From 5941e7e5ccbad08dafb9a5dc6caf0c97e11d8eb3 Mon Sep 17 00:00:00 2001 From: Nigel Tao Date: Sun, 24 Jan 2021 21:58:00 +1100 Subject: [PATCH] Rename sse128/sse128i to x86_sse42/x86_m128i --- internal/cgen/builtin.go | 2 +- internal/cgen/expr.go | 2 +- internal/cgen/statement.go | 8 ++-- lang/ast/ast.go | 2 +- lang/builtin/builtin.go | 50 ++++++++++---------- lang/check/resolve.go | 4 +- lang/check/type.go | 10 ++-- lang/token/list.go | 16 ++++--- std/png/decode_filter_sse128.wuffs | 74 +++++++++++++++--------------- 9 files changed, 85 insertions(+), 83 deletions(-) diff --git a/internal/cgen/builtin.go b/internal/cgen/builtin.go index 90101d94..6f21abd8 100644 --- a/internal/cgen/builtin.go +++ b/internal/cgen/builtin.go @@ -141,7 +141,7 @@ func (g *gen) writeBuiltinCall(b *buffer, n *a.Expr, sideEffectsOnly bool, depth b.writes("&empty_io_buffer") return nil } - case t.IDSSE128I: + case t.IDX86M128I: return g.writeBuiltinCPUArch(b, recv, method.Ident(), n.Args(), sideEffectsOnly, depth) } } diff --git a/internal/cgen/expr.go b/internal/cgen/expr.go index bffba4ba..9d6fdf60 100644 --- a/internal/cgen/expr.go +++ b/internal/cgen/expr.go @@ -579,7 +579,7 @@ var cTypeNames = [...]string{ t.IDTokenReader: "wuffs_base__token_buffer*", t.IDTokenWriter: "wuffs_base__token_buffer*", - t.IDSSE128I: "__m128i", + t.IDX86M128I: "__m128i", } const noSuchCOperator = " no_such_C_operator " diff --git a/internal/cgen/statement.go b/internal/cgen/statement.go index 3f53bd4b..b77b55cf 100644 --- a/internal/cgen/statement.go +++ b/internal/cgen/statement.go @@ -258,18 +258,18 @@ func (g *gen) writeStatementChoose(b *buffer, n *a.Choose, depth uint32) error { } func cpuArchCNames(asserts []*a.Node) (caMacro string, caName string, caAttribute string) { - sse128 := false + sse42 := false for _, o := range asserts { o := o.AsAssert() if !o.IsChooseCPUArch() { continue } switch o.Condition().RHS().AsExpr().Ident() { - case t.IDSSE128: - sse128 = true + case t.IDX86SSE42: + sse42 = true } } - if sse128 { + if sse42 { return "X86_64", "sse128", "__attribute__((target(\"sse4.2\")))" } return "", "", "" diff --git a/lang/ast/ast.go b/lang/ast/ast.go index a6a5b243..bb4c562a 100644 --- a/lang/ast/ast.go +++ b/lang/ast/ast.go @@ -455,7 +455,7 @@ func (n *Assert) IsChooseCPUArch() bool { return false } switch rhs.Ident() { - case t.IDSSE128: + case t.IDX86SSE42, t.IDX86AVX2: return true } return false diff --git a/lang/builtin/builtin.go b/lang/builtin/builtin.go index 5f2f3bdd..596f3d30 100644 --- a/lang/builtin/builtin.go +++ b/lang/builtin/builtin.go @@ -243,7 +243,7 @@ var Types = []string{ // ---- - "sse128_i", + "x86_m128i", } var Funcs = []string{ @@ -532,35 +532,35 @@ var Funcs = []string{ "pixel_swizzler.swizzle_interleaved_transparent_black!(" + "dst: slice u8, dst_palette: slice u8, num_pixels: u64) u64", - // ---- sse128_i + // ---- x86_m128i - "sse128_i.load_u32!(a: u32)", - "sse128_i.load_u64!(a: u64)", - "sse128_i.load_slice128!(a: slice base.u8)", - "sse128_i.truncate_u32() u32", - "sse128_i.truncate_u64() u64", - "sse128_i.store_slice128!