rsx: Optimize vertex decoder to generate simpler code

- Significantly improves compilation speed by simplifying most of the code and doing something similar to LICM.
  * Actual decoding is now vectorized and performed in one step rather than in a loop.
  * Switches inside loops are removed and replaced with simple comparison. Generates much nicer (and smaller) GCN bytecode.
This commit is contained in:
kd-11 2020-09-25 21:48:10 +03:00 committed by kd-11
parent 259844f4f3
commit a14a358b73

View File

@ -213,44 +213,17 @@ namespace glsl
static void insert_vertex_input_fetch(std::stringstream& OS, glsl_rules rules, bool glsl4_compliant=true)
{
std::string vertex_id_name = (rules == glsl_rules_opengl4) ? "gl_VertexID" : "gl_VertexIndex";
std::string vertex_id_name = (rules != glsl_rules_spirv) ? "gl_VertexID" : "gl_VertexIndex";
//Actually decode a vertex attribute from a raw byte stream
OS <<
"struct attribute_desc\n"
"{\n"
" uint type;\n"
" uint attribute_size;\n"
" uint starting_offset;\n"
" uint stride;\n"
" uint frequency;\n"
" bool swap_bytes;\n"
" bool is_volatile;\n"
" bool modulo;\n"
"};\n\n"
"uint get_bits(const in uint x, const in uint y, const in uint z, const in uint w, const in bool swap)\n"
"{\n"
" if (swap) return (w | z << 8 | y << 16 | x << 24);\n"
" return (x | y << 8 | z << 16 | w << 24);\n"
"}\n\n"
"uint get_bits(const in uint x, const in uint y, const in bool swap)\n"
"{\n"
" if (swap) return (y | x << 8);\n"
" return (x | y << 8);\n"
"}\n\n"
"int preserve_sign_s16(const in uint bits)\n"
"{\n"
" //convert raw 16 bit value into signed 32-bit integer counterpart\n"
" if ((bits & 0x8000u) == 0)\n"
" return int(bits);\n"
" else\n"
" return int(bits | 0xFFFF0000u);\n"
"}\n\n"
"#define get_s16(v, s) preserve_sign_s16(get_bits(v, s))\n\n";
"#define VTX_FMT_SNORM16 0\n"
"#define VTX_FMT_FLOAT32 1\n"
"#define VTX_FMT_FLOAT16 2\n"
"#define VTX_FMT_UNORM8 3\n"
"#define VTX_FMT_SINT16 4\n"
"#define VTX_FMT_COMP32 5\n"
"#define VTX_FMT_UINT8 6\n\n";
// For intel GPUs which cannot access vectors in indexed mode (driver bug? or glsl version too low?)
// Note: Tested on Mesa iris with HD 530 and compilant path works fine, may be a bug on Windows proprietary drivers
@ -266,9 +239,8 @@ namespace glsl
" case 2: vector.z = scalar; return;\n"
" case 3: vector.w = scalar; return;\n"
" }\n"
"}\n";
"}\n\n"
OS <<
"uint ref(const in uvec4 vector, const in int index)\n"
"{\n"
" switch(index)\n"
@ -278,7 +250,7 @@ namespace glsl
" case 2: return vector.z;\n"
" case 3: return vector.w;\n"
" }\n"
"}\n";
"}\n\n";
}
else
{
@ -288,73 +260,107 @@ namespace glsl
}
OS <<
"struct attribute_desc\n"
"{\n"
" uint type;\n"
" uint attribute_size;\n"
" uint starting_offset;\n"
" uint stride;\n"
" uint frequency;\n"
" bool swap_bytes;\n"
" bool is_volatile;\n"
" bool modulo;\n"
"};\n\n"
"uint gen_bits(const in uint x, const in uint y, const in uint z, const in uint w, const in bool swap)\n"
"{\n"
" return (swap) ?\n"
" bitfieldInsert(bitfieldInsert(bitfieldInsert(w, z, 8, 8), y, 16, 8), x, 24, 8) :\n"
