[AMDGPU] Increase LDS to 320K on gfx1250

llvm/docs/AMDGPUUsage.rst

@@ -5598,6 +5598,8 @@ The fields used by CP for code objects before V3 also match those specified in
                                                        roundup(lds-size / (128 * 4))
                                                      GFX950
                                                        roundup(lds-size / (320 * 4))
+                                                     GFX125*
+                                                       roundup(lds-size / (256 * 4))
 
      24      1 bit   ENABLE_EXCEPTION_IEEE_754_FP    Wavefront starts execution
                      _INVALID_OPERATION              with specified exceptions

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -1548,7 +1548,7 @@ def FeatureGFX11 : GCNSubtargetFeatureGeneration<"GFX11",
 
 def FeatureGFX12 : GCNSubtargetFeatureGeneration<"GFX12",
   "gfx12",
-  [FeatureFP64, FeatureAddressableLocalMemorySize65536, FeatureMIMG_R128,
+  [FeatureFP64, FeatureMIMG_R128,
    FeatureFlatAddressSpace, Feature16BitInsts,
    FeatureInv2PiInlineImm, FeatureApertureRegs,
    FeatureCIInsts, FeatureGFX8Insts, FeatureGFX9Insts, FeatureGFX10Insts,
@@ -1977,6 +1977,7 @@ def FeatureISAVersion11_5_3 : FeatureSet<
 
 def FeatureISAVersion12 : FeatureSet<
   [FeatureGFX12,
+   FeatureAddressableLocalMemorySize65536,
    FeatureLDSBankCount32,
    FeatureDLInsts,
    FeatureDot7Insts,
@@ -2019,6 +2020,7 @@ def FeatureISAVersion12_50 : FeatureSet<
   [FeatureGFX12,
    FeatureGFX1250Insts,
    FeatureCUStores,
+   FeatureAddressableLocalMemorySize327680,
    FeatureCuMode,
    Feature64BitLiterals,
    FeatureLDSBankCount32,

llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp

@@ -1103,7 +1103,11 @@ void AMDGPUAsmPrinter::getSIProgramInfo(SIProgramInfo &ProgInfo,
   ProgInfo.DX10Clamp = Mode.DX10Clamp;
 
   unsigned LDSAlignShift;
-  if (STM.getFeatureBits().test(FeatureAddressableLocalMemorySize163840)) {
+  if (STM.getFeatureBits().test(FeatureAddressableLocalMemorySize327680)) {
+    // LDS is allocated in 256 dword blocks.
+    LDSAlignShift = 10;
+  } else if (STM.getFeatureBits().test(
+                 FeatureAddressableLocalMemorySize163840)) {
     // LDS is allocated in 320 dword blocks.
     LDSAlignShift = 11;
   } else if (STM.getFeatureBits().test(

llvm/lib/Target/AMDGPU/AMDGPUFeatures.td

@@ -30,6 +30,7 @@ class SubtargetFeatureAddressableLocalMemorySize <int Value> : SubtargetFeature<
 def FeatureAddressableLocalMemorySize32768 : SubtargetFeatureAddressableLocalMemorySize<32768>;
 def FeatureAddressableLocalMemorySize65536 : SubtargetFeatureAddressableLocalMemorySize<65536>;
 def FeatureAddressableLocalMemorySize163840 : SubtargetFeatureAddressableLocalMemorySize<163840>;
+def FeatureAddressableLocalMemorySize327680 : SubtargetFeatureAddressableLocalMemorySize<327680>;
 
 class SubtargetFeatureWavefrontSize <int ValueLog2> : SubtargetFeature<
   "wavefrontsize"#!shl(1, ValueLog2),

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

@@ -1160,6 +1160,8 @@ unsigned getAddressableLocalMemorySize(const MCSubtargetInfo *STI) {
     return 65536;
   if (STI->getFeatureBits().test(FeatureAddressableLocalMemorySize163840))
     return 163840;
+  if (STI->getFeatureBits().test(FeatureAddressableLocalMemorySize327680))
+    return 327680;
   return 0;
 }
 
