/** * @brief lp_build_fetch_rgba_aos_array * * \param format_desc describes format of the image we're fetching from * \param dst_type output type * \param base_ptr address of the pixel block (or the texel if uncompressed) * \param offset ptr offset */ LLVMValueRef lp_build_fetch_rgba_aos_array(struct gallivm_state *gallivm, const struct util_format_description *format_desc, struct lp_type dst_type, LLVMValueRef base_ptr, LLVMValueRef offset) { struct lp_build_context bld; LLVMBuilderRef builder = gallivm->builder; LLVMTypeRef src_elem_type, src_vec_type; LLVMValueRef ptr, res = NULL; struct lp_type src_type; memset(&src_type, 0, sizeof src_type); src_type.floating = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FLOAT; src_type.fixed = format_desc->channel[0].type == UTIL_FORMAT_TYPE_FIXED; src_type.sign = format_desc->channel[0].type != UTIL_FORMAT_TYPE_UNSIGNED; src_type.norm = format_desc->channel[0].normalized; src_type.width = format_desc->channel[0].size; src_type.length = format_desc->nr_channels; assert(src_type.length <= dst_type.length); src_elem_type = lp_build_elem_type(gallivm, src_type); src_vec_type = lp_build_vec_type(gallivm, src_type); /* Read whole vector from memory, unaligned */ if (!res) { ptr = LLVMBuildGEP(builder, base_ptr, &offset, 1, ""); ptr = LLVMBuildPointerCast(builder, ptr, LLVMPointerType(src_vec_type, 0), ""); res = LLVMBuildLoad(builder, ptr, ""); lp_set_load_alignment(res, src_type.width / 8); } /* Truncate doubles to float */ if (src_type.floating && src_type.width == 64) { src_type.width = 32; src_vec_type = lp_build_vec_type(gallivm, src_type); res = LLVMBuildFPTrunc(builder, res, src_vec_type, ""); } /* Expand to correct length */ if (src_type.length < dst_type.length) { res = lp_build_pad_vector(gallivm, res, src_type, dst_type.length); src_type.length = dst_type.length; } /* Convert to correct format */ lp_build_conv(gallivm, src_type, dst_type, &res, 1, &res, 1); /* Swizzle it */ lp_build_context_init(&bld, gallivm, dst_type); return lp_build_format_swizzle_aos(format_desc, &bld, res); }
/** * Fetch a pixel into a 4 float AoS. * * \param format_desc describes format of the image we're fetching from * \param ptr address of the pixel block (or the texel if uncompressed) * \param i, j the sub-block pixel coordinates. For non-compressed formats * these will always be (0, 0). * \return a 4 element vector with the pixel's RGBA values. */ LLVMValueRef lp_build_fetch_rgba_aos(struct gallivm_state *gallivm, const struct util_format_description *format_desc, struct lp_type type, LLVMValueRef base_ptr, LLVMValueRef offset, LLVMValueRef i, LLVMValueRef j) { LLVMBuilderRef builder = gallivm->builder; unsigned num_pixels = type.length / 4; struct lp_build_context bld; assert(type.length <= LP_MAX_VECTOR_LENGTH); assert(type.length % 4 == 0); lp_build_context_init(&bld, gallivm, type); /* * Trivial case * * The format matches the type (apart of a swizzle) so no need for * scaling or converting. */ if (format_matches_type(format_desc, type) && format_desc->block.bits <= type.width * 4 && util_is_power_of_two(format_desc->block.bits)) { LLVMValueRef packed; /* * The format matches the type (apart of a swizzle) so no need for * scaling or converting. */ packed = lp_build_gather(gallivm, type.length/4, format_desc->block.bits, type.width*4, base_ptr, offset); assert(format_desc->block.bits <= type.width * type.length); packed = LLVMBuildBitCast(gallivm->builder, packed, lp_build_vec_type(gallivm, type), ""); return lp_build_format_swizzle_aos(format_desc, &bld, packed); } /* * Bit arithmetic */ if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN && (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB || format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) && format_desc->block.width == 1 && format_desc->block.height == 1 && util_is_power_of_two(format_desc->block.bits) && format_desc->block.bits <= 32 && format_desc->is_bitmask && !format_desc->is_mixed && (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED || format_desc->channel[1].type == UTIL_FORMAT_TYPE_UNSIGNED)) { LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4]; LLVMValueRef res; unsigned k; /* * Unpack a pixel at a time into a <4 x float> RGBA vector */ for (k = 0; k < num_pixels; ++k) { LLVMValueRef packed; packed = lp_build_gather_elem(gallivm, num_pixels, format_desc->block.bits, 32, base_ptr, offset, k); tmps[k] = lp_build_unpack_arith_rgba_aos(gallivm, format_desc, packed); } /* * Type conversion. * * TODO: We could avoid floating conversion for integer to * integer conversions. */ if (gallivm_debug & GALLIVM_DEBUG_PERF && !type.floating) { debug_printf("%s: unpacking %s with floating point\n", __FUNCTION__, format_desc->short_name); } lp_build_conv(gallivm, lp_float32_vec4_type(), type, tmps, num_pixels, &res, 1); return lp_build_format_swizzle_aos(format_desc, &bld, res); } /* * YUV / subsampled formats */ if (format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) { struct lp_type tmp_type; LLVMValueRef tmp; memset(&tmp_type, 0, sizeof tmp_type); tmp_type.width = 8; tmp_type.length = num_pixels * 4; tmp_type.norm = TRUE; tmp = lp_build_fetch_subsampled_rgba_aos(gallivm, format_desc, num_pixels, base_ptr, offset, i, j); lp_build_conv(gallivm, tmp_type, type, &tmp, 1, &tmp, 1); return tmp; } /* * Fallback to util_format_description::fetch_rgba_8unorm(). */ if (format_desc->fetch_rgba_8unorm && !type.floating && type.width == 8 && !type.sign && type.norm) { /* * Fallback to calling util_format_description::fetch_rgba_8unorm. * * This is definitely not the most efficient way of fetching pixels, as * we miss the opportunity to do vectorization, but this it is a * convenient for formats or scenarios for which there was no opportunity * or incentive to optimize. */ LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(gallivm->builder))); char name[256]; LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context); LLVMTypeRef pi8t = LLVMPointerType(i8t, 0); LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context); LLVMValueRef function; LLVMValueRef tmp_ptr; LLVMValueRef tmp; LLVMValueRef res; LLVMValueRef callee; unsigned k; util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_8unorm", format_desc->short_name); if (gallivm_debug & GALLIVM_DEBUG_PERF) { debug_printf("%s: falling back to %s\n", __FUNCTION__, name); } /* * Declare and bind format_desc->fetch_rgba_8unorm(). */ function = LLVMGetNamedFunction(module, name); if (!function) { /* * Function to call looks like: * fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j) */ LLVMTypeRef ret_type; LLVMTypeRef arg_types[4]; LLVMTypeRef function_type; ret_type = LLVMVoidTypeInContext(gallivm->context); arg_types[0] = pi8t; arg_types[1] = pi8t; arg_types[2] = i32t; arg_types[3] = i32t; function_type = LLVMFunctionType(ret_type, arg_types, Elements(arg_types), 0); function = LLVMAddFunction(module, name, function_type); LLVMSetFunctionCallConv(function, LLVMCCallConv); LLVMSetLinkage(function, LLVMExternalLinkage); assert(LLVMIsDeclaration(function)); } /* make const pointer for the C fetch_rgba_float function */ callee = lp_build_const_int_pointer(gallivm, func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm)); /* cast the callee pointer to the function's type */ function = LLVMBuildBitCast(builder, callee, LLVMTypeOf(function), "cast callee"); tmp_ptr = lp_build_alloca(gallivm, i32t, ""); res = LLVMGetUndef(LLVMVectorType(i32t, num_pixels)); /* * Invoke format_desc->fetch_rgba_8unorm() for each pixel and insert the result * in the SoA vectors. */ for (k = 0; k < num_pixels; ++k) { LLVMValueRef index = lp_build_const_int32(gallivm, k); LLVMValueRef args[4]; args[0] = LLVMBuildBitCast(builder, tmp_ptr, pi8t, ""); args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels, base_ptr, offset, k); if (num_pixels == 1) { args[2] = i; args[3] = j; } else { args[2] = LLVMBuildExtractElement(builder, i, index, ""); args[3] = LLVMBuildExtractElement(builder, j, index, ""); } LLVMBuildCall(builder, function, args, Elements(args), ""); tmp = LLVMBuildLoad(builder, tmp_ptr, ""); if (num_pixels == 1) { res = tmp; } else { res = LLVMBuildInsertElement(builder, res, tmp, index, ""); } } /* Bitcast from <n x i32> to <4n x i8> */ res = LLVMBuildBitCast(builder, res, bld.vec_type, ""); return res; } /* * Fallback to util_format_description::fetch_rgba_float(). */ if (format_desc->fetch_rgba_float) { /* * Fallback to calling util_format_description::fetch_rgba_float. * * This is definitely not the most efficient way of fetching pixels, as * we miss the opportunity to do vectorization, but this it is a * convenient for formats or scenarios for which there was no opportunity * or incentive to optimize. */ LLVMModuleRef module = LLVMGetGlobalParent(LLVMGetBasicBlockParent(LLVMGetInsertBlock(builder))); char name[256]; LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context); LLVMTypeRef f32x4t = LLVMVectorType(f32t, 4); LLVMTypeRef pf32t = LLVMPointerType(f32t, 