" bitfieldInsert(bitfieldInsert(bitfieldInsert(x, y, 8, 8), z, 16, 8), w, 24, 8);\n"
"}\n\n"
"uint gen_bits(const in uint x, const in uint y, const in bool swap)\n"
"{\n"
" return (swap)? bitfieldInsert(y, x, 8, 8) : bitfieldInsert(x, y, 8, 8);\n"
"}\n\n"
"vec4 sext(const in ivec4 bits)\n"
"{\n"
" // convert raw 16 bit values into signed 32-bit float4 counterpart\n"
" bvec4 sign_check = lessThan(bits, ivec4(0x8000));\n"
" return _select(bits | ivec4(0xFFFF0000), bits, sign_check);\n"
"}\n\n"
"float sext(const in int bits)\n"
"{\n"
" return (bits < 0x8000) ? float(bits) : float(bits | 0xFFFF0000); \n"
"}\n\n"
"vec4 fetch_attribute(const in attribute_desc desc, const in int vertex_id, usamplerBuffer input_stream)\n"
"{\n"
" vec4 result = vec4(0., 0., 0., 1.);\n"
" vec4 scale = vec4(1.);\n"
" bool reverse_order = false;\n"
"\n"
" const int elem_size_table[] = { 2, 4, 2, 1, 2, 4, 1 };\n"
" const float scaling_table[] = { 32768., 1., 1., 255., 1., 32767., 1. };\n"
" const int elem_size = elem_size_table[desc.type];\n"
" uvec4 tmp;\n"
"\n"
" int n;\n"
" int i = int((vertex_id * desc.stride) + desc.starting_offset);\n"
"\n"
" const vec4 scale = scaling_table[desc.type].xxxx;\n\n"
" uvec4 tmp, result = uvec4(0u);\n"
" vec4 ret;\n"
" int n, i = int((vertex_id * desc.stride) + desc.starting_offset);\n\n"
" for (n = 0; n < desc.attribute_size; n++)\n"
" {\n"
" tmp.x = texelFetch(input_stream, i++).x;\n"
" if (elem_size == 2)\n"
" {\n"
" tmp.y = texelFetch(input_stream, i++).x;\n"
" tmp.x = get_bits(tmp.x, tmp.y, desc.swap_bytes);\n"
" tmp.x = gen_bits(tmp.x, tmp.y, desc.swap_bytes);\n"
" }\n"
" else if (elem_size == 4)\n"
" {\n"
" tmp.y = texelFetch(input_stream, i++).x;\n"
" tmp.z = texelFetch(input_stream, i++).x;\n"
" tmp.w = texelFetch(input_stream, i++).x;\n"
" tmp.x = get_bits(tmp.x, tmp.y, tmp.z, tmp.w, desc.swap_bytes);\n"
" }\n"
"\n"
" switch (desc.type)\n"
" {\n"
" case 0:\n"
" //signed normalized 16-bit\n"
" mov(scale, n, 32767.);\n"
" case 4:\n"
" //signed word\n"
" mov(result, n, preserve_sign_s16(tmp.x));\n"
" break;\n"
" case 1:\n"
" //float\n"
" mov(result, n, uintBitsToFloat(tmp.x));\n"
" break;\n"
" case 2:\n"
" //half\n"
" mov(result, n, unpackHalf2x16(tmp.x).x);\n"
" break;\n"
" case 3:\n"
" //unsigned byte\n"
" mov(scale, n, 255.);\n"
" case 6:\n"
" //ub256\n"
" mov(result, n, tmp.x);\n"
" reverse_order = desc.swap_bytes;\n"
" break;\n"
" case 5:\n"
" //cmp\n"
" result.x = preserve_sign_s16((tmp.x & 0x7FFu) << 5);\n"
" result.y = preserve_sign_s16(((tmp.x >> 11) & 0x7FFu) << 5);\n"
" result.z = preserve_sign_s16(((tmp.x >> 22) & 0x3FFu) << 6);\n"
" result.w = 1.;\n"
" scale = vec4(32767., 32767., 32767., 1.);\n"
" break;\n"
" }\n"
" tmp.x = gen_bits(tmp.x, tmp.y, tmp.z, tmp.w, desc.swap_bytes);\n"
" }\n\n"
" mov(result, n, tmp.x);\n"
" }\n\n"
" // Actual decoding step is done in vector space, outside the loop\n"
" if (desc.type == VTX_FMT_SNORM16 || desc.type == VTX_FMT_SINT16)\n"
" {\n"
" ret = sext(ivec4(result));\n"
" }\n"
"\n"