@@ -3340,8 +3342,8 @@ bool isDPALU_DPP(const MCInstrDesc &OpDesc, const MCSubtargetInfo &ST) {
 }
 
 unsigned getLdsDwGranularity(const MCSubtargetInfo &ST) {
-  // Currently this is 128 for all subtargets
-  return 128;
+  return ST.hasFeature(AMDGPU::FeatureAddressableLocalMemorySize327680) ? 256
+                                                                        : 128;
 }
 
 bool isPackedFP32Inst(unsigned Opc) {

llvm/test/CodeGen/AMDGPU/extra-lds-size.ll

@@ -6,6 +6,8 @@
 ; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx950 < %s | FileCheck -check-prefix=GFX950-MESA %s
 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX1200-PAL %s
 ; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1200 < %s | FileCheck -check-prefix=GFX1200-MESA %s
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1250 < %s | FileCheck -check-prefix=GFX1250-PAL %s
+; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1250 < %s | FileCheck -check-prefix=GFX1250-MESA %s
 
 ; Check EXTRA_LDS_SIZE in SPI_SHADER_PGM_RSRC2_PS.
 
@@ -29,6 +31,11 @@
 ; GFX1200-MESA: .long 45100
 ; GFX1200-MESA-NEXT: .long 1024
 
+; GFX1250-PAL: '0x2c0b (SPI_SHADER_PGM_RSRC2_PS)': 0x200
+
+; GFX1250-MESA: .long 45100
+; GFX1250-MESA-NEXT: .long 512
+
 @lds = internal addrspace(3) global [4096 x i8] poison
 
 define amdgpu_ps void @global_store_saddr_uniform_ptr_in_vgprs(i32 %voffset) {

llvm/test/CodeGen/AMDGPU/lds-limit-diagnostics-gfx1250.ll

@@ -0,0 +1,13 @@
+; RUN: not llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s 2>&1 | FileCheck -check-prefix=ERROR %s
+
+; GFX1250 supports upto 320 KB LDS memory.
+; This is a negative test to check when the LDS size exceeds the max usable limit.
+
+; ERROR: error: <unknown>:0:0: local memory (327684) exceeds limit (327680) in function 'test_lds_limit'
+@dst = addrspace(3) global [81921 x i32] undef
+
+define amdgpu_kernel void @test_lds_limit(i32 %val) {
+  %gep = getelementptr [81921 x i32], ptr addrspace(3) @dst, i32 0, i32 100
+  store i32 %val, ptr addrspace(3) %gep
+  ret void
+}

llvm/test/CodeGen/AMDGPU/lds-size-hsa-gfx1250.ll

@@ -0,0 +1,72 @@
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1250 < %s | FileCheck -check-prefix=MESA %s
+
+; GFX1250 supports upto 320 KB configurable LDS memory.
+; This test checks the min and max size of LDS that can be allocated.
+
+@lds.i8 = addrspace(3) global i8 undef
+@lds.array.i8 = addrspace(3) global [327679 x i8] undef
+@lds.i16 = addrspace(3) global i16 undef
+@lds.array.i16 = addrspace(3) global [163839 x i16] undef
+@lds.i32 = addrspace(3) global i32 undef
+@lds.array.i32 = addrspace(3) global [81919 x i32] undef
+
+; GCN-LABEL: test_lds_i8:
+; GCN: .amdhsa_group_segment_fixed_size 1
+; GCN: ; LDSByteSize: 1 bytes/workgroup
+; MESA: granulated_lds_size = 1
+define amdgpu_kernel void @test_lds_i8(i8 %val) {
+  store i8 %val, ptr addrspace(3) @lds.i8
+  ret void
+}
+
+; GCN-LABEL: test_lds_i16:
+; GCN: .amdhsa_group_segment_fixed_size 2
+; GCN: ; LDSByteSize: 2 bytes/workgroup
+; MESA: granulated_lds_size = 1
+define amdgpu_kernel void @test_lds_i16(i16 %val) {
+  store i16 %val, ptr addrspace(3) @lds.i16
+  ret void
+}
+
+; GCN-LABEL: test_lds_i32:
+; GCN: .amdhsa_group_segment_fixed_size 4
+; GCN: ; LDSByteSize: 4 bytes/workgroup
+; MESA: granulated_lds_size = 1
+define amdgpu_kernel void @test_lds_i32(i32 %val) {
+  store i32 %val, ptr addrspace(3) @lds.i32
+  ret void
+}
+
+; GCN-LABEL: test_lds_array_i8:
+; GCN: .amdhsa_group_segment_fixed_size 327680
+; GCN: ; LDSByteSize: 327680 bytes/workgroup
+; MESA: granulated_lds_size = 320
+define amdgpu_kernel void @test_lds_array_i8() {
+  %gep = getelementptr inbounds [327679 x i8], ptr addrspace(3) @lds.array.i8, i32 0, i32 5
+  %val = load i8, ptr addrspace(3) %gep
+  store i8 %val, ptr addrspace(3) @lds.i8
+  ret void
+}
+
+; GCN-LABEL: test_lds_array_i16:
+; GCN: .amdhsa_group_segment_fixed_size 327680
+; GCN: ; LDSByteSize: 327680 bytes/workgroup
+; MESA: granulated_lds_size = 320
+define amdgpu_kernel void @test_lds_array_i16() {
+  %gep = getelementptr inbounds [163839 x i16], ptr addrspace(3) @lds.array.i16, i32 0, i32 10
+  %val = load i16, ptr addrspace(3) %gep
+  store i16 %val, ptr addrspace(3) @lds.i16
+  ret void
+}
+
+; GCN-LABEL: test_lds_array_i32:
+; GCN: .amdhsa_group_segment_fixed_size 327680
+; GCN: ; LDSByteSize: 327680 bytes/workgroup
+; MESA: granulated_lds_size = 320
+define amdgpu_kernel void @test_lds_array_i32() {
+  %gep = getelementptr inbounds [81919 x i32], ptr addrspace(3) @lds.array.i32, i32 0, i32 20
+  %val = load i32, ptr addrspace(3) %gep
+  store i32 %val, ptr addrspace(3) @lds.i32
+  ret void
+}