(a: slice base.u8)", + "x86_m128i.load_u32!(a: u32)", + "x86_m128i.load_u64!(a: u64)", + "x86_m128i.load_slice128!(a: slice base.u8)", + "x86_m128i.truncate_u32() u32", + "x86_m128i.truncate_u64() u64", + "x86_m128i.store_slice128!(a: slice base.u8)", // TODO: generate these methods automatically? - "sse128_i.create_mm_set1_epi8(a: u8) sse128_i", + "x86_m128i.create_mm_set1_epi8(a: u8) x86_m128i", - "sse128_i._mm_abs_epi16() sse128_i", - "sse128_i._mm_add_epi16(b: sse128_i) sse128_i", - "sse128_i._mm_add_epi8(b: sse128_i) sse128_i", - "sse128_i._mm_and_si128(b: sse128_i) sse128_i", - "sse128_i._mm_avg_epu8(b: sse128_i) sse128_i", - "sse128_i._mm_blend_epi16(b: sse128_i, imm8: u32) sse128_i", - "sse128_i._mm_blendv_epi8(b: sse128_i, mask: sse128_i) sse128_i", - "sse128_i._mm_cmpeq_epi16(b: sse128_i) sse128_i", - "sse128_i._mm_min_epi16(b: sse128_i) sse128_i", - "sse128_i._mm_packus_epi16(b: sse128_i) sse128_i", - "sse128_i._mm_slli_si128(imm8: u32) sse128_i", - "sse128_i._mm_srli_si128(imm8: u32) sse128_i", - "sse128_i._mm_sub_epi16(b: sse128_i) sse128_i", - "sse128_i._mm_sub_epi8(b: sse128_i) sse128_i", - "sse128_i._mm_unpacklo_epi8(b: sse128_i) sse128_i", - "sse128_i._mm_xor_si128(b: sse128_i) sse128_i", + "x86_m128i._mm_abs_epi16() x86_m128i", + "x86_m128i._mm_add_epi16(b: x86_m128i) x86_m128i", + "x86_m128i._mm_add_epi8(b: x86_m128i) x86_m128i", + "x86_m128i._mm_and_si128(b: x86_m128i) x86_m128i", + "x86_m128i._mm_avg_epu8(b: x86_m128i) x86_m128i", + "x86_m128i._mm_blend_epi16(b: x86_m128i, imm8: u32) x86_m128i", + "x86_m128i._mm_blendv_epi8(b: x86_m128i, mask: x86_m128i) x86_m128i", + "x86_m128i._mm_cmpeq_epi16(b: x86_m128i) x86_m128i", + "x86_m128i._mm_min_epi16(b: x86_m128i) x86_m128i", + "x86_m128i._mm_packus_epi16(b: x86_m128i) x86_m128i", + "x86_m128i._mm_slli_si128(imm8: u32) x86_m128i", + "x86_m128i._mm_srli_si128(imm8: u32) x86_m128i", + "x86_m128i._mm_sub_epi16(b: x86_m128i) x86_m128i", + "x86_m128i._mm_sub_epi8(b: x86_m128i) x86_m128i", + "x86_m128i._mm_unpacklo_epi8(b: x86_m128i) x86_m128i", + "x86_m128i._mm_xor_si128(b: x86_m128i) x86_m128i", } var Interfaces = []string{ diff --git a/lang/check/resolve.go b/lang/check/resolve.go index c9cd8f7f..c1e2c31c 100644 --- a/lang/check/resolve.go +++ b/lang/check/resolve.go @@ -75,7 +75,7 @@ var ( typeExprDecodeFrameOptions = a.NewTypeExpr(0, t.IDBase, t.IDDecodeFrameOptions, nil, nil, nil) - typeExprSSE128I = a.NewTypeExpr(0, t.IDBase, t.IDSSE128I, nil, nil, nil) + typeExprX86M128I = a.NewTypeExpr(0, t.IDBase, t.IDX86M128I, nil, nil, nil) typeExprSliceU8 = a.NewTypeExpr(t.IDSlice, 0, 0, nil, nil, typeExprU8) typeExprTableU8 = a.NewTypeExpr(t.IDTable, 0, 0, nil, nil, typeExprU8) @@ -125,7 +125,7 @@ var builtInTypeMap = typeMap{ t.IDDecodeFrameOptions: typeExprDecodeFrameOptions, - t.IDSSE128I: typeExprSSE128I, + t.IDX86M128I: typeExprX86M128I, } func (c *Checker) parseBuiltInFuncs(m map[t.QQID]*a.Func, ss []string) error { diff --git a/lang/check/type.go b/lang/check/type.go index cde48c3c..90a2875f 100644 --- a/lang/check/type.go +++ b/lang/check/type.go @@ -25,7 +25,7 @@ import ( type cpuArchBits uint32 const ( - cpuArchBitsSSE128 = cpuArchBits(0x00000001) + cpuArchBitsX86SSE42 = cpuArchBits(0x00000001) ) func calcCPUArchBits(n *a.Func) (ret cpuArchBits) { @@ -35,8 +35,8 @@ func calcCPUArchBits(n *a.Func) (ret cpuArchBits) { continue } switch o.Condition().RHS().AsExpr().Ident() { - case t.IDSSE128: - ret |= cpuArchBitsSSE128 + case t.IDX86SSE42: + ret |= cpuArchBitsX86SSE42 } } return ret @@ -46,8 +46,8 @@ func (q *checker) tcheckCPUArchBits(cab cpuArchBits, typ *a.TypeExpr) error { if qid := typ.Innermost().QID(); qid[0] == t.IDBase { need := cpuArchBits(0) switch qid[1] { - case t.IDSSE128I: - need = cpuArchBitsSSE128 + case t.IDX86M128I: + need = cpuArchBitsX86SSE42 } if (cab & need) != need { return fmt.Errorf("check: missing cpu_arch for %q", typ.Innermost().Str(q.tm)) diff --git a/lang/token/list.go b/lang/token/list.go index 0ab1da5f..87916d86 100644 --- a/lang/token/list.go +++ b/lang/token/list.go @@ -661,14 +661,14 @@ const ( // -------- 0x300 block. minBuiltInCPUArch = 0x300 - maxBuiltInCPUArch = 0x32F + maxBuiltInCPUArch = 0x33F - // 0x30? are reserved for NEON. + // [0x300 ..= 0x31F] are reserved for ARM/NEON. - IDSSE128 = ID(0x310) - IDSSE128I = ID(0x311) + IDX86SSE42 = ID(0x320) + IDX86AVX2 = ID(0x321) - // 0x32? are reserved for AVX256. + IDX86M128I = ID(0x330) // -------- @@ -1072,8 +1072,10 @@ var builtInsByID = [nBuiltInIDs]string{ // -------- 0x300 block. - IDSSE128: "sse128", - IDSSE128I: "sse128_i", + IDX86SSE42: "x86_sse42", + IDX86AVX2: "x86_avx2", + + IDX86M128I: "x86_m128i", IDLoadU32: "load_u32", IDLoadU64: "load_u64", diff --git a/std/png/decode_filter_sse128.wuffs b/std/png/decode_filter_sse128.wuffs index 1e911f88..ac5661c4 100644 --- a/std/png/decode_filter_sse128.wuffs +++ b/std/png/decode_filter_sse128.wuffs @@ -20,30 +20,30 @@ // than the non-SIMD one. // // pri func decoder.filter_1_distance_3_sse128!(curr: slice base.u8), -// choose cpu_arch >= sse128, +// choose cpu_arch >= x86_sse42, // { // var c : slice base.u8 -// var x128 : base.sse128_i -// var a128 : base.sse128_i +// var x128 : base.x86_m128i +// var a128 : base.x86_m128i // // iterate (c = args.curr)(length: 4, advance: 3, unroll: 2) { // x128.load_u32!(a: c.peek_u32le()) -// x128 = x128._mm_add_epi8!(b: a128) +// x128 = x128._mm_add_epi8(b: a128) // a128 = x128 // c.poke_u24le!(a: x128.truncate_u32()) // } else (length: 3, advance: 3, unroll: 1) { // x128.load_u32!(a: c.peek_u24le_as_u32()) -// x128 = x128._mm_add_epi8!(b: a128) +// x128 = x128._mm_add_epi8(b: a128) // c.poke_u24le!(a: x128.truncate_u32()) // } // } pri func decoder.filter_1_distance_4_sse128!(curr: slice base.u8), - choose cpu_arch >= sse128, + choose cpu_arch >= x86_sse42, { var c : slice base.u8 - var x128 : base.sse128_i - var a128 : base.sse128_i + var x128 : base.x86_m128i + var a128 : base.x86_m128i iterate (c = args.curr)(length: 4, advance: 4, unroll: 2) { x128.load_u32!(a: c.peek_u32le()) @@ -61,21 +61,21 @@ pri func decoder.filter_1_distance_4_sse128!