num_args++] = derivs->ddy[i];
+ }
+ }
+
+ assert(num_args <= LP_MAX_TEX_FUNC_ARGS);
+
+ tex_ret = LLVMBuildCall(builder, function, args, num_args, "");
+ bb = LLVMGetInsertBlock(builder);
+ inst = LLVMGetLastInstruction(bb);
+ LLVMSetInstructionCallConv(inst, LLVMFastCallConv);
+
+ for (i = 0; i < 4; i++) {
+ texel_out[i] = LLVMBuildExtractValue(gallivm->builder, tex_ret, i, "");
+ }
+}
+
+
+/**
+ * Build texture sampling code.
+ * Either via a function call or inline it directly.
+ */
+void
+lp_build_sample_soa(struct gallivm_state *gallivm,
+ const struct lp_static_texture_state *static_texture_state,
+ const struct lp_static_sampler_state *static_sampler_state,
+ struct lp_sampler_dynamic_state *dynamic_state,
+ struct lp_type type,
+ boolean is_fetch,
+ unsigned texture_index,
+ unsigned sampler_index,
+ LLVMValueRef context_ptr,
+ const LLVMValueRef *coords,
+ const LLVMValueRef *offsets,
+ const struct lp_derivatives *derivs, /* optional */
+ LLVMValueRef lod_bias, /* optional */
+ LLVMValueRef explicit_lod, /* optional */
+ enum lp_sampler_lod_property lod_property,
+ LLVMValueRef texel_out[4])
+{
+ if (USE_TEX_FUNC_CALL) {
+ lp_build_sample_soa_func(gallivm,
+ static_texture_state,
+ static_sampler_state,
+ dynamic_state,
+ type,
+ is_fetch,
+ texture_index,
+ sampler_index,
+ context_ptr,
+ coords,
+ offsets,
+ derivs,
+ lod_bias,
+ explicit_lod,
+ lod_property,
+ texel_out);
+ }
+ else {
+ lp_build_sample_soa_code(gallivm,
+ static_texture_state,
+ static_sampler_state,
+ dynamic_state,
+ type,
+ is_fetch,
+ texture_index,
+ sampler_index,
+ context_ptr,
+ coords,
+ offsets,
+ derivs,
+ lod_bias,
+ explicit_lod,
+ lod_property,
+ texel_out);
+ }
+}
+
+
void
lp_build_size_query_soa(struct gallivm_state *gallivm,
const struct lp_static_texture_state *static_state,