llvm/test/CodeGen/AMDGPU/lds-size-pal-gfx1250.ll

@@ -0,0 +1,61 @@
+; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1250 < %s | FileCheck -check-prefix=PAL %s
+
+; GFX1250 supports upto 320 KB configurable LDS memory.
+; This test checks the min and max size of LDS that can be allocated.
+
+; PAL: .shader_functions:
+; PAL: test_lds_array_i16:
+; PAL: .lds_size:       0x50000
+; PAL: test_lds_array_i32:
+; PAL: .lds_size:       0x50000
+; PAL: test_lds_array_i8:
+; PAL: .lds_size:       0x50000
+; PAL: test_lds_i16:
+; PAL: .lds_size:       0x2
+; PAL: test_lds_i32:
+; PAL: .lds_size:       0x4
+; PAL: test_lds_i8:
+; PAL: .lds_size:       0x1
+
+@lds.i8 = addrspace(3) global i8 undef
+@lds.array.i8 = addrspace(3) global [327679 x i8] undef
+@lds.i16 = addrspace(3) global i16 undef
+@lds.array.i16 = addrspace(3) global [163839 x i16] undef
+@lds.i32 = addrspace(3) global i32 undef
+@lds.array.i32 = addrspace(3) global [81919 x i32] undef
+
+define amdgpu_gfx void @test_lds_i8(i8 %val) {
+  store i8 %val, ptr addrspace(3) @lds.i8
+  ret void
+}
+
+define amdgpu_gfx void @test_lds_i16(i16 %val) {
+  store i16 %val, ptr addrspace(3) @lds.i16
+  ret void
+}
+
+define amdgpu_gfx void @test_lds_i32(i32 %val) {
+  store i32 %val, ptr addrspace(3) @lds.i32
+  ret void
+}
+
+define amdgpu_gfx void @test_lds_array_i8() {
+  %gep = getelementptr inbounds [327679 x i8], ptr addrspace(3) @lds.array.i8, i32 0, i32 5
+  %val = load i8, ptr addrspace(3) %gep
+  store i8 %val, ptr addrspace(3) @lds.i8
+  ret void
+}
+
+define amdgpu_gfx void @test_lds_array_i16() {
+  %gep = getelementptr inbounds [163839 x i16], ptr addrspace(3) @lds.array.i16, i32 0, i32 10
+  %val = load i16, ptr addrspace(3) %gep
+  store i16 %val, ptr addrspace(3) @lds.i16
+  ret void
+}
+
+define amdgpu_gfx void @test_lds_array_i32() {
+  %gep = getelementptr inbounds [81919 x i32], ptr addrspace(3) @lds.array.i32, i32 0, i32 20
+  %val = load i32, ptr addrspace(3) %gep
+  store i32 %val, ptr addrspace(3) @lds.i32
+  ret void
+}