" result /= scale;\n"
" return (reverse_order)? result.wzyx: result;\n"
" else if (desc.type == VTX_FMT_FLOAT32)\n"
" {\n"
" ret = uintBitsToFloat(result);\n"
" }\n"
" else if (desc.type == VTX_FMT_FLOAT16)\n"
" {\n"
" tmp.x = bitfieldInsert(result.x, result.y, 16, 16);\n"
" tmp.y = bitfieldInsert(result.z, result.w, 16, 16);\n"
" ret.xy = unpackHalf2x16(tmp.x);\n"
" ret.zw = unpackHalf2x16(tmp.y);\n"
" }\n"
" else if (desc.type == VTX_FMT_UINT8 || desc.type == VTX_FMT_UNORM8)\n"
" {\n"
" ret = vec4(desc.swap_bytes? result.wzyx : result);\n"
" }\n"
" else //if (desc.type == VTX_FMT_COMP32)\n"
" {\n"
" result = uvec4(bitfieldExtract(result.x, 0, 11),\n"
" bitfieldExtract(result.x, 11, 11),\n"
" bitfieldExtract(result.x, 22, 10),\n"
" uint(scale.x));\n"
" ret = sext(ivec4(result) << ivec4(5, 5, 6, 0));\n"
" }\n\n"
" if (desc.attribute_size < 4)\n"
" {\n"
" ret.w = scale.x;\n"
" }\n\n"
" return ret / scale; \n"
"}\n\n"
"attribute_desc fetch_desc(const in int location)\n"
@ -368,7 +374,7 @@ namespace glsl
" // [32-60] starting offset\n"
" // [60-61] swap bytes flag\n"
" // [61-62] volatile flag\n"
" // [62-63] modulo enable flag\n";
" // [62-63] modulo enable flag\n\n";
if (rules == glsl_rules_opengl4)
{
@ -378,7 +384,7 @@ namespace glsl
" int sub_block = (location & 1) << 1;\n"
" uvec2 attrib = uvec2(\n"
" ref(input_attributes_blob[block], sub_block + 0),\n"
" ref(input_attributes_blob[block], sub_block + 1));\n";
" ref(input_attributes_blob[block], sub_block + 1));\n\n";
}
else
{
@ -389,25 +395,20 @@ namespace glsl
OS <<
" attribute_desc result;\n"
" result.stride = attrib.x & 0xFFu;\n"
" result.frequency = (attrib.x >> 8) & 0xFFFFu;\n"
" result.type = (attrib.x >> 24) & 0x7u;\n"
" result.attribute_size = (attrib.x >> 27) & 0x7u;\n"
" result.starting_offset = (attrib.y & 0x1FFFFFFFu);\n"
" result.swap_bytes = ((attrib.y >> 29) & 0x1u) != 0;\n"
" result.is_volatile = ((attrib.y >> 30) & 0x1u) != 0;\n"
" result.modulo = ((attrib.y >> 31) & 0x1u) != 0;\n"
" result.stride = bitfieldExtract(attrib.x, 0, 8);\n"
" result.frequency = bitfieldExtract(attrib.x, 8, 16);\n"
" result.type = bitfieldExtract(attrib.x, 24, 3);\n"
" result.attribute_size = bitfieldExtract(attrib.x, 27, 3);\n"
" result.starting_offset = bitfieldExtract(attrib.y, 0, 29);\n"
" result.swap_bytes = bitfieldExtract(attrib.y, 29, 1) != 0;\n"
" result.is_volatile = bitfieldExtract(attrib.y, 30, 1) != 0;\n"
" result.modulo = bitfieldExtract(attrib.y, 31, 1) != 0;\n"
" return result;\n"
"}\n\n"
"vec4 read_location(const in int location)\n"
"{\n"
" attribute_desc desc = fetch_desc(location);\n"
" if (desc.attribute_size == 0)\n"
" {\n"
" //default value\n"
" return vec4(0., 0., 0., 1.);\n"
" }\n\n"
" int vertex_id = " << vertex_id_name << " - int(vertex_base_index);\n"
" if (desc.frequency == 0)\n"
" {\n"
@ -421,8 +422,8 @@ namespace glsl
" else\n"
" {\n"
" vertex_id /= int(desc.frequency); \n"
" }\n"
"\n"
" }\n\n"
" if (desc.is_volatile)\n"
" return fetch_attribute(desc, vertex_id, volatile_input_stream);\n"
" else\n"