0); LLVMTypeRef pi8t = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0); LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context); LLVMValueRef function; LLVMValueRef tmp_ptr; LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4]; LLVMValueRef res; LLVMValueRef callee; unsigned k; util_snprintf(name, sizeof name, "util_format_%s_fetch_rgba_float", format_desc->short_name); if (gallivm_debug & GALLIVM_DEBUG_PERF) { debug_printf("%s: falling back to %s\n", __FUNCTION__, name); } /* * Declare and bind format_desc->fetch_rgba_float(). */ function = LLVMGetNamedFunction(module, name); if (!function) { /* * Function to call looks like: * fetch(float *dst, const uint8_t *src, unsigned i, unsigned j) */ LLVMTypeRef ret_type; LLVMTypeRef arg_types[4]; LLVMTypeRef function_type; ret_type = LLVMVoidTypeInContext(gallivm->context); arg_types[0] = pf32t; arg_types[1] = pi8t; arg_types[2] = i32t; arg_types[3] = i32t; function_type = LLVMFunctionType(ret_type, arg_types, Elements(arg_types), 0); function = LLVMAddFunction(module, name, function_type); LLVMSetFunctionCallConv(function, LLVMCCallConv); LLVMSetLinkage(function, LLVMExternalLinkage); assert(LLVMIsDeclaration(function)); } /* Note: we're using this casting here instead of LLVMAddGlobalMapping() * to work around a bug in LLVM 2.6. */ /* make const pointer for the C fetch_rgba_float function */ callee = lp_build_const_int_pointer(gallivm, func_to_pointer((func_pointer) format_desc->fetch_rgba_float)); /* cast the callee pointer to the function's type */ function = LLVMBuildBitCast(builder, callee, LLVMTypeOf(function), "cast callee"); tmp_ptr = lp_build_alloca(gallivm, f32x4t, ""); /* * Invoke format_desc->fetch_rgba_float() for each pixel and insert the result * in the SoA vectors. */ for (k = 0; k < num_pixels; ++k) { LLVMValueRef args[4]; args[0] = LLVMBuildBitCast(builder, tmp_ptr, pf32t, ""); args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels, base_ptr, offset, k); if (num_pixels == 1) { args[2] = i; args[3] = j; } else { LLVMValueRef index = lp_build_const_int32(gallivm, k); args[2] = LLVMBuildExtractElement(builder, i, index, ""); args[3] = LLVMBuildExtractElement(builder, j, index, ""); } LLVMBuildCall(builder, function, args, Elements(args), ""); tmps[k] = LLVMBuildLoad(builder, tmp_ptr, ""); } lp_build_conv(gallivm, lp_float32_vec4_type(), type, tmps, num_pixels, &res, 1); return res; } assert(0); return lp_build_undef(gallivm, type); }
/** * @brief lp_build_fetch_rgba_aos_array * * \param format_desc describes format of the image we're fetching from * \param dst_type output type * \param base_ptr address of the pixel block (or the texel if uncompressed) * \param offset ptr offset */ LLVMValueRef lp_build_fetch_rgba_aos_array(struct gallivm_state *gallivm, const struct util_format_description *format_desc, struct lp_type dst_type, LLVMValueRef base_ptr, LLVMValueRef offset) { struct lp_build_context bld; LLVMBuilderRef builder = gallivm->builder; LLVMTypeRef src_vec_type; LLVMValueRef ptr, res = NULL; struct lp_type src_type; boolean pure_integer = format_desc->channel[0].pure_integer; struct lp_type tmp_type; lp_type_from_format_desc(&src_type, format_desc); assert(src_type.length <= dst_type.length); src_vec_type = lp_build_vec_type(gallivm, src_type); /* Read whole vector from memory, unaligned */ ptr = LLVMBuildGEP(builder, base_ptr, &offset, 1, ""); ptr = LLVMBuildPointerCast(builder, ptr, LLVMPointerType(src_vec_type, 0), ""); res = LLVMBuildLoad(builder, ptr, ""); LLVMSetAlignment(res, src_type.width / 8); /* Truncate doubles to float */ if (src_type.floating && src_type.width == 64) { src_type.width = 32; src_vec_type = lp_build_vec_type(gallivm, src_type); res = LLVMBuildFPTrunc(builder, res, src_vec_type, ""); } /* Expand to correct length */ if (src_type.length < dst_type.length) { res = lp_build_pad_vector(gallivm, res, dst_type.length); src_type.length = dst_type.length; } tmp_type = dst_type; if (pure_integer) { /* some callers expect (fake) floats other real ints. */ tmp_type.floating = 0; tmp_type.sign = src_type.sign; } /* Convert to correct format */ lp_build_conv(gallivm, src_type, tmp_type, &res, 1, &res, 1); /* Swizzle it */ lp_build_context_init(&bld, gallivm, tmp_type); res = lp_build_format_swizzle_aos(format_desc, &bld, res); /* Bitcast to floats (for pure integers) when requested */ if (pure_integer && dst_type.floating) { res = LLVMBuildBitCast(builder, res, lp_build_vec_type(gallivm, dst_type), ""); } return res; }