(curr: slice base.u8), // 3, distance = 3) doesn't actually bench faster than the non-SIMD one. // // pri func decoder.filter_3_distance_3_sse128!(curr: slice base.u8, prev: slice base.u8), -// choose cpu_arch >= sse128, +// choose cpu_arch >= x86_sse42, // { // etc // } pri func decoder.filter_3_distance_4_sse128!(curr: slice base.u8, prev: slice base.u8), - choose cpu_arch >= sse128, + choose cpu_arch >= x86_sse42, { var c : slice base.u8 var p : slice base.u8 - var x128 : base.sse128_i - var a128 : base.sse128_i - var b128 : base.sse128_i - var p128 : base.sse128_i - var k128 : base.sse128_i + var x128 : base.x86_m128i + var a128 : base.x86_m128i + var b128 : base.x86_m128i + var p128 : base.x86_m128i + var k128 : base.x86_m128i if args.prev.length() == 0 { k128 = k128.create_mm_set1_epi8(a: 0xFE) @@ -124,7 +124,7 @@ pri func decoder.filter_3_distance_4_sse128!(curr: slice base.u8, prev: slice ba // Filter 4: Paeth. pri func decoder.filter_4_distance_3_sse128!(curr: slice base.u8, prev: slice base.u8), - choose cpu_arch >= sse128, + choose cpu_arch >= x86_sse42, { // See the comments in filter_4_distance_4_sse128 for an explanation of how // this works. That function's single loop is copied twice here, once with @@ -135,16 +135,16 @@ pri func decoder.filter_4_distance_3_sse128!(curr: slice base.u8, prev: slice ba var c : slice base.u8 var p : slice base.u8 - var x128 : base.sse128_i - var a128 : base.sse128_i - var b128 : base.sse128_i - var c128 : base.sse128_i - var p128 : base.sse128_i - var pa128 : base.sse128_i - var pb128 : base.sse128_i - var pc128 : base.sse128_i - var smallest128 : base.sse128_i - var z128 : base.sse128_i + var x128 : base.x86_m128i + var a128 : base.x86_m128i + var b128 : base.x86_m128i + var c128 : base.x86_m128i + var p128 : base.x86_m128i + var pa128 : base.x86_m128i + var pb128 : base.x86_m128i + var pc128 : base.x86_m128i + var smallest128 : base.x86_m128i + var z128 : base.x86_m128i // ยง The advance is 3, not 4. iterate (c = args.curr, p = args.prev)(length: 4, advance: 3, unroll: 2) { @@ -202,20 +202,20 @@ pri func decoder.filter_4_distance_3_sse128!(curr: slice base.u8, prev: slice ba } pri func decoder.filter_4_distance_4_sse128!(curr: slice base.u8, prev: slice base.u8), - choose cpu_arch >= sse128, + choose cpu_arch >= x86_sse42, { var c : slice base.u8 var p : slice base.u8 - var x128 : base.sse128_i - var a128 : base.sse128_i - var b128 : base.sse128_i - var c128 : base.sse128_i - var p128 : base.sse128_i - var pa128 : base.sse128_i - var pb128 : base.sse128_i - var pc128 : base.sse128_i - var smallest128 : base.sse128_i - var z128 : base.sse128_i + var x128 : base.x86_m128i + var a128 : base.x86_m128i + var b128 : base.x86_m128i + var c128 : base.x86_m128i + var p128 : base.x86_m128i + var pa128 : base.x86_m128i + var pb128 : base.x86_m128i + var pc128 : base.x86_m128i + var smallest128 : base.x86_m128i + var z128 : base.x86_m128i iterate (c = args.curr, p = args.prev)(length: 4, advance: 4, unroll: 2) { // Load the pixel